Lecture Notes in Networks and Systems 211

Saad Motahhir
Badre Bossoufi   Editors

Digital
Technologies
and Applications
Proceedings of ICDTA 21, Fez, Morocco
Lecture Notes in Networks and Systems

Volume 211

Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland

Advisory Editors
Fernando Gomide, Department of Computer Engineering and Automation—DCA,
School of Electrical and Computer Engineering—FEEC, University of Campinas—
UNICAMP, São Paulo, Brazil
Okyay Kaynak, Department of Electrical and Electronic Engineering,
Bogazici University, Istanbul, Turkey
Derong Liu, Department of Electrical and Computer Engineering, University
of Illinois at Chicago, Chicago, USA; Institute of Automation, Chinese Academy
of Sciences, Beijing, China
Witold Pedrycz, Department of Electrical and Computer Engineering,
University of Alberta, Alberta, Canada; Systems Research Institute,
Polish Academy of Sciences, Warsaw, Poland
Marios M. Polycarpou, Department of Electrical and Computer Engineering,
KIOS Research Center for Intelligent Systems and Networks, University of Cyprus,
Nicosia, Cyprus
Imre J. Rudas, Óbuda University, Budapest, Hungary
Jun Wang, Department of Computer Science, City University of Hong Kong,
Kowloon, Hong Kong
The series “Lecture Notes in Networks and Systems” publishes the latest
developments in Networks and Systems—quickly, informally and with high quality.
Original research reported in proceedings and post-proceedings represents the core
of LNNS.
Volumes published in LNNS embrace all aspects and subfields of, as well as new
challenges in, Networks and Systems.
The series contains proceedings and edited volumes in systems and networks,
spanning the areas of Cyber-Physical Systems, Autonomous Systems, Sensor
Networks, Control Systems, Energy Systems, Automotive Systems, Biological
Systems, Vehicular Networking and Connected Vehicles, Aerospace Systems,
Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power Systems,
Robotics, Social Systems, Economic Systems and other. Of particular value to both
the contributors and the readership are the short publication timeframe and the
world-wide distribution and exposure which enable both a wide and rapid
dissemination of research output.
The series covers the theory, applications, and perspectives on the state of the art
and future developments relevant to systems and networks, decision making,
control, complex processes and related areas, as embedded in the fields of
interdisciplinary and applied sciences, engineering, computer science, physics,
economics, social, and life sciences, as well as the paradigms and methodologies
behind them.
Indexed by SCOPUS, INSPEC, WTI Frankfurt eG, zbMATH, SCImago.
All books published in the series are submitted for consideration in Web of
Science.

More information about this series at http://www.springer.com/series/15179

Saad Motahhir Badre Bossoufi
•

Editors

Digital Technologies
and Applications
Proceedings of ICDTA 21, Fez, Morocco

123
Editors
Saad Motahhir Badre Bossoufi
ENSA Faculty of Sciences
Sidi Mohamed Ben Abdellah University Sidi Mohamed Ben Abdellah University
Fez, Morocco Fez, Morocco

ISSN 2367-3370 ISSN 2367-3389 (electronic)

Lecture Notes in Networks and Systems
ISBN 978-3-030-73881-5 ISBN 978-3-030-73882-2 (eBook)
https://doi.org/10.1007/978-3-030-73882-2
© The Editor(s) (if applicable) and The Author(s), under exclusive license
to Springer Nature Switzerland AG 2021
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
We are honored to dedicate the proceedings
of ICDTA’21 to all the participants and
committees of ICDTA’21.
Foreword

It is with deep satisfaction that I write this Foreword to the proceedings of the
ICDTA’21 organized by École Nationale des Sciences Appliquées and Faculté des
sciences which belong to SMBA University, Fez, Morocco, on January 29 and 30,
2021.
This conference was bringing together researchers, academics, and professionals
from all over the world, experts in digital technologies and their applications. This
conference particularly encouraged the interaction of research students and devel-
oping academics with the more established academic community in an informal
setting to present and to discuss new and current work. The papers contributed the
most recent scientific knowledge known in the field of digital technologies and their
applications. Their contributions helped to make the conference as outstanding as it
has been. The organizing and technical program committees put much effort into
ensuring the success of the day-to-day operation of the meeting.
We hope that this book proceedings will further stimulate research in digital
technologies such as artificial intelligence, Internet of things, embedded systems,
network technology, information processing and their applications, in several areas
including hybrid vehicles, renewable energy, robotic, COVID-19, etc.
We feel honored and privileged to serve the best recent developments to you
through this exciting book proceedings.
We thank all authors and participants for their contributions.

Saad Motahhir
Badre Bossoufi

vii
Preface

This conference proceedings volume contains the written versions of most of the
contributions presented during the ICDTA’21. The conference provided a setting
for discussing recent developments in a wide variety of topics including artificial
intelligence, Internet of things, embedded systems, network technology, informa-
tion processing and their applications, in several areas such as hybrid vehicles,
renewable energy, robotic, and COVID-19. The conference has been a good
opportunity for participants from various destinations to present and discuss topics
in their respective research areas.
ICDTA’21 tends to collect the latest research results and applications on digital
technologies and their applications. It includes a selection of 166 papers from 328
papers submitted to the conference from universities and industries all over the
world. All of accepted papers were subjected to strict peer reviewing by 2–4 expert
referees. The papers have been selected for this volume because of quality and the
relevance to the conference.
ICDTA’21 would like to express our sincere appreciation to all authors for their
contributions to this book. We would like to extend our thanks to all the referees for
their constructive comments on all papers; especially, we would like to thank
organizing committee for their hardworking. Finally, we would like to thank the
Springer publications for producing this volume.

Saad Motahhir
Badre Bossoufi

ix
Acknowledgments

We request the pleasure of thanking you for taking part in the first edition of the
International Conference on Digital Technologies and Applications (ICDTA’21).
We are very grateful for your support, so thank you everyone for bringing your
expertise and experience around the conference and engaging in such fruitful,
constructive, and open exchanges throughout the two days of the ICDTA’21.
We would like to extend our deepest thanks and gratitude to all the speakers for
accepting to join us from different countries. Thank you for being such wonderful
persons and speakers. Again, thanks for sharing your insight, knowledge, and
experience.
Of course, this event could not be that successful without the effort of the
organizing and technical program committees. Therefore, Pr. Badre and I would
like to express our sincere appreciation to all of you who generously helped us.
We would like to especially thank all the participants for the confidence and trust
you have placed in our conference. We hope we lived up to your highest
expectations.
Finally, our acknowledgement would be incomplete without thanking the big-
gest source of support. Our deepest gratitude goes to Pr. REDOUANE MRABET,
President of Sidi Mohamed Ben Abdellah University, Prof.
ABDERRAHIM LAHRACH, Director of the National School of Applied Sciences,
Prof. MOHAMMED BELMLIH, Dean of the Faculty of Sciences,
Professor ABDELMAJID SAKA, Deputy Director of the National School of
Applied Sciences, and Professor MOHAMMED ELHASSOUNI, Vice Dean of
Faculty of Sciences. Thank you all for your support and for being all-time open to
hosting such events.

Saad Motahhir
Badre Bossoufi

xi
Contents

Artificial Intelligence, Machine Learning and Data Analysis

New Reconfiguration of the Radial Distribution Network by Using
the Chaotic Mapping and the Success Rate to Improve
the PSO Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Meriem M’dioud, Ismail ELkafazi, and Rachid Bannari
Convolutional Neuronal Networks for Tumor Regions Detection
in Histopathology Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Mohammed Lamine Benomar, Nesma Settouti, Rudan Xiao,
Damien Ambrosetti, and Xavier Descombes
Early Detection of Wheat Diseases in Morocco by Convolutional
Neutral Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Kenza Aitelkadi, Noureddine Chtaina, Saber Bakouri,
and Mohamed Belebrik
Genetic Algorithm Design Optimization for Non-standard
Spur Gears . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Samya Belarhzal and El Mostapha Boudi
A Robust Method for Face Classification Based on Binary Genetic
Algorithm Combined with NSVC Classifier . . . . . . . . . . . . . . . . . . . . . . 47
M. Ngadi, A. Amine, B. Nassih, Y. Azdoud, and A. El-Attar
A Proposed Solution to Road Traffic Accidents Based on Fuzzy
Logic Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Halima Drissi Touzani, Sanaa Faquir, Saloua Senhaji, and Ali Yahyaouy
Multi-objective Archimedes Optimization Algorithm for Optimal
Allocation of Renewable Energy Sources in Distribution Networks . . . . 65
Ahmad Eid and Hassan El-Kishky

xiii
xiv Contents

The Recurrent Neural Network for Program Synthesis . . . . . . . . . . . . . 77

Achraf Berrajaa and El Hassane Ettifouri
MaskNet: CNN for Real-Time Face Mask Detection Based on Deep
Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Ismail Nasri, Mohammed Karrouchi, Hajar Snoussi,
Abdelhafid Messaoudi, and Kamal Kassmi
Machine Learning System for Fraud Detection. A Methodological
Approach for a Development Platform . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Salma El Hajjami, Jamal Malki, Mohammed Berrada, Harti Mostafa,
and Alain Bouju
U-Net: Deep Learning for Extracting Building Boundary Collected
by Drone of Agadir’s Harbor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Tarik Chafiq, Hayat Hachimi, Mohammed Raji, and Soufiane Zerraf
Analysis and Classification of Plant Diseases Based
on Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Assia Ennouni, Noura Ouled Sihamman, My Abdelouahed Sabri,
and Abdellah Aarab
A Novel Approach to Data Augmentation for Document Image
Classification Using Deep Convolutional Generative
Adversarial Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Aissam Jadli, Mustapha Hain, and Abderrahman Jaize
Knowledge Driven Paradigm for Anomaly Detection from Tweets
Using Gated Recurrent Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
S. Manaswini, Gerard Deepak, and A. Santhanavijayan
Stacked Deep Learning LSTM Model for Daily Solar Power Time
Series Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Soufiane Gaizen, Ouafia Fadi, and Ahmed Abbou
A Comparison Study of Machine Learning Methods for Energy
Consumption Forecasting in Industry . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Mouad Bahij, Moussa Labbadi, Mohamed Cherkaoui, Chakib Chatri,
and Soufian Lakrit
The Review of Objectives, Methods, Tools, and Algorithms
for Educational Data Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Mohamed Timmi, Adil Jeghal, Said EL Garouani, and Ali Yahyaouy
Brain-Computer Interface: A Novel EEG Classification for Baseline
Eye States Using LGBM Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
Said Abenna, Mohammed Nahid, and Abderrahim Bajit
A Light Arabic POS Tagger Using a Hybrid Approach . . . . . . . . . . . . . 199
Khalid Tnaji, Karim Bouzoubaa, and Si Lhoussain Aouragh
Contents xv

RSEPUA: A Recommender System for Early Predicting

University Admission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
Inssaf El Guabassi, Zakaria Bousalem, Rim Marah, and Aimad Qazdar
Forecasting Students’ Academic Performance Using Different
Regression Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Inssaf El Guabassi, Zakaria Bousalem, Rim Marah, and Aimad Qazdar
Computational Analysis of Human Navigation Trajectories
in a Spatial Memory Locomotor Task . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Ihababdelbasset Annaki, Mohammed Rahmoune, Mohammed Bourhaleb,
Jamal Berrich, Mohamed Zaoui, Alexandre Castilla Ferro,
and Alain Berthoz
An Electrocardiogram Data Compression-Decompression Technique
Based on the Integration Filtering of Peaks’ Amplitude Signal . . . . . . . 245
Skander Bensegueni

Internet of Things, Embedded Systems, Blockchain and Security

Narrowband IoT Evolution Towards 5G Massive MTC:
Complete Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
Adil Abou El Hassan, Abdelmalek El Mehdi, and Mohammed Saber
A Fuzzy Ontology Driven Integrated IoT Approach
for Home Automation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
Levin Varghese, Gerard Deepak, and A. Santhanavijayan
Towards an IoT/Big-Data Platform for Data Measurements,
Collection and Processing in Micro-grid Systems . . . . . . . . . . . . . . . . . . 279
Abdellatif Elmouatamid, Youssef Alidrissi, Radouane Ouladsine,
Mohamed Bakhouya, Najib Elkamoun, Mohammed Khaidar,
and Khalid Zine-Dine
Smart Hospitals and Cyber Security Attacks . . . . . . . . . . . . . . . . . . . . . 291
Yassine Chahid, Mohammed Benabdellah, and Nabil Kannouf
Design of an Automatic Control and IoT Monitoring System
for Dual Axis Solar Tracker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
Fatima Cheddadi, Hafsa Cheddadi, Youssef Cheddadi, Fatima Errahimi,
and Ikram Saber
Analysis Jamming Attack Against the Protocol S-MAC
in IoT Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
Imane Kerrakchou, Sara Chadli, Mohamed Emharraf,
and Mohammed Saber
Simulation and Analysis of Jamming Attack in IoT Networks . . . . . . . . 323
Imane Kerrakchou, Sara Chadli, Amina Kharbach, and Mohammed Saber
xvi Contents

A Brief Survey on Internet of Things (IoT) . . . . . . . . . . . . . . . . . . . . . . 335

Fatima Zahra Fagroud, Lahbib Ajallouda, El Habib Ben Lahmar,
Hicham Toumi, Ahmed Zellou, and Sanaa El Filali
Design and Implementation of Smart Irrigation System Based
on the IoT Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
Asmae Hafian, Mohammed Benbrahim, and Mohammed Nabil Kabbaj
Lightweight Key Exchange Algorithms for MQTT Protocol
in Its Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
Anwar D. Alhejaili and Omar H. Alhazmi
IoT Design and Realization of a Supervision Device for Photovoltaic
Panels Using an Approach Based on Radiofrequency Technology . . . . . 365
Zaidan Didi and Ikram El Azami
ESP8266 Wireless Indoor Positioning System using Fingerprinting
and Trilateration Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
Safae El Abkari, Jamal El Mhamdi, Abdelilah Jilbab,
and El Hassan El Abkari
High Performance Predictive Control for Permanent Magnet
Synchronous Machine Drive: FPGA-Based Implementation . . . . . . . . . . 387
Badre Bossoufi and Ahmed Lagrioui
High-Level Synthesis Implementation of Monocular SLAM
on Low-Cost Parallel Platforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399
Ayoub Mamri, Mohamed Abouzahir, Mustapha Ramzi,
and Mohamed Sbihi
Improving Multispectral Image Processing for Real-Time
Agricultural Indices Tracking Using an Embedded System . . . . . . . . . . 411
Amine Saddik, Rachid Latif, and Abdelhafid El Ouardi
Design and Implementation of an Open Source and Low-Cost
Nanosatellite Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
Ahmed Hanafi, Aziz Derouich, Mohammed Karim, and Assiya Lemmassi
Design and Realization of Fire Safety System for Controlling
and Monitoring a Siren Using Arduino Uno . . . . . . . . . . . . . . . . . . . . . 433
Abdennabi Morchid, Mustapha El Alaoui, Rachid El Alami,
Hassan Qjidaa, Karim El Khadiri, and Youness Mehdaoui
Water-Purifying Distiller with a Cooling System Controlled
by a Photovoltaic Panel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
Kamilia Mounich, Aicha Wahabi, and Mohammed Chafi
Security at the Application Layer and the Physical Layer in 5G . . . . . . 457
Hafida Amgoune and Tomader Mazri
Contents xvii

Blockchain Based Security and Privacy in VANETs . . . . . . . . . . . . . . . 469

Sana Iqbal, Narmeen Zakaria Bawany, and Ayesha Zulfiqar
Dynamic Data Protection for Mobile Agents:
XACML/ABAC Policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483
Hassan Razouki
Securing the Storage of Passwords Based on the MD5
HASH Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495
Hamza Touil, Nabil El Akkad, and Khalid Satori
Is Blockchain Technology Applicable in Small
and Medium-Sized Enterprises? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
Zakariya Chabani, Salima Hamouche, and Raed Said
Anomaly Detection from Log Files Using Multidimensional
Analysis Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
Yassine Azizi, Mostafa Azizi, and Mohamed Elboukhari

Web-Based Techniques and E-Learning

A Review on Content Based Recommender Systems in Tourism . . . . . . 527
Asmae Bentaleb, Younès El Bouzekri El Idrissi, and Ayoub Ait Lahcen
Towards Semantic Web Services Density Clustering Technique . . . . . . . 543
Naoufal El Allali, Mourad Fariss, Hakima Asaidi,
and Mohamed Bellouki
KSTAR: A Knowledge Based Approach for Socially Relevant Term
Aggregation for Web Page Recommendation . . . . . . . . . . . . . . . . . . . . . 555
Deepak Surya, Gerard Deepak, and A. Santhanavijayan
MKSTD: Multiple Knowledge Sourcing Approach for Topic
Directories Construction and Linking Web Pages Using
Fusion Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
Gerard Deepak and A. Santhanavijayan
A Fuzzy Near Neighbors Approach for Arabic Text Categorization
Based on Web Mining Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575
Mounir Gouiouez
Recommendation System Based on Semantic Web Approach . . . . . . . . 585
Sara Belattar, Otman Abdoun, and Haimoudi El Khatir
Sentiment Analysis of Moroccan Tweets Using Text Mining . . . . . . . . . 597
Moncef Garouani, Hanae Chrita, and Jamal Kharroubi
Contribution of Digital Collaboration and E-Learning
to the Implementation of Smart Mobility in Morocco . . . . . . . . . . . . . . 609
Mohammed Mouhcine Maaroufi, Laila Stour, and Ali Agoumi
xviii Contents

MOOCs Semantic Interoperability: Towards Unified

and Pedagogically Enriched Model for Building a Linked
Data Repository . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 621
Hanane Sebbaq and Nour-eddine El Faddouli
IBN TOFAIL UNIVERSITY: From Classical University
to Smart University . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633
Khawla Mazwa and Tomader Mazri
Collaborative Metacognitive Technique in Smart Tutoring System . . . . 643
Hibbi Fatima-Zohra, Abdoun Otman, and Haimoudi Elkhatir
Enhancing Education System with a Q&A Chatbot: A Case Based
on Open edX Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 655
Rachid Karra and Abdelali Lasfar
A Multi-agent and Content-Based Course Recommender System
for University E-learning Platforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 663
Meryem Amane, Karima Aissaoui, and Mohammed Berrada

Image and Information Processing

Sentiment Classification for Moroccan Arabic Using
the Lexicon-Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675
Chorouk Ferroud and Abdeljalil Elouardighi
Forecasting Tourism Demand in Marrakech with SQD-PCA-SVR . . . . 685
Houria Laaroussi, Fatima Guerouate, and Mohamed Sbihi
Toward Improving Arabic Text Preprocessing
in Sentiment Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 695
Mohcine Maghfour and Abdeljalil Elouardighi
IIMDR: Intelligence Integration Model for Document Retrieval . . . . . . 707
S. Aditya, P. Muhil Aditya, Gerard Deepak, and A. Santhanavijayan
Effective Application of Information Technology Tools for Real-Time
Project Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 719
Mahmoud Ershadi, Marcus Jefferies, Peter Davis,
and Mohammad Mojtahedi
HCODF: Hybrid Cognitive Ontology Driven Framework for Socially
Relevant News Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
V. Adithya, Gerard Deepak, and A. Santhanavijayan
Understanding Customer Adoption of E-banking . . . . . . . . . . . . . . . . . . 741
Jaride Chama, Taqi Ahmed, and El Hachimi Mohamed Amine
Amalgamation of Novel Objective Function and Multi-sink Solution
for a Reliable RPL in High Traffic Monitoring Applications . . . . . . . . . 749
Abdelhadi Eloudrhiri Hassani, Aicha Sahel, and Abdelmajid Badri
Contents xix

Wireless Sensor Network Fault Detection and Isolation

for Smart Irrigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 759
Nassima Jihani, Mohammed Nabil Kabbaj, and Mohammed Benbrahim
Multi-resolution Texture Analysis for Osteoporosis Classification . . . . . 769
Meriem Mebarkia, Abdallah Meraoumia, Lotfi Houam,
Seddik Khemaissia, and Rachid Jennane
High Performance Decoding by Combination of the Hartmann
Rudolph Decoder and Soft Decision Decoding by Hash Techniques . . . . 781
Hamza Faham, My Seddiq El Kasmi Alaoui, Saïd Nouh,
and Mohamed Azzouazi
Automated Assessment of Question Quality on Online
Community Forums . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 791
Harish Rithish, Gerard Deepak, and A. Santhanavijayan
QFRDBF: Query Facet Recommendation Using Knowledge Centric
DBSCAN and Firefly Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 801
Deepak Surya, Gerard Deepak, and A. Santhanavijayan
SentiFusion (SF): Sentiment Analysis of Twitter Using
Fusion Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813
Sampurn Anand, Chaya Vijaykumar, Gerard Deepak,
and A Santhana Vijayan
Using Internet of Things to Increase Efficient Collaboration
in PLM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825
Narjiss Tilioua, Fatima Bennouna, and Zakaria Chalh
ECC Image Encryption Using Matlab Simulink Blockset . . . . . . . . . . . 835
Sara Chillali and Lahcen Oughdir
Development of Large Chaotic S-boxes for Image Encryption . . . . . . . . 847
Younes Qobbi, Abdeltif Jarjar, Mohamed Essaid, and Abdelhamid Benazzi
Evaluation of Feature Extraction Methods Combined with Support
Vector Machines for Powerline Component Recognition
in Aerial Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 859
Jamila Garfaf, Lamyae Fahmani, and Hicham Medromi
A Novel Brain Tumor Detection Approach Based on Fuzzy C-means
and Marker Watershed Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 871
Hanae Moussaoui, Mohamed Benslimane, and Nabil El Akkad
Image Retrieval Based on MPEG-7 Feature Selection Using
Meta-heuristic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 881
Naoufal Machhour and M’barek Nasri
xx Contents

A Powerful and Efficient Method of Image Segmentation Based

on Random Forest Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 893
Zahra Faska, Lahbib Khrissi, Khalid Haddouch, and Nabil EL Akkad

Digital Technologies in Battle Against COVID-19

Towards Overlapped Objects Recognition Using Capsule Network . . . . 907
Merieme Mansouri, Samia Benabdellah Chaouni, and Said Jai Andaloussi
The Spread of the Corona Virus Disease (Covid-19) and the Launch
of 5G Technology in China: What Relationship . . . . . . . . . . . . . . . . . . . 919
Abdelhakim Moutaouakil, Younes Jabrane, Abdelati Reha,
and Abdelaziz Koumina
Analysis of Covid-19 Impact on Gender (Male and Female) in South
Asian Country Pakistan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 925
Saba Malik, Mubbra Habib, Mehmood Ahmed Husnain Hashmi,
Muhammad Tariq Saeed, Anwaar ul Huda, Syeda Anam Zahra,
Muhammad Nisar, Muhammad Sajid, Shahbaz Ahmad Shahzad,
Saddam Hussain, and Muhammad Sakandar Majid
Deep Residual Convolutional Neural Network Based Detection
of Covid-19 from Chest-X-Ray Images . . . . . . . . . . . . . . . . . . . . . . . . . . 939
Valaparla Rohini, M. Sobhana, Ch. Smitha Chowdary, Mukesh Chinta,
and Deepa Venna
A New Optimized Approach to Resolve a Combinatorial Problem:
CoronaVirus Optimization Algorithm and Self-organizing Maps . . . . . . 947
Omayma EL Majdoubi, Farah Abdoun, and Otman Abdoun
Study of a Dual-Function Intelligent Electronic Pin to Help
Compliance with Security Measures Related to Covid-19 . . . . . . . . . . . . 959
Laince Pierre Moulebe, Abdelwahed Touati, Eric Obar Akpoviroro,
and Nabila Rabbah
DenTcov: Deep Transfer Learning-Based Automatic Detection
of Coronavirus Disease (COVID-19) Using Chest X-ray Images . . . . . . 967
Youssra El Idrissi El-Bouzaidi and Otman Abdoun

Network Technology
Explainable Deep Learning Model for COVID-19 Screening in Chest
CT Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 981
Mostafa El Habib Daho, Amin Khouani, Mohammed El Amine Lazouni,
and Sidi Ahmed Mahmoudi
Contents xxi

A Novel Tool for Automating the Trace File Analysis in Vehicular

Ad Hoc Networks with Multi-agents Method . . . . . . . . . . . . . . . . . . . . . 991
Sanaa Achchab, Souad El Houssaini, Amal Tmiri, Souad Ajjaj,
and Mohammed-Alamine El Houssaini
A Fuzzy On/Off Switching Strategy for Green Cellular Networks . . . . . 1003
Soufiane Dahmani, Mohammed Gabli, Abdelhafid Serghini,
and El Bekkaye Mermri
Multi-band Planar-Inverted-F-Antenna Design for WIFI WIMAX
and WLAN Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1013
Asma Khabba, Layla Wakrim, Saida Ibnyaich,
and Moha M’Rabet Hassani
A New Approach for Evaluating the Performance of AODV Routing
Protocol in VANETs Using the Plackett-Burman Method . . . . . . . . . . . 1021
Souad Ajjaj, Souad El Houssaini, Mustapha Hain,
and Mohammed-Alamine El Houssaini
Design and Simulation of a New Dual Band Fractal Antenna
for TM Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1033
Abdelhakim Moutaouakil, Younes Jabrane, Abdelati Reha,
Abdelaziz Koumina, and Nadia El Makoussi
Design and Simulation of Three Bands Planar Inverted F Antenna
Array for 5G Communication Systems . . . . . . . . . . . . . . . . . . . . . . . . . 1039
Sara Arsalane, Nabil Arsalane, and Mounir Rifi
RFID System for Hospital Monitoring and Medication Tracking
Using Digital Signature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1051
Safae El Abkari, Soufiane Kaissari, Jamal El Mhamdi, Abdelilah Jilbab,
and El Hassan El Abkari
Building a Domain Ontology for the Construction Industry:
Towards Knowledge Sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1061
Hayat El Asri, Fatine Jebbor, and Laila Benhlima
Optimization of the Ad Hoc Network by Using Hybridization
of Genetic Algorithm with a Two-Optimization Algorithm . . . . . . . . . . 1073
Hala Khankhour, Otman Abdoun, and Jâafar Abouchabaka
Design of a Microstrip Antenna Two-Slot for Fifth Generation
Applications Operating at 27.5 GHz . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1081
Salah-Eddine Didi, Imane Halkhams, Mohammed Fattah, Younes Balboul,
Said Mazer, and Moulhime El Bekkali
Performance Improvement of Corner-Truncated Sierpinski Carpet
Fractal Antenna Using U-tree Slots for Vehicular Communications
and Other Bands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1091
Fatima Ez-Zaki, Abdelilah Ghammaz, and Hassan Belahrach
xxii Contents

A Novel System Based V2V Communications to Prevent Road

Accidents in Morocco . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1101
Zakaria Sabir and Aouatif Amine
Detection of Underground Cavities Using Electromagnetic GPR
Method in Bhalil City (Morocco) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1111
Oussama Jabrane, Driss El Azzab, Mahjoub Himi, Mohammed Charroud,
and Mohammed El Gettafi
An Adapted Routing Protocol for Mobile Edge Computing . . . . . . . . . . 1121
Ali Ouacha

Advanced Technologies in Energy and Electrical Engineering

Improved Geographic Routing Protocol for Wireless
Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1135
Kenza Redjimi, Mohammed Redjimi, and Mehdi Boulaiche
Supervision and Monitoring of Photovoltaic Systems Using Siemens
PLC and HMI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1147
Ahmed Bouraiou, Ammar Neçaibia, Saad Motahhir,
Mohammed Salah Bouakkaz, Issam Attoui, Nadir Boutasseta,
Rachid Dabou, and Abderrezzaq Ziane
Maximum Power Point Tracking Using SEPIC Converter and Double
Diode Solar Cell Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1159
Khalid Chennoufi and Mohammed Ferfra
Global Maximum Power Point Tracking Using Genetic Algorithm
Combined with PSO Tuned PID Controller . . . . . . . . . . . . . . . . . . . . . . 1171
Mohammed Salah Bouakkaz, Ahcene Boukadoum, Omar Boudebbouz,
Abdel Djabar Bouchaala, Nadir Boutasseta, Issam Attoui,
Ahmed Bouraiou, and Saad Motahhir
Experimental Validation of a Neuro-Fuzzy Model of Open
Circuit Voltage and Maximum Power Voltage Case
of Amorphous Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1181
Brahim Bouachrine, Mustapha Kourchi, Driss Yousfi, Kaoutar Dahmane,
Mhand Oubella, and Mohamed Ajaamoum
Simulation of Natural Ventilation on Building with Solar Chimney
Under Climatic Conditions of Errachidia Morocco Zone . . . . . . . . . . . . 1191
Herouane Aboubakr, Thami Ait Taleb, and Mourad Taha Janan
Investigating the Effect of the Air Gap of a Solar Air Heater Intended
for an Indirect Drying System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1205
Mourad Salhi, Dounia Chaatouf, Benyounes Raillani, Samir Amraqui,
and Ahmed Mezrhab
Contents xxiii

Experimental Investigation of Efficiency and Dynamic Losses

for a Constructed Solar Boost Converter . . . . . . . . . . . . . . . . . . . . . . . . 1213
Weam El Merrassi, Abdelouahed Abounada, and Mohamed Ramzi
Improvement of the Non-linear Control Strategy of a Wind Turbine
by a High-Gain Observer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1223
Azeddine Loulijat, Najib Ababssi, and Mohamed Makhad
Optimization of DFIG Wind Turbine Power Quality Through
Adaptive Fuzzy Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1235
Hamid Chojaa, Aziz Derouich, Mohammed Taoussi, Othmane Zamzoum,
and Mourad Yessef
Use of a Doubly-Fed Induction Generator for the Conversion
of Wind Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1245
Wahabi Aicha, Elmoudden Abdelhadi, and Senhaji Rhazi Kaoutar
Improvement of the Wind Power System Based on the PMSG . . . . . . . 1255
Btissam Majout, Badre Bossoufi, Mohammed Karim,
and Chakib El Bekkali
Control of Wind Water Pumping Using Input-Output Feedback
Linearization Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1267
Atarsia Loubna, Toufouti Riad, and Meziane Salima
Improved Hybrid Control Strategy of the Doubly-Fed Induction
Generator Under a Real Wind Profile . . . . . . . . . . . . . . . . . . . . . . . . . . 1279
Mourad Yessef, Badre Bossoufi, Mohammed Taoussi, Ahmed Lagrioui,
and Hamid Chojaa
Flow-Oriented Control Design of Wind Power Generation System
Based on Permanent Magnet Synchronous Generator . . . . . . . . . . . . . . 1291
Nada Zine Laabidine, Chakib El Bakkali, Karim Mohammed,
and Badre Bossoufi
Households Energy Consumption Forecasting with Echo
State Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1305
Wadie Bendali, Mohammed Boussetta, Ikram Saber, and Youssef Mourad
Management of Multi-agents in a Smart Grid Network
with the Python Using the Contract Net Protocol . . . . . . . . . . . . . . . . . 1317
Saber Ikram, El Bachtiri Rachid, Bendali Wadie, Boussetta Mohammed,
El hammoumi Karima, and Cheddadi Fatima
Smart Grid Operation Using Hybrid AI Systems . . . . . . . . . . . . . . . . . . 1327
Chakib Alaoui and Hajar Saikouk
xxiv Contents

Economic Comparison Between Two Hybrid Systems

(Wind-Hydrogen) and (Wind-Hydroelectric) for Electricity
Production in Socotra, Yemen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1341
Saif Serag, Outhman Elbakkali, and Adil Echchelh
A New Hybrid System Used to Connect PV to a Microgrid . . . . . . . . . . 1353
Alexandru Dusa, Petru Livinti, Daniel Ciprian Balanuta,
and Gelu Gurguiatu
Coordinated Control and Optimization Dispatch of a Hybrid
Microgrid in Grid Connected Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1365
Marouane Lagouir, Abdelmajid Badri, and Yassine Sayouti
Smart Energy Management System: SCIM Diagnosis and Failure
Classification and Prediction Using Energy Consumption Data . . . . . . . 1377
Oussama Laayati, Mostafa Bouzi, and Ahmed Chebak
Intelligent Technique Proposed for Nonlinear Inductor Modelling
for DC/DC Converters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1387
Rafika El idrissi, Ahmed Abbou, Abderrahim Taouni,
and Mohcine Mokhlis
High Order Sliding Mode Power Control of Doubly Fed Induction
Generator Based on Wind Turbine System . . . . . . . . . . . . . . . . . . . . . . 1399
Yahya Dbaghi, Sadik Farhat, Mohamed Mediouni, Hassan Essakhi,
and Aicha Elmoudden
Robust Voltage Control for Four-Phase Interleaved DC-DC Boost
Converter for Electric Vehicle Application . . . . . . . . . . . . . . . . . . . . . . . 1409
Mohamed Barara, M. R. Barakat, Nabil Elhaj, and Ghania Belkacem
MPPT and Pitch Angle Based on Neural Network Control
of Wind Turbine Equipped with DFIG for All Operating Wind
Speed Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1421
Moussa Reddak, Ayoub Nouaiti, Anass Gourma, and Abdelmajid Berdai
Cairo Metro Power Analysis and Harmonic Mitigation . . . . . . . . . . . . . 1433
Ahmed Abd Elsadek, O. A. Monem, Zaki Matar, Mokhtar Hussien,
and Yasser Elsayed
Dual Fuzzy Direct Power Control for Doubly-Fed Induction
Generator: Wind Power System Application . . . . . . . . . . . . . . . . . . . . . 1445
Hala Alami Aroussi, El Mostafa Ziani, Manale Bouderbala,
and Badre Bossoufi
DFIG-WECS Non Linear Power Controls Application . . . . . . . . . . . . . 1455
Manale Bouderbala, Badre Bossoufi, Hala Alami Aroussi,
Mohammed Taoussi, and Ahmed Lagriou
Contents xxv

Behaviour of Concrete and Steel for the Construction of Storage

Tanks: Comparative Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1467
Salma El Aamery, El Hassan Achouyab, and Hassan OuajjiOuajji
Sensorless Control of DC-DC Converter Using Integral State
Feedback Controller and Luenberger Observer . . . . . . . . . . . . . . . . . . . 1477
Djamel Taibi, Toufik Amieur, Mohcene Bechouat, Moussa Sedraoui,
and Sami Kahla
Optimization and Improvement of the Efficiency of a Drying System
Based on the Dimensional Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 1489
Dounia Chaatouf, Benyounes Raillani, Mourad Salhi, Samir Amraqui,
and Ahmed Mezrhab
Monitoring and Control System of a Hybrid Micro-CSP/Biomass
Boiler System for Traditional Baths . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1497
Said Lamghari, Hassan Hamdi, Mehdi Khaldoun, Mickael Benhaim,
Fatima Ait Nouh, and Abdelkader Outzourhit
Low Capacity Diffusion Absorption Refrigeration:
Experimental Investigations and Thermal Characterization . . . . . . . . . . 1505
Ikram Saâfi, Ahmed Taieb, and Ahmed Bellagi
The Synchronized Electrical Charge Extraction Regulator
for Harvesting Energy Using Piezoelectric Materials . . . . . . . . . . . . . . . 1517
Youssef El Hmamsy, Chouaib Ennawaoui, Ikrame Najihi,
and Abdelowahed Hajjaji
Calculation of the Magnetic Field in the Vicinity of the Overhead
Transmission Lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1529
S. Houicher and R. Djekidel
Implementation of the Vehicle Speed and Location Monitoring System
to Minimize the Risk of Accident . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1539
Mohammed Karrouchi, Ismail Nasri, Hajar Snoussi,
Abdelhafid Messaoudi, and Kamal Kassmi
Comparison Study of the Resonant Inductive Power Transfer
for Recharging Electric Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1549
Naima Oumidou, Ali Elkhatiri, Sara Khalil, Moussa Labbadi,
and Mohamed Cherkaoui
Magnetic Chargers in Electrical Models: Operating Principle
and Efficiency Analysis of an Inductively Coupled Power
Transfer System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1561
Ali Elkhatiri, Naima Oumidou, Moussa Labbadi, Mouna Lhayani,
and Mohamed Cherkaoui
xxvi Contents

Improve the Energy Harvesting Alternatives Using the Bond Graph

Approach for Powering Critical Autonomous Devices . . . . . . . . . . . . . . 1573
Souad Touairi and Mustapha Mabrouki
Online Energy Management Strategy of FC/UC Hybrid Electric
Vehicle Based on Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1585
Nada Rifai, Jalal Sabor, and Chakib Alaoui
The Potential Outcomes of Artificial Intelligence Applied
to the Powered Two-Wheel Vehicle: Analytical Review . . . . . . . . . . . . . 1595
F. Jalti, B. Hajji, and A. Mbarki
The Design of an EEG Bio-Amplifier for the Primary Motor Cortex
Zone in 180 nm CMOS Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1607
Younes Laababid, Karim El Khadiri, Hassan Qjidaa, Ahmed Lakhssassi,
and Ahmed Tahiri
Design of a Non-inverting Buck-Boost Converter Controlled
by Voltage-Mode PWM in TSMC 180 nm CMOS Technology . . . . . . . 1619
Anas Boutaghlaline, Karim El Khadiri, Hassan Qjidaa, and Ahmed Tahiri
Low-Power High-Speed On-Chip 5 to 1 Serializer
in 180 nm Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1631
Aicha Menssouri, Karim El Khadiri, Hassan Qjidaa, Ahmed Lakhssassi,
and Ahmed Tahiri
Computing and Analyzing Through Silicon Via (TSV) Noise Coupling
in 3D-IC Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1643
Khaoula Ait Belaid, Hassan Belahrach, Hassan Ayad, and Fatima Ez-zaki
Design of a CMOS Bandgap Reference Voltage Using the OP AMP
in 180 nm Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1655
Ahmed Rahali, Karim El Khadiri, Zakia Lakhliai, Hassan Qjidaa,
and Ahmed Tahiri

Mechatronic, Robotic and Control System

Analytical Study of the CMOS Active Inductor . . . . . . . . . . . . . . . . . . . 1665
Imane Halkhams, Wafae El Hamdani, Said Mazer, Moulhime El Bekkali,
and Mohammed Fattah
SM Versus PI Tunning for PWM Rectifier for SEIG in Wind Energy
Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1675
Ouafia Fadi, Soufiane Gaizen, and Ahmed Abbou
Improved DTC of the PID Controller by Using Genetic Algorithm
of a Doubly Fed Induction Motor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1687
Said Mahfoud, Aziz Derouich, Najib El Ouanjli, Mohammed Taoussi,
and Mohammed El Mahfoud
Contents xxvii

State Feedback Control of DC-DC Converter Using LQR Integral

Controller and Kalman Filter Observer . . . . . . . . . . . . . . . . . . . . . . . . . 1699
Djamel Taibi, Toufik Amieur, Mohcene Bechouat, Sami Kahla,
and Moussa Sedraoui
Backstepping and Indirect Vector Control for Rotor Side Converter
of Doubly Fed-Induction Generator with Maximum Power
Point Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1711
Elmostafa Chetouani, Youssef Errami, Abdellatif Obbadi,
and Smail Sahnoun
Integral Backstepping Power Control of DFIG Based Nonlinear
Modeling Using Voltage Oriented Control . . . . . . . . . . . . . . . . . . . . . . . 1725
Mourad Loucif, Abdelkader Mechernene, and Badre Bossoufi
Smart Monitoring PID of Solenoid Response Based
on Fiber Squeezer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1735
Said Amrane, Abdallah Zahidi, Nawfel Azami, Mostafa Abouricha,
Naoual Nasser, and Mohamed Errai
Fuzzy Logic Based MPPT Control of an Isolated Hybrid
PV-Wind-Battery Energy System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1745
Hadjer Abadlia and Riad Toufouti
Implementation of a Digital Control for PV Power Systems . . . . . . . . . 1757
Badreddine Lahfaoui
Towards a Digital Modeling of the Optimal Mechanical Properties
of a Green Eco Composite Based on Renewable Resources . . . . . . . . . . 1765
Aziz Moumen, Abdelghani Lakhdar, Mustapha Jammoukh,
and Khalifa Mansouri
Experimental Evaluation of MEMS Accelerometers Integrated
into Smartphones: A Case Study of Bearing Condition Monitoring . . . . 1777
Abdelbaset Ait Ben Ahmed, Abdelhamid Touache, Abdelhadi El Hakimi,
and Abderrahim Chamat
MPPT Using Adaptive Genetic-Fuzzy Logic Control for Wind
Power System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1787
Chakib Alaoui, Hajar Saikouk, and Anass Bakouri
Development of an Automated Measurement Platform
for an Electrical Traction System Test Bench . . . . . . . . . . . . . . . . . . . . 1799
Amine El Houre, Driss Yousfi, Zakaria Bourzouk, and Mohammed Chaker
Simulation of a Guidance Law on a ROS-Based Nonholonomic
Mobile Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1809
Nabil Rezoug and Mokhtar Zerikat
xxviii Contents

Speed Sensorless Direct Torque Control of Doubly Fed Induction

Motor Using Model Reference Adaptive System . . . . . . . . . . . . . . . . . . 1821
Mohammed El Mahfoud, Badre Bossoufi, Najib El Ouanjli, Said Mahfoud,
Mourad Yessef, and Mohammed Taoussi

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1831

Artificial Intelligence, Machine Learning
and Data Analysis
New Reconfiguration of the Radial
Distribution Network by Using the
Chaotic Mapping and the Success Rate to
Improve the PSO Algorithm

Meriem M’dioud, Ismail ELkafazi, and Rachid Bannari

Abstract The inertia weight is one of the inspection parameters that improve the
reliability of the Particle Swarm Optimization (PSO). To define this inertia, several
techniques have been suggested to improve the quality of the PSO algorithm. In this
paper, the aim objective is to determine the new reconfiguration of the radial system
by using two strategies: the chaotic descending inertia weight, and the combination
strategy of the chaotic inertia weight and the success rate technique aiming to improve
the result. To define the reliability and the solidity of these suggested techniques,
IEEE 33 bus is used, and the results are compared with other recent studies where
they have also used the particle swarm optimization by using other strategies to define
the inertia weight parameter. These strategies of inertia weight used in this paper are
selected for their features of speed convergence and precision of solution that is near
to the optimal solution.

Keywords Radial system distribution · Particle swarm optimization · Load flow

analysis · Power losses · Chaotic inertia weight · Success rate

1 Introduction

In recent years, several companies have been oriented towards the research about
the new strategies, to improve and optimize energy exploitation by finding suitable
solutions, so these new strategies help to minimize the losses. This paper focuses on
the distribution network’s reconfiguration. Many researchers have proposed different
methods based on mathematical programming for solving problems of losses and
cost minimization.

M. M’dioud (B) · R. Bannari

Laboratory Engineering Sciences Ensa, Ibn Tofail University Kenitra, Kenitra, Morocco
I. ELkafazi
Laboratory SMARTILAB, Moroccan School of Engineering Sciences, EMSI, Rabat, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 3

S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_1
4 M. M’dioud et al.

In this vision, the authors of [1], have used an evolutionary technique using the
particle swarm optimization, to find the optimal solution of restoration of radial power
distribution. The authors of [2], have solved the same problem using the minimum
spanning tree, and in the end, they have compared their solution with the solution of
the authors [1]. For the authors of [3], they have considered that the reactive power is
the weights of lines to reduce losses. On the other hand, the authors of [4] have chosen
the genetic algorithm that provides good quality solutions in reduced execution time.
In this article, i will use an improved and adaptative PSO algorithm, thanks to
its high speed to converge towards an optimal solution, it easy to implement and it
focuses on simple mathematical equations. For testing the solidity and the reliability
of this suggested method, this paper uses the IEEE 33 bus, and compare the results
with other recent studies.
In this vision, the authors of [5], have used in their study, the Particle Swarm
Optimization based on the decreasing inertia weight w to update position and velocity
of the particle then they have used the Backward Forward Sweep to do load flow
analysis. For the authors [6], they have used also the linear decreasing inertia weight
by applying the sigmoid transformation to limit the velocities. On another side, the
authors of [7] have applied the linear decreasing weight by canceling the wend in the
second term. The authors of [8] have tried to solve this problem of reconfiguration
by using a modified shark smell optimization, that based on the same idea of particle
swarm optimization, and he concludes that this method solves the problem in a
significant time and help to improve the profile voltage at each node.
Considering the authors of [9], where they have made a comparative study between
the different strategies used to calculate the inertia weight parameter, and to validate
the performance of their study they based on five famous mathematic equations,
at the end of their study, they have concluded that the chaotic inertia weight is
the best strategy for better accuracy, however, the random inertia weight strategy
is best for better efficiency. On the other hand, the authors of [10] have based on
the chaotic inertia weight by combined the swarm success rate parameter with the
chaotic mapping to set the new inertia weight coefficient, aiming to verify the quality
and the efficiency of this proposed method, they tested their strategy by searching the
solution of these function (Griewank, Rastrigin, Rosenbrock, Schaffer f6, Sphere),
and they have concluded that the swarm success rate is a useful tool to improve the
reliability of any swarm focused on optimization algorithms.
Hence, in this paper, the aim objective is to adjust the inertia weight parameter by
using the two tools described in the previous paragraph (the chaotic inertia weight and
the swarm success rate combined with the chaotic mapping). To study this problem,
we have divided the article into four sections. Section two presents the objective
function and define the constraints. Describe the algorithm of the chosen method
and present the case study. Section three discusses and analyses the results with the
other recent works. And in the five and the final section, i conclude the research and
i determine the future works.
New Reconfiguration of the Radial Distribution Network … 5

2 Problematic Reformulation

As described in the previous section, the peak demand period is when all resources
are operating at maximum, and among the drawback of this period is that give rise to
unnecessary expenses for the electric companies resulting from losses. As the load
increases, these losses are more significant. Aiming to reduce the losses, this paper is
focused on reconfiguration of the radial system by using two kinds of inertia weight
to implement the PSO algorithm.

2.1 Objective Function

In this paper, to calculate losses, it seems important to use the following expression:

Pj = Rk ∗ Ik 2 (1)
k∈S

With S is the set of the network’s lines and Ik is the current at line k, Rk is the
current at line k.
This problem of the loss’s optimization solved under the following constraints
[11]:
Kirchhoff’s law:

I∗X=0 (2)

Where; I: row vector of current and X: incidence matrix of graph (Xij = 0 if there
are no arcs between i and j; Xij = 1 else);
Tolerance limit:

|(Vin − Vi )/Vin | ≤ εimax (3)

Vin nominal voltage at node i, Vi is the voltage at node i, and εimax is tolerance
limit at node i [4] (±5% for HTA and +6%/−10%BT).
Admissible current constraint:

Ik ≤ Ik,maxadm (4)

Ik : current of the line k and Ik,maxadm : current maximum admissible of the line k.
Radial topology constraint:
For the reasons to have a simple, inexpensive operation and keeping the security
and protection of distribution power grid, the radial configuration is requested. It
means that each loop should have an open line. To check this topology, the following
constraints should be considered:
Total number of main loops:
6 M. M’dioud et al.

Nmain loops = (Nbr − Nbus ) + 1 (5)

Where Nbr is the total branches of the network, Nbus is the total number of buses
and Nmain loops are the total number of loops in the network.
The Total number of sectionalizing switches

Nbr = Nbus − 1 (6)

The total number of open switches must be equal to the number of main loops in
the network.
To solve this problem, it seems useful to break up this problem into two important
elements, the first element concerning the reconfiguration of the radial distribution
system using the PSO algorithm and the second element concerning the Newton-
Raphson methods to apply the power flow module to check the set of the constraints.
This load flow method is chosen due to its feature of a fast convergence rate [12].

2.2 PSO Algorithm

PSO is a metaheuristic of optimization, invented by the authors [13]. This optimiza-

tion tool is focused on the collaboration of individuals among themselves. In this
study, thanks to the characteristic of the simple travel rules in the solution space, the
particles can gradually find the global minimum. This algorithm works following
these steps:
Step 1: in the beginning, initialize the number of particles and the number of tie
lines by respecting the condition of the system is in radial nature.
Step 2: Initialize iteration number (maxiter), inertia coefficient (wmin and wmax),
and acceleration coefficients (C1 and C2), the initial velocity of each particle is
randomly generated.
Step 3: Identify the search space for each D dimension (all possible configuration).
Step 4: apply newton-Raphson method [12] to load flow analysis.
Step 5: define the best value among all pbest values.
Step 6: find the global best and identify the new tie switches.
Step 7: update the velocity and new position for each D dimension of the ith
particle using the following equation:
Select a random number z in the interval [0, 1] and use a chaotic mapping by
using the logistic mapping to set inertia weight coefficient [9]:

z(iter + 1) = 4 ∗ z(iter) ∗ (1 − z(iter)) (7)

or use the success rate [10]:

    
1 f  Pbestti  < i
f  Pbestt−1 
Successti = (8)
0 f Pbestti ≥ i
f Pbestt−1
New Reconfiguration of the Radial Distribution Network … 7

n
Succrate = Successti /n (9)
i=1

z(iter + 1) = 4 ∗ Succrate ∗ (1 − Succrate ) (10)

According to the choice of the inertia weight calculation strategy, adjust and
calculate the inertia coefficient by using [13]:

W = (wmin − wmax) ∗ (maxiter − iter)/maxiter+ wmax*z(iter+1) (11)

After adjusting the inertia weight by selecting one of the previous equations that
define this parameter:
Update the velocity by using this equation [13]:

Vi(iter + 1) = W ∗ Vi (iter) + C1 ∗ r1 ∗ (Pi (iter) − Xi(iter)) + C2 ∗ r2 ∗ (G(t) − Xi (iter))

(12)

Update the position by using this equation [13]:

Xi (iter + 1) = xi (iter) + vi (iter + 1) (13)

Define the new fitness function value for the new position [13]:
  
Pbestti f xit+1  > Pbestit
Pbestt+1 = (14)
t
xit+1 f xit+1 ≤ Pbestit

Define the global best by using [13]


Gbest = min Pbestit } (15)

Step 8: until iter = maxiter, go to step 4. Else print the optimal results.
Step 9: display result.

2.3 Test Case

In IEEE graph has 33 buses. There are 37 branches, 32 close switches, and 5 open
switches. The red line indicates the tie-line switches. In the initial case, we assume
that {33-34-35-36-37} is the set of the opened lines. The IEEE 33 features (electrical
and topological) are given in the reference [14], the following figure shows the
architecture of the network at the initial case (Fig. 1):
The respective loops, hence, dimensions, the search space for respective dimen-
sion and the parameters of PSO are given in [8].
8 M. M’dioud et al.

Fig. 1 Distribution system IEEE “33 node before reconfiguration”

Fig. 2 IEEE 33 bus after reconfiguration using chaotic inertia weight

3 Test and Results

After the implementation of the IEEE 33 bus data in MATLAB, we obtained the new
radial distribution that presented in the next schema (Figs. 2 and 3):
In this case, and after reconfiguration by using the chaotic inertia weight, it’s
clear that the new tie lines are 7-9-14-32-37. For the second suggested method using
the combination of the success rate and the chaotic inertia weight, the tie lines are
7-11-14-28-36. To define the updated loss value of the new distribution network
and improve the voltage profile at each node, the Newton Raphson method [12]
(applicated by using Matpower version 4.1). The next table summarizes the result of
New Reconfiguration of the Radial Distribution Network … 9

Fig. 3 Network distribution after reconfiguration by using PSO with combination of the success
rate and the chaotic inertia weight

Table 1 Comparative study

Base case [5] [7] [8] PSO using the PSO using the
chaotic inertia combination of
weight success rate and
chaotic inertia
weight
Tie switch 33-34-35-36-37 7-9-14-28-31 9-7-14-37-32 7-9-14-32-37 7-9-14-32-37 7-11-14-28-36
Real power 208.459 148.3075 136.3 138.9276 138.9275 143.9227
losses (kW)
Vmin(p.u) 0.91075 0.9395 0.940 0.94234 0.9423 0.93682
Node 18 32 22 32 32 18

this paper and gives the result of other studies to make difference and conclude the
feature and the benefits of this study.
As presented in the above table, concerning the tie switch, for the case of recon-
figuration by using the chaotic inertia weight, the new tie lines are 7-9-14-32-37,
while the new tie lines in the case of using the combination of success rate and the
chaotic inertia are 7-11-14-28-36. When we talk about the role of these strategies to
reduce losses, we notice that the losses in the case of PSO using the chaotic inertia
weight equal to 138.9276 kW, and in the case of the using a combination of success
rate and the chaotic inertia weight the losses become 143.9227 kW. These values are
lesser than the losses in the base case (208.459 kW) and the case [5] where the value
of losses is 148.3075 kW. But they are greater than the value found by [7] which
equals 136.3 kW.
For the voltage profile side, the Table 1 proves that in the case of PSO using the
chaotic inertia weight, the minimum voltage profile equal to 0.9423 p.u at node 32,
and in the case of PSO using the combination of success rate and the chaotic inertia
weight the losses equal to 0.93682 p.u at node 18.these values are much improved
10 M. M’dioud et al.

Fig. 4 Voltage profile before and after the reconfiguration using chaotic inertia weight

than the base case 0.91075 p.u at bus 18, and also improved than [5] where the
minimum voltage profile is 0.9395 p.u at bus 32. Also, this value found by using the
chaotic inertia weight is perfect than [7], where the minimum voltage in this study
equal to 0.940 p.u at node 22.
To conclude, it seems clear that the PSO using the chaotic inertia weight gives the
best result than the PSO using the combination of the success rate and the chaotic
inertia weight. This first proposed strategy gives the same result as [8]. However, this
first suggested method reduces the losses with voltage profile improved than [7].
The following figure shows the difference between the voltage profile in the case
of the chaotic inertia weight method and in the base case (Fig. 4):
The Fig. 5 shows the difference between the voltage profile in the case of the
combination method and in the base case. In addition, it is important to note that this
first proposed algorithm takes 46.68s and the second one takes about 55.178566 s
for that they run.
New Reconfiguration of the Radial Distribution Network … 11

Fig. 5 Voltage profile before and after the reconfiguration using the combination of success rate
and the chaotic inertia weight

4 Conclusion

This paper is concerned with a study of two kinds of strategies to determine the
inertia weight, the chaotic inertia weight, and the combination strategy. The main
objective of this research is to define a new reconfiguration of the radial distribution,
aiming to reduce losses and improve voltage profile.
To check the reliability and the performance of these strategies, we use IEEE
33 bus, to test these strategies. The results found are compared with another recent
research focused on the PSO algorithm. According to the results, it’s clear that using
the chaotic inertia weight help to reduce losses better than the combination strategy
of the success rate and the chaotic inertia weight. In the framework of the profile
voltage improvement, these two strategies found the same result. But the combination
strategy takes more time than the chaotic inertia weight strategy to execute.
Finally, the chaotic inertia weight is the best strategy for better reconfiguration
and improved losses. In the next study, I will use this perfect strategy to find the best
allocation and sizing of a distributed generation.

References

1. Oliveira LW, Oliveira EJ, Silva IC, Gomes FV, Borges TT, Marcato AL, Oliveira ÂR (2015)
Optimal restoration of power distribution system through particle swarm optimization. In: IEEE
Eindhoven PowerTech, p 13
12 M. M’dioud et al.

2. Mohamad H (2019) Power system restoration in distribution network using kruskal’s algorithm.
Indonesian J Electr Eng Comput Sci 16(1):8
3. Sudhakar TD (2017) Loss minimization in distribution network by using Prim’s algorithm. Int
J Appl Eng Res 9(24):15
4. Tomoiagă B, Chindriş M, Sumper A, Villafafila-Robles R, Sudria-Andreu A (2013) Distribution
system reconfiguration using genetic algorithm based on connected graphs. Electric Power Syst
Res J 104(216–225):10
5. Sudhakara Reddy AV, Damodar Reddy M (2016) Optimization of network reconfiguration by
using particle swarm optimization. In: 1st IEEE international conference on power electronics,
intelligent control and energy systems (ICPEICES)
6. Tandon A, Saxena D (2014) Optimal reconfiguration of electrical distribution network using
selective particle swarm optimization algorithm
7. Atteya II, Ashour H, Fahmi N, Strickland D (2017) Radial distribution network reconfiguration
for power losses reduction using a modified particle swarm optimisation
8. Juma SA (2018) Optimal radial distribution network reconfiguration using modified shark smell
optimization. Pan African University Institute for Basic Sciences Technology and Innovation
9. Bansal JC, Singh PK, Saraswat M, Verma A, Jadon SS, Abraham A (2011) Inertia weight
strategies in particle swarm optimization
10. Arasomwan MA, Adewumi AO (2013) On adaptive chaotic inertia weights in particle swarm
optimization
11. Enacheanu FB (2008) outils d’aide à la conduite pour les opérateurs des réseaux. https://tel.
archives-ouvertes.fr/tel-00245652
12. Sharma A, Saini M, Ahmed M (2017) Power flow analysis using NR method. In: International
conference on innovative research in science, technology and management
13. Kennedy J, Eberhart R (1995) Particle swarm optimization. In: IEEE international conference
on neural networks, proceedings, vol 4
14. Baran ME, Wu FF (1989) Network reconfiguration in distribution systems for loss reduction
and load balancing. EEE Trans Power Deliv 4(2):7
Convolutional Neuronal Networks
for Tumor Regions Detection
in Histopathology Images

Mohammed Lamine Benomar, Nesma Settouti, Rudan Xiao,

Damien Ambrosetti, and Xavier Descombes

Abstract Recent advances in deep learning and image sensors present the oppor-
tunity to unify these two complementary fields of research for a better resolution of
the image classification problem. Deep learning provides image processing with the
representational power necessary to improve the performance of image classification
methods. The deep networks are considered as the most powerful tool in visual recog-
nition, notably on the automated analysis of histological slides using Whole Slide
Image (WSI) scanning techniques. In this paper, we proposed a comparative study
analysis for efficient localization of tumor regions in kidney histopathology WSIs
by employing state-of-the-art convolutional deep learning networks such as ResNet
and VGGNet which can be fine-tuned to the target problem and exploit the powerful
transfer learning techniques that save low computational costs. Experimental results
conducted on an independent set of 17.300 images derived from 12 WSIs show
that ResNet50 and VGG19 combined with VGG16 achieve almost 97% and 96% of
accuracy respectively, for classifying tumor and normal tissues to identify relevant
regions of interest (ROIs).

Keywords Convolutional neuronal network · Whole Slide Image (WSI) · Renal

Cell Carcinoma (RCC) · Histopathology image

M. L. Benomar (B)
University of Ain Temouchent Belhadj Bouchaib, Ain Temouchent, Algeria
e-mail: [email protected]
M. L. Benomar · N. Settouti
Biomedical Engineering Laboratory, University of Tlemcen, Tlemcen, Algeria
e-mail: [email protected]
R. Xiao · X. Descombes
Inria, Sophia Antipolis, Biot, France
e-mail: [email protected]
X. Descombes
e-mail: [email protected]
D. Ambrosetti
Laboratoire Central d’Anatomo Pathologie, CHU Nice/Hôpital Pasteur, Nice, France
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 13

S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_2
14 M. L. Benomar et al.

1 Introduction

Histopathological WSI analysis is the microscopic examination of biological tissues

in order to study the manifestations and evolution of a disease, highlighting the
interactions between different cells and other components of the tissue. This study
helps to determine what type of cancer it is and what treatments to use. Pathologists
must observe in detail wide-field images, images of hundreds of millions of pixels,
in order to extract characteristic areas representative of the patient’s condition. In
recent years, numerous contests for the automatic processing of such images have
been launched [1]. The medical field, and more specifically computational pathology
has recently received increasing attention, partially due to the competition emerging,
where deep learning approaches have been awarded first place in various challenges:
MoNuSAC2020 [2], BACH18 [3], ANHIR2019 [4], CAMELYON17 [5], CAME-
LYON16 [6] and TUPAC16 [7]. In general, competition winners have used one or
more architectures that have performed particularly well in other popular image
recognition competitions. These deep models allow end-to-end learning to extract
and classify. The extracted features are then automatically determined during the
learning process to optimize the classification.
Despite the very interesting results obtained in the execution of certain tasks [8–
11] using deep learning approaches, the underlying theoretical aspects are not yet
well understood and we still have not a very clear understanding of the architectures
or networks that offer better performance than others. It is therefore difficult to say
what structure, how many layers or even the number of neurons that must be available
on each layer in order to obtain good results for a very specific task. In addition, the
lack of knowledge about the choice of critical values, namely the learning rate and
the reliability of the optimization, etc…. Therefore, the design of the architecture
has always been done in an ADHOC way.
In order to reduce the field of investigation and direct users to the best network, we
propose a comparative study to find the accurate Convolutional Neuronal Network
RCC (renal cell carcinoma) region tumors detection in histopathology images. In this
work, we propose a comparative study of several deep learning architectures among
these approaches, we are particularly interested in two deep networks: ResNet [12]
the winner of the ILSVRC 2015 challenge and VGGNet [13], the finalist of the
ILSVRC 2014 competition. Both networks are widely used and known for their
power, with architectures that are far from a traditional CNN.
The remainder of the paper is organized as follows. In Sect. 2, we provide
an overview of the related deep learning works for tumor regions detection in
histopathology images, Sect. 3 describes the proposed approach with different deep
learning architectures. Section 4 presents the image dataset and discuss the exper-
imental results achieved by the models. Finally, conclusions and perspectives for
future work are reported in Sect. 5.
Convolutional Neuronal Networks for Tumor Regions Detection … 15

2 Related Works

Research on deep learning through convolutional and other neural networks has paved
the way for a representation machine learning approach based on an unsupervised
extraction of descriptors (low, medium and high level), i.e. independent of any human
intervention that could affect these performances. In recent years, several papers have
been published showing the potential of deep learning in digital histopathology.
According to the review of state-of-the-art deep learning approaches that have been
used in the context of histopathological image analysis presented by Srinidhi et al.
[1], among the supervised learning techniques, there are three major canonical deep
learning models based on the nature of tasks that are solved in digital histopathology:
classification, regression and segmentation-based models.

Classification: related to pixel-wise/sliding-window classification-based

approaches in a local-level tasks (based on a region) with methods based on
multi-scale CNN [14] and transfer learning approaches [15], or global-level
tasks (image-level prediction like WSI disease grading) [16] with the popular
attention-based models [17].

Regression: consists in detecting or localizing objects by directly regressing the

likelihood of a pixel, as the position of cells [18] or nuclei [19] or cancer severity
score prediction with constrained CNN models [20]. Most of deep regression
models proposed in the literature are mainly based on either Convolutional Neuronal
Networks [21] or Fully Convolutional Networks architectures [22].

Segmentation: related mainly to fully convolutional network (FCN) based approach

[23] and its variants, which are generally formulated as a semantic segmentation task
[24] or instance aware semantic segmentation task [25] in computer vision.

Most deep learning methods applied in digital pathology have been successfully
evaluated in renal clear cell carcinoma [26], breast cancer [11], cervical cancer
[16], prostate cancer [27], where several multiple publicly databases are available.
However, there is a need for a relevant one which is more significant and represen-
tative of real clinical task. In this paper, we proposed a histopathological Kidney
database, it includes 17.300 images issue from 12 WSIs Hematoxylin and eosin
(H&E) stained slides of Kidney tumor collected at the University Hospital Center of
Nice (France).
Whilst large part of the literature deals with tumor regions segmentation [28, 29]
or cell and nuclei morphology analysis [26, 30]. Meanwhile, [31] investigate the
vascular network on the same dataset study in this paper, in scanning the network
irregularity indicating the severity of the cancer i.e. malignant tumor and inversely,
the regularity of the network for low grade tumor.
In our case, we are particularly interested in the highest level of magnification,
and in the segmentation of tumor regions, although not requiring the intervention of
an expert, this task could gain in precision to become sufficiently reliable and thus
avoid having use of manual segmentation. We propose to study two deep learning
architectures, VGGNet and ResNeT which their weight configuration is publicly
16 M. L. Benomar et al.

available and has been used in many other applications and challenges as a basic
functionality extractor. However, these two architectures include several parameters,
which can be a bit difficult to manage and interesting to study. Indeed, VGGNet 16 &
19 consists of 16 and 19 convolutional layers respectively and are very attractive due
to their very uniform architecture. Like the canonical approach of LeNet and AlexNet,
with only 3 × 3 convolutions, but lots of filters. It is currently the most popular choice
in the community for extracting functionality from images. The ResNet residual
neural network is characterized by an architecture with “jump connections” and
presents a very heavy batch normalization. These jump connections are also called
gated 15 units or gated recurrent units and have a strong similarity to recent successful
elements applied in RNNs.

3 Proposed Approach

In this paper, we evaluate different deep learning networks to deal with histopathology
images task that may assist pathologists in locating tumor regions and allowing
further complex image analysis exams to be performed only on relevant regions
within the entire slide image (WSI) as can be seen from Fig. 1, thereby achieving
more computational efficiency increase pathologists throughput.
To develop this system for identifying tumor tissues on histopathological images,
our approach focused on data augmentation and the implementation of ResNet and
a combination of VGG19 & VGG16. To perform data augmentation, the whole
slide image was divided into patches by cropping (500 × 500) regions from the high
resolution image to increase the amounts of training images [32, 33], where the goal is

Fig. 1 Tumor Regions Detection Pipeline. 1 Tissue whole slide images (WSI); 2 Extracted patches
from whole slide images; 3 Feeding the Deep Network with the Normal and Cancerous slides
patches extracted by sliding window; 4 Patches classification as Tumor or Not-tumor; 5 Binary
mask generated where each pixel of the map corresponds to a (500 × 500 pixels) patch in the
input WSI; 6 Small patches removed using morphological operations. The resulting image is then
displayed by sliding window to obtain the final tumor map for the entire whole slide image
Convolutional Neuronal Networks for Tumor Regions Detection … 17

to classify the WSI patches as belonging to tumor regions (malignant) or normal tissue
(benign class). Additionally, we applied affine (low-level) transformations to modify
images without biasing the classification by using more morphological operators.
Besides the data augmentation, computational burden and memory load are also
impacted; consequently, each patch (sub-image) was processed independently.
The major strength of ResNet architecture lies in its ability to train success-
fully very deep neural networks with over 150 layers, and addresses the problem of
vanishing gradients (i.e. the repeated multiplication can make the gradient extremely
small when back-propagating to earlier layers) by providing shortcuts connection
at each layer of the network allowing the gradient to pass through the layers more
easily [12]. As discussed, we investigate the way of applying popular models such as
the VGG network [13] due to its simplicity and success in the ImageNet challenge.
VGG network uses small kernels (only 3 × 3) in every convolution layer in order
to extract fine features present in images. Additionally, a significant improvement
can be achieved by pushing the depth to 16–19 weight layers. Therefore, to correctly
classify tumor and other tissue regions from histopathology whole slide images we
combine both VGG16 and VGG19 to improve the performance. Indeed, combining
different feature extraction methods increases the classification accuracy [34].

4 Experiments and Results

4.1 Data Collection

The histopathology dataset used is composed of 17.300 images distributed to 2

classes (ROI tumor and not-ROI) with (500 × 500) pixels originate from a set of 12
digitized hematoxylin and eosin (H&E) stained slides of Kidney tumor collected at
the University Hospital Center of Nice (France).
The WSIs were acquired using Leica SCN400 scanner on 400 × magnification
with a resolution of 0.25 µm per pixel and stored in ‘.scn’ format (RGB encoding
with 8 bits per channel), the images are around 100 000 × 100 000 pixels in size. The
histopathology whole slide image (WSI) contains different types of deviations from a
healthy structure which do not necessarily form a continuous spectrum of changes and
their detection can be difficult for pathologists. Therefore, pathologists examined all
slides to identify and delineate manually tumor regions for the evaluation of models’
performance. Consequently, each slide has a labeled ground truth image where the
tumor tissue regions are marked in black. Additionally, Fig. 2 shows various patches
samples from our whole slide images dataset. Finally, the dataset is split into training
and test set. We randomly partitioned 14.300 of these images (about 82.5% of the
dataset) for model training, and the remaining 3.000 images (about 17.5% of the
dataset) for model testing. Furthermore, the 14.300 images used for training were
split into a training set of 12.000 images and a validation set of 2.300 images.
18 M. L. Benomar et al.

Fig. 2 Histopathologic patches examples resize to 224 × 224 pixels at 400× magnification

4.2 Experimental Setup

We implemented ResNet-50 architecture pre-trained on ImageNet dataset weights

with resized square patches from 500 × 500 to 224 × 224 for inputs since the shape of
the input image is defined as 224 × 224 pixels and reducing the image size nearly by
half does not significantly impact convolutional models accuracy for histopathology
images analysis [32]. Data augmentation techniques are widely used in medical
image analysis for model generalization. To this end, we performed rotation (90°)
and random flipping across the horizontal and vertical axes on the training images
dataset.
Furthermore, before inputting a patch into the model, the color channels were
normalized by rescaling all the values to [0 1] interval. This makes the training
approach more convenient by reducing memory and processing needs, and prevents
falling into exploding gradients problem without large impact on prediction results.
We train the ResNet-50 on the histopathology WSI dataset, however, during the
process we freeze ResNet50 layer until “res5a branch2a” of the convolutional part
in order to train only the last few layers, where the number 5 indicates the block and
the letter “a” the sub-block.
The second architecture based on VGG19 and VGG16 fusion, to fine-tuning
networks we use the same image patch size 224 × 224 pixels and rescaling all the pixel
values to [0 1] interval in order to decrease the colored intensive rate. Additionally, to
limit over-fitting we apply the same data augmentation transformations (rotation and
flip) to modify original images and avoid biasing the classification process. Unlike
the ResNet50 setup, VGG19 and VGG16 were initialized with random weights and
all layers are not freezing in order to retrain it on our histopathology dataset with
low learning rate.
For training the RestNet and VGGs models, a stochastic gradient optimizer was
used to fine-tune the networks and cross-entropy as a loss function. Regarding the
model’s parameters, batch size and epochs numbers were set to 32 and 20 respectively,
with an initial learning rate of 1e−4 and decay according to the epochs. Note that the
learning rate is one of the most important hyper-parameters that needs to be tuned
when building the model, since it controls the adjustment speed of our network
weights with respect to the local minima. Moreover, to avoid over-fitting during the
experiments, the early stopping method is required, hence the average number of
epochs was between 8 and 12 as the training stopped when the validation accuracy
Convolutional Neuronal Networks for Tumor Regions Detection … 19

did not improve for a certain number of epochs (3 epochs). The tests for patch
classification were implemented using Keras library and trained in 25 and 29 min
for ResNet50 and VGG19 & VGG16 respectively, on Colab (Google Colaboratory)
GPU environment.

4.3 Results

The two deep networks described in the previous section were evaluated on the
histopathology whole slide images dataset and the experiments were carried out on
the training, validation and test subset, where we apply the exact same preprocessing
by rescaling images before evaluating the network in order to reach maximum perfor-
mance. The number of trainable parameters for each model used are 40.666.881 for
ResNet50 and 34.740.097 for VGG19 and VGG16.
Figures 3 and 4 present the training and validation set loss of ResNet50 and VGGs
models respectively, the variance is low ensuring that our models are not over-fitted.
Additionally, the accuracy results show that both ResNet50 and VGG19 & VGG16
models perform well and correctly classify cancerous regions. However, ResNet50
achieved slightly better performance than VGG19 & VGG16 models, as shown in
Figs. 3 and 4 we plotted the Receiver Operating Characteristic (ROC) Curve of our
models, where it achieved an area under the curve (AUC) equal to 0.972 and 0.967
respectively.
These results are in good agreement with the ResNet50 model which use the
transfer learning techniques that can improve the classification performance with a
limited training dataset. Moreover, Deep Residual Network uses bottleneck residual

Fig. 3 ResNet50 training performances

Fig. 4 VGG training performances

20 M. L. Benomar et al.

Fig. 5 ROIs identification results (from left to right): input image, ground truth image, ResNet50
and VGG19&VGG16 result respectively

block design to mitigate the vanishing gradient problem by allowing alternate shortcut
path for gradient to flow through.
The fusion from VGG19 and VGG16 convolutional layers also achieved impor-
tant improvements by incrementally adapting the pre-trained features and correlate
the sizes of the receptive fields of the convolutional layers with the basic image
elements (nucleus, vessels or a particular tissue pattern) to increase the recogni-
tion accuracy on tumor vs not-tumor regions classification in histopathology images
dataset. Besides the training and validation graphs, the visual results are also impor-
tant as an evaluation tool. As shown in Fig. 5, the two networks provide important
precision in locating the tumor regions for a given whole slide image and assure
that no critical region is overlooked during diagnosis. Subsequently, for future work
some tumor characteristics must be indicated in order to retain particular regions,
since the model provides tumor regions not delimited by the physician in the ground
truth.

5 Conclusion

In this study, we compare ResNet and VGG convolutional deep learning models that
can identify RCC (renal cell carcinoma) regions from normal tissue in histopathology
whole slide images. The CNN’s successfully classify the cancerous regions where the
ResNet50 achieves the best performance accuracy by adding shortcuts connections
to address the accuracy degradation caused by gradient vanishing effect. Moreover,
transfer learning and data augmentation can be considered to overcome the lack of
data in medical field. As a future work, a post-processing step is necessary to obtain
tumor probability maps for the whole slide image. Also, we need to visualize the
overall information captured by the networks and the critical learned features in
differentiation of cancer regions to further improve the results by adding more data.
This opens the possibility to perform classification of the extracted regions into RCC
subtypes.

Acknowledgements The authors would like to thank the Directorate-General of Scientific

Research and Technological Development (Direction Géenérale de la Recherche Scientifique et
Convolutional Neuronal Networks for Tumor Regions Detection … 21

du Développement Technologique, DGRSDT, URL: www.dgrsdt.dz, Algeria) for the financial

assistance towards this research.

References

1. Srinidhi CL, Ciga O, Martel AL (2019) Deep neural network models for computational
histopathology: a survey. arXiv preprint arXiv:1912.12378
2. Verma R, Kumar N, Patil A, Kurian NC, Rane S, Sethi A (2020) Multi-organ nuclei
segmentation and classification challenge 2020
3. Aresta G, Araujo T, Kwok S, Chennamsetty SS, Safwan M, Varghese A, Marami B, Prastawa
M, Chan M, Donovan MJ, Fernandez G, Zeineh J, Kohl M, Walz C, Ludwig F, Braunewell S,
Baust M, Vu QD, To MNN, Kim E, Kwak JT, Galal S, Sanchez-Freire V, Brancati N, Frucci
M, Riccio D, Wang Y, Sun L, Ma K, Fang J, Koné I, Boulmane L, Campilho A, Eloy C,
Polonia A, Aguiar P (2018) BACH: grand challenge on breast cancer histology images. CoRR
abs/1808.04277
4. Borovec J, Kybic J, Arganda-Carreras I, Sorokin DV, Bueno G, Khvostikov AV, Bakas S, Chang
EI, Heldmann S, Kartasalo K, Latonen L, Lotz J, Noga M, Pati S, Punithakumar K, Ruusuvuori
P, Skalski A, Tahmasebi N, Valkonen M, Venet L, Wang Y, Weiss N, Wodzinski M, Xiang Y,
Xu Y, Yan Y, Yushkevic P, Zhao S, Munoz-Barrutia A (2020) ANHIR: automatic non-rigid
histological image registration challenge. IEEE Trans Med Imaging 1 (2020)
5. Bandi P, Geessink O, Manson Q, Dijk MV, Balkenhol M, Hermsen M, Bejnordi BE, Lee B,
Paeng K, Zhong A, Li Q, Zanjani FG, Zinger S, Fukuta K, Komura D, Ovtcharov V, Cheng S,
Zeng S, Thagaard J, Dahl AB, Lin H, Chen H, Jacobsson L, Hedlund M, Cetin M, Halici E,
Jackson H, Chen R, Both F, Franke J, Kusters-Vandevelde H, Vreuls W, Bult P, van Ginneken
B, van der Laak J, Litjens G (2019) From detection of individual metastases to classification
of lymph node status at the patient level: the camelyon17 challenge. IEEE Trans Med Imaging
38(2):550–560
6. Ehteshami Bejnordi B, Veta M, Johannes van Diest P, van Ginneken B, Karssemeijer N, Litjens
G, van der Laak JAWM, the CAMELYON16 Consortium (2017) Diagnostic assessment of deep
learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA
318(22), 2199–2210
7. Veta M, Heng YJ, Stathonikos N, Bejnordi BE, Beca F, Wollmann T, Rohr K, Shah MA,
Wang D, Rousson M, Hedlund M, Tellez D, Ciompi F, Zerhouni E, Lanyi D, Viana MP,
Kovalev V, Liauchuk V, Phoulady HA, Qaiser T, Graham S, Rajpoot NM, Sjoblom E, Molin J,
Paeng K, Hwang S, Park S, Jia Z, Chang EI, Xu Y, Beck AH, van Diest PJ, Pluim JPW (2018)
Predicting breast tumor proliferation from whole-slide images: the TUPAC16 challenge. CoRR
abs/1807.08284
8. Qaiser T, Rajpoot NM (2019) Learning where to see: a novel attention model for automated
immunohistochemical scoring. IEEE Trans Med Imaging 38:2620–2631
9. Serag A, Ion-Margineanu A, Qureshi H, McMillan R, Saint Martin MJ, Diamond J, O’Reilly
P, Hamilton P (2019) Translational ai and deep learning in diagnostic pathology. Front Med
6:185
10. Dimitriou N, Arandjelovi O, Caie PD (2019) Deep learning for whole slide image analysis: an
overview. Front Med 6:264
11. Singh OV, Choudhary P, Thongam K (2020) A study on deep learning for breast cancer detection
in histopathological images. In: Nain N, Vipparthi SK, Raman B (eds) Computer vision and
image processing. Springer, Singapore, pp 36–48
12. He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition. In: 2016
IEEE conference on computer vision and pattern recognition, CVPR 2016, Las Vegas, NV,
USA, 27–30 June 2016. IEEE Computer Society, pp 770–778
22 M. L. Benomar et al.

13. Simonyan K, Zisserman A (2015) Very deep convolutional networks for large-scale image
recognition. In: Bengio Y, LeCun Y (eds) 3rd International conference on learning representa-
tions, ICLR 2015, San Diego, CA, USA, 7–9 May 2015, Conference Track Proceedings
14. Albarqouni S, Baur C, Achilles F, Belagiannis V, Demirci S, Navab N (2016) Aggnet: deep
learning from crowds for mitosis detection in breast cancer histology images. IEEE Trans Med
Imaging 1313–1321
15. Valkonen M, Isola J, Ylinen O, Muhonen V, Saxlin A, Tolonen T, Nykter M, Ruusuvuori P
(2020) Cytokeratin-supervised deep learning for automatic recognition of epithelial cells in
breast cancers stained for ER, PR, and Ki-67. IEEE Trans Med Imaging 39(2):534–542
16. Halicek M, Shahedi M, Little JV, Chen AY, Myers LL, Sumer BD, Fei B (2019) Head and neck
cancer detection in digitized whole-slide histology using convolutional neural networks. Sci
Rep 9(1):14043
17. Zhang Z, Chen P, McGough M, Xing F, Wang C, Bui M, Xie Y, Sapkota M, Cui L, Dhillon J,
Ahmad N, Khalil FK, Dickinson SI, Shi X, Liu F, Su H, Cai J, Yang L (2019) Pathologist-level
interpretable whole-slide cancer diagnosis with deep learning. Nat Mach Intell 1(5):236–245
18. Xie Y, Xing F, Shi X, Kong X, Su H, Yang L (2018) Efficient and robust cell detection: a
structured regression approach. Med Image Anal 44:245–254
19. Xie Y, Kong X, Xing F, Liu F, Su H, Yang L (2015) Deep voting: a robust approach toward
nucleus localization in microscopy images. In: Navab N, Hornegger J, Wells WM, Frangi AF
(eds) Medical image computing and computer-assisted intervention - MICCAI 2015. Springer,
Cham, pp 374–382
20. Sirinukunwattana K, Raza SEA, Tsang Y, Snead DRJ, Cree IA, Rajpoot NM (2016) Locality
sensitive deep learning for detection and classification of nuclei in routine colon cancer
histology images. IEEE Trans Med Imaging 35(5):1196–1206
21. Naylor P, Lae M, Reyal F, Walter T (2019) Segmentation of nuclei in histopathology images
by deep regression of the distance map. IEEE Trans Med Imaging 38(2):448–459
22. Xing F, Cornish TC, Bennett T, Ghosh D, Yang L (2019) Pixel-to-pixel learning with weak
supervision for single-stage nucleus recognition in Ki67 images. IEEE Trans Biomed Eng
66(11):3088–3097
23. Long J, Shelhamer E, Darrell T (2015) Fully convolutional networks for semantic segmentation.
In: 2015 IEEE conference on computer vision and pattern recognition (CVPR), pp 3431–3440
24. Chan L, Hosseini MS, Rowsell C, Plataniotis KN, Damaskinos S (2019) Histosegnet: semantic
segmentation of histological tissue type in whole slide images. In: The IEEE international
conference on computer vision (ICCV), October 2019
25. Qu H, Riedlinger G, Wu P, Huang Q, Yi J, De S, Metaxas D (2019) Joint segmentation and
fine-grained classification of nuclei in histopathology images. In: 2019 IEEE 16th international
symposium on biomedical imaging (ISBI 2019), pp 900–904
26. Tian K, Rubadue CA, Lin DI, Veta M, Pyle ME, Irshad H, Heng YJ (2019) Automated clear
cell renal carcinoma grade classification with prognostic significance. PLoS ONE 14(10):1–16
27. Nagpal K, Foote D, Liu Y, Chen C, Wulczyn E, Tan F, Olson N, Smith J, Mohtashamian A,
Wren JH, Corrado G, MacDonald R, Peng L, Amin M, Evans A, Sangoi A, Mermel C, Hipp
J, Stumpe M (2019) Development and validation of a deep learning algorithm for improving
Gleason scoring of prostate cancer. Nature Partner J (NPJ) Digit Med 2:48
28. Cheng J, Mo X, Wang X, Parwani A, Feng Q, Huang K (2017) Identification of topological
features in renal tumor microenvironment associated with patient survival. Bioinformatics
34(6):1024–1030
29. de Bel T, Hermsen M, Litjens G, van der Laak J (2018) Structure instance segmentation in
renal tissue: a case study on tubular immune cell detection. In: Stoyanov D, Taylor Z, Ciompi
F, Xu Y, Martel A, Maier-Hein L, Rajpoot N, van der Laak J, Veta M, McKenna S, Snead D,
Trucco E, Garvin MK, Chen XJ, Bogunovic H (eds) Computational pathology and ophthalmic
medical image analysis. Springer, Cham, pp 112–119
30. Tabibu S, Vinod PK, Jawahar CV (2019) Pan-renal cell carcinoma classification and survival
prediction from histopathology images using deep learning. Sci Rep 9(1):10509
Convolutional Neuronal Networks for Tumor Regions Detection … 23

31. Zubiolo A, Debreuve E, Ambrosetti D, Pognonec P, Descombes X (2016) Is the vascular

network discriminant enough to classify renal cell carcinoma? In: 14th international workshop
on content-based multimedia indexing, CBMI 2016, Bucharest, Romania, 15–17 June 2016.
IEEE, pp 1–6
32. de Matos J, de Souza Britto A, de Oliveira LES, Koerich AL (2019) Texture CNN for histopatho-
logical image classification. In: 2019 IEEE 32nd international symposium on computer-based
medical systems (CBMS), pp 580–583
33. Spanhol FA, Oliveira LS, Petitjean C, Heutte L (2016) Breast cancer histopathological image
classification using convolutional neural networks. In: 2016 international joint conference on
neural networks (IJCNN), pp 2560–2567
34. Thuy MBH, Hoang VT (2020) Fusing of deep learning, transfer learning and GAN for breast
cancer histopathological image classification. In: Le Thi HA, Le HM, Pham Dinh T, Nguyen
NT (eds) Advanced computational methods for knowledge engineering. Springer, Cham, pp
255–266
Early Detection of Wheat Diseases
in Morocco by Convolutional Neutral
Network
Kenza Aitelkadi , Noureddine Chtaina, Saber Bakouri,
and Mohamed Belebrik

Abstract Convolutional Neural Network is a deep learning method that covers

several fields, including the agronomy sector. This new method provides fast and
very encouraging results for diagnosis disease. As part of this work, we built a model
of early diagnosis of the wheat plant in Moroccan region based on the Convolutional
Neutral Network. Our goal is to allow rapid recognition of contaminated plants at
early stages of appearance in order to provide the necessary treatment and limit prop-
agation damage. Our study begins with a set image collection of wheat plants in the
field and in the laboratory. We collect two type of images. Images of healthy plants
and others that display three diseases, namely Brown rust, Septoria and Powdery
Mildew. A first phase of data processing consisted in segmenting and augmenting
these images. The next step is the model architecture choice, the network training
and finally the evaluation phase. The validation of our model by test images, allowed
us to reach 97% as an accuracy level in the wheat diseases diagnosis.

Keywords Early diagnosis · Wheat disease · Convolutional neural network

1 Introduction

Plant diseases cause considerable crops damage, which greatly reduces produc-
tion. Several types of diseases are destructive [1] especially for the cereals that
are the subject of our study. In Morocco, wheat is the most consumed cereal, its
consumption is estimated at 173 kg/year/person with a production in 2016 of 2.7
Mt [2]. However, this production is affected by many biotic constraints; including
cryptogamic diseases. Septoria and brown rust caused by Zymoseptoria tritici and

K. Aitelkadi (B) · S. Bakouri · M. Belebrik

School of Geomatic Sciences and Surveying, Hassan II Agronomic and Veterinary Institute,
Rabat, Morocco
e-mail: [email protected]
N. Chtaina
Plant Production Department, Hassan II Agronomic and Veterinary Institute, Rabat, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 25

S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_3
26 K. Aitelkadi et al.

Puccinia triticina, respectively, are our main wheat diseases [3]. The symptoms of
these two leaf diseases are specific, they cause significant losses [4]. In order to avoid
agricultural yield damage, protecting wheat from disease is essential to guarantee
the quality and quantity of crops [5]. An effective protection strategy must begin
with early detection of the disease in order to choose the appropriate treatment at the
right time to prevent its spread [6]. Usually, in the absence of an agricultural warning
system, the diagnosis of diseases is done in two stages. The first consists of peri-
odic inspections to detect anomalies in plant growth. The second consists in taking
samples of diseased plants which will be analyzed in the laboratory by phytopathol-
ogists [1]. For farmer, continuous monitoring is difficult and time consuming. These
limitations make important the automation of an early diagnosis diseases for crop
protection [5].
In this context, the image processing by machine learning approaches provides
quick and easily usable solutions. For disease diagnosis, the classifying images by
machine learning are divided in two classical methods and deep forms. Classification
methods based on classical learning use a limited number of images [7]. They allow
iteration more quickly and therefore open the possibility of trying several techniques
in a shorter time [8].These algorithms are easier to interpret and understand [9].
However, their accuracy are lower than those of deep learning methods, in particular
the CNN method. The introduction of deep learning techniques in agriculture, espe-
cially in phytopathology field [10], has only started to take place in the last six years,
to a rather limited extent. The author in reference [11] presented in their research
the importance and potentialities of deep learning in agriculture and plant disease.
The research [10] presented a new method for identifying rice diseases based on
CNN techniques, using a data set of 500 natural images of healthy and diseased rice
leaves. Reference [12] used deep learning architectures for the diagnosis of various
plant diseases. Experimental results of [13] demonstrate that VGG-FCN-VD16 and
VGG-FCN-S system outperforms conventional CNN architectures for wheat disease
recognition. In this work, we try to experiment the performance of new CNN architec-
tures in Moroccan wheat disease context using our specific image and by comparing
results of small and large database.

2 Material and Methods

2.1 Data

Data acquisition is the first preparing step. Data can be acquired from the web, using
a database, or manually by taking pictures with RGB cameras. They are prepared and
inoculated in plant protection laboratory of the Agronomic and Veterinary Hassan
II Institute. We have taken pictures using smartphone camera with 27 Mega Pixels.
Three diseases are chosen: brown rust, Septoria leaf spot and powdery mildew. We
obtained 100 images for each disease. Figure 1 shows an example of the images.
Early Detection of Wheat Diseases in Morocco … 27

Fig. 1 Sample images of affected wheat leaves: a Healthy, b Septoria, c Powdery mildew, d Brown
rust

For training and testing our deep learning algorithm, it is important to separate the
images. The goal is to keep some of the data outside of training, which will be used to
test the reliability of the model. The most popular separation is 80% for training data
and 20% for test data. This is to make sure that the model can work in unexpected
cases.

2.2 Software Tools

We chose Python as the programming language. It’s an easy language to learn. It

has simple syntax and has many libraries [14]. In the context of deep learning, we
used the “Keras” library on python. This library uses tensors to do the various deep
learning calculations which are mathematical objects generalizing scalars, vectors
and matrices to higher dimensions. Also, we chose the Kaggle free workspace since
it makes easy the integration and management of large data. It allows deep learning
models to converge successfully. The storage space of our workspace is 16 GB of
disk space and 16 GB of the Central Processing Unit RAM and 16 GB of the Graphics
Processing Unit RAM.

2.3 Methodology

In this work, we expose the adopted methodology whose main objective is the
construction of early detection model for wheat diseases diagnosis. Our approach
begins by the image processing. Next, we choose the most suitable model architec-
ture for the recognition of wheat diseases. We will also study the effect of hyper
parameters variations and their influences on the model. To finish, we present and
discuss the obtained results.

Image Processing
Image Labeling
28 K. Aitelkadi et al.

Fig. 2 Segmentation process

Image labeling consists of associating the images with labels, for example associating
the image of wheat disease with its label which is the name of this disease. These
labels are a knowledge base for learning the model. The purpose of the model will
be to associate the appropriate label with each test image. So you have to make sure
you create correct labels.
Image Transformation
Neural networks perform better when the input data has common information, the
same background and the same image size. Our images are characterized by the
variation of the backgrounds. Following this variation we decided to proceed with
segmentation to homogenize the background, this procedure is explained in Fig. 2.
The segmentation was done using OpenCV graphics library. We started by
converting the RGB image to an HSV format. Subsequently, we applied an edge
detection algorithm. Finally we created a mask to fill the detected part. Figure 3
shows an image before and after segmentation.
Wheat Leaf Segmentation According to the Verticality
In order to duplicate the number of images and allow our algorithm to recognize
the different parts of a wheat leaf, we proceeded to a second segmentation. It is a
separation step of the three leaf parts according to the verticality. Figure 4 shows the
obtained images. We got three new images. The first representing the top of the leaf,
the second the middle and the third the bottom of the leaf.
Resize and Data Augmentation
We have resized the wheat images to 400 * 300 format because the high resolu-
tion of the images requires powerful equipment for training. Large image sizes
bring to memory problems and can stop training. Image resizing was done using
the OpenCV graphics library and a resizing algorithm [15]. In order to generalize the
wheat diseases recognition on other images with different orientations and positions
of the leaves, we adopted an augmentation step. The purpose of the augmentation is

Fig. 3 Segmentation result:

a before segmentation,
b after segmentation
Early Detection of Wheat Diseases in Morocco … 29

Fig. 4 Segmentation result according to the leaf verticality: a top, b middle, c bottom

to increase the data amount. This augmentation is done by several transformations:

rotations, translations, vertical and horizontal symmetries and horizontal and vertical
stretching. The transformed images will be associated with training data.

Convolutional Neural Network Architecture

One of the deep learning approaches is the Convolutional Neural Network. CNN
models use multiple layers. The input layers, the hidden layers and the output layer.
The depth of the model depends on the network layers number, especially the hidden
layers. The adopted model depends on several parameters, namely the type of data, its
volume, its quality and the type of processing or result expected by the network. The
richness and the performance level of the architectures evolves rapidly and changes
according to the need. In this work, we tried to find the most used architectures, in
a similar context, to solve our problem of diagnosing wheat diseases. We preferred
to experiment architectures which are often used in the field of human, animal and
vegetable diseases diagnosis. We choose to experiment three architectures which are:
“Tairu” architecture [16], “Alexnet” architecture and “LeNet” architecture [12].
The Tairu architecture is a CNN architecture that uses five convolutional layers.
It is implemented for training converges faster. This is useful because the architec-
ture is deep and the number of parameters logically increases which will require a
long training period. This architecture showed very good performance for “PlantVil-
lage” database during our experimentation on tomato diseases diagnosis.”Alexnet”
is popular architecture. Its model includes five convolutional layers and 3 fully
connected layers. In a single convolutional layer, there are usually several filters
of the same size. The activation function used is ReLU. For the first and second
convolutional layers, the activation function is followed by a local normalization
step before pooling. LeNet architecture is a simple model of 3 layers with a convolu-
tion activation function ReLU, followed by max-pooling layers. The main advantage
of this architecture is that it is efficient even if you don’t have a large amount of data.
It’s characterized by a reduced number of layers and therefore it is relatively shallow
learning.
In addition of architecture choice, we should determine the hyper-parameters
before the learning process. A right determination avoids the over-learning problem
30 K. Aitelkadi et al.

and provides better prediction accuracy. Indeed, when a model has learned too much
the peculiarities of training data, it has a very high success rate (up to 100%). Instead
of setting generalized learning on the data test, the model projects closely training
results which provides an unstable model. There are no generic or statistical means
to determine the best hyper-parameters. It’s based on experimentation with empirical
values. The hyper-parameters that we study in this document are:
• The size of image filter
• The Number of Epochs
• The size of the images

3 Results

3.1 Effect of Hyper Parameters Variation

Hyper-parameters are parameters whose value is defined before the start of the
learning process. They cannot be learned directly from data and must be prede-
fined. Unfortunately, there are still no statistical means to determine the best. So we
experimented with different values. We train our model and we choose at the end the
parameters that provide the best precision.

Filter Size
A filter is a matrix of values. It is generally small and often square. It is used to
show some characteristics of a given image (color, outline, brightness, sharpness,
etc.). By scanning the image in successive steps, according to the values and the size
of the filter matrix, we obtain a new image more or less modified. The purpose of
this process is to reveal some characteristics of the image. It is not best to use very
small filters as this can cause over-learning. We tested LeNet model which contains
three convolutional layers. Through tests we carried out on the size, we evaluated the
precision obtained. If a number of pixels is required to recognize the object, large
filters should be used. If the objects present a small part of the image, the small filters
are better [17]. Since the symptoms of the diseases are in a small part of the images
(for example small spots for septoria), it seems logical to use small filters to extract
the characteristics. The obtained results confirm this hypothesis (Fig. 5). By reducing
the size of the filter, we obtained a better precision, we reached a precision of 97%
with a filter size (3 × 3), 92% for (12 × 12) and 89% for (40 × 40).
We found similar results when we tested the small sizes of filters in the neural
network for the tomato diseases diagnosis.

The Number of Epochs

The epoch number is the number of times the algorithm identifies the complete data
set. Thus, each time the algorithm has all the samples in the dataset, an epoch is
over [18]. We have tried to vary the number of epochs, between 1 and 100. Figure 6
Early Detection of Wheat Diseases in Morocco … 31

Fig. 5 Two filters of

different sizes on an image
of a diseased leaf (case of
septoria)

achieved by the “matplotlib” visualization library, shows the evolution of two details
as a function of the epoch number. The first accuracy is that of the training data
and the second is the precision of the tested data. In all the graphs presented in this
document, the x-axis refers to the number of epochs and the y-axis corresponds to
the reached prediction rate.
By growing the number of epochs, the training takes more time and the precision
increases but stagnates from epoch 40. We also notice that the blue curve (training
data) exceeds the red curve (test curve) in the thirtieth epoch which means that
our model did over learning and therefore good performance on training data and
poor generalization on other data. We therefore worked with 25 epochs because this
number offers good precision and at the same time prevents us an over-learning.

The Image Size

The adopted image size depends on the network size and the capacity of the GPU
(Graphics Processing Unit) used to be able to train it. It should add a lot of reasonable
size in GPU memory. We tested three different sizes of images to compare the details
obtained with the optimal training time: 300 * 400, 256 * 256 and 50 * 50. We
obtained 97.61% for 300 * 400 size with an interval time of 768 ms per epoch. Using
the reduced size of 256 * 256, we obtained an accuracy of 97.32% with an interval of
424 ms per epoch. Finally, the small size of 50 * 50 gives us an accuracy of 94.04%
with 33 ms per epoch. So it offers a very short duration, but the accuracy is relatively
low compared to other sizes.

Fig. 6 The accuracy

depending on the epoch’s
number
32 K. Aitelkadi et al.

3.2 Choice of Model

Tairu Architecture
Experimentation with this architecture provided us a 90.47% prediction accuracy.
This is a slightly low diagnostic accuracy. However, despite the precision obtained,
we had an over-learning.
AlexNet Architecture
The “Alexnet” architecture achieved an accuracy of 94.04%. We preferred to test for
100 epochs to see the accuracy evolution of the precision. Precision increased until
epoch 60 where it stabilizes.
LeNet Architecture
By using the LeNet architecture, we obtained an accuracy of 97.61%. The precision
increases according to the time.
The synthesis of our experimentation with the three architectures mentioned
above, allowed us to adopt the appropriate model according to the precision of
predictions obtained (Fig. 7).
We obtained the best predictions from the LeNet model. Indeed, its architecture
has a limited depth which makes it more efficient when we have less data. In the case
of our study, we obtained a prediction accuracy of 97.61% since we have few images.
We could only use 400 images obtained after the segmentation and augmentation
operations. The results also show the effectiveness of CNN architectures regardless
of the data masses. The good precision of certain architectures requires thousands
of images in order to adapt to their depths (in the case of Tairu and AlexNet archi-
tectures). Others provide more relevant results with average mass of images (case
of LeNet architecture). Although it is true, it is not always easy to get millions of
images and to process them on ordinary machines or servers.

Fig. 7 The obtained accuracy according to the architectures

Early Detection of Wheat Diseases in Morocco … 33

4 Discussion

It should be noted that phytopathologists do not need machine learning methods

to accurately interpret the condition of plants and provide assumptions about the
diseases. Rather, think about using them to confirm their hypotheses or to make an
early diagnosis. This could be done by a mobile application using the deep learning
algorithm. For example, if the disease begins to spread in an area and has not spread
over the entire field, machine learning algorithms can effectively help phytopathol-
ogists to accelerate decision-making. This tool offers new possibilities that were
not available before. First, early diagnosis and detection of the disease by the farmer
himself will allow the farmer to know the type of disease and make an early diagnosis.
He can then immediately apply treatments or contact an expert and ensure rapid inter-
vention. Secondly help the expert to confirm his hypothesis. Through this document,
we have tried to simplify the various experiments made either on hyper parameters
or on network architectures, without noting the mathematical formulations since our
main concern was to try to obtain good details on the detection of diseases. This
was accomplished with a prediction accuracy of 97.62% of the three wheat diseases
experienced. Figure 8 shows the results of an iteration of our prediction model on
a septoria leaf. The possibilities offered [healty leaf, brown rust, powdery mildew,
septoria]. We obtained a perfect score for the appearance of septoria (score of 1
or otherwise 100%). The probability that the leaf will be affected by the other two
diseases and the probability that it will be healthy is very low and tends to zero.
Following these results, we conclude that the convolutional neural networks hold
great promise for the detection of diseases and allow very interesting applications
such as the intelligent use of pesticides. In this sense, if deep learning algorithms
use both drone images and terrestrial images of a crop field, the farmer will be able
to guide the phytosanitary treatment of affected areas in an intelligent and optimal
way. However, like any new technology and experiment, some difficulties arise,
especially the complexity of the calculations generated by different models. This
complexity makes the training phase time consuming. In order to improve the work
carried out and with a view to developing other researches, we recommend setting
up an international database of plant disease images which will serve as training data

Fig. 8 The scores obtained

by our prediction model
respectively from left to
right: healthy leaves, brown
rust, Oidum and Septoria
leaf spot
34 K. Aitelkadi et al.

for the deep learning model. The addition of other parameters could improve the
accuracy such as the location, the plant age, etc.

5 Conclusion

In this work, we have integrated deep learning for the early detection of certain
wheat diseases that spread in Morocco. The process requires RGB images of sick
and sick leaves. We have concluded from the results of our experiment that the
use of convolutional neural networks is very promising. We managed to predict
the various diseases of wheat at 97.62%. Particular attention is considered in the
availability of training data, the adequate choice of network structures and adopted
hyper-parameters.

References

1. Blancard D (2012) Tomato diseases: identification, biology and control. Manson Publishing,
London, 688 p. eBook ISBN 9780429158445
2. USDA Foreign Agricultural Service. Morocco: grain and feed update (2016). https://gain.fas.
usda.gov/Recent%20GAIN%20Publications/Grain%20and%20Feed%20Annual_Rabat_Mor
occo_3-30-2016.pdf
3. Lyamani A (1989) Fréquence et sévérité des maladies d’orge, blé dur et blé tendre dans les
principales régions céréalières du Maroc. Rapport d’activité 1988–89. INRA/Centre régional
Settat, Maroc, pp 75–76
4. Eyal Z, Sharen AL, Prescott JM, Van Ginkel M (1987) The Septoria diseases of wheat: concepts
and methods of disease management. CIMMYT, Mexico, 52 p
5. Brahimi M, Boukhalfa K, Moussaoui A (2017) Deep learning for tomato diseases: classification
and symptoms visualization. Appl Artif Intell 31(4):1–17
6. Al Hiary H, Bani SA, Reyalat M, Braik M, ALRahamneh Z (2011) Fast and accurate detection
and classification of plant diseases. Int J Comput Appl 17:31–38
7. Zubiolo A (2017) Extraction de caractéristiques et apprentissage statistique pour l’imagerie
biomédicale cellulaire et tissulaire. PhD thesis, p 134
8. Tzu-Liang T, ReddyAletia H, Zhong H, Yongji K (2015) E-quality control: a support vector
machines approach. J Comput Des Eng 2015(2):91–101
9. Padmavath J (2012) Logistic regression in feature selection in data mining. University SRM,
India, pp 1–3
10. Yang X, Guo T (2017) Machine learning in plant disease research. Eur J BioMed Res 3:6–9.
https://doi.org/10.18088/ejbmr.3.1.2016m
11. Kamilaris A, Francesc X, Prenafeta B (2018) Deep learning in agriculture: a survey. Comput
Electron Agric 147:70–90
12. Ferentinos KP (2018) Deep learning models for plant disease detection and diagnosis. Comput
Electron Agric 145:311–318
13. Lu J, Hu J, Zhao G, Mei F, Zhang C (2017) An in-field automatic wheat disease diagnosis
system. Comput Electron Agric 142:369–379
14. Karkare P (2019) First Steps — Deep Learning using Python and Keras. https://medium.com/
x8-the-ai-community/first-steps-deep-learning-using-python-and-keras-3ae6bcb37d0e
Early Detection of Wheat Diseases in Morocco … 35

15. Wang SY, Wang O, Zhang R, Owens A, Efros AA (2020) CNN-generated images are surpris-
ingly easy to spot... for now. In: Proceedings of the IEEE conference on computer vision and
pattern recognition, vol 7
16. Tairu O (2018) plant disease detection using convolutional neural network. https://towardsda
tascience.com/plant-ai-plant-disease-detection-using-convolutional-neural-network-9b58a9
6f2289
17. Oliveira DCD (2016) How do we choose the filters for the convolutional layer of a Convolu-
tion Neural Network (CNN)? https://www.researchgate.net/post/How_do_we_choose_the_fil
ters_for_the_convolutional_layer_of_a_Convolution_Neural_Network_CNN
18. Sharma A (2018) What is an epoch in deep learning? https://www.quora.com/What-is-an-
epoch-in-deep-learning
Genetic Algorithm Design Optimization
for Non-standard Spur Gears

Samya Belarhzal and El Mostapha Boudi

Abstract Spur gear design is one of the understood subjects that can not be treated
by limited classical methods. For this reason, an optimization process by the genetic
algorithm was conducted in order to minimize the structure volume. Mechan-
ical problems such as gear designs present many variables to analyze and several
constraints to respect. In order to gain in terms of weight cost and transmission,
Genetic Algorithms are used to improve the design process of the gearbox and find
the optimum gear parameters. In this paper, the influence of the profile shift factor
on one spur gear volume and clearance volume is studied including the bottom clear-
ance volume equation in the fitness function. The optimal results of the model with
corrected pinion and gear are compared with the standard spur gear model without a
profile shift factor in bottom clearance equation volume or structure equation volume.

Keywords Genetic algorithms · Optimization · Volume · Profile shift factor ·

Bottom clearance · Spur gear

1 Introduction

Genetic algorithms are frequently used for optimization purposes to solve theoretical
understanding problems. They have been often used in different fields of science,
engineering, or biomechanics [1]. Genetic algorithms adopt the approach of organic
evolution, inspired by populations of individuals with a high quality of genetic mate-
rials. Collecting and modelling different organic evolutions lead to form the basis of
GA [2]. The greatest application of this technique is in the optimization sector, as
discussed by De Jong [3].
Thomas Bäck, Ulrich Hammel, and Hans-Paul Schwefel talked about the history
of evolutionary computation, their objectives, and different structures, including
genetic algorithms [4].

S. Belarhzal (B) · E. M. Boudi

Mechanical Department, Mohammadia School of Engineers, Mohammed V University, Rabat,
Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 37

S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_4
38 S. Belarhzal and E. M. Boudi

In the mechanical field, the gearbox is known for its primary role in power trans-
mission for machines. They are often used in rough environments leading to severe
damages, system failures, and high maintenance costs [5]. For this reason, mechan-
ical designers are always recommended to improve designs to give better results and
reduce costs.
Cevdet Gologlu used GA to minimize the volume of a parallel axis gearbox with
two-stage helical gear trains. He compared the design results with static and dynamic
penalty function to the solutions of a deterministic design already developed [6]. Said
Golabi and Javad Jafari Fesharaki analyzed the volume/weight minimization of one,
two, and three-stage gear train. Depending on different values of input power, the
hardness of gears and, the gear ratio, the results were compared to previous works
solutions [7].
Particle swarms (PS) are used too on many occasions in gearbox volume optimiza-
tions. Ketan Tamboli and Sunny Patel took the volume geometry for a pair of helical
gear reducer as an objective function. The results were enough to improve perfor-
mance requirements [8]. Even if sometimes the particle swarm optimization (PSO)
solutions outperform the GA results on the gearbox minimization volume/weight in
some points, Genetic algorithms still the most used method in gears optimization
design [9].
One of the most interesting parameters to study in the gears optimization field
is the profile shift factor and his influence on different aspects of design, stress,
and transmission capacity. Diez-Ibarbia studied the impact of profile shift factor on
the efficiency of spur gears. He concluded that increasing the profile shift factor
reduces the transmission efficiency [10]. Coming to Genetic Algorithm optimization
Paridhi Rai and Aman Agrawal included the profile shift as a design variable, with
the module, the face width, and the number of teeth in their volume optimization
of the helical gear, using Real Coded Genetic Algorithm (RCGA). They found that
optimum volume with profile shift factor is lower than a volume without it [11].
Daniel Miler and Antonio Loncar treated the impact of profile shift factor on
spur gear pair optimization using the Genetic Algorithm. The pair gear volume was
considered as an objective function. Its a function of addendum diameter instead of
the pitch diameter to highlight the impact of shift coefficient in the objective function.
The obtained results were confirmed by the specialized commercial software results.
The inclusion of the profile shift as a variable leads to a lower volume and weight
[12]. The same authors have made a multi-objective GA optimization on spur gears
based on volume and efficiency as objective functions. The aim was to decrease the
power loss and design volume. The optimum results were achieved by reducing the
module, the face width, and raising the profile shift and teeth number [13].
A lot of work has been achieved on gearbox design. In this work, the authors
studied the impact of the shift coefficient on one pair of spur gear volume by the
process of genetic algorithms.
Two models of the geometry structure are compared. The first model with a simple
volume equation is calculated as a function of pitch diameter, the face width b and,
the transmission ratio i. The second model included the profile shift factor in the
bottom clearance equation to make clear his influence on the optimization results.
Genetic Algorithm Design Optimization … 39

The pinion and gear shift coefficients x 1 and x 2 are taken as function variables in this
model besides the module m, the face width b, and the pinion teeth number Z 1 in the
first and second model.
The novelty of this paper is the inclusion of the bottom clearance volume equation
in the structure volume equation for more accurate results and to spotlight more the
influence of the shift coefficient on optimization results.

2 Problem Definition

The aim of this paper is the volume minimization of one simple single-stage spur
gear system transmission, then analyze the effect of the bottom clearance volume
equation, and profile shift factor of the pinion and the gear on the optimization result.

2.1 Gear Basic Volume Equation for the First Model

The basic equation of the gear volume can be written as below:

 
V = f (m, b, Z 1 ) = π b (m Z )2 − d B2 /4 (1)

2.2 Gear Volume Equation with Bottom Clearance Equation

Volume Using the Profile Shift Factor X for Second
Model

When the reference line of the basic rack is at a distance (r + y) from the spur
gear center, and conjugate with the gear involute part of the profile teeth. Then the
spur gear became a corrected gear (Non-standard), with a profile shift y [14]. The
clearance is the difference between the addendum of one gear and the addendum of
the mating gear [15].
In the solid-structure of spur gear volume computation, the tooth part is the most
complicated part to calculate. That’s because the tooth part begins from the root
circle to the pitch circle. The bottom clearance is created when the bottom of the first
gear tooth and the tip of the second operation gear tooth don’t match (Fig. 1). Sa is
the tooth tip thickness of the mating spur gear (Eq. (3)). To simplify the calculations
Sa is considered as thickness or broad of the clearance cube. Z is the number of gear
teeth, and b is the gear face width.
The clearance volume V c is described by the Eq. (2), with r is the pitch radius,
r a is the addendum radius, r b is the base radius, α is the pressure angle, and S is the
thickness of the teeth at the pitch circle.
40 S. Belarhzal and E. M. Boudi

Fig. 1 The bottom clearance between two spur gears teeth in contact

Vc = Sa ∗ h ∗ b ∗ Z (2)

ra
Sa = S ∗ − 2ra ∗ (invαa − invα) (3)
r
The relation between the profile shift of gear and the thickness of its teeth is
presented in the Eq. (4):
π 
S=m∗ + 2xtanα (4)
2
 
Z Z Z
ra = m ∗ + x + 1 ; r = m ∗ ; r = m ∗ ∗ cosα (5)
2 2 2
rb
invαa = tanαa − αa ; αa = ar ccos (6)
ra

h is the height of the bottom clearance. It presents the radial distance between the
top of the first spur gear and the bottom of the second gear in contact. For corrected
spur gears the height of the bottom clearance is written as:

h = (0, 25 − 2x) ∗ m (7)

Using the bottom clearance volume while taking the profile shift factor in
consideration, transform the equation of gear volume into equation right bellow:
 
V = f (m, b, Z 1 ) = π b (m Z )2 − d B2 /4 − Vc (8)

3 Method

This optimization research combine between genetic algorithms method and ISO
6336:2006 Standards for spur gear strength calculations. The author’s first objective
Genetic Algorithm Design Optimization … 41

was to study the effect of the profile shift factor on structure volume. For this reason,
besides modules, pinion number of teeth, and teeth face width used in the first model,
pinion profile shift factor and wheel profile shift factor were included as functions
variable for the second model, using genetic algorithms. The optimization results for
the first and the second model were compared.
Corrected gears (with positive profile shift factor) use an extended part of the
involute to form the tooth flank compared to standard gears; (without profile shift
factor), which leads to the creation of the clearance between tooth profiles in contact
while meshing. In this paper, the authors analyze the influence of positive values
of shift coefficient on the clearance between teeth in contact. The bottom clearance
equation was considered as a fitness function besides the structure volume equation
for a multi-objective optimization using genetic algorithms. The aim is to consider
the variation of the entire volume structure as well as the bottom clearance volume
according to profile shift modifications.
The same input power, input speed, and shaft diameter are used for the three
models. The module, the face width, the number of pinion teeth, and the profile shift
factor for the gear are limited, to have possible results in practical situations. Data
about boundary conditions are presented in Table 1. The profile shift factor for the
pinion is related to the tip tooth thickness by the equation:

x1max = 0.4m (9)

3.1 Genetic Algorithm

The objective of this paper is the minimization of the spur gear volume, which is
considered as an objective function for the genetic algorithm. To come up with the
results, the genetic algorithm toolbox of software MATLAB is used [16].
The Table 2 shows the chosen input power, torque, and rotational speed. The
gear number of teeth Z 2 is calculated using the transmission ratio i. The material
considered for this structure is the standard steel with ρ = 7830 kg/m3 as density.
The pinion and gear shaft diameters are fixed on 20 mm.
In this paper, the population size adopted is 1000 individuals with a random
vector for the initial population variables instead of the uniform vectors [17]. For
more precision, the optimization was repeated five times for each volume model.

3.2 The Constraint Functions and Strength Calculation

Most spur gear failures came from bending stress on the root of the teeth and surface
durability limited by the contact stress between teeth in contact while meshing [18].
For this reason, the contact stress and the bending stress are considered as constraints
42 S. Belarhzal and E. M. Boudi

Table 1 Optimization
Input variables Range
boundary conditions
Module m [mm] [2, 10]
Pinion number of teeth Z 1 [18, 30]
Face width b [14, 65]
Pinion profile shift factor x 1 [0.2, 0.7]
Gear profile shift factor x 2 [0.2, 0.4]

Table 2. GA optimization
Parameter Value
input data
Pinion shaft diameter [mm] 20
Gear shaft diameter [mm] 20
Input power [w] 5000
input speed [tr/min] 1500
Poisson’s ratio 0.3
Elasticity factor [GPa] 210
Torque [N.mm] 35000
Transmission ratio i 2.5
Maximum bending stress σ F max [MPa] 750
Maximum contact stress σ Hmax [MPa] 450

to satisfy the gear strength conditions (10), and (11):


Wt i + 1
g1 (x) = Z H Z E Z ε Z β K A K v K Hβ K H α − σ Hlim (10)
bd i
Ft
g2 (x) = Z F Z S Z ε Z β − σ F Lim (11)
b.m
The contact ratio factor Z ε is calculated for each generation since it is directly
influenced by the profile shift factor value [19]. The helix factor Z β is equal to 1 for
the spur gears [20]. The zone factor Z H , Elasticity factor Z E , the stress factor Z S , and
the form factor Z F are calculated using the ISO 6336:2006 Standards.
k is the face width coefficient, with 6 < k < 12 and b = k.m, which lead to tooth
face with constraints (12), and (13):

b
g3 (x) = 6 − (12)
m
b
g4 (x) = − 12 (13)
m
Genetic Algorithm Design Optimization … 43

Changing the profile shift value could increase or decrease the contact ratio CR
value [21], therefore with a greater number of CR corrections the gearing action is
disturbed by tooth gear tip interference [22]. To avoid this undesirable phenomenon
the tooth interference constraints are used (14), and (15):

17 − Z 1
g5 (x) = − x1 (14)
17
17 − Z 2
g6 (x) = − x2 (15)
17
i is the transmission ratio.
Z2
i= = 2, 5 (16)
Z1

4 Results and Discussion

The optimization without profile shift factor converged faster. The program running
was for more than 30 times and didn’t present any deviation. First of all, the gener-
ation number fixed on 500 generations, then reduced to 100 generations to decrease
the computation time. The computation time could be influenced too by the initial
population. Generally, using pseudo-random numbers for initial population gives
accurate results in less time [23].
The results of the genetic algorithm optimization for both equation volume of
one-stage spur gear are presented in Table 3. It showed that the smallest volume was
obtained by including the bottom clearance volume and the profile shift factor in the
spur gear equation volume. It was reduced by 48%, the volume minimization is clear.
Increasing the profile shift factor by the amount of a positive value leads to a
higher clearance volume between the gear tooth flank as shown in Fig. 2. It can be
justified by the difference in shape between standard and corrected spur gear teeth.
During the meshing process, a prolonged part of the involute is taken as a flank.
Which explains the larger space between gear and pinion in contact.
As shown in Fig. 3, the Pareto front illustrates the best optimization solutions
between global structure volume and clearance volume. As the volume increases, so
does the clearance volume.
44 S. Belarhzal and E. M. Boudi

Table 3. Output AG optimization variables

Function volume variables Standard gears Corrected gears
Number of teeth of the pinion Z1 27 25
The module m 4,7 4,2
The face width b 46 33
The face width coefficient k 9,8 7,8
The profile shift factor of the pinion x1 0 0,34
The profile shift factor of the gear x2 0 0,14
The structure volume V [mm3] 5114600 2640900
The clearance volume %V c 0,15 0,18

Fig. 2 Impact of profile shift factor on the percent of clearance volume compared to the global
structure volume

Fig. 3 Pareto front of GA multi-optimization for global structure volume and clearance volume
for corrected spur gear pair
Genetic Algorithm Design Optimization … 45

5 Conclusion

Gears are the most important part in the transmission power field. They are widely
working in the industry. For this reason, many researchers worked on how to improve
their different aspects. Gears design is a complicated problem to treat. It contains
several variables and deals with different constraints. Genetic Algorithms present a
metaheuristic approach to treat complicated gearbox design that cannot be solved
with classical methods.
In this work, the effect of the profile shift factor for pinion and gear on the genetic
algorithm volume and bottom clearance volume optimization of one-stage of the
spur gear are studied. The results are significant since the low volume was obtained
using the profile shift factor on the spur gear volume equation. With a difference of
48% less than the first model using only the module m, the face width b, and the
pinion profile shift factor Z 1 as function variables.
Including the clearance volume equation helps to have more accurate optimization
volume results and avoid interference problems that appear with corrected gears since
V c increases while increasing the profile shift factor.
It is to be noted that this study has a limitation concerning the constraint functions
involved in the genetic algorithm. Spur gear’s dynamic properties were not considered
but will be developed in future studies.

References

1. Bensghaier A, Romdhane L, Benouezdou F (2012) Multi-objective optimization to predict

muscle tensions in a pinch function using genetic algorithm. CR Mec 340:139–155
2. Back T (1966) Evolutionary Algorithms in Theory and Practice: Evolution Strategies,
Evolutionary Programming, Genetic Algorithms. Oxford University Press, Oxford (1966)
3. De Jong KA (1992) Are genetic algorithms function optimizers? Elsevier, The Netherlands,
pp 3–13
4. Back T, Hammel U, Schwefel HP (1997) Evolutionary computation: comments on the history
and current state. IEEE Trans Evol Comput 1:3–17
5. Graja O, Zghal B, Dziedziech K, Chaari F, Jablonski A, Barszcz T, Haddar M (2019) Simulating
the dynamic behavior of planetary gearbox based on improved Hanning function. CR Mec
374:49–61
6. Gologlu C, Zeyveli M (2009) A genetic approach to automate preliminary design of gear drives.
Comput Ind Eng 57:1043–1051
7. Golabi S, Fesharaki JJ, Yazdipoor M (2014) Gear train optimization based on minimum
volume/weight design. Mech Mach Theory 73:197–217
8. Tamboli K, Patel S, George PM, Sanghvi R (2014) Optimal design of a heavy duty helical gear
pair using particle swarm optimization technique. Procedia Technol 14:513–519
9. Savsani V, Rao RV, Vakharia DP (2010) Optimal weight design of a gear train using particle
swarm optimization and simulated annealing algorithms. Mech Mach Theory 45:531–541
10. Diez-Ibarbia A, Fernandez del Rincon A, Iglesias M, de-Juan A, Garcia P, Viadero F (2015)
Efficiency analysis of spur gears with a shifting profile. Meccanica 51:707–723
11. Rai P, Agrawal A, Saini ML, Jodder C, Barman AG (2018) Volume optimization of helical
gear with profile shift using real coded genetic algorithm. Procedia Comput Sci 133:718–724
46 S. Belarhzal and E. M. Boudi

12. Miler D, Lončar A, Žeželj D, Domitran Z (2017) Influence of profile shift on the spur gear pair
optimization. Mech Mach Theory 117:189–197
13. Miler D, Žeželj D, Lončar A, Vučković K (2018) Multi-objective spur gear pair optimization
focused on volume and efficiency. Mech Mach Theory 125:185–195
14. Michael GM (2013) Effect of change of contact ratio on contact fatigue Michael Gmariam
these. Addis Ababa University (2013)
15. Colbourne JR (1987) The Geometry of Involute Gears. Springer, New York
16. Mei W, Na J, Yang F, Shen G, Chen J (2016) The optimal design method and standardized
mathematical model of tooth profile modification of spur gear. Math Probl Eng 1–7. https://
doi.org/10.1155/2016/6347987
17. Maaranen H, Miettinen K, Mäkelä MM (2004) Quasi-random initial population for genetic
algorithms. Comput Math Appl 47:1885–1895
18. Samya B, Bachir A, Boudi EM, Amarir I (2019) The effect of addendum factor on contact ratio
factor and contact stress for spur gears. In: 2019 7th international renewable and sustainable
energy conference (IRSEC), Agadir, Morocco, pp 1–6. https://doi.org/10.1109/IRSEC48032.
2019.9078205
19. Samya B, Boudi EM (2020) Profile shift factor’s effect on contact stress and natural frequencies
of spur gears using the finite element method. Int Rev Model Simul (IREMOS) 13:1–12
20. Samya B, Boudi EM, Bachir A, Amadane Y (2020) Analysis of profile shift factor’s effect on
bending stress of spur gears using the finite element method. In: 2020 IEEE 6th international
conference on optimization and applications (ICOA), Beni Mellal, Morocco, pp 1–6. https://
doi.org/10.1109/ICOA49421.2020.9094486
21. Gebremariam M, Thakur A, Leake E, Tilahun D (2018) Effect of change of contact ratio on
contact fatigue stress of involute spur gears. Int J Current Eng Technol 8:719–731
22. Maiti R, Roy AK (1966) Minimum tooth difference in internal-external involute gear pair.
Mech Mach Theory 31:475–485
23. Maaranen H, Miettinen K, Penttinen A (2007) On initial populations of a genetic algorithm
for continuous optimization problems. J Glob Optim 37:405. https://doi.org/10.1007/s10898-
006-9056-6
A Robust Method for Face Classification
Based on Binary Genetic Algorithm
Combined with NSVC Classifier

M. Ngadi, A. Amine, B. Nassih, Y. Azdoud, and A. El-Attar

Abstract Image features selection consists on reducing the number of features of

an image, used for training or testing a classifier, by eliminating irrelevant, noisy
and redundant data without decreasing significantly the prediction accuracy of the
classifier. In this paper we consider the problem of feature selection. We give a quick
technique dependent on a binary Genetic Algorithm (GA) combined with Neigh-
boring Support Vector Classifier NSVC. More concretely, we utilize the k-Nearest
Neighbors KNN classifier as a fitness function for each feature selection decision
during the evaluation step of the genetic algorithm. GA-NSVC is implemented and
tested on BOSS and MIT-CBCL Face dataset. The results show its robustness and
high performances.

Keywords NSVC · KNN · Feature selection · Genetic algorithm

1 Introduction

The issue of choosing a subset of factors can be viewed as an inquiry in a space of

speculations (called a bunch of potential arrangements). The nature of a chose subset
is evaluated by an exhibition model. On account of a directed characterization issue,
this basis is regularly the exactness of a classifier (classification rate) worked from
the arrangement of chose factors.
Choosing a subset of qualities disposes of insignificant and excess data relying
upon the standards utilized.
GA is probably the most recent strategy in the field of the element determination
[1, 2]. It comes to locate a particular capacity that permits a decent separation among
chromosomes and characterize the hereditary administrators [3, 4].
Our objective is to evaluate the ability of the GA-NSVC for the improvement of
the classification rate of images.

M. Ngadi (B) · A. Amine · B. Nassih · Y. Azdoud · A. El-Attar

Systems Engineering Laboratory, ENSA, Ibn Tofail University of Kenitra, Kenitra, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 47

S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_5
48 M. Ngadi et al.

We have chosen to use the NSVC of the classification function [5]. Next, to
demonstrate the effectiveness of our proposed algorithm GA-NSVC, we have tested
on two real Face datasets.
Our trials on the distinctive datasets show that our strategy prevailing with regards
to lessening the quantity of highlights while keeping up great precision surpassing
99%.
The remainder of this article is coordinated as follows: The essential ideas of
GA are given in area Sect. 2. Section 3 presents the NSVC classifier. In Sect. 4 we
conduct a comprehensive experimental study and we conclude the article and give
perspectives to our work in section.

2 Genetic Algorithm (GA)

Genetic algorithm is an iterative algorithm that search for the optimum of a constant
size population [3, 6]. The constant population size leads to a phenomenon of
competition between chromosomes.
The making of another populace from the past is finished by:

Selection
The choice is a cycle by which a chromosome is replicated in the new populace
dependent on the estimations of the capacity to advance for this chromosome.

Crossover
The straightforward intersection comprises on, right off the bat picking several chro-
mosomes with a likelihood p and afterward in a subsequent advance, cutting the agent
directs in an indistinguishable arbitrary situation in the two guardians. This produces
two “head fragments” and two “tail portions”. At last, we drive two guardians’ tail
portions to get two youngsters who acquire a few attributes of their folks (Fig. 1a).

Mutation
A transformation is characterized as the in-rendition of somewhat in a chromosome
(Fig. 1b). This is comparable to haphazardly change the estimation of a boundary.
Transformations go about as commotion and forestall the development to freeze.
They permit guaranteeing worldwide (just as nearby) search, contingent upon the
weight and the quantity of pieces moved. Moreover, they guarantee numerically that
the general ideal will be accomplished.

Fitness Function
The fundamental objective of highlight choice is to utilize less highlights to get the
equivalent or better presentation [1, 7]. The inquiry’s procedure will likely discover
a component subset boosting this capacity. The hybrid and change capacities are the
primary administrators that haphazardly sway the wellness esteem.
A Robust Method for Face Classification … 49

Fig. 1 Genetic crossover and mutation process

Fitness = K N N _{accuracy}

3 Neighboring Support Vector Classifier

SVM is first presented by Vapnik and partners for the issues of arrangement and
relapse can be viewed as another preparation method dependent on conventional
straight, polynomial and Spiral Premise Capacity (RBF).
In any case, the suspicion that the preparation information are indistinguishably
produced from obscure likelihood circulations may restrict the use of SVM to the
issues of regular daily existence.
To loosen up the supposition of indistinguishable circulation, the NSVC [8,
9] utilizes a bunch of vicinal centers capacities fabricated dependent on directed
grouping in the element space actuated by the portion.
Signifying a non-direct change of the info space X to a high-dimensional space
utilizing a piece work as:

 : Rn → F

xi → (xi ), i = 1, . . . , l

Where (xi ) is the changed point xi .

All preparation information focuses are circulated in c regions/groups in the
component space, where k (z) is the focal point of mass of the k-th region dwelling
50 M. Ngadi et al.

in F. This is a comparable portrayal to bunching dependent on the trademark space

of k-means:


l
k = αki z i , k = 1, 2, . . . , c (1)
i=1

Where c is the quantity of groups, αki are the boundaries to be characterized by

the bunching strategy (SKDA) and zi = yi (xi ) signifies the information focuses
named in the component space [10, 11].
That is, the issue of bunching by SKDA [12, 13] is formulated as minimizing.

F = J − TH (2)

Where J is the distortion function, T are the Lagrange multiplier and H is

Shannon entropy [14, 15].
To decide the αki parameter, the free energy work F is limited regarding the
likelihood of affiliation, which is identified with the Gibbs dispersion.

p(zi )p(k |zi )

αki = l     (3)
j=1 p zj p k |zj

Let characterize the mono- and bi-vicinal kernels as:


l
(x) = yi αki K(x, xi ) + b , k = 1, 2, . . . . . . , K (4)
i=1


l 
l
 
Mkm (x) = yi yj αki αmj K xi , yj + b , k, m = 1, 2, . . . . . . , K (5)
i=1 j=1

The decision boundary is:


c
f(x) = βk yk Lk (x) + b (6)
k=1
A Robust Method for Face Classification … 51

Fig. 2 Typical face images of BOSS dataset

Fig. 3 Face pictures used for classification

4 Results and Discussion

4.1 Datasets Description

BOSS Dataset contains 2000 images with 949 faces and 1051 non-faces. Figure 2
shows some face pictures in the preparation and test sets.
MIT-CBCL Face Dataset: is a preparation set comprising on 6,977 trimmed
pictures (2,429 countenances and 4,548 nonfaces). Figure 3 shows some face pictures
in the preparation and test sets.

4.2 NSVC Classifier

The classification accuracy without dimension reduction on the two datasets is given
in Table 1:
52 M. Ngadi et al.

Table 1 The order precision

Dataset Kernel Parameter kernel Accuracy (%)
results With NSVC
BOSS Poly q=2 99.95
MIT-CBCL Poly q=2 99.79

Table 2 Boundaries utilized

GA parameter Value
in GA
Population size 100
Genome length 20
Population type bitstrings
Fitness function KNN
Number of generations 100
Crossover Arithmetic crossover
Crossover probability 0.8
Mutation Uniform mutation
Mutation probability 0.1

The NSVC can improve the quality of classification, confirming that NSVC is
performing well.

4.3 Genetic Algorithm Based Feature Selection

In view of the GA setup (Table 2), the accompanying outcomes were acquired. The
deliberately picked wellness work empowered the GA to limit grouping mistake
from KNN.
Figures 4 and 5 shows the development of the fitness estimation of the best chro-
mosome in every age. This curve shows that the error in the classification for the best
individual that represents the subset of features to select decreases with changes in
population at every generation.

4.4 GA and NSVC Classifier Parameters

We used the GA approach to select a subset of features for the NSVC classifier
(Fig. 6).
The characterization precision on the two subsets is given in Table 3:
Our experiments on different databases show that our method has succeeded in
reducing the number of features while maintaining a good precision which exceeds
99%.
A Robust Method for Face Classification … 53

Fig. 4 GA-KNN simulation diagram on BOSS dataset

Fig. 5 GA-KNN simulation diagram on MIT-CBCL dataset

Finally, we can confirm the effectiveness of our method to select the most relevant
features and give a better classification.
54 M. Ngadi et al.

Fig. 6 The phases of our methodology

Table 3 Precision results

Subset Kernel Parameter kernel Accuracy (%)
With GA-NSVC
BOSS Linear – 99.80
MIT-CBCL Poly q=2 99.37

5 Conclusion

In this paper another methodology dependent on GA for feature selection is intro-

duced. The calculation of the fitness function is done using the KNN classifier for
each primitive.
This method was tested on two real datasets using NSVC classifier. The results
show the robustness of our approach GA-NSVC, with a precision exceeding 99%.
Our objective sooner rather than later is to proceed with the investigation of the
AG to test it on a few datasets from other examination regions, and attempt to locate
the best trade off among precision and execution time.

References

1. Babatunde O et al (2014) A genetic algorithm-based feature selection. Int J Electron Commun

Comput Eng 5(4):2278–4209
2. Oluleye B, Leisa A, Leng J, Dean D (2014) Zernike moments and genetic algorithm: tutorial
and application. Br J Math Comput Sci 4(15):2217–2236
A Robust Method for Face Classification … 55

3. Amine A, El Akadi A, Rziza M (2009) Aboutajdine D (2009) GA-SVM and mutual information
based frequency feature selection for face recognition. INFOCOMP J Comput Sci 8(1):20–29
4. Gunavathi C, Premalatha K (2014) Performance analysis of genetic algorithm with KNN and
SVM for feature selection in tumor classification. Int J Comput Electr Autom Control Inf Eng
8(8):1490–1497
5. Yang X, Cao A, Song Q, Schaefer G, Su Y (2014) Vicinal support vector classifier using
supervised kernel-based clustering. Artif Intell Med 60(3):189–196
6. Huang C-L, Wang C-J (2006) A GA-based feature selection and parameters optimization for
support vector machines. Expert Syst Appl 31:231–240
7. Mathworks T (2013) Statistics toolbox user’s guide. The MathWorks, Inc, Natick
8. Ngadi M, Amine A, Hachimi H, El-Attar A (2016) A new optimal approach for breast cancer
diagnosis classification. Int J Imaging Robot 16(4):25–36
9. Cao A, Song Q, Yang X (2004) Mammographic mass detection by vicinal support vector
machine. In: IEEE international joint conference on neural networks, Budapest, Hungary, vol
3. IEEE, pp 1953–1958
10. Ngadi M, Amine A, Nassih B, Hachimi H, El-Attar A (2016) The performance of LBP and
NSVC combination applied to face classification. Appl Comput Intell Soft Comput 2016:10.
Article ID 8272796
11. Ngadi M, Amine A, Nassih B, Hachimi H, El-Attar A (2017) Uniformed two local binary
pattern combined with neighboring support vector classifier for classification. Int J Artif Intell
15(2):102–115
12. Ngadi M, Amine A, Hachimi H, El-Attar A (2017) Vicinal support vector classifier: a novel
approach for robust classification based on SKDA. Pattern Recognit Image Anal 27(3):403–411
13. Ngadi M, Amine A, Nassih B, Hachimi H (2019) A highly efficient system for mammographic
image classification using NSVC algorithm. Procedia Comput Sci 148:135–144
14. Nassih B, Ngadi M, Amine A, El-Attar A (2018) New proposed fusion between DCT for
feature extraction and NSVC for face classification. Cybern Inf Technol 18(2):89–97
15. Ngadi M, Amine A, Nassih B, Hachimi H, El-Attar A (2018) Intelligent classification of heart
disease using neighboring support vector classifier. Int J Tomogr Simul 31(4):54–64
A Proposed Solution to Road Traffic
Accidents Based on Fuzzy Logic Control

Halima Drissi Touzani, Sanaa Faquir, Saloua Senhaji, and Ali Yahyaouy

Abstract Road Traffic accidents are one of the main problems causing death and
fatal injuries in all countries. Several general suggestions were presented to the
ministry of transport to try to overcome this problem. This paper discusses a specific
approach based on fuzzy logic control to train semi-autonomous cars to prevent
accidents from occurring. These types of cars equipped with many devices along
with decision making algorithms based on learning will make cars more intelligent
to take the right decision when no reaction of the driver is noticed.

Keywords Road accidents · Fuzzy logic control · Data analytics ·

Semi-autonomous cars

1 Introduction

National Highway Traffic Safety Administration (NHTSA), states that more than
90% of the crashes are caused by human errors, and 40% of deadly ones imply drunk,
drugged, tired or distracted drivers [1, 2]. Therefore the technology of Autonomous
vehicles has the possibility to impact vehicle safety, traffic jam and travel behavior
[2], so to avoid potential collision and vehicle accidents lot of studies has been applied
on cars system to became automatic or semi-automatic, Triveni et al. equipped the
car with an ultrasonic sensor which will continuously track for any obstacles from
the front side, In this way they can ensure that a safe distance is always maintained
between the two cars and thus accident can be avoided [3].
Keeping up with evolution, these last years’ vehicles are being more equipped
with AEB (Automatic Emergency Braking) systems as stated in Erik et al. article,
they have also described one of the latest AEB systems called Collision Warning with

H. D. Touzani (B) · S. Faquir · A. Yahyaouy

LISAC Laboratory, Faculty of Science Dhar Mehraz, Sidi Mohamed Ben Abdallah University,
Fez, Morocco
S. Faquir · S. Senhaji
LSED Laboratory, Faculty of Sciences of Engineering, Private University of Fez, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 57

S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_6
58 H. D. Touzani et al.

Full Auto Brake and Pedestrian Detection (CWAB-PD) which helps drivers avoid
both rear-end and pedestrian accidents by providing a warning and, if necessary,
automatic braking using full braking power [4].
To find solutions to traffic accidents, data should be collected and analyzed to
discover insights from it and present solutions to the government. In the previous
paper, a research study on road accidents in the country of Morocco was conducted
using a real database provided by the Ministry of transport and equipment’s and
composed of 1953 Instances of data on road accidents taken in 2014. The data
has been maintained using data transformation techniques such as filters and then
analyzed by three different methods, MS Excel, K-means classifier. The two methods
showed almost similar results presenting reasonable causes on different types of
accidents which can help to provide valuable suggestions to the Ministry to reduce
road accidents in the country like for example [5]:
• Install roundabouts that help reduce collisions and control speed.
• Build overpasses and underpasses in the places knowing more traffic.
Since the above mentioned suggestions are general and mostly related to infras-
tructure, a second solution is proposed which relies on using the data gathered to train
Semi-autonomous cars to be able to take decisions. These types of cars are equipped
with many devices along with decision making algorithms based on learning such
as deep learning or linguistic variables such as fuzzy logic will make cars more
intelligent to take the right decision when no reaction of the driver is noticed.
Fuzzy logic system have the potential to improve the performance of a system
by applying reasoning and are able to describe complex systems in linguistic terms
rather than numerical values [6]. This approach is very suitable in several areas within
many researches as it represents data by rules in place of using specific equations
[7].
Fuzzy logic has been excelled in various domain such as medicine, Jason et al.
describe how this method works by representing its application in a simplified model
of fluid resuscitation of intensive care unit patients [8]. Blood pressure has been
used for regulation of depth of anesthesia by fuzzy logic control and in some cases
exceeded a human operator [9]. In energy domain, Faquir et al. adopted Fuzzy logic
to control the energy between production and consumption and also confirm the
availability of power on demand for an isolated house in the city of Essaouira in
Morocco [10]. In electric vehicle Rachid et al. apply fuzzy control to ensure a longer
life for a hybrid electrical vehicle by controlling the flow of energy inside [11].
We propose in this paper an approach based on Fuzzy logic control to help semi
autonomous cars to decide and take action when it is needed.
A Proposed Solution to Road Traffic Accidents … 59

Table 1 Results provided by K-means algorithm

Day Night
Rate Type of Initial Movement Rate Type of Initial Movement
collision shock collision shock
15% By the side Front By inserting 17% By the Back Same sense
back
12% By the side Front Against the 10% Other Front Against the
Left direction direction
12% By the side Front By inserting 8% By the side Front Same sense
right
9% By the side Front By inserting 6% Front Back Crossing the
left left road
7% Front Front Left changing 5% By the Front Same sense
wire back

2 Summary of Data Analytics Phase

We based our work on the previous result which analyzed 1953 instances using
statistics graphs and k-means classifier to find similarities between data and extract
different types of accidents that frequently happen and important shock points [5].
The most important similarity criteria determined by K-means algorithm are
described in Table 1. Based on the previous statistical analysis it was found that
the majority of accidents occur during the day, so we decided to partition the results
by day and night, we noticed also that the k-means algorithm could identify the most
important attributes that represents the 10 different classes and one of the attributes
was day/night. Other attributes such as the type of collision, the initial shock point
and the movement of the vehicles in the road were also considered as criteria to sepa-
rate different classes between them, so the majority of accidents that have occurred
in the day have a type of collision by the side of the car, front and by entering the
road. On the other hand, the accidents that happened at night took place between
cars of the same sense [5].

3 Fuzzy Logic Control

Fuzzy logic is a method of reasoning convenient to natural language and human

thinking than the traditional logical systems. Basically, it provides an effective means
of capturing the approximate, inexact nature of the real world [12]. Historically, the
term of fuzzy logic was used for the first time by the professor Lotfi Zadeh in 1956,
when he observed that the logic of conventional computer was not able to manipulate
data which represent subjective or unclear human ideas [13].
Fuzzy logic architecture is composed from four main parts as represented in Fig. 1.
60 H. D. Touzani et al.

Fig. 1 Architecture of fuzzy logic

• Rules: Fuzzy Inference Rules are the most important operation of the fuzzy infer-
ence system, including a number of inference rules connecting the input fuzzy
variables to the output fuzzy variables [7], and the experts offered all the rules
and the IF-THEN conditions to control the decision making system [13].
• Fuzzification: where the fuzzifier performs the act of fuzzification, it transform
a numeric value to a fuzzy set [14], and for further processing the crisp inputs
are measured by sensors and passed into the control system. For example Human
body temperature, pressure, etc. [13].
• Inference Engine: in this step we could determine the degree of match between
fuzzy input and the rules. Based on the % match, it helps to determine which rules
need implement according to the given input field. And then the applied rules are
combined to develop the control actions [13].
• Defuzzification: which is an important operation in the process of fuzzy set,
because it transforms a fuzzy set to numeric value [14].

4 Semi-autonomous Car

A semi-autonomous car is a vehicle that dispenses with the action of the driver
in many areas. It is equipped with several sensors and software that automatically
cooperate using artificial intelligence. It will use various techniques to sense its
surrounding, such as radar, Lidar, and camera.
Lidar (Light Sensing and Telemetry), It is a technology which is based on remote
sensing that calculates distance by illuminating a target with a light beam and analyzes
the reflected light. It is mounted on the roof of the vehicle on a cylindrical enclosure
which rotates 360° [15].
Lidar consists of a transmitter, a mirror and a receiver, It can also use laser,
ultraviolet and infrared light to create images of objects.
RADAR (Radio Detection and Ranging) is becoming an important automotive
technology. This equipment helps to estimate the mutual speed of the vehicle and
the object, based on electromagnetic waves. Radar technologies generally used are:
frequency-modulated continuous wave radar or FMCW, pulse Doppler radar [16].
A Proposed Solution to Road Traffic Accidents … 61

VIDEO CAMERAS: These are attached to the top of the front window, near the
rearview mirror and reproduce real-time 3D images of the road in front of you. These
are adopted to detect traffic lights and signs, unforeseen everything like pedestrians,
or animals…. [17].
A semi-autonomous car will equipped with such devices along with decision
making algorithms based on learning such as deep learning or linguistic variables
such as fuzzy logic will make cars more intelligent to take the right decisions when
no reaction of the driver is noticed.

5 Proposed Solutions

Based on the results we found by both Excel graphs and K-means classifier algorithm,
we provided the following suggestions to the ministry of transportation to help reduce
accidents:
• Widen the most frequented roads, and install warning signs that alert drivers
who exceed the speed limit, especially where there are schools, Public services,
hospitals…
• More control over non-controlling intersection and giving more priority to turning.
• Install roundabouts that help reduce collisions and control speed.
• Build overpasses and underpasses in the places knowing more traffic.
Aging causes a loss of meaning. For example, an elderly person sees a sign only
65 m away, while a 20-year-old sees it at 100 m. So, it is important for the ministry
to work on the visibility and legibility of signage [5].
Concerning the behavior of the driver it remains relative to the education received
by each individual as well as to the values acquired by each one of them.
Since all of these suggestions are general and most of them requires building a new
infrastructure, a second solution is proposed which relies on using the data gathered
to train Semi-autonomous cars that can take decisions like breaking, reducing speed,
and stopping when it is needed in case the driver didn’t take any related action
and this to limit accidents and help ensure greater road safety. This system will be
implemented by the fuzzy logic control.

6 The Proposed Approach

Based on results presented earlier in the data analytics phase; a scenario was generated
to help a semi autonomous car take right decisions when there is a bad or no reaction
from the driver which will prevent accidents from happening.
The Flowchart presented in Fig. 2. represents the parameters which are Type of
collision, initial shock point, Movement, braking system in S (second), and depending
on the case an action will be generated, if the driver takes action its will be no
62 H. D. Touzani et al.

Fig. 2 Flowchart presenting our approach

reaction from the semi-autonomous car otherwise the car will take action either
braking immediately, braking and attention left or right or reducing speed.
This scenario will be Implemented using fuzzy logic control since the decisions
can be based on linguistic values.

7 Conclusion

Thousands of people die in traffic crashes yearly; this involves not only loss of human
life but also property damage, this makes road safety one of the most studied subjects
nowadays. In this paper, a data analytics was done by three different tools to analyze
a big database of road traffic accidents collected by the ministry of transportation in
Morocco. The purpose of the analysis was to identify the hidden parts in the data
and locate the most important initial points and the types of intersection in the road
that mostly causes accidents. As a better solution, a semi-autonomous car equipped
A Proposed Solution to Road Traffic Accidents … 63

with different types of sensors and cameras connected to intelligent models can help
the car take the right decisions when there is no driver reaction. This paper presented
an algorithm based on fuzzy logic to consider different scenarios and try to help the
car take the right decision and prevent the accident from occurring when the driver
doesn’t react.
Our future work will be the implementation of the algorithm and try it on different
scenarios.

References

1. NHTSA | National Highway Traffic Safety Administration. https://www.nhtsa.gov/. Accessed

27 Jun 2020
2. Fagnant DJ, Kockelman K (2015) Preparing a nation for autonomous vehicles: opportunities,
barriers and policy recommendations. Transp Res Part A: Policy Pract 77:167–181
3. Shinde T (2013) Car anti-collision and intercommunication system using communication
protocol. Int J Sci Res (IJSR) 2(6):187–191
4. Coelingh E, Eidehall A, Bengtsson M (2010) Collision warning with full auto brake and
pedestrian detection - a practical example of automatic emergency braking. In: IEEE conference
on intelligent transportation systems, proceedings, ITSC, pp 155–160
5. Drissi Touzani H, Faquir S, Yahyaouy A (2020) Data mining techniques to analyze traffic acci-
dents data: case application in Morocco. In: The fourth international conference on intelligent
computing in data sciences ICDS 2020 conference, pp 1–3
6. Zadeh LA (1975) The concept of a linguistic variable and its application to approximate
reasoning-III. Inf Sci 9(1):43–80
7. Driss M, Saint-Gerand T, BenSaïd A, Benabdeli K, Hamadouche MA (2013) A fuzzy logic
model for identifying spatial degrees of exposure to the risk of road accidents (Case study of
the Wilaya of Mascara, Northwest of Algeria). In: 2013 international conference on advanced
logistics and transport, ICALT 2013, pp 69–74
8. Bates JHT, Young MP (2003) Applying fuzzy logic to medical decision making in the intensive
care unit. Am J Respir Crit Care Med 167(7):948–952
9. Zbinden AM, Feigenwinter P, Petersen-Felix S, Hacisalihzade S (1995) Arterial pressure
control with isoflurane using fuzzy logic. Br J Anaesth 74(1):66–72
10. Faquir S, Yahyaouy A, Tairi H, Sabor J (2016) Energy management in a hybrid PV/wind/battery
system using a type-1 fuzzy logic computer algorithm. Int J Intell Eng Inform 4(3–4):229–244
11. El Amrani R, Faquir S, Yahyaouy A, Tairi H (2018) Modelling and implementation of an
energy management simulator based on agents using optimised fuzzy rules: Application to an
electric vehicle. Int J Innovative Comput Appl 9(4):203–215
12. Lee CC (1990) Fuzzy logic in control systems: fuzzy logic controller—part I. IEEE Trans Syst
Man Cybern 20(2):404–418
13. Guru99. Fuzzy Logic Tutorial: What is, Application & Example. https://www.guru99.com/
what-is-fuzzy-logic.html. Accessed 08 Oct 2020
14. Roychowdhury S, Pedrycz W (2001) A survey of defuzzification strategies. Int J Intell Syst
16(6):679–695
15. Ondruš J, Kolla E, Vertaľ P, Šarić Ž (2020) How do autonomous cars work? Transp Res Procedia
44(2019):226–233
16. Sarkan B, Stopka O, Gnap J, Caban J (2017) Investigation of exhaust emissions of vehicles
with the spark ignition engine within emission control. Procedia Eng 187:775–782
17. Yun HS, Kim TH, Park TH (2019) Speed-bump detection for autonomous vehicles by lidar
and camera. J Electr Eng Technol 14(5):2155–2162
Multi-objective Archimedes
Optimization Algorithm for Optimal
Allocation of Renewable Energy Sources
in Distribution Networks

Ahmad Eid and Hassan El-Kishky

Abstract This paper presents the optimal allocation of the distributed generations
(DGs) into the distribution networks (DNs) using the Archimedes Optimization Algo-
rithm (AOA). The AOA optimizes three DGs operating at unity power factor repre-
senting the PV renewable energy sources (RES). The standard 33-bus DN is used as
a test system to verify the AOA’s effectiveness in both single-objective and multi-
objective optimizations. In addition to the overall power loss elimination, the total
voltage deviation (TVD) of the DN is reduced by the optimum allocation of three
DGs. In multi-objective optimization, the Pareto Optimal Front (POF) method is
adopted to determine the non-dominated solutions. A fuzzy linear function decides
the Best Compromise Solution (BCS) between the points set by the POF. The AOA’s
obtained results in both single and multi-objective optimizations are compared to
the Particle Swarm Optimization (PSO) and Atom Search Optimization (ASO) algo-
rithms. The three algorithms are effective in solving the optimization problem or
both single and multiple dimensions. Moreover, the AOA outperforms the ASO and
PSO algorithms in different case studies.

Keywords Archimedes optimization algorithm · Multi-objective optimization ·

Distributed generations

A. Eid (B)
Department of Electrical Engineering, Aswan University, Aswan 81542, Egypt
e-mail: [email protected]
Department of Electrical Engineering, College of Engineering, Qassim University, Unaizah
56452, Saudi Arabia
H. El-Kishky
Electrical Engineering Department, The University of Texas at Tyler, Tyler, TX 75799, USA
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 65

S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_7
66 A. Eid and H. El-Kishky

1 Introduction

In recent years, environmental issues, fuel cost volatility, energy liberalization, and
technological developments have contributed to the rise in distributed generation
(DG) investment. This development has generated fantastic benefits but produced
a range of difficulties in planning and operating distribution networks (DNs). The
primary objective of inserting a DG into any DN is to provide extra real and/or reactive
power feeding the load that reduces the total power supplied by the mains. More
advantages to utilities of DG allocations include power loss and TVD reductions,
enhancing the stability of DNs, increasing the system reliability [1–3]. At the same
time, DG technologies provide a secure operation of the electricity markets during
the deregulation of the power industry. Integration of DG units into DNs plays a
vital role in providing support for reactive power, spinning reserve, and frequency
control. However, poorly designed and incorrectly controlled DGs will contribute to
reversing power flows, increasing power losses, and resulting in feeders’ overload.
Recently, many researchers applied different metaheuristic optimization algo-
rithms to allocate DG units into DNs. In [4, 5], the authors used the particle
swarm optimization (PSO) algorithm or enhanced versions to reduce the distribu-
tion networks’ power losses. Grey wolf optimizer (GWO) algorithm is adopted in
[6, 7] to minimize the power losses and reactive losses and enhance the voltage
profiles of different DNs. A bacterial foraging optimization algorithm (BFOA) has
been applied in [8, 9] to optimally determine the size and site of multiple DG units
with DNs. The study also included allocating STATCOM devices with different load
models for power loss minimization, voltage profile improvement, and increasing
the reliability of the systems. A hybrid of ant and artificial bee colonies is applied
in [10] for optimal placement and sizing of different DG sources to minimize the
power losses and energy costs and enhance voltage stability. An enhanced artificial
ecosystem-based optimization (EAEO) algorithm is developed and applied in [11]
for optimal allocations of multiple DG units to reduce the power losses and TVD and
enhance the system stability. The analysis included the application of the DG under
fixed, unity, and optimal power factors. The authors of [12] applied a combination
of the genetic multi-objective solver (GMOS) with linear programming to optimize
the DG units attached to DNs for reducing energy price arbitrage, energy losses and
enhance the power quality of the systems. In [13] adaptive PSO and modified grav-
itational search algorithm (GSA) are adopted to optimize the size and locations of
different DG units with single and multi-objective problems. The main objectives
were to reduce the total power losses and TVD and to enhance voltage stability.
A hybrid salp swarm algorithm (SSA) with the loss sensitivity is used in [14] for
renewable energy integration considering the annual load growth.
The authors in [15] have proposed a newly published algorithm known as
Archimedes Optimization Algorithm (AOA). It is one of the most recent nature-
inspired algorithms formulated from the physics Archimedes law of gravity. It
imitates the concept of buoyant force imposed on an entity, partly or entirely
submerged in a fluid, which is proportional to the weight of the fluid being displaced.
Multi-objective Archimedes Optimization Algorithm … 67

The application of AOA for DG allocation in distribution systems has not been
considered yet, which encourages us to implement it to deal with these optimization
problems.
In this paper, the AOA is implemented to determine the optimal sizes and locations
of three DG units operating with UPF to reduce the power losses and TVD of the 33-
bus DN. The obtained results of AOA are compared favorably with other algorithms
such as PSO and ASO. The results show the superiority of the AOA algorithm in the
single- and multi-objective optimization case studies.

2 Problem Formulation

The DG units’ integration to the DNs enhances their performance in terms of loss
reduction and improvement of the voltage profile. The AOA algorithm is imple-
mented to optimally determine the sizes and locations of three DG units to minimize
the total power losses and the total voltage deviation. The DG type includes the
PV operating at unity power factor (PF). The Forward/Backward sweep method is
used for solving the DN before and after the DG integrations. For single-objective
optimization, the branch power loss is calculated as:
br
Ploss = Ibr
2
Rbr (1)

where Ibr represents the branch current and Rbr represents the branch current.
The total power loss of the system, Ploss, is:
 N br
Ploss = 3 br
Ploss (2)
1

where N br is the number of branches and three accounts for the three-phase
system, the power loss objective function is expressed as:

floss = min(Ploss) (3)

The TVD is the sum of the voltage differences between the bus voltages and unity,
as:
N b
TVD = |Vi − 1.0| (4)
i=1

where N b represents the number of buses, and Vi is the bus voltage, the TVD
objective function is as:

f T V D = min(T V D) (5)

For multi-objective (MO) optimization, both functions are optimized together as:
68 A. Eid and H. El-Kishky

f M O = min[ floss , f T V D ] (6)

In this case, the non-dominated points are determined using the Pareto Optimal
Front (POF) method. The fuzzy-based function selects the BCS in the POF collection
of points. For each objective function f i with a solution range of k points, the linear
membership function µik is calculated as:
⎧
⎪
⎨1 f i ≤ f imin
f imax − f i
µik = f imin ≤ f i ≤ f imax (7)
⎪
⎩
f imax − f imin
0 fi ≥ f imax

The normalized function µk for each solution k, is as follows:

 Nob
µik
µ =  Nnd i=1
k
 Nob (8)
k=1 i=1 µik

where f imin is the minimum value while f imax is the maximum value of the i th
objective function; Nob is the objective function numbers and Nnd is the number of
the non-dominated solutions. The minimum value of this set is the BCS of the POF
points.

3 Optimization Algorithms

The newly published AOA [15] is implemented here to optimize the three DG units
in order to minimize power loss and TVD. The performance of the AOA is compared
to the ASO [16] and PSO [13] algorithms. The details of each algorithm are found
in the respective literature.

4 Results and Discussions

In this section, the obtained results of the adopted optimization algorithms are demon-
strated and compared. The study includes the sizing and siting of two different RES
of PVs and WTs in the 33-bus RDS to decrease the total power losses and total
voltage deviation. Three RESs are optimized, working at unity power factor (UPF).
The following case studies are investigated.
Multi-objective Archimedes Optimization Algorithm … 69

23 24 25 26 27 28 29 30 31 32 33

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

19 20 21 22

Fig. 1 The 33-bus RDS layout

4.1 Test System

The paper discusses the DG allocations using different optimization algorithms to

the standard 33-bus radial distribution system. This system has 32 branches and 33
nodes, as shown in Fig. 1, with a total load of 3715 kW and 2300 kVar with a voltage
level of 12.66 kV. The complete system data is presented in [17].

4.2 Simulation Strategies

In this paper, the optimal placement and sizing of three PVs and three WTs are
proposed using the AOA algorithm to reduce the losses and TVD. The study includes
both single-objective optimization (SOO) and multi-objective optimization (MOO)
based on a fuzzy function to determine the BCS among the POF points. In all case
studies, the obtained results of the AOA are compared to ASO and PSO algorithms.

4.3 Simulation Results of Single-Objective Functions

In this section, the single-objective functions of power loss and TVD are optimized
separately using the AOA, ASO, and PSO algorithms. The obtained results are
compared in terms of total power loss, reactive loss, and TVD. The comparison
also includes statistical analysis showing the average, median, standard deviation,
and standard error for each algorithm.

Power Loss Reduction: The three optimization algorithms PSO, ASO, and AOA,
are simulated and compared, as listed in Table 1 with the base case for minimum loss
objective function. Investigating the table shows that the AOA algorithm obtains the
lowest power loss of 72.7837 kW while it is a little bigger of 72.9759 kW by using the
70 A. Eid and H. El-Kishky

Table 1 Optimal sizing of three UPF DGs with power loss reduction as the objective function
Parameters No DGs PSO ASO AOA
Power loss (kW) 210.9822 72.7951 72.8291 72.7837
Reactive loss (kVar) 143.0313 50.6530 50.7958 50.6530
TVD 1.8044 0.6164 0.6185 0.6164
Minimum voltage (pu) 0.9038 0.9686 0.9683 0.9687
DG size & location – 801.80/13 742.03/14 801.79/13
1091.31/24 1081.62/24 1091.31/24
1053.60/30 1114.29/30 1053.59/30

ASO algorithm and 72.7951 kW for the PSO case. The power loss for the base case
is 210.9822 kW. Allocating three DGs working at UPF of optimal sizes, as shown
in Table 1 at optimal buses 13, 24, and 30, reduces the power loss by 65.5% and
64.6% for reactive power loss. Moreover, TVD is reduced by 65.8%, which means
improving the system’s voltage profile. As a result, the minimum voltage of the base
case of 0.9038 pu has been elevated to 0.9687 pu after DGs allocations. The complete
voltage profile is shown in Fig. 2. The node voltages are improved at all buses after
the DG allocations. Although the three algorithms’ power loss values are relatively
equal, the AOA outperforms the other algorithms regarding the convergence speed
of the objective function against the iteration number, as shown in Fig. 3. The PSO
and ASO algorithms come in the second and third stages, respectively, compared to
the AOA algorithm.
Statistical analysis is performed to compare the performance of the three algo-
rithms for solving the DG allocation problem. Every algorithm solves the system 30
consecutive times, allocating the three DGs to minimize the power loss. The average,
median, standard deviation, and standard error are calculated for each algorithm and
listed in Table 2. It is clear from the table that AOA is the best algorithm in all listed
parameters. These results show the superiority of the AOA algorithm compared to
others.

Reduction of TVD: In the same manner as the previous case of power loss reduction,
the three algorithms are simulated to reduce the TVD as a single-objective function.
The TVD values are 0.0715, 0.0693, and 0.0636 V, respectively, for the PSO, ASO,
and AOA, as listed in Table 3. The AOA outperforms the other two algorithms of
ASO and PSO. The allocation of three DGs also contributes to decrease the power
losses and enhance the voltage profile, as shown in Fig. 4. The bus voltages are
all improved (except the slack bus) and approximately equal to unity, where the
minimum bus voltage is 0.9943 pu. The speed convergence of the three algorithms
is shown in Fig. 5. It is clear that the AOA is much faster and more accurate than the
other PSO and ASO algorithms. A statistical analysis based on 30 consecutive runs
of the three algorithms to minimize the TVD as presented in Table 4. The statistical
parameters include average, mean, median, standard deviation, and standard error.
Multi-objective Archimedes Optimization Algorithm … 71

without DGs with DGs

1.02
1
Voltage, pu 0.98
0.96
0.94
0.92
0.9
0.88
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34
Bus number

Fig. 2 Voltage profile before and after DG allocations with power loss reduction

Fig. 3 Convergence speed

of different algorithms for
minimum power loss

Table 2 Statistical analysis

Parameters PSO ASO AOA
for power loss reduction
Average 72.8058 73.3148 72.7846
Median 72.8005 73.1597 72.7837
Standard deviation 20.372E-3 0.48053 1.781E-3
Standard error 3.719E-3 0.08773 0.325E-3

4.4 Simulation Results of Multi-objective (MO) Functions

This section presents the multi-objective (MO) optimization of both the power losses
and the T V D of the distribution system. The PSO, ASO, and AOA algorithms are
applied to the 33-bus system to optimize the two objective functions. The decision-
maker decides the best operation point depending on the system situations and
72 A. Eid and H. El-Kishky

Table 3 Optimal sizing of three UPF DGs with TVD reduction as the objective function
Parameters No DGs PSO ASO AOA
TVD 1.8044 0.0715 0.0693 0.0636
Power loss (kW) 210.98 162.18 114.88 113.18
Reactive loss (kVar) 143.03 114.91 79.89 78.95
Minimum voltage (pu) 0.9038 0.9637 0.9944 0.9943
DG size/location – 1350.6/12 1170.0/13 1168.9/13
1481.7/24 1762.3/24 1558.7/24
1759.6/30 1782.1/30 1847.9/30

without DGs with DGs

1.02
1
0.98
Voltage, pu

0.96
0.94
0.92
0.9
0.88
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34
Bus number

Fig. 4 Voltage profile before and after DG allocations with TVD reduction

Table 4 Statistical analysis

Parameters PSO ASO AOA
for TVD reduction
Average 0.076021 0.078834 0.068913
Median 0.072150 0.078150 0.070401
Standard deviation 9.211E-3 7.906E-3 2.654E-3
Standard error 1.682E-3 1.444E-3 0.484E-3

attributed conditions. The power loss (Ploss) and T V D objectives are optimized by
optimally sizing and sitting different DG units.
The MO-AOA, MO-ASO, and MO-PSO algorithms are used to optimize a single
DG operating at UPF. The objective is to minimize both the power losses and the
T V D using POF techniques accompanied by the fuzzy distribution function. The
POF set of points and the BCS points for all the algorithms are shown in Fig. 6.
Although the BCS points for the three algorithms are close, the MO-AOA has the
minimum values for Ploss and T V D. With two DGs, the optimal solutions with their
BCS points are shown in Fig. 6. For three DGs optimization, the POF points with the
Multi-objective Archimedes Optimization Algorithm … 73

Fig. 5 Convergence speed

of different algorithms for
minimum TVD

Fig. 6 Optimal solutions for optimizing a single-DG (left) and two-DG (right)

BCS points of the different algorithms are shown in Fig. 7. In this case, the MO-ASO
and MO-PSO solutions are close together regarding the BCS points. The MO-AOA
has a lower BCS for both objectives, and hence, it beats the other algorithms. If one
investigates Figs. 6, 7 carefully, it is clear that the expansion ranges of the AOA
when optimizing the TVD and power losses are larger than with the PSO and ASO
algorithms. Subsequently, the AOA finds more non-dominated solutions than the
other algorithms. Moreover, the effectiveness of the AOA appears more clearly with
increasing the number of DG units. The BCS is much lower for AOA than for other
algorithms with three DG units, as shown in Fig. 7. Hence, the AOA is more suitable
for large DNs operations.
74 A. Eid and H. El-Kishky

Fig. 7 Optimal solutions of

different algorithms for
optimizing three DGs

5 Conclusions

The optimal sizing and siting of multiple distributed generations are determined
using the new metaheuristic Archimedes optimization algorithm (AOA) to reduce
total power losses and total voltage deviation. The AOA is implemented in a single-
objective as well as multi-objective optimizations. It is found that the allocation of
the DG units to the distribution networks decreases the power losses and enhances
the voltage profile of the system. The AOA’s obtained results surpass those obtained
by the ASO or PSO algorithms regarding the optimized functions or convergence
speed.

References

1. Home-Ortiz JM, Pourakbari-Kasmaei M, Lehtonen M, Sanches Mantovani JR (2019) Optimal

location-allocation of storage devices and renewable-based DG in distribution systems. Electr
Power Syst Res 172:11–21
2. Nduka OS, Pal BC (2018) Quantitative evaluation of actual loss reduction benefits of a
renewable heavy DG distribution network. IEEE Trans Sustain Energy 9(3):1384–1396
3. Muthukumar K, Jayalalitha S (2016) Optimal placement and sizing of distributed generators
and shunt capacitors for power loss minimization in radial distribution networks using hybrid
heuristic search optimization technique. Int J Electr Power Energy Syst 78:299–319
4. Prakash DB, Lakshminarayana C (2016) Multiple DG placements in distribution system for
power loss reduction using PSO algorithm. Procedia Technol 25:785–792
5. Eid A, Abdel-Akher M (2019) Power loss reduction using adaptive PSO in unbalanced distribu-
tion networks. In: 2019 21st International Middle East Power Systems Conference, MEPCON
2019 - Proceedings, pp 675–680
6. Sanjay R, Jayabarathi T, Raghunathan T, Ramesh V, Mithulananthan N (2017) Optimal alloca-
tion of distributed generation using hybrid grey Wolf optimizer. IEEE Access 5:14807–14818
7. Sultana U, Khairuddin AB, Mokhtar AS, Zareen N, Sultana B (2016) Grey wolf optimizer
based placement and sizing of multiple distributed generation in the distribution system. Energy
111:525–536
Multi-objective Archimedes Optimization Algorithm … 75

8. Devabalaji KR, Ravi K (2016) Optimal size and siting of multiple DG and DSTATCOM in
radial distribution system using bacterial foraging optimization algorithm. Ain Shams Eng J
7(3):959–971
9. Mohammadi M, Rozbahani AM, Montazeri M (2016) Multi criteria simultaneous planning
of passive filters and distributed generation simultaneously in distribution system considering
nonlinear loads with adaptive bacterial foraging optimization approach. Int J Electr Power
Energy Syst 79:253–262
10. Kefayat M, Lashkar Ara A, Nabavi Niaki SA (2015) A hybrid of ant colony optimization and
artificial bee colony algorithm for probabilistic optimal placement and sizing of distributed
energy resources. Energy Convers Manag 92:149–161
11. Eid A, Kamel S, Korashy A, Khurshaid T (2020) an enhanced artificial ecosystem-based
optimization for optimal allocation of multiple distributed generations. IEEE Access 8:178493–
178513
12. Jannesar MR, Sedighi A, Savaghebi M, Anvari-Moghadam A, Guerrero JM (2020) Optimal
multi-objective integration of photovoltaic, wind turbine, and battery energy storage in
distribution networks. J Energy Manag Technol 4(4):76–83
13. Eid A (2020) Allocation of distributed generations in radial distribution systems using adaptive
PSO and modified GSA multi-objective optimizations. Alexandria Eng J 59(6):4771–4786
14. Abdel-Mawgoud H, Kamel S, Yu J, Jurado F (2019) Hybrid Salp Swarm Algorithm for inte-
grating renewable distributed energy resources in distribution systems considering annual load
growth. J King Saud Univ Comput Inf Sci (2019, in press)
15. Hashim FA, Hussain K, Houssein EH, Mabrouk MS, Al-Atabany W (2020) Archimedes opti-
mization algorithm: a new metaheuristic algorithm for solving optimization problems. Appl
Intell 51(3):1531–1551
16. Zhao W, Wang L, Zhang Z (2019) Atom search optimization and its application to solve a
hydrogeologic parameter estimation problem. Knowl-Based Syst 163:283–304
17. Aman MM, Jasmon GB, Bakar AHA, Mokhlis H (2014) A new approach for optimum simul-
taneous multi-DG distributed generation units placement and sizing based on maximization
of system loadability using HPSO (hybrid particle swarm optimization) algorithm. Energy
66:202–215
The Recurrent Neural Network
for Program Synthesis

Achraf Berrajaa and El Hassane Ettifouri

Abstract The program synthesis is a new and ambitious field consists of generating
from short and natural descriptions a source code. The program synthesis is rapidly
becoming a popular research problem, interesting potential has been offered by using
artificial intelligence to support new tools in almost all areas of program analysis
and software engineering. The goal of our long-term research project is to enable
and facilitate any user to create an application from a description written in natural
language by specifying the need for a complete system. This involves carrying out
a study of the user’s needs, the design and implementation of an intelligent system
allowing the automatic realization of an Informatic project (architecture, initialization
scripts, configuration, etc.) expressed in natural language. We propose a recurrent
neural network with LSTM cells and we publish a dataset of more than 145 765
questions and their best answer, especially in Java. The performance of the proposed
model is very interesting such that the success rate is 94%.

Keywords Artificial intelligence · The program synthesis · RNN · LSTM and new
dataset

1 Introduction

The dream of using a natural language to generate source code in a programming

language has existed almost as long as the programming task itself. Natural language
as a support for computer programming would be universally accessible and allow the
automation of an application [1]. However, the ambiguity and diversity of the text, the

A. Berrajaa (B)
INSA Euro-Mediterranean UEMF, Fez, Morocco
e-mail: [email protected]
E. H. Ettifouri
Novelis Lab, Paris, France

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 77

S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_8
78 A. Berrajaa and E. H. Ettifouri

compositional nature of the source code, and the architecture of the layers in the soft-
ware make it too difficult to generate the source code from natural language descrip-
tions. Artificial intelligence offers interesting potential to support new tools in almost
all areas of program analysis and software engineering. In recent years, artificial intel-
ligence has grown exponentially in the field of natural language processing (NLP) and
software engineering. For example, researchers have developed for several languages
several tools for analyzing and extracting useful information in digital documents
in order to generate automatic coding [2]. However, the differences between the
linguistic structures of different languages and the compositional nature of the code
do not always make it possible to extend the use of the programs developed for a
given language.
Program synthesis occupies an important place in the field of artificial intelligence,
and consists of automatically building a program from a specification [2]. It is a field
of multidisciplinary research, which makes collaborate linguists, computer scientists,
logicians, lexicographers or translators. Computing has long regarded programs as
data objects. With the rise of machine learning, there is renewed interest in treating
programs as data to feed into learning algorithms. However, programs have special
features, including several layers of structure, such as the contextless syntax structure
of a program, name and type constraints that are not without context, and program
semantics. As a result, programs do not start in a form that is immediately compatible
with most standard learning techniques [3]. Overall, the program synthesis consists
in designing a solution allowing the automatic generation of the source code of an
application based on descriptions and specifications written in natural languages
(French/English) taking into account the context and the field of activity of this
application, which forces the proposal of solutions of artificial intelligence [4].
This paper presents a recurrent neural network (RNN) with LSTM cells with a
learning process based on a new data set from the program synthesis. The new dataset
(Corpus) is based on over 145 765 Java Q&A from different topics from the “Stack
OverFlow” website.
In the rest of the paper, we start with the Sect. 2, reviewing recent work on program
synthesis, also a description of some algorithms which are considered to be the most
used and presenting models of machine learning and deep learning. In Sect. 3, we
will present the problem statement, then in Sect. 4 we present the construction of the
corpus and we describe our network of recurrent neurons with LSTM cells. In the
Sect. 5, we discuss our results and we will conclude with some thoughts on future
work.

2 Related Work

Linking a natural language description with source code has many topical applica-
tions, such as program synthesis, comment or text prediction, code autocompletion,
prediction of properties of a program, bugs detection and anomalies, the automation
of the migration and the analysis of the quality of the code. The goal of program
The Recurrent Neural Network for Program Synthesis 79

synthesis is to create and generate complete programs based on a [1] specification.

Generally, a specification is a set of requirements to be met by the program. The
models proposed in the program overview assumed that a specification is the desired
user intention to be achieved, which in many cases has also proven to be more compli-
cated than writing the program itself. Synthesis of programs is too difficult a field of
research. Its challenge can be summed up in two main components: the size of the
program’s search space and the diversity of the user’s expression. In other words, the
main challenge is to search a large space for possible programs to find a program that
matches the specification. This has several lines of research in inductive synthesis.
This field has made many progress in different communities, including program-
ming languages and artificial intelligence (machine learning and deep learning). The
first automation approaches were proposed in [5]. The basic idea behind this kind
of models was to use theorem provers to build a proof of the user specification, then
use the proof to provide the corresponding logic program. Currently, researchers
have considered incomplete or partial specifications, such as [6] input/output pairs
or a natural language description. When the specification is a natural language
description, this is the task of semantic analysis.
Semantic analysis methods have been implemented by the NLP community, they
are based on data that attempts to map natural languages to structured logical forms
that can be executed by simple computers. These logical forms are general purpose
meaning representations [7], instructions or orders for personal assistants or robots
[8] or formalisms for querying knowledge bases [9]. These proposals have a great
advantage that they are comparable to data, compared to programming languages
used by developers, the languages specific to these areas targeted by this research
have a relatively simple syntax and scheme. In 2016, Ling et al. [4] proposed a data-
driven code generation approach for high-level programming languages like Java and
Python. However, unlike most semantic analysis jobs, it does not take into account
that the source code must be welldefined programs in the target syntax.
The deep learning community in turn sparked interest in the program synthesis.
Deep neural networks (Deep learning) have two major advantages for modeling and
translating natural language. With the invention of [10] word embedding, informal
instruction sequences have become a possible entry for neural models. The recurrent
neural networks (RNN) have made it possible to make great progress in terms of
modeling the structure of natural language [3]. The “Seq2Seq: sequence-to-sequence
neural network” sequence-to-sequence models [11] with new methods such as the
attention technique [12] allowed the results of translation to go beyond automatic
expression-based in production systems such as Google Translate [13].
Until recently, neural modeling of natural language dominated the idea that neural
networks do not require any information about the syntactic structure of sentences.
The outstanding results of Deep Learning in learning representation and in analyzing
structure suggested using a simple sequence of words as input for a neural model
and allowing it to learn a representation. Of all other important characteristics re-
establishing an error gradient (based on the gradient backpropagation). However,
recent results of using syntactic structures for machine translation [14] and reading
comprehension [15] exceed the results of Seq2Seq models. To analyze the syntactic
80 A. Berrajaa and E. H. Ettifouri

structures, recursive neural networks [16, 17] and a grammar of recurrent neural
networks [18] were used. Recursive neural networks have also shown good results
for the analysis of [19] and dependency [20] semantics. Reed and Freitas [21] and
Neelakantan et al. [22] introduce differentiable neuronal architectures for program
induction. This is an interesting emerging area of research [23], but one that does
not yet adapt to the kinds of problems that the programming language and software
engineering community is considering [24].
The generation of source code from descriptions and sentences written in natural
language (Text-to-Code) aims to help developers, end-users, to easily generate
programs. There is growing interest in this area, with applications such as converting
natural language to: Excel software macros [25], Java expressions [26], Shell
commands [27], if-then simple programs [8], regular expressions [28], SQL queries
[29], variable names, methods and their types [30, 31] and entity names [32].
This work presents a network of recurrent neurons (RNN) with LSTM cells with
a learning process based on a new dataset (Corpus) that contains more than 145
765 questions and their best response specifically in Java in order to automatically
generate source code for java applications with an n-tier (multilayer) architecture.

3 Problem Statement

The creation of an application or software, meeting a user need, is a project that

requires several steps as: framing meetings of the need, functional and technical
design, implementation of the technical architecture, development of the applica-
tion, code reviews, tests and validations, correction of quality anomalies and various
technical and functional bugs.
The project requires the intervention of several profiles: project manager, technical
experts, architect, developers and test and validation teams. Our objective is to design
a solution allowing the automatic generation of the source code of an application from
sentences (specifications and descriptions) written in natural languages (program
synthesis) in order to minimize this cost. This objective involves the resolution of
several major technological and scientific challenges:
• Application and implementation of approaches allowing the machine to under-
stand the user’s need, expressed in natural language (text) in a specific field, by
exploiting the results of the state of the art and by participating through scientific
contributions to the resolution of the limitations encountered.
• Automatic generation of a computer project meeting user needs (Architecture,
configuration, initialization scripts, …).
• Generation of algorithms that take into account the constraints and requirements
expressed in the user’s need in order to integrate them into the generated project.
We will focus our work on the generation of applications and programs with
language java with an n-tier (multilayer) architecture, for that we propose a network
of recurrent neurons (RNN) with LSTM cells with a learning process based on a
The Recurrent Neural Network for Program Synthesis 81

new set of data more than 145 765 questions and their best answer, especially in
Java. This dataSet represents the learning data of our recurrent neural network with
LSTM.

4 The Proposed Contribution

4.1 Stack OverFlow

The dataset we propose in this article is based on the dataset provided by the “Stack
Overflow” site. We rely on it because it is both the largest and the newest platform for
exchange between computer programming experts. Stack Overflow offers questions
and answers on a wide range of topics related to computer programming (java, c, c++,
hp, javascript, python, …). It is part of the Stack Exchange site network. The proposed
dataset is based on the questions and their best answers available in stackoverflow.

4.2 Data Preparation

The main goal behind creating a new data set is to enable the scientific community to
evaluate new techniques using the same test conditions. Also, this dataset represents
an interesting and recent sample of questions and answers from computer experts
which will allow it to be a good learning sample, and it will be a strong point for a
learning model because it will be learned from experts in the field. These benefits
are usually not yet offered by the program synthesis researchers.
The dataset we use in this article is based on all the data provided by the Stack
OverFlow Questions and Answers website via the Kaggle web platform in a pre-
release version. We used this dataset because it is both an interesting and recent
sample of the questions and answers from the Stack OverFlow site. Initially, we did
the extraction of dataset containing a 1.2 m sample of questions and answers since
2008 until October 2016 [33] and to have a larger and updated sample with the new
scripts, we have added a second dataset with 143 690 questions and answers since
October 2016 until September 2020. Then we processed the data of the two datasets
checking the validity of the fields (empty fields and quotation marks) and we filtered
the lines taking only the questions and their best response from the java language,
furthermore we combined the two samples to have a single structure. The result is a
dataset of 145 765 questions and their best answer specifically in Java. For the latter
to be usable for machine learning.
82 A. Berrajaa and E. H. Ettifouri

4.3 LSTM Neural Networks

Long Short-Term Memory (LSTM) is an alternative solution was proposed in [34]:

The network architecture is modified such that the vanishing gradient problem is
explicitly avoided, whereas the training algorithm is left unchanged. But what is the
strong point of LSTM? why are we using it in our implementation?
A common LSTM unit is composed of a cell, a forget gate, an input gate and an
output gate. The cell remembers values over arbitrary time intervals and the three
gates regulate the flow of information into and out of the cell. LSTM networks are
well-suited to classifying, processing and making predictions based on time series
data, since there can be lags of unknown duration between important events in a time
series. And also as we explained LSTMs were developed to deal with the exploding
and vanishing gradient problems that can be encountered when training traditional
RNNs. Figure 1 shows the internal architecture of an LSTM.
– Forget gate: it’s the power to forget information. In a long sequence concretely to
the classic neural network where you have to memorize all the information on the
long term. In the LSTM the power to forget non-useful information is represented
by the function ft = α (Wf ∗ [ht − 1, xt] + bf), where α is a sigmoid function,
Wf is the weights and [ht − 1, xt] is the concatenation of the two vectors ht − 1
and xt.
– Input gate: is the responsible to add relevant information, the function that will
offer new information is Ct = Ct − 1 ∗ ft + it ∗ At where it = α (Wi ∗ [ht − 1,
xt] + bi) and At = tanh (Wc ∗ [ht − 1, xt] + bc).
– Output gate: Finally, this operation allows to define the current state of the cell.
So far we have used the status of the president cell, we have forgotten information
and we have added new information to the memory. We still have to define the
state of the current cell. This is summarized by the following function:

ht = α (W o ∗ [ht − 1, xt] + bo) ∗ tanh(Ct).

Fig. 1 LSTM recurrent neural networks

The Recurrent Neural Network for Program Synthesis 83

4.4 Network Architecture and Choice of Parameters

Although there are several differences and several architectures in neural network
language models that have been successfully applied so far, all present and share
some basic principles like:
The input sequence (the input words), in our implementation it is encoded by an
encoding from 1 to K, where K is the unique number of letters in our dataSet. This
involves listing single letters and the code using the One Hot Encoding [35]. Note
that the sequenes are generated in 64 batch with 10% diversity (in the objective to
avoid overfitting).
The model makes a prediction of a sequence where the first element of each
sequence represents the next letter of the input. Concerning the network topology, an
architecture with five hidden layers has been applied each with 128 LSTM cells, or
each uses the “relu” activation function. While the output layer, a “softmax” activation
function is used to produce properly normalized probability values in order to choose
a single output letter.
As a learning criterion, Adam optimizers is used with a precision of 0,001. It
is generally advisable to normalize the input data of a neural network [36], which
means that a linear transformation is applied so that the data has zero mean and unit
variance. Only one normalization was to use is to keep all the letters in lower case.
Otherwise we kept the symbols because they represent part of the source code. The
following algorithm summarizes the main steps.

Algorithm 1 : The proposed RNN architecture with LSTM

- Initialization of the weights by random drawing according to a uniform law on [0;1].
- Normalization: keep the text in lowercase.
- Codage: list all elements in text, represent each one by a number and coded it using
the One Hot Encoding.
- Creating an RNN network: Creating seven layers, the six hidden layers has been ap-
plied each with 128 LSTM cells. While the output layer, a "softmax" activation func-
tion is used to predict the continuation of the text.
while counter ≤ iterMax do
1. Batch generation with 10% of noise.
2. Have the model trained.
3. Adam optimizers is used with a precision of 0.001 in order to assign to each
entry a responsibility in the overall error.
4. Update each weight
end while
84 A. Berrajaa and E. H. Ettifouri

5 Experimental Results

In this section, we present the types and percentage of data in our dataSet, also the
learning rate achieved by our neural network.

The Content of the dataSet

In order to identify the categories of questions asked by StakOverFlow users for the
Java language, we ran through all the questions in the dataset and symbolized each
question, then counted the occurrences of each to rank the words by categories. Each
class contains words with convergent occurrences. Surprisingly, we find subjects of
different categories. For example, there are general topics with words like “error”
or this class deals with the problems of errors. Also there are topics specific to the
exception problem or references to specific techniques (for example “java mail”) or
topics that can rarely occur (like “keyevent” or “jaxrs”). We recognize categories
of questions that cannot be explained with code because they do not always coin-
cide with identifiers. These are emerging code properties, such as library or version
problems.
By performing the tokenization of the title and whenever possible, we analyzed
the java code snippets, we identify the fact that 83% of the text, 16.95% contains
both text and java code and that the questions 0.05% only contain java code. This
reaffirms that our dataSet contains a diversity to present a specification.

The Learning and Test Rate

The learning rate makes it possible to get an idea of the quality of the model. In the
following figure we present a graph that represents the learning rate (94%) and the
test rate (87%).

Fig. 2 The learning rate and the loss rate

The Recurrent Neural Network for Program Synthesis 85

6 Conclusion

In order to design and propose a model that meets the need for program synthesis.
Specifically, the generation of frontend/backend web applications based on Java tech-
nology and which respect an architecture with n levels (multilayer). We proposed
a project architecture in order to meet the different user needs expressed in natural
language on different software engineering domains. We started by introducing a
recurrent neural network with LSTM cells and we publish a set of learning data
of more than 145 765 questions and their best answer, especially in Java. The
performance of the proposed model is very interesting such that the success rate
is 94%.
The perspectives are set by proposing new methods, new neural network
architectures and implementing them in GPUs to improve the learning face.

References

1. Gulwani S, Polozov O, Singh R (2017) Program synthesis. Found Trends Program Lang 4(1–
2):1–119
2. Berrajaa A, Ettifouri EH, Dahhane W, Bouchentouf T, Rahmoun M (2019) Nl2code: a corpus
and semantic parser for natural language to code. In: International conference on smart
information and communication technologies. Springer, Cham, pp 592–599
3. Jozefowicz R, Vinyals O, Schuster M, Shazeer N, Wu Y (2016) Exploring the limits of language
modeling. arXiv preprint arXiv:1602.02410
4. Ling W, Grefenstette E, Hermann KM, Kočiský T, Senior A, Wang F, Blunsom P (2016) Latent
predictor networks for code generation. arXiv preprint arXiv:1603.06744
5. Manna Z, Waldinger RJ (1971) Toward automatic program synthesis. Commun ACM
14(3):151–165
6. Lau T (2001) Programming by demonstration: a machine learning approach. Doctoral
dissertation
7. Banarescu L, Bonial C, Cai S, Georgescu M, Griffitt K, Hermjakob U, Knight K, Koehn
P, Palmer M, Schneider N (2013) Abstract meaning representation for sembanking. In:
Proceedings of the 7th linguistic annotation workshop and interoperability with discourse,
pp 178–186
8. Quirk C, Mooney R, Galley M (2015) Language to code: learning semantic parsers for if-this-
then-that recipes. In: Proceedings of the 53rd annual meeting of the association for computa-
tional linguistics and the 7th international joint conference on natural language processing, vol
1, pp 878–888
9. Berant J, Chou A, Frostig R, Liang P (2013) Semantic parsing on freebase from question-
answer pairs. In: Proceedings of the 2013 conference on empirical methods in natural language
processing, pp 1533–1544
10. Bengio Y, Ducharme R, Vincent P, Jauvin C (2003) A neural probabilistic language model. J
Mach Learn Res 3:1137–1155
11. Sutskever I, Vinyals O, Le QV (2014) Sequence to sequence learning with neural networks.
In: Advances in neural information processing systems, pp 3104–3112
12. Bahdanau D, Cho K, Bengio Y (2014) Neural machine translation by jointly learning to align
and translate. arXiv preprint arXiv:1409.0473
13. Wu Y, Schuster M, Chen Z, Le QV, Norouzi M, Macherey W, Krikun M, Cao Y, Gao Q,
Macherey K, Klingner J (2016) Google’s neural machine translation system: bridging the gap
between human and machine translation. arXiv preprint arXiv:1609.08144
86 A. Berrajaa and E. H. Ettifouri

14. Chen H, Huang S, Chiang D, Chen J (2017) Improved neural machine translation with a
syntax-aware encoder and decoder. arXiv preprint arXiv:1707.05436
15. Xie P, Xing E (2017) A constituent-centric neural architecture for reading comprehension.
In: Proceedings of the 55th annual meeting of the association for computational linguistics
(Volume 1: Long Papers), vol 1, pp 1405–1414
16. Goller C, Kuchler A (1996) Learning task-dependent distributed representations by backpropa-
gation through structure. In: Proceedings of international conference on neural networks (ICNN
1996), vol 1, pp 347–352
17. Socher R, Lin CC, Manning C, Ng AY (2011) Parsing natural scenes and natural language with
recursive neural networks. In: Proceedings of the 28th international conference on machine
learning (ICML 2011), pp 129–136
18. Dyer C, Kuncoro A, Ballesteros M, Smith NA (2016) Recurrent neural network grammars.
arXiv preprint arXiv:1602.07776
19. Tai KS, Socher R, Manning CD (2015) Improved semantic representations from treestructured
long short-term memory networks. arXiv preprint arXiv:1503.00075
20. Zhu C, Qiu X, Chen X, Huang X (2015) A re-ranking model for dependency parser with
recursive convolutional neural network. arXiv preprint arXiv:1505.05667
21. Scott R, de Freitas N (2016) Neural programmer-interpreters. In: Proceedings of the
international conference on learning representations (ICLR)
22. Neelakantan A, Le QV, Sutskever I (2015) Neural programmer: inducing latent programs with
gradient descent. arXiv preprint arXiv:1511.04834
23. Kant N (2018) Recent advances in neural program synthesis. arXiv preprint arXiv:1802.02353
24. Gaunt AL, Brockschmidt M, Singh R, Kushman N, Kohli P, Taylor J, Tarlow D (2016) Terpret:
a probabilistic programming language for program induction. arXiv preprint arXiv:1608.04428
25. Gulwani S, Marron M (2014) Nlyze: interactive programming by natural language for spread-
sheet data analysis and manipulation. In: Proceedings of the 2014 ACM SIGMOD international
conference on management of data. ACM, pp 803–814
26. Gvero T, Kuncak V (2015) Synthesizing java expressions from free-form queries. Acm Sigplan
Notices 50(10):416–432
27. Lin XV, Wang C, Zettlemoyer L, Ernst MD (2018) Nl2bash: a corpus and semantic parser for
natural language interface to the linux operating system. arXiv preprint arXiv:1802.08979
28. Kushman N, Barzilay R (2013) Using semantic unification to generate regular expressions
from natural language. In: Proceedings of the 2013 conference of the North American chapter
of the association for computational linguistics: human language technologies, pp 826–836
29. Zhong V, Xiong C, Socher R (2017) Seq2sql: generating structured queries from natural
language using reinforcement learning. arXiv preprint arXiv:1709.00103
30. Alon U, Zilberstein M, Levy O, Yahav E (2019) code2vec: learning distributed representations
of code. Proc ACM Program Lang 3(POPL):40
31. Gao S, Chen C, Xing Z, Ma Y, Song W, Lin SW (2019) A neural model for method name gener-
ation from functional description. In: 2019 IEEE 26th international conference on software
analysis, evolution and reengineering (SANER). IEEE, pp 414–421
32. Parvez MR, Chakraborty S, Ray B, Chang KW (2018) Building language models for text with
named entities. arXiv preprint arXiv:1805.04836
33. StackOverflow (2008–2016) Stacksample: 10 percent of stack overflow questions and answers.
https://www.kaggle.com/stackoverflow/stacksample
34. Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9(8):1735–1780
35. Buckman J, Roy A, Raffel C, Goodfellow I (2018) Thermometer encoding: one hot way to
resist adversarial examples
36. Bishop CM (1995) Neural networks for pattern recognition. Oxford University Press, Oxford
MaskNet: CNN for Real-Time Face Mask
Detection Based on Deep Learning
Techniques

Ismail Nasri, Mohammed Karrouchi, Hajar Snoussi, Abdelhafid Messaoudi,

and Kamal Kassmi

Abstract Coronavirus disease 2019 (COVID-19) is currently spreading in several

countries around the world. The wearing of face masks during the COVID-19
pandemic is one of major protection method that has received varying recommen-
dations from different public health agencies and governments. In this paper we
have proposed a face mask detector based on deep learning techniques to classify
each face as with mask or without mask in real-time. This study will focus on the
Prajna Bhandary dataset. In this work a comparison between transfer learning and
training from scratch has been provided, this comparison based on accuracy, model
size and computer time. We have also analyzed the role of number of images dataset
and learning rate in the classification accuracy and training time. In additions, the
extracted model can achieve an accuracy of more than 98% and can be implemented
in embedded system and used for many applications.

Keywords COVID-19 · Face mask detection · Convolutional neural networks ·

Transfer learning · Training from scratch

1 Introduction

The World Health Organization and several countries have recommended that masks
can limit the spread of certain respiratory viral diseases, including COVID-19 [1].
This recommendation is meant to reduce the spread of the disease by asymptomatic
and pre-symptomatic individuals and is a complementary measure to established
preventive measures such as social distancing. In this paper we have proposed a
solution for mask detection that can be very effectively used at airports to detect
travelers without masks and at other public places like Malls, schools, etc. Figure 1

I. Nasri (B) · M. Karrouchi · H. Snoussi · A. Messaoudi · K. Kassmi

Electrical Engineering and Maintenance Laboratory, High School of Technology,
Mohammed First University, BP. 473, Oujda, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 87

S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_9
88 I. Nasri et al.

Fig. 1 General architecture of face mask detection process

presents the general architecture of face mask detection, in three phases: face detec-
tion, feature extraction, and classification. From the existing method of face recog-
nition, we have used the Viola-Jones algorithm [2, 3] to detect and crop faces from
images. We get the image with the face and run it through the Viola-Jones algorithm.
If the face is found, the classifier will give the region of interest of the face, and pass
it to the proposed CNN. CNN with Feature Detection Layers are used to extract the
deep features and those features are passed to Classification Layers. Softmax layer
in CNN classifies the images as With Mask or Without Mask in real time with the
probabilities of each classes.
The remainder of this paper is organized as follows. Section 2 introduces a brief
description of the related work. The Sect. 3 describes the proposed solution and
approach adopted to prepare the developed CNN models. The obtained results are
discussed in Sect. 4. Finally, in Sect. 5, conclusions, as well as suggestions for future
research, are presented.

2 Related Work

In this section, a review of the previous methods and approaches for classification
images based on deep learning and computer vision techniques will be provided.
Ququ Chen and Lei Sang [4] proposed Face mask recognition to prevent fraud and
to limit the problem of identifying the face and mask in the field of financial security
precaution. For that, OpenCV [5] and Dlib [6] are used to recognize the face and
extract it, and they can predict whether the image of which they test is a human face
or a mask. Loey et al. [7] propose a hybrid model using deep and classical machine
learning for face mask detection using Resnet50. In this study, we detect the region
of interest for helping proposed algorithms to learn more precise features for the
classification. For that, we use the cascade object detector uses the Viola-Jones algo-
rithm for detecting and extracting face from images, the images extracted will act
as the datasets for training and testing the Convolutional Neural Networks CNNs
proposed. The Viola-Jones object detection algorithm [2, 3] is the first object detec-
tion framework used in real-time, which was proposed by Paul Viola and Michael
MaskNet: CNN for Real-Time Face Mask Detection … 89

Fig. 2 With mask and without mask image samples from Prajna Bhandary dataset

Jones in 2001. It can be used to detect various object classes, but it was mainly used
for the problem of face detection.

3 Proposed Method

This section provides an overview of the proposed method: Dataset description and
approach adopted to prepare the developed CNN models that will be able to classify
whether the person is wearing a mask or not.

3.1 Dataset Description

This study will focus on the Prajna Bhandary dataset [8]. It contains the full compo-
nent of the dataset for training and validation the proposed CNN model. For dataset
creation, Prajna takes normal images of faces and adds face masks to them using
custom computer vision Python script (Fig. 2).

3.2 Proposed Approach

In the rest of this section, an overview of the proposed approach (see Fig. 3) to prepare
a convolutional neural network model will be provided.
The proposed approach consists of five main steps:
90 I. Nasri et al.

Fig. 3 Approach proposed to prepare a deep learning model

• Step 1: Load and Explore images: Loading images from Prajna Bhandary dataset
for processing.
• Step 2: Detecting and Cropping faces from images: In the second step, we use the
cascade object detector that uses the Viola-Jones algorithm to detect and crop faces
from images (see Fig. 8). These images will be used to train and test the proposed
models. In the present study, datasets utilized are explained in two categories as
training and test sets, 70% for training and 30% for validation and testing.
• Step 3: Creating and Configuring Network Layers: In this step, we define the
convolutional neural networks CNN [9] architecture MaskNet.
• Step 4: Model training and testing: The cropped face from dataset will provided
as the input to the training algorithm, detailed in (algorithm 1).
• Step 5: Extract the model: Finally, the CNN model can be saved as a file and used
to classify images with the predicted label and probability. Figure 4 illustrates the
flowchart of the proposed algorithm for mask detection.

3.3 Transfer Learning

For transfer learning, four pre-trained network such as convolutional neural networks
CNN namely AlexNet [10], VGG16 [11], GoogLeNet [12] and ResNet18 [13] were
used to classify images as with mask or without mask. Table 1 shows an overview
of the pretrained CNN used in this work. AlexNet is a CNN that contains eight
layers; 3 fully connected layers and 5 convolutional layers. The AlexNet can classify
images into 1000 object classes, such as a laptop, pen and a variety of objects. This
network was trained on ImageNet database, with inputs at size 227 × 227 × 3 images.
VGG16 consists of 16 layers; 13 convolutional layers and 3 fully connected layers;
the convolutional layers are followed by the rectified linear unit (ReLU) layer for
faster and more effective training, this layer performs a threshold operation to each
element of the input by mapping any negative values to zero and maintaining positive
values. The image input size of VGG16 is 224 × 224 × 3. GoogLeNet is a CNN
that that contains 22 layers, each convolutional layer followed by pooling layer for
simplifies the output by reducing the number of parameters that the network needs
to learn about.
MaskNet: CNN for Real-Time Face Mask Detection … 91

Fig. 4 Flowchart of the

proposed algorithm

Table 1 Overview of the pretrained CNN used in this work

Network Year Parameters Nbr layers Size input
AlexNet 2012 60M 8 227 × 227 × 3
VGG16 2014 138M 16 224 × 224 × 3
GoogLeNet 2014 7M 22 224 × 224 × 3
ResNet 18 2015 11M 18 224 × 224 × 3

Figure 5 shows the training progress plot for transfer learning using AlexNet,
and Fig. 6 shows the classification result with the predicted label and the predicted
probability.
92 I. Nasri et al.

Fig. 5 The training progress plot for transfer learning

Fig. 6 Classification images with the predicted label and the predicted probability %

3.4 Training from Scratch

For training from scratch, we creating and configuring network layers by defining the
convolutional neural network architecture. In the proposed CNN model MaskNet,
we use 3 convolutional layers and one fully connected layer. Softmax classifier is
used to classify images as With Mask or Without Mask. The network was trained by
the following algorithm. Figure 7 shows the architecture of the proposed deep CNN
MaskNet.
MaskNet: CNN for Real-Time Face Mask Detection … 93

Fig. 7 Proposed deep CNN MaskNet

Algorithm 1 : Training from Scratch

Input: Processed images dataset.
Output: Learned CNN model MaskNet.
1. Load and Explore Image data from My PC.
2. Split data into training and test sets.
3. Create and Configure Network Layers by defining the convolutional neural network
architecture.
4. Specify Training Options.
5. Train the Network Using Training Data (imdsTrain).
6. Review Network Architecture (see Fig. 7).
7. Compute accuracy.

4 Experimental Results

4.1 Dataset Preparation

The processor for training and test processing platform was a 3.6 GHz Intel (R)
Core (TM) Core i5-8350U with 8 GB memory and 256 GB SSD hard disk, and
the development platform for the algorithm was Matlab R2018b. About the dataset
processing, the faces was detected and cropped from images (see Fig. 8) using the
Viola-Jones algorithm as it appears in the Table 2 below.

4.2 Result and Analysis

Once the face mask detector is trained from scratch based on our proposed CNN
model, we can then proceed on to loading our CNN mask detector, performing face
detection first, and then classify each face as with or without mask. For testing the
accuracy of the proposed model we use a dataset including images and videos of
subjects of various ethnicities, scenarios, and conditions.
94 I. Nasri et al.

Fig. 8 Extracted image samples by the Viola-Jones algorithm from dataset

Table 2 Overview of images

Dataset Nbr images Nbr images extracted by
processed for each class
Viola-Jones
With mask 690 648
Without mask 686 677
Total 1376 1325

In this part, a comparison of performances of the proposed CNN model MaskNet

and pretrained AlexNet is presented. The two models were trained and evaluated by
the same number of images dataset (1325 images). The following Table 3 shows the
network performance of these models.
Training the model from scratch and achieving reasonable results requires a lot
of effort and computer time. The reason for that is we should test the performance
of the network. If it is not adequate, we should try modifying the CNN architecture
and adjusting some of the training options and retraining. The training time of the
CNN Scratch model is 15 min and 14 s. Experimental results show that size of the
proposed CNN model is small (2.12 Mbit) comparing to AlexNet (621Mbit) size, but
results of proposed CNN model are more accurate than pretrained AlexNet results.
The accuracy rate of the developed model is almost 99% for proposed CNN model
and 96% for AlexNet.
In this part, a comparison between the pretrained networks is provided. The model
performance was evaluated by modifying the number of images datasets for training
the CNN model (Fig. 9).

Table 3 Comparison between the proposed models (CNN Transfer & CNN Scratch)
Model Accuracy Computer time Size Nbr images
AlexNet 96% 12 min 4 s 621Mbit 1325
MaskNet 99% 15 min 14 s 2.12Mbit 1325
MaskNet: CNN for Real-Time Face Mask Detection … 95

100

80

Accuracy % 60

40

20

0
VGG16 ResNet18 GoogLeNet AlexNet MaskNet
100 91 97 80 75 95
500 93 98 91 89 98
1325 96 98 97 96 99

Fig. 9 Accuracy of classification with different image number of dataset

The accuracy of classification obtained using VGG16 net, ResNet18, GoogLeNet,

and AlexNet were 96%, 98%, 97%, 96% respectively for 1325 images. Achieving
a higher accuracy requires a large dataset. The Accuracy rate of all CNN models is
very important for a number of images of 1325.
Table 4 shows the model size of the pretrained network and the proposed CNN
MaskNet. The model size of VGG16 was 528 Mbit with 138M of parameters, and
GoogLeNet was 94.2 Mbit with 7M parameters. The obtained results show that the
size of the model is directly proportional to the number of parameters.
Figure 10 provides the computer time of the pretrained networks and the proposed
CNN with different number of dataset. The results shows that the computer time

Table 4 Comparison between the CNN models

CNN model VGG16 net ResNet-18 GoogLeNet AlexNet MaskNet
Model size (Mbit) 528 121 94.2 621 2.12

Fig. 10 Training Time of 45

CNNs with different number 100
40
of dataset
500
35
Computer Time (min)

30 1325

25
20
15
10
5
0
AlexNet MaskNet VGG16 GoogLeNet ResNet18
96 I. Nasri et al.

120% MaskNet 120% AlexNet

99% 98% 96%
100% 94% 100% 95% 93%
85%

Accuracy %
Accuracy %
80% 80%

60% 60% 60%

40% 40%

20% 20%

0% 0%
0.0001 0.001 0.01 0.1 0.0001 0.001 0.01 0.1

Fig. 11 Analysis of classification accuracy for varying learning rate value

depends on a lot of factors including the number of image dataset, networks

architecture, and the number of parameters, etc.
The CNN models used in this work achieve a higher accuracy when the learning
value is 0.0001. This parameter controls how quickly the model adapted to the
problem. A small learning rate value can lead to good results in terms of accu-
racy but with long computer time, whereas when the learning rate increased to 0.1,
we have too fast learning with an unstable training process and less accuracy rate.
Figure 11 below shows the classification accuracy of MaskNet and AlexNet for
varying Learning Rate value.

5 Conclusion

In this work, we have proposed a method for face mask detection using deep learning
techniques. The face is detected using the Viola-Jones algorithm. The proposed
convolutional neural network with Feature Detection Layers is used to extract the
deep features, and those features are passed to Classification Layers. A Softmax
layer in the CNN provides the classification output as With Mask or Without Mask
with the probabilities for each class. The proposed model has been trained and eval-
uated using the Prajna Bhandary dataset by two commonly used approaches for deep
learning: transfer learning and training from scratch. For that, we have used four
pre-trained networks, and one proposed CNN, namely MaskNet. The performances
of the models such as accuracy and computer time have been evaluated by modifying
the number of images dataset and the learning Rate value. Further work will focus
on the implementation of the proposed CNN MaskNet in an embedded system for
mask detection in the era of the COVID-19 pandemic.
MaskNet: CNN for Real-Time Face Mask Detection … 97

References

1. World Health Organization. Coronavirus disease (COVID-19) advice for the public. https://
www.who.int/emergencies/diseases/novel-coronavirus-2019/advice-for-public. Accessed 20
Oct 2020
2. Viola P, Jones M (2001) Rapid object detection using a boosted cascade of simple features.
In: The 2001 IEEE computer society conference on computer vision and pattern recognition,
CVPR 2001, vol 1. IEEE
3. Viola P, Jones M (2001) Robust real-time object detection. Int J Comput Vis 4(34–47):4
4. Chen Q, Sang L (2018) Face-mask recognition for fraud prevention using Gaussian mixture
model. J Vis Commun Image Represent 55:795–801
5. Bradski G, Kaehler A (2008) Learning OpenCV: computer vision with the OpenCV library.
O’Reilly Media, Inc.
6. Dlib C++ toolkit, 08 January 2018. https://dlib.net/
7. Loey M, Manogaran G, Taha MHN, Khalifa NEM (2020) A hybrid deep transfer learning
model with machine learning methods for face mask detection in the era of the COVID-19
pandemic. Measurement 167:108288
8. Prajna Bhandary dataset. Mask classifier. https://github.com/prajnasb/observations/tree/mas
ter/experiements/data
9. O’Shea K, Nash R (2015) An introduction to convolutional neural networks. arXiv pre-print
arXiv:1511.08458
10. Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification with deep convolutional
neural networks. In: Advances in neural information processing systems
11. Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556
12. Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, Erhan D, Vanhoucke V, Rabinovich
A (2015) Going deeper with convolutions. In: Proceedings of the IEEE conference on computer
vision and pattern recognition, pp 1–9
13. He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition. In:
Proceedings of the IEEE conference on computer vision and pattern recognition
Machine Learning System for Fraud
Detection. A Methodological Approach
for a Development Platform

Salma El Hajjami, Jamal Malki, Mohammed Berrada, Harti Mostafa,

and Alain Bouju

Abstract The democratization and massification use of credit cards lead inexorably
to a high number of fraudulent transactions. Generally, the fraud detection is part
of the anomaly detection problem. In this field, current approaches and techniques
are constantly looking for optimized solutions to detect anomalies. Faced with a
massive and growing data volume, these methods are put to the test, and thus lead
to a large number of undetected anomalies. Real time fraud detection requires the
design and implementation of scalable techniques capable of ingesting and analyzing
massive amounts of data continuously. Recent advances in storage, data analytics
processing, and open-source solutions open up new perspectives in the anomaly
detection field and in particular fraud. In this article, we are interested in the design of
a fraud detection system (FDS) based on open-sources Big Data technologies. Thus, a
general methodology is proposed based on the formalization, the implementation and
the technical design of a platform for fraud detection. The formalization part consists
of four layers: distributed storage, data processing, model building, and finally the
model evaluation. The implementation part uses Spark distributed data processing
system. In particular, we are based on its framework dedicated to machine learning,
called MLlib. The technical design part of the platform is based on the latest Big
Data technologies such as Hadoop, Yarn, Livy etc.

Keywords Machine learning · Anomaly detection · Fraud detection · Big data ·

Spark and Hadoop platforms

Projet UE-FEDER PLAIBDE https://plaibde.ayaline.com

S. El Hajjami (B) · M. Berrada

IASSE Laboratoire, ENSA, USMBA, Fès, Marocco
e-mail: [email protected]
J. Malki · A. Bouju
L3i Laboratoire, La Rochelle Université, La Rochelle, France
H. Mostafa
LIMS Laboratoire, FSDM, USMBA, Fès, Marocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 99

S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_10
100 S. El Hajjami et al.

1 Introduction

In recent years, the volume of credit card transactions has increased considerably
due to the popularization of their uses, the rapid development of associated services,
such as e-commerce, e-finance and all mobile payments. The large-scale adoption of
credit card coupled with the development of different use cases where transactions
take place without rigid verification and supervision inevitably result in large losses
[1]. In this area, it becomes necessary, even crucial, to immediately process the
collected data to detect any potential fraud. However, traditional detection tools
are unable to handle the captured data volume [2]. Therefore, they cannot detect
anomalies and threats. It has therefore become essential to overcome the bottlenecks
of existing techniques by using a new generation of artificial intelligence approaches,
such as those based on machine learning. These approaches must work closely with
techniques for data processing, known as Big Data. They must also provide near
real-time responses.
Ultimately, we believe that Machine Learning (ML) models can provide a reliable
solution to the anomaly detection problem and therefore its application to the fraud
detection [3]. Depending on the needs and scope of ML, there are several scenarios,
how these models can be constructed and applied. Recent work in the field of ML
highlights two kinds of important platforms: development platform and deployment
platform.
Development platforms dramatically reduce the time and cost of building ML
models. They guarantee the organization of all levels of maturity of these models.
They can be out of the box or tailor-made, based on open-sources or commercial
software. Some development platforms can act as model deployment platforms as
well, and in a few cases, model life cycle management platforms, but their core is
running model building pipelines. The deployment platform brings the models to
the production environment. Suppose an ML model has been created and evaluated
on a given development platform, the next step is to use that model in a production
application. The latter can be a simple web application, embedded in an IoT device,
or a service or micro-service forming part of a complex architecture. This is the role
of the deployment platform.
The two platforms are linked by what is known as the machine learning pipeline.
Indeed, the life cycle of a model should not stop at the development and evaluation
stage. It is put into production and then returned to the development stage as much
as possible to be refined and improved.
In this work, we present a methodology for building ML models in a development
platform approach based on open-source software. This platform dedicated to a fraud
detection system (FDS) brings two important contributions:
– A fraud detection methodology which contains most of the design ideas common
in the latest FDS, which greatly facilitates the integration of detection algorithms
into the workflow.
– A four-layer framework that includes a distributed storage layer, data processing
layer, model building layer, and finally model evaluation layer.
Machine Learning System for Fraud Detection … 101

– A development platform whose central core is based on Hadoop and Spark. With
these technologies, we are able to build a scalable and reliable system.
The remainder of the paper is organized as follows. In Sect. 2, data and system
challenges for fraud detection are described. Section 3 describes our methodology
for building ML models for fraud detection. The results of the experimental evalua-
tion are illustrated in Sect. 4. Section 5 gives all the details about the development
platform. Finally, conclusion and future work terminate our article.

2 Fraud Detection: Data and System Challenges

The design of an effective, real-time, and scalable based FDS is subjected to several
data and system challenges enumerated as follows [4–7]:
– Data imbalance: generally, fraudulent transactions represent less than 0.05\% of
total transactions. This ratio results in the very imbalanced dataset. A good FDS
should be able to handle asymmetric distributions of data.
– Data overlap: some fraudulent patterns behave like normal ones. A good FDS
should be able to detect fraudulent transactions that mimic legitimate transactions.
– Concept Drift: FDS targeting fraudulent behavior suffer from the fact that in the
real world, the profile of normal and fraudulent behavior changes over time. So, a
model that performs well over a period of time can quickly become obsolete and
produces inaccurate predictions. Therefore, a good FDS should be dynamic and
able to adapt to changes in fraudulent patterns.
– Evaluation metric: evaluation metric used for fraud detection techniques should
be chosen with care. Indeed, some measures such as accuracy are not suitable for
asymmetric distribution.
– Misclassification cost problem: cost of misclassifying each transaction varies.
Fraud detection is a very cost sensitive area. Generally, undetected fraudulent
transactions are much more serious and costly than the detection of normal
behavior as fraud. Therefore, a good FDS should prioritize transactions with
a higher misclassification cost.
– Lack of Real-Time FDS: most of the existing FDS reported in the literature work
on archival data to drive future security policies. The fraud must be detected and
blocked-in real time to avoid future fraud.

3 Machine Learning Methodology and Application

to Fraud Detection

The fraud detection is a Big Data problem. Our ML methodology and application
to fraud detection takes into account all the data properties intrinsic to a Big Data
102 S. El Hajjami et al.

system. It includes four main layers: data collection and storage, data processing,
model building and evaluation before pre-production.

3.1 Data Collection and Storage

This layer consists of collecting data transactions made by customers. In their

raw state, they contain information classified as confidential. The vast majority of
payment systems provide interfaces to access transaction data. Legal data protection
rules must be observed.
The quality of ML models is measured, among other things, by the quality of the
data from which they are built. Therefore, knowing how to use good data collection
practices is essential to develop high performing models. The data must be, as much
as possible, error-free and contain information relevant to the task at hand [2].

3.2 Data Processing

Data processing is an important and time-consuming task due to its importance

to overall performance. The purpose of data processing is to create what is called
prepared data. Data processing includes various operations:
1. Data cleaning: deletion or correction of records containing corrupted or invalid
values, or for which a large number of columns are missing.
2. Data transformation: converting a numeric characteristic to a categorical char-
acteristic and converting categorical characteristics to a numeric representation.
Some models only work with numeric or categorical characteristics, while others
can handle characteristics of different types.
3. Feature Selection: selecting a subset of the input features for training the model
while ignoring irrelevant or redundant ones.
4. Data Sampling: sampling the dataset before training the predictive model in
order to have more balanced data.

3.3 Model Building

Model development begins by partitioning the datasets into one dataset used to train
a model, another dataset used to test the trained model. This splitting of the data
ensures that the model does not remember a particular subset of data. There are two
steps in a learning model building:
Machine Learning System for Fraud Detection … 103

Table 1 Kaggle credit card

Transactions Majority class Minority class Columns
fraud dataset details
284 807 284 315 492 31

– Training phase: mainly consists of the building or adjusting a model. The funda-
mental goal is to learn a pattern of trends that lend themselves well to generaliza-
tion to new data instead of simply memorizing the data he was able to see during
his training.
– Test phase: once the model is trained, it is important to check whether it behaves
correctly on new examples that are not used for training the model. To do this, the
model is used to predict the response on the test dataset. Then, the predicted target
is compared to the actual response to measure the performance of the model.

3.4 Model Evaluation

It’s important to test, measure and monitor the performance of a predictive model. We
must then define measures to be used for performance evaluation. The assessment
metrics used depends on several factors. Such as, the task of modeling (classification,
regression or segmentation), the context of the problem we are trying to solve as
well as the data distribution. These metrics are also used to compare the model’s
performance and select those that give the best performance.

4 Credit Card Fraud Detection Use Case: Experimentation

and Evaluation

4.1 Data Collection

For data collection, we use the Kaggle Credit Card Fraud Detection dataset1 to
illustrate these different steps. It contains the transactions carried out by credit card
during two days of September 2013 by holders of European cards. The Table 1
provides statistics for the dataset and shows that the minority class (fraud) accounts
for 0.172% of all transactions. Therefore, this dataset is very imbalanced [8]. It
contains 31 digital features. Because some of the input features contain financial
information, the PCA transformation of 28 digital input features (named V 1 , …,
V 28 ) was performed due to privacy concerns. Three of the given features have not
been transformed. Time feature displays the time between the first transaction and
each other transaction in the dataset. The Amount feature is the value of the amount

1 Anonymized credit card transactions labeled as fraudulent or genuin From Kaggle https://www.
kaggle.com/mlg-ulb/creditcardfraud.
104 S. El Hajjami et al.

spent in a single credit card transaction. The Class feature represents labels and takes
two values: value 1 for fraudulent transaction and value 0 otherwise.

4.2 Data Processing

During the processing phase, we proceed to remove missing data. Then, it was
analyzed and all features except Amount and Time were scaled using the PCA
transformation technique. Therefore, the Time and Amount columns are scaled and
normalized to ensure consistency.
For selection features, we are only interested in data features that are able to
separate the two classes (fraudulent, normal). Visualization techniques can be helpful
in this process. Consider the example shown in Fig. 1 showing the class distribution
for some features of our dataset. We can see for V 12 a significant divergence in the
distribution of the two classes. It is therefore a feature with high predictive power. We
therefore keep it when building models. Similarly, we can see for feature V 13 that the
distribution of normal transactions (majority class) corresponds to the distribution of
fraudulent transactions (minority class). This feature cannot effectively contribute to
the separation between the two classes, we eliminate it from the dataset. We carried
out this process for all the 28 features. As a result, 11 relevant features were selected
for our experiments: V 3 , V 4 , V 9 , V 10 , V 11 , V 12 , V 14 , V 16 , V 17 , V 18 and V 19 .
The main difficulties are the skewness and data overlap in fraud detection
cases. Our goal is to correctly classify the fraudulent transactions. So, we use
One Side Behavioral Noise Reduction (OSBNR) [9, 10] as a sampling approach.
This approach manages behavioral noise in order to improve the classification of
fraudulent transactions.

Fig. 1 Class distribution histogram on some features

Machine Learning System for Fraud Detection … 105

4.3 Model Building

Since there is no rule of thumb for dividing a data set into training and test sets,
we choose the 70/30 rule for training/test sets. Several studies reported Random
Forest and Multilayer Perceptron to get the best performance [1, 3, 11, 12], this is
one of the reasons we are adopting Random Forest and Multilayer Perceptron in our
experiments:
– Random forest (RF) is an algorithm that consist of many decision trees. This
algorithm works best when there are more trees in the forest. Each decision tree
in the forest gives results. These results are merged in order to obtain a more
precise and stable prediction [13].
– MultiLayer perceptron (MLP) is an artificial neural network with direct action
which is made up of at least 3 layers of nodes: entry layer, hidden layer and exit
layer. Each node uses an activation function. The activation function calculates
the weighted sum of its inputs and adds a bias. This allows us to decide which
neuron should be removed and not taken into account in the external connections.

4.4 Model Evaluation

Various metrics can be used to measure the predictive accuracy of a model. In our
case, the major challenge is to tackle the imbalance problem, since legitimate trans-
actions are much more numerous than fraudulent transactions (less than 1% of total
transactions). This problem often leads to extremely high accuracy where a model
can reach up to 99% of the prediction accuracy, ignoring the 1% of minority class
cases. In other words, accuracy does not reflect reality in this data imbalance case.
The ML model performance is evaluated using AUC (Area Under The Curve)
of the ROC Curve (Receiver Operating Characteristics) [14]. The ROC curve is
generated by plotting the true positives rate TPR, against the false positives rate
FPR, on all decision thresholds. The true positive rate TPR is the proportion of
real positives (fraud) that are correctly identified as the positive class, and the false
positive rate FPR measures the proportion of real positives (fraud) that are wrongly
classified. AUC (1) is a concise measure of the performance of the ROC curve with
a single value between 0 and 1, where a perfect model has a score close to 1. In
addition, AUC has been shown to be effective for class imbalance [14, 15].

AUC = (1 + TPR − FPR)/2 (1)

106 S. El Hajjami et al.

Fig. 2 AUC results for MLP 1

AUC Metric
Relevant Features
Case (80%,20%)
Case (70%,30%)
0.98
Case (60%,40%)

0.96

0.94

0.92

AUC Metric
0.9

0.88

0.86

0.84

0.82

0.8
MLP MLP_CNN MLP_ENN MLP_AKNN MLP_RENN MLP_TL MLP_OSS MLP_NCR MLP_OSBNR

ML Models used

Fig. 3 AUC results for RF 1

AUC Metric
Relevant Features
Case (80%,20%)

0.98 Case (70%,30%)

Case (60%,40%)

0.96

0.94

AUC Metric 0.92

0.9

0.88

0.86

0.84

0.82

0.8
RF RF_CNN RF_ENN RF_AKNN RF_RENN RF_TL RF_OSS RF_NCR RF_OSBNR

ML Models used

4.5 Results Analysis

Figures 2 and 3 show some results analysis done for both MLP anf RF models
using the AUC metric. They also show the model’s performance of a new approach
called OSBNR. In~\cite{Salma2020b,el2020machine}, MLP and RF algorithms are
combined with different resampling methods to study the performance of the resulting
ML models.

5 ML Model Development Platform

5.1 General Architecture

Horizontal scaling ensures the platform profitability when data and processing
requirements vary between models. Adding more machines to a cluster would not
be of much value on its own. What we need is a system that can take advantage of
horizontal scalability and that runs on multiple machines seamlessly, regardless of
the number of machines in the cluster. The choice of a distributed system for the ML
model development platform is then necessary because it operates transparently on
a cluster of machines and automatically manages the necessary resources. Figure 4
shows our cluster formed by a master node and three slave nodes. The cluster is
Machine Learning System for Fraud Detection … 107

Fig. 4 General architecture

of the ML model
development platform

managed by a private cloud-type platform within the L3i laboratory as part of the
FEDER-PLAIBDE project2 . It is based on:
– Operating system: Ubuntu-16.04.3-server-amd64
– Hard disk: type SSD 500 Go
– Processors: 16 Go de RAM (4 sockets, 4 cores)
– Network: 2x10 Go SFP +

5.2 Distributed Data System

In this design, we made the choice of data storage in raw format. Thus, the system
does not depend on any data model. It is therefore natural that the choice of Apache
Hadoop Distributed File System (HDFS) is necessary to manage our distributed
data system. HDFS is a well-known example of a distributed file system in the big
data ecosystem. In HDFS, we use a «master/slave» architecture consisting of a single
NameNode which manages the distributed file system and several DataNodes, which
generally reside on each node of the cluster and manage the physical disks attached
to this node as well as data that is stored physically.

5.3 Distributed Treatment System

In our ML model development platform, we use Apache Spark. It is a versatile

distributed processing engine capable of handling large volumes of data. The versa-
tility of Apache Spark is that it is suitable for a wide variety of large-scale use cases.
In this work, we are particularly interested in its machine learning library called
«Spark ML».

2 https://plaibde.ayaline.com/
108 S. El Hajjami et al.

Spark uses a «maître/esclave» architecture with a central coordinator called

«Driver» and a set of executable workflows called «Executors» located on different
nodes of the cluster. Each application written in Apache Spark gives rise to a pilot
program or «Driver». It segments a Spark application into tasks which are then parti-
tioned between the worker nodes in the distributed cluster. The pilot program also
creates a SparkContext that tells the application how to connect to the cluster and its
underlying services. The slave nodes are where the computational processing phys-
ically occurs. Typically, they are located on the same nodes where the underlying
data is available. These nodes generate processes called «Spark Executor». They
are responsible for performing computational tasks and storing locally cached data.
Executors communicate with the pilot program to receive scheduled functions which
are then executed.

5.4 Spark ML Algorithms

Spark ML is an essential software brick of our ML model development platform. It

offers all the services dedicated to the implementation of these models. In Apache
Spark, can distinguish two APIs:
– Spark ML: based on the DataFrame API. It is a distributed, column-oriented data
structure suitable for learning algorithms.
– Spark MLlib: based on RDD API (Resilient Distributed Datasets). This is a
distributed, object-oriented data representing data.
As part of our ML model development platform, we use Spark ML. At the heart
of it are several layers of software for different services.

5.5 Integrated Development Environment

In our ML model development platform, we use a development environment based on

Jupyter Lab. It is a web application for recording the entire process of development,
testing, evaluating and documenting ML models. The Jupyter product was originally
developed as part of the IPython project. It allows interactive development in several
languages. The name Jupyter itself is derived from the combination of Julia, Python,
and R. Our development environment provides three cores:
– PySpark: for applications written in Python2.
– PySpark3: for applications written in Python3.
– Spark: for applications written in Scala.
Machine Learning System for Fraud Detection … 109

6 Conclusion and Future Work

This work presents a machine learning methodology and a development platform.

We presented the fraud detection problem and approached its resolution using an ML
approach. Also, we showed that this approach requires the design and implementation
of an adequate platform. The presented solution is based on open-source solutions
from the world of Big-Data. Our future work covers other aspects of ML modeling
and deployment platform.

Acknowledgment This work is carried out thanks to the support of the European Union through
the PLAIBDE project of the FEDER-FSE operational program for the Nouvelle-Aquitaine region,
France.

References

1. Hajjami SE, Malki J, Berrada M, Bouziane F (2020) Machine learning for anomaly detection.
Performance study considering anomaly distribution in an imbalanced dataset. In: Cloudtech
2020. IEEE
2. Roh Y, Heo G, Whang SE (2018) A survey on data collection for machine learning: a big data
- AI integration perspective. arXiv:abs/1811.03402
3. Dal Pozzolo A, Caelen O, Le Borgne YA, Waterschoot S (2014) Bontempi G (2014) Learned
lessons in credit card fraud detection from a practitioner perspective. Expert Syst Appl
41(10):4915–4928
4. Maes S, Tuyls K, Vanschoenwinkel B, Manderick B (2002) Credit card fraud detection using
Bayesian and neural networks. In: Proceedings of the 1st international Naiso congress on neuro
fuzzy technologies, pp 261–270
5. Şahin YG, Duman E (2011) Detecting credit card fraud by decision trees and support vector
machines
6. Adewumi AO, Akinyelu AA (2017) A survey of machine-learning and nature-inspired based
credit card fraud detection techniques. Int J Syst Assur Eng Manage 8(2):937–953
7. Puh M, Brkić L (2019) Detecting credit card fraud using selected machine learning algorithms.
In: 42nd international convention on information and communication technology, electronics
and microelectronics. IEEE, pp 1250–1255
8. Dal Pozzolo A, Caelen O, Johnson RA, Bontempi G (2015) Calibrating probability with
undersampling for unbalanced classification. In: IEEE symposium series on computational
intelligence. IEEE, pp 159–166
9. Hajjami SE, Malki J, Bouju A, Berrada M (2020) Machine learning facing behavioral noise
problem in an imbalanced data using one side behavioral noise reduction: application to a fraud
detection. J Comput Inf Eng
10. Hajjami SE, Malki J, Bouju A, Berrada M (2020) A machine learning based approach to
reduce behavioral noise problem in an imbalanced data: application to a fraud detection. In:
International conference on intelligent data science technologies and applications (IDSTA).
IEEE, pp 11–20
11. Bhattacharyya S, Jha S, Tharakunnel K, Westland JC (2011) Dataminingforcredit card fraud:
a comparative study. Decis Support Syst 50(3):602–613
12. Dal Pozzolo A, Boracchi G, Caelen O, Alippi C, Bontempi G (2015) Credit card fraud detec-
tion and concept-drift adaptation with delayed supervised information. In: International joint
conference on Neural networks. IEEE, pp 1–8
110 S. El Hajjami et al.

13. Breiman L (2001) Random forests. Mach Learn 45(1):5–32

14. Bekkar M, Djemaa HK, Alitouche TA (2013) Evaluation measures for models assessment over
imbalanced data sets. J Inf Eng Appl 3(10) (2013)
15. Jeni LA, Cohn JF, De La Torre F (2013) Facing imbalanced data–recommendations for the use
of performance metrics. In: 2013 Humaine association conference on affective computing and
intelligent interaction. IEEE, pp 245–251
U-Net: Deep Learning for Extracting
Building Boundary Collected by Drone
of Agadir’s Harbor

Tarik Chafiq, Hayat Hachimi, Mohammed Raji, and Soufiane Zerraf

Abstract Identification of a specific object in an image might be a trivial task for

humans but are often quite challenging for machines. Recently the field has witnessed
groundbreaking research with cutting-edge results. However, for real-world applica-
tion problems of this research remain a challenge. The approach used is based on a
training model from a Dataset, and this model will be used in all processing to detect
homes from sample images. All the images were extracted from the unmanned aerial
vehicle (UAV) recordings. This paper presents a method to detect segmentation of
the Building footprint using U-Net architecture; in order to build footprints without
needing manual digitizing with higher accuracy.

Keywords Image classification · U-NET · Deep Learning · UAV

1 Introduction

The automation of work carried out by the human being in order to reduce the
execution time, quality of production, and precision has always made dreams to
researchers in the computer science field while simplifying workers’ routine activi-
ties and helping [1, 2] companies provide next-generation services. In recent years,
Artificial intelligence has progressed rapidly, approaching or even exceeding human
precision in certain cases, object detection represents one of the most significant tasks
[3] in the field of computer vision and building footprint detection and outlining satel-
lite or drone imagery is a very useful tool in many types of applications [4], from
population mapping to the monitoring of illegal development, from urban expansion
control to planning a prompt and more reliable rescue response within the case of
catastrophic events. Detecting building footprints in optical, multi-spectral satellite
or drone data is not easy to unravel in a general way due to the acute heterogeneity

T. Chafiq (B) · H. Hachimi · M. Raji · S. Zerraf

Faculty of Sciences, Ben M’sick-University Hassan II, Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 111
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_11
112 T. Chafiq et al.

of material [5], shape, spatial and spectral patterns that will result in diverse envi-
ronmental conditions and construction activities embedded in various parts of the
world.
Deep learning, also known as artificial intelligence approaches to self-learning,
provides new doors for data processing and vision of machines [6]. The possi-
bilities of object detection and pattern recognition are currently being revolution-
ized by convolutional neural networks (CNN) that are used in remote sensing
[7]. CNN enables efficient analysis of image textures, i.e. the contextual signal of
multiple neighboring pixels, unlike common pixel-based methods. The self-learning
capabilities of CNN allow these textures to be analyzed effectively [7].
Even in complex urban scenes, accurate information on urban objects automat-
ically detected by drone images is a fundamental approach to obtaining a reliable
classification result [6]. At this stage, the technologies of remote sensing have become
a promising tool and are continuing to grow at an unparalleled pace [6, 8, 9]. Due
to modern platforms and sensors, such as very high-resolution satellite or unmanned
aerial vehicles (UAVs), the availability of optical Earth observation data revealing
both spatial details and high temporal urban patterns is increasing [10]. Using deep
learning methods, this optical feature space can be expanded further with details on
the 3D urban structure [11, 12]; Effective techniques are required to make full use
of this source of knowledge for urban mapping.

2 Related Works

Some of the previous work contributions use the images segmentation for the auto-
matic detection of buildings based on High Resolution Satellite (VHRS) images
are many [13–18]; in this section, two examples are cited: the works of Menarched
[17], whom was present the potential of SVMs to automatically extract buildings
in suburban area using Very High Resolution Satellite (VHRS) images. Then, SVM
classifier was used to extract buildings. The proposed method has been applied on a
suburban area in Tetouan city (Morocco) and 83.76% of existing buildings have been
extracted by only using color features [17]. Regarding the work of Vakalopoulou [19],
they work for the approach exploits a decomposed interconnected graphical model
formulation where registration similarity constraints are relaxed in the presence of
change detection wish gave the accuracy of 74% [19].
U-Net: Deep Learning for Extracting Building Boundary Collected … 113

3 Materiel and Method

3.1 Study Area

The study of the harbor area of Agadir (Fig. 1), which is the fishermen’s village south
of Safi, Morocco. It is 30 km south of Safi, and 90 km north of Essaouira. The active
maritime population in AGADIR is estimated at 1000 sailors including fishing related
activities. The study area is covered by 1218 images with a surface area of 15 ha and a
maximum altitude of 10 m. Determination of the coordinates of the boundary markers
the first was carried out in relation to the Merchich (national Projected coordinate
system), which adopts the Lambert conformal conical projection.

4 Methodology

In this stage, ArcGIS Pro will be used to classify the UAV using statistical classifi-
cation methods. As defined above, deep learning is a form of machine learning that
depends on many layers of non-linear processing listed in a model for feature iden-
tification and pattern recognition [20]. For object recognition and image processing,
Deep Learning models can be combined with ArcGIS [21]. Using the Classification
Training Sample Manager tools to generate training samples of entities or objects of
interest in ArcGIS Pro, and then using the “Label Object for Deep Learning” tool to

Fig. 1 Map of the Agadir harbor area

114 T. Chafiq et al.

Fig. 2 Schematic diagram, showing overall methodology of the work

label them. Finally converted to a format for the in-depth learning structure. These
training examples are used to train the model using the “Train Deep Learning Model”
rid analysis tool or the ArcGIS API for Python. The trained model associated with
the model definition (and possibly the Python raster function script file) is packaged
and shared as a dlpk (deep learning package) element, which is then used to generate
the raster analysis inference tools. The inference tool extracts specific features or
classifies pixels in the image. The dlpk element can be used multiple times as input
to the raster analysis tool, making it easy to evaluate multiple images at different
locations and over different time periods after model formation (Fig. 2).

4.1 Data Preparation

The orthophoto of the Tan-Tan harbor (Fig. 3) will be used as a training dataset, then
the orthophoto’s harbor of Agadir as a test of the accuracy of the model, and produce
the vector data. In fact, the training dataset has been exported using a tool available
U-Net: Deep Learning for Extracting Building Boundary Collected … 115

Fig. 3 BF on the
Orthophoto of Tan-Tan’s
harbor

Fig. 4 Thematic raster 8bit

unsigned of Tan’s harbor

in ArcGIS Pro which is called Export Training Data For Deep Learning in order to
generate a thematic raster 8bit unsigned (Fig. 4).
The preparation of data can be a very long process involving the selection and
massaging of training in the format needed for any deep learning model [22, 23].
Typical data processing command lines include the splitting of the data into
training and test (validation) sets, various data techniques are implemented, data
structures are created to load the data into the model, batch sizing is specified, etc.
All these long-term activities can be automated by ArcGIS, and the training dataset
exported by ArcGIS is read directly by the planning process.

4.2 Model Architecture

Convolutional Neural Networks (hereafter referred to as CNN) can be seen as MLPs

neural networks particularly adapted to 2D signal processing. These networks were
inspired by the work of Hubel and Wiesel on the visual cortex in mammals [24].
116 T. Chafiq et al.

Fig. 5 The architecture of a convolutional neural network [27]

The first CNNs date back to the 1980s with the work on the Necognitron by K.
Fukushima [25], but it is in the 1990s that these networks became popularized with
the work of Y LeCun & al about character identification [26].
A CNN is a network of neurons whose layers are linked not by a matrix operation
but by a two-dimensional convolution. NDCs allow the direct processing of large
amounts of data such as images for several reasons:
– The images are spatially correlated (the values of adjacent pixels are generally
very close) and the convolution layers make it possible to create links between
these spatially correlated data.
– Convolution layers are usually followed by sub-sampling (pooling) layers that
greatly reduce the size of the data.
A CNN architecture is formed by a stack of independent processing layers (Fig. 5):

4.3 Modeling Approached

Instead of developing a model from scratch, an existing convolutional neural network

model is used for the segmentation of the U-net image, originally designed for the
segmentation of biomedical image [28, 29]. Once formed, the network was capable
of producing a binary at the classification pixel level (whether under construction or
not) with high accuracy.
U-net is a particular form of FCN [30] that has gained a great deal of interest in
the segmentation of biomedical images using a reduced data collection but is known
to be very effective for the pixel-by-pixel classification of satellite and aerial images.
U-net has been created by Olaf Ronneberger [28]. For the segmentation of
biomedical images (Fig. 6). The architecture contains two paths.
The first path is the path of contraction or encoder [31] used to catch the context
in the image. The encoder is a conventional stack full of clustering layers.
The second path is the symmetrical expansion path or decoder [32] which is used
to allow precise localization using transposed convolutions. It is therefore a fully
U-Net: Deep Learning for Extracting Building Boundary Collected … 117

Fig. 6 U-NET architecture [28]

convolutional network (FCN) end-to-end, i.e. it contains only convolutional layers

and no dense layers, allowing it to accept images of any size.

4.4 Optimal Learning Rate

Tuning hyper-parameters means optimisation of profound learning. We will find an

’optimal learning rate’ for our training data template in this phase. The learning rate
is an important parameter in the stage of training the model. An extremely high rate
of learning would contribute to the convergence of our paradigm into an unstable
approach. we have found optimum research speeds at which a robust model can be
trained fast enough.

4.5 Evaluating Model Accuracy

The complete training data was divided into training set and validation was defined
at the preparatory data level, and the validation set is by default 0,22 or 20% and the
rest of the training data is 80%. Here, it lets us know how well our models generalize
data that they have never observed, and how they avoid overfitting of the model. As
we can see in this case, our model (Fig. 7) can also categorize pixels in construction
or not with an accuracy of 96%.
118 T. Chafiq et al.

Evaluang Model Accuracy

100%

80%

60%

40%

20%

0%
1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930

train_loss valid_loss accuracy

Fig. 7 Evaluating model accuracy

Fig. 8 Thematic raster 8bit

unsigned of Agadir’s harbor

5 Results

5.1 Preview Results

In this step, the Geoprocessing tool which called Classify Pixels Using Deep Learning
is available in ArcGIS Pro was used to process the images and generate a classified
raster (Fig. 8).
The results of this tool will take the form of a classified raster that contains both
background and building footprints.

5.2 Extracting Building Footprints

To extract building footprint, the Pixel Classification tool using deep learning to
segment the imagery through the model and post-processed the resulting raster in
ArcGIS Pro. For post-processing the classified imagery, the Model Builder (MB) has
been used, which is a visual programming language for building workflows [33].
U-Net: Deep Learning for Extracting Building Boundary Collected … 119

Fig. 9 Building footprints

extracted using automated
process

The results of the building footprint model generated on Agadir’s Harbor (Fig. 9);
this is few subsets of results overplayed on the drone recording. furthermore, it works
well when any space divides buildings from each other.

6 Conclusion

The model described in this paper offers many advantages among which the ability
to elucidate better identification and automatic detection of building-footprint based
on the Training-Model. Moreover, it shows that the benefit of using improved ground
truth labels for building segmentation without human intervention is a testament of
the progress made in AI. Furthermore, such goals have been completed by making the
system capable of recognizing buildings in different situations, achieving a precision
rate of 96% after a training period on images acquired and processed for this purpose.
This would also enable future research to focus on adding a third dimension to
generate an indexable object 3D model that could give more value and precision to
the building- footprint.

References

1. Naya F, et al (2006) Workers’ routine activity recognition using body movements and location
information. In: 2006 10th IEEE international symposium on wearable computers. IEEE
2. Chafiq T, Ouadoud M, Elboukhari K (2020) Covid-19 forecasting in Morocco using FBprophet
Facebook’s Framework in Python. Int J 9(5)
3. Ding P et al (2018) A light and faster regional convolutional neural network for object detection
in optical remote sensing images. ISPRS J Photogram Remote Sens 141:208–218
4. Martin F-M et al (2018) Using single-and multi-date UAV and satellite imagery to accurately
monitor invasive knotweed species. Remote Sens 10(10):1662
5. Darraj B, Gourja B, Faiq A, Zerraf S, Tace EM, Belaaouad S (2019) The use of concept maps
to evaluate learning strategies for nursing students. Int J Innov Technol Explor Eng 9(1)
6. Kattenborn T, Eichel J, Fassnacht FE (2019) Convolutional Neural Networks enable efficient,
accurate and fine-grained segmentation of plant species and communities from high-resolution
UAV imagery. Sci Rep 9(1):1–9
120 T. Chafiq et al.

7. Huang B, et al (2018) Large-scale semantic classification: outcome of the first year of Inria
aerial image labeling benchmark. In: IGARSS 2018–2018 IEEE international geoscience and
remote sensing symposium. IEEE
8. Corbane C et al (2015) Remote sensing for mapping natural habitats and their conservation
status–New opportunities and challenges. Int J Appl Earth Obs Geoinf 37:7–16
9. White JC et al (2016) Remote sensing technologies for enhancing forest inventories: A review.
Can J Remote Sens 42(5):619–641
10. Hirschmugl M et al (2017) Methods for mapping forest disturbance and degradation from
optical earth observation data: a review. Current Forestry Reports 3(1):32–45
11. Kattenborn T et al (2019) UAV data as alternative to field sampling to map woody invasive
species based on combined Sentinel-1 and Sentinel-2 data. Remote Sens Environ 227:61–73
12. Fassnacht FE, et al (2017) Estimating stand density, biomass and tree species from very high
resolution stereo-imagery–towards an all-in-one sensor for forestry applications? Forestry: Int
J Forest Res 90(5):613–631
13. Porway J, Wang Q, Zhu SC (2010) A hierarchical and contextual model for aerial image parsing.
Int J Comput Vision 88(2):254–283
14. Lin Y et al (2014) Rotation-invariant object detection in remote sensing images based on
radial-gradient angle. IEEE Geosci Remote Sens Lett 12(4):746–750
15. Zhang F et al (2016) Weakly supervised learning based on coupled convolutional neural
networks for aircraft detection. IEEE Trans Geosci Remote Sens 54(9):5553–5563
16. Moranduzzo T, Melgani F (2014) Detecting cars in UAV images with a catalog-based approach.
IEEE Trans Geosci Remote Sens 52(10):6356–6367
17. Benarchid O, Raissouni N (2013) Support vector machines for object based building extraction
in suburban area using very high resolution satellite images, a case study: Tetuan, Morocco.
IAES Int J Artif Intell 2(1)
18. Liu Z et al (2016) Ship rotated bounding box space for ship extraction from high-resolution
optical satellite images with complex backgrounds. IEEE Geosci Remote Sens Lett 13(8):1074–
1078
19. Vakalopoulou M, et al (2015) Simultaneous registration and change detection in multitemporal,
very high resolution remote sensing data. In: Proceedings of the IEEE conference on computer
vision and pattern recognition workshops
20. Liu W et al (2017) A survey of deep neural network architectures and their applications.
Neurocomputing 234:11–26
21. Zhang X et al (2020) How well do deep learning-based methods for land cover classification and
object detection perform on high resolution remote sensing imagery? Remote Sens 12(3):417
22. Chafiq T, et al (2018) Application of data integrity algorithm for geotechnical data quality
management12(8)
23. Ouadoud M, Chafiq T, Chkouri MY (2018) Designing an IMS-LD model for disciplinary
information space of learning management system. In: Proceedings of the 3rd international
conference on smart city applications (2018)
24. Hubel DH, Wiesel TN (1962) Receptive fields, binocular interaction and functional architecture
in the cat’s visual cortex. J Physiol 160(1):106
25. Fukushima M (1980) Dirichlet forms and Markov processes, vol. 23. Elsevier
26. LeCun Y, Denker JS, Solla SA (1990) Optimal brain damage. In: Advances in eural information
processing systems
27. Krut P (2019) MNIST Handwritten digits classification using a Convolutional
Neural Network (CNN). https://towardsdatascience.com/mnist-handwritten-digits-classific
ation-using-a-convolutional-neural-network-cnn-af5fafbc35e9. Accessed 12 Apr 2020
28. Ronneberger O, Fischer P, Brox T (2015) U-net: Convolutional networks for biomedical image
segmentation. In: International conference on Medical image computing and computer-assisted
intervention. Springer
29. Ouadoud M, et al (2019) Generate a meta-model content for collaboration space of learning
management system compatible with IMS-LD
U-Net: Deep Learning for Extracting Building Boundary Collected … 121

30. Zhou X et al (2017) Deep learning of the sectional appearances of 3D CT images for anatomical
structure segmentation based on an FCN voting method. Med Phys 44(10):5221–5233
31. Guo X, et al (2017) Deep clustering with convolutional autoencoders. In: International
conference on neural information processing. Springer
32. Forney GD Jr (1974) Convolutional codes III. Sequential decoding. Inf Control 25(3):267–297
33. Allen DW (2011) Getting to know ArcGIS modelBuilder. Esri Press
Analysis and Classification of Plant
Diseases Based on Deep Learning

Assia Ennouni, Noura Ouled Sihamman, My Abdelouahed Sabri,

and Abdellah Aarab

Abstract Plant diseases identification is generally done by visual evaluation. The

diagnosis quality depends strongly on professional knowledge. However, agricultural
expertise is not easily learned. To overcome this issue, automatic analysis and clas-
sification of plant diseases through image processing and artificial intelligence is an
encouraging solution and can reduce the lack of agricultural knowledge. This alliance
between image processing and machine learning and/or deep learning approaches
has produced very good results in medical imaging and has enabled the develop-
ment of robust systems to assist in the diagnosis of several diseases. Deep learning
(DL) and Machine learning approaches became the most promising tools for image
analysis and classification. The objective of our work is to conduct a comparative
study between different deep learning architectures for the analysis and classifica-
tion of plant diseases. Four DL-based architectures have been implemented such as
MobileNet, AlexNet, Inception V3, and VGG16. Simulation results have shown that
the MobileNet architecture, even simple, has allowed better classification accuracy.

Keywords Plant diseases detection · Plant diseases classification · Image

classification · Deep learning · Machine learning · Convolutional neuron network

1 Introduction

Plant diseases are considered to be alterations of plants that affect and may damage
their vital functions. Plant diseases are caused mainly by viruses, fungi or bacteria.
Plant diseases have a great impact on agricultural productivity. Plant diseases have
increased considerably in recent years, according to the FAO (Food and Agriculture

A. Ennouni (B) · N. O. Sihamman · M. A. Sabri

Department of Computer Science, LISAC Laboratory, Faculty of Sciences Dhar-Mahraz,
University Sidi Mohamed Ben Abdellah, Fez, Morocco
A. Aarab
Department of Physics, LISAC Laboratory, Faculty of Sciences Dhar-Mahraz, University Sidi
Mohamed Ben Abdellah, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 123
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_12
124 A. Ennouni et al.

Organization of the United Nations). This is mainly due to climate change, glob-
alization, and the low resilience of production systems, … Early detection of plant
diseases based on visible symptoms on leaves is one of the main challenges to protect
cultures. This mainly depends on the farmer’s knowledge and expertise. However,
not all actors in the agricultural domain are well experienced and especially small
farmers. To overcome this problem and reduce the lack of agricultural knowledge
problem, automatic detection, and recognition of plant diseases combining image
processing and artificial intelligence is a promising solution [1]. Machine Learning
(ML) and Deep Learning (DL) techniques have become more promising approaches
in the medical field because of their ability to learn from reliable and discriminating
visual features [2–4]. For their success, ML and DL have been used in many fields
including agriculture [5, 6]. Artificial Intelligence (AI) and the Internet of Things
(IOT) have been used to develop the concept of Intelligent Agriculture [7, 8].
The objective of our research is to propose a comparative study between four
different deep learning architectures for the analysis and classification of plant
diseases. Namely MobileNet, AlexNet, Inception V3, and VGG16. These archi-
tectures have been implemented and tested. Simulation results have shown that the
MobileNet architecture, even simple, has allowed better classification accuracy.
The rest of this article is as follows: in the following section, we will present the
different plant diseases and then present the different DL-based image classification
architectures and more precisely the classification of plant diseases. We will present
in detail the different DL architectures used and describe our dataset. Implementation
results will be presented afterward. And we end this work with a conclusion.

2 Plant Diseases

Plants are living beings. They are one of the three main groups in which living things
are usually classified (the other two are animals and fungi). Unlike other living beings,
most plants are fixed to the ground by roots, which make them highly dependent on
the conditions of their environment [9].
Plant diseases are damages that disrupt or modify their vital duties. All plant
species are exposed to diseases. Some plants are immune to certain kinds of diseases
while others are particularly prone to them [10]. Plant diseases can be classified as
either they are infectious or not or to the causal agent. Therefore, leaf diseases are
either fungal, bacterial, or viral [9, 10].

2.1 Bacterial Plant Diseases

Bacterial plant infection symptoms are very similar to the symptoms of fungal plant
diseases. Infections are recognized by the presence of leaf spots, wilting, burns,
cankers, scabies, and rot. They can appear on fruits, leaves, stems, and/or flowers.
Analysis and Classification of Plant Diseases Based on Deep Learning 125

Fig. 1 Bacterial leaf blight on wheat and Bacterial strand test on cut stems

Fig. 2 Sclerotinia Infected

by Soybeans

When they affect plants, pathogenic bacteria can be responsible for serious and
potentially disastrous diseases [11]. The following figure presents two bacterial plant
diseases (Fig. 1).

2.2 Fungal Plant Diseases

Fungi cover rust, mildew, mildew, fungus, and others are the cause of diseases that
are not risky but are harmful. There are 144000 known species. Some fungi are used
in medicine. A distinction is made between those that live in the soil or water and that
considered parasitic for plants and animals [12]. An example of the fungal disease
Sclerotinia is presented in Fig. 2.

2.3 Viral Plant Diseases

Organisms can be infected with viruses or viroids. Viruses differ from viroids by their
protein layer. Both are mainly transmitted by insects or nematodes. The symptoms
of this infection vary and depend on the type of virus infecting. The viral disease can
affect all areas of the plant (see Fig. 3). Plant viruses are mainly isometric (polyhedral)
or stem-shaped [13].
126 A. Ennouni et al.

Fig. 3 Necrotic spot virus

on chili leaves

3 Image Classification Based on Deep Learning

Computer vision has made prodigious steps in recent years thanks to machine learning
and deep learning [3, 3]. CNN (Convolutional Neural Network), denoted also as
ConvNet, is a classification algorithm that can be used as both machine learning (one
layer) and deep learning (many layers). CNN does not require much pre-processing
as other classification algorithms.
A standard CNN architecture is generally composed of a succession of convolu-
tion and pooling layers and at the end there one or more fully connected layers. An
average pooling layer replaces the fully connected in some architectures. To optimize
the performance of the CNN deep learning architecture, some regulatory units like
dropout and batch normalization are incorporated. Designed Deep Learning archi-
tectures differ essentially by the arrangement of the CNN components which play
a fundamental role in their performance. A large number of parameters have to be
defined for a given deep learning architecture such as the layers number, the neurons
number in each layer, the size of the filters mask, the weights of the neurons, the
activation function to be used, the best learning rate obtained, … [25, 25].
Several CNN architectures have been proposed in the literature. The following
table lists the most successful architectures, specifying for each the number of
parameters, the authors, the year, and the depth (Table 1).

Table 1 The most successful CNN architectures

Architectures References Year Nb of parameters Depth
LeNet-5 [14] 1998 60,000 5
AlexNet [15] 2012 60 million 8
VGG [16] 2014 138 million 19
GoogleNet [17] 2015 4 million 22
Inception V3 [18] 2015 23 million 159
Inception V4 [19] 2016 35 million 70
ResNet [20] 2016 25 million 152
MobileNetV1 [21] 2017 4.2 million 28
MobileNetV2 [22] 2018 3.47 million 53
Analysis and Classification of Plant Diseases Based on Deep Learning 127

Fig. 4 VGG16 macro

architecture

4 Plant Diseases Classification

In this work, we will conduct a comparative study between 4 Deep Learning architec-
tures for plant diseases classification. The chosen architectures are AlexNet, Incep-
tionV3, VGG16, and MobileNet. The database used contains images of diseased
crop leaves classified into 38 classes and healthy ones [23].

4.1 VGG

K. Simonyan and A. Zisserman proposed a convolutional neural network called

VGG in their paper published in 2015 entitled “Very Deep Convolutional Networks
for Large-Scale Image Recognition” [16]. VGG is characterized by its simplicity
because it uses 3 × 3 convolutional stacked layers on top of each other at an increasing
depth.
The reduction in volume size is managed by max-pooling which is used as a down-
sampling strategy. Two fully connected layers (each layer contains 4,096 nodes) are
then followed by the “softmax” activation function. Two versions were proposed;
VGG16 and VGG19. Where 16 and 19 refers to the number of layers in the network.
In 2014, the 16 and 19 layer-networks were considered as very deep architectures.
Figure 4 shows the VGG16 macro-architecture.

4.2 Inception V3

Authors Szegedy et al. introduced in 2015 the micro-architecture “Inception” in their

article entitled “Going deeper with convolutions” [17]. The original implementation
of this architecture was called GoogLeNet. Later releases were then called Inception
128 A. Ennouni et al.

Fig. 5 Inception modules with dimension reduction [17]

VX where X refers to the number of the version Google published. Figure 5 presents
the GoogLeNet modules.
The purpose of the start-up module is to calculate 1 × 1, 3 × 3, and 5 × 5
convolutions in the same network module. The output is then concatenated and feeds
into the next layer of the network. Compared to VGG and ResNet the Inception V3
weights (96 MB) being the smallest.

4.3 MobileNet

MobileNet architecture is a lightweight deep neural network that provides a robust

model for embedded and mobile applications [22]. It is a simplified architecture that
uses separable convolutions. Figure 6 shows the MobileNet architecture where we
can see the depth-wise separable convolutions.

Fig. 6 MobileNet architecture

Analysis and Classification of Plant Diseases Based on Deep Learning 129

Fig. 7 AlexNet architecture

4.4 AlexNet

In 2012, AlexNet considerably outperformed all the prior competitors and won the
challenge by reducing the top-5 error from 26% to 15.3%. The second-place top-5
error rate, which was not a CNN variation, was around 26.2% (Fig. 7).
The network had a very similar architecture as LeNet by Yann LeCun et al. but was
deeper, with more filters per layer, and with stacked convolutional layers. It consisted
11 × 11, 5 × 5, 3 × 3, convolutions, max pooling, dropout, data augmentation, ReLU
activations, SGD with momentum. It attached ReLU activations after every convo-
lutional and fully-connected layer. AlexNet was trained for 6 days simultaneously
on two Nvidia Geforce GTX 580 GPUs which is the reason for why their network is
split into two pipelines. AlexNet was designed by the SuperVision group, consisting
of Alex Krizhevsky, Geoffrey Hinton, and Ilya Sutskever [23].

Plant Village Dataset Description

In our study, we will use the Plant Village (PV) Dataset for the evaluation of the deep
learning architectures. PV dataset is one of the most used dataset for the evaluation
of plant disease classification algorithms. It contains images of infected and healthy
crop leaves on a uniform background. In our simulations, we will use an offline
augmentation of the dataset which is still available in GitHub [23]. The PV dataset
contains 87K RGB images classified into 38 subsets. The dataset is provided with
training and validation subsets.

Performance Measures
In the pattern recognition field, information retrieval and automatic classification,
accuracy (or positive predictive value) are the proportion of relevant items among all
proposed items; recall (or sensitivity) is the proportion of relevant items proposed
among all relevant items. These two notions thus correspond to a design and a rele-
vance measure. Precision can thus be seen as a measure of accuracy or classification
quality, while recall is a completeness measure or classification quantity. In the multi-
class context (where the number of classes is more than 1), the recall and precision
average over all classes is calculated by the macro-mean. It calculates first the recall
and precision over each class and then calculates the average of the precision and
recall over the classes:
130 A. Ennouni et al.

n
pr ecision i
pr ecision = i=1
(1)
n
n
i=1 r ecall i
r ecall = (2)
n
n
accuracy i
r ecall = i=1
(3)
n
Where n is the number of classes and i the current class.
TPi TPi TPi + TNi
precisioni = recalli = accuracyi =
TPi + FPi TPi + FNi TPi + TNi + FPi + FNi

TP is The True Positives number, FP is the False Positives number, FP is the False
Positives number, and FN is the False Negatives number.

5 Experimental Results

Simulations have been tested on a Windows 10 computer with Intel Core i7 3.6 GHz
CPU and 16 GB of RAM.
In our study, we chose 4 different depth sizes CNN architectures: VGG16, Incep-
tionV3, AlexNet and MobileNet. The number of epoch for the 4 DL architectures is
25. Table 2 presents the evolution of training and validation accuracies, Training and
validation losses, precision, recall, and the accuracy of each of the 4 DL architectures.
As it is a multi-classification problem with 38 classes in total, we can conclude
that the 4 architectures performed well and allowed an acceptable classification
rate. The classification errors are detected particularly at the level of the “tomato
late blight" class. This is mainly due to a confusion between several classes, which
are mainly tomato related diseases, namely: “early tomato blight, septoria leaf spot
tomato blight, and tomato target spot”. Even visually, as shown in Fig. 8, the different
diseases represent visual similarities on tomato leaves. We have also noticed that
there are visual similarities between other diseases which did not allow an excellent
classification. To remedy this problem, it is possible to look for other information
that would allow the discrimination of these different diseases and to consider using
ensemble learning by concatenating features extracted from DL architectures and
handcrafted features.
The classification accuracies of Alexnet and MobileNet are better than Inception
V3. However, since the precision and recall values obtained by MobilNet are higher
than Alexnet, we can conclude that the model build using MobileNet architecture is
the most suitable to better classify plant disease.
Analysis and Classification of Plant Diseases Based on Deep Learning 131

Fig. 8 Examples of classification confusion of leaf samples

Table 2 The evolution of Training and validation accuracies, Training and validation losses,
precision, recall, and the accuracy of each of the 3 DL architectures
Training and validaƟon Accuracy Training and ValidaƟon Loss P R A
AlexNet

0.86 0.96 0.95

VGG16

0.89 0.96 0.92

IncepƟon V3

0.89 0.95 0.91

MobileNet

0.94 0.97 0.95

132 A. Ennouni et al.

6 Conclusion

Farmers usually identify plant diseases by visual assessment. The quality of the diag-
nosis depends strongly on the knowledge of the profession. However, human exper-
tise is not easily acquired by all actors in the agricultural world. Image processing
and analysis in combination with artificial intelligence represent an effective solution
to assist in the recognition and classification of plant diseases. Deep learning and
especially convolutional neural networks (CNN) was recently used in the medical
field and sees itself as a very promising solution. In this work, several DL-based
architectures such as AlexNet, VGG16, Inception V3, and MobileNet have been
tested and compared using the Plants Village database. The MobileNet architecture,
even though very simple in comparison with the others, showed very encouraging
results.

References

1. Prakash RM, Saraswathy PG, Ramalakshmi KH, Mangaleswari TK (2017) Detection of

leaf diseases and classification using digital image processing, In: International conference
on innovations in information, embedded and communication systems (ICIIECS), pp 1–4,
Coimbatore
2. Sabri MA, Filali Y, Ennouni A, Yahyaouy A, Aarab A (2019) An overview of skin lesion
segmentation features engineering, and classification. In: Intelligent decision support systems:
applications in signal processing, De Gruyter, Berlin, Boston, pp 31–52
3. Filali Y, EL Khoukhi H, Sabri MA, Aarab A (2020) Efficient fusion of handcrafted and pre-
trained CNNs features to classify melanoma skin cancer. Multimed Tools Appl (2020)
4. Sabri MA, Filali Y, Khoukhi H, Aarab A (2020) Skin cancer diagnosis using an improved
ensemble machine learning model. In: 2020 international conference on intelligent systems
and computer vision (ISCV), Fez, Morocco, pp 1–5
5. Sue HL, Hervé G, Pierre B, Alexis J (2020) New perspectives on plant disease characterization
based on deep learning. Comput Electron Agri 170:105220. ISSN 0168–1699
6. Shiv RD, Anand SJ (2012) Detection and classification of apple fruit diseases using complete
local binary patterns. In: IEEE computer and communication technology (ICCCT), 2012 third
international conference on computer and communication technology, Allahabad, pp 346–351
7. Badreddine M, El-Bay B, Samia B, Salim C (2020) A study of LoRaWAN protocol performance
for IoT applications in smart agriculture. Comput Commun 164:148–157. ISSN 0140–3664
8. Fanyu B, Xin W (2019) A smart agriculture IoT system based on deep reinforcement learning.
Fut Gen Comput Syst 99:500–507. ISSN 0167–739X
9. El-Sayed AMA, Rida SZ, Gaber YA (2020) Dynamical of curative and preventive treatments
in a two-stage plant disease model of fractional order. Chaos Sol Fractals 137:109879. ISSN
0960–0779
10. Vijai S, Namita S, Shikha S (2020) A review of imaging techniques for plant disease detection.
Artif Intell Agric 4:229–242. ISSN 2589–7217
11. Sujeet V, Tarun D (2016) A novel approach for the detection of plant diseases. Int J Comput
Sci Mob Comput 5(7):44–54. ISSN: 2320–088X
12. Britannica. https://www.britannica.com/science/plant-disease/Epiphytotics. Accessed 19 Dec
2020
13. Saleem MH, Potgieter J, Mahmood Arif K (2019) Plant disease detection and classification by
deep learning, Plants 8(11): 468
Analysis and Classification of Plant Diseases Based on Deep Learning 133

14. Ohioline. https://ohioline.osu.edu/factsheet/plpath-gen-5. Michigan State University Exten-

sion. Accessed 19 Dec 2020
15. LeCun Y, Bottou L, Bengio Y, Haffner P (1998) Gradient-based learning applied to document
recognition. Proc IEEE 86:2278–2324
16. Krizhevsky A, Sutskever I, Hinton GE (2012) ImageNet classification with deep convolutional
neural networks. Adv Neural Inf Process Syst 1–9
17. Simonyan K, Zisserman A (2015) Very deep convolutional networks for large-scale image
recognition. ICLR 75:398–406
18. Szegedy C, Liu W, Jia Y, et al (2015) Going deeper with convolutions. In: 2015 IEEE conference
on computer vision and pattern recognition (CVPR). IEEE, pp 1–9
19 Szegedy C, Vanhoucke V, Ioffe S, et al (2016) Rethinking the inception architecture for
computer vision. In: Proceedings of the IEEE computer society conference on computer vision
and pattern recognition. IEEE, pp 2818–2826
20. Szegedy C, Ioffe S, Vanhoucke V (2016) Inception-v4, Inception-ResNet and the impact of
residual connections on learning. arXiv Prepr arXiv160207261v2 131:262–263
21. He K, Zhang X, Ren S, Sun J (2015) Deep residual learning for image recognition. Multimed
Tools Appl 77:10437–10453
22. Sandler M, Howard A, Zhu M, Zhmoginov A, Chen LC (2018) MobileNetV2: inverted resid-
uals and linear bottlenecks. In: 2018 IEEE/CVF conference on computer vision and pattern
recognition. IEEE, pp 4510–4520
23. Sandler M, Howard A, Zhu M, Zhmoginov A, Chen L-C (2019) MobileNetV2: inverted
residuals and linear bottlenecks. arXiv:1801.04381
24. Github. https://github.com/spMohanty/PlantVillage-Dataset. Accessed 19 Dec 2020
25. Kafi M, Maleki M, Davoodian N (2015) Functional histology of the ovarian follicles as deter-
mined by follicular fluid concentrations of steroids and IGF-1 in Camelus dromedarius. Res
Vet Sci 99:37–40
26. Shin H-CC, Roth HR, Gao M et al (2016) Deep convolutional neural networks for computer
aided detection: CNN architectures, dataset characteristics and transfer learning. IEEE Trans
Med Imaging 35:1285–1298
A Novel Approach to Data Augmentation
for Document Image Classification Using
Deep Convolutional Generative
Adversarial Networks

Aissam Jadli, Mustapha Hain, and Abderrahman Jaize

Abstract Data augmentation is a procedure where new samples are generated

from the training dataset by applying various techniques and algorithms to improve
machine and deep learning models’ accuracy and generalization ability. The recent
advances in deep learning and computer vision techniques have made scanned docu-
ment classification a painless and straightforward process. However, such approaches
require a lot of labeled data before training and validating the classifiers, which can
be done by augmenting the existing dataset by different means. In this contribu-
tion, we explored using a system based on deep convolutional adversarial networks
(DCGAN) to generate fake document images using an existing scanned documents
dataset. Moreover, we compared conventional data augmentation techniques (rota-
tion, zoom, random cropping, etc.) to DCGAN-based augmentation in a document
classification context. The newly generated data could then be used along with the
labeled data to train a Convolutional Neural Network (CNN) to classify scanned docu-
ment images. This experiment compared the performances of a system trained on
different datasets (original labeled data, GAN-augmented dataset, hybrid augmented
dataset). Its results revealed the effectiveness of the proposed approach.

Keywords Document classification · Data augmentation · Image classification ·

Deep learning · DCGAN

1 Introduction

Enterprise Resource Planning (ERP) is a modularized software designed and imple-

mented to manage various internal and external processes of a company, including
Client relationship management (CRM), Supplier relationship management (SRM),
finances and accounting, procurement, manufacturing, logistics, inventory, etc. The

A. Jadli (B) · M. Hain

SISM Team, I2SI2E Laboratory, ENSAM Casablanca, University Hassan II, Casablanca, Morocco
A. Jaize
LAVET Laboratory, FST Settat, University Hassan I, Settat, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 135
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_13
136 A. Jadli et al.

daily workflow of an organization usually generates a large volume of documents

covering various transactions. Moreover, most popular ERP softwares have a docu-
ment management module able to index, retrieve, store, and process saved documents.
Within such a module, document classification is a useful feature seeing that different
documents may need other processing pipelines. In this context, limited labeled data
can make building accurate classifiers a problematic task due to insufficient training
samples. Consequently, many approaches were thoughtfully researched and explored
to solve the lack of labeled data issues in document classification, such as transfer
learning, data augmentation, etc. [1].
Additionally, recent advances in generative modeling introduced new ways to
generate new realistic images that plausibly look like they come from the same
dataset and can be used for performance optimization.
This paper examined using a Deep Convolutional Generative Adversarial
Networks (DCGAN) approach to augment real-world scanned documents. Generated
data is then used alongside original images to solve a document classification task by
training the classifier used on different datasets and assessing the classification perfor-
mances. This paper will be structured as follows: the second section goes through this
paper’s context and the addressed problematic; the third section defines DCGANs
concepts and technical details, while the fourth section details proposed experiments,
and the fifth section presents the experiment results and draws conclusions.

2 Paper Context

This section introduces the basic concepts of this paper, namely, explaining the
ERP ecosystem and features, determining our problematic boundaries, and finally
exploring related works.

2.1 Enterprise Resources Planning (ERP)

ERP is a modularized and packaged software designed around the company’s

vital processes (e.g. Client Relationship Management, finance and accounting,
manufacturing, etc.) using a central database to manage efficiently and effectively
resources usage and information processing. Additionally, different other modules
exist to support ERP architecture, such as messaging, user support, and document
management, etc.
A Novel Approach to Data Augmentation … 137

2.2 Problematic

Document management is handled using a central document management module to

store, index, and process scanned document images in most modern ERP solutions.
Usually, these documents are processed using a technique called Optical Character
Recognition (OCR) to extract useful pieces of information from them. This being
said, to speed up the process, a preliminary document classification step can be
interesting to optimize recognition results. Usually, a document classification task
is a supervised learning task that needs sufficient labeled data to build accurate
and performant models [4], meaning that the training dataset is ordinarily corre-
lated to the model’s performance and stability. The problem of insufficient datasets
is commonly encountered and can be tackled using different solutions (i.e. data
augmentation, transfer learning, etc.). Generative modeling through DCGANs can
help extend existing dataset with newly generated samples, improving classification
models using existing images.

2.3 Related Works

Document image classification uses different approaches and can be based on super-
vised or unsupervised feature learning, which usually requires a large amount of
labeled training data.
For example, Sherif et al. [2] presented an Exemplar-based unsupervised learning
technique to analyze document layout to improve the document image classification
efficiency without the need to expand the volume of annotated data, Jun et al. [3]
studied the use of Faster RCNN (fast regional convolution for object detection)
architecture to automatically classify documents’ internal structure, Kosaraju et al.
[4] proposed a novel structure-based convolutional neural network for document
layout analysis, called DoT-Net. DoT-Net is a multiclass classifier that can effectively
identify document regions to solve document classification tasks.
Similarly, The recent advances in Generative adversarial Networks (GAN) led
research to explore extensive use for such a concept in different domains and various
purposes [5]. Mansourifar et al. [6] introduced a new concept called Virtual Big
Data to address the problem of insufficiency of training data in many research areas,
Xu et al. [7] compared using GAN-based data augmentation for underwater image
classification with other data-augmentation techniques, Lee et al. [8] proposed an iris
image augmentation based on a conditional generative adversarial network (cGAN)
and a method for improving recognition performance that uses this augmentation
method. Meanwhile, Wang et al. [9] propose a GAN-based solution to augment
training data for improved performance of palmprint recognition.
138 A. Jadli et al.

3 DCGAN-Based Data Augmentation

3.1 Data Augmentation

Data augmentation is a technique widely used in deep learning because overfit-

ting is more likely to occur in training that is based on supervised learning using
inadequate or insu cient dataset. Efficient data augmentation methods may increase
inter-class distances, lessen intraclass distances, and improve performance [9]. Stan-
dard data augmentation techniques that have been implemented in different deep
learning frameworks include random geometric translations, rotations, horizontal and
vertical flips, the application of Gaussian noise, random image cropping, horizontal
and vertical shifting, and in/out zooming [10]. These geometric transform-based data
augmentation methods result in insufficient data variation. Consequently, they are
not fit for solving performance declines due to the input scarcity issue, including the
overfitting problem.

3.2 Deep Convolutional Generative Adversarial Networks

The Generative Adversarial Network architecture and the first empirical demonstra-
tion of the approach were proposed in 2014 by Goodfellow et al. [11], describing
an architecture summarily involving a generator model that takes points from a
latent space (random noise) and generates an image, and a discriminator model that
classifies images as either real (authentic) or fake (output by the generator) (Fig. 1).
The two models are arranged in a competition (in a game theory perception)
where the generator model (Fig. 2) seeks to fool the discriminator model (Fig. 1),
and the discriminator is provided with both examples of real and generated samples
and seeks to distinguish between fake and real images.

Fig. 1 Example of a 64 x 64 discriminator architecture

A Novel Approach to Data Augmentation … 139

Fig. 2 Example of a 64 × 64 generator architecture [12]

4 Our Contribution

The proposed framework is a DCGAN-based system to generate document images

from existing real-world documents. Generated images are used to augment an
existing dataset for a document classification task using a Convolutional Neural
Network (CNN) classifier. The goal is to assess if the generated (fake) document
images hold enough features and variance to effectively augment the training dataset.

Fig. 3 Overview of conducted experiment

140 A. Jadli et al.

4.1 Overview

The experiment discussed in this paper (Fig. 3) can be divided into two different
parts:
• Using DCGAN to generate images of the scanned documents;
• construct a model for visual document classification using generated and real data.
A generator model (G) is trained on images of real-world company’s scanned
documents. The process starts from a latent vector (usually a normal distributed
random vector). It follows with the generator generating an image based on real vector
distribution. The generated image is forwarded alongside the authentic images to the
discriminator model, trying to tell the fake pictures from the real ones. The calculated
loss helps back-propagate the errors improving performance over iterations until the
discriminator cannot tell the real from the generated document images.
Therefore, the previously trained generator is used to generate a subset of fake
documents to create a new dataset, which is made of 50% of real documents and
50% of dummy document images.
A Convolutional Neural Network (CNN) algorithm is used in the next step to create
classification models out from each dataset and compare them based on models’
accuracy to assess if we can train classification models using data generated by
DCGANs.

4.2 Materials and Methods

Dataset Images used in these experiments are scanned real-world documents

provided by different organizations obtained from a sponsoring company archive.
Additionally, these documents cover various pipelines and business processes. We
selected samples of different documents classes and labeled them manually into four
categories:
• Electronic Invoices (EI): Electronic documents generated by software;
• Handwritten Invoices (HI): Documents filled manually by an employee;
• Checks (CH): Bank Checks;
• Receipts (RT).
The distribution of sample classes of the experiment’s dataset is summarized in
Table 1.
For experiment results generalization purposes on different situations, the scanned
document images were taken from various geometric conditions (angles, distances,
rotations, skewness, etc.) and in different lighting conditions (shadows, brightness,
etc.) using a standard smartphone camera. Figure 4 presents samples of different
classes.
A Novel Approach to Data Augmentation … 141

Table 1 Distribution of
Document class Sample count
samples classes and the
number of samples E.I 180
HI 109
CH 106
RT 103
Total 498

Fig. 4 Sample documents from dataset

The objective of this paper is to assess the impact of augmenting the existing
training dataset on the accuracy of the classifier in a document classification task.
The classifier used is a Convolutional Neural Network (CNN) using a simple archi-
tecture with an image input size of (64, 64) with a single channel (grayscale images).
Different techniques can be used to tackle the model’s over-fitting issue, such as max
norm constrains, L2/L1 regularization, and dropout. After comparing these methods,
dropout was simpler to use and very efficient. It was used with a rate of 0.5 despite
roughly doubling the number of iterations required to converge[10].

5 Results and Discussion

5.1 Data Generation

Despite the small size of the dataset (498 images distributed on four classes), the
generator showed good performances at generating fake document images that looked
plausibly real (Fig. 5). The generator performed in EI and HI classes better than CH
and RT classes because of the difference between the samples count in each category
(EI case) and the simplicity of standard layouts (HI case),
142 A. Jadli et al.

Fig. 5 Sample images

generated using DCGAN

5.2 Data Augmentation and Performance Comparison

The second part of the experiment was using a CNN algorithm to classify the
documents on 3 distinct datasets:
• D1: contains only real authentic document images;
• D2: contains original real dataset augmented with classical data augmentation
techniques;
• D3: contains original real dataset alongside images generated in part 1 (using
DCGAN);
• D4: contains original real dataset using the conventional and DCGAN data
augmentation simultaneously.
The models built were standard CNN using the same architecture over the three
datasets. Table 2 shows the model’s performance based on accuracy.
Based on overall results, we can assess that model trained on the conventional
augmented dataset (83% accuracy) performed better than the model trained on the
original dataset (77%), and DCGAN augmented dataset (79%). Due to low vari-
ance in geometric data augmentation techniques, this can lead to model overfitting.
Combining the two approaches helped improve the model accuracy (89%), which
is a considerable performance boost even without hyper-parameter tuning or opti-
mization. Figures 6 and 7 show the confusion matrix of the model trained using the
DCGAN augmented dataset (CNN3) and the model trained using the combination
of conventional techniques and DCGAN (CNN4).

Table 2 Performance
Augmentation technique Accuracy (%)
comparison between
constructed models Raw Dataset 77.1
Classical Augmentation 83.3
DCGAN-based Augmentation 79.2
Conventional + DCGAN Augmentation 89.6
A Novel Approach to Data Augmentation … 143

Fig. 6 Confusion Matrix of

CNN3

Fig. 7 Confusion Matrix of

CNN4

6 Conclusion and Perspectives

This paper investigates the use of machine-generated data to substitute real data in
training machine learning algorithms. The experimental results show that despite a
small dataset size (498 documents), the use of DCGAN based data augmentation can
be a good approach in the document classification context. We can recommend this
approach to be used in practice in a lack of labeled data situation for document image
144 A. Jadli et al.

classification in an ERP ecosystem (i.e. Odoo). We plan to explore new state-of-the-

art DCGAN architectures and more specialized machine learning algorithms in our
future work and investigate new ways to improve document image classification.

References

1. Jadli A, Hain M (2020) Automatic document classification using deep feature selection and
knowledge transfer, April 2020. https://doi.org/10.1109/IRASET48871.2020.9092256
2. Abuelwafa S, Pedersoli M, Cheriet M (2019) Unsupervised exemplar-based learning for
improved document image classification. IEEE Access 7:133738–133748. https://doi.org/10.
1109/ACCESS.2019.2940884
3. Jun C, Suhua Y, Shaofeng J (2019) Automatic classification and recognition of complex docu-
ments based on Faster RCNN. In: 2019 14th IEEE international conference on electronic
measurement and instruments, ICEMI 2019, November 2019, pp 573–577. https://doi.org/10.
1109/ICEMI46757.2019.9101847
4. Kosaraju SC, et al (2019) DoT-Net: document layout classification using texture-based CNN.
In: Proceedings of the international conference on document analysis and recognition, ICDAR,
September 2019, pp 1029–1034. https://doi.org/10.1109/ICDAR.2019.00168
5. Jadli A, Hain M (2020) Toward a deep smart waste management system based on pattern
recognition and transfer learning. In: 3rd international conference on advanced communication
technologies and networking, CommNet 2020, September 2020, pp 1–5. https://doi.org/10.
1109/CommNet49926.2020.9199615
6. Mansourifar H, Chen L, Shi W (2019) Virtual big data for GAN based data augmentation. In:
Proceedings - 2019 IEEE international conference on big data, big data 2019, December 2019,
pp 1478–1487. https://doi.org/10.1109/BigData47090.2019.9006268
7. Xu Y, Zhang Y, Wang H, Liu X (2017) Underwater image classification using deep convo-
lutional neural networks and data augmentation. In: 2017 IEEE international conference
on signal processing, communications and computing, ICSPCC 2017, December 2017, vol
2017–January, pp 1–5. https://doi.org/10.1109/ICSPCC.2017.8242527
8. Lee MB, Kim YH, Park KR (2019) Conditional generative adversarial network-based data
augmentation for enhancement of iris recognition accuracy. IEEE Access 7:122134–122152.
https://doi.org/10.1109/ACCESS.2019.2937809
9. Wang G, Kang W, Wu Q, Wang Z, Gao J (2019) Generative Adversarial Network (GAN) based
data augmentation for palmprint recognition, January 2019. https://doi.org/10.1109/DICTA.
2018.8615782
10. Antoniou A, Storkey A, Edwards H (2017) Data augmentation generative adversarial networks.
arXiv. arXiv, 12 November 2017. https://arxiv.org/abs/1711.04340v3. Accessed 17 Jan 2021
11. Goodfellow I, et al (202) Generative adversarial networks. Commun ACM 63(11): 139–144.
https://doi.org/10.1145/3422622
12. Radford A, Metz L, Chintala S (2016) Unsupervised representation learning with deep
convolutional generative adversarial networks
Knowledge Driven Paradigm
for Anomaly Detection from Tweets
Using Gated Recurrent Units

S. Manaswini, Gerard Deepak, and A. Santhanavijayan

Abstract Online social media has become of great importance in the recent past.
It has proven to be a medium for connectivity as well as effortless publicity. Given
the vast scope of activities that can be performed with the help of social media, it
can also be misused. An anomaly has no concrete definition but the best description
of an anomaly would be to identify it as the one that exhibits abnormal behavior. In
this paper two events have been taken into consideration, namely black lives matter
and demonetization, that took great popularity over social media platforms in recent
years and received mixed emotions and behavioural patterns from the public. The
response to these events have been observed on twitter and an anomaly in the trend
has been predicted in the course of a selected time frame using a proposed system
model to accomplish the process efficiently. The deep learning model uses a Gated
Recurrent Unit classifier along with a domain ontology to enhance the process of
prediction. A comparative graph of other unsupervised training models have also
been included to display the efficiency accomplished through the system designed.
The overall accuracy of the trained system for demonetization is observed to be
95.43% and black lives matter is seen to be 96.18%.

Keywords Anomaly Detection · Deep learning · Domain ontology · Gated

recurrent Unit · Online social media · Twitter

1 Introduction

Technology has come to play a vital role in the connectivity of the global community.
With its speeding advancements in the discovery of new software environments to

S. Manaswini
Department of Computer Science and Engineering, SRM Institute of Science and Technology,
Ramapuram, Chennai, India
G. Deepak (B) · A. Santhanavijayan
Department of Computer Science and Engineering, National Institute of Technology,
Tiruchirappalli, India
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 145
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_14
146 S. Manaswini et al.

bind people across the globe, there are so many platforms that are available online
today, that gives people a chance to encourage thoughts, opinions, talent promotion,
news, forum for publicity of various on-going events, awareness propagation, busi-
ness and so on. To list a few of the former- Facebook, Instagram, Twitter top the charts
for maximum networking with regard to such content. Whereas anomalies are those
that exhibit irregularity in a sequence of data. They are often categorized as illegal,
sarcastic, undesired or counter-productive to the ideology. Anomalous behavior has
both positive and negative impacts. Such data can be helpful in the prediction of
natural disasters, disease out-breaks, emergency detection but on the contrary it can
also be used to promote malevolent activities such as terrorist attacks, fraudulence
using social media. It also takes into consideration, controversial standing of events
and provides a platform to express political opinions for influencing the mob and
trending topics. On a broad spectrum, social media anomalies are driven by either
single individual—‘point anomaly’ or groups of people—group or tribe anomaly
where latter generally refers to malicious behavior of hidden groups of people. These
anomalies can be described as a change in interaction patterns, change in method
of interaction and therefore can be classified into three kinds, based on the pattern
on input of the anomaly namely- content, structural and behavioral. This paper shall
be identifying two such scenarios- ‘Demonetization’ and ‘Black lives matter’ over
selective time frames. Tweets from specific time frames extracted during the peak of
both the issues have been analysed to detect anomalies in content.

Motivation: Any political action or global issue implies uncalled for public attention
wherein the entire cause of implication is twisted by individualistic opinions and the
original motive is often lost. The issues considered above impacted in ways that have
resulted in contrasting opinions. And it is interesting to determine and classify tweets
that stand against/out of topic or those that have exhibited anomalous behaviour over
the given time. The black lives matter protests have been observed over the time
frame of two months- May 20 to July 10, 2020. The mentioned time frame witnessed
a peak in the number of tweets regarding the issue. Demonetization was observed
over five months- November 10 2016 to April 20 2017.

Contribution: This paper proposes a methodology that makes use of Gated Recur-
rent Units (GRU), a kind of Recurrent Neural network (RNN) with a gating mech-
anism. A domain ontology [1, 2] derived from an RDF knowledge base has further
been compared to the former using an integration of cosine similarity and Web Point-
wise Mutual Information (Web PMI) that helps determine the prediction error or the
dissimilarity in comparison which will in turn denote the anomaly in the crawled
dataset. Twitter datasets have been specifically chosen due to the high political usage
and media attention which is crucial to the topic addressed and tweets can be conve-
niently made distinct during pre-processing for easier classification. GRU being an
improvised take on the RNN, have been observed to resolve a preliminary drawback
of the standard RNNs. GRU has resolved the dependency issue of the loss of infor-
mation in recurrent networks and has regulation over the input using structures called
‘gates’ to produce improved prediction performance over the constructed time series
model. The proposed approach has been improvised with an overall F-measure of
96.8% in black lives matter and 95.47% in demonetization.
Knowledge Driven Paradigm for Anomaly Detection … 147

Organization: The flow of the rest of the paper is as follows -: Sect. 2 consists
of related work to the field of study and experimentation. Section 3 drafts the
proposed architecture for the system of anomaly detection. Section 4 consists of
the implementation and experimental observation. Section 5 includes conclusion
and references.

2 Related Work

Sujatha et al. [3] has elaborately explained about various OSM anomalies. Behavioral
anomalies have been observed over a twitter data set on demonetization policy using
Self-Supervised learning. 86% of the public opinion was seen to support demoneti-
zation. Guo et al. [4] have put forth a Gated Recurrent Unit based Gaussian Mixture
VAE system for detection of outliers on n-dimensional time-series data called GGM-
VAE and they claim that their scheme exceeds in performance when compared to
the state-of-the-art abnormality detection model and has produced up to 5.7, 7.2%
increased accuracy accuracy and F1 score respectively, in comparison to existing
schemes. Liu et al. [5] have put together a literature review on the anomaly detection
methodologies based on social media which includes an overview of the data types
and social media data attributes, network data, temporal data anomaly detection and
also includes description of state-of-the-art technique on massive social media data.
Cho et al. [6] proposed a neural network model called RNN Encoder–Decoder that
contains a double RNN. Both were jointly trained to maximize the conditional prob-
ability of a sequence target. This paper denotes the origin of GRU. Watson et al. [7]
developed an automated outlier detection model built over supervised Long Short-
Term Memory [LSTM] and statistical analysis. Zhang et al. [8] used deep learning
methodologies to social media data for detection of traffic accidents. Deep Belief
Network (DBN) and LSTM were implemented on the extracted features and DBN
showcased an accuracy of 85%. Results displayed superiority of DBN over Support
Vector Machines (SVM) and supervised Latent Dirichlet allocation (sLDA). In [9–
20] several semantic and ontological approaches in support of the proposed work
has been discussed.

3 Proposed System Architecture

The proposed architecture as shown in Fig. 1 consists of three notable segments. The
first segment includes data pre-processing and classification using GRU. The second
segment involves the formulation of a domain ontology using an RDF knowledge
base that has been filtered out of heterogeneous fact resources. The third segment
compares the knowledge base to the tweets using an integrated cosine similarity
and Web PMI module. The very aim of this system design is to verify that the key
148 S. Manaswini et al.

indices derived from the RDF knowledge base, drawn into an ontology, is compared
to the tweets classified by GRU in order to point out an outlier in the extracted data
sequence. The first phase of the process involves scraping of the tweets through
‘Twint’, an API that facilitates scrapping of twitter.
Data pre-processing is a very vital step as the data extracted needs to be trans-
formed into a workable format. Twitter data consists of multiple symbols like ‘@’
‘#’, abbreviations, etc. that are difficult for the machine to contextually comprehend
in language processing applications. Therefore, all text was converted to lowercase
and tokenization was performed on the sequence of strings. Stemming and stop words
were removed and lemmatization is carried out to retain words but at the same time
preserve contextual meaning. On completion of the pre-processing of tweets, it is
passed into the Term Frequency- Inverse Document Frequency [TF-IDF] [21].
Prediction models are highly dependent on the feature selection process. In text
classification, features are a subset of the words in a given string. In order to classify
the ‘importance’ of a word in a document TF-IDF is used. An anomaly can only
be arbitrarily defined. In order to capture abnormal behaviour the data sequence
first needs to be in the form of a computable input in order to be analysed. TF-
IDF is computed using the product of ‘the number of occurrences of a term’ and
the ‘inverse document frequency of a word across a set of documents’. In common
terms the latter computes how less valuable a term is in the set of documents. It is
seen to approach 0 when a specific term is seen to occur frequently and otherwise
approaches 1. Equations (2) and (3) shows how Eq. (1), the final TF-IDF equation
is derived.

TF-IDF = Term Frequency × Inverse DocumentFrequency (1)

Fig. 1 System architecture

Knowledge Driven Paradigm for Anomaly Detection … 149

Nt
TF(t, d) = (2)
Nd
ND
IDF(t, D) = loge (3)
Dt
The next step is to test and classify the pre-processed tweets and dataset using
the GRU [4] classifier. The output of the classifier would be the top 20–25% of
the categorized tweet data words based on the term classes initially derived from
tweets. Further categorization, classification and comparison will be based on the top
categories. The gated recurrent neural network is an improvised RNN architecture.
It consists of two vectors or gates- reset gate and update gate. These two gates are
responsible for the regulation of information that passes on to the output gate. GRU
has been chosen as it conveniently addresses the vanishing gradient problem or the
issue of retention of necessary information in more common terms. The update gate
decides how much of the information from the preceding steps are carried onto to the
succeeding steps. The reset gate decides how much of the information is eliminated
or forgotten. The main activation functions involved in the adopted model of GRU
includes Sigmoid and tanh.
Assuming a new input ‘x’ is inserted to the network at time ‘t’, and a variable ‘y’
stores information until t-1 units, then the product of ‘x’ and its weight is summed
with the product of y and its weight and the summation is passed into the sigmoid
function which acts as the activation function and is used to contain the value between
0 and 1. Similar to the process of the update gate (1), reset gate has the exact same
procedure but the way in which the weights and gates will be used varies in terms of
function and design. When the important part of information is nearing the reset gate
is assigned a value close to 0 in order to forget the unnecessary information from the
traversal until the current point.
The current memory content is then determined by the product of the input ‘x’
and its weight that is summed with the Hadamard product of the reset gate and ‘U.y’
passed into the nonlinear tanh activation function. The final memory at time step
‘t’ is determined by the Hadamard product of the update gate and ‘y’ summed with
the product of (1−z) and the calculated current memory content ‘Ct’ and is passed
through tanh function. From (6) the final memory content at step ‘t’ is given in (7).
Equation (6) makes use of (5), Eq. (7) makes use of (66) and (4).

Update gate(zt ) = σ(x.Wz + y.Uz ) (4)

Reset gate(Rt ) = σ(x.Wr + y.Ur ) (5)

 
Current memory content C1t = tanh(x.W + Rt  U.y) (6)

Final memory content(Ct ) = tanh(Zt  y + (1 − Zt )C1t (7)

150 S. Manaswini et al.

The second phase involves various fact sources like prevalent news daily, Google
news API, telegraph, RSS news feed and other verified resources that have been
crawled over a specified time line and tokenized to collect facts or specific terms
required for the formulation of the knowledge base. The RDF knowledge base [22]
lays on the Subject-Predicate-Object format. The source of facts is the predicate, fact
pretext is the subject and the fact context and summarized facts form the object. From
the heterogeneous sources, all the ‘keywords’ are crawled and gathered to form an
indexed fact base that leads to depiction of an in-depth domain ontology containing
all the indexed terms from the fact base. This is then compared to the categorized
tweets using the cosine similarity combined with Web PMI.
The third phase helps compare the categorized terms with the domain ontology
to help in the prediction of anomalies. The similarity function implemented is the
cosine similarity [23] along with Web PMI. Irrespective of the size of the document
cosine similarity is used to compare them. Cosine similarity is the cosine of the angle
between two vectors in a multi-dimensional space. The smaller the angle the more
the similarity.

0, H (A ∩ B) ≤ C
Web PMI (A, B) =
log2 H (A)( −NH (B)
H A∩B ) (8)
, other wise
N N

Web PMI just like every other PMI is a metric to measure the correlation between
two terms. It is observed to be directly proportional to the frequency of occurrence of
the two terms or their similarity and is defined between 0 and 1. Increasing similarity
implies that the value tends towards 1. For the purpose of anomaly prediction, the
terms from the categories classified by the GRU and those indices from the domain
ontology are compared. Because we aim on prediction of the outliers, we set the
threshold to 0.25, and any value <0.25 depicts an anomaly and >0.25 implies its
commonality in occurrence. Web PMI can be numerically defined as shown in Eq. (8),
where, A and B are the terms for comparison and N is the total number of occurrence
of synonymous term pairs and page count of the term-pair along with the page count
for individual terms is denoted by H.

4 Implementation

For the detection of anomalies the proposed algorithm has been run on a Windows
10 Operating System supported by an Intel Core i7 8th gen processor with 16 GB
RAM. The code has been trained over python version 3.9.0 using Google Collab-
oratory, which has been stacked on multiple GPUs and hosted by Jupyter note-
book. The library files included for the implementation of the proposed design
are textblob, pandas, mathplotlib, scikit learn, and NLTK. Two events have been
considered for the experiment—Black lives matter and Demonetization. Five other
Knowledge Driven Paradigm for Anomaly Detection … 151

training approaches apart from the proposed system have been adopted for obser-
vational purposes namely, self supervised learning, SVM, logistic regression, and
RNN. Precision (%), Recall (%), Accuracy (%), F-measure (%) and False Negative
Rate (FNR) has been calculated from the selected dataset.
From Fig. 2, it is evident that the proposed model has achieved an accuracy of
96.18% showing an increase of 6.04% from the adopted baseline, self supervised
learning model. The experimental precision is observed to be 98.89%, denoting
a 6.71% increase. In consideration of the baseline model [1], even though self-
supervised models have shown improved accuracy when compared to most of the
conventional unsupervised models, labelling of data is still an issue and has to be
manually annotated. Random Forest produces an accuracy of 86.43% and has a
precision of 87.15%, which clearly denotes an under-fitting model. This could be
because of the absence of discrete labels, large data sets and large memory space and
processing time that is caused by the complexity of the ensemble of the decision trees
that tests the significance of the individual variables, leading to undesired accuracy
and precision. Similarly, SVM does not perform well with large datasets and is seen
to underperform in cases that have overlapping target classes and therefore shows an
accuracy of 84.12% and a precision of 86.77%.
The Linear Regression model is seen to have an accuracy of 91.19% and a preci-
sion of 93.89%. Though there is a significant increase when compared to Random
Forest, this model assumes a linear relationship between dependent and independent
attributes that highly affects the outliers as the boundaries are tight and it might falter
in an inconsistent dataset, therefore the recall rate dipping to 86.76%. RNN shows
an accuracy rate of 91.79% but has a precision of 94.33%. One of the main prob-
lems of a standard RNN is the retention of information through the succeeding steps
and computation could be slow but with GRU the vanishing gradient problem stands
resolved and further integration with cosine similarity and Web PMI improves perfor-
mance hence proving to be the best suitable model in comparison. Similarly, for the
event demonetization as shown in Fig. 3, the accuracy measure of the proposed model

Black lives maer

Proposed Knowledge Centric Approach… 98.89 94.81 96.18 96.8

RNN 94.33 89.41 91.79 91.8

Logisc Regression 93.89 86.76 91.19 90.18

SVM 86.77 82.12 84.12 84.38

Random Forest 87.15 85.33 86.43 85.22

Self-Supervised Learning [Basepaper] 92.18 89.32 90.14 90.72

0 50 100 150 200 250 300 350 400 450

Precision % Recall % Accuracy % F-Measure %

Fig. 2 Performance comparison of the proposed approach for black lives matter
152 S. Manaswini et al.

Demonezaon
Proposed Knowledge Centric Approach… 96.15 94.81 95.43 95.47
RNN 89.12 92.44 91.14 90.75
LogiscRegression 88.16 90.12 90.01 89.13
SVM 80.81 82.18 81.18 81.49
Random Forest 84.33 87.43 85.81 85.85
Self-Supervised Learning [Basepaper] 86.69 90.43 88.11 88.52

0 50 100 150 200 250 300 350 400 450

Precision Recall Accuracy F-Measure

Fig. 3 Performance comparison of the proposed approach for demonetization

FNR:BLACK LIVES vs DEMONETISATION

Proposed Knowledge Centric Approach… 0.05190.0519
RNN 0.1059 0.075
LogiscRegression 0.1324 0.0988
SVM 0.1782 0.1782
Random Forest 0.1467 0.1257
Self-Supervised Learning [Basepaper] 0.1068 0.0957

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

FNR1 FNR2

Fig. 4 FNR—Black lives matter Vs Demonetization

is observed to be 95.43% and precision to be 96.15% indicating improved perfor-

mance when compared to the other self-supervised and unsupervised methodologies.
Figure 4 depicts the FNR of both the events.

5 Conclusions

On an experimental basis it is interesting to know the scope of outliers in a given

dataset. Especially, when real life events of influence are taken into consideration the
identification of anomalies can help serve various political causes such as identifying
protests and public opinions. The GRU classifier in the system design has proved more
effective than most of the unsupervised models exhibiting an accuracy of 96.18%
in the event ‘black lives matter’ and 95.43% in the event ‘demonetization’. The
proposed model shows enhancement in detecting anomaly by the inclusion of a
cosine similarity and Web PMI that determines the dissimilarity providing us with
improved precision in the results of prediction.
Knowledge Driven Paradigm for Anomaly Detection … 153

References

1. Deepak G, Santhanavijayan A (2020) OntoBestFit: a Best-Fit Occurrence Estimation strategy

for RDF driven faceted semantic search. Comput Commun 160:284–298
2. Pushpa CN, Deepak G, Kumar A, Thriveni J, Venugopal KR (2020) OntoDisco: improving web
service discovery by hybridization of ontology focused concept clustering and interface seman-
tics. In: 2020 IEEE international conference on electronics, computing and communication
technologies (CONECCT). IEEE, July 2020, pp 1–5
3. Kokatnoor SA, Krishnan B (2020) Self-supervised learning based anomaly detection in online
social media. Int J Intell Eng Syst (INASS)
4. Guo Y, Liao W, Wang Q, Yu L, Ji T, Li P (2018) Multidimensional time series anomaly
detection: a GRU-based gaussian mixture variational auto encoder approach. Asian conference
on machine learning (ACML) (2018)
5. Liu Y, Chawala S (2017) Social media anomaly detection: challenges and solutions. In: 10th
ACM international conference (2017)
6. Cho K, van Merrienboer B, Gulcehre C, Bahdanau D, Bougares F, Schwenk H, Bengio Y
(2014) Learning phrase representations using RNN encoder–decoder for statistical machine
translation, EMNLP
7. Jia W, Shukla RM, Sengupta S (2019) Anomaly detection using supervised learning and
multiple statistical methods. In: 18th IEEE international conference on machine learning and
applications (ICMLA)
8. Zhanga Z, He Q, Gaod J, Nic M (2017) A deep learning approach for detecting traffic accidents
from social media data
9. Kumar A, Deepak G, Santhanavijayan A (2020) HeTOnto: a novel approach for conceptualiza-
tion, modeling, visualization, and formalization of domain centric ontologies for heat transfer.
In: 2020 IEEE International Conference on Electronics, Computing and Communication
Technologies (CONECCT). IEEE, July 2020, pp 1–6
10. Deepak G, Kasaraneni D (2019) OntoCommerce: an ontology focused semantic framework
for personalised product recommendation for user targeted e-commerce. Int J Comput Aided
Eng Technol 11(4–5):449–466
11. Gulzar Z, Anny Leema A, Deepak G (2018) Pcrs: personalized course recommender system
based on hybrid approach. Procedia Comput Sci 125(2018):518–524
12. Deepak G, Teja V, Santhanavijayan A (2020) A novel firefly driven scheme for resume parsing
and matching based on entity linking paradigm. J Discrete Math Sci Cryptograp 23(1):157–165
13. Haribabu S, Sai Kumar PS, Padhy S, Deepak G, Santhanavijayan A, N. Kumar D (2019) A
novel approach for ontology focused inter- domain personalized search based on semantic set
expansion. In: 2019 Fifteenth International Conference on Information Processing (ICINPRO),
Bengaluru, India, pp 1–5. https://doi.org/10.1109/ICInPro47689.2019.9092155
14. Deepak G, Kumar N, VSN Sai Yashaswea Bharadwaj G, Santhanavijayan A (2019) OntoQuest:
an ontological strategy for automatic question generation for e-assessment using static and
dynamic knowledge. In: 2019 fifteenth international conference on information processing
(ICINPRO). IEEE, pp 1–6
15. Santhanavijayan A, Kumar DN, Deepak G. A semantic-aware strategy for automatic speech
recognition incorporating deep learning models. In: Intelligent system design. Springer,
Singapore, pp 247–254
16. Deepak G, et al (2019) Design and evaluation of conceptual ontologies for electrochemistry as
a domain. In: 2019 IEEE international WIE conference on electrical and computer engineering
(WIECON-ECE). IEEE (2019)
17. Deepak G, Priyadarshini JS (2018) Personalized and Enhanced Hybridized Semantic Algorithm
for web image retrieval incorporating ontology classification, strategic query expansion, and
content-based analysis. Comput Electr Eng 72:14–25
18. Deepak G, Priyadarshini JS, Babu MH (2016). A differential semantic algorithm for query
relevant web page recommendation. In: 2016 IEEE International Conference on Advances in
Computer Applications (ICACA). IEEE, October 2016, pp 44–49
154 S. Manaswini et al.

19. Kaushik IS, Deepak G, Santhanavijayan A (2020) QuantQueryEXP: a novel strategic approach
for query expansion based on quantum computing principles. J Discrete Math Sci Cryptography
23(2):573–584
20. Santhanavijayan A, Kumar DN, Deepak G. A novel hybridized strategy for machine translation
of indian languages. Soft Comput Signal Process 363
21. Qaiser S, Ali R (2018) Text mining: use of TF-IDF to examine the relevance of words to
documents. Int J Comput Appl (0975 – 8887) 181(1)
22. Stoermer H, Bouquet P, Palmisano I, Redavid D (2017) A context-based architecture for RDF
knowledge bases: approach, implementation and preliminary results. In: Springer international
conference on web reasoning and rule systems
23. Otsuka Y (1936) The faunal character of the Japanese Pleistocene marine Mollusca, as evidence
of the climate having become colder during the Pleistocene in Japan. Bull Biogeograph Soc
Jpn 6(16):165–170. AGCJ.
Stacked Deep Learning LSTM Model
for Daily Solar Power Time Series
Forecasting

Soufiane Gaizen, Ouafia Fadi , and Ahmed Abbou

Abstract Several researches have been concentrated on predicting issues in many

appl cation areas. In machine learning field, recurrent neural networks have been
successfully employed. Such models were suggested to tackle time- dependent
learning problems. The goal of this paper is to afford a credible forecasting method-
ology for solar power, which will reduce the inaccuracies involved in forecasting the
near future generation. In this proposed methodology, the stacked LSTM models with
different configurations are applied to track the dynamic performance of the solar
energy. The second objective of this analysis is to assess the effect of the stacked
LSTM hidden layers by comparing the predicted outcomes of each model. We eval-
uate the proposed approach with three forecasting algorithms. The simulation results
show that the value of the MAPE, using stacked LSTM model, starts from 10.5%
(one hidden layer) and decreases to 0.442% when four hidden layers are applied.
Thus, the use of Stacked-LSTM with multiple layers provides a further decrease in
the fore- casting error relative to other approaches.

Keywords Solar power · Prediction · Machine learning · RNN · LSTM ·

Stacked-LSTM

1 Introduction

Nowadays, solar power is the most important sources for electrical power generation
[1, 2]. Solar photovoltaic produces energy from sunlight by converting solar radi-
ance into electricity [3]. The efficiency of PV installations is influenced by climate
parameters such as temperature [4]. photovoltaic energy relies on weather condi-
tions and solar irradiance, which add to photovoltaic generation heterogeneity and
instability. As stated by the IEA (International Energy Agency), The total power of
renewable electricity is projected to increase by more than 1TW during the period
starting from 2018 to 2023 [5]. Solar photovoltaic constitutes more than ½ of this

S. Gaizen (B) · O. Fadi · A. Abbou

Mohammadia School of Engineers, Mohammed 5 University, P.B. 765, Rabat, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 155
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_15
156 S. Gaizen et al.

extension and dominates the increase of renewable capacity. Taking into account the
ongoing development of solar PV sector, the electrical grid is facing the challenge of
the aforementioned heterogeneity and instability in power production. It exists many
alternatives to resolve these problems as energy storage installations to mitigate the
discontinuity of photovoltaic production [6]. Moreover, knowing how much photo-
voltaic energy would be generated, may extremely decrease power plants operating
costs and increase the resilience of the grid, thus, predicting PV power production is
a challenging task because it relies considerably on external circumstances like PV
radiation and climate conditions.
Solar power prediction strategies are primarily parted into 3 groups in the liter-
ature: physical, statistical, and machine learning configurations [7–15]. Machine
learning configurations are capable of managing complex problems [16]. It contains
3 structures: mentioned in these references [17, 18], one of these structures is the
Artificial Neural Network [18]. This structure is currently the most widely employed
strategies for time series prediction. Most traditional PV power prediction models are
confined to expose the correlation between the data, but these models are not capable
of discovering a deep correlation, uncovering hidden pattern and reveal significant
data. For the tremendous data from the recent generating stations, the application of
traditional methods isn’t conducive to ensuring reliable forecasting. Over a decade,
deep learning strategies have appeared as effective machine learning technology
[19]. They are becoming increasingly prevalent, thanks to their strong ability to
explain dependencies in time series data. Several deep learning configurations have
been recommended in these references [20–22], One of these configurations, is the
Recurrent Neural Networks [22] that takes advantage of the sequential nature of
input data. RNNs are employed to model time-dependent in timeseries data. In the
time series forecasting, RNNs provide good results, which have been proven in
several researches [23–25]. The LSTM (Long Short-Term Memory) networks is a
kind of RNNs that can accommodate information retrieval for longer time periods
[26]. Furthermore, it is largely employed for time series forecasting applications,
which is ideally tailored to problems with solar PV generation and wind power
forecasting [27]. Long short-term memory permits the memorization of the weights
that are forward and back-propagated across layers [28]. LSTM-based RNNs are an
appealing alternative for modeling consecutive data such as time series since they
integrate qualitative knowledge from past inputs. LST memory has promising oper-
ations in solar power forecasting as a novel machine learning algorithm [6, 29]. By
preserving the structure of memory cell, LSTM could sustain essential features that
should be recognized throughout the learning procedure. Thus, applying LSTM to
forecast solar power permits not only to achieve the correlation over uninterrupted
hours but also to derive its long-term behavior patterns [6]. To forecast photovoltaic
power, Abdel-Nasser [30] suggests the application of LST memory to reliably esti-
mate the generated power for solar systems. The results have exposed that the use
of LSTM provides a decrease in the forecasting error relative to other techniques.
Qing [6] suggested a new method of forecasting hourly day-to-day solar power
using LSTM-based weather forecasting. The suggested algorithm utilizes the hourly
weather as the intrant parameter, and the hourly solar power values of the similar
Stacked Deep Learning LSTM Model … 157

scheduled day as the output parameter. Due to the assumption that time dependency
is considered, the results reveal that the suggested learning algorithm is more precise
than other three models. Srivastava [29] analyzed the ability of LSTM to forecast
solar power time series, his study demonstrated the robustness of LSTM, also this
model outperforms GBR and FFNN for day-to-day solar power time series fore-
casting. Liu [31] suggested a new deep learning approach using wavelet analysis to
forecast wind speed. LST memory is applied to forecast the low frequencies of sub-
layer, and the high frequencies are forecasted by Elman-neural-network. The mixed
configuration has satisfactory performance compared to 11 traditional approaches.
Under complex weather conditions, Yuet [32] suggested an LST memory-based
approach to forecast short-term global horizontal irradiance, the study demonstrated
that LST memory surpasses ARIMA, SVR, and NN configurations, especially on
cloudy weather. Despite the mentioned LST memory techniques don’t thoroughly
analyzed experimentally the impact of various variables and structures, but these vari-
ables would influence the precision of prediction. Our article is provided to assess
the impact of the depth of the LSTM networks namely Stacked-LSTM networks.
The original time sets are separated into training and testing series. In addition, all
the outcomes of the forecasts obtained are applied to the persistence comparison
model for a proper benchmark evaluation. As a result, the Stacked-LST Memory
configuration with multiple hidden layers show high performances in the daily solar
energy forecasting.

2 Predictive Patterns

2.1 Database

The solar radiation data used in this work are from the NASA prediction of worldwide
energy resource website for the region of Fes/Morocco. The solar radiation data is
recorded within the date range from 01/01/2017 to 01/01/2019. Every 30 min, the
solar power data is registered. this recorded value was converted to daily average
data. Averaging these values would make the signal smoother, and the algorithms
are easier to learn. We observed that the forecast of solar radiation on a daily average
basis is more reliable than the prediction of solar radiation of 30 min ahead. The
data on solar radiation is standardized and the outliers and incomplete values are
excluded.

2.2 Recurrent Neural Network Pattern

For series prediction issues, RNNs are the most widely used Neural Network archi-
tecture. It also achieved particular attention in the natural language process domain.
158 S. Gaizen et al.

Fig. 1 Basic example for RNN.

Alike to Artificial neural networks, RNNs are considered universal approximators

[33]. Unlike ANNs, the feedback loops of the recurrent neural network implicitly
handle the temporal order, as well as the time dependency of the sequences [33]. The
RNN is a type of ANNs, where the links between nodes creates a directed graph along
a temporal sequence. And this permits it to display a temporal dynamic attitude.
Figure 1 illustrates the fundamental concept of the RNNs. Xt represents the intrant
layer at the time-step ‘t’, ht is the hidden state at ‘t’. It should be remembered that
the hidden state is often referred to be the network memory. It is determined on the
basis of the previously hidden state and the intrant layer at the same step.
Equation 1 [33] represents a general formula to determine the hidden state ht .

0, t =0
ht = (1)
δ(τxt , xt ), other wise

δ is a non-linear function. The actualized hidden state is achieved as follows:

h t = F(τxt , uh t−1 ) (2)

F is a “tanh” function (hyperbolic tangent). Generally, when using recurrent

neural networks patterns, it is not straight forward to catch long-term time-series
dependencies. To overcome this restriction, LST Memory Networks models were
designed.

2.2.1 LST Memory Recurrent Neural Network Pattern

The LST Memory Neural Network was suggested by [34] to prevent long-term
dependency difficulties through a targeted architecture. Unlike typical NNs, LSTM
hold a memory block to store temporary input data information.
Figure 2 illustrates the hidden layer of the LSTM, that is recognized as the LST
Memory cell.
Applying a gate structure, the LST Memory cell binds the intrant layer Xt , to the
output layer ht . There are three gates in the LST Memory cell: Intrant gate It , Output
gate Ot , and Forget gate ft . The input cell state Čt , the current cell output state Ct,
and the output state of the previous irritation, are used in the training procedure.
Using the relations given by [30], the input state, the output state and the gates can
be defined. The architecture of the LSTM can be defined as follows [30]:
Stacked Deep Learning LSTM Model … 159

Fig. 2. Architecture of
LSTM

   
f t = σ w f × h t−1 , xt + b f (3)

   
i t = σ wi × h t−1 , xt + bi (4)

   
ct = tanh wc × h t−1 , xt + bc (5)

ct = f t × ct−1 + i t × ct (6)

   
ot = σ wo × h t−1 , xt + bo (7)

h t = ot × tanh(ct ) (8)

Where wf , wi , wc and wo are weight matrices. bf , bi , bc and bo present bias vectors.

Čt is a current state that is given by xt and ht−1 via the tanh layer. σ () is the function
of activation.

2.3 Evaluation Metrics

Three error metrics were suggested to measure the efficiency of the predicting model.
The Root mean square error calculates the discrepancy between the real and the fore-
casted variables. Lower RMSE indicates better forecasting outcome, and is described
as [35]:


1
N
 2
RMSE = X f − Xa (9)
N i=1

However, the Mean Absolute Percentage Error (MAPE) represents the ratio of
error to the real value [35].
 
1
X f − Xa
N
M AP E = (10)
N i=1 |X a |
160 S. Gaizen et al.

Where |X a | presents the mean of the actual solar power data.

R-Square (R2 ): R2 measures the proportion of the total variance described by the
fitting regression model, R2 is included in the interval [0, 1]. If R2 equal to 0, the
configuration fits weakly; if R2 equal to 1, the configuration holds zero errors. R2 is
determined as follow [35]:
N  2
i=1 X f − Xa
R =1−
2
N  2 (11)
i=1 X a − Xa

3 Proposed Approach

The suggested approach is designed to forecast solar power time series using deep
stacked LSTM. Figure 3 outlines the essential steps of the suggested approach.
The suggested approach comprises four primary points:
1. Collect Solar power data from the NASA prediction of universal energy resource
website for the region of Fes/Morocco. The solar power series are assembled
into training and testing period.

Start Data collection

Data Pre-processing

Data standardization

Training and valida- Improve the perfor-

tion data mance and the depth
of the model.

Test Data Build and train LSTM

Model

Good Model No

yes

Test and validate

Model

End

Fig. 3 Flowchart of the proposed forecasting method

Stacked Deep Learning LSTM Model … 161

2. Pre-process and clean up the data by eliminating outliers and imputing incom-
plete values. There are several essential steps in the preprocessing of data, such
as data cleaning, data conversion, and the selection. Data cleaning and conver-
sion are techniques used to eliminate outliers and standardize data in such a way
that they take a shape that could be conveniently used to construct a model.
3. Normalize the solar power initial data.
4. Train, check and verify the Stacked-LST memory configuration. Several statis-
tical metrics are employed to measure the precision of the configuration. If the
precision is not good, the stacked LSTM layer number is increasing to assess
the performance and the depth of the stacked-LST memory model.

4 Results and Discussion

The future PV power is derived from the historical solar time series using the Stacked-
LST memory configuration. MAPE, RMSE and R2 are the performance assessors
for the suggested technique. In this analysis, 80% of data is applied to adjust the
configurations and the remainder is applied for assessing purposes. In addition, the
efficacy of the evolved configuration has been matched to the persistence reference
configuration. The persistence reference configuration efficiency in the daily solar
forecast was described in Fig. 4 as 1.027 for the RMSE, 27.7% for the MAPE and
0.668 for the R2 [35].
Using the stacked LST Memory configuration with four hidden layers, Fig. 5
shows the predicted solar energy values. As seen, as the depth of learning is increased,
it is apparent that a strong match between the expected and real value exists. For the
stacked LST Memory configuration, the precision of the forecast was increased by
98.7% over the persistence reference configuration.
Figure 6 demonstrates the scattering plot of the expected solar energy using the
Stacked LST Memory configuration over the real values. As shown in this figure,
the traced data points generally correlated closer towards the fitting line.

Fig. 4 Forecasted solar energy values of the Persistence Reference configuration vs. the initial
dataset [35]
162 S. Gaizen et al.

Fig. 5 Stacked LST Memory forecasting performance with 4 hidden layers

Fig. 6 Plot scattering of the expected results of the Stacked LST Memory configuration with four
hidden layers compared to the initial test data.

Table 1 Solar energy

N˚ of hidden layer RMSE MAPE R2
forecasting outcomes of the
Stacked LST Memory per day Persistence model 1.027 27.7% 0.668
ANN 0.24 6.8% 0.943
SARIMA 0.29 7.4% 0.935
1 0.38 10.5% 0.921
2 0.16 4.42% 0.956
3 0.09 2.486% 0.979
4 0.016 0.442% 0.997

Table 1 displays the daily solar energy forecast results for the Stacked LST
Memory configuration. When evaluating the error values in this table, the RMSE
of 0.38, the MAPE of 10.5% and an R2 of 0.921 were found for the LSTM model
with one hidden layer. RMSE of 0.16, MAPE of 4.42% and R2 of 0.956 were LSTM
model with two hidden layers. RMSE of 0.09, MAPE of 2.486% and R2 of 0.979
were found for LSTM model with three hidden layers. RMSE of 0.016, MAPE of
0.442% and R2 of 0.997 were found for LSTM configuration with four hidden layers.
Stacked Deep Learning LSTM Model … 163

5 Conclusion

Our work provides evaluation analysis of various hybrid configuration to predict solar
energy. The precision of the forecasting is highly enhanced over to the persistence-
reference configuration. Our analysis reveals that the inclusion of hidden layers of the
LST Memory configuration improves the precision of prediction. This is illustrated
by the MAPE value beginning from 10.5% (one hidden layer) and dropping to 0.442%
when four hidden layers are included. The suggested analysis is intended to alleviate
the core problem of PV energy forecasting.

References

1. Sobri S, Koohi-Kamali S, Rahim NA (2018) Solar photovoltaic generation forecasting methods:

a review. Energy Conv Manage 156:459–497 (2018)
2. Harrou F, Taghezouit B, Sun Y (2019) Improved $ k $ NN-based monitoring schemes for
detecting faults in PV systems. IEEE J Photovolt 9(3):811–821
3. Lianos P (2017) Review of recent trends in photoelectrocatalytic conversion of solar energy to
electricity and hydrogen. Appl Catal B 210:235–254
4. Gürtürk M, Benli H, Ertürk NK (2020) Determination of the effects of temperature changes
on solar glass used in photovoltaic modules. Renew Energy 145:711–724
5. Renewables 2018. Power. 8 October 2018. https://www.iea.org/renewables2018/power/.
Accessed 16 Sept 2020
6. Qing X, Niu Y (2018) Hourly day-ahead solar irradiance prediction using weather forecasts
by LSTM. Energy 148:461–468
7. Sobri S, Koohi-Kamali S, Rahim NA (2018) Solar photovoltaic generation forecasting methods:
a review. Energy Conv Manage 156
8. Lopes FM, et al (2018) Short-term forecasts of GHI and DNI for solar energy systems operation:
assessment of the ECMWF integrated forecasting system in southern Portugal. Solar Energy
170:14–30 (2018)
9. Alonso-Montesinos J, Batlles FJ, Portillo C (2015) Solar irradiance forecasting at one-minute
intervals for different sky conditions using sky camera images. Energy Convers Manage
105:1166–1177
10. Miller SD, et al (2018) Short-term solar irradiance forecasting via satellite/model coupling.
Solar Energy 168:102–117 (2018)
11. Halabi LM, Mekhilef S, Hossain M (2018) Performance evaluation of hybrid adaptive neuro-
fuzzy inference system models for predicting monthly global solar radiation. Appl Energy
213:247–261
12. Hou W, Xiao J, Niu LY (2016) Analysis of power generation capacity of photovoltaic power
generation system in electric vehicle charging station. Electr Eng 4:53–58 (2016)
13. Miao S, et al (2018) Markov Chain model for solar farm generation and its application to
generation performance evaluation. J Clean Prod 186:905–917 (2018)
14. Agoua XG, Girard R, Kariniotakis G (2017) Short-term spatiotemporal forecasting of
photovoltaic power production. IEEE Trans Sustain Energy 9(2):538–546 (2017)
15. Massidda L, Marrocu M (2017) Use of multilinear adaptive regression splines and numerical
weather prediction to forecast the power output of a PV plant in Borkum, Germany. Sol Energy
146:141–149
16. Hossain M, et al (2017) Application of extreme learning machine for short term output power
forecasting of three grid-connected PV systems. J Clean Prod 167:395–405
164 S. Gaizen et al.

17. Lin G-Q, et al (2020) An improved moth-flame optimization algorithm for support vector
machine prediction of photovoltaic power generation. J Cleaner Prod 253:119966
18. Lu H, Ma X (2020) Hybrid decision tree-based machine learning models for short-term water
quality prediction. Chemosphere 249:126169
19. Hinton GE, Osindero S, Teh Y-W (2006) A fast learning algorithm for deep belief nets. Neural
Comput 18(7):1527–1554
20. Ogawa S, Mori H (2020) PV output forecasting by deep Boltzmann machines with SS-PPBSO.
Electrical Eng Jpn
21. Zang H, et al (2020) Application of functional deep belief network for estimating daily global
solar radiation: a case study in China. Energy 191:116502 (2020)
22. Jung Y, et al (2020) Long short-term memory recurrent neural network for modeling temporal
patterns in long-term power forecasting for solar PV facilities: Case study of South Korea. J
Cleaner Prod 250:119476
23. Kumar D, et al (2020) Forecasting of solar and wind power using LSTM RNN for load frequency
control in isolated microgrid. Int J Model Simul 1–13
24. Park N, Ahn HK (2019) Multi-layer RNN-based short-term photovoltaic power forecasting
using IoT dataset. In: 2019 AEIT International Annual Conference (AEIT). IEEE
25. Srivastava T, Vedanshu, Tripathi MM (2020) Predictive analysis of RNN, GBM and LSTM
network for short-term wind power forecasting. J Stat Manage Syst 23(1):33–47 (2020)
26. Zhao F, et al (2020) Time-sequenced flow field prediction in an optical spark-ignition direct-
injection engine using bidirectional recurrent neural network (bi-RNN) with long short-term
memory. Appl Thermal Eng 115253
27. Liu S (2019) Wind/solar power forecasting using improved LSTM neural networks. Diss
28. Hochreiter S, Schmidhuber J (1996) LSTM can solve hard long-time lag problems. Adv Neural
Inf Process Syst 9:473–479
29. Srivastava S, Lessmann S (2018) A comparative study of LSTM neural networks in forecasting
day-ahead global horizontal irradiance with satellite data. Sol Energy 162:232–247
30. Abdel-Nasser M, Mahmoud K (2019) Accurate photovoltaic power forecasting models using
deep LSTM-RNN. Neural Comput Appl 31(7):2727–2740
31. Liu H, Mi X-W, Li Y-F (2018) Wind speed forecasting method based on deep learning strategy
using empirical wavelet transform, long short-term memory neural network and Elman neural
network. Energy Convers Manage 156:498–514
32. Yu Y, Cao J, Zhu J (2019) An LSTM short-term solar irradiance forecasting under complicated
weather conditions. IEEE Access 7:145651–145666
33. Schäfer AM, Zimmermann HG (2006) Recurrent neural networks are universal approxima-
tors. In: International conference on artificial neural networks. Springer, Berlin, Heidelberg,
September 2006, pp 632–640
34. Hochreiter S (1997) JA1 4 rgen Schmidhuber. Long Short-Term Memory. Neural Computat
9(8)
35. Gaizen S, Fadi O, Abbou A (2020) Solar power time series prediction using wavelet analysis.
Int J Renew Energy Res (IJRER) 10(4):1764–1773
A Comparison Study of Machine
Learning Methods for Energy
Consumption Forecasting in Industry

Mouad Bahij, Moussa Labbadi, Mohamed Cherkaoui, Chakib Chatri,

and Soufian Lakrit

Abstract The rapid expansion in the industrial sector requires good control of
energy consumption, whereas the prediction of the consumption is the most impor-
tant factor for energy management as well as environmental protection. This paper
presents a model for predicting energy consumption using ML machine learning
models such as Linear Regression (LR), Support Vector Machine (SVM), Deci-
sion Tree (DT), Artificial Neural Networks (ANN), and Recurrent Neural Networks
(RNN) are used to predict energy consumption based various attributes as inputs.
In order to evaluate the effectiveness of the proposed approaches, a simulation is
carried out using the Python software. Finally, according to simulation results, the
Linear Regression (LR) method is efficient and give better performances.

Keywords Energy consumption forecasting · Machine learning · SVM · LR ·

DT · ANN · RNN

1 Introduction

Energy forecasting is increasingly important to protect the environment, in order to

ensure the efficiency of the energy management and the reduction of greenhouse
gas emissions. In the context, the energy forecasting used in many areas such as
industrial sectors, buildings, and transport. MOROCCO focuses on enhancing their
energy transition to green growth. This transition based on several factors such as
energy efficiency, which presents an optimal solution in the planning and reducing

M. Bahij (B) · M. Labbadi · M. Cherkaoui · C. Chatri · S. Lakrit

Engineering for Smart and Sustainable Systems Research Center, Mohammed V University,
Mohammadia School of Engineers (EMI), Rabat, Morocco
M. Labbadi
e-mail: [email protected]
M. Cherkaoui
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 165
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_16
166 M. Bahij et al.

costs consumption as well as minimizing energy consumption in the industrial sector

[1].
The Machine Learning presents good results for the time series and regression
problems. In order to improve the energy efficiency, it is necessary to conduct
the statistics, to an analysis of historical energy data, and to predict future energy
consumption [2]. These machine learning algorithms are able to process a large
volume of data in the aim to solve the problem devoted to energy prediction [3].
In industry sector, the evolution of energy is composed according to the following
steps: technical calculation, comparative analysis based on simulation models and
statistical modeling and Machine Learning.
Indeed, the ML process operates in accordance with the principle of the black
box, which find the relationship between the various input data characteristics and
the output objectives using data historical processing. The ML algorithm shows great
potential for the energy modeling and the evaluation of the production in the industrial
sector, for example, the ANN is used for predictions based on a mapping between data
inputs and data outputs in the system [4]. For overcoming the problems of precision
on prediction in the ANN, The RNN shows a better adaptability in the processing of
energy consumption data, the structure of which consists of three layers; the input
layer, and the hidden layer as well as the output layer [5]. The algorithm DT is used
a mechanism to partition the data into groups according to a diagram in order to train
the model [6].
The SVM model presents a robust solution to the problems of the regression and
the classification [7]. The RL is a technique which combines one or several variables
from the historical data base to predict a result according to other variables [8].
This paper proposes a comparative study with five approaches based on Machine
Learning for estimation and analysis energy consumption in industry. The remainder
of this paper is organized as follows. In Sect. 2, the related work is described. In
Sect. 3, the background is developed. The simulation results are discussed in Sect. 4.
Finally, Sect. 5 is devoted to conclusions.

2 Related Work

Generally, the ML algorithm became more used for training prediction models from
existing data compared with conventional methods. This section discusses the papers
linked to this study.
The authors in [9] develop a model for energy consumption forecasting according
to three statistical learning methods included regression trees, RL and multivariate
adaptive regression (MARS).
In paper [10], the authors propose an algorithm based on the ANN for prediction
of consumption load in the building from a set of hourly energy data measures of
consumption of 93 households for 2 months. For the forecasting of energy consump-
tion in the building or industry, multiple linear regression methods, the ANN, and
the genetic algorithms are used to process for data, in order to generate the desirable
A Comparison Study of Machine Learning Methods … 167

prediction model [11]. In article [12], the ANN used to predict the consumption in
the building for the cooling, the heat and the consumption and emission of carbon
dioxide.
In [13], the SVM method is compared with others for a tunnel settlement prediction
problem. This approach is most used in prediction problems [14]. The authors in [6]
used the improved TD as a predictive tool, where she found that this technique had
good results for a small to medium-sized dataset.
The DT presents an interpretable architecture, efficient for solving the problems
the regression and the classification, which the DT does not require a large number
of parameters and the results obtained are easy to in the classification model.
From this brief rated work, a model for energy consumption forecasting in indus-
trial depends on historical data, which comparison between the different algorithms
in order to evaluate the most efficacious methods in this study.

3 Background

3.1 Recurrent Neural Networks

The RNN is a new technique recently used in the processing of time series of energy
consumption, where their structure is developed by the ANN network as it combines
contextual information from previous entries.
According to the memory mechanism the information is transmitted to the network
one by one and the network nodes store their state at a time stage and use it to inform
the next stage [5]. The {x1, x2,…, xn} input sequence is modeled at the RNN level
using Eq. 1 looping:

h i = f (h i−1 , xi ) (1)

According to the three layers are RNN, the calculation of each layer is described
in detail below:
a Input layer
Input layer: The word sequence of the text at the input layer is of the sequence
{x1 , x2 . . . . . . .xn },where the current input vector mapping is {x1∧ , x2∧ . . . . . . .xn∧ }.
b Hidden layer
∧
At the input layer the vector of the sequence xn−1 , as well as h t−1 describe the hidden
state at the previous moment of the output vector. The output of the cache state is
expressed according to Eq. 2:
 ∧

h = σ W hx xn−1 + W hh hn−1 + bh (2)
168 M. Bahij et al.

Where σ means the activation function of each element, and bh presents the vector
bias of the hidden layer.
c Output layer
The output layer compute is based on the current cache layer memory state expression
as input. Equation 3 describes the calculation process:
 
yt = so f tmax U y h t + by ∈ R V 
P(xn /x1 , x2 . . . . . . .xn−1 ) (3)

From where yt presents the probable output displayed, and U y h t defines the weight
matrix between the hidden state and the output layer.

3.2 Artificial Neural Networks

In the ML field, neural networks are widely known as an energy estimation tool in
building and industry. According to their strong learning ability the ANN are more
used to solve the non-linear problem [4].
The ANN composed of the following layers; the input layer contains 4 neurons
reserved for attributes, and the hidden layer is composed of 10 neurons, finally, 1
neuron in the output that expresses the prediction of energy. Figure 1 shows the
Model ANN. The construction of the ANN algorithm according to six steps;
• Preparation of the systems data base.
• Initialization of the weights.
• From the input value of previous neurons calculates the desired output value.
• Improve the output value by acting on the average square root factor between the
output components.
• Data training phase using back-propagation algorithm.
• Forecasting model formulation.

be
Data input

Energy Consumption
forecasting

bs

Output layer

Input layer Hidden layer

Fig. 1 Structure of ANN

A Comparison Study of Machine Learning Methods … 169

3.3 Support Vector Machine

In regression and classification problems, the SVM presents a strong robustness to

solve non-linear system problems with good solutions by a few data samples. The
SVM seeks to minimize the learning error value and a higher level of confidence in
value [2].
In this study, this algorithm is used to predict the energy consumption of an
industrial entity according to a training process based on historical data. The SVM
makes the relation between the inputs of the system according to Eq. 3:

F(xi ) = w, ϕ(xi ) + b

1  l

Minimise = ω2 + C ξ + ξi∗ (4)
2 i=1

where ∗, ∗, w, ϕ(xi), and b ∈ R, describe the dot product, the weight, the nonlinear
mapping function, and the bias.

3.4 Regression Linear

Regression analysis is a traditional technique used to model predictions [7]. A

multiple linear regression model can be expressed according to Eq. 5:

Y = ϕ0 + ϕ1 b1 + ϕ2 b2 + ϕ3 b3 . . . . . . . + ϕ P b p + θ (5)

where the function Y presents the output variable, ϕi for the regression parameters,
bi presents inputs values, and θ is the random error.

3.5 Decision Tree

Decision trees are widely proposed as the optimal solution for classification prob-
lems, or the architecture of the decision tree provide good flexibility, efficiency and
precision for the generation of the prediction model [6].
In fact, their mechanism allows data to be partitioned using a flow chart to predict
desirable consumption. For a large amount of data, the DT is suitable to discover
hidden rules, but in classifiers it may be unstable in relation to the noise present in
the learning data.
When learning the model, the data processing undergone a process has a root node
and 2 branch nodes, in order divided into the other sub-nodes so that the final result
170 M. Bahij et al.

of the DT approach is shows the terminal node. In the paper, this method is used to
predict the outcome from the database provided by following certain decision rules.

4 Simulations Results and Discussion

4.1 Data Analysis

In this work, the database established during training of the prediction model consists
of the assigned disputes including temperature, humidity, lighting and consumption
at the production level.
The overall diagram proposed system is described in Fig. 2. The data set estab-
lished in the model formation consists of two parts, one for training and the rest for
testing, this database undergoes such a process of normalization and standardization
described in Eq. 6:

x−μ
x= (6)
σ
Where μ describes the mean of characteristic vectors, σ this standard deviation,
x is the original characteristic vector and x is the vector of characteristics after
normalization. Based on the difference between the expected results and the actual
data, the efficacy of the approach used is determined.

Historical data of energy consumption

in industrial
Pre-processing and Feature

Data input analysis

extraction

Dataset training Dataset testing

ML Approche Model

Estimation

Fig. 2 Forecasting model

A Comparison Study of Machine Learning Methods … 171

4.2 Simulation Results

In this part, the simulation and analysis results of RNN, ANN, SVM, DT and RL algo-
rithms are expressed. The approaches are designed in the Python environment with
the PyTorch tensor library. Then, to examine actual and predicted energy consump-
tion using ANN, RNN, DT, SVM, and LR approaches are shown in Figs. 3, 4, 5, 6
and 7 figures respectively.
The performance of these algorithms used in the theme of precision are of average
absolute percentage Error (MAPE) described in Eq. 6:
n  
100%  y − yi∧ 
M AP E = (6)
n I =0  yi 

The energy consumption profile is defined by the correlation between the actual
and predicted energy consumption presented in Fig. 3. At the level of the simulation
using the ANN approach, we can see that the forecast curve accurately follows the
energy consumption compared to the real consumption. Then, the forecasting error
is less than 0.2.
In the RNN, the prediction error is of order 0.9, which means that this algo-
rithm is less efficient than the ANN. Figure 4 expresses the behaviour of the energy
consumption between the measured real and the profile estimated by the model.
The performance of the SVM is demonstrated in Fig. 5, where the error value
does not exceed 0.05 of the references, that expresses that the SVM has a strong
training capacity. In Fig. 5, it is well defined that the estimated energy consumption
is close to the curve of the real consumption.
For DT, the error is 0.5, which the variation in the energy consumption profile
between the actual consumption and the estimated is shown in Fig. 6

Fig. 3 Energy profile estimated by ANN

172 M. Bahij et al.

Fig. 4 Energy profile estimated by RNN

Fig. 5 Energy profile estimated by SVR

Fig. 6 Energy profile estimated by DT

A Comparison Study of Machine Learning Methods … 173

Figure 7 expresses the forecast of the energy consumption profile which is similar
to the actual consumption. The error obtained is of reference value 0.6 × 10−3 in the
LR approach.
The performance criteria used in the comparison of the five approaches are
the following: the absolute percentage of error MAPE, the training score and the
prediction score are describing in Table 1.
Moreover, The LR training score is 100% implies that higher performance algo-
rithm. The score is 99.999% in the SVM and the same in DT method. But those are
less inferior to the first RL algorithm.
The RNN has a value of 99.708% and the ANN has a score of 99.506%, which
presents the least efficient algorithm in the prediction of the model.
In Similarly, the value of the mean square error is 4.202 for ANN, which is lower
than the RNN approach of the order of 1.56. The error is 5.75 10–2 in DT method.
Then, the SVM is the value of 5.05 10−4 which expresses that the ANN and RNN
algorithms perform less than the two methods preceding. The LR most effective
approach in model training with a lower order error 3.45 10−28 than all approaches
proposed.

Fig. 7 Energy profile estimated by LR

Table 1 The performance

Approaches Training Score Predict Score MAPE
parameters for the five
(%) (%)
approaches
RL 100.00 100.00 3.45 10 −28
SVM 99.999 99.999 5.05 10–4
DT 99.999 99.990 5.75 10–2
RNN 99.970 99.970 1.5600
ANN 99.506 99.270 4.2026
174 M. Bahij et al.

5 Conclusion

This study focuses on the development of ML approaches for the prediction of

energy consumption in the Morocco industrial sector. The different ML methods,
including the linear regression (LR), artificial neural network (ANN), recurrent neural
network (RNN), support vector machine (SVM), and the decision tree (DT) have
been described and modelled. Moreover, the training of the consumption forecasting
model depends on the database of the attributes which influences consumption in the
industry.
The simulation results demonstrate that the linear regression offers better perfor-
mance compared to other methods. However, the SVM is less than the LR but, it is
better compared to the DT, ANN, and RNN.

References

1. Rapport annual. https://www.cese.ma/rapport-annuel-2019. Accessed 21 Nov 2020

2. Zhao HX, Magoulès F (2012) A review on the prediction of building energy consumption.
Renew Sustain Energy Rev 16(6):3586–3592
3. Wei Y, Zhang X, Shi Y, Xia L, Pan S, Wu J, Han M, Zhao X (2018) A review of data-driven
approaches for prediction and classification of building energy consumption. Renew Sustain
Energy Rev 82:1027–1047
4. Ahmad AS, Hassan MY, Abdullah MP, Rahman HA, Hussin F, Abdullah H, Saidur R (2014)
A review on applications of ANN and SVM for building electrical energy consumption
forecasting. Renew Sustain Energy Rev 33:102–109
5. Chakib C, Mohammed O (2018) Sensorless control of the PMSG in WECS using artificial
neural network and sliding mode observer. In: International symposium on advanced electrical
and communication technologies (ISAECT). Rabat, Morocco, pp 1–6
6. Patil S, Kulkarni U (2019) Accuracy prediction for distributed decision tree using machine
learning approach. In: 2019 3rd international conference on trends in electronics and informatics
(ICOEI). IEEE. Tirunelveli, India, April 2019, pp 1365–1371
7. Kavitha S, Varuna S, Ramya R (2016) A comparative analysis on linear regression and
support vector regression. In: 2016 online international conference on green engineering and
technologies (IC-GET). IEEE. Coimbatore, India, November 2016, pp 1–5
8. Salam A, El Hibaoui A (2018) Comparison of machine learning algorithms for the power
consumption prediction: case study of Tetouan city. In: 2018 6th International Renewable and
Sustainable Energy Conference (IRSEC). IEEE. Rabat, Morocco, December 2018, pp 1–5
9. Williams KT, Gomez JD (2016) Predicting future monthly residential energy consumption
using building characteristics and climate data: a statistical learning approach. Energy Build
128:1–11
10. Rodrigues F, Cardeira C, Calado JMF (2014) The daily and hourly energy consumption and load
forecasting using artificial neural network method: a case study using a set of 93 households
in Portugal. Energy Procedia 62:220–229
11. Jeong G, Park S, Lee J, Hwang G (2018) Energy trading system in microgrids with future
forecasting and forecasting errors. IEEE Access 6:44094–44106
12. Sossan F, Namor E, Cherkaoui R, Paolone M (2016) Achieving the dispatchability of distri-
bution feeders through prosumers data driven forecasting and model predictive control of
electrochemical storage. IEEE Trans Sustain Energy 7(4):1762–1777
A Comparison Study of Machine Learning Methods … 175

13. Hu M, Li W, Yan K, Ji Z, Hu H (2019) Modern machine learning techniques for univariate

tunnel settlement forecasting: a comparative study. Math Prob Eng
14. Pai PF, Hong WC (2005) Support vector machines with simulated annealing algorithms in
electricity load forecasting. Energy Convers Manage 46(17):2669–2688
The Review of Objectives, Methods,
Tools, and Algorithms for Educational
Data Mining

Mohamed Timmi, Adil Jeghal, Said EL Garouani, and Ali Yahyaouy

Abstract The remarkable growth in online learning resources, educational soft-

ware, the usage of internet in the educational context, and the creation of national
student information databases have produced extensive educational data repositories.
Today, virtual educational systems have grown drastically and have led us to store a
huge amount of potential data from a myriad of sources, with various formats and
distinct levels of complexity. Particularly, lock-downs have established the funda-
mental necessity of online learning tools. For this reason, each educational aim has a
range of specificities and characteristics. In fact, it goes along with numerous educa-
tional features requiring a specific formula to the extraction, processing and analysis
of data. This is reflected in the variety and specificity of Educational Data (EDM)
Mining tactics that are put in place to solve a definite educational assignment. The
pivotal intention of this paper is to put forth a detailed scrutiny on the previously
conducted studies related to educational data mining.

Keywords Educational data mining · EDM · Classification · Clustering and

prediction

M. Timmi (B) · S. EL Garouani · A. Yahyaouy

LISAC Laboratory, Faculty of Sciences Dhar El Mahraz, Sidi Mohamed Ben Abdellah University,
Fez, Morocco
e-mail: [email protected]
S. EL Garouani
e-mail: [email protected]
A. Yahyaouy
e-mail: [email protected]
A. Jeghal
LISAC Laboratory, National School of Applied Sciences, Sidi Mohamed Ben Abdellah
University, Fez, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 177
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_17
178 M. Timmi et al.

1 Preliminaries

This article discusses a study related to the field of data mining and how it can be
diligently implemented within the educational field. Besides, it expands upon the
central features of educational data mining processes and the objectives of these
operations, along with some rudiments that pave the way for an effective application
of data mining algorithms.
Historically, data mining is not novel, because it mainly relies on statistics that
date back to the middle of the 18th century. The current state of data mining is the
result of substantial efforts that have been accumulated for many years [1], yet it is
worth noting that this area has witnessed a remarkable development recently.
Returning to the roots of educational data mining, we find that research in
this area is relatively new. Although research and analysis of data, resulting from
educational software, started a long time ago, knowledge exploration in educa-
tional data mining (EDM) has been referred to, in various conferences and scien-
tific journals, as a separate field itself in recent times. This claim has also been
advocated in a wide range of reputable conferences on educational technologies,
namely the international conferences revolving around the following matters: educa-
tional data mining/artificial intelligence in education/intelligent systems/user
modeling, adaptation and personalization (UMAP).

2 Educational Data Mining

Educational data mining is defined in a variety of ways. However, the one that is put
forward in the Journal of Educational Data Mining is fundamentally insightful. In
this sense, the journal regards EDM as the process whereby methods are developed
to explore idiosyncratic data that come from educational environments [2], and the
use of these methods to deepen our own understanding of learners, and to reflect
them on the environment in which they operate. Correspondingly, another definition
considers EDM as the use of information mining techniques for specific types of
data extracted from educational environments to either deal with or solve important
education related issues. As far as we can tell, both definitions focus on extracting
knowledge from educational data to improve education systems [3]. The exploration
of knowledge in the field of education can occur by means of piecing the following
elements together: computer science, education and statistics [4]. With that in mind,
the intersection of the previous three areas forms other sub-areas, such as learning
analytics, e-learning and data mining [5] (Fig. 1).
Educational data mining is a multidisciplinary arena, for it uses various methods
and techniques, namely statistics, data mining, information retrieval systems, sugges-
tion systems, machine learning, science education, teaching principles and others.
Notwithstanding, the process of selecting these techniques depends predominantly
on the studied educational problem [6].
The Review of Objectives, Methods, Tools, and Algorithms … 179

Fig. 1 Areas in relation with

EDM

Fig. 2 Educational data mining process

The stages through which the process of exploration in educational systems goes
can be diagrammed as follows[7]: (Fig. 2).
It is also worth mentioning that there are other areas where data mining is applied,
such as criminal investigations, industrial engineering, fraud detection and customer
relationship management.

3 Selection of Features

The selection of features consists in choosing a small subset of features that are
prerequisite to describe the target concept by means of prediction, classification,
regression or clustering …. (It depends on the method used to explore the educa-
tional data). The selected subset is chosen with the assumption that the input data
contains a range of irrelevant features, and its removal would not affect in any way
the accuracy of the classification model (of the method used), as a matter of example.
Obviously, this is advantageous in the sense that it allowed a remarkable reduction in
computational complexity, better interpretability of the model and lessened data fit
180 M. Timmi et al.

by improving generalization. A wide spectrum of machine learning algorithms can

be utilized for this purpose [8], and the best feature propositions are listed below for
reference:
Student’s gender, physically disabled, family size, student’s living milieu, number
of brothers and sisters, student’s family situation, ownership of a vehicle, notes/notes
obtained in high school, grade, section, type of school, location of school, home study,
the parents’ education/occupation and monthly income.

4 Objectives of Educational Data Mining (EDM)

Exploration objectives can be categorized in educational data according to the types

of end users who benefit from the knowledge derived from exploration as follows
[2, 9]:

4.1 Learners

Providing notes and advice to students, meeting the needs of the students, working
to increase the students’ productivity and academic level, etc.

4.2 Teachers

Expand their understanding of the students’ nature, the progress of the educational
process, and get them to update their teaching methods and performance.

4.3 Researchers

Develop and compare data extraction techniques to intensify their focus on optimal
suggestions for each problem or specific pedagogical tasks, assess the effectiveness
of learning when using different methods, environments. etc.

4.4 Administrators

Come up with alternative canals to arrange resources (human and material) and assess
their educational establishments.
The Review of Objectives, Methods, Tools, and Algorithms … 181

It is also important to admit that the process of involving all stakeholders and
providing them with the results of the analysis is a critical task that should be consci-
entiously managed. Actually, some of these interests may end up taking an exclu-
sively different turn. Hypothetically, the use of learning analytics (LA) by personnel
managers to introduce technology into teaching methods might be misperceived, as
teachers would regard that as an attempt to evaluate or judge their teaching methods.
Additionally, the same goes for students who may feel uncertain to use personal data
for their assessment and classification. As far as we can see, this can produce an
unexpected consequence; teachers and students alike will be unwilling to use these
new technologies, or to partake in the various stages of the analysis.
Previously, the objectives were classified based on the types of end users. However,
it is challenging to categorize the objectives of knowledge exploration integrally
in education systems using one criterion, particularly, when all the objectives are
linked to several users. In this regard, another trustworthy research has described the
classification of exploration objectives in educational data as follows [9]:
• Student modeling: user modeling in the education system requires providing a
detailed account about the characteristics of students as well as their situation.
It is primordial to have an all-inclusive understanding of their knowledge, skills,
experiences and motivations. On the other hand, one needs to examine certain
types of problems, which can negatively derail the educational process, namely
the use of ineffective educational resources. Consequently, the objective is to
create or improve the student’s model based on the available information.
• Observe the students’ level and their results: the aim is to anticipate the outcomes
of every educational procedure by measuring how active the interactions in the
classroom are.
• Search for new learning suggestions: the aim is to suggest new content and better
tasks that are in conformity with the students’ current level/situation.
• A profound exploration of the learner’s behavior: This objective can actually vary.
Based on the previously delivered answers, the classification of students has to
make recourse to numerous factors, namely their personal profiles.
• Communicate with stakeholders: the objective is to assist teachers and get
managers to analyze the activities accredited to students.
• The analysis of the terrain structure: the goal is to define the terrain structure to
know the optimal content that can be added to the material and choose the best
possible sequence for the educational process. This can be achieved by devel-
oping the skill of predicting the student’s performance and determine the ideal
educational environment that would guarantee a fulfilled learning experience.
• The Improvement of educational content: it is closely linked to the two previous
objectives. The aim here is to decide on how to improve the educational material
(content, activities, etc.) using student information.
• Study the effects of educational support that is put forward by education systems.
It is noteworthy that the previously named objectives of educational data mining
are meant to improve and computerize education systems. More significantly, it can
182 M. Timmi et al.

be proclaimed that student modeling is the main key to achieve many objectives and
tasks through the exploration of educational data [7].

5 Methods Used for Educational Data Mining (EDM)

Attaining the objectives of educational data mining tends to call on a plenty of

traditional methods of data mining (clustering, classification, association analysis,
etc.) They are supposed to be classified as follows [2, 7, 10–12].

5.1 Prediction

The goal of this method is to develop a model that allows the inference of a particular
aspect of this data (called a predicate variable) in a set of aspects within the same data
reservoir (called predictive variables). The applications of educational data mining
using this method attempt to anticipate the student’s academic success by reviewing
their behavior in the classification-based virtual environment. Besides, going about
predicting the results of students who have completed the same teaching unite is
manifestly useful to back up those who may fail to fulfill it. The KNN algorithm
(K nearest neighbors) is among the techniques that are highly recommended in this
respect. It should be pointed out that the algorithms used in this area, are virtually
expressed in the interactions of students with the available sources in the form of
multidimensional rays [12, 13].

5.2 Clustering

This method implies, in the fewest possible terms, that the classified samples which
encompass data are naturally similar to each other. A cluster is widely defined as a
group of identical objects that differ from those found in other clusters. It is broadly
agreed upon that the efficiency of the cluster upsurges suggestively depending on
how similar or different it is from the elements contained in other clusters. This
mechanism is postulated to discover similar features between different behavioral
patterns manifested by students [13].

5.3 Relationship Mining

It refers to the examination of variables that constitute a data sample. This connection
can come in different forms, such as association rules mining, sequential pattern
The Review of Objectives, Methods, Tools, and Algorithms … 183

mining, correlation mining, and casual data mining. In the realm of educational data
mining, “relationship mining” is manipulated to determine the possible connections
between students’ online activities and their grades [13].

5.4 Distillation

This method includes portraying data in a manner that allows the user to rapidly
detect and categorize the data features. This technology chiefly depends on data
Summarization, visualization, and the use of Interactive Interfaces to demonstrate
useful information as well as supporting decision-making [13].

5.5 Outliers Detection

This method ruminates on the discovery of certain data segments with a marked
difference, compared to other casual genres of samples. It is predominantly applied
to explore educational data in order to reveal common behaviors displayed by the
involved stakeholders. Besides, it has been proven a practical medium to spot students
with learning anomalies [13].

5.6 Social Network Analysis

The structural analysis of this method is looking forward to studying the relationship
between individuals rather than considering their behavioral characteristics. From
an empirical viewpoint, social relationships are merely transcribed contractually to
establish multiple social links. This elementary segment can manifest itself, in terms
of educational data exploration, to interpret and analyze relationships in collaborative
assignments [9, 12, 13].

5.7 Process Mining

This particular procedure devotes itself to the extraction of knowledge-related

processes, using the system’s recorded information to create an obvious visual depic-
tion of the whole operation. In fact, it entails three well-calibrated sub-fields: confor-
mance, checking and model discovery. In the Learning Data Syndicate, the explo-
ration process can be put forth to reflect on the students’ behavior and its probable
impacts on their studies.
184 M. Timmi et al.

5.8 Text Mining

This method sheds light on extrapolating useful styles, patterns, and rules from
unstructured texts, namely HTML files, discussion messages, and emails. The Tasks
comprise text categorization, text compilation, analyzing the emotional insights
within a text, synthesizing documents and modeling relationships between diverse
objects. Essentially, the main reason behind using text mining is to examine the
content of discussions, forums, chats, web pages, documents… etc. [12, 13].

6 Algorithms Used for Educational Data Mining (EDM)

Educational data mining algorithms derive from a number of sciences. It is more or

less an amalgamation of several scientifically oriented disciplines, such as statistics,
mathematics, logic, science learning, artificial intelligence / expert systems, pattern
recognition, and machine science, long with intelligent and non-traditional sciences.
Educational data mining phase has attracted a considerable amount of attention in
the research community over the past decade, with the intention of developing scal-
able algorithms and adapting to increasing quantities of educational data in research
for significant cognitive models. Below are the most commonly used data mining
algorithms [11, 14]:

6.1 K-means

“K-means” is a reiterative algorithm that attempts to divide the dataset into K pre-
determined and non-overlapping subgroups (clusters). Here, every data point is only
affiliated to one group [15]. The pseudocode of K-means algorithm as follows [16]:
The Review of Objectives, Methods, Tools, and Algorithms … 185

6.2 Apriori

Apriori algorithm is given by R. Agrawal and R. Srikant in 1994 for finding frequent
set of items in a dataset for Boolean association rule. Surprisingly enough, this
significant appellation was given to the algorithm for its reliance on prior knowledge
of frequent itemset properties [15, 17].

6.3 EM Algorithm

EM algorithm is a combination of two distinctive features, expectation and maxi-

mization. It is an iterative method intended to find as much similarities as possible
in statistical models, where these latter depend on unobserved latent variables [18].

6.4 PageRank Algorithm

Google Search famously adopts PageRank algorithm to categorize web pages in their
search engine results [19].

6.5 C4.5 Algorithm

C4.5 algorithm is specially designed to generate a decision tree[15]. The pseudocode

of Tree C4.5 is shown as follow [20]:
186 M. Timmi et al.

7 The Available Tools for Educational Data Mining (EDM)

The applications of educational data mining require a considerable number of tools.

In truth, the choice of the appropriate tool depends on the nature of the studied data,
their size and the precise objective. In this scope, the most frequently used tools are
the following [14]:

7.1 Rapid Miner

Rapid Miner is widely known for being a readymade, open source, no-coding required
software, which helps in giving advanced analytics [11].

7.2 Weka

Weka is a range of machine learning algorithms for data mining tasks [11, 15].

7.3 Orange

Orange-This tool comes with visually programmed environments. It deals with tools
of data importing, data dragging and dropping widgets, as well as providing links to
connect dissimilar widgets together so that the workflow can be completed.

7.4 R

R- Is widely defined as a commonly used free software environment for statistical

computing and graphics.

7.5 Knime-Primarily

Knime-Primarily used for data preprocessing. It is a powerful tool with GUI that
shows the network of data nodes.
The Review of Objectives, Methods, Tools, and Algorithms … 187

7.6 Tanagra

Tanagra is a free open source of data mining software put in place for mere academic
and research purposes.

7.7 XL Miner

XL Miner-It is the only comprehensive data mining add-in for Excel, with neural
nets, classification and regression trees, logistic regression, linear regression, Bayes
classifier, K-nearest neighbors, discriminant analysis, association rules, clustering,
principal components, and a lot more.

7.8 Microsoft Excel

Microsoft Excel also provides maximum methods for statistical analysis.

8 Conclusion

Educational data mining (EDM) has shown itself as an enormously fertile research
area, considering its remarkable potentials to contribute to a deeper understanding of
teaching, learning and motivational processes in individual and collaborative teaching
environments.
Currently, EDM’s main contributions are focused on two main axes:
• Data analysis and creation of models to sharpen the collective understanding of
the learning process.
• Developing more effective and innovative methods to support learning and
enhance the qualitative dimension of using educational web applications (E-
Learning, M-Learning).
Today, around the world, many intelligent tutoring systems use EDM techniques
to provide more personalized and better learning.

References

1. Şahin M, Yurdugül H (2020) Educational data mining and learning analytics: past. Present
Future 9:121–131
188 M. Timmi et al.

2. Baker RS, Yacef K (2009) The state of educational data mining in 2009: a review and future
visions. JEDM| J Educ Data Min 1(1), 3–17
3. Romero C, Ventura S (2017) Educational data science in massive open online courses. Wiley
Interdisciplinary Rev Data Min Knowl Discov 7(1):e1187
4. Romero C, Ventura S (2020) Educational data mining and learning analytics: an updated survey.
WIREs Data Min Knowl Disc. https://doi.org/10.1002/widm.1355
5. Salloum SA, Alshurideh M, Elnagar A, Shaalan K (2020) Mining in educational data: review
and future directions. In: Joint European-US workshop on applications of invariance in
computer vision. Springer, pp 92–102
6. Romero C, Ventura S (2020) Educational data mining and learning analytics: an updated survey.
WIREs Data Min Knowl Disc 10(3):e1355. https://doi.org/10.1002/widm.1355
7. Anjum N, Badugu S (2020) A study of different techniques in educational data mining. In:
Advances in decision sciences, image processing, security and computer vision. Springer,
Cham, pp 562–571
8. Acharya A, Sinha D (2014) Application of feature selection methods in educational data mining.
Int J Comput Appl 103(2)
9. Mitrofanova YS, Sherstobitova AA, Filippova OA (2019) Modeling smart learning processes
based on educational data mining tools. In: Smart education and e-learning. Springer, pp 561–
571
10. Peña-Ayala A (2014) Educational data mining: applications and trends
11. Kumar AD, Selvam RP, Kumar KS (2018) Review on prediction algorithms in educational data
mining. Int J Pure Appl Math 118(8):531–537
12. Bakhshinategh B, Zaïane O, Elatia S, Ipperciel D (2017) Educational data mining applications
and tasks: a survey of the last 10 years. Educ Inf Technol 23. https://doi.org/10.1007/s10639-
017-9616-z
13. Aldowah H, Al-Samarraie H, Fauzy WM (2019) Educational data mining and learning analytics
for 21st century higher education: a review and synthesis. Telematics Inform 37:13–49. https://
doi.org/10.1016/j.tele.2019.01.007
14. Shende A, Thakare MS, Byagar S, Joshi MAN (2020) The review of different data mining tools,
techniques and algorithms for the data mining in education. Our Heritage 68(27):290–296
15. Dhingra K, Sardana KD (2017) Educational data mining: a review to its future vision. Int J
Technol Transfer Commer 15(3):309–321
16. Olukanmi P, Nelwamondo F, Marwala T (2019) Rethinking k-means clustering in the age of
massive datasets: a constant-time approach. Neural Comput Appl 1–23
17. Bhandari A, Gupta A, Das D (2015) Improvised apriori algorithm using frequent pattern tree
for real time applications in data mining. Procedia Comput Sci 46:644–651. https://doi.org/10.
1016/j.procs.2015.02.115
18. Kurban H, Jenne M, Dalkilic MM (2017) Using data to build a better EM: EM* for big data.
Int J Data Sci Anal 4(2):83–97
19. Agryzkov T, Curado M, Pedroche F, Tortosa L, Vicent JF (2019) Extending the adapted
pagerank algorithm centrality to multiplex networks with data using the Pagerank two-layer
approach. Symmetry 11(2):284
20. Wati, M, Budiman E, Kridalaksana AH, Haviluddin H, Dengen N, Purnawansyah P (2018)
Performance of decision tree C4. 5 algorithm in student academic evaluation
Brain-Computer Interface: A Novel EEG
Classification for Baseline Eye States
Using LGBM Algorithm

Said Abenna, Mohammed Nahid, and Abderrahim Bajit

Abstract This work has been carried out in the development of the best BCI systems
to recognize EEG visual system signal helps tired drivers, the elderly, patients, and
the disabled to drive on good safety condition. The method used in this article based
on an application of rhythms (delta, theta and alpha) on EEG acquisition channels to
increase the accuracy of the prediction system, after using delta rhythms on selected
electrodes using ExtraTrees algorithm, we notice an increase in accuracy from 95.46
to 99.97%, also this system has a good prediction speed of 12560 samples per second
during testing. Therefore, instead of minimizing the number of acquisition electrodes,
these bad electrodes can be used to increase the performance of the BCI system using
rhythms (delta, theta and alpha).

Keywords Brain-computer interface (BCI) · Electroencephalogram (EEG) ·

Visual system · Feature extraction · Classification

1 Introduction

The brain computer interface systems were launched in 1924 with the first acqui-
sition of EEG signals by Hans Berger. Each organ connects to a specific area in
the brain via the neural system, the distribution of electrical activity in the brain
refers to a sensation or movement of an organ, so these signals transmit through a
neurons network between organ and its analysis area in the brain [10]. The aim is to
provide a direct link between observed human intentions in the brain activities and
computer control, an important feature of current BCI systems is the high complexity
of their functional extracts compared to their simple classifications (usually linear)
[8]. In the human brain, there are billions of neurons and these neurons commu-
nicate with each other using weak electrical signals, using these electrical signals

S. Abenna (B) · M. Nahid

Faculty of Science and Technology, Hassan II University, Mohammedia, Morocco
A. Bajit
National School of Applied Sciences, Ibn Toufail University, Kenitra, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 189
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_18
190 S. Abenna et al.

brain activity is measured and these signals are called EEG signals [2], the main
purpose of using EEG signals is to diagnose brain diseases [3], most medical appli-
cations are based on BCI systems, because it can be used to prevent harmful effects
of smoking, alcohol consumption, it also helps detect and diagnose the brain disor-
ders, sleep disorders and tumours, the BCI can also be used to restore the health of
people with a cerebral vascular accident, disability and mental disorders [7]. The
EEG signals is recorded by several electrodes placed on specific areas from scalp,
the EEG signals have a distinctive high resolution of up to one millisecond, which is
still not possible with the latest imaging technologies such as computerized tomog-
raphy (CT) or magnetic resonance imaging (MRI) [13]. The EEG technology has
been progressively used in the auxiliary diagnosis of diseases such as schizophrenia,
mild cognitive impairment, epilepsy and Alzheimer’s disease. Above all, a close
relationship has been found between brain and depression, a cognitive ability of
depressed patients changes with mood, and these changes affect EEG [4], the diffi-
culties required in EEG signals cannot be controlled because many activities take
place in the brain at same time, the use of adaptive factors ap-pears to be beneficial
for biological symptoms such as electroencephalogram [11], there have been many
EEG classification studies in recent years, these studies used different classification
techniques, collected their performance and evaluated different collections of char-
acteristics, among these categories, k-nearest neighbor (k-NN), linear discriminant
algorithm (LDA), support vector machine (SVM), artificial neural network (ANN)
have become popular [6, 11]. In this paper, we have been interested in developing a
prediction system using EEG signals for eye states (open or closed), such that this
system is based on three main steps as a sequel, the selection of electrodes that allows
to determine the most important electrodes for classification and extraction of char-
acteristics based in the application of delta rhythms, theta and alpha on selected bad
electrodes, the final step is the use of the LGBM algorithm to classify EEG signals.
As a maximum accuracy value obtained in this work is in the order of 99.97%. The
rest of this article is organized as follows, Sect. 2 present the data used, feature
extraction and classifier proposed for system, the results obtained and discussed are
presented in Sect. 3, while Sect. 4 provides conclusions and an overview of future
work.

2 Methods

2.1 Dataset

This paper uses statistical data from an acquisition of EEG signals for eye states
(open or closed) using 14 non-invasive electrodes (Fig. 1a) from UCI Database [15].
The signals acquired using Emotive EEG Neuroheadset device use the labels ‘0’ to
indicates when eyes are open and ‘1’ to indicates when eyes are closed, the total
Brain-Computer Interface: A Novel EEG Classification … 191

Fig. 1 Electrodes position (a) and the visual system (b)

Fig. 2 Proposed general architecture for baseline eye states recognition system

number of data is 14980 samples, such as training data use 80% of the total data and
the rest 20% is used as testing data (Fig. 2).

2.2 Feature Extraction

Electroencephalography signals are noisier, so they have poor accuracy during a

direct classification, for this one using some algorithms for feature extraction. In this
work, we used the rhythms Delta (0.5 to 4 Hz), Theta (4 to 7 Hz) and Alpha (7 to
13 Hz) to minimize the noise generated by the electrodes before classification, to get
good results, we applied three types of filter passes the band in parallel, a first filter
uses delta rhythms, the second filter merge the two rhythms delta and theta (0.5 to
7 Hz), and the third uses alpha rhythms, such as delta and theta rhythms are applied
to the wrong electrodes and alpha on the selected medium.
192 S. Abenna et al.

2.3 Feature Selection

This step is very important in determining the electrodes that are weak system recog-
nition accuracy. There are many feature selection algorithms, in this work we used
ExtraTrees (ET) [12] based on SelectKBest and chi2 algorithms to select important
channels.

2.4 Classification

The classification of EEG signals is a very important step to resume the quality of
the BCI system, from many algorithms, the LGBM algorithm was chosen because of
its effectiveness. Like LGBM (Light-Gray Box Model) is a gradient-enhancing tool
that uses leaf tree growth for learning, which speeds up the formation and accuracy
of better results, even when you use less memory [9].

2.5 Evaluation Metrics

The principle measures are used to evaluate the performance of prediction in this
paper are:
With TP, FP, TN and FN are calculated from the confusion matrix of prediction
during testing (Fig. 3b):
-TP (True Positive): refers to a case where an alarm is raised when eyes open.
-TN (True Negative): refers to a case where an alarm is not raised when eyes
closed.
-FN (False Negative): refers to a case where an alarm is not raised when eyes
open.
-FP (False Positive): refers to a case where an alarm is raised when eyes closed.

Fig. 3 Evaluation metrics (a) and confusion matrix for binary class (b)
Brain-Computer Interface: A Novel EEG Classification … 193

Fig. 4 Channels selection (a), correlation matrix between electrodes (b)

3 Results

3.1 Feature Selection Results

The results of this paper were obtained using python3 on a core i-5 computer with
4-CPU and 6-Gb in RAM. Figure 4a illustrates a classification of electrodes from the
bad electrode used (P8) to the best electrode (O1), the Fig. 4b shows the correlation
values between the electrodes. we notice the existence of links between different
electrodes of each group (AF3, P8, F8), (F7, F3), (FC5, O1), (P7, AF4) and (O2, T8),
compared to Fig. 1b, each set of electrodes belongs to the same nerve which implies
the same signal characteristics and ‘1’ as correlation value between the electrodes,
which shows the good quality of the signals acquired by Emotive EEG Neuroheadset
device.

3.2 Classification Results

The classification of EEG data is the most important step in evaluating results. For
this, a set of metric values such as accuracy (ACC), sensitivity (SEN), specificity
(SP), precision (PR), kappa-score (k), F1-score (F1), mathews correlation coeffi-
cient (MCC), mutual information (MI), zero on loss (ZOL), pretreatment time (Ta),
classification time (Tc) and prediction time (Tp) were used.
Table 1 illustrates the classification results using the LGBM algorithm and rhythms
on the selection channels, this table shows the increase in classification performance
when applying the rhythms on several electrodes, so the classification time remains
constant, in such a way the accuracy values increased from 95.46% when not using
Table 1 Classification results with LGBM
194

NEEG Feature selection Classifier performances

Delta + Theta Alpha ACC (%) SE (%) SP (%) PR (%) F1 (%) K (%) MCC (%) MI ZOL Ta (s) Tc (s) Tp (s)
All All Nan 99.97 99.96 100.0 99.97 99.97 99.93 99.93 0.69 1 0.90 1.47 0.04
6 AF3, F3, FC5, P8, Nan 99.80 99.82 99.56 99.78 99.80 99.60 99.60 0.68 6 0.84 1.56 0.04
FC6, AF4
6 AF3, F3, FC5, P8, Nan 99.77 99.77 99.71 99.76 99.76 99.53 99.53 0.67 7 1.02 1.63 0.04
FC6, AF4
6 AF3, F3, FC5, P8, Nan 99.73 99.74 99.64 99.73 99.73 99.46 99.46 0.67 8 0.97 1.54 0.04
FC6, AF4
6 F3, FC5, T7, FC6, Nan 99.73 99.72 99.85 99.74 99.73 99.46 99.46 0.67 8 0.86 1.68 0.04
F4, AF4
5 F3, FC5, FC6, F4, Nan 99.70 99.72 99.56 99.69 99.70 99.39 99.39 0.67 9 0.96 1.63 0.04
AF4
5 FC5, P8, FC6, F8, Nan 99.70 99.69 99.71 99.70 99.70 99.39 99.39 0.67 9 0.96 1.63 0.04
AF4
5 AF3, FC5, P7, F8, Nan 99.70 99.69 99.78 99.71 99.70 99.39 99.39 0.67 9 1.02 1.55 0.04
AF4
5 AF3, FC5, T7, Nan 99.67 99.67 99.56 99.66 99.66 99.33 99.33 0.67 10 0.85 1.67 0.04
FC6, AF4
4 F3, O2, FC6, AF4 Nan 99.73 99.73 99.78 99.74 99.73 99.46 99.46 0.67 8 0.90 1.55 0.04
4 AF3, P8, T8, FC6 Nan 99.73 99.73 99.78 99.74 99.73 99.46 99.46 0.67 8 0.86 2.24 0.04
4 AF3, P8, FC6, AF4 Nan 99.67 99.66 99.71 99.67 99.66 99.33 99.33 0.67 10 1.02 1.73 0.04
4 T7, FC6, F4, AF4 Nan 99.63 99.63 99.56 99.63 99.63 99.26 99.26 0.67 11 0.96 1.68 0.04
3 FC5, F4, AF4 Nan 99.63 99.63 99.63 99.63 99.63 99.26 99.26 0.67 11 1.53 2.49 0.04
(continued)
S. Abenna et al.
Table 1 (continued)
NEEG Feature selection Classifier performances
Delta + Theta Alpha ACC (%) SE (%) SP (%) PR (%) F1 (%) K (%) MCC (%) MI ZOL Ta (s) Tc (s) Tp (s)
3 F3, O2, FC6 Nan 99.53 99.52 99.56 99.53 99.53 99.06 99.06 0.66 14 1.44 2.36 0.04
3 FC5, P7, FC6 Nan 99.50 99.49 99.49 99.50 99.50 98.99 98.99 0.66 15 1.43 2.55 0.05
3 FC5, P7, AF4 Nan 99.47 99.47 99.27 99.45 99.45 98.92 98.92 0.66 16 1.49 1.94 0.04
2 FC6, F4 Nan 99.47 99.46 99.42 99.46 99.46 98.92 98.92 0.66 16 0.81 1.73 0.05
2 P7 F8 99.43 99.41 99.56 99.44 99.43 98.86 98.86 0.65 17 0.85 2.04 0.05
2 FC6 F8 99.43 99.44 99.27 99.42 99.43 98.86 98.86 0.65 17 0.84 1.97 0.05
2 P7, F8 Nan 99.37 99.38 99.13 99.35 99.36 98.72 98.72 0.65 19 0.84 1.67 0.05
1 F3 Nan 99.33 99.33 99.20 99.32 99.33 98.66 98.66 0.65 20 0.81 1.86 0.05
1 P7 Nan 99.20 99.20 99.05 99.19 99.19 98.39 98.39 0.64 24 1.47 2.09 0.06
1 T8 Nan 99.17 99.14 99.27 99.17 99.16 98.32 98.32 0.64 25 0.93 2.65 0.11
1 FC6 Nan 99.17 99.15 99.19 99.17 99.16 98.32 98.32 0.64 25 0.83 1.87 0.05
0 Nan Nan 95.46 95.44 94.84 95.41 95.43 90.86 90.86 0.51 136 0.79 3.56 011
Brain-Computer Interface: A Novel EEG Classification …
195
196 S. Abenna et al.

Table 2 Comparative results with literature

Method NEEG Results
ACC (%) SE (%) SP (%)
Zhou et al. (2019) [15] Delta-RFC All 91.9 Nan Nan
Zhou et al. (2019) [15] Delta-RFC 5 99.8 Nan Nan
Zhou et al. (2019) [15] Delta-RFC 4 99.6 Nan Nan
Zhou et al. (2019) [15] Delta-RFC 3 99.6 Nan Nan
Hamilton et al. (2015) [5] K* + RRF All 97.40 96.4 98.3
Hamilton et al. (2015) [5] K* All 97.30 96.5 98.0
Hamilton et al. (2015) [5] ada All 97.20 96.3 97.9
Hamilton et al. (2015) [5] RRF All 95.10 92.5 97.1
Reddy et al. (2016) [14] MLN All 94.92 Nan Nan
Saghafi et al. (2017) [1] Delta-Theta-LR All 88.2 Nan Nan
Saghafi et al. (2017) [1] Delta-Theta-ANN All 82.4 Nan Nan
Saghafi et al. (2017) [1] Delta-Theta-SVM All 70.6 Nan Nan
This work Delta-Theta-LGBM All 99.97 99.96 100.0
Delta-Alpha-LGBM 5 99.83 99.83 99.85
Delta-Theta-LGBM 4 99.73 99.73 99.78
Delta-Theta-LGBM 3 99.63 99.63 99.63
Delta-Theta-LGBM 2 99.47 99.46 99.42
Delta-Theta-LGBM 1 99.33 99.33 99.20

feature extraction to 99.97% when using feature extraction. Compared to the results
with literature in Table 2, we notice that the wrong method was used by [1] using
the Delta-Theta-SVM method with an accuracy value of 70.6%, on the other hand at
our work, the accuracy value takes 99.97% and the specificity takes the ideal value
(100%). So to calculate the speed of prediction of the system during testing, one
divides the total number of test data over the duration of pretreatment and prediction,
to calculate the total number of test data from Table 1, we divide the value of the
ZOL on the error rate (100-ACC (%)), also the preparation time and prediction can
be calculated by a simple addition between 20% of Ta and Tp (0.2 * Ta + Tp). After
calculating the average speed of system prediction, it is found that its value takes
12560 samples per second.
Every literature work suffering some technical deficiency to not find good results,
the Zhou et al. [15] method use Delta rhythms only in the stage of feature extraction
of EEG data as in our work, but these rhythms can delete an important set of data
for classification, so we must add the rhythms Theta bandwidth as in the method of
Seghafi et al. [1] so the filter passes band uses bandwidth between 0.5 and 7 Hz to
add more character to the training data, with the possible minimum of noise, we also
notice the possibility of improving the quality of classification when merging two
algorithms as in the case of the Hamilton et al. [5] method, such that we notice that
Brain-Computer Interface: A Novel EEG Classification … 197

the accuracy value increases by 95.10% and 97.30 to 97.40% for this UCI dataset
classification, so we notice that RRF and LGBM finds the same accuracy value
without feature extraction algorithm, which shows that this work makes a good
correction of the different methods of literature.

4 Conclusion

In conclusion, this work improves the recognition of eye states, such as accuracy takes
the value 99.97% using the LGBM algorithm and low rhythms of EEG signals from
bad electrodes selected for the feature extraction, and compared to other methods,
we notice that our system is reliable with a high average prediction speed of 12560
samples per second. We hope that this study helps other researchers develop good
BCI systems to help and improve human life. For the future work, we will now
improve a new method for the feature extraction and classification steps of motor
movement-imagery recognition.

References

1. Saghafi A, Tsokos CP, Goudarzi M, Farhidzadeh H (2017) Random eye state change detection
in real-time using EEG signals. Expert Syst Appl. https://doi.org/10.1016/j.eswa.2016.12.010
2. Khosla A, Khandnor P, Chand T (2020) A comparative analysis of signal processing and
classification methods for different applications based on EEG signals. Biocybernet Biomed
Eng 5–6. https://doi.org/10.1016/j.bbe.2020.02.002
3. Emrah A, Turker T, Sengul D (2020) A Tunable-Q wavelet transform and quadruple symmetric
pattern based EEG signal classification method. Medical Hypotheses 1–2. https://doi.org/10.
1016/j.mehy.2019.109519
4. Cai H, Qu Z, Li Z, Zhang Y, Hu X, Hu B (2020) Featurelevel fusion approaches based on
multimodal EEG data for depression recognition. Inf Fusion 1–2. https://doi.org/10.1016/j.inf
fus.2020.01.00
5. Hamilton CR, Shahryari S, Rasheed KM (2015) Eye state prediction from EEG data using
boosted rotational forests. In: 2015 IEEE 14th international conference on machine learning
and applications. https://doi.org/10.1109/ICMLA.2015.89
6. Nicolae C, Nirvana P, Marius L (2020) Reading into the mind’s eye: Boosting automatic
visual recognition with EEG signals. Neurocomputing. https://doi.org/10.1016/j.neucom.2019.
12.076
7. Edla DR, Ansari MdF, Chaudhary N, Dodia S (2018) Classification of facial expressions from
EEG signals using Wavelet Packet Transform and SVM for wheelchair control operations.
Procedia Comput Sci 1–2. https://doi.org/10.1016/j.procs.2018.05.081
8. Farquhar J (2009) A linear feature space for simultaneous learning of spatio-spectral filters in
BCI. Neural Networks J. https://doi.org/10.1016/j.neunet.2009.06.035
9. Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, Ye Q, Liu TY (2017) LightGBM: a highly
efficient gradient boosting decision tree. In: 31st conference on neural information processing
systems, Long Beach, CA, USA
10. Eitan N, Alex F, Dan F (2020) Real-time EEG classification via coresets for BCI applications.
Eng Appl Artif Intell. https://doi.org/10.1016/j.engappai.2019.103455
198 S. Abenna et al.

11. Parvinnia E, Sabeti M, Zolghadri Jahromi M, Boostani R (2014) Classification of EEG Signals
using adaptive weighted distance nearest neighbor algorithm. Comput Inf Sci 1–2. https://doi.
org/10.1016/j.jksuci.2013.01.001
12. scikit-learn user guide, Release 0.20.4, 2019, pp 255–257
13. Amin SU, Alsulaiman M, Muhammad G, Mekhtiche MA, Shamim Hossain M (2019) Deep
Learning for EEG motor imagery classification based on multi-layer CNNs feature fusion.
Future Gen Comput Syst 1–2. https://doi.org/10.1016/j.future.2019.06.027
14. Reddy TK, Behera L (2016) Online eye state recognition from EEG data using deep architec-
tures. In: 2016 IEEE international conference on systems, man, and cybernetics. https://doi.
org/10.1109/SMC.2016.7844325
15. Zhou Z, Li P, Liu J, Dong W (2019) A novel RealTime EEG based eye state recognition system.
Springer. https://doi.org/10.1007/978-3-030-06161-6 17
A Light Arabic POS Tagger Using
a Hybrid Approach

Khalid Tnaji, Karim Bouzoubaa, and Si Lhoussain Aouragh

Abstract Part Of Speech (POS) tagging is the ability to computationally deter-

mine which POS of a word is activated by its use in a particular context. It is
a useful preprocessing tool in many natural languages processing (NLP) applica-
tions. In this paper, we expose a new Arabic POS Tagger based on the combina-
tion of two main modules: the 1st order Markov and a decision tree models. These
two modules allow improving existing POS Taggers with the possibility of tagging
unknown words. The tag set used for this POS is an elementary tag set composed
of 4 tags {noun, verb, particle, punctuation} that are sufficient for some NLP appli-
cations but greatly help increasing the accuracy. The POS tagger has been trained
with the NEMLAR corpus. The experiment results demonstrate its efficiency with
an overall accuracy of 98% for the full system.

Keywords Natural Language Processing · Arabic text processing · POS tagging ·

Machine learning · Decision trees

1 Introduction

The world of the Internet has known a huge and continuous growth regarding the
Arabic content in terms of texts, videos, images, etc. Statistics show1 that this content
has increased by 9348% between the year of 2000 and 2020. As a result, developing
Arabic Natural Language Processing (ANLP) tools is required to process this huge

K. Tnaji (B) · K. Bouzoubaa

Mohammadia School of Engineers, Mohammed V University in Rabat, Rabat, Morocco
e-mail: [email protected]
K. Bouzoubaa
e-mail: [email protected]
S. L. Aouragh
Faculty of Law, Economics and Social Sciences, Salé, Mohammed V University in Rabat, Rabat,
Morocco
1 https://www.internetworldstats.com

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 199
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_19
200 K. Tnaji et al.

and growing content. In fact, there are many applications that have been already
developed for Arabic such as Part-of-speech (POS) taggers. This latter is an essential
tool in many NLP applications such as word sense disambiguation [18], question
answering [11] and information retrieval [9].
However, the number of Arabic POS taggers that can be used as rapid and free
tools is still low [1, 7, 8]. In addition, they suffer from three major problems. The
first and foremost problem concerns words having more than one possible tag when
they are used in different contexts. The second problem concerns words not found
in the annotated training set since the majority of the POS taggers can tag only
words that actually exist in the training set. Therefore, these unknown words lead
to a major problem in any tagging system and always decrease its performance.
Finally, as will be detailed in the literature review, Arabic researchers use different
and non-standard tag sets such as [4, 10, 14]. This situation leads to the difficulty of
comparing or combining these POS taggers.
Therefore, our objective is to develop a new tool for Arabic POS tagging fixing the
above-mentioned drawbacks. This POS tagger is developed using a hybrid approach
combining a statistical one to deal with ambiguous words and a symbolic one to
cope with unknown words. Finally, instead of specifying a full and very detailed tag
set, we limit our tag set to four generic tags (noun, verb particle and punctuation)
since they are enough in many kinds of applications. We consider then our new POS
tagger as a light Arabic POS tagger.
The rest of this paper is organized as follows. The next section presents related
works concerning POS tagging techniques for unknown and ambiguous words.
Section 3 provides the steps for developing our POS tagger. The results of experi-
ments are described and commented in Sect. 4. The exploitation of this system is
described in Sect. 5. Finally, some conclusions and future works are given in Sect. 6.

2 Related Works

Research on POS tagging has a long history where numerous techniques have been
successfully applied. Some of them focus on the use of linguistic rules. Others
consider POS tagging as a classification problem: the tags are the classes and a
machine learning algorithm is used to assign each occurrence of a word to one class
based on the evidence from the context.
Different Arabic taggers have recently been proposed. Some of them use super-
vised machine learning approaches such as [1] while others use hybrid approaches,
combining the rule-based approach and the statistical one such as [2]. Some of these
POS taggers handle unknown words such as [13]. A detailed review about the most
recent works related to Arabic POS taggers can be found in [16].
The major problem in all these Arabic POS taggers is the absence of a standard tag
set. For example, the number of tags varies between 6 tags in the case of Columbia
Arabic Treebank POS tag set [10], and 2000 tags in the case of Penn Arabic Treebank
(PATB) [14]. This situation makes it impossible to compare or combine these existent
A Light Arabic POS Tagger Using a Hybrid Approach 201

systems. A comparison about some of the existing Arabic tag sets can be found in
[5] where we can notice the big difference between the tag sets because each one is
developed for a specified application.
Some studies have been done to develop an Arabic standard tag set such as [5, 12]
that aims to design detailed hierarchical levels of the Arabic tag set categories and
their relationships.
Relating to the rapidity of the Arabic POS tagging systems, few works report on
this measure despite its importance. Among the systems that have a known execution
time, we find farasa POS tagger [7] which is able to process 2000 words per second.

3 Our Approach

Let us remind that our objective is to develop a POS tagger with unknown words
handling. To avoid the non-standardization of the tag sets and to improve the POS
execution time, one solution consists in reducing the tag set containing the basics
{noun, Verb, Particle, Punctuation}.
The development of this POS tagger is based on a supervised machine learning
approach exploiting an annotated corpus. This system is developed using the Hidden
Markov Model (HMM) combined with a decision tree guessing model based on word
substring information.

3.1 Data Set

The data set used for this POS tagger is the NEMLAR corpus. This corpus was
produced within the NEMLAR project2 [3] and consists of about 500,000 words of
Arabic text from 13 different categories, aiming to achieve a well-balanced corpus
that offers a representation of the variety in morphology, syntactic, semantic and
pragmatic features of modern Arabic language such as political news, political debate
or Arabic literature.
The tag set existing in the NEMLAR corpus is detailed, and since we typically
need only the 4 elementary tags, we performed a simple modification on this corpus
to map the used NEMLAR tag set to our four basic tags. For instance, adjectives,
nouns and proper nouns are all mapped to the basic noun tag.

2 https://www.nemlar.org
202 K. Tnaji et al.

3.2 Processing

A POS tagging system provides the most likely tag for each word of the sentence. Our
system is based on the hidden Markov model and the Viterbi algorithm. The HMM
[15] is a sequence model. A sequence model or sequence classifier is a model whose
job is to assign a tag or class to each unit in a sentence, thus mapping a sequence
of words to a sequence of tags. An HMM is a probabilistic sequence model: given
a sequence of words, it computes a probability distribution over possible sequences
of POS tags and chooses the best tag sequence by using the Viterbi algorithm.
However, HMM can only tag the words existing in the training data. To tag the
unknown words we have to pursue another technique. Indeed, our approach is to
exploit the morphology of words (word substring information) to associate with each
one its most likely tag. For instance, in the Arabic word. “‫( ”متفاهمون‬mtfAhmwn)3
[6], the three characters (‘‫( ’ف‬f), ‘‫( ’ه‬h)) and (‘‫( ’م‬m)) are root radical characters.
Therefore, they can be replaced by other three Arabic characters such as {‘‫( ’ق‬q),
‘‫( ’د‬d), ‘‫( ’ص‬S)} to produce the other Arabic word “‫( ”متصادقون‬mtSAdqwn) which
have different meaning but both words are nouns. The other characters at the begin-
ning and the end of the word represent the augmented characters. Linguistically
speaking each augmented character can be only one of these Arabic ten characters
{‘ ‫( ’ا‬A), ‘‫( ’ه‬h), ‘‫( ’ی‬y), ‘‫( ’ن‬n), ‘‫( ’و‬w), ‘‫( ’م‬m), ‘‫( ’ت‬t), ‘‫( ’ل‬l), ‘‫( ’أ‬O), ‘‫( ’س‬s)}.
The augmented characters, also known as affixes, associated with their position in
the word play a critical role in determining the possible POS tag of the word. For
example, Arabic words “‫( ”یتصالح‬ytSAlH), “‫( ”یتصاعد‬ytSAEd), “‫( ”یتعامل‬ytEAml) all
have three augmented characters {‘‫( ’ی‬y), ‘‫( ’ت‬t)} and {‘‫( ’ا‬A)} in the first, second
and the fourth position and all these words are verbs. Therefore, our objective is to
extract automatically these classification rules that allow classifying a word from its
augmented which represents the affixes4 . The extraction of these rules is done by
using a decision tree algorithm. To build our decision tree algorithm.
The idea is simple and consists in building a decision tree by learning all over the
corpus and using as an attribute vector containing the prefix (first and second letters),
the suffix (last and before last letter) and the size of the word which replaces the infix
to avoid the complexity of extracting it. Consequently, the non-terminal nodes in our
decision tree are the attributes (prefix, suffixes or word’s size). The edges represent
the possible values for each attribute: Arabic letters or a number which represent the
size of the word. Finally, the terminal nodes are the POS tags.
A Light Arabic POS Tagger Using a Hybrid Approach 203

3.3 Combining the Two Approaches

The Objective is tagging a sentence composed of several known and/or unknown

words:
• If all the words are known, we can use the Viterbi algorithm to estimate the most
likely tag sequence by using the emission and transition probabilities.
• If there is one or more unknown words, this leads to zero probabilities being
assigned to unseen words in the emission matrix, causing the probability of the
whole sequence to be set to zero.
To eliminate these zeroes in the emission matrix, we use the distribution of prob-
ability given by our decision tree-guessing model. The POS Tagging procedure is
done as indicated in (Fig. 1). We must indicate that on a reduced tag set and a large
corpus the transition probabilities will not cause a problem because virtually there
will be no transitions with null probability.
For instance, let us consider the following sentence: “‫( ”یأکل الولد من الطیبات‬yOkl
Alwld mn AlTybAt). In the first step, the emission matrix of the input sentence is
constructed, that means assigning a probability of words having each tag. Therefore,
we have to extract this probability for each word of the sentence. If the word is

Fig. 1 Procedure of sentences tagging

3 Transliterated using Buckwalter [6].

4 Affixes represent the prefixes, infixes and suffixes indicate substrings that come respectively at the

beginning middle and at the end of a word.

204 K. Tnaji et al.

Table 1 Test on the

Test Result
ambiguity problem.
Total number of tested words 101991
Words correctly classified 100737 (99%)
Words incorrectly classified 1254 (1%)

present in the vocabulary, the vector of word/tag emission probability is obtained

directly. Otherwise, we apply our decision tree guessing model that receives the
prefix, suffix and the length of the unknown word as input. For example, if the word
“‫( ”الطیبات‬AlTybAt) is unknown the input parameters for the prediction model are the
prefix “‫( ”ال‬Al), the suffix “‫( ”ات‬At) and the word’s size (7). The model generates
a probability distribution of this word on all tags and this distribution is used as
an emission probability vector to complete the emission probability matrix of the
sentence.

4 Experiments

The test is subdivided into four evaluation series where the first three use the Nemlar
corpus. In the first one, we test the efficiency of our system on the ambiguous words
using only HMM. In the second, we test our decision tree guessing model and for
the third series of evaluation, we test the full POS tagging system. Finally, we test
the POS tagging system using another corpus.

4.1 Evaluation 1

To evaluate the efficiency of our system on ambiguous words, we test the POS tagger
which depends only on the hidden Markov model. Therefore, it is necessary to use
a test data with no unknown words. Statistics and results are presented in Table 1.
Concerning words, the precision of the system is close to perfection. A precision
of 99% is very high, but it is justifiable. First, all the words exist in the vocabulary and
we are using a reduced tag set. In addition to that we can neglect the punctuation that
can never be misclassified. Therefore, the remaining 1% error is practically between
the verbs, nouns and the particles.

4.2 Evaluation 2

For this experiment, we performed a test by using 5-fold-cross-validation. Table 2

below represents the evaluation results.
A Light Arabic POS Tagger Using a Hybrid Approach 205

Table 2 Decision tree cross

Test Result
validation evaluation
Total number of tested words 539,465
Words correctly classified 527,290 (97.74%)
Words incorrectly classified 12,175 (2.26%)

Table 3 Total evaluation of

Test Result
the POS tagger
Total number of Words 85,492
Words incorrectly classified 1,362 (1.6%)
Number of unknowns Words 908
Unknowns words incorrectly classified 292 (31%)
Execution time 4,473 words/second

Despite that the POS tag of some words depends on its context in the sentence
and/or on the diacritics, this model provides an excellent accuracy.

4.3 Evaluation 3

To evaluate the full POS tagging system, we process with the train-test split technique.
In the first step we divide the full dataset into sentences. In the second step, we take
80% of the sentences for training and we keep 20% for the test. Table 3 summarizes
the statistics and results.
As we notice the number of misclassified words is extremely insignificant. This
is due to the very good quality of our model but also to the elementary tag set used.
Concerning unknown words, the number of misclassified words represents 31% of
the total unknown words that seems relatively high but technically it’s caused by the
hidden Markov model and/or the decision tree guessing model.
Finally, relating to the rapidity of this POS Tagger note that our model is extremely
rapid and could process about 4,473 words per second (with i3 processor and
4 GB/RAM) in comparison with farasa POS tagger [7] which can process about
2,000 word/second (with i7 processor and 16 GB/RAM).

4.4 Evaluation 4

In this series of evaluation, we tested our system on the WikiNews5 corpus. It contains
manually-segmented and POS tagged words from articles on seven domains: politics,

5 Corpus available at https://github.com/kdarwish/Farasa.

206 K. Tnaji et al.

Table 4 Evaluation of the POS tagger using WikiNews corpus

Test Result
Total number of tested words 9,940
Number of unknown words 1758 (17.68%)
Words correctly classified 9,548 (96.05%)
Words incorrectly classified 392 (3.95%)
Execution time 4,000 words/second

Fig. 2 POS Tagging results interface

economics, health, science and technology, sports, arts, and culture. We started with
a preprocessing step that transforms the 16 used tags into our reduced tag set. Table 4
summarizes the statistics and results. For instance, the WikiNews tags noun, adjective
are mapped into the noun tag.
Once again, we notice the good accuracy and rapidity of this POS tagging system,
despite the considerable number of unknown words.

5 Exploitation

This system has been developed in java and has been integrated into the Safar frame-
work6 . This framework integrates several tools to process the Arabic language at
the morphological, syntactic and semantic levels containing parses, lemmatizers or
segmenters [17]. The framework is available both as a web interface7 and as an API.
Figure 2 shows the use of the POS tagger web interface. After selecting “Safar light
POS Tagger” from the input “Select Tagger”, the user can either tag a text by entering
it to the target zone or by directly importing a text file from the machine. The tagging

6 http://arabic.emi.ac.ma/safar/
7 http://arabic.emi.ac.ma:8080/SafarWeb/
A Light Arabic POS Tagger Using a Hybrid Approach 207

Fig. 3 POS Tagging program example

results can be visualized directly or they can be generated as an Excel, CSV or XML
file.
An example of using the POS tagger via SAFAR API is illustrated in Fig. 3. To do
so, we import the package “safar-modern_standard_arabic-basic-syntax-posTagger”.
The following code tags a text with only two instructions and then outputs the POS
for every word in the input text.

6 Conclusions and Future Works

This work was done in the perspective of having a rapid and efficient ANLP tool.
Indeed, we have developed a POS tagger based on a hybrid approach containing two
main modules: a statistical module based on HMM, and the second one based on
decision trees. For the tag set we chose an elementary tag set to ensure accuracy and
rapidity which allowed us to develop a very rapid and accurate POS tagging system
with unknown words handling.
The evaluation of each module gives satisfying results (accuracy of 99% for HMM
and 97% for the tree-guessing model) and after the assembly of the two modules for
the complete POS tagger the results obtained are encouraging and the execution time
is very low.
As perspectives for the improvement of this work, we plan to enlarge the vocabu-
lary and increase the order of the hidden Markov model. It is also worth focusing on
the disparity of the POS tag set of the Arabic language and propose a standardized
version of both the tag set and the corresponding system.

References

1. Al Shamsi FG (2006) A hidden Markov model-based POS tagger for Arabic. In: Proceeding
of the 8th international conference on the statistical analysis of textual data, France, pp 31–42
2. Albared MO (2009) Arabic part of speech disambiguation, pp 517–532
208 K. Tnaji et al.

3. Attia MM (2005) Specifications of the Arabic Written Corpus produced within th NEMLAR
project
4. Atwell ES (2008) Development of tag sets for part-of-speech tagging
5. Atwell MS (2013) A standard tag set expounding traditional morphological features for Arabic
language part-of-speech tagging. Edinburgh University Press
6. Buckwalter Arabic Transliteration. (n.d.). https://www.qamus.org/transliteration.htm. 20 Oct
2020
7. Darwish K, Mubarak H, Abdelali A, Eldesouki M (2017) Arabic POS tagging: don’t abandon
feature engineering just yet. In: Proceedings of the third arabic natural language processing
workshop, pp 130–137. https://doi.org/10.18653/v1/W17-1316
8. Diab M, Hacioglu K, Jurafsky D (2004) Automatic tagging of Arabic text: from raw text
to base phrase chunks. In: Proceedings of HLT-NAACL 2004: short papers. association for
computational linguistics, pp 149–152
9. Dinçer BT, Karaoğlan B (2004) He effect of part-of-speech tagging on IR performance for
Turkish. In: Aykanat C, Dayar T, Körpeoğlu İ (eds.), Computer and Information Sciences—
ISCIS 2004, Springer, pp 771–778. https://doi.org/10.1007/978-3-540-30182-0_77
10. Habash NF (2009) Syntactic annotation in Columbia Arabic Treebank. In: 2nd International
Conference on Arabic Language Resources & Tools MEDAR. Cairo
11. Hammo B, Abu-Salem H, Lytinen SL, Evens M (2002) QARAB: A: question answering system
to support the Arabic language. In: Proceedings of the ACL-02 workshop on Computational
approaches to semitic languages. July 2002
12. Imad Zeroual AL (2017) Towards a standard Part of Speech tagset for the Arabic language. J
King Saud Univ Comput Inf Sci 171–178
13. Albared M, T-M O-S-A (2005) probabilistic Arabic part of speech tagger with unknown words
handling. J Theor Appl Inf Technol
14. Maamouri MA (2004) Developing an Arabic treebank: methods, guidelines, procedures, and
tools. In: Proceedings of the 20th international conference on computational linguistics
15. Rabiner L, Juang B (1986) An introduction to hidden Markov models. IEEE ASSP Mag 3(1):4–
16
16. Salameh S (2018) A review of part of speech tagger for Arabic Language. International Journal
of Computation and Applied Sciences IJOCAAS, 4, 4–5, June 2018. Darwish K, Mubarak H
(n.d.). Farasa: A New Fast and Accurate Arabic Word Segmenter. 5
17. Jaafar Y, Bouzoubaa K (2015) Arabic natural language processing from software engineering
to complex pipelines. In: Cicling 2015, Cairo, Egypt, April 2015
18. Zouaghi A, Merhbene L, Zrigui M (2012) Combination of information retrieval methods with
LESK algorithm for Arabic word sense disambiguation. Artif Intell Rev 38:257–269. https://
doi.org/10.1007/s10462-011-9249-3
RSEPUA: A Recommender System
for Early Predicting University
Admission

Inssaf El Guabassi, Zakaria Bousalem, Rim Marah, and Aimad Qazdar

Abstract Machine Learning allows us to reduce the human error probability by

providing very strong recommendations, predictions, and decisions based on only
the input data. For that reason, it has become one of the most important and common
aspects of the digital world. Different application areas adapt and adopt Machine
Learning techniques in their systems such as medicine, finance, marketing, busi-
ness intelligence, healthcare, etc. In our case, we aim to design a recommender
system based on Machine Learning techniques in the field of Education. Therefore,
the purpose of this research work is to provide a recommender system for early
predicting university admission based on four Machine Learning algorithms namely
Linear Regression, Decision Tree, Support Vector Regression, and Random Forest
Regression. The experimental results showed that the Random Forest Regression is
the most suitable Machine Learning algorithm for predicting university admission.
Also, the Cumulative Grade Point Average (CGPA) is the most important parameter
that influences the chance of admission.

Keywords Linear regression · Decision Tree · Support vector regression ·

Random forest regression · Predictive model · Educational data mining

1 Introduction

Machine Learning is a subset of artificial intelligence (AI) that enable computers

to automatically improve through experience, today it is becoming an increasing
part of our daily lives because its applications are extending to different areas like
agriculture, finance, electronic commerce, logistics, marketing, security, shopping,

I. El Guabassi (B) · R. Marah

Faculty of Sciences, Tetouan, Morocco
Z. Bousalem
Faculty of Sciences and Technology, Settat, Morocco
A. Qazdar
ISI Laboratory FS Semlalia UCA, Marrakech, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 209
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_20
210 I. El Guabassi et al.

etc. Moreover, Machine Learning enables an increasing number of applications that

were not possible before. In the area of education, the adoption of Machine Learning
is also accelerating. In recent decades, several researchers and scientists show their
interest in the application of Machine Learning in an educational context [1–3].
Therefore, some examples of the topics that are studied by different researchers are:
• Dropout prediction in e-learning courses
• Early prediction of student outcomes
• Prediction of academic performance associated with internet usage behaviors
• Predicting at-risk university students in a virtual learning environment
• Mental stress detection in university students
• Dropout early warning systems for high school students
• Analyze and support mediation of student e-discussions
• Predicting MOOC dropout over weeks
• Identifying students’ inquiry planning
• Detection of learner learning style
• Etc.
In this sense, early predicting university admission is also considered an important
topic for not only the new graduate students but also for the university. Unfortunately,
newly graduate students usually are not aware of admission requirements to a post-
graduate program at the university. Thus, they, therefore, spend their precious time
and money focusing on things that won’t increase their chances of admission to
graduate programs. The real and major issues of this topic are: Firstly, which the best
Machine Learning algorithm for predicting university admission? Secondly, what
most important parameters can affect the chance of admission?
The main aim of this paper is to provide a Recommender System for Early
Predicting University Admission based on Machine Learning algorithms. Hence,
the purposes are threefold. The first is to apply several supervised Machine Learning
algorithms (i.e., Linear Regression, Support Vector Regression, Decision Tree
Regression, and Random Forest Regression) on our dataset. The second purpose is to
compare and evaluate algorithms used to create a predictive model based on various
evaluation metrics. The last purpose is to determine the most important parameters
that influence the chance of admission. It should be noted that the present research
work uses the dataset for graduate students at the University of California in Los
Angeles.
The outline of the present paper is as follows: Sect. 2 presents recent studies
regarding the specified area. The settings and basic definitions are briefly described in
Sect. 3. Section 4 concentrates on the proposed methodology. In Sect. 5 our proposed
recommender system is presented. Section 6 contains results and discussion. Finally,
Sect. 7 presents the main conclusions considering some future research directions.
RSEPUA … 211

2 Related Work

Due to the rapidly growing interest in the field of education [4–6], there are several
research studies have been conducted on predicting university admission based on
several factors using supervised or unsupervised Machine Learning algorithms.
Xiaojun Wu & Jing Wu [7] conducted a study on predicting students’ selection
criteria in non-native language MBA admission based on Ridge regression, SVM,
Random forest, GBDT. Al Ghamdi et al. [8] applied three classification methods
which are linear regression, decision tree, and logistic regression model to auto-
matically predict postgraduate admission. Nandal [9] developed Student Admission
Predictor (SAP) based on Deep Learning techniques namely Decision Tree, Support
Vector Machine, Gaussian Naive Bayes, Linear Regression, Random Forest, and
Deep Neural Networks. Zhao et al. [10] explained their study based on a quantitative
Machine Learning approach to predict master students’ admission in professional
institutions. They used standard SVM, S3VM, and SVM+, as well as their dataset,
which is collected from Northeastern University’s MS in Computer Science (MSCS)
program.

3 Settings and Basic Definitions

In this section we will discuss the three main elements to achieve our purpose, these
elements are Recommender Systems, Algorithms used, and evaluation methods.

3.1 Recommender Systems

Simply put, a recommender system is an application intended to offer user items

that may be of interest to him according to his profile. Recommender systems are
used in particular on online sales websites. They allow e-commerce merchants to
automatically highlight products likely to interest visitors. Recommender algorithms
can be divided into three categories which are:
• Content-Based Filtering [11]: This type of recommender system is based on
profiles. In fact, we build profiles for users as well as for products.
• Collaborative Filtering [12]: This method tries to find a group of users who have
the same tastes and preferences as the target user.
• Knowledge-based systems [13]: This type of recommender system recommends
items based on the knowledge user.
212 I. El Guabassi et al.

3.2 Algorithms

There are many algorithms for predictive modeling Machine Learning. In the next
sections, we will present the algorithms used to build predictive models.

Linear Regression [14] is the most important algorithm in the field of Machine
Learning, especially supervised learning. It is a way to model a relationship between
a dependent variable and one or more independent variables. It consists of finding a
regression line straight line through the points.

Decision Tree [15] is the most widely used classification and prediction technique. It
is a tree structure, where each internal node with outgoing edges indicates a condition
on an attribute, each branch is an outcome of the test, and each leaf terminal node
represents a class label.

Support Vector Regression (SVR) [16] is also a very popular Machine Learning
technique used in both classification and regression. It is similar to Linear Regression
with only a few minor differences. SVR allows defining how much error is acceptable
in our predictive model and will find an appropriate line to fit the data.

3.3 Evaluation Methods

Evaluating a model is a core part of building an effective Machine Learning model.

There are many methods of evaluation that can be used. In the following, we will
discuss the three main metrics which we will use in our evaluation.

R-Squared (R2 or the coefficient of determination) [18] is an indicator that allows

judging the quality of simple linear regression. It measures the fit between the
model and the observed data or how well the regression equation is to describe
the distribution of points.

Mean Square Error (MSE) [18] is the arithmetic mean of the squares of the predic-
tions between the model and the observations. This is the value to be minimized in
the context of a single or multiple regressions.

Root Mean Square Error (RMSE) is a standard way to measure the error in model
evaluation studies. It is the square root of the mean of the square of all of the errors.

After briefly determining the settings and describing basic definitions, in the next
section, we will present the methodology used to develop our recommender System
RSEPUA.
RSEPUA … 213

Machine Learning
Algorithms

Dataset Data
Preparation
Modeling

Deployment
&
Evaluation
Integration

Fig. 1 Methodology used to our recommender System RSEPUA

4 Proposed Methodology

Admission to graduate schools is the most difficult and important step. The major
problem students often fall into when applying to graduate schools is they are not
aware of admission requirements. The purpose of this research work is to provide
a recommender system for early predicting university admission based on Machine
Learning algorithms. The methodology used in this work is composed of six steps.
Figure 1 presents an overview of this methodology.

Dataset: The data used in this research work is collected from a second version of
the data set named “Graduate Admission 2” [19]. It is inspired by the University
of California in Los Angeles Graduate Dataset. This data is available publicly even
on Kaggle which offers free datasets for training and evaluation. It is comprised of
500 rows with 8 features. The dataset contains various features which are considered
important during the enrollment and admission for master’s programs.

Table 1 represents an overview of dataset features used for training and testing.
214 I. El Guabassi et al.

Table 1 Dataset features

Feature Description Type
GRE score Graduate Record Examinations ( i.e., from 260 to 340) Quantitative
TOEFL score Test of English as a Foreign Language ( i.e., from 0 to 120) Quantitative
University rating University Rating ( i.e., from 1 to 5) Quantitative
SOP Statement of Purpose ( i.e., from 1 to 5) Quantitative
LOR Letter of Recommendation ( i.e., from 1 to 5) Quantitative
CGPA Cumulative Grade Point Average ( i.e., from 0 to 10) Quantitative
Research Research Experience ( i.e., 0 or 1) Quantitative
Chance of admit Probability of getting admitted ( i.e., from 0 to 1) Quantitative

Data Preparation: It is referred to as “data preprocessing”. It represents one of

the most crucial steps in all Machine Learning projects because it involves data
collection, formatting data, Improving data quality, feature engineering, and labeling.

Modeling: This step involves the conception of different Machine Learning algo-
rithms (e.g., regression, classification, clustering, etc.) that can be used for predicting
university admission.

Machine Learning Algorithms: It represents the algorithms used to build our

predictive model, which are Linear Regression (LR), Decision Tree (DT), Support
Vector Regression (SVR), and Random Forest Regression (RFR).

Evaluation: This step is a core part of building our Machine Learning model. There
are different metrics of evaluations that can be used. The evaluation metrics used in
this research work are Mean Square Error, Root Mean Square Error, and R-squared.

Deployment and Integration: It is all of the tasks that make our recommender
system RSEPUA available for use.

5 Proposed Recommender System

The recommender system proposed in this paper is articulated in two parts as shown
in Fig. 2. The first part is the User-based datamining that contains sample data.
The second part focuses on Machine Learning algorithms, that use different regres-
sion algorithms (i.e., Linear Regression, Decision Tree, Support Vector Regression,
and Random Forest Regression) to Build a predictive model in order to predict
students’ admission for higher education. It should be noted that the parameters used
in this study are GRE Score, TOEFL Score, University Rating, SOP, LOR, CGPA, and
Research Experience. After presenting the proposed recommender system RSEPUA,
in the next section we will present the discussion of the obtained results.
RSEPUA … 215

User Model

User-based datamining

Machine Learning Algorithm

Modeling Evaluation

GRE Score
Predictive Model TOEFL Score
University Rating
SOP
LOR
Prediction of Graduate
CGPA
Admission Research

Fig. 2 Proposed recommender system RSEPUA

6 Results and Discussion

In this present work, we used Supervised Machine Learning algorithms to build

a predictive model for early predicting university admission. Hence, the prediction
model is based on four regression algorithms which are Linear Regression (LR), Deci-
sion Tree (DT), Support Vector Regression (SVR), and Random Forest Regression
(RFR). All those algorithms implementation is available in the XLSTAT environment
[20].

6.1 Correlation Between the Parameters Profile

and the Chance of Admission

We used the Pearson Correlation Coefficient (PCC) [21] to evaluate the linear corre-
lation between the variables in our dataset. Indeed, this coefficient allows us to reduce
the number of features and determine the relationship between different parameters
profile (i.e., GRE Score, TOEFL Score, University Rating, SOP, LOR, CGPA, and
Research Experience) and the chance of admission. Table 2 illustrates the results
obtained for the Pearson Correlation Coefficients. Figure 3 illustrates a correlation
216 I. El Guabassi et al.

Table 2 Correlation matrix (Pearson)

Variables GRE score TOEFL University SOP LOR CGPA Research Chance of
score rating admit
GRE score 1 0,827 0,635 0,613 0,525 0,826 0,563 0,810
TOEFL 0,827 1 0,650 0,644 0,542 0,811 0,467 0,792
score
University 0,635 0,650 1 0,728 0,609 0,705 0,427 0,690
rating
SOP 0,613 0,644 0,728 1 0,664 0,712 0,408 0,684
LOR 0,525 0,542 0,609 0,664 1 0,637 0,373 0,645
CGPA 0,826 0,811 0,705 0,712 0,637 1 0,501 0,882
Research 0,563 0,467 0,427 0,408 0,373 0,501 1 0,546
Chance of 0,810 0,792 0,690 0,684 0,645 0,882 0,546 1
admit

Table 3 Evaluation of the

MSE RMSE R2
results
LR 0.003322947 0.057645010 0.82190074
DT 0.003992939 0.063189710 0.784975338
SVR 0.003636979 0.060307369 0.821388466
RFR 0.003004176 0.065606216 0.885856381

map of the table values. The results obtained demonstrate a high correlation between
CGPA, GRE score, and TOEFL score and the chance of admission.

6.2 Comparison of Machine Learning Algorithms

for Predicting University Admission

Table 3 represents a comparison of four Machine Learning algorithms namely

Linear Regression (LR), Decision Tree (DT), Support Vector Regression (SVR),
and Random Forest Regression (RFR) for predicting university admission. The eval-
uation metrics used in this comparison are Mean Square Error (MSE), Root Mean
Square Error (RMSE), and R-squared (R2 ).
As Table 2 shown, it is clear that Random Forest Regression (RFR) provides better
performance because it has a low MSE, low RMSE, and high R2 score.
RSEPUA … 217

40

35

Mean increase error 30

25

20

15

10

0
GRE Score University LOR Research
RaƟng
Variables

Fig. 3 Parameters influence the chance of admission

6.3 Most Important Parameters Influence the Chance

of Admission

After identifying the most suitable Machine Learning algorithm for predicting univer-
sity admission, it is therefore evident to ask which parameters have the most predictive
power. Indeed, the parameters with high importance are the engines of the prediction
and their values have a strong and significant impact on the outcome values. For
that reason, after training and validating a random forest, the variable importance
(VIMP) is calculated based on Mean increase error.
The normalized variable importance measure for each variable is represented in
Fig. 3. As we can see, the most important parameter is the Cumulative Grade Point
Average (CGPA). Hence, it is clear that there is a very strong link between the CGPA
and the chance of admission.

7 Conclusion and Future Work

In the 21st century, University education is becoming a key pillar of social and
economic life. In fact, it plays a major role not only in the educational process but
also in ensuring two important things which are the prosperous career and finan-
cial security. However, predicting university admission can be especially difficult
because the students are not aware of admission requirements. For that reason, the
218 I. El Guabassi et al.

main objective of this research work is to build a predictive model for early predicting
university admission. Thus, the contributions were threefold: The first was to apply
several Supervised Machine Learning algorithms (i.e., Linear Regression, Support
Vector Regression, Decision Tree Regression, and Random Forest Regression) on our
dataset. The second purpose was to compare and evaluate algorithms used to create a
predictive model based on various evaluation metrics. The last purpose was to deter-
mine the most important parameters that influence the chance of admission. The
experimental results showed that the Random Forest Regression is the most suitable
Machine Learning algorithm for predicting university admission. Also, the under-
graduate GPA (CGPA) is the most important parameter that influences the chance of
admission. The major directions for future work are: (i) Applying techniques such
as clustering and artificial neural networks to have better predicting. (ii) Utilizing
dataset with massive size and diverse features to tackle the issue of scalability. (iii)
Exploiting few hybrid feature selection algorithms.

References

1. Chui KT, Fung DCL, Lytras MD, Lam TM (2020) Predicting at-risk university students
in a virtual learning environment via a machine learning algorithm. Comput Hum Behav
107:105584
2. Qazdar A, Er-Raha B, Cherkaoui C, Mammass D (2019) A machine learning algorithm frame-
work for predicting students performance: a case study of baccalaureate students in Morocco.
Educ Inf Technol 24(6):3577–3589
3. El Guabassi I, Al Achhab M, Jellouli I, El Mohajir BE (2016) Recommender system for
ubiquitous learning based on decision tree. In: 2016 4th IEEE International Colloquium on
Information Science and Technology (CiSt), pp 535–540. IEEE
4. Guabassi IE, Achhab MA, Jellouli I, Mohajir BEE (2016) Towards adaptive ubiquitous learning
systems. Int J Knowl Learn 11(1):3–23
5. El Guabassi I, Al Achhab M, Jellouli I, Mohajir BEE (2018) Personalized ubiquitous learning
via an adaptive engine. Int J Emerg Technol Learn (iJET) 13(12):177–190
6. Bousalem Z, El Guabassi I, Cherti I (2018) Toward adaptive and reusable learning content
using XML dynamic labeling schemes and relational databases. In: International Conference
on Advanced Intelligent Systems for Sustainable Development, pp 787–799. Springer, Cham.
7. Wu X, Wu J (2019) Criteria evaluation and selection in non-native language MBA students
admission based on machine learning methods. J Ambient Intell Hum Comput 1–13
8. AlGhamdi A, Barsheed A, AlMshjary H, AlGhamdi H (2020) a machine learning approach
for graduate admission prediction. In Proceedings of the 2020 2nd international conference on
image, video and signal processing, pp 155–158
9. Nandal P (2020) Deep learning in diverse computing and network applications student
admission predictor using deep learning. SSRN 3562976
10. Zhao Y, Lackaye B, Dy JG, Brodley CE (2020) A quantitative machine learning approach to
master students admission for professional institutions. International Educational Data Mining
Society
11. Di Noia T, Mirizzi R, Ostuni VC, Romito D, Zanker M (2012) Linked open data to support
content-based recommender systems. In: Proceedings of the 8th international conference on
semantic systems, pp 1–8
12. Herlocker JL, Konstan JA, Terveen LG, Riedl JT (2004) Evaluating collaborative filtering
recommender systems. ACM Trans Inf Syst (TOIS) 22(1):5–53
RSEPUA … 219

13. Burke R (2000) Knowledge-based recommender systems. Encyclopedia of library and

information systems 69(Supplement 32):175–186
14. Montgomery DC, Peck EA, Vining GG (2012) Introduction to linear regression analysis, vol
821. Wiley, Hoboken
15. Safavian SR, Landgrebe D (1991) A survey of decision tree classifier methodology. IEEE Trans
Syst Man Cybern 21(3):660–674
16. Smola AJ, Schölkopf B (2004) A tutorial on support vector regression. Stat Comput 14(3):199–
222
17. Liaw A, Wiener M (2002) Classification and regression by randomForest. R news 2(3):18–22
18. Miles J R squared, adjusted R squared. Wiley StatsRef: Statistics Reference Online (2014)
19. Willmott CJ, Matsuura K (2005) Advantages of the mean absolute error (MAE) over the root
mean square error (RMSE) in assessing average model performance. Climate Res 30(1):79–82
20. Acharya MS, Armaan A, Antony AS (2019) A comparison of regression models for prediction
of graduate admissions. In: 2019 international conference on computational intelligence in data
science (ICCIDS), pp. 1–5. IEEE
21. Addinsoft X (2015) Data analysis and statistics with MS Excel. Addinsoft, NY, USA. xlstat
available at https://www.xlstat.com/en/home
22. Benesty J, Chen J, Huang Y, Cohen I (2009) Pearson correlation coefficient. In: Noise reduction
in speech processing, pp 1–4. Springer, Heidelberg
Forecasting Students’ Academic
Performance Using Different Regression
Algorithms

Inssaf El Guabassi, Zakaria Bousalem, Rim Marah, and Aimad Qazdar

Abstract In recent years, predicting student’s academic performance is the main

objective of all educational institutions. Numerous research works show that machine
learning can be a highly efficient technology to meet this objective. In this research
work, our first purpose is to compare several machine learning algorithms for
predicting students’ academic performance. The machine learning algorithms used
for comparison are ANCOVA, Logistic Regression, Support Vector Regression, Log-
linear Regression, Decision Tree Regression, Random Forest Regression, and Partial
Least Squares Regression. Our second purpose is to evaluate the seven algorithms
used using the various evaluation metrics. The experimental results showed that the
Log-linear Regression provides a better prediction, closely followed by ANCOVA.

Keywords Predictive model · Academic performance · Log-linear regression ·

Machine learning · Regression · Educational data mining

1 Introduction

In the real world, with a remarkable growth within the universe of measured data
warehouse sizes, analyzing the data in order to extract useful information is becoming
a necessity and a rich subject for several research projects [1]. Many application
areas adopt machine learning methods in their systems namely telecommunication,
shopping platforms, restaurants, economy, marketing, tourist targets, and medicine.
Over the last two decades, machine learning has entered the e-learning space as well
[2–5]. Thus, several machine learning algorithms have been exploited by researchers
to predict hidden patterns from educational settings [6–8]. The early prediction of

I. El Guabassi (B) · R. Marah

Faculty of Sciences, Tetouan, Morocco
Z. Bousalem
Faculty of Sciences and Technology, Settat, Morocco
A. Qazdar
ISI Laboratory FS Semlalia UCA, Marrakech, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 221
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_21
222 I. El Guabassi et al.

students at risk for academic failure is a crucial step in the learning process. Thus,
it must be made as soon as possible during the period of learning. The prediction of
student performance is necessary for:
• Providing a good quality education for students as well as for tutors;
• Reducing the number of dropouts;
• Increasing the rate of school completion;
• Improving educational outcomes for all students.
However, the real and major problems are: How to identify the “weak” students
who will need additional help to improve their performance? Also, which the suit-
able machine learning algorithm (i.e. predictive model) for predicting students’
academic performance? This research work evaluates and compares the effectiveness
of different machine learning algorithms. While there are many regression algorithms
for creating predictive models, this work concentrates on seven of them, which are
(1) ANCOVA, (2) Logistic Regression, (3) Support Vector Regression, (4) Log-
linear Regression, (5) Decision Tree Regression, (6) Random Forest Regression, and
(7) Partial Least Squares Regression. The outline of the present paper is as follows:
Sect. 2 contains recent studies regarding the specified area. The dataset used is briefly
described in Sect. 3. Section 4 concentrates on our methodology. Section 5 presents
the results and discussion. Finally, Sect. 6 presents the main conclusions considering
some future research directions.

2 Related Literature

In recent decades, many studies on predicting the performance of students by using

machine learning algorithms have been proposed.
Bravo-Agapito et al. [13] explained their study based on the prediction of
802 undergraduate student’s academic performance in e-learning. They used three
machine learning techniques, namely factor analysis, cluster analysis, and multiple
linear regressions. They concluded the “age” is a factor that affects the academic
achievement of the student. Gray and Perkins [14] conducted a study on predicting
student outcomes in the fourth week of the fall semester using several machine
learning methods. Hamsa et al. [15] applied two classification techniques which
are decision tree and fuzzy genetic algorithm to predict the student’s performance
both for the Bachelor and Master degree students. Their study focused on two disci-
plines which are Computer Science, and Electronics and Communication. Hussain
et al. [16] described a performance study on predicting student difficulties during
learning. They have also used machine learning algorithms such as artificial neural
networks, support vector machines, logistic regression, Naïve Bayes classifiers, and
decision trees. Their results show that artificial neural networks and support vector
machines are the most suitable algorithms to predict the performance of a student.
Karthikeyan et al. [17] investigated the performance of the students by developing
a hybrid educational data mining model called HEDM. Their model combines two
Forecasting Students’ Academic Performance … 223

techniques which are the J48 Classifier and Naive Baye’s classification. Their results
show that HEDM outperforms the results obtained in EDM.
In summary, many existing studies have made significant results in educational
data mining. However, most of them use classification methods for predicting
students’ academic performance. Moreover, there was very little focus on inter-
actional and parental involvement features.

3 Dataset

The data used for this work’s experimentation is collected from a first version of
the dataset named “Students’ Academic Performance Dataset (xAPI-Edu-Data)”
[18, 19]. It is, therefore, an open-source dataset available publicly on the Kaggle
dataset repository for academic and research purposes. The primary source of this
dataset used is from Elaf Abu Amrieh, Thair Hamtini, and Ibrahim Aljarah. There are
three main categories of the dataset features named “Demographic characteristics”,
“Academic background characteristics”, and “Behavioral characteristics”. Table 1
contains an overview of dataset features used for training and testing. It contains
three fields: feature, description, and type. It should be noted that there are two
major feature types, named “Nominal” and “Numeric”.

4 Methodology

Increasingly, E-learning has become an important tool for teaching and learning
around the world. Further, Learners have the opportunity to switch to distance
learning in various scientific areas anytime and anywhere [9, 20]. It is therefore
evident that many researchers work on the various aspects of e-learning [10–12].
The identification of the “weak” students and the factors affecting students’ academic
performance is a crucial step for successful learning. Hence, in the present paper, our
aim is evaluating student’s academic performance and identifying the factors that
influence academic performance using supervised machine learning algorithms.
This research work focuses on the following steps:
• Applying several machine learning algorithms which are ANCOVA, Logistic
Regression, Support Vector Regression, Log-linear Regression, Decision Tree
Regression, Random Forest Regression, and Partial Least Squares Regression.
• Comparing and evaluating machine learning algorithms for identifying which are
most suitable by using several evaluation metrics which are Mean Square Error
(MSE), Root Mean Square Error (RMSE), and R-squared (R2 ).
Many experiments were conducted in seven major steps depending on the regres-
sion methods namely ANCOVA, Logistic Regression (Logit-R), Support Vector
Regression (SVR), Log-linear Regression (Log-LR), Decision Tree Regression
224 I. El Guabassi et al.

Table 1. An overview of dataset features

Feature Description Type
Demographic features Gender Gender of Student (i.e., Male Nominal
or Female)
Nationality Nationality of Student (e.g., Nominal
Morocco, Kuwait, Lebanon,
Jordan, Egypt, USA, etc.)
Place of birth Country of birth for the Nominal
student (e.g., Morocco,
Kuwait, Lebanon, Jordan,
Egypt, USA, etc.)
Academic background Educational Stages The educational level of the Nominal
features student (i.e., Lowerlevel,
MiddleSchool or HighSchool)
Performance Levels The performance level of the Nominal
student
Section ID Classroom of the student (i.e., Nominal
A, B or C)
Topic Course topic (i.e., English, Nominal
Spanish, French, IT, Maths,
Chemistry, Biology, Quran, or
Geology)
Semester Semester of the year (i.e., Nominal
First or Second)
Behavioral features Parent responsible Parent responsible for student Nominal
(i.e., mother or father)
Raised hand Number of times the student Numeric
raised hand in the classroom
(i.e., from 0 to 100)
Visited resources Number of times the student Numeric
visited a course content (i.e.,
from 0 to 100)
Viewing announcements Number of times the student Numeric
checked the new
announcements (i.e., from 0 to
100)
Discussion groups Number of times the student Numeric
participated in discussion
groups (i.e., from 0 to 100)
Parent Answering Survey Parent answered the surveys Nominal
which are provided from
school or not (i.e., Yes or No)
Parent School Satisfaction The degree of parent Nominal
satisfaction for the school
(i.e., Yes or No)
Student Absence Days The number of absence days Nominal
for each student (i.e.,
above-7or under-7)
Forecasting Students’ Academic Performance … 225

Table 2. Evaluation of the

Value
ANCOVA algorithm
Mean Square Error (MSE) 0.157256464
Root Mean Square Error (RMSE) 0.396555752
R-squared (R2 ) 0.71890384

Table 3. Evaluation of the

Value
Logit-R algorithm
Mean Square Error (MSE) 0.156250000
Root Mean Square Error (RMSE) 0.395284708
R-squared (R2 ) 0.73799242

(DTR), Random Forest Regression (RFR), and Partial Least Squares Regression
(PLS-R). These regression methods were applied using the XLSTAT environment
[28]. In the following, the experimental result of each algorithm is presented.

4.1 Ancova

ANCOVA (ANalysis of COVAriance) [21] is used to compare the means of a variable-

response between two or more groups taking into account the variability of other
variables, called covariates. Thus, we decompose the total variance of the sample
into two partial variances, the inter-class variance, and the residual variance, and we
compare these two variances. Table 2 represents the evaluation metrics obtained after
applying the ANCOVA Algorithm on our dataset.
In basic terms, the R-Squared is the proportion of variability obtained by the
mathematical model compared to the total variability observed.

4.2 Logistic Regression (Logit-R)

Logistic Regression or Logit Regression (Logit-R) [22] is a statistical approach that

can be used to assess and characterize the relationships between a binary response
variable (e.g., wine/lose, alive/dead, pass/fail), and one, or more, explanatory vari-
ables, which can be categorical (e.g., nationality), or continuous numeric (e.g., note).
We evaluate our dataset using the Logistic Regression Algorithm. Table 3 represents
an overview of the results of the evaluation.
If R2 tends towards 1, it means that the cloud of points narrows around the
regression line.
226 I. El Guabassi et al.

Table 4. Evaluation of the

Value
SVR algorithm
Mean Square Error (MSE) 0.212447120
Root Mean Square Error (RMSE) 0.460919863
R-squared (R2 ) 0.6271547

Table 5. Evaluation of the

Value
Log-LR algorithm
Mean Square Error (MSE) 0.158611894
Root Mean Square Error (RMSE) 0.398261088
R-squared (R2 ) 0.71667276

4.3 Support Vector Regression (SVR)

Support Vector Regression (SVR) [23] is a binary classification algorithm. Just like
the Logistic Regression. If we take the image above, we have two classes (e.g.,
suppose these are e-mails, and Spam mails are in red and non-spam emails are in
blue).
The table above (Table 4) shows the evaluation metrics obtained after evaluating
the Support Vector Regression Algorithm. Mean Square Error (MSE) is the arithmetic
mean of the squares of the predictions between the model and the observations. This
is the value to be minimized in the context of a single or multiple regressions.

4.4 Log-Linear Regression (Log-LR)

The log-linear regression (Log-LR) [24] is a prediction model that applies when
the target variable Y is a counting variable (number of occurrence of events during
a lap of time). Table 5 contains the evaluation metrics obtained after applying the
Log-linear regression Algorithm.
Root Mean Square Error (RMSE) represents a standard way to measure the error
in the predictive model. It is the square root of the mean of the square of all of the
errors.

4.5 Decision Tree Regression (DTR)

The decision trees (DTR) [25] are classification rules that base their decision on a
suite of tests associated with attributes, the tests being organized in a tree structure.
We evaluate our dataset using the Decision Tree Algorithm. Table 6 represents an
overview of the results of the evaluation.
Forecasting Students’ Academic Performance … 227

Table 6. Evaluation of the

Value
DTR algorithm
Mean Square Error (MSE) 0.195293449
Root Mean Square Error (RMSE) 0.441920184
R-squared (R2 ) 0.65025193

Table 7. Evaluation of the

Value
RFR algorithm
Mean Square Error (MSE) 0.171994444
Root Mean Square Error (RMSE) 0.414722128
R-squared (R2 ) 0.69480482

Table 8. Evaluation of the

Value
PLS-R algorithm
Mean Square Error (MSE) 0.205659323
Root Mean Square Error (RMSE) 0.453496773
R-squared (R2 ) 0.63238366

4.6 Random Forest Regression (RFR)

Random Forest Regression (RFR) [26] is a supervised learning algorithm that

combines multiple predictions to make a more accurate prediction than a single
model (Table 7).
The table above represents the evaluation metrics obtained after applying the
Random Forest Regression Algorithm.

4.7 Partial Least Squares Regression(PLS-R)

Partial Least Squares Regression (PLS-R) [27] is a flexible statistical technique appli-
cable to any form of data. It allows to model the relationships between inputs and
outputs, even when the inputs are correlated and noisy, the outputs multiple and the
inputs more numerous than the observations.
Our predictive model based on Decision Trees is evaluated using several evalu-
ation metrics such as Mean Square Error, Root Mean Square Error, and R-squared
(see Table 8).
228 I. El Guabassi et al.

Table 9. Experimental results

MSE RMSE R2
ANCOVA 0.157256464 0.396555752 0.71890384
Logit-R 0.156250000 0.395284708 0.73799242
SVR 0.212447120 0.460919863 0.6271547
Log-LR 0.158611894 0.398261088 0.71667276
DTR 0.195293449 0.441920184 0.65025193
RFR 0.171994444 0.414722128 0.69480482
PLS-R 0.205659323 0.453496773 0.63238366

5 Results and Discussion

The table below, therefore, represents summary results for the seven algorithms used
in this research work. The evaluation metrics used in this experiment are Mean Square
Error (MSE), Root Mean Square Error (RMSE), and R-squared (R2 ). It should be
noted that RMSE is just the square root of the MSE (Table 9).
After rigorously evaluating all the seven algorithms on the 480 students of our
dataset, we compare the performances in order to determine the most suitable predic-
tive model. According to the experimental results, it is clear that Log-linear Regres-
sion (Log-LR) provides better performance because it has a low MSE, low RMSE,
and high R2 score. On the other hand, we observed that Support Vector Regression
(SVR) isn’t suitable for predicting students’ academic performance because it has a
high MSE, high RMSE, and low R2 score.
Given the R2 = 73% of the variability of the dependent variable, Class is explained
by the 16 explanatory variables. The remainder of the variability is due to other
explanatory variables that have not been considered during the present experiment
research.
The chart below (see Fig. 1) indicates the predicted values versus the observed
values. Also, Confidence intervals for the mean allow the detection of potential
outliers.
The histogram below (see Fig. 2) represents the standardized residuals versus the
performance. It indicates that the residuals grow with the Performance. As we can
see in Fig. 2 the residuals bar chart allows to quickly showing the residuals that are
out of the range [2, 2].
Forecasting Students’ Academic Performance … 229

Pred(Performance) - Performance
5

3
Performance

0
-1 0 1 2 3 4 5

-1
Pred(Performance)

Fig. 1. Predicted values versus the observed values

Standardized resudials/Performance

Obs457
Obs433
Obs409
Obs385
Obs361
Obs337
Obs313
Observations

Obs289
Obs265
Obs241
Obs217
Obs193
Obs169
Obs145
Obs121
Obs97
Obs73
Obs49
Obs25
Obs1
-4 -2 0 2 4
Standardized resudials

Fig. 2. Standardized residuals versus the performance

230 I. El Guabassi et al.

6 Conclusion and Future Work

In recent years, the world’s population is increasingly demanding to predict the future
with certainty, predicting the right information in any area is becoming a necessity.
One of the ways to predict the future with certainty is to determine the possible
futures. In this sense, machine learning is a way to analyze huge datasets in order to
make strong predictions or decisions. In this research work, our main purpose was
to compare several machine learning algorithms for predicting student’s academic
performance. Certainly, this research work has some limitations. That’s why the
major directions for future work could focus on the following: Firstly, applying
techniques such as clustering and artificial neural networks to have better predicting.
Secondly, utilizing dataset with massive size and diverse features to tackle the issue
of scalability. The final area that can be improved is exploiting few hybrid feature
selection algorithms.

References

1. Kalaivani S, Priyadharshini B, Nalini BS (2017) Analyzing student’s academic performance

based on data mining approach. Int J Innov Res Comput Sci Technol 5(1):194–197
2. Moubayed A, Injadat M, Shami A, Lutfiyya H (2020) Student engagement level in e-learning
environment: clustering using k-means. Am J Distance Educ 1–20
3. Alenezi HS, Faisal MH (2020) Utilizing crowdsourcing and machine learning in education:
literature review. Educ Inf Technol 1–16
4. El Guabassi I, Al Achhab M, Jellouli I, El Mohajir BE (2016, October) Recommender system
for ubiquitous learning based on decision tree. In: 2016 4th IEEE international colloquium on
information science and technology (CiSt), pp 535–540. IEEE
5. Hew KF, Hu X, Qiao C, Tang Y (2020) What predicts student satisfaction with MOOCs: a
gradient boosting trees supervised machine learning and sentiment analysis approach. Comput
Educ 145:103724
6. Qazdar A, Er-Raha B, Cherkaoui C, Mammass D (2019) A machine learning algorithm frame-
work for predicting students performance: a case study of baccalaureate students in Morocco.
Educ Infn Technol 24(6):3577–3589
7. Huang AY, Lu OH, Huang JC, Yin CJ, Yang SJ (2020) Predicting students’ academic perfor-
mance by using educational big data and learning analytics: evaluation of classification methods
and learning logs. Interact Learn Environ 28(2):206–230
8. Waheed H, Hassan SU, Aljohani NR, Hardman J, Alelyani S, Nawaz R (2020) Predicting
academic performance of students from VLE big data using deep learning models. Comput
Hum Behav 104:106189
9. El Guabassi I, Al Achhab M, Jellouli I, Mohajir BEE (2018) Personalized ubiquitous learning
via an adaptive engine. Int J Emerg Technol Learn (iJET) 13(12):177–190
10. Syed AM, Ahmad S, Alaraifi A, Rafi W (2020) Identification of operational risks impeding
the implementation of eLearning in higher education system. Educ Inf Technol 1–17
11. Bousalem Z, El Guabassi I, Cherti I (2018, July) Toward adaptive and reusable learning content
using XML dynamic labeling schemes and relational databases. In: International conference
on advanced intelligent systems for sustainable development, pp 787–799. Springer, Cham
12. El Guabassi I, Bousalem Z, Al Achhab M, EL Mohajir BE (2019) Identifying learning style
through eye tracking technology in adaptive learning systems. Int J Electr Comput Eng (2088–
8708):9
Forecasting Students’ Academic Performance … 231

13. Bravo-Agapito J, Romero SJ, Pamplona S (2020) Early prediction of undergraduate student’s
academic performance in completely online learning: a five-year study. Comput Human Behav
106595
14. Gray CC, Perkins D (2019) Utilizing early engagement and machine learning to predict student
outcomes. Comput Educ 131:22–32
15. Hamsa H, Indiradevi S, Kizhakkethottam JJ (2016) Student academic performance prediction
model using decision tree and fuzzy genetic algorithm. Proc Technol 25:326–332
16. Hussain M, Zhu W, Zhang W, Abidi SMR, Ali S (2019) Using machine learning to predict
student difficulties from learning session data. Artif Intell Rev 52(1):381–407
17. Karthikeyan VG, Thangaraj P, Karthik S (2020) Towards developing hybrid educational data
mining model (HEDM) for efficient and accurate student performance evaluation. Soft Comput
1–11
18. Amrieh EA, Hamtini T, Aljarah I (2016) Mining educational data to predict student’s academic
performance using ensemble methods. Int J Database Theory Appl 9(8):119–136
19. Amrieh EA, Hamtini T, Aljarah I (2015, November) Preprocessing and analyzing educational
data set using X-API for improving student’s performance. In: 2015 IEEE Jordan conference
on applied electrical engineering and computing technologies (AEECT), pp. 1–5. IEEE
20. El Guabassi I, Bousalem Z, Al Achhab M, Jellouli I, Mohajir BEE (2018) Personalized adaptive
content system for context-aware ubiquitous learning. Proc Comput Sci 127:444-453
21. Rutherford A (2001) Introducing ANOVA and ANCOVA: a GLM approach Sage
22. Kleinbaum DG, Dietz K, Gail M, Klein M, Klein M (2002) Logistic regression. Springer, New
York
23. Smola AJ, Schölkopf B (2004) A tutorial on support vector regression. Stat Comput 14(3):199–
222
24. Heien DM (1968) A note on log-linear regression. J Am Stat Assoc 63(323):1034–1038
25. Safavian SR, Landgrebe D (1991) A survey of decision tree classifier methodology. IEEE Trans
Syst Man Cybern 21(3):660–674
26. Liaw A, Wiener M (2002) Classification and regression by randomForest. R news 2(3):18–22
27. Geladi P, Kowalski BR (1986) Partial least-squares regression: a tutorial. Anal Chim Acta
185:1–17
28. Addinsoft X (2015) Data analysis and statistics with MS Excel. Addinsoft, NY, USA. xlstat
available at https://www.xlstat.com/en/home
Computational Analysis of Human
Navigation Trajectories in a Spatial
Memory Locomotor Task

Ihababdelbasset Annaki, Mohammed Rahmoune, Mohammed Bourhaleb,

Jamal Berrich, Mohamed Zaoui, Alexander Castilla Ferro,
and Alain Berthoz

Abstract In this paper, we use computational tools (Cipresso P, Matic, A, Giakoumis

D, Ostrovsky Y (2015) Advances in computational psychometrics. Comput Math
Methods Med. Article ID 418683. https://doi.org/10.1155/2015/418683.5) to explore
human navigation through an example of a visuomotor spatial memory locomotor
task, the Walking Corsi task (WCT) variant from a well-known table test known as the
Corsi Block Tapping task [(CBT) [2] and [15]. This variant was performed using the
“Virtual Carpet” ™ experimental setup. The subjects had to memorize a succession
of the position of targets projected on the ground and reproduce sequences of 2 to 9
targets by walking to each. The trajectory of the head was recorded and processed
from a kinematic point of view. Generic tools that computational data analytics
provides and through computer simulations by replicating visually this data allowed
categorization of the different features of the behavior of the subjects providing a
new powerful tool for both normal and pathological behavior characterization.

Keywords Data analytics · Computer simulation · Virtual reality · Artificial

intelligence · Human navigation · Corsi block tapping test

I. Annaki (B) · M. Rahmoune · M. Bourhaleb · J. Berrich

Embedded Systems, Electronics, and IT Team, Research Laboratory in Applied Sciences,
National School of Applied Sciences, PB 669, 60000 Oujda, Morocco
e-mail: [email protected]
M. Rahmoune
e-mail: [email protected]
M. Zaoui · A. C. Ferro · A. Berthoz
Collège de France, CIRB, 11, Place Marcelin-Berthelot, 75231 Paris Cedex 05, France
Institut de Médecine Expérimentale (IME), Paris, France

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 233
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_22
234 I. Annaki et al.

1 Introduction

Among the most fundamental human tasks is to navigate from an initial position to
a destination. This process is performed several times a day by each of us. In fact,
navigation is characterized as organized and goal-oriented environmental movements
[3]. This includes the physical act of action and the cognitive dimensions of decision-
making and an itinerary follow-up. These two navigational elements are known as
locomotion, spatial orientation, and navigation [4].
To assess individuals’ efficiency at the highest level of locomotive control, i.e.,
navigation [5]. Neuropsychologists use a locomotor version of a table memory task
called the Corsi Block Tapping task (CBT) [15]. They use the Walking Corsi Task
(WCT) and its variants [2] and [6], which is closely related to navigation. The
performance data resulting from the WCT has been analyzed previously in studies
concerning the pathological deficits in adults and children, but the present work brings
new methods for a Kinematic analysis of the trajectories and mental manipulation
of viewpoint changes using [7].
The previous analysis of trajectories that a human may take to reach a distant
position and turn in a given direction revealed that the formation of trajectories
during goal-oriented locomotion in humans is, in practice, stereometric and kinemat-
ically stereotypical [8, 9]. Furthermore, robotics engineers and researchers are more
concerned with implementing pathfinding algorithms that mimic human behavior
and can also contribute to optimizations based on various parameters, such as time
and space [10].
In neurosciences however classical statistical approaches have been used to test
differences in behavior and navigation strategies between individuals and groups.
This has prompted our attempt to develop more effective methods to help clinicians
in sorting out normal subjects and patients or for the study of the development of
sensory-motor and cognitive abilities [11].
From this perspective, the present paper aims to explore the extracted data
outcomes of the WCT test from a kinematic point of view with data analytics
using the potential of python and computer simulations developed with Unity. It
represents, a new effort toward integrating in-depth artificial intelligence, computer
vision, and deep learning approaches, and stepping into computer simulations by
taking advantage of the capacity of virtual reality. It is a specific contribution to
the wider developing field using information contained in experimental data and
combining information from different sources and different nature and coping with
a large amount of data (Big-Data Problem) [12]. Also, the effort here to generate a
synthetic data analysis based on real samples provided from clinical trials using data
augmentation models [13]. It also is applied to the virtual environment for simulations
or enhancements of the neuropsychological assessment [14].
Computational Analysis of Human Navigation Trajectories … 235

2 Materials and Method

2.1 Neuropsychological Assessments

The Corsi visuospatial memory protocol was used in a series of studies using different
technologies. These protocols were all inspired by the Corsi Table Block Tapping test
(CBT). A locomotor version of this table test was developed using simple targets.
Then two new technologies were developed the “Magic Carpet” and the “Virtual
Carpet”. They allowed more complex measurements of the behavior and more flexi-
bility in the presentation of the targets and kinematic measurements of the trajectories
and behavior of the subjects.

Corsi Block-Tapping Test (CBT)

The CBT [15], is a table spatial memory visuo-manual test. it consists of nine blocks
(4.5 × 4.5 cm) mounted to the baseboard (30 × 25 cm) in a dispersed series. Both
short-and long-term visual-spatial memory has been tested. The experimenter taps
with a finger a series of blocks at a rate of one block per 2 s, in which the subject
must tap the block sequence in the same order as it was introduced. Block sequences
progressively increase in length (starting with a 2-block series). The score is the
number of blocks correctly recalled in the most extended sequence (“block span”).

The Walking Corsi Test (WCT)

The WCT [16, 17], a locomotor version of the CBT (3 × 2.5 m; scale 1:10 of the
CBT) developed by C. Guariglia and Laura Piccardi in Rome. It was mounted in
an empty room. In this test, the participant must walk to visit various locations.
The WCT consists of nine squares placed on a carpet. Both the examiner and the
participants began from the same origin. The experimenter demonstrated the series
by walking on the carpet and stopping for 2 s on each rectangle. The subject had to
repeat the same pattern as the experimenter by walking and stopping on the squares
in the way.

The Magic Carpet (E-WCT)

An electronic version of the WCT was created in Paris by the Berthoz Lab [18]. The
protocol was the same as in the WCT, but the technology differed. The equipment
was called the “Magic Carpet”. It has the same dimensions as the WCT (2.50 × 3 m).
But nine electronically controlled plastic white tiles (30 × 30 cm) were mounted on a
carpet and constituted targets for the Corsi protocol. It was named the “Magic Carpet
Test” (MCT). Every tile was 10 mm thick and has had a luminous white surface of
75 mm at the top and six pressure sensors, placed under the surface. The tiles were
connected to an electronic computer-controlled system that turned the luminosity of
the tiles on or off and sensed the fact that a person was walking on the tiles. The
computer managed each sequence at a rate of 2 s per tile and record the participant’s
236 I. Annaki et al.

output—specifically, the number of correct tiles achieved in the right order. The
series to be memorized is presented by illuminating the tiles in series consecutively
at one tile per 2 s (each tile lights up for 2 s and then shuts off before the next tile
is lit). At the end of the experiment, the auditory alarm signals to the participant
that the pattern detected while walking must be repeated. This Magic Carpet Corsi
protocol was used for several studies on normal and pathological adults and children
and the processing of correct target acquisitions, delays, and several kinds of errors
were computed.

The Virtual Carpet™

A new version of the carpet was designed by Mohamed Zaoui and used by the Berthoz
lab. [19] for the application of the Walking Corsi protocol.

Projection of Virtual Tiles

Instead of real tiles on the ground virtual white tiles (10 cm square) were projected on
the ground by a video projector. This was done using a PowerPoint image program
allowing great flexibility in the organization of the series of tiles illuminated. This
created a virtual environment using the Unity Game Engine and replicating the WCT
with 2 to 9 target sequences.
Programs of control of the image sequential appearance were built with open-
source software. Blender. The protocol repeated the WCT protocol. They were
designed in the Blender Game Engine to control motions and capture experimental
data. created a virtual environment using the Unity Game Engine.

Motion Capture of Head Trajectories

The trajectories of the head during locomotion to the target were recorded using a
virtual reality device: The HTC VIVE VR-Headset fitted with the SteamVR tracking
system. These included a sensor placed on the head of the subject to measure head
position, direction, and velocity in the horizontal plane. The experimental station
consists of two synchronized computers for the generation of images and data acqui-
sition. The protocol included a calibration of the target’s positions and the subject’s
motions with precision around 1 cm.

2.2 Experimental Data

The batch contains a calibration file that sets the coordinates of the targets relative to
the current subject under focus, and each corresponding folder contains the derived
data sets relative to the trajectory and locomotion.
Computational Analysis of Human Navigation Trajectories … 237

2.3 Participants

The data collected were provided without further detail on the subjects, from gender
and age to actual status so that the analysis would not have any a priori bias. However,
this sample was subject to the usual protocols, and participants were made aware that
their data would be analyzed, and they could have access to the results. Besides, the
details have been obscured to make the data more accessible and to examine it
extensively without any background restrictions.

2.4 Visual Replication

Using Matplotlib, a robust library for developing static, interactive, and immersive
Python visualizations and based on our theoretical claims. We visually repeated the
sequence on which our findings were based. The reproduction of the trajectory and
the targets with animation allowed us to understand the kinematic behavior of the
subjects during each session (Fig. 1).
This plot represents the experimental array seen from above each target is shown
as a square. The departure point is on the right. The subject had to go to two targets
and return. Head trajectory is shown, and the blue circles indicate that the subject
stayed successfully on both targets.

Fig. 1 Example of reproduction of subject trajectory

238 I. Annaki et al.

2.5 Target Detection Algorithm

Several computations have been carried out on the data to establish a firm inter-
pretation and find criteria to help us draw statements about the various subjects.
We developed an algorithm to detect the various targets visited. This was done by
computing the head positions during the trajectory and from this set of numbers
calculate a walking velocity along the trajectory (tangential velocity). To obtain the
information that the subject had reached a target we compare the position data with
a target calibration field in which the X and Y components of the nine targets were
given. This allowed us to also compute the duration in seconds during which the
subject had stopped on the target. This file implements nine target positions, as each
dataset in the user folder reflects a series of scenarios.
Modern techniques sallow computational approaches of these data [19] and [20].
Using “Pandas” a data analysis and manipulation tool library, in particular the concept
of DataFrame, each file is stored as a data frame to avoid any conflict between the
file types. Each target is also flagged to be able to settle on the series assigned to
each subject. The crucial step in the algorithm is to compare the distance between
the positions of the subject on the trajectory and the location of the target extracted
from the calibration file, to be able to determine whether the target was visited by
the subject during the session, and to track the kinematic behavior until the subject is
close to the target, or until the entire session is over. This potentially can also indicate
if the subject missed a target.
Using all these data we could classify the different subjects according to the values
of velocity, and time on various targets. We also distinguished the behavior when the
trajectory included only 2 targets or more up to Nine which is a much more difficult
task for instance in the Table Block test a span of 7 is generally the maximum that
most normal subjects can memorize.

3 Results

In this section, we provide a summary of the production classifications, concentrating,

as an example, on one batch the others have been processed and evaluated similarly.

3.1 Classification of Subjects Based on the Length

of the Sequence

We took as a criterion that a subject had visited a target if his position was within a
confidence circle of 30 × 30 cm, around the value given by the calibration file. A
velocity slowdown was also taken as a criterion of target reaching. A first numerical
calculation has allowed us to identify a classification of experiments according to
Computational Analysis of Human Navigation Trajectories … 239

the number of targets visited by the subjects. The result is a classification of the
data according to the sequence’s length (number of visited targets). It indicates four
classes. These classes referred to reflect the distribution of subjects according to the
series correctly replicated by the subject. Specifically, we find classes in which the
participant repeats situations of series lengths of 2, 3, 4, and 5. As a side effect, more
classes were recognized, however, the number of subjects in these classes does not
allow us to extract significant information.

3.2 Overall General Classification Based on Average Speed

and Time Spent Within the Target

We aggregate the subject’s data to retrieve their average speed along the entire trajec-
tory and the time, they remain on a target before leaving the confidence area as
mentioned in the target detection algorithm. The global analysis allows us to evaluate
the performance of a subject throughout the whole process and determine aggregates
related to speed. These aggregates are a comparison threshold allowing us to classify
the subjects into 3 categories:
– The fastest subjects.
– The moderately fast subjects.
– The slowest subjects.
The aggregate of the global average time that the subject spends within a
target allows us to compare the coherence between speed and time and allows us
to identify anomalies and classify the subjects into 2 categories:
– Subjects spending important time within a target.
– Subjects moderately spending below or closer time to the global average of time
spent within a target (Fig. 2).

3.3 A Detailed Classification Based on Average Speed

and Time Spent Within the Target for Each Class
Sequence Length

The same approach is used for general analysis. For each class, we aggregate the data
of the subject to extract its average speed and the time spent in the target. Every class
helps us to calculate the class’s speed and time spent within the target aggregates to
extract the same categories as in the general classification.
The detailed analysis allows us to identify subject participation and attendance at
the level of each class and to refine the overall analysis carried out (Fig. 3).
240 I. Annaki et al.

Fig. 2 General Combination Chart representation of average speed as a red line chart and time
spent on the target as a blue-sky bar chart. The green reference line represents the overall average
speed (All subjects), the red one is the average speed of slow subjects (the qualification of slow
subjects is based on the subject’s speed below the overall average speed) and the blue line identifies
the average speed of fast subjects (as cited for the slow subjects, the fast ones are those who got a
speed over the overall average speed)

Fig. 3 Example: Class 3 combination chart representation of average speed as a red line chart
and time spent on the target as a blue-sky bar chart. The green reference line represents the class
average speed (All subjects attending the scenarios with 3 targets visited), the red is the average
speed of slow subjects (the qualification of slow subjects is based on the subject’s speed below the
class average speed) and the blue identifies the average speed of fast subjects (as cited for the slow
subjects, the fast ones are those who got a speed over the class average speed)
Computational Analysis of Human Navigation Trajectories … 241

4 Discussion

The purpose of this research was to identify the different behaviors of subjects who
performed the Walking Corsi Protocol on the Virtual Carpet set up using a kinematic
approach instead of basing the research on pre-established scoring strategies in the
literature. Commonly, scoring is based on the maximum length of the series that the
subject can achieve correctly. Here we characterized the behavior and performance
used the subject using average speed to complete the sequences and the time spent
on the targets before heading to another goal. Moreover, we were able to identify
some incoherence between the subject’s pace and the spent time. Further studies are
needed to generalize this approach and analyze its efficiency and accuracy.
This is a very global approach that will complement other data analysis of the
behavior, but it can be very useful to provide criteria that give us insights about
the subjects who can repeat the sequence easily, and other subjects who find it
challenging, particularly when the sequence is getting longer or requires mental
rotations, etc. None of the existing publications on these tests has brought this tool
and it is worth trying to use it in young children or with patients or elderly persons
and eventually it may be very useful to measure quantitatively learning or functional
recuperation during rehabilitation.

5 Conclusion

The current paper provides the first attempt at characterizing the behavior of human
subjects in a complex visuospatial task used to test memory and navigation capabili-
ties. The aim is to integrate emerging technology into the psychometric methods that
will give neuropsychologists greater insight into new technology, such as big data,
machine learning, and computer vision [20]. Knowing that psychometric clinical
trials require a significant amount of time and planning, it can take months to years
to collect data. Then data science can offer new approaches to generate synthetic
data based on the samples extracted from the psychometric tests. Neuropsycholog-
ical assessments are valuable instruments for telling cognitive impairment diagnosis.
However, decoding neuropsychological assessments requires experts and is often
time-consuming. To streamline the use of psychometric evaluations in clinical envi-
ronments, machine learning algorithms are being built and tested using multi-center
neuropsychological data [21].

Acknowledgments We thank Dr. Bernard Cohen, Paris France, to let us use the preliminary test
data obtained in cooperation with him to test the present method to evaluate its validity for the use
of patient assessments.
242 I. Annaki et al.

References:

1. Cipresso P, Matic, A, Giakoumis D, Ostrovsky Y (2015) Advances in computational psycho-

metrics. Comput Math Methods Med. Article ID 418683. https://doi.org/10.1155/2015/418
683.5
2. Tedesco M, Bianchini F, Piccardi L, Clausi S, Berthoz A, Molinari M, Guariglia C, Maria
L (2017) Does the cerebellum contribute to human navigation by processing sequential
information? Neuropsychology 31. https://doi.org/10.1037/neu0000354
3. Montello D, Sas C (2006) Human factors of wayfinding in navigation. International encyclo-
pedia of ergonomics and human factors. https://doi.org/10.1201/9780849375477.ch394
4. Irmischer I, Clarke K (2018) Measuring and modeling the speed of human navigation.
Cartogrph Geogr Inf Sci 45:177–186. https://doi.org/10.1080/15230406.2017.1292150
5. Belmonti V, Cioni G, Berthoz A (2016) Anticipatory control and spatial cognition in locomotion
and navigation through typical development and in cerebral palsy. Dev Med Child Neurol
58(Suppl 4):22–27. https://doi.org/10.1111/dmcn.13044
6. Corsi PM (1998) Human memory and the medial temporal region of the brain (Ph.D.) McGill
University (1972). Berch DB, Krikorian R, Huha EM (1998) The Corsi block-tapping task:
methodological and theoretical considerations. Brain Cogn 38(3):317–338. https://doi.org/10.
1006/brcg.1998.1039
7. Meilinger T, Berthoz A, Wiener JM (2011) The integration of spatial information across
different viewpoints. Memory Cogn 39:1042–1054. https://doi.org/10.3758/s134210110088-x
8. Hicheur H, Pham QC, Arechavaleta G, Laumond JP, Berthoz A (2007) The formation of
trajectories during goal-oriented locomotion in humans. I. Stereotyped behavior. Eur J Neurosci
26(8):2376–2390. https://doi.org/10.1111/j.14609568.2007.05836.x
9. Pham QC, Hicheur H, Arechavaleta G, Laumond JP, Berthoz A (2007) The formation of
trajectories during goal-oriented locomotion in humans. II. A maximum smoothness model.
Eur J Neurosci 26(8):2391–2403. https://doi.org/10.1111/j.14609568.2007.05835.x
10. Pham Q (2014) A general, fast, and robust implementation of the time-optimal path param-
eterization algorithm. In: IEEE Trans Robot 30(6):1533–1540. https://doi.org/10.1109/TRO.
2014.2351113
11. Kang MJ, Kim SY, Na DL et al. (2019) Prediction of cognitive impairment via deep learning
trained with multi-center neuropsychological test data. BMC Med Inform Decis Mak 19:231.
https://doi.org/10.1186/s12911-019-0974-x
12. Gorriz J, Ramírez J, Ortiz A, Martínez-Murcia, F, Segovia F, Suckling J, Leming M, Zhang
Y-D, Álvarez-Sánchez J, Bologna G, Bonomini M, Casado F, Charte D, Charte F, Contreras
R, Cuesta-Infante A, Duro R, Fernández-Caballero A, Fernandez E, Ferrández J (2020) Arti-
ficial intelligence within the interplay between natural and artificial Computation: advances
in data science, trends and applications. Neurocomputing. https://doi.org/10.1016/j.neucom.
2020.05.078
13. What Is Synthetic Data? | Unite.AI. https://www.unite.ai/what-is-synthetic-data/
14. Cipresso P, Serino S, Riva G (2016) Psychometric assessment and behavioral experiments
using a free virtual reality platform and computational science. BMC medical informatics and
decision making 16:37. https://doi.org/10.1186/s12911-016-0276-5
15. Corsi PM (1972) Human memory and the medial temporal region of the brain (Ph.D.). McGill
University
16. Piccardi L, Iaria G, Ricci M, Bianchini F, Zompanti L, Guariglia C (2008) Walking in the Corsi
test: which type of memory do you need? Neurosci Lett 432:127–131. https://doi.org/10.1016/
j.neulet.12.044
17. Piccardi L, Leonzi M, D’Amico S, Marano A, Guariglia C (2014) Development of navigational
working memory: evidence from 6- to 10-year-old children. Br J Dev Psychol 32:205–217.
https://doi.org/10.1111/bjdp.12036
18. Perrochon A, Kemoun G, Dugué B, Berthoz A(2014) Cognitive impairment assessment through
visuospatial memory can be performed with a modified walking Corsi test using the ‘Magic
Carpet’. Dementia Geriatric Cogn Disord Extra 4:1–13. https://doi.org/10.1159/000356727
Computational Analysis of Human Navigation Trajectories … 243

19. Berthoz A, Zaou, M (2015) New paradigms and tests for evaluating and remediating visuospa-
tial deficits in children. Dev Med Child Neurol 57(Suppl 2):15–20. https://doi.org/10.1111/
dmcn.12690
20. Elgendy N, Elragal A (2016) Big data analytics in support of the decision-making process.
Procedia Comput Sci 100:1071–1084. https://doi.org/10.1016/j.procs.2016.09.251
21. Kang MJ, Kim SY, Na DL et al. (2019) Prediction of cognitive impairment via deep learning
trained with multi-center neuropsychological test data. BMC Med Inform Decis Mak 19:231.
https://doi.org/10.1186/s12911-019-0974-x.
An Electrocardiogram Data
Compression-Decompression Technique
Based on the Integration Filtering
of Peaks’ Amplitude Signal

Skander Bensegueni

Abstract In this paper, a novel electrocardiogram data compression/decompression

technique is presented, based on transforming the electrocardiogram record to peaks’
amplitude signal. The proposed method reduces the number of non-zero amplitude of
the signal and thus reduce the data dimension. At the decompression step and in order
to recover the signal, an integration filtering is applied in aim to link the non-zero
amplitudes together and reconstruct the signal. This technique was applied on the
48 records of the MIT-BIH arrhythmia database and its performance was evaluated
through calculation of metrics such as the compression ratio and the mean-square-
error difference percentage. The average values of these last two metrics were equal
respectively to 17.83 and 0.62, indicating that the proposed method reduces the
dimension of the signal well by preserving its waveforms and segments.

Keywords Electrocardiogram data · Compression-decompression · Integration

filter · Peak’ amplitude signal

1 Introduction

The electrocardiogram (ECG) data are the result of the electric potential recording of
the myocardium. It’s used to monitor the heart’s health. The need for long duration
recording or permanent monitoring of the electrocardiogram signal is gaining more
and more attention for its ability to record pathological events appearing suddenly.
The compression of ECG signal is very important in health monitoring, medical diag-
nosis aid and decision systems for patients with cardiovascular diseases. By reducing
the size of the cardiac activity record, it will be easy to store and transmit data, which
allow the analysis and processing of this data distantly and by automatic tools.
The construction of an efficient ECG telemonitoring system can be achieved
by reducing the amount of transmitted data and minimizing the computational

S. Bensegueni (B)
National Polytechnic School, BP 75, A, Nouvelle Ville RP, Constantine, Algeria

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 245
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_23
246 S. Bensegueni

complexity [1]. In the paper of Polanía et al. [2], a reducing dimensionality tech-
nique was proposed by exploiting the representational power of restricted Boltz-
mann machines (RBMs) to model the probability distribution of the sparsity pattern
of transmitted ECG signals. Several ECG data compression techniques have also
been proposed in the literature. Techniques using wavelets give good performances;
such as the dimensionality reduction approach applied in [3] by selecting elementary
components from redundant wavelet dictionaries via a greedy strategy. The statis-
tical methods also were used to compress ECG data as in [4] by varying discrete
cosine transform thresholds applied on different principal directions of the ECG
data. A sparse representation-based compression algorithm was proposed by Grossi
et al. [5] through exploiting the repetition of common patterns of ECG segments as
natural basis.
The proposed compression algorithm is based on the application of transformation
of the ECG signal into peaks’ amplitudes signal which will contain a reduced number
of non-zeros amplitudes to be transmitted. This paper is divided into several sections;
starting from the development of the new theoretical basics as: the slop change
coefficients and the integrating filter, going through the application of these concepts
to compress and decompress the electrocardiogram signal, and finishing by results
and conclusions.

2 Theory

2.1 Slop Change Coefficients

We calculate the slop change (SC) coefficients by using (1) in order to decom-
pose the signal x(t) into coefficients characterized by their sensitivity to steepness
changes where lower slopes have coefficients’ absolute values lower than those of
sharper slopes. The SC coefficients are constructed by calculating the integral of the
combination of two scale signals with opposite signs [6].

k k+1/2
SC(k) = − ∫ x(t − k)dt + ∫ x(t − k)dt (1)
k−1/2 k

However, the sign of the coefficients is inversely dependent on slopes’ direction. A

minimum of the signal is represented by a positive-negative couple of SC coefficients,
and a maximum is represented by a negative-positive couple of SC coefficients (see
Fig. 1.b).
The SC coefficients should be sub-sampled by 2. So, in order to recover the size
of samples and to detect the optimal values of the signal (represented as showing
before by SC coefficients couples of min.-max. or max.-min), we apply a high-pass
interpolation filter g on the SC coefficients as following:
An Electrocardiogram Data Compression-Decompression Technique … 247
+∞
y(k) = SC(l)g(l − 2k) (2)
k=−∞

where the high-pass filter g(k) is given by:

√ √
g(k) = 2δ(k − 1) − 2δ(k − 2) (3)

so, the resulting signal will be calculating by:

l +∞
y(k) = − ∫ g(l − 2k)x(t − l)dt
l−1/2 l=−∞
(4)
l+1/2 +∞
+ ∫ g(l − 2k)x(t − l)dt
l k=−∞

The resulting discrete signal y(k) will detect, by exploiting the proprieties of
high-pass filter, the min.-max. (resp. max-min.) couples of SC coefficients and will
interpolate the samples by 2. The maxima (resp. minima) of the signal x(k) are
represented by two successive positive (resp. negative) y(k) samples (Fig. 1.c).
In a final step, we calculate the peaks’ amplitude (PA) of the signal, by the
following procedure:
1) If y(k) and y(k + 1) have the same sign, then: P A(k + 1) = 1
2) If y(k) and y(k + 1) have opposite signs, then: P A(k + 1) = 0.

(a) (b)

(c) (d)

Fig. 1 Changing in slopes detection: (a) double triangular signal, (b) SC coefficients, (c) high-pass
interpolation filter and (d) PA signal
248 S. Bensegueni

The PA signal will detect all the samples positions of the changes of slopes’
directions (i.e. position of optima values of the signal).

2.2 Integrator Filter

The purpose of the integrator filter is to merge the amplitudes of separate non-zero
samples by connecting them with regressive or progressive lines. In practice, this
integrating filter is calculated using a moving window defined by the following
equation:

1  N −1
y(k) = x(k − l) (5)
N l=0

where N is the width of the integration window.

The width N of this low-pass filter is calculated so as to connect the desired points
of the signal without missing other desired samples by using a so large filter width.

3 Compression Stage

The compression algorithm is based on the application of transformation of the ECG

signal into peaks’ amplitudes signal which will contain a reduced number of non-
zeros amplitudes to be transmitted. This peaks’ amplitudes signal represents the
change of slopes while non-zero peaks are representative of the existent details of
the ECG signal. However, before applying this transformation, we should minimize
the high-frequencies of the signal because they are characterized by fast variation of
slopes and may affect the high performance of our procedure. So, to reduce them,
we use wavelets based high frequency filter as in [6].
The following sub-sections represent the steps of the compression algorithm:

3.1 Slop Change Coefficients of ECG Signal

The slop change coefficients of the ECG signal are generated by applying (1), in order
to detect all the changes of the directions of this signal. The resulting coefficients
will represent peaks by two successive coefficients with the same sign, and the
other samples of the signal will be represented by a sequence of coefficients having
alternating opposite signs.
An Electrocardiogram Data Compression-Decompression Technique … 249

Fig. 2 Detecting of
changing in direction of the
ECG signal. All signal peaks
are represented by two
successive coefficients with
the same sign

3.2 Signal Time Scale Recovery

After calculating the SC coefficients, the time scale of this resulting signal, will be
reduced to half the time size of the original signal. Then, to recover the initial signal
size, we apply a high-pass interpolation filter g on SC coefficients by using (4).
The resulting signal will detect the min.-max. (resp. max.-min.) couples of SC
coefficients and will interpolate the samples by 2 (Fig. 1.d). The local optima values
of the ECG signal will be represented by two successive positive (resp. negative)
samples.

3.3 Generation of Null Samples

The final step of the proposed compression method is to generate a peaks’ amplitudes
(PA) signal which produces a signal with amplitudes equal to those of the ECG signal
for the filtered SC min.-max. and max.-min. couples (i.e. at the peak positions of the
ECG signal) and zero amplitudes for the other samples.
Figure 3.b shows the PA of an electrocardiogram signal. We note that the samples
of the ECG signal with changes of slopes’ directions are detected and replaced at
peaks positions by pulses having the same amplitudes values. The remaining samples
are replaced by zeros.
The non-zeros PA coefficients coded on 32 bits (simple precision) and the null
coefficients coded on one 1 bit, will be stocked or transmitted to the decompression
bloc as shown in Fig. 4.

4 Decompression Stage

In this stage, the received coefficients will be used to reconstruct the ECG signal
by using an integrator filter in order to merge the amplitudes of separate non-zero
samples and connecting them with regressive or progressive lines by using (5) and
250 S. Bensegueni

replacing the x(k) samples by the received PA coefficients. But before that, we should
convert the one-bit zero samples to 32 bits (simple precision) samples.
In general, the choice of the width of the integrator filter window should be
approximately the same as the smallest ECG characteristic wave. Actually, the width
of the window is determined empirically, and for a sample rate of 360 Hz, the window
is 7 samples width (approximately equal to 19 ms).

4.1 Electrocardiogram Database

In this paper, I use the MIT-BIH arrhythmia database, which is composed of 48

records of 30 min each. Few records (from 100 to 124) have been selected randomly
from selected patients and the other (from 200 to 234) contain plots with major
arrhythmic events. Several records were obtained from the modified derivation DII
by placing the electrodes on the chest. The other ones were obtained from precordial
leads V 1, V 2 and V 5 [7].

4.2 Evaluation Criteria

To evaluate the performance of ECG data compression/decompression algorithm, I

have chosen the following criteria [3, 8]:
Percentage Root-Mean-Square Difference. The (PRD) evaluates the quality of
the reconstructed signal compared to the original and is calculated as follows:

n
i=1 (ecg(i) − ecgr (i))2
P RD =  (6)
n
i=1 (ecg(i))
2

where ecg is the original signal and ecgr is the reconstructed signal from the
decompression part;
Compression Ratio. The (CR) is one of the most used measures for evaluating
compression/decompression techniques performances. It is defined by:

N umber o f bits in the original dataset

CR = (7)
N umber o f bits in the r econstr ucted dataset

Quality Score. The (QS) is used to evaluate the trade-off between the compression
ratio and the PRD by using the following ratio:

CR
QS = (8)
P RD
An Electrocardiogram Data Compression-Decompression Technique … 251

(a)

(b)

(c)

Fig. 3 Changing in slopes detection: a the ECG signal, b the PA signal and c in red, the
reconstruction of the ECG signal by using the high-pass interpolation filter

Computational Speed. The (CS) is defined as the speed at which the computer
execute the code and process data. It’s calculated by:

N umber o f coded samples

CS = (9)
Computation time

The CS is usually defined in millions of instructions per second (MIPS).

252 S. Bensegueni

Decompression
Slop Change steps
Original ECG
Coefficients of ECG
signal
Signal
7 samples wide
integrator filter
Signal Time Scale
Recovery

Reconstructed
Generation of Null ECG signal
Compression Samples
steps

Fig. 4 Compression/Decompression bloc diagram

The previously defined criteria were applied on data from the MIT-BIH database,
using the Matlab environment installed on a SSD drive of an i5 3rd generation laptop
with 2.60 GHz 2 cores and 4 threads processor, and equipped with a 8 GB of DDR3.

4.3 Results of Compression/Decompression Technique

The proposed method was applied on all the 30 min ECG records of the MIT-BIH
database, starting by calculating the slop change coefficients of these ECG signals
and by recovering their time scales. In a final step, the generation of null samples is
achieved in order to reduce the dimension of the records. Table 1 shows the evaluation
criteria values of the compression/decompression method of the 48 MIT-BIH records.
In Fig. 5, we can see the subplots of the reconstructed and the original signal of
segments of records: 100, 221 and 223. These figures show that the reconstructed
and the original signals are almost identical even in segments with sudden changes
in slope.From the previous results, we can verify that the mean value of the PRD
is equal to 0.62. This value shows that the reconstructed signals are acceptable and
the existent distortions are negligible and unimportant for many biomedical uses,
especially if we compare this value with the CR mean value equal to 17.83 which
represents a good ratio of signal compression.
An Electrocardiogram Data Compression-Decompression Technique … 253

Table 1 Evaluation criteria values of the compression/decompression method.

Records PRD CR QS CS (MIPS)
100 0.49 15.88 32.40 2.00
101 0.68 18.02 26.50 1.37
102 0.46 17.88 38.87 1.46
103 0.67 19.62 29.28 1.80
104 0.69 18.60 26.96 1.74
105 0.68 18.46 27.15 1.80
106 0.74 19.42 26.24 1.77
107 0.98 27.16 27.71 1.74
108 0.44 14.60 33.18 1.73
109 0.77 19.62 25.48 1.72
111 0.51 19.16 37.59 1.87
112 0.28 16.88 60.29 1.80
113 0.71 19.28 27.15 1.59
114 0.40 15.44 38.60 1.65
115 0.53 18.44 34.79 1.84
116 0.53 17.54 33.09 1.76
117 0.36 15.40 42.78 1.79
118 0.67 15.04 22.45 1.78
119 0.60 17.34 28.90 1.73
121 0.36 13.64 37.89 1.43
122 0.47 16.84 35.83 2.00
123 0.42 15.82 37.67 1.37
124 0.52 16.56 31.85 1.46
200 0.84 15.70 18.69 1.82
201 0.61 18.16 29.77 1.70
202 0.70 18.28 26.11 1.74
203 0.75 16.50 22.00 1.62
205 0.71 16.08 22.65 1.53
207 0.59 20.06 34.00 1.35
208 0.69 18.12 26.26 1.61
209 0.56 15.44 27.57 1.14
210 0.64 17.78 27.78 1.27
212 0.61 15.46 25.34 1.19
213 0.74 19.62 26.51 1.18
214 0.83 19.64 23.66 1.04
215 0.69 16.04 23.25 1.82
(continued)
254 S. Bensegueni

Table 1 (continued)
Records PRD CR QS CS (MIPS)
217 0.88 23.66 26.89 2.04
219 0.56 19.10 34.11 1.94
220 0.60 20.98 34.97 2.01
221 0.65 17.98 27.66 2.09
222 0.42 15.82 37.66 2.03
223 0.78 21.92 28.10 2.10
228 0.59 16.00 27.12 1.28
230 0.73 19.40 26.57 1.26
231 0.68 18.48 27.18 1.39
232 0.47 14.22 30.26 1.36
233 0.76 19.24 25.32 1.39
234 0.55 15.34 27.89 1.30

(a) (b)

(c)

Fig. 5 Original and reconstructed signals of MIT-BIH records segments: a 100, b 221 and c 223
An Electrocardiogram Data Compression-Decompression Technique … 255

5 Conclusion

In this paper, an electrocardiogram data compression-decompression technique was

proposed, by applying a transformation of the ECG to a peaks’ amplitude signal,
and hence, reducing the number of non-zero amplitude of the signal. At the decom-
pression step an integrator filter was applied in order to link the non-zero amplitudes
and reconstruct the signal. This compression method was very efficient in term of
reducing the dimension of the signal by a CR average value equal to 17.83 and
without too much distorting the signal with a PRD average value equal to 0.62.

References

1. Lalos AS, Alonso L, Verikoukis C (2014) Model based compressed sensing reconstruction
algorithms for ECG telemonitoring in WBANs. Digital Signal Process 35:105–116
2. Polanía LF, Plaza RI (2018) Compressed sensing ECG using restricted Boltzmann machines.
Biomed Signal Process Control 45:237–245
3. Rebollo-Neiraa L, Cerná D (2019) Wavelet based dictionaries for dimensionality reduction of
ECG signals. Biomed Signal Process Control 54:1–9
4. Bensegueni S, Bennia A (2016) ECG signal compression using a sinusoidal transformation of
principal components. Int J Softw Eng Appl 10:59–68
5. Grossi G, Lanzarotti R, Lin J (2015) High-rate compression of ECG signals by an accuracy-
driven sparsity model relying on natural basis. Digital Signal Process 45:96–106
6. Bensegueni S A new method for electrocardiogram features extraction using slope change
coefficients (under review)
7. Moody GB (1992) ECG database applications guide. Harvard University
8. Tiwari A, Falk TH (2019) Lossless electrocardiogram signal compression: a review of existing
methods. Biomed Signal Process Control 51:338–346
Internet of Things, Embedded Systems,
Blockchain and Security
Narrowband IoT Evolution Towards 5G
Massive MTC: Complete Performance
Evaluation

Adil Abou El Hassan , Abdelmalek El Mehdi , and Mohammed Saber

Abstract As the emergence of the 5G wireless network is expected to significantly

revolutionize the domain of communication, its design should consider the Internet of
Things (IoT) among the principal orientations to support the massive Machine-Type
Communication (mMTC). The emerging IoT applications supporting mMTC, need
new requirements other than throughput such as coverage, latency, power consump-
tion and connection density. To this purpose, the third Generation Partnership Project
(3GPP) has introduced a novel cellular IoT technology known as Narrowband IoT
(NB-IoT). This paper aims to determine the NB-IoT system configuration and deploy-
ment required to fully meet the 5G mMTC requirements. A complete performance
evaluation of NB-IoT against the 5G mMTC requirements is presented. It is shown
through the simulation results that these requirements can be met but under certain
conditions relating to the system configuration and its deployment.

Keywords NB-IoT · 5G · mMTC · Performance evaluation · 3GPP

1 Introduction

The Internet of Things (IoT) market for the 5G ecosystem is based on innovative tech-
nologies that efficiently support massive Machine-Type Communications (mMTC).
The 5G system is a next-generation wireless network that is based on New Radio
(NR) technology and is also designed to provide connectivity to a massive number

A. Abou El Hassan (B) · A. El Mehdi · M. Saber

Mohammed First University Oujda, National School of Applied Sciences, SmartICT Lab, Oujda,
Morocco
e-mail: [email protected]
A. El Mehdi
e-mail: [email protected]
M. Saber
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 259
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_24
260 A. Abou El Hassan et al.

of devices. The aim of the mMTC deployment is to connect up to billions of devices,

such as remote outdoor and indoor sensors through a cloud-based system.
The purpose of the 5G system design is to support three use case areas: enhanced
Mobile Broadband (eMBB), massive Machine Type Communication (mMTC), as
well as ultra Reliable Low Latency Communication (uRLLC) which is designed to
fulfill the requirements of critical Machine Type Communication (cMTC) [1]. The
advantage of the 5G system is the flexibility of its structure, which makes it possible
to cover many use cases with a single integrated system, by using a new feature
which is network slicing based on two new technologies that are Network Function
Virtualization (NFV) and Software-Defined Networking (SDN) [2].
The third Generation Partnership Project (3GPP) has introduced in Release 13
(Rel-13) specifications, a new licensed low power wide area (LPWA) technology
known as Narrowband IoT (NB-IoT) [3]. It is built from the existing Long Term
Evolution (LTE) base system, and supports IoT applications that do not require high
throughput like smart metering, smart parking and smart electric vehicle charging
system.
The 3GPP Rel-13 core specifications for NB-IoT were finalized in June 2016
[3]. About the Rel-14 and Rel-15 enhancements, they were completed in June 2017
and June 2018 respectively [3], whereas Rel-16 improvements are underway and
scheduled for completion in 2020 [1].
The 3GPP design aims for Rel-13 were low complexity and low cost devices,
long battery life, and support of a large number of devices. In addition the coverage
enhancement to reach devices in poor coverage conditions. To this purpose, two
power saving techniques have been implemented to reduce the power consump-
tion: Power Saving Mode (PSM) and extended Discontinuous Reception (eDRX)
introduced in Rel-12 and Rel-13 respectively [4, 5].
This paper aims to determine the NB-IoT system configuration and deployment
required to fully meet the 5G mMTC requirements based on the complete perfor-
mance evaluation of NB-IoT. In Sect. 2, an overview of NB-IoT is presented. It is
followed in Sect. 3, by a complete performance evaluation against the 5G mMTC
requirements in terms of coverage, throughput, latency, battery lifetime and connec-
tion density. It is shown through the simulation results that these requirements can be
met but under certain conditions regarding the system configuration and its deploy-
ment. The enhancements provided by the recent 3GPP releases are also discussed.
Finally, Sect. 4 concludes the paper.

2 NB-IoT Overview

2.1 3GPP Specifications and Numerology of NB-IoT

To ensure deeper coverage, the bandwidth occupied by the NB-IoT carrier is 180 kHz
corresponding to an one Physical Resource Block (PRB) of 12 subcarriers in an LTE
system [4]. There are three operation modes to deploy NB-IoT: as a stand-alone
Narrowband IoT Evolution Towards 5G Massive MTC … 261

carrier, in guard-band of an LTE carrier and in-band within an LTE carrier, which
offers operating mode flexibility of NB-IoT [6].
In order to coexist with LTE system, NB-IoT uses orthogonal frequency divi-
sion multiple access (OFDMA) in downlink with the identical subcarrier spacing
of 15 kHz and frame structure as LTE [4]. Whereas NB-IoT uses in uplink single-
carrier frequency division multiple access (SC-FDMA) and two transmission modes
which are the multi-tone and single-tone transmissions to ensure both high capacity
and maximum coverage for NB-IoT device with a single antenna [4, 6]. Multi-tone
transmission uses the same 15 kHz subcarrier spacing and 0.5 ms slot duration as
LTE. While single-tone transmission supports two numerologies which use 15 kHz
and 3.75 kHz subcarrier spacings with 0.5 ms and 2 ms slot durations respectively
[7]. Also, the restricted QPSK and BPSK modulation schemes are used in downlink
and uplink [6].
Two device categories Cat-NB1 and Cat-NB2 are defined by NB-IoT which corre-
spond to the device categories introduced in Rel-13 and Rel-14 respectively. The
maximum transport block size (TBS) supported in uplink by Cat-NB1 is only 1000
bits compared to 2536 bits for Cat-NB2. For downlink, the maximum TBS supported
by Cat-NB1 is only 680 bits compared to 2536 bits for Cat-NB2 [3].

2.2 Signals and Physical Channels of NB-IoT

The signals and channels used in downlink (DL) are as follows: Narrowband Primary
Synchronization Signal (NPSS), Narrowband Secondary Synchronization Signal
(NSSS), Narrowband Reference Signal (NRS), Narrowband Physical Broadcast
Channel (NPBCH), Narrowband Physical Downlink Shared Channel (NPDSCH)
and Narrowband Physical Downlink Control Channel (NPDCCH) [4].
Since NB-IoT uses only one PRB in a slot, the downlink signals and channels
are therefore multiplexed only in the time domain. The NB-IoT downlink frame is
structured as follows: subframe #0 of every frame carries NPBCH, subframe #1 to
#4, and subframe #6 to #8 of every frame carry NPDCCH or NPDSCH. NPSS is
carried by subframe #5 of every frame, whereas NSSS is carried by subframe #9
of every even numbered frame. While subframe #9 of every odd numbered frame
carries NPDCCH or NPDSCH [7].
In the uplink (UL), only one signal and two channels are used: Demodu-
lation Reference Signal (DMRS), Narrowband Physical Uplink Shared Channel
(NPUSCH) and Narrowband Physical Random Access Channel (NPRACH) [4].
Two formats are used for NPUSCH which are: Format 1 (F1) and Format 2 (F2).
NPUSCH F1 is used by user equipment (UE) to carry uplink user’s data to the
evolved Node B (eNB), and it supports both single-tone and multi-tone transmis-
sions [5, 7]. Whereas, NPUSCH F2 is used to carry uplink control information (UCI),
such as Hybrid Automated Repeat reQuest-Acknowledgement (HARQ-ACK), and
it supports only single-tone transmission [5, 7].
262 A. Abou El Hassan et al.

Table 1. Simulation and

Parameter Value
system model parameters.
System bandwidth 10 MHz
Channel model Tapped Delay Line
(TDL-iii/NLOS)
Doppler spread 2 Hz
NB-IoT mode of operation Guard-band
eNB Rx/Tx 4/2 and 4/4 only for NPSS/NSSS
transmissions
Device Rx/Tx 1/1

For cell access, the UE must first synchronize with the eNB using NPSS and
NSSS signals to achieve time and frequency synchronization with the network and
cell identification. Then, it receives Narrowband Master Information Block (MIB-
NB) and System Information Block 1 (SIB1-NB) carried by NPBCH and NPDSCH
respectively from eNB to access the system. Thereafter the UE starts random access
procedure by using single-tone transmission of NPRACH with 3.75 kHz subcar-
rier spacing, and waits for the scheduling uplink grant on NPDCCH which is used
to convey Downlink Control Information (DCI) for uplink, downlink and paging
scheduling [4, 5].

3 Performance Evaluation Against 5G mMTC

Requirements

In Rel-15, 3GPP has defined five targets for 5G mMTC corresponding to the following
performances: coverage defined by the maximum coupling loss (MCL), throughput,
latency, battery life and connection density [8]. In this section, a complete evaluation
of these five performances is presented to verify whether the 5G mMTC targets are
fulfilled by NB-IoT. In addition, the enhancements of performance provided by the
recent 3GPP releases are also discussed.

3.1 Coverage

The MCL is a common measure to define the level of coverage a system can support.
It is depending on the maximum transmitter power (PTX ), the required signal-to-
interference-and-noise ratio (SINR), the receiver noise figure (NF) and the signal
bandwidth (BW) [9]:

MCL = PTX − (SINR + NF + N0 + 10 log10 (BW)) (1)

N0 is the thermal noise density, which is a constant equal −174 dBm/Hz.

Narrowband IoT Evolution Towards 5G Massive MTC … 263

Table 2. Downlink and uplink coverage of NB-IoT.

Assumptions Downlink physical channel Uplink physical channel
for simulation NPBCH NPDCCH NPDSCH NPRACH NPUSCH NPUSCH
F1 F2
TBS [Bits] 24 23 680 – 1000 1
Acquisition 1280 512 1280 205 2048 32
time [ms]
BLER 10% 1% 10% 1% 10% 1%
Max transmit 46 46 46 23 23 23
power [dBm]
Transmit 35 35 35 23 23 23
power/carrier
[dBm]
Noise figure 7 7 7 5 5 5
NF [dB]
Channel 180 180 180 3.75 15 15
bandwidth
[kHz]
SINR [dB] −14.5 −16.7 −14.7 −8.5 −13.8 −13.8
MCL [dB] 163.95 166.15 164.15 164.76 164 164

Based on the simulation assumptions given in Table 1 according to [10] and using
(1) to calculate MCL, Table 2 provides the NB-IoT channels coverage to achieve the
MCL of 164 dB which corresponds to the 5G mMTC coverage requirement to be
supported [8].
Table 2 also shows the required acquisition time and block error rate (BLER)
associated with each channel to achieve the targeted MCL of 164 dB.
From Table 2, we note that to reach the MCL of 164 dB at the appropriate BLER, it
is necessary to use the time repetition technique of the simulated channels. However,
NPUSCH F1 is considered as the limiting channel, since it uses the longest acquisition
time to achieve the MCL of 164 dB. Whereas NPDCCH must be configured with
512 repetitions to achieve the targeted BLER of 1%, knowing that the maximum
configurable repetition number for NPDCCH is 2048 [5]. Therefore, NB-IoT can
support device operations in extreme coverage.

3.2 Throughput

The downlink and uplink throughputs are obtained according to the NPDSCH and
NPUSCH F1 scheduling cycles respectively and based on the acquisition times shown
in Table 2.
264 A. Abou El Hassan et al.

Fig. 1. NPDSCH scheduling cycle (Rmax = 512; G = 4) at the MCL

Fig. 2. NPUSCH F1 scheduling cycle (Rmax = 512; G = 1.5) at the MCL

Figure 1 depicts the NPDSCH scheduling cycle according to [10], where the
NPDCCH user-specific search space is configured with a maximum repetition factor
Rmax of 512 and a relative starting subframe periodicity G of 4.
Based on BLER and TBS given in Table 2 and using an overhead (OH) of 5
bytes corresponding to the radio protocol stack, a MAC-layer throughput (THP) in
downlink of 281 bps is achieved with the following formula:

(1 − B L E R)(T B S − O H )
THP = (2)
N P DCC H Period
The NPUSCH F1 scheduling cycle depicted in Fig. 2 corresponds to scheduling
of NPUSCH F1 transmission once every fourth scheduling cycle according to [10],
which ensures a MAC-layer THP in uplink of 281 bps according to the formula (2)
using the BLER and TBS given in Table 2 and OH of 5 bytes.
As part of 3GPP Rel-15, 5G mMTC requires that downlink and uplink throughputs
supported at the MCL of 164 dB must be at least 160 bps [8]. As can be seen,
the MAC-layer throughput meets the 5G mMTC requirement in both downlink and
uplink. It should be noted that the BLER targets associated with each channel require
the acquisition times shown in Table 2. Therefore, by choosing to operate with
NPDSCH and NPUSCH F1 BLER levels above 10% and the use of new Cat-NB2
device introduced in 3GPP Rel-14 which supports a larger TBS in both downlink
and uplink with two simultaneous HARQ processes [5], should further enhance the
throughput levels.
Narrowband IoT Evolution Towards 5G Massive MTC … 265

3.3 Latency

The latency should be evaluated for two cases: the first case is the use of the Radio
Resource Control (RRC) Resume procedure depicted in Fig. 3, while the second
case consists in use of new Early Data Transmission (EDT) procedure that has been
introduced in Rel-15 and depicted in Fig. 4. The EDT procedure allows the device to
terminate the transmission of small data packets earlier in RRC-idle mode without
switching to RRC-connected mode.
The latency evaluation is based on the same radio related assumptions and the
system model given in Table 1 by using the data and signaling flows corresponding
to RRC Resume and EDT procedures. Table 3 shows the used packet sizes and the
evaluated latency at the 164 dB MCL according to [10].
When the RRC Resume procedure is used, we obtain a latency of 9 s, compared
to 5.8 s when the EDT procedure is used. As can be seen the 5G mMTC target of

Fig. 3. NB-IoT RRC

Resume procedure

Fig. 4. NB-IoT EDT

procedure
266 A. Abou El Hassan et al.

Table 3. Packet sizes and results of latency evaluation

RRC resume procedure EDT procedure
Random Access Response: 7 bytes Random Access Response: 7 bytes
Msg2 Msg2
RRC Conn. Resume Request: 11 bytes RRC Conn. Resume Request: 11 + 105 bytes
Msg3 Msg3 + UL report
RRC Conn. Resume: Msg4 19 bytes RRC Conn. Release: Msg4 24 bytes
RRC Conn. Resume 22 + 200 bytes
Complete: Msg5 + RLC Ack
Msg4 + UL report
RRC Conn. Release 17 bytes
Latency 9s Latency 5.8 s

10 s latency at the MCL of 164 dB defined in 3GPP Rel-15 [8] is met, for both the
RRC Resume and EDT procedure. However, the best latency of 5.8 s is obtained by
using the EDT procedure, through multiplexing the user’s data with Message 3 on
NPUSCH F1 channel as shown in Fig. 4.

3.4 Battery Life

The RRC resume procedure is used in battery life evaluation instead of the EDT
procedure, since EDT procedure does not support uplink TBS larger than 1000 bits
which requires long transmission times. The packet flow used to evaluate battery life
is the same as shown in Fig. 3 where DL data corresponds to application acknowl-
edgment regarding receipt of UL report by the eNB. Four levels of device power
consumption are defined, including transmission, reception, Idle-Light sleep corre-
sponding to device in RRC-Idle mode or RRC-Connected mode but not actively
receiving or transmitting, whereas Idle-Deep sleep corresponds to power saving
mode. The device power consumption for each level and assumed traffic model
according to Rel-14 scenario are given in Table 4. In addition, an Active Timer of
20 s is included after connection release where the device is in Idle-Light sleep,
to monitor the control channel NPDCCH according to [11]. The Rayleigh fading
Extended Typical Urban (ETU) channel model with a Doppler shift of 1Hz is used
to evaluate the battery life according to [11].
The battery life in years is calculated using the following formula according to
[9]:

  Batter y energy capacit y

Battery life years = E day
(3)
365 × 3600
Narrowband IoT Evolution Towards 5G Massive MTC … 267

Table 4. Assumptions of
Message format
NB-IoT battery life evaluation
UL report 200 bytes
DL Application Acknowledgment 20 bytes
Report periodicity Once every 24 h
Device power consumption
Transmission power (PTx ) 500 mW
Reception power (PRx ) 80 mW
Idle - Light sleep power (PILS ) 3 mW
Idle - Deep sleep power (PIDS ) 0.015 mW

Knowing that Eday is the device’s energy consumed per day, expressed in Joule and
calculated as follows:
 
E day = (PT x × TT x + PRx × TRx + PI L S × TI L S ) × Nr ep + (PI DS × 3600 × 24)
(4)

Where TTx , TRx and TILS correspond to overall times given in seconds for trans-
mission, reception and Idle-Light sleep respectively and obtained from the assumed
transmission times according to [11] and the packet flow shown in Fig. 3, while Nrep
corresponds to the number of reports per day.
A device of 23-dBm power class with an 5 Wh battery and a base station of
46-dBm power class are used for guard-band and in-band operation modes, while
for stand-alone operation mode the power class of the base station is only 43 dBm
according to [11]. Based on the assumed transmission times given in [11] and using
the formulas (3) and (4), the evaluated battery lifes to achieve MCL of 164 dB in
in-band, guard-band and stand-alone operation modes are 11.4, 11.6 and 11.8 years
respectively. Knowing that the 5G mMTC requires battery life beyond 10 years at
the MCL of 164 dB, supposing an energy storage capacity of 5Wh [8]. Therefore,
NB-IoT achieves the targeted battery life in all operation modes.
Furthermore, to increase battery life further, the Narrowband Wake-Up Signal
(NWUS) that was introduced in 3GPP Rel-15 can be implemented, since it allows
the UE to remain in idle mode until informed to decode NPDCCH channel for a
paging occasion, thereby achieving energy saving.

3.5 Connection Density

The 5G mMTC target on connection density that is also part of the International
Mobile Telecommunication targets for 2020 and beyond (IMT-2020), requires the
support of one million devices per square kilometer in four different urban macro
scenarios [8]. These scenarios are based on two channel models (UMA A) and (UMA
268 A. Abou El Hassan et al.

Table 5. System level

Parameter Value
simulation assumptions of
urban macro scenarios Frequency band 700 MHz
LTE system bandwidth 10 MHz
Cell structure Hexagonal grid with 3 sectors
per size
NB-IoT operation mode In-band
Pathloss model UMA A, UMA B
eNB power and antennas 46 dBm–2Tx/2Rx
configuration
UE power and antennas 23 dBm–1Tx/1Rx
configuration

B) and two distances of 500 and 1732 m between adjacent cell sites denoted by ISD
(inter-site distance) [12].
Based on the simulation assumptions given in Table 5 and the non-full buffer
system level simulation to evaluate connection density according to [13], Fig. 5 shows
the latency required at 99% reliability to deliver 32 bytes payload as a function of the
connection attempts intensity (CAI) to be supported, corresponding to the number
of connection attempts per second, cell and PRB. It should be noted that the latency
shown in Fig. 5 includes the idle mode time to synchronize to the cell and read the
MIB-NB and SIB1-NB by the device.
Knowing that each device must submit a connection attempt to the system peri-
odically, we can calculate the connection density to be supported (CDS) per cell area
using the following formula:

C AI · C A P
C DS = (5)
A
Where CAP is the periodicity of device’s connection attempts
√ given in seconds
and the hexagonal cell area A is calculated by A = I S D 2 · 3/6.
To evaluate the connection density per PRB and square kilometer depicted in
Fig. 6, and corresponding to the overall number of devices that successfully transmit
a payload of 32 bytes accumulated over two hours with the required latency, the CDS
values are obtained from formula (5) using the CAI values of Fig. 5 and a periodicity
of connection attempts of two hours.
As can be seen from Fig. 6, in the two scenarios corresponding to 500 m ISD,
more than 1.2 million devices per PRB and square kilometer can be supported by
an NB-IoT carrier with a maximum 10 s latency. However, only 94000 and 68000
devices per PRB and square kilometer can be supported in the scenarios using the
(UMA B) and (UMA A) channel models respectively with an ISD of 1732 m within
the 10-s latency limit.
Since, in the case of 1732 m ISD, the density of base stations is 12 times lower
than with 500 m ISD. Therefore, this difference in base station density results in
Narrowband IoT Evolution Towards 5G Massive MTC … 269

Fig. 5. Intensity of
connection attempts in
relation to latency

Fig. 6. Connection density

in relation to latency

differences of up to 18 times between the connection densities relating to the scenarios

of 500 and 1732 meters ISD.

4 Conclusion

To conclude, this paper shows that the five targets of 5G mMTC are fulfilled by NB-
IoT. However, the evaluation results show that the performance achieved is highly
dependent on certain conditions relating to system configuration and deployment,
such as the number of repetitions configured for channels transmission and the density
of base stations, which is significantly dependent on the inter-site distance ISD. Since
the connection density evaluation shows that the target of 5G mMTC is only achieved
if ISD is 500 m. In addition, 3GPP Rel-15 allows 5G NR to coexist with carriers when
270 A. Abou El Hassan et al.

NB-IoT devices are already deployed in massive IoT applications. Thus, NB-IoT can
be considered as a key technology of the 5G system for massive IoT applications,
such as smart metering systems allowing measurement of water, gas and electricity
consumption.

References

1. Ghosh A, Maeder A, Baker M, Chandramouli D (2019) 5G evolution: a view on 5G cellular

technology beyond 3GPP release 15. IEEE Access 7:127639–127651. https://doi.org/10.1109/
ACCESS.2019.2939938
2. Barakabitze AA, Ahmad A, Mijumbi R, Hines A (2020) 5G network slicing using SDN and
NFV: a survey of taxonomy, architectures and future challenges. Comput Netw 167:106984.
https://doi.org/10.1016/j.comnet.2019.106984
3. Ratasuk R, Mangalvedhe N, Xiong Z, Robert M, Bhatoolaul D (2017) Enhancements of narrow-
band IoT in 3GPP Rel-14 and Rel-15. In: 2017 IEEE conference on standards for communi-
cations and networking (CSCN), pp 60–65. IEEE. https://doi.org/10.1109/CSCN.2017.808
8599
4. Feltrin L, Tsoukaneri G, Condoluci M, Buratti C, Mahmoodi T, Dohler M, Verdone R
(2019) Narrowband IoT: a survey on downlink and uplink perspectives. IEEE Wirel Commun
26(1):78–86. https://doi.org/10.1109/MWC.2019.1800020
5. Rastogi E, Saxena N, Roy A, Shin DR (2020) Narrowband internet of things: a comprehensive
study. Comput Netw 173:107209. https://doi.org/10.1016/j.comnet.2020.107209
6. Mahmood A, Zafar S (2019) Performance analysis of narrowband internet of things (NB-IoT)
deployment modes. In: 2019 22nd international multitopic conference (INMIC), pp 1–8. IEEE.
https://doi.org/10.1109/INMIC48123.2019.9022748
7. Wang YPE, Lin X, Adhikary A, Grovlen A, Sui Y, Blankenship Y, Bergman J, Razaghi HS
(2017) A primer on 3GPP narrowband internet of things. IEEE Commun Mag 55(3):117–123.
https://doi.org/10.1109/MCOM.2017.1600510CM
8. 3GPP (2018): TR 38.913, 5G: study on scenarios and requirements for next generation access
technologies Release 15, version 15.0.0. Technical report, ETSI, September. https://www.etsi.
org/deliver/etsi_tr/138900_138999/138913/15.00.00_60/tr_138913v150000p.pdf. Accessed
21 Jan 2021
9. 3GPP (2015): TR 45.820 v13.1.0: Cellular system support for ultra-low complexity and low
throughput Internet of Things (CIoT) Release 13. Technical report, 3GPP Organizational Part-
ners (ARIB, ATIS, CCSA, ETSI, TTA, TTC), November. https://www.3gpp.org/ftp/Specs/arc
hive/45_series/45.820/45820-d10.zip. Accessed 21 Jan 2021
10. Ericsson (2019): R1–1907398, IMT-2020 self evaluation: mMTC coverage, data rate, latency
& battery life. Technical report, 3GPP TSG-RAN WG1 Meeting #97, May. https://www.3gpp.
org/ftp/TSG_RAN/WG1_RL1/TSGR1_97/Docs/R1-1907398.zip. Accessed 21 Jan 2021
11. Ericsson (2017): R1–1705189, Early data transmission for NB-IoT. Technical report, 3GPP
TSG RAN1 Meeting #88bis, April. https://www.3gpp.org/ftp/TSG_RAN/WG1_RL1/TSGR1_
88b/Docs/R1-1705189.zip. Accessed 21 Jan 2021
12. ITU-R (2017): M.2412-0, Guidelines for evaluation of radio interface technologies for IMT-
2020. Technical report, International Telecommunication Union (ITU), October. https://www.
itu.int/dms_pub/itu-r/opb/rep/R-REP-M.2412-2017-PDF-E.pdf. Accessed 21 Jan 2021
13. Ericsson (2019): R1-1907399, IMT-2020 self evaluation: mMTC non-full buffer connec-
tion density for LTE-MTC and NB-IoT. Technical report, 3GPP TSG-RAN WG1 Meeting
#97, May. https://www.3gpp.org/ftp/TSG_RAN/WG1_RL1/TSGR1_97/Docs/R1-1907399.
zip. Accessed 21 Jan 2021
A Fuzzy Ontology Driven Integrated IoT
Approach for Home Automation

Levin Varghese, Gerard Deepak, and A. Santhanavijayan

Abstract Major advancements and development, in the field of Internet and other
technologies is now prevalent in many modern areas and dominating the way of
life. This in turn has allowed us to connect everyday objects, such as sensing and
actuating devices, to the digital world. These intelligent devices empower the dawn
of state-of-the-art services and applications for a multitude of various domains and
help in improving the complete utilization of the given resources. In this paper, an
interoperable, low cost, energy efficient Internet of Things platform using ontologies
for a smart home is being implemented. The said system promises to controls the
home appliances from anywhere i.e. offers remote operability and also provides
the collected data to Third party programs if need be, so that the Third parties can
creates tailor- made applications and services that cater to the need of the consumer.
The system promises to reduce human intervention, enable automation and provide
access to devices remotely from any place. Also, it allows for the collected data to be
sent to third parties for generation of third- party services, and also for the analysis,
and overall improvement of the system.

Keywords Fuzzy ontologies · Intelligent systems · Internet of Things · Smart

home

1 Introduction

The term “smart-home” refers to a residence which is equipped with a number of

devices that help in sensing and automating tasks which are normally handled by
humans. A growing trend sees people trying to add voice commands or artificial
intelligence into these designs, so that the tasks which are required to be carried
out, are done with much greater ease, and if possible, in the absence of a human
being, i.e. certain decisions which need to be made are done automatically. Turning

L. Varghese · G. Deepak (B) · A. Santhanavijayan

Department of Computer Science and Engineering, National Institute of Technology
Tiruchirappalli, Tiruchirappalli, Tamil Nadu, India
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 271
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_25
272 L. Varghese et al.

a house into a smart home, for some, could mean buying a wireless speaker, while
it can mean different for others, such as connecting several various products for
creating a so called smart environment or home. Home automation aims to enhance
the comfort, energy consumption efficiency and security in domestic scenarios [6].
We aim to reduce the human intervention, secure access control to home devices
from anywhere, provide smart homes data for application services as well as for
analysis, and improve the utilization of resources [5].

Motivation: As our resources such as electricity, drinking water, food and other
natural resources are getting scarcer and more expensive, there is a need to keep
the utilization of these resources in check. In order to reduce the wastage of these
resources and to ensure the safety of the people within a household, there is a great
scope and need for technologies, such as smart homes, which help to monitor and
save the resources and also try to keep the household secure from many mishaps,
such as fires, theft, etc.

Contribution: This paper is being done in two phases, in the first phase, identification
of the topic, research work on the topic and a simulation is being carried out. In the
second phase the implementation and ontology part of the paper is carried out. The
overall system can further be used in various other applications, with addition of
various other appliances. The system monitors room temperature, detects smoke
(mainly carbon dioxide), it also detects fire and tries to prevent any damages using
other appliances such as water sprinkler, alarm, etc. Further, all these devices can be
remotely controlled through cloud and are secured using an ID and password.

Organization: This paper is organized in the following way, in Sect. 2 literature

review is done. In Sect. 3, the System Architecture is proposed. Implementation,
Performance Evaluation and Results are discussed in Sect. 4 and finally, conclusion
and future works is discussed in Sect. 5.

2 Related Work

In [1], the authors, have integrated the semantic characteristics into the smart home
system paradigm and have created an ontology based semantic service model which
they call OSSM and that is what acts as the middleware for their system. In [2],
the authors use a structure based on an ontological model to facilitate the mindless
arrangement of fitting uses, instead of traditionally programming each household.
In [3], the authors have successfully provided an answer for securing households
by taking decisions dynamically with help of intrusive devices. In [4], the authors
suggest a smart home, using Raspberry Pi and IoT, it is done by connecting cameras
and sensing devices into a web app. In [5], the authors aim to reduce the involvement
of humans, ensure secure access to home appliances from any place, also provide
analytical services and help improve the system. In [6], the authors propose an inter-
face through which context awareness and decision making is carried out using
A Fuzzy Ontology Driven Integrated IoT Approach … 273

distance based fuzzy algorithm. In [7], the authors have used a fuzzy logic approach
for real- time checking and analysis of the data collected from temperature sensor,
heart beat sensor and gas sensor. Data collected is used to evaluate if dangers exist.

3 Proposed System Architecture

This IoT based smart home is built using a powerful tool, namely Raspberry Pi 3
Model B+ microcomputer. It is very cooperative when it comes to minimizing the
required system components. This system uses a multitude of sensors such as motion
detector, fire sensor, smoke sensor, etc. All these sensing appliances are interfaced
using the GPIO pins present on Raspberry Pi board. The general architecture of the
overall system can be seen in Fig. 1 below. A simulation of this said architecture has
been carried out in Cisco Packet Tracer tool.

4 Implementation

The application is completed using Python as the language for programming in

Jupyter notebook. Fuzzy ontology is used for the ontology part of the project and the
decision-making part is carried out with the help of triangular membership functions.
The general algorithm can be seen in Table 1 below. After running the algorithm
mentioned in Table 1 below, we get the output seen in Fig. 2 below.
The simulation worked flawlessly and also helped us to understand the potential of
the system in a very nice and informative manner. For the implementation part of the
project, we used a DHT22 sensor, which is sensor to sense humidity and temperature,
and an air quality sensor, CCS811. The sensors and actuators worked instantly when

Fig. 1 Architecture of the system

274 L. Varghese et al.

certain events were triggered, ensuring the minimization of any materialistic loss, as
well as, any loss of life.
We ran the fuzzy ontology for the scenario when there is a gas leak and or a fire
in the house and plotted the graph for the same, only one example is shown in Fig. 2
below, but this was carried out for a dataset of values for temperature and smoke,
and great results were achieved.

Table 1 General algorithm for Fuzzy ontology

Input: Crisp input values (values read from sensors)
Output: Crisp output values (values sent to actuators based on decision made)
Begin
Step 1: Initialization of terms and linguistic variables.
temperature = ctrl.Antecedent(np.arange(10, 60, 10), 'temperature')
Step 2: Construct MF (Membership Function or Functions)
temperature['Cool'] = fuzz.trimf(op_sense.universe, [10, 20, 30])
smoke['Acceptable'] = fuzz.trimf(op_sense.universe, [10, 20, 30])
Step 3: Initialization of rule/rules (if-then conditions)
rule1 = ctrl.Rule(temperature['Cool'] & smoke['Acceptable'],
op_sense['sec1'])
Step 4: Create control system (CS) and control system simulation (CSS)
object
Step 5: Using membership functions convert crisp input to fuzzy input
OPIot.input['temperature']= 35
OPIot.input['smoke']= 35
Step 6: Compute the fuzzy output by feeding the fuzzy input to the CSS
OPIot.compute()
Step 7: Convert the fuzzy output and then convert to crisp output.
print (OPIot.output['op_sense'])
op_sense.view(sim=OPIot)
End

Fig. 2 Fuzzy Ontology

output
A Fuzzy Ontology Driven Integrated IoT Approach … 275

The value 27.377, which we got in Fig. 2 above, indicates that our environment is
relatively in a safe zone, but at the same time, caution is to be taken as the environment
is still slightly above the acceptable environment (i.e. 23.5). So a prompt is sent to
the user to ensure if things are alright or not.
The value 21.60 which we got in Fig. 3 below, indicates that our environment is
in a very safe zone, though a little caution is to be expected as you can see, sector 2 is
also slightly shaded. Although the environment is stable, just some minor supervision
is required. This is only the case during the early stages of deployment. Once the
system learns the patterns of the consumer, it will make decisions accordingly.
The value 33.50 which we got in Fig. 4 below, indicates that our environment
is in the cautionary zone as well as in the danger zone. As you can see, sector 2 is
now shaded almost 75% and sector 3 is shaded around 25%. Also, after seeing the
inputs and checking the rules which were created by the user, code can be found
in Appendix A.1, we can infer that the temperature is more than what the user has
deemed cautionary, and it is this value of temperature which has caused the system
to be in sector 2 as well as sector 3. The amount of smoke is within the cautionary
limit. So, both parameters need to be monitored, and until the system is resolved,
preventive measures need to be taken i.e. triggering alarm, opening window/s etc.
Similarly, many other values were read, and he graphs for them were plotted and
inferred from.
Accuracy was used as a metric for the system. Equation 1 below, shows how the
accuracy is calculated.

Correct Outcomes
Accuracy = (1)
Total no. of outcomes

Fig. 3 System output when

temp value is in cautionary
zone
276 L. Varghese et al.

Fig. 4 System output when

temp and smoke value are
slightly above normal

The system was made to run a total of 2046 trials (Total no. of outcomes), out of
which 1957 trials were perfect (Correct Outcomes) while others were wrong, due to
wear and tear of the sensors. These achieved values imply that our system has about
96% accuracy, which is very good.

5 Conclusions

In this paper we have simulated one feasible solution for smart home system using
IoT and fuzzy ontology. One of our primary targets was to find an affordable and
practical solution for a real-time smart home system, and we were able to achieve it
by using the sensing devices along with the Raspberry Pi. Certain events were created
to simulate mishaps which could potentially cause damage to property and life, but
we were successfully able to negate any damage by designing a fast, responsive and
interactive smart home system. The fuzzy ontology part also yielded great results,
which has allowed us to complete the project successfully. This paper can be easily
scaled up or down by addition or deletion of sensors. Furthermore, a more visual
and informative, GUI (Graphical User Interface) can be used, based on the users
requirements, to help understand the system in a better and easy way.

References

1. Yang Y, Wang Z, Yang Y, Li M, Yang Y (2012) An ontology based semantic service model in
smart home environment. In: proceedings of 2010 3rd international conference on computer
A Fuzzy Ontology Driven Integrated IoT Approach … 277

and electrical engineering (ICCEE 2010 no 1)

2. Xu J, Lee YH, Tsai WT, Li W, Son YS, Park JH, Moon KD (2009) Ontology-based smart home
solution and service composition. In: 2009 International Conference on Embedded Software
and Systems, pp 297–304. IEEE, May
3. Javale D, Mohsin M, Nandanwar S, Shingate M (2013) Home automation and security system
using android ADK. International J Electron Commun Comput Technol (IJECCT) 3:382–385
4. Ganta Dr, Vinod A, Priyanka N, Ch. K (2019) iot based web controlled home automation using
Raspberry PI. Int J Sci Res Sci Eng Technol 229–234. https://doi.org/10.32628/IJSRSET19
6246
5. Iqbal A (2018) Interoperable internet-of-things platform for smart home system using webof-
objects and cloud. Sustain Cities Soc 38:636–646
6. Suryaprakash S, Mathankumar M, Ramachandran R (2017) IOT based home automation system
through adaptive decision making fuzzy algorithm. Res J Eng Technol 8(3):268–272
7. Abdulrazak B, Chikhaoui B, Gouin-Vallerand C, Fraikin B (2010) A standard ontology for
smart spaces. Int J Web Grid Serv 6(3):244–268
8. Deepak G, Teja V Santhanavijayan A (2020) A novel firefly driven scheme for resume parsing
and matching based on entity linking paradigm. J Discrete Math Sci Cryptograph 23(1):157–
165
9. Deepak G, Santhanavijayan A (2020) Onto best fit: a best-fit occurrence estimation strategy
for RDF driven faceted semantic search. Comput Commun
10. Kumar N, Deepak G, Santhanavijayan A (2020) A novel semantic approach for intelligent
response generation using emotion detection incorporating NPMI Measure. Procedia Comput
Sci 167:571–579
11. Deepak G, Kumar N, Santhanavijayan A (2020) A semantic approach for entity linking
by diverse knowledge integration incorporating role-based chunking. Procedia Comput Sci
167:737–746
12. Haribabu S, Kumar PSS, Padhy S, Deepak G, Santhanavijayan A, Kumar N (2019) A novel
approach for ontology focused inter- domain personalized search based on semantic set expan-
sion. In 2019 fifteenth international conference on information processing (ICINPRO), pp 1–5.
IEEE, December
13. Deepak G, Kumar N, Bharadwaj GVSY, Santhanavijayan A (2019) OntoQuest: an ontological
strategy for automatic question generation for e-assessment using static and dynamic knowl-
edge. In: 2019 fifteenth international conference on information processing (ICINPRO), pp
1–6. IEEE
14. Kaushik IS, Deepak G, Santhanavijayan A (2020) QuantQueryEXP: a novel strategic approach
for query expansion based on quantum computing principles. J Discrete Math Sci Cryptograph
23(2):573–584
15. Varghese L, Deepak G, Santhanavijayan A (2019) An IoT analytics approach for weather
forecasting using Raspberry Pi 3 Model B+. In: 2019 fifteenth international conference on
information processing (ICINPRO), pp 1–5. IEEE, December
16. Deepak G, Priyadarshini S (2016) A hybrid framework for social tag recommendation using
context driven social information. Int J Soc Comput Cyber-Phys Syst 1(4):312–325
Towards an IoT/Big-Data Platform
for Data Measurements, Collection
and Processing in Micro-grid Systems

Abdellatif Elmouatamid , Youssef Alidrissi , Radouane Ouladsine ,

Mohamed Bakhouya , Najib Elkamoun, Mohammed Khaidar,
and Khalid Zine-Dine

Abstract The resent infrastructure of the buildings is becoming socio-technical

systems that integrate different heterogeneous entities (e.g., sensors, actuators,
lighting, HVAC (Heating, Ventilation, and Air Conditioning), occupants, renew-
able energy, storage systems), which could interact dynamically and in a collective
manner to balance between energy efficiency, occupants’ comfort, sustainability,
and adaptability. This new infrastructure is known as the “Micro-grid” (MG) system
concept. However, the main challenge for this infrastructure is real-time monitoring
and data processing, which requires the use of new information and communication
technologies. In addition, incorporating mechanisms and techniques are required in
order to have buildings more energy-efficient while ensuring occupants’ comfort by
allowing entities interaction for suitable actions (e.g., turning On/Off HVAC and
lighting, balancing the fluctuation between power production and consumption). In
this work, a new holistic architecture of smart buildings is presented by improving
the main layers of MG systems. This architecture is proposed in order to integrate all
buildings’ aspects with the main trade-off is to efficiently manage the building while
maintaining a suitable occupants’ comfort. In fact, an MG system is structured into
three layers following the proposed holistic architecture. More precisely, we shed
more light on the MG system’s layer by integrating recent IoT/Big-Data technologies
for data gathering, processing, and control. A set of sensors are installed for power
measurement in the MG system while the measured data is gathered and transmitted
to an IoT/Big-Data platform for analysis, processing, and storing. The main aim is to
store a historical of data (e.g., power production/consumption, weather conditions)
from the installed MG system with open-access tools.

A. Elmouatamid (B) · Y. Alidrissi · R. Ouladsine · M. Bakhouya

College of Engineering and Architecture, LERMA Lab, International University of Rabat,
11100 Sala ElJadida, Morocco
e-mail: [email protected]
A. Elmouatamid · N. Elkamoun · M. Khaidar
Faculty of Sciences, STIC Laboratory, CUR”EnR&SIE, Chouaïb Doukkali University,
24000 El Jadida, Morocco
K. Zine-Dine
Faculty of Sciences, Mohammed V University, 10000 Rabat, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 279
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_26
280 A. Elmouatamid et al.

Keywords Micro-grid · Smart buildings · IoT/Big-data platform · Information and

communication technologies · Renewable energy sources · Energy management

1 Introduction

The integration of RESs (renewable energy sources) for large-scale of electrical

energy production has recently accelerated because of evident climate change, insuf-
ficiency of fossil resources, and greenhouse gas emissions problem. RES are clean
and eco-friendly sources and their abundance and renewable nature are among the
most important factors for their integration into SG (smart grid) networks. Mainly, the
simple consumer of electricity is becoming a producer of electricity due to the use of
these RESs with the possibility to store and consume locally the electricity. However,
these green energy sources come with new challenges, mainly their seamless inte-
gration with existing electrical networks. In addition, another important challenge
for this new electricity infrastructure is real-time monitoring and data processing,
which requires new ICT (information and communication technologies) based infras-
tructures (Fig. 1). The main aim is to ensure sustainable and reliable renewable
energy generation systems [1]. Therefore, this integration of ICTs, energy distribu-
tion systems, as well as distributed energy generation systems (e.g., RESs), creates
what is commonly named “Smart Grid”. In fact, SG represents the new smart elec-
trical network, since it brings the flexibility to integrate new electrical services, such
as electrical vehicles, and enables consumers to be energy producers by integrating
RESs using a bidirectional communication network [2]. In fact, due to the develop-
ment of IoT (internet of things) infrastructures and its related intelligent services, the
electricity grid has new capabilities to monitor, manage, and control its components
and then takes advantage of sophisticated bidirectional interactions.

Fig. 1 Global architecture: from SG to smart MG

Towards an IoT/Big-Data Platform for Data Measurements … 281

Moreover, the ICTs integration enables various smart and automatic services,
such as smart metering infrastructure, smart control and management for D/R
(demand/response) balance, advanced electricity marketing, and intelligent energy
storage for electrical vehicles integration. However, despite this progress, some
research work stated that the SG is experiencing new challenges. Mainly, the SG
is based on the actual infrastructures of power distribution grids, which are limited
by the unidirectional exchange of electricity [3]. Therefore, face these challenges;
other concepts have been developed together with the revolution of SG, such as the
IoE (internet of energy), the IoT, and the IoS (internet of services) [4]. Especially,
the development and the emergence of smart MG systems could resolve some of the
abovementioned SG challenges. MG could simplify the management of electrical
energy, from centralized to distribute EM. In addition, in MG systems, different types
of energy can be managed locally with the possibility to interconnect different MGs
in a distributed manner.
In this work, a new holistic architecture of smart buildings is presented by
improving the main layers of MG. The proposed MG system presents a smart and
active building that combines ICTs/Big-data infrastructure, RESs/storage systems,
EM/control strategies, and electrical power grids. Mainly, the main component of
the energetic system is presented, which is installed in our deployed MG system. We
focused on the modeling of each component of MG layers, especially for sensing and
measurement using our ICTs/Big-data infrastructure. The main contribution of this
work is to propose open-access tools to build an IoT/Big-Data platform for researches
in MG system. The installed platform collects the data from MG system (e.g., power
flow, temperature, humidity, wind speed) using a set of installed sensors. The gath-
ered data is used to develop predictive control approaches for energy management in
MG system using the new ICTs methods (e.g., machine learning algorithms, IoT).

2 MG Layers Structure for Energy Efficient Buildings

The MG system is structured into three horizontal layers: passive building layer (e.g.,
building envelope and insulation, architecture design), active building systems layer
(e.g., HVAC system, Lighting), and RESs system layer (e.g., PV, wind, storage).
These layers are monitored by one vertical layer for communication and ICTs inte-
gration. This vertical layer integrates mainly an IoT/Big-Data platform in order to
measure, analyze, predict, and forecast actions depending on the actual and predicted
context. In particular, our MG system is a smart and active building that combines
ICTs/Big-data infrastructure, RESs/storage systems, EM/control strategies, and elec-
trical power grids. This new concept of a building is more interactive for both
consumers and energy producers. In fact, consumers will reduce the cost of their
energy consumption based on the used control approaches, which take into account
the real-time cost of the power and the predictive power generation, for efficient D/R
management [4, 5]. Moreover, this MG structure offers the possibility to integrate
new buildings’ services, such as electrical vehicles, which can be used as a storage
282 A. Elmouatamid et al.

device to compensate the energy in the building by integrating the “Grid-to-Vehicle

& Vehicle-to-Grid” techniques.

2.1 Communication Technology Used for MG System

The ICTs have a fundamental role in the performance of MG systems operation and
its practical deployment. In an MG system, a large number of distributed compo-
nents, such as the end-users, generators, and energy storage systems, require a reli-
able, sophisticated, and fast communication infrastructure to reach the mandatory
of such systems. Therefore, effective communication systems should be normalized
to enable real-time exchange of data in MG systems in order to collect the neces-
sary information used for the EM (energy management) of the system. Mainly, the
communication technologies that are used actually in MG systems can be classified
into two categories, wired and wireless technologies. Table 1 presents the different
communication technologies, which can be used in the MG system with the specified
international standards [6, 7].

Table 1 Communication technologies for MG building systems

Type Technology Standard
WIRED Power line PLC ISO/ IEC 14908-3
Serial RS-422, RS-485, RS-232 IEC-61968
Ethernet Ethernet IEEE 802.3 IEC-61970
Bus-based CANBus
Modbus
Profibus IEC61850
SCADA PLC IEEE STD. 1815- 2012, IEEE
RS 60870
Ethernet IEEE 802.3
WIRELESS WLAN Wifi IEC61850
Lifi IEC61850
Bluetooth IEEE 802.15.1
ZigBee IEC-61968
IEC-61970
WiMax IEEE 802.16j/m
Cellular networks 3G IEC61850
4G
5G
Towards an IoT/Big-Data Platform for Data Measurements … 283

2.2 Data Sensing and Communication

The building’s transition to a smart MG system requires a distributed communication

as well as a distributed system for the management. For instance, the IEEE Standards
Association permitted the OpenFog reference architecture as an official standard that
is improved by the OpenFog Consortium incorporating Microsoft, Intel, Princeton
University, Dell, Cisco Systems, and ARM Holdings, under the name of IEEE 1934
[8], for IoT communication in MG systems. The standard classifies the interac-
tion of interconnected units for streaming and real applications into three classes:
i) Machine-to-Machine: any component can send and receive data or command to
and from any other one using the capabilities of IoT, ii) People-to-Machine: the
users can control each machine, which can be adapted depending on the occupants’
comfort, and iii) People-to-People: the users can exchange data through a cooper-
ation mechanism. In our studies, sensors are used to collect contextual data, which
are submitted to our remote platform for storage or further analytics. Mainly, current
and voltage sensors are installed to gather data from RESs, storage devices, and load
consumption. Generally, three types of sensors outputs can be specified, current loop
output, voltage output, and digital output. The sensors based on current loop output
are calibrated over a specified range against various engineering units to convert the
real value, such as the temperature, pressure, and current, to an equivalent current
value measured by the deployed microcontroller. The output of the individual cali-
bration certificate is compared to a mathematical manipulation improved in order
to obtain the absolute value. The operation of the loop is straightforward character-
ized by the conversation of the sensors’ output signal to a proportional current, with
the less current (generally on mA) presents the sensors’ zero-level output and the
high current represents the sensor’s full-scale output. Therefore, the voltage output
sensors can have various voltage outputs (5 V, 10 V, from −5 V to +5 V and more)
that are calibrated over this range against various units. The output is compared to
the individual calibration certificate for the specified measurement range for which
the sensor is selected. Generally, two types of voltage output signals are considered,
single-ended measurements and differential. Single-ended measurement considers
the voltage difference between a wire and the ground and the noise is only on the
positive wire, therefore, it is still measured along the output voltage of the sensor.
The differential voltage signal is known as “floating”, in which no reference to the
ground is considered. The measured signal is taken as the voltage difference between
the two wires. The noise elimination is considered as a main benefit of this method
because it is added to both wires and can then be filtered out by the common-mode
rejection of the data acquisition system. However, other types of sensors use digital
signal rather than analog signals of current and voltage because it is more robust
face to the noises and the measurement perturbs. The two most common digital
communications are RS-232 and RS-485. The RS-485 does not recommend any
communications protocol, it only specifies the electrical characteristics of the gener-
ator and the receiver. The main examples of communication protocols employing
the RS-485 are Canbus, Modbus, Profibus, and SDI-12.
284 A. Elmouatamid et al.

Fig. 2 Monitoring platform architecture

Therefore, the main advantage of the RS-485 system is the flexibility to transport
the signal. It mains that each sensor is defined by its individual address offering the
possibility to be attached to one cable. Multiple sensors use one port on the data
logger saving significantly the installation cable and data logging costs. Mainly, our
deployed MG system is equipped with component for measuring and collecting both
power consumption and production. In fact, a set of current and voltage sensors are
installed for power measurement as shown in Fig. 2. The system is equipped by an
Arduino, which allows collecting the data from different sensors and send it to a
micro-computer (Raspberry pi) connected directly to the IoT/Big-Data platform. In
fact, sensors transmit analog signals to the microcontroller, which converts them into
numerical data.
The next section introduces our MG system’s architecture. In particular, we high-
lighted the necessity of integrating recent ICTs and IoT/Big-Data technologies for
gathering external and internal data, which have been used to generate predictive
actions (e.g., regulating the room temperature by forecasting the building’s occu-
pancy, ventilation speed variation according to the forecasted CO2 , intelligent and
predictive control of energy flows management using forecasted power production,
consumption, and battery SoC).

2.3 IoT/Big-Data Platform for Data Collection

The aim was to develop a research testsite integrating the different components of
an MG system. As shown in Fig. 2, the system integrates PV panels, WT, batteries,
and the utility grid connected together in order to supply electricity to the loads in
Towards an IoT/Big-Data Platform for Data Measurements … 285

the building according to actual contexts. The system is monitored by an IoT/Big-

Data platform, which is used to collect, analyze, and store the data for EM and
control strategies development. Moreover, several scenarios are deployed in order
to develop a research platform that considers the concept of MG systems with the
different components of the different layers.
Real-time and context awareness information could be exploited for developing
predictive and adaptive context-driven control approaches using recent IoT/Big-data
technologies together with real-time and ML (Machine-learning) algorithms [9, 10].
The communication platform is composed of four main layers: i) sensors/actuators
layer, ii) data acquisition layer, iii) data processing layer, iv) data visualization/storage
layer. These layers are installed together with further services and applications for
context-driven control (Fig. 3). The MG is mainly equipped with a component for
measuring the different necessary parameters (e.g., current, voltage, temperature,
wind speed), for interacting with the passive and the active equipment, for regu-
lating the comfort for the occupancy, and for managing the power production and
consumption. In fact, a set of sensors is installed depending on the desired scenarios.
Regarding the data acquisition layer, a Kaa application is developed (i.e., IoT tech-
nique) [11] which is used to receive data from deployed sensors. We have also used
the MQTT that is a publish-subscribe-based protocol for IoT applications. For data
processing and storage, Storm services are used [12].
Mainly a topology composed of Spouts and Bolts was designed and developed
to allow receiving and processing streaming data from sensors. The spouts receive
the data from the Kaa application, and then transmit it to the Bolts for processing
and storage into the database (e.g., MongoDB) for further in-depth analysis. The
services layer includes real-time visualization and storage together with the control
of active equipment and RESs power production and consumption monitoring and
management. The platform was used for data gathering and processing of internal and
external building’s context. For instance, it was used to have occupant information
(e.g., number, presence, behavior, activities) in the building, since is a key input
for control approaches in energy-efficient buildings (e.g., active systems control)

Fig. 3 IoT/Big-Data platform architecture

286 A. Elmouatamid et al.

[9]. In fact, comprehensive occupancy information could be integrated to improve

the performance of occupancy-driven control of HVAC, lighting, and ventilation
systems.

3 Experimental Platform of MG Systems

Approaches developed in the frame of this work are deployed and tested in real-sitting
scenarios in two MG systems. The first, named EEBLab (Energy Efficient Building
Laboratory) is used as a testbed for testing scenarios and validation. The second MG
system is a real house, which is deployed during Solar Decathlon Africa competition,
and it is deployed in Green Energy Park in Benguerir, Morocco. Both MG systems
are used actually as a research platform for MG system development. The main aim is
to measure, collect, analyze, and store the different data in the deployed MG systems.
Mainly, the collected data are, especially, the indoor/outdoor temperature, the indoor
CO2 variation, the weather conditions, and the power production/consumption.
Therefore, the measured data are used to develop different scenarios as the occu-
pant’s detection, the comfort amelioration, and the EM/control for the deployed
RESs/storage systems. In this part, we focus on the different measurements realized
for the energetic system. Regarding the external context, we have deployed a weather
station. In fact, for several scenarios, we need to collect internal and external context
data. We have built a weather station near the WT and PVs in order to have as precisely
as possible the data for wind speed, direction, irradiation, temperature, and humidity.
The weather station was deployed and used to collect the data for real-time visualiza-
tion and processing for further usage by other building’s services and applications.
Moreover, D/R control strategies are therefore required for an efficient management
of energy flows by considering the intermittent RES generation and the delay might
occur between RESs power generation and the actual building’s demand. The main
aim is to develop a control card to test the different studied control strategies for EM.
Unlike existing systems, which are used as a black-box to collect and manage the
energy in MG, the deployed control card allows us to measure, monitor, manage,
and deploy our algorithms [13]. Therefore, to develop an optimal and reliable control
strategy, multiple objective functions should be taken into account, like the continuity
of the power generated to the loads, the electrical energy cost, the smoothing of RESs
production„ and the C/D cycle of the batteries [13]. For that, a control strategy should
be deployed to satisfy the constraints designed by the optimization functions. The
communication infrastructure is employed for total electrical energy measurement
and management. The installed infrastructure provides the autonomous operation
with the required measurements, decisions, and command by gathering data through
the sensors and producing the commands for the Hardwar/Software (Hw/Sw) card
presented in Fig. 4, which is connected to the control switches used in the hybrid
system [10].
For the different scenarios, the proposed IoT/Big-Data platform could be used to
measure different parameters in the MG system. Depending on the studied scenario,
Towards an IoT/Big-Data Platform for Data Measurements … 287

Fig. 4 The deployed Hw/Sw control card

Fig. 5 Power measurement

scenario in our deployed MG
system platform

suitable sensors are selected and can be connected to this platform for data collec-
tion, monitoring, and processing. Mainly, the measured parameter is used to control
efficiently the deployed MG system. Moreover, a case study is presented for a large-
scale installation in which the developed control card and the IoT/Big-Data platform
are used to measure and store the data collected by the deployed current and voltage
sensors for the total installed RESs, battery storage system, and load consumption.
As shown in Fig. 5, the green curve presents the power generated from the eight PV
panels for 24 h. This power is calculated by measuring the PV current and voltage
variability during the day, which depends on the weather conditions (e.g., tempera-
ture, irradiance) changeability. At the same time, the battery SoC is calculated using
our battery characterization system installed in the MG system. Moreover, the power
consumption is measured and stored for the same period. These parameters are the
main input for the EM strategy.
However, due to the limited number of pages of this work, the obtained results
are summarized in Table 2 for the different measurements realized in the deployed
MG system focusing on that are used for EM in the deployed hybrid RESs system.
Table 2 presented the obtained results for 10 days of the test. Actually, the house
288 A. Elmouatamid et al.

Table 2 Energy balance in the deployed MG system

is used as a research platform and it is planned to deploy the test control scenarios,
tested in a medium-scale, for large and real rural MG systems.

4 Conclusions

The main aim of this work is to introduce an MG platform by focusing on commu-

nication and data monitoring concepts. The MG platform connects the building’s
components using sensing/actuating, IoT/Big-Data technologies in order to leverage
real-time gathering and data processing for RESs production and loads’ consump-
tion. The platform was deployed and several scenarios have been tested and eval-
uated. The proposed data gathering techniques can be an interesting reference for
the researches to have their proper platform for data collection that can be used in
the research works while the deployed mechanisms are open sources. In the future
works, several scenarios will be developed using the deployed IoT/Big-Data plat-
form, especially, that concerns EM and control, the active/passive equipment control,
and the occupant’s comfort management. In addition, machine-learning algorithms
will be developed and tested in real scenarios using real data to train the proposed
algorithms.

Acknowledgments This work is supported by MIGRID project (grant 5-398, 2017–2019), which
is funded by USAID under the PEER program. It is also partially supported by HOLSYS project,
which is funded by IRESEN (2020–2022).

References

1. Kabalci E, Kabalci Y (2019) From smart grid to internet of energy. Academic Press
2. Asaad M, Ahmad F, Alam MS, Sarfraz M (2019). Smart grid and Indian experience: a review.
Resour Pol 101499
Towards an IoT/Big-Data Platform for Data Measurements … 289

3. Wang K, Hu X, Li H, Li P, Zeng D, Guo S (2017) A survey on energy internet communications

for sustainability. IEEE Trans Sustain Comput 2(3):231–254
4. Tsiatsis V, Karnouskos S, Holler J, Boyle D, Mulligan C (2018) Internet of Things: technologies
and applications for a new age of intelligence. Academic Press, San Francisco
5. Elmouatamid A, NaitMalek Y, Bakhouya M, Ouladsine R, Elkamoun N, Zine-Dine K, Khaidar
M (2019) An energy management platform for micro-grid systems using Internet of Things
and Big-data technologies. Proc Inst Mech Eng Part I J Syst Control Eng 233(7):904–917
6. Dede A, Della Giustina D, Massa G, Cremaschini L (2016) Toward a new standard for secondary
substations: the viewpoint of a distribution utility. IEEE Trans Power Delivery 32(2):1123–1132
7. Parhizi S, Lotfi H, Khodaei A, Bahramirad S (2015) State of the art in research on microgrids:
a review. IEEE Access 3:890–925
8. Chiang M, Zhang T (2016) Fog and IoT: An overview of research opportunities. IEEE Internet
Things J 3(6):854–864
9. Elkhoukhi H, NaitMalek Y, Bakhouya M, Berouine A, Kharbouch A, Lachhab F, Essaaidi M
(2019) A platform architecture for occupancy detection using stream processing and machine
learning approaches. Concur Comput Pract Exp e5651
10. Elmouatamid A, NaitMalek Y, Ouladsine R, Bakhouya M, Elkamoun N, Khaidar M, Zine-
Dine K (2020a) A micro-grid system infrastructure implementing IoT/big-data technologies
for efficient energy management in buildings. Submitted to ATSPES’1 (Advanced Technologies
for Solar Photovoltaics Energy Systems), Book Chapters (Springer Book).
11. Kaa (2020) IoT technique. https://www.kaaproject.org/. Accessed 10 Oct 2020
12. Storm (2020) Apache Storm. https://storm.apache.org/. Accessed 10 Oct 2020
13. Elmouatamid A (2020) MAPCAST: an adaptive control approach using predictive analytics
for energy balance in micro-grid systems. Int J Renew Energy Res (IJRER) 10(2):945–954
14. Elmouatamid A, Ouladsine R, Bakhouya M, El Kamoun N, Khaidar M, Zine-Dine K (2021)
Review of control and energy management approaches in micro-grid systems. Energies
14(1):168
Smart Hospitals and Cyber Security
Attacks

Yassine Chahid, Mohammed Benabdellah, and Nabil Kannouf

Abstract Until now, the Covid-19 pandemic was infected more than 37,167,291
people around the world in 212 countries and territories, killed 1,071,181 people
since its appearance in December in Wuhan, of which more than 7,876,942 people in
the United States, 6,979,434 people in the India and 5,057,958 people in Brazil
(Novel Coronavirus (2019-nCoV) situation reports - World Health Organization
(WHO), 2019 Novel Coronavirus (2019-nCoV) in the U.S -. U.S. Centers for Disease
Control and Prevention (CDC), symptoms of Novel Coronavirus (2019-nCoV) –
CDC, china Travel Advisory - U.S. State Department, accessed January 31, 2020.).
This large number of infected people caused saturation in hospitals and put a big
pressure on the intensive care units and hospital beds. While the medical staff are
fighting to save the lives of these patients, others are playing with these lives by
making cyber-attacks on research centers and health institutions, recently case was
in September 2020, a woman in Germany died during a ransomware attack on the
Duesseldorf University Hospital. Smart hospitals are one of the most affected because
they use its technology (Smart hospitals) that generates a lot of personal and crit-
ical data, these hospitals are equipped with “things” that are used for patient moni-
toring, maintenance, remote operation and control like wearables, smart pills, smart
beds, remote monitoring systems, RTHS (Real-time Health Systems), biosensors,
glucose measurement devices, robots, equipment monitoring devices, and more. In
this paper, we will propose a new secured system model to keep these hospital
systems safe against cyber-attacks by using verification and classification of treat-
ment requests and controlling the patients’ health using “sensors” to ensure the safety
of the requested operations.

Keywords Coronavirus · Internet of Things · Security · Cyber-attacks · Hospitals

Y. Chahid (B) · M. Benabdellah

Faculty of Sciences, ACSA Laboratory, Mohamed First University, Oujda, Morocco
N. Kannouf
ENSA Al-Hoceima, DSCI, LSA Laboratory, University of Abdelmalek Essaadi, Tétouan,
Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 291
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_27
292 Y. Chahid et al.

1 Introduction

The Internet of things continues to enter our lives slowly every day and affects
many areas [2]. Healthcare is one of the richest domains that contain the biggest
connected objects, thermometers, ultrasounds, automatic respirators, glucose moni-
tors, digestible sensors, smartwatch, electrocardiograms and others, have become
connected and letting patients track their health and facilitate the doctor’s work by
automating some tasks like operations, drugs injections or insulin control [3]. Several
hospitals have started to use smart beds, which can be equipped with sensors like
injectors or blood collectors, insulin pens, automatic respirators and others, in order
to facilitate the patient treatment operation and provide great support without the need
for a doctor or nurse intervene [4]. The internet of things in the field of healthcare can
also help patients to take their medicament with the right doses directly from their
homes without having to stay in the hospital, this can also guarantee doctors better
monitoring their patients and having an alert if ever there is the slightest concern
[5]. IoT market in healthcare is predicted to exceed $10 billion by 2024 [6], which
means billions of critical and personal data that will circulate on networks. In 2016,
73 million healthcare devices were connected worldwide. In 2020, they will be 161
million [7]. Growth will mainly be driven by 3 trends: the rising average age of the
world’s population, the prevalence in some countries of diseases requiring regular
monitoring (such as diabetes) and the growing demand for quantifiable fitness solu-
tions. This paper is organized as follows: in the first section, we give a state of the art
of the MIMT attack and the related work already done by other research to improve
security against this attack, in the second section we will present a new model that
can help us to improve security inside of smart hospitals, the last section will discuss
the results obtained.

1.1 Internet of Things and COVID-19

The large amount of data generated by the IoT in the healthcare field forces us to fear,
because this information attracts attackers and make this data accessible everywhere
in the world if we do not apply strong security against these attacks. According to [8]
Large numbers of domain names related to Covid-19 have been registered in recent
weeks, more than 50% of which are thought to be part of “lure and decoy” operations.
Phishing attempts to have soared by over 600% since the end of February, including
traditional impersonation scams but also business email compromise (BEC) and
extortion attacks [9]. Other e-mail campaigns are asking users to transfer money
to sites claiming to be collecting charitable donations or to fake e-commerce sites
offering personal protective equipment.
Smart Hospitals and Cyber Security Attacks 293

1.2 Cyber-Attack Examples

The healthcare domain is not immune to these attacks, below are some events
examples made during the COVID-19:
1. May 2020, the FBI issues Flash Alert MI-000124-MW, covering specific
indicators for COVID-19 phishing email campaigns.
2. April 2020, Attacks/campaigns using COVID-themed lures has continued to
increase over the last week.
3. April 2020, several Android-focused campaigns were observed spreading the
Anubis and Cerberus banking trojan to victims seeking additional information
on Coronavirus in their area.
4. March 2020, a fake email using CORONAVIRUS information disguised as an
alert from the “Department of Health” was propagated.
5. April 2020, fake charity and donation have become more and more frequent
since the onset of the pandemic.

2 Related Work

Many efforts have been performed to develop a secured IoT healthcare system. In
this section, we cite some works that are reviewed and summarized.
Fathi Ibrahim Salih et al. have proposed an IoT security risk management model
for healthcare industry [10], as a solution, this paper proposes an enhanced IoT
risk management model for healthcare with consideration of three risk categories:
Secured Technology, Human Privacy and Trustable Process and Data. Bahar Fara-
hani et al. discuss applicability of IoT in healthcare field by presenting an architecture
of IoT eHealth ecosystem [11]. Behailu Negash et al. [12] presents a chapter focuses
on a smart e-health gateway implementation for use in the Fog computing layer, also,
connecting a network of such gateways, both in home and in hospital use. Simple
healthcare scenarios are presented. The features of the gateway in the Fog implemen-
tation are discussed and evaluated. Nasser S. Abouzakhar et al. [13] have published
a paper that presents different risks and security threats in the IoT, some security
measures and solution models as well as a vulnerability assessment approach for the
IoT in the healthcare field. Florian Kammüller [14] show in his paper, how to derive
formal specifications of secure IoT systems by presenting a process that uses the risk
assessment strategy of attack trees on infrastructure models, his proposed process in
the paper uses two features to refine a system specification until expected security
and privacy properties can be proved. He illustrates, in the final part of his paper, the
stepwise application of the proposed process in the Isabelle Insider framework on the
case study of an IoT healthcare system. Maria Almulhim and Noor Zaman [15] have
published research that proposes a secure group-based lightweight authentication
scheme for IoT based E-health applications, the proposed model provides mutual
294 Y. Chahid et al.

authentication and energy efficient, and computation for healthcare IoT based appli-
cations, which will use elliptic curve cryptography (ECC) principles that provide
mentioned featured of suggested model.

3 The Proposed System

The most dangerous attack in healthcare is that which relies on the use of sensors
(insulin pens, automatic respirators, drugs injection…etc.) to inject higher/lower
doses according to the patient’s health condition, stop automatic respirators or modify
the volume of oxygen transmitted, or even the attack at the same time of remote
surgery, these attacks can cause immediate patient death. To cope with this type
of attacks, we are going to propose a secure system model which is based on the
classification of the requested operation types for patients and to control access or
any treatment according to these categories.
In this table (Table 1), we give an example of attack which is based on the use of
Man-In-The-Middle (MITM) which is very well-known in the attack’s types, MITM
is an attack type when an attacker positions himself in the middle of the conversation
and precisely between a user and an application to imitate one of the parties, making
it appear as if a normal exchange of information is established.

3.1 Man in the Middle Attack

The goal of this attack is to steal personal data: login credentials, credit card numbers
or account details. Victims are generally the financial applications, e-commerce
websites or smart hospitals. The MITM attack come in many types [16].

Table 1 Healthcare cyberattack workflow

Heading level Example
1 The attacker tries to connect to the network
2 It creates a Man-In-The-Middle (MITM) attack
3 It launches the injection action of a mortal dose of chloroquine to the patient
using the drugs injector
4 The mortal dose is injected directly without any verification
5 The patient died immediately
Smart Hospitals and Cyber Security Attacks 295

3.2 The Proposed System

The proposed model will be positioned between the hospital’s external server and
the client, which in our case is the hospital (Fig. 1).
The system will aim to retrieve requests and be able to process them according to
their severity, the system will play the orchestrator role between any patient’s equip-
ment like electrocardiogram (ECG), Electromyography (EMG), Electrogastrogram
(EGG), Insulin pens…etc.) using tags [17, 18] (Fig. 2).
To explain this algorithm, we will analyze each step (Table 2):

4 Implementation and Results Analysis

Before continuing reading this article, it should be noted that this paper is a computer
science paper and does not encourage the use of any medication type against COVID-
19 (We used the chloroquine just as an example).

4.1 The Prerequisites

The proposed system is implemented using Python Web APIs and MySQL as
database for storing logs and processed requests. We simulate a smart hospital,
patients, and we regulated the categories as follows:

Fig. 1 Proposed secured system model

296 Y. Chahid et al.

BEGIN
needed_value 0
isSafe false
levels empty_list
values empty_table_params
levelList Get_category_levels
// The system verify the criticity of the operation
critic verify_criticity(RequestedOperationType, values)
tag_level_action
IF isSafe
do_operation
ELSE
PRINT "Potential critical action”
PRINT "Lanching internal checks”
// Get chloroquine, glucose, tension values of the patient
values Get_values(RequestedOperationType)
IF level 2
needed_value calculate_needed_value
do_action(needed_value)
ELSE
reject_action
send_report_to_doctor
save_operation_logs
END IF
END IF
END

Fig. 2 Proposed system algorithm

Table 2 Healthcare cyberattack workflow

Step Example
1 The attacker tries to connect to the network
2 It launches the injection action of a mortal dose of chloroquine to the patient using the
drugs injector
3 The system retrieves the list of categories and tries to classify this request according to
his criteria: nature of the request (injection of chloroquine dose, oxygen increasing,
blood recovery…), the parameters of the request (dose value, value for increasing the
oxygen rate…)
4 The system checks if the request is considered as safe or not
5 If the demand is considered as critic, the system launches the internal verification using
sensors (Make an ECG test for example, calculate the blood pressure…etc.)
6 Based on these checks, the system can accept or reject the demand
7 The system logs and send a report to a doctor
Smart Hospitals and Cyber Security Attacks 297

Table 3 Patient’s status with

Case Requested chloroquine dose Patient’s health
injected dose values without
system implementation 1 1000 Died
2 5000 Critic
3 5000 Critic
4 120 Normal
5 3000 Died
6 430 Critic
7 1000 Died
8 1000 Died
9 100 Normal
10 2670 Died

1. If the action is of the injection type then automatically it is considered as level 2,

if the value of certain parameters such as glucose exceeds 500 ml then the action
goes to level 3. like oxygen for example, the action in this case is considered
level 3.
2. If the action is not an injection or stop, it is considered level 1.
All these rules can be marked and stored in a simple text file rules.txt that it must be
encrypted and be a Read Only to prevent its modification from any intruder. The calls
in this system are Asynchronous calls, because if ever the actions are considered as
critical, the system will launch an internal check (glucose check, tension check…etc.)
which will take a little time before completing the action. We have modeled 5 cases
of attacks, in each case we put different information about the action type and the
value of the demand and below the summary table (Table 3):
As we can see in the table of results, we simulated a chloroquine injection test on
10 patients, each patient has an initial dose on his body. We have patients who needs
chloroquine injection and others do not need that. Without implementation of the
system, the attacker injected into the 10 patient’s blood an important and fatal dose,
these doses were immediately killed most patients (5/10), worsening the situation of
the (3/10) others while (2/10) were able to survive. Below a graph which shows the
graph of the evolution of the patients after the attack (Fig. 3):
The table (Table 4) shows the results of the patient’s health after the implementa-
tion of the system.
As we can see, the system can prevent the injection of most of the doses by
classifying the requests by category and degree of severity, then triggered the internal
reverification of chloroquine for most patients before authorizing the execution of the
injection. The system was able to save these patients by refusing 7 requests (Fig. 4).
As we see, the system rejects 80% of the requests because they are considered
like attacks.
The result now provides evidence to use the proposed model in order to enhance
smart hospitals security, these results confirm that this a good choice for using an
internal system that can check MIMT sent requests before injecting these doses.
298 Y. Chahid et al.

Fig. 3 Patient’s status graphic with injected dose values without system implementation

Table 4 Patient’s status with injected dose values in case of system implementation
Case Requested chloroquine dose Patient’s health
1 1000 Normal
2 5000 Normal
3 5000 Normal
4 120 Normal
5 3000 Normal
6 430 Critic
7 1000 Normal
8 1000 Normal
9 100 Normal
10 2670 Normal

When comparing our results to those of older studies, it must be pointed out that
our proposed model can be integrated with any third system, also, the model can
improve the security of smart hospitals against MITM attacks because he is the last
who check requests and it is the closest to patient and this can guarantee that no one
can modify data after the check.
Smart Hospitals and Cyber Security Attacks 299

Fig. 4 Patient’s status graphic with injected dose values with system implementation

5 Conclusion

In summary, based on this experimental and theoretical study of this cyberattack,

we can say that the proposed system is used to increase the level of security against
IoT attacks, even if it will be difficult to launch a MITM attack, we must work
on its security. It must be known that this solution can be applied also in other
IoT automated fields like transport or industry and it can be combined with other
algorithms to guarantee the three pillars of IoT that are integrity availability and
confidentiality of data.

References

1. Novel Coronavirus (2019-nCoV) situation reports - World Health Organization (WHO), 2019
Novel Coronavirus (2019-nCoV) in the U.S -. U.S. Centers for Disease Control and Prevention
(CDC), symptoms of Novel Coronavirus (2019-nCoV) – CDC, china Travel Advisory - U.S.
State Department. Accessed 31 Jan 2020
2. Chahid Y, Benabdellah M, Azizi A (2017) Internet of things security. In: International confer-
ence on wireless technologies, embedded and intelligent systems (WITS). https://doi.org/10.
1109/WITS.2017.7934655
3. Al-Refaie A, Chen T, Judeh M (2018) Optimal operating room scheduling for normal and
unexpected events in a smart hospital. Oper Res Int J 18:579–602. https://doi.org/10.1007/s12
351-016-0244-y
4. Hurysz T, Romisher A, Campbell D, Ottenstein Maia (2020) Utilizing the smart rooms device
to improve patient experience in hospitals. Phase 1. Paper 12. https://jdc.jefferson.edu/si_dh_
2022_phase1/12
5. Gandhi DA, Ghosal M (2018) Second international conference on inventive communication
and computational technologies (ICICCT). https://doi.org/10.1109/ICICCT.2018.8473026
300 Y. Chahid et al.

6. Internet of Things (IoT) in Utility - Global Market Research and Forecast, 2015–2025 Report
ID: HeyReport6861 | Number of pages: 80 | Publish Date: Feb 2019 | Publisher: HeyReport |
Category: ICT and Media
7. Internet of Things (IoT) in Healthcare Market Size, Share & Trends Analysis Report By Compo-
nent (Service, System & Software), By Connectivity Technology (Satellite, Cellular), By End
Use (CRO, Hospital & Clinic), By Application, and Segment Forecasts, 2019–2025, Report
ID: 978-1-68038-857-2
8. https://blog.checkpoint.com/2020/03/19/covid-19-impact-as-retailers-close-their-doors-hac
kers-open-for-business/
9. Phil Muncaster UK / EMEA News Reporter, Infosecurity Magazine https://www.infosecurity-
magazine.com/news/covid19-drive-phishing-emails-667
10. Salih FI et al. (2019) IoT security risk management model for healthcare industry. Malaysian
J Comput Sci [S.l.] 131–144. ISSN 0127-9084. https://doi.org/10.22452/mjcs.sp2019no3.9.
Accessed 10 May 2020
11. Farahani B, Firouzi F, Badaroglu VCM, Constant N, Mankodiya K (2017) Towards fog-driven
IoT eHealth: promises and challenges of IoT in medicine and healthcare. Elsevier. https://doi.
org/10.1016/j.future.2017.04.036
12. Negash B et al. (2018) Leveraging fog computing for healthcare IoT. In: Rahmani A, Liljeberg
P, Preden JS, Jantsch A (eds) Fog computing in the internet of things. Springer, Cham. https://
doi.org/10.1007/978-3-319-57639-8_8
13. Abouzakhar NS, Jones A, Angelopoulou O (2017) Internet of things security: a review of risks
and threats to healthcare sector. In: IEEE international conference on internet of things (iThings)
and IEEE green computing and communications (GreenCom) and IEEE cyber, physical and
social computing (CPSCom) and IEEE smart data (SmartData). https://doi.org/10.1109/iTh
ings-GreenCom-CPSCom-SmartData.2017.62
14. Kammüller F (2019) Combining secure system design with risk assessment for IoT health-
care systems. In: IEEE international conference on pervasive computing and communications
workshops (PerCom workshops). https://doi.org/10.1109/PERCOMW.2019.8730776
15. Almulhim M, Zaman N (2018) International conference on advanced communication tech-
nology (ICACT). https://doi.org/10.23919/ICACT.2018.8323802
16. Chahid Y, Benabdellah M, Azizi A (2017) Internet of things protocols comparison, architecture,
vulnerabilities and security: State of the art. In: ACM international conference proceeding
series, a65
17. Kannouf N, Labbi M, Benabdellah M, Azizi A (2018) Security of information exchange
between readers and tags. In: Maleh Y, Ezzati A, Belaissaoui M (eds.) Security and privacy in
smart sensor networks, pp 368–396. IGI Global, Hershey. https://doi.org/10.4018/978-1-5225-
5736-4.ch016
18. Kannouf N, Douzi Y, Benabdellah M, Azizi A (2015) Security on RFID technology. In: 2015
international conference on cloud technologies and applications (CloudTech), Marrakech, pp
1–5. https://doi.org/10.1109/CloudTech.2015.7336997
19. Eddy M, Perlroth N (2020) Cyber Attack Suspected in German Woman’s Death, New York
Times, 18 September 2020
Design of an Automatic Control and IoT
Monitoring System for Dual Axis Solar
Tracker
Fatima Cheddadi, Hafsa Cheddadi, Youssef Cheddadi, Fatima Errahimi,
and Ikram Saber

Abstract This paper develops a combination of the implementation of a dual-axis

solar tracker and real-time monitoring system based on the use of the Internet of
Things. To obtain maximum efficiency, the photovoltaic panel must be constantly
oriented perpendicular to the sun’s rays. However, the position of the sun varies
continuously during the day from east to west along the azimuthal and during the
seasons along a zenithal axis. For this reason, we have designed and implemented
a dual-axis solar tracker that allows real-time tracking of the apparent movement of
the sun to maximize the power extraction. The proposed sun tracker is equipped with
a system of monitoring to supervise and evaluate its performances, by collecting
the electrical parameters from the PV panels, and the meteorological parameters
from its surface. Both the control and the monitoring is established using the ESP32
embedded board, and the designed support of solar tracker is made using a 3D printer.
The dual-axis solar tracker performances are compared with a fixed system.

Keywords Photovoltaic system · Internet of Things · ESP32 · Monitoring · Dual

axis solar tracker

1 Introduction

Due to the high demand of energy and the ever-increasing use of fossil resources,
the whole world has led to the development of the alternative solution and have
moved towards using renewable energy such as solar, wind, biomasses and hydraulic
energy to decrease the uses of fossil resources. In addition to meet the growing energy
demand, another advantage of the use of renewable energy is to reduce CO2 emissions
and the earth warming.

F. Cheddadi (B) · H. Cheddadi · Y. Cheddadi · F. Errahimi

SIGER, USMBA, Fez, Morocco
e-mail: [email protected]
I. Saber
TSI, USMBA, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 301
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_28
302 F. Cheddadi et al.

In this way, the photovoltaic (PV) energy is one of the cleanest energy resources
and the most ecological type of energy to use [1]. It is directly converted into electrical
energy by the use of PV modules. However, the harvested power mainly depends on
climatic variations namely the temperature and irradiance, where the need to track
the maximum power point (MPP). For this purpose, there are two strategies to extract
the maximum power. The first one is an electronic strategy based on the called MPPT
methods that try to find optimal operation points of the PV generators without any
movement. The second is based on electro-mechanical trackers that track the sun
movement during the day.
In the literature, several researchers have developed and implemented different
types of maximum power point tracking controller. These controllers differ from
each other in terms of many parameters, namely the number of required sensors,
convergence rate, rapidity of tracking the MPP when the climatic condition suddenly
change, complexity, and the cost of implementation [2].
Some of the MPPT strategies are based on adjusting the duty cycle of the DC-DC
converter [2–8]; the duty cycle plays the role of an adaptation stage between the load
resistance seen by the PV system and the internal PV module.
The other strategies are based on the implementation of solar trackers [9, 10],
where there are two main families of solar trackers: the passive ones and the actives
that include the single-axis and double-axis trackers.
The energy supplied by the PV panel is highly dependent on the amount of sunlight
absorbed by the panel. This amount depends on the orientation of the panel in relation
to the sun. In order to collect the maximum energy, the PV panel should be constantly
oriented perpendicular to the sunlight [11].
Every day this angle decreases or increases, the surface area (m2 ) of the panel
exposed to the rays decreases and therefore the power produced decreases, hence
the importance of the orientation of the panels in relation to the position of the sun
[12]. Another factor that influences the performance of the PV panel is the tilt angle,
which is the angle formed by the plane of the solar panel in relation to the horizontal.
The purpose of PV monitoring systems is to offer continuously a clear information
to improve energy efficiency, many researchers have proposed a monitoring system
of PV station, which help also on fault detection [13–15]. Authors in [16] have
presented a real-time monitoring system, which contain an alert system to a remote
user, when there is a deviation of solar power.
The main use of the system of monitoring in this work is to highlight the effec-
tiveness of the dual axis solar tracker by comparing the measured power from the
system of tracking with a fixed system.
The paper is organized as follows: Sect. 2 presents the proposed tracker. Section 3
presents the PV monitoring system. Results and a discussion are described in Sect. 4.
The paper is concluded in Sect. 5.
Design of an Automatic Control and IoT Monitoring System … 303

2 Dual Axis Solar Tracker Based on Light Sensors

In order to collect the maximum amount of energy, the PV panel must be constantly
oriented perpendicularly to the sun’s rays; this can be ensured by a solar tracking
system allowing to follow the sun throughout the day.
This method is based on light sensors or photo-resistors to determine the position
of the sun at any time, allowing instantaneous tracking, and including optimal panel
orientation. In this type of techniques, tracking is carried out independently of the
zone in which the system will be located.
The two types of solar trackers that exist are either single-axis or double-axis,
which are either are based on the same principle of pursuit illustrated in Fig. 1.
The technique used for this study is based on the difference in incident illumination
on photo-resistance or LDR (Light-Dependent-Resistor) light sensors separated by
opaque walls and placed on the PV panel. The control and command circuit compares
the signals emitted by the sensors by calculating their differences and then sends
pulses back to the motor to reposition the panel perpendicular to the sunlight.

LDR1
M1
Vertical
LDR2 System of
control
LDR3 (ESP32)
M2
Horizontal
LDR4

Fig. 1 General principle of tracking control

LDR LDR
Top left top right

LDR LDR
bottom bottom right
left

Fig. 2 Position of the four LDRs

304 F. Cheddadi et al.

Start

The initial position of the sun

TopRight=analogRead (LDR To-

Read the analog value of each pRight)
LDR sensor TopLeft=analogRead (LDR TopLeft)
BottRight =analogRead (LDR Bot-
tRight)

Calculate average values AvTop= (TopRight + TopLeft) / 2

AvBott= (BottRight + BottLeft) / 2
AvRight= (TopRight + BottRight) / 2
AvLeft= (TopLeft + BottLeft) / 2

Calculate the differences (azi-

DiffElev= diffVerti= AvTop - AvBott
muth && elevation) DiffAzim=diffHoriz= AvRight -
AvLeft

Yes
| DiffAzim | Stop the left-right servomotor

NO
Yes
The left-right servomotor moves
DiffAzim > 0
the panel to the right.
NO

Left-right servomotor moves the

panel to the left

NO
Yes
| DiffElev | <50 Stop the up-down servomotor

NO
Yes Up-down servomotor moves
DiffElev > 0
the panel upwards

The Up-down servomotor moves the

panel downwards

Fig. 3 Flowchart of the double-axis tracking method

Design of an Automatic Control and IoT Monitoring System … 305

Fig. 4 Solar tracker

prototype

Starting from the previous general principle, to ensure a complete tracking of the
solar path, 4 LDRs distributed in the 4 corners are used.
The technique used consists in checking the equality of illumination of four light
sensors distributed in the four quadrants represented in Fig. 2.
The operating principle is based on the tracking algorithm described in the
following flowchart Fig. 3.
In order to prototype our solar tracker Fig. 4, we use the 3D printing, which
is recognized as one of the biggest technological revolutions in the world today.
It represents a very important new technique for the manufacturing processes of
three-dimensional solid objects.

3 Photovoltaic Monitoring System

In this part, we will present the components necessary for the realization of the proto-
type. Some components collect data (INA219 as a current/voltage sensor, BH1750
as an illumination sensor, DHT11 as a temperature sensor). Other components act as
microcontrollers that process and analyze the data and provide wireless connectivity
(dual core ESP32), and for monitoring PV production, a web server is embedded to
store the daily collected data using the sensors to follow the metrics of our monitoring
system (ThingSpeak).
In this proposed conceptual system, the ESP32 DEVKITV1 board acts as a micro-
controller that processes incoming data from various sensors. The ESP32 is a series
of low-cost, low-power systems with Wi-Fi and Bluetooth integrated in a single chip.
This card is chosen because it reduces the cost of the monitoring system. The
ESP32 card is based on the microcontroller Xtensa LX6 dual core 32-bit processor
Tensilica, it has built-in Wi-Fi and Bluetooth, it runs 32-bit programs. The clock
frequency can go up to 240 MHz and has a RAM of 512 KB. It also offers a wide
variety of peripherals, such as: CAN, CNA, UART, SPI, I2C, it is equipped with an
306 F. Cheddadi et al.

Current
100
Current (mA)

80
60
40
20
0
13:26 13:33 13:40 13:48 13:55 14:02
Time

Voltage
14
12
Voltage (V)

10
8
6
4
2
0
13:26 13:33 13:40 13:48 13:55 14:02
Time

Power
1200
1000
Power (mW)

800
600
400
200
0
-20013:26 13:33 13:40 13:48 13:55 14:02
Time

Fig. 5 Current, voltage and power monitored

integrated Hall effect sensor and an integrated temperature sensor.so, it can support
monitoring and tracking programs.
The developed monitoring system is capable to measure in real time the energy
produced by a photovoltaic station composed of 3 polycrystalline PV panels placed
Design of an Automatic Control and IoT Monitoring System … 307

in parallel capable to generate 1.6 W of power and 12 V of voltage for each one at
the STC. To follow up and verify their correct operation and to analyze and have an
overview of the entire system, the figures above present the results of the monitoring
on 28/10/2020 in Fez Morocco.
Three parameters namely Power, current, and voltage, are monitored using the
experimental prototype Fig. 5. This latter is connected using WIFI and visualized
via ThingSpeak platform.
Between 13:50 and 13h55 the panels was exposed to the shadow which results
in a decrease in the amount of irradiation measured and consequently a decrease in
voltage, current and power have been detected.

4 Results and Discussion

The main objective of this work is to highlight the effectiveness of the dual axis solar
tracker by comparing the measured power from the tracking system with the fixed
system.
We have imitated the behavior and the movement of the sun, which is emulated
by a lamp in five positions as shown in the Fig. 6, for both systems (fixed and with
tracking system) at the temperature 21 °C.
We have taken measurements in the darkness. Figure 7 shows the electrical
measurements in real time, namely the current, voltage and power.
Figure 7 displays the dashboard with different widgets to present and visualize
the results of data collected using the solar tracker (blue curve) and compare them
with fixed system (red curve).
As it is mentioned, the system of control integrates an active method of tracking
to follow and track in real time the sun position in order to maximize the generated
power from the panel.
As it is shown in Fig. 7 and Table 1, even if the position of the sun changes, the
solar tracker is able to follow its position while extracting and keeping the maximum
power. While in case of using the fixed station, the power is maximum only when
the sun is in position 3 which corresponds to the midday productivity.
The increase of energy extracted by the solar tracker could be calculated by the
Eq. 1:

Fig. 6 Scenario of the sun position adopted during the experimentation

308 F. Cheddadi et al.

Power with solar tracker Power with fixed system

25

20

15
Power (mW)

10

0
20:02 20:09 20:16 20:24 20:31 20:38 20:45 20:52 21:00 21:07 21:14
-5
Time

4
Voltge (V)

0
20:02 20:09 20:16 20:24 20:31 20:38 20:45 20:52 21:00 21:07 21:14
Time

Voltage with solar tracker Voltage with fixed system

5
Current (mW)

0
20:02 20:09 20:16 20:24 20:31 20:38 20:45 20:52 21:00 21:07 21:14
Time

Current with solar tracker Current with fixed system

Fig. 7 Comparison between current, voltage and power monitored using solar tracker and fixed
system
Design of an Automatic Control and IoT Monitoring System … 309

Table 1 Power monitored from fixed system and using the proposed solar tracker
Position POS 1 POS 2 POS 3 POS 4 POS 5 Total
Fixed system (P1) 0 mW 4 mW 22 mW 4 mW 0 mW 30 mW
Solar tracker (P2) 20 mW 22 mW 22 mW 21 mW 18 mW 103 mW

  
P1 − P2
e f f iciency =  × 100 = 70.87% (1)
P1

We found the sum of power extracted by the fixed-axis by the three panels is equal
to 30 mW, and the sum of power generated using the proposed dual axis sun tracker
is equal to 103 mW, which means that the average efficiency of the used tracker is
equal to 70.87%.

5 Conclusion

The objective of this work was to present the results of the implementation hard-
ware of the active tracking method, based on the difference in illuminance of the
light collectors. The proposed sun tracker ensure a continuous maximum of power
extraction.
We implemented in parallel a real-time monitoring system to supervise remotely
the performance of the prototyped sun tracker and to validate the use of solar tracker.
This latter shows good results in comparison to the fixed-axis PV panels with an
important increase in average of 70.87%.

References

1. Youssef C, Fatima E (2018) A technological review on electric vehicle DC charging stations

using photovoltaic sources. In: IOP conference series: materials science and engineering, vol
353, no 1, pp 012014, 10. https://doi.org/10.1088/1757-899X/353/1/012014
2. Cheddadi Y, Errahimi F, Es-sbai N (2018) Design and verification of photovoltaic MPPT
algorithm as an automotive-based embedded software. Solar Energy 171:414-425. https://doi.
org/10.1016/j.solener.2018.06.085
3. Cheddadi Y, Cheddadi F, Errahimi F, Es-Sbai N (2017) Extremum Seeking Control-based
Global maximum power point tracking algorithm for PV array under partial shading conditions.
In: 2017 international conference on wireless technologies, embedded and intelligent systems
(WITS), Fez, Morocco, pp 1–6. https://doi.org/10.1109/WITS.2017.7934653.
4. Abdel-Salam M, El-Mohandes M-T, Goda M (2018) An improved perturb-and-observe based
MPPT method for PV systems under varying irradiation levels. Solar Energy 171:547–561.
https://doi.org/10.1016/j.solener.2018.06.080
5. Al-Dhaifallah M, Nassef AM, Rezk H, Nisar KS (2018) Optimal parameter design of fractional
order control based INC-MPPT for PV system. Solar Energy 159:650-664. https://doi.org/10.
1016/j.solener.2017.11.040.
310 F. Cheddadi et al.

6. Bayod-Rújula Á-A, Cebollero-Abián J-A (2014) A novel MPPT method for PV systems with
irradiance measurement. Solar Energy 109:95–104. https://doi.org/10.1016/j.solener.2014.
08.017
7. Bukar AL, Tan CW (2019) A review on stand-alone photovoltaic-wind energy system with fuel
cell: system optimization and energy management strategy. J Clean Prod. 221:73-88. https://
doi.org/10.1016/j.jclepro.2019.02.228
8. Bradai R et al. (2017) Experimental assessment of new fast MPPT algorithm for PV systems
under non-uniform irradiance conditions. Appl Energy 199:416-429. https://doi.org/10.1016/
j.apenergy.2017.05.045
9. Bounechba H, Bouzid A, Snani H, Lashab A (2016) Real time simulation of MPPT algorithms
for PV energy system. Int J Electr Power Energy Syst 83:67–78. https://doi.org/10.1016/j.ije
pes.2016.03.041
10. Awasthi A, et al. (2020) Review on sun tracking technology in solar PV system. Energy Rep
6:392-405. https://doi.org/10.1016/j.egyr.2020.02.004
11. Hafez AZ, Yousef AM, Harag NM (2018) Solar tracking systems: technologies and trackers
drive types – a review. Renew Sustain Energy Rev 91:754–782. https://doi.org/10.1016/j.rser.
2018.03.094
12. Mousazadeh H, Keyhani A, Javadi A, Mobli H, Abrinia K, Sharifi A (2009) A review of prin-
ciple and sun-tracking methods for maximizing solar systems output. Renew Sustain Energy
Rev 13(8):1800–1818. https://doi.org/10.1016/j.rser.2009.01.022
13. Adhya S, Saha D, Das A, Jana J, Saha H (2016) An IoT based smart solar photovoltaic remote
monitoring and control unit, p 5
14. Lopez-Vargas A, Fuentes M, Vivar M (2019) IoT application for real-time monitoring of solar
home systems based on ArduinoTM with 3G connectivity. IEEE Sensors J. 19(2):679–691.
https://doi.org/10.1109/JSEN.2018.2876635
15. Madeti SR, Singh SN (2017) Monitoring system for photovoltaic plants: a review. Renew
Sustain Energy Rev 67:1180–1207. https://doi.org/10.1016/j.rser.2016.09.088
16. Cheddadi Y, Cheddadi H, Cheddadi F, Errahimi F, Es-sbai N (2020) Design and implementation
of an intelligent low-cost IoT solution for energy monitoring of photovoltaic stations. SN Appl
Sci 2(7):1165. https://doi.org/10.1007/s42452-020-2997-4
Analysis Jamming Attack Against
the Protocol S-MAC in IoT Networks
Imane Kerrakchou , Sara Chadli , Mohamed Emharraf ,
and Mohammed Saber

Abstract The Internet of Things (IoT) is one of the growing developments of the
recent period that has to a great extent pulled in industry and the scholarly community.
Life sans IoT is indispensable. In order to disperse any questions about its far-reaching
reception, IoT positively requires both actually and consistently right answers to
guarantee fundamental security and protection. This paper gives a brief description
of IoT and the different technologies utilized, for example, WSN, and inspects known
attacks against remote sensor systems like a Jamming attack of DoS type. Among the
current MAC protocols, the S-MAC protocol is proposed to analyze the performances
of the WSN in case of a Jamming attack using the OMNeT++ test system. Distinctive
application scenarios have been assessed. Performance parameters, for example,
consumed energy, lifetime, system loading, and packet reception rate are the principal
factors taken in our investigation.

Keywords Internet of Things (IoT) · Attack · Jamming · DoS · S-MAC · Wireless

sensor network (WSN)

I. Kerrakchou (B) · S. Chadli · M. Emharraf · M. Saber

Mohammed First University Oujda, ENSA Oujda, SmartICT Lab, Oujda, Morocco
e-mail: [email protected]
S. Chadli
e-mail: [email protected]
M. Emharraf
e-mail: [email protected]
M. Saber
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 311
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_29
312 I. Kerrakchou et al.

1 Introduction

The Internet of Things (IoT) is an evolving technique of immense social, technolog-

ical, and economic value. Present forecasts about the effects of IoT are very remark-
able. It is expected that 100 billion linked IoT devices will be utilized by 2025. It
would also have a world economic effect of over $11 trillion [1]. The Internet of
Things (IoT) could make every aspect of our lives easier. The prospect of a more
comfortable daily life, safer road traffic, a more environmentally energy supply, and a
healthier lifestyle promote its development. Possible applications cover a wide range
of areas of life [2]. The primary thought of IoT is to empower programmed corre-
spondence between objects of regular daily existence (things that have been given
capacities for recognizing, transmitting, and handling data) using the Internet and to
connect a local network of objects to a global network utilizing standardized proto-
cols. Intelligent objects (for example things) can be communicating using different
varied network technologies, for example, Wireless Sensor Networks (WSN), Radio
Frequency Identification (RFID), Wireless Local Area Networks (WLAN), and cell
systems (3G, 4G, LTE, and 5G) [3]. IoT makes things and individuals equipped for
associating with anybody and anything, anywhere, utilizing any kind of media and
service.
IoT system faces various dangers and vulnerabilities [4], including the commu-
nication canal, which permits the attacker to mess with information in travel. The
detection device has limited computing power and low storage capacity. Sensor nodes
are not furnished with sealed equipment because of cost requirements. On the off
chance that a huge level of system nodes, even a couple of basic nodes, are assaulted,
the life of the network might be decreased to a few days. The main cause of energy
waste in these sensor systems is the radio subsystem. The data link layer, particularly
the MAC (Media Access Control) protocol, is liable for dealing with the radio. In this
manner, the MAC protocol should maintain the radio in a non-powered standby state
as far as could reasonably be expected. With this in mind, many searches in the field
of sensor node energy conserving concentrate on MAC protocols. For this reason,
this paper will examine the sensor MAC (S-MAC) protocol [5]. Two scenarios are
proposed in order to compare network performances with and without Jamming
attack.
The rest of this document is presented as follows. Section 2 defines the security
objective of IoT and the classification of attacks. Section 3 provides an overview of the
functioning of the S-MAC protocol and the attack targeting this protocol. Section 4
explores the impact of Jamming attacks against the MAC protocol presented in
Sect. 3 by describing their implementation on the simulator. Finally, Sect. 5 presents
a conclusion.
Analysis Jamming Attack Against the Protocol S-MAC in IoT Networks 313

2 IoT Security

2.1 IoT Security Objective

The general objective is to protect the whole system that represents an IoT installation
against a wide range of attacks. The most granular security prerequisites frequently
alluded to as security characteristics [6], are:
– Confidentiality: A concept to ensure that information can only be read by
authorized persons.
– Availability: the objective is to guarantee the continuity of network services
against DoS attacks.
– Integrity: Set of means and techniques to limit the modification of data to
authorized persons.
– Privacy: the objective is to prevent the disclosure of confidential information to
malicious entities and the illegal collection of sensitive information about entities.

2.2 IoT Attacks Classification

In the course of the most recent times, the IoT network was confronted with various
attacks that made the makers and consumers aware of the need to utilizing IoT
equipment with more caution. This section explains the various types of attacks in IoT,
which are physical and cyber-attacks where this last comprise of active and passive
attacks (see Fig. 1). Cyber-attack is a danger that aims at diverse IoT equipment in a
wireless system by breaking the network for manipulation of the user’s information
(i.e., delete, destroy, steal, corrupt). Furthermore, physical attacks allude to the attacks
that interrupt the service and physically harm the IoT equipment.
The following paragraph presents the two kinds of cyber-attacks according to the
gravity in IoT system:

Fig. 1 The different types

and surfaces of attack in IoT
314 I. Kerrakchou et al.

– Active attack: in order to interrupt certain services and manipulate the configura-
tion of the system, an active attack happen by an intruder that gets to the network
and its relating information. There are various methods of attacking the security
of IoT devices, including disruptions, modifications, and interventions. Among
active attacks, we can cite Jamming, Sybil, DoS, MITM, Spoofing, etc.
– Passive attack: this attack attempts to accumulate user information with no his
assent, using this data to break the security of his private information. Traffic
analysis and Eavesdropping are the two fundamental elements to play out a passive
attack using an IoT system.
This subsection presents the generic IoT architecture, which consists primarily of
three layers (see Fig. 1), in order to describe the possible attacks according to each
layer. The IoT surface attacks are classified as follows [7]:
– Perception surface attacks: the attackers may easily penetrate the physical layer
of the IoT system, because the Micro-controllers, Sensors, actuators are a few
pieces of the physical system that are utilized for communication, identification,
and collecting information, and these elements are very susceptible to Jamming,
radio interference, and DoS. Furthermore, physical attacks are the most disturbing
for perception surfaces [8].
– Network surface attacks: In IoT systems, the physical devices are associated
through network services, including wireless networks. This wireless sensor
network is a primary objective for various forms of attacks because user infor-
mation moves overtly with no solid secure protocol. Attacks on the surface of
networks are likely to occur through MITM, DoS, Sybil, Routing, Spoofing, and
so on.
– Application surface attacks: The application layer manages various types of
applications, which allow remote control, and access to IoT devices. This IoT
equipment is linked to the network with clouds and servers. They utilize software-
based applications that allow malicious developers to dispatch distinctive malware
attacks in order to get to IoT devices with or without user authorization. Therefore,
attacks targeting this layer are DoS, Data corruption, etc.

3 Jamming Attack on S-MAC Protocol

3.1 The Functioning of S-MAC Protocol

For wireless sensor networks, the MAC protocols in which nodes have a sequence
of active and inactive periods are called activity rate protocols. The activity rate
represents the percentage of time spent in the period of activity. Protocols with activity
rate may be classified into two main types. The first type is that of synchronous MAC
protocols, in which all nodes have their period of activity simultaneously or planned.
The second category is that of asynchronous MAC protocols, in which the nodes
have activity periods independent. There are many MAC protocols, among which
Analysis Jamming Attack Against the Protocol S-MAC in IoT Networks 315

Fig. 2 Sequence of activity and inactivity periods in S-MAC

we can cite: S-MAC, T-MAC, B-MAC, etc. In this document, we will talk about
the S-MAC, which is the synchronous protocol. Nodes switch between periods of
activity and periods of inactivity by a fixed action rate. During the active period, a
node can send and get data. The inactive period corresponding to the deactivation of
the radio transmitter/recipient to spare energy. The active period is itself divided into
a synchronization period utilizing a SYNC frame, and a data trade period. The nodes
keep up a table with the wake-up times of their neighbors. At the point when a node is
active, it listens to the channel for a while. If it hears a SYNC frame, it synchronizes
to the schedule of the node transmitting this SYNC frame. If it hears no frame after
this time, it embraces its schedule and communicates it to its neighborhood. All
nodes that follow this schedule form a virtual cluster. At the point when a node has a
frame to transmit, it utilizes a reservation mechanism such as RTS/CTS (for Request
To Send/ Clear To Send) and ACK (Acknowledgement) based on the IEEE 802.11
norm (see Fig. 2) [9].

3.2 Jamming Attack

Jamming is a type of attack that was planned to intentionally intervene with the typical
functioning of the wireless device, at the perception layer, all the more exactly in the
MAC layer, by saturating the canal through the injection and persistent transmission
of data or control package, causing abnormalities in the transmission and reception
of data. What’s more, this attack reduces the performance of the system by increasing
power consumption, memory, bandwidth, etc. Generally speaking, this sort of attack
is associated with Denial of Service attacks (DoS) [10].
However, the Jamming attack could be classified according to the level of knowl-
edge of the protocol used to apply it. Without knowledge of the protocol utilized in
the network, attacks are very restricted in their effects on network performances. On
the other hand, knowing the communication protocol utilized in the system can be
destructive to a WSN. With complete information of the MAC protocol, an attacker
can create big traffic to increase the most extreme effect from Jamming attacks.
316 I. Kerrakchou et al.

4 Jamming Attack Implementation and Results

In our situation, we are going to launch two types of attacks. First, we assume that the
attacker does not know the protocol utilized. For this reason, he will use DATA packets
to launch this attack. Therefore, he starts replaying these data packets throughout the
network aiming at the sink. Secondly, we suppose that the attacker knows the protocol
used in the network. He analyzes the traffic to determine the protocol utilized in the
system, which is the S-MAC. Then, he starts replaying explicit data messages by
targeting the sink. In this case, the attacker utilizes retransmission of DATA packets
correctly programmed in the S-MAC protocol in order to waste more energy from the
network. Moreover, even in the case of Jamming, as indicated by the MAC protocol,
the lifetime of a WSN can be long due to the idle cycle that disables the nodes,
forcing the attacker to scrambling the system for all simulation time.

4.1 Simulation OMNeT++

First of all, the simulator used in our work is OMNeT++. It is a powerful network
simulator and especially useful when designing and validating new protocols, and
exploring new or exotic scenarios.
In this simulation experience, two scenarios are recreated that have almost the
same qualities with the difference that the first scenario (see Fig. 3) is a classic
situation with no malicious nodes or attacks on the network. We talk about the
normal case.
The second scenario (see Fig. 4) represents the attack of the jammer on the nodes.
This scenario contains one jammer that injects unauthorized traffic into the network
and affects the WSN, which has no particular mechanism for detecting or fore-
stalling Jamming attacks. In this scenario, we simulate the two types of attack that
we explained in the paragraph above. In the first type (no knowledge of the protocol),
the attacker continuously transmits DATA packets in the transmission channel. In the

Fig. 3 Scenario 1—normal

case
Analysis Jamming Attack Against the Protocol S-MAC in IoT Networks 317

Fig. 4 Scenario 2—network

with Jamming attack

second type (knowledge of the protocol), the attacker synchronizes with all the nodes
and continuously transmits DATA packets during the activity period.
The jammer used in this scenario is a no-mobile jammer that allows DATA
messages to be transmitted to the sink in order to inject a specific number of packets
into the network and subsequently cause it to fail. These Jamming attacks are one
of the most effective types of Jamming because they reduce the life of the network.
The jammer parameters are shown in Table 1 below:
The simulation model examined utilizes a star topology, where communication
takes place between the nodes and the sink. The three simulations consist of 11 fixed
nodes. Every node is equipped with two AA batteries, enough for a long working
span. The duration of the simulation is set at 100 s. So as to identify the effect of the
attack, the simulation time stays consistent as well as the total number of nodes for
both scenarios. The simulation parameters are shown in Table 2 below:
The principal objective for simulating scenario 1 is to determine the current state of
the network under ordinary circumstances, which will allow us to make comparisons
with scenario 2 and distinguish the effect of a Jamming attack on the network.
However, the goal of the two attack simulations is to determine the degree of the
Jamming attack effect on the network performances.

Table 1 Simulation
Parameters Values
parameters for Jammer node
No. of jammers 1
Transmit Power (mW) 57.42
Trajectory Fixed
Data rate (pps) 50
Data packet size (bytes) 250
End time (s) End of simulation
318 I. Kerrakchou et al.

Table 2 Simulation
Parameters Values
parameters
Simulation time (s) 100
Simulation area (m) 40 × 40
Topology Star
No. of nodes 11
Mobility model No mobility
Transmit power (mW) 46.2
Data rate (pps) 5
Data packet size (bytes) 100
End Time End of simulation
Protocol S-MAC
SYNC packets size (bytes) 11
Frame time (ms) 610
Contention period (ms) 10

4.2 Results and Discussion

In this section, network performance behavior is analyzed with and without Jamming
attacks. For the second scenario, both attack simulations used a higher packet size,
packet rate, and transmission power compared to scenario 1 where no attack is imple-
mented. Based on the results of the simulation experiences, the following outcomes
are obtained.
Figure 5 presents the consumed energy by all the nodes, for the three situations.
For the normal case, we notice that the S-MAC has better performance. The energy
consumption is normal, which means that the network is functioning properly. At the
point when the network is subject to a Jamming attack, with or without knowledge of
the protocol used in the network, the system has a much higher energy consumption
compared to the result obtained in scenario 1.
In our example, the role of the sink (node 0) is to receive and collect packets from
the entire network, so it does not transmit data packets. What explains its high energy
consumption in the case of the no knowledge of the protocol compared to the normal
case, is the reception of the traffic generated by the attacker. On the other hand,
we notice that in the case of the knowledge of the protocol used, the consumption
is much higher because of the reception and processing of a large number of fake
packets sent. The knowledge of the protocol used in the network allows the attacker
to behave like an ordinary node, which allows him to deceive the sink and let it
process its fake packets.
Moreover, the continuous sending of these data packets allows an attacker to
force other nodes to send more frames than expected since a communication failure
is usually followed by several more attempts. With each new attempt, the nodes
consume even more power.
Analysis Jamming Attack Against the Protocol S-MAC in IoT Networks 319

Fig. 5 Energy consumption per node (J)

Once the power consumption is high, network life will decrease. The life of a
network is defined as the average time between the deployment of the network and
the moment when the power supply to the nodes is exhausted. Figure 6 shows the
estimated lifetime of the network for the three situations. In the case of the attack,
the life of the network is much shorter than in the ideal case. This is due to the
Jamming attack, which increased traffic, resulting in wasted energy and therefore
reduced network life.
Figure 7 represents the number of DATA packets sent by all the network sensors
and received by the Sink. We can see that sending and receiving packets work
correctly in the normal case. Once the attack is implemented, the number of packets
sent and received decreases.
This decrease in the number of packets sent is due to the high throughput broadcast
in the network by the malicious node. The attacker increased traffic and made the
transmission channel busy, which prevented other nodes in the network from sending
their packets properly and communicating with the Sink. As far as packet reception
is concerned, we can see that the Sink can no longer correctly receive the packets sent
to it by the legitimate nodes due to its occupation with receiving the fake packet and,

Fig. 6 Estimated network lifetime (days)

320 I. Kerrakchou et al.

Fig. 7 Number of sent and received packets in the network

Fig. 8 Packets reception rate per node

above all, with processing the fake packets in the case of knowledge of the protocol
used. Which explains the gross decrease in the reception rate. Figure 8 shows the
decrease in packet reception rate by each node.
From the results obtained, it is deduced that both types of attacks harm network
performance. On the other hand, the attack with knowledge of the protocol used
in the system is much more effective compared to an attack without knowledge,
because the attacker, who already knows the protocol used, diffuses the traffic in the
network following all the rules of the MAC protocol used for collision avoidance,
synchronization, etc.

5 Conclusion

Most current IoT security research is concerned with confidentiality and integrity
data, to a great extent disregarding data availability. Without the capacity to secure the
physical medium in which communication happens, sensor networks are susceptible
Analysis Jamming Attack Against the Protocol S-MAC in IoT Networks 321

to a progression of potential attacks that focus on quickly depleting the batteries of

sensor nodes and rendering the network unusable.
This paper provides an overview of IoT, specifically WSN, with a focus on
Jamming attacks and the power efficiency of wireless systems. The four parame-
ters analyzed to determine network performances are power consumption, lifetime,
the network load, and packet reception rate. The effects on these performances using
the S-MAC protocol and the impacts of Jamming attacks on wireless communication
networks were studied using the OMNET++ simulator.
Future work will investigate a defensive system and discover ways to apply it to
currently available detection devices so as to develop specific mechanisms to secure
them against this type of attack and to plan new energy-efficient MAC protocols.

References

1. Manzoor A, Braeken A, Kanhere SS, Ylianttila M, Liyanage M (2020) Proxy re-encryption

enabled secure and anonymous IoT data sharing platform based on blockchain. J Netw Comput
Appl 176:102917
2. Hassija V, Chamola V, Saxena V, Jain D, Goyal P, Sikdar B (2019) A survey on IoT security:
application areas, security threats, and solution architectures. IEEE Access 7:82721–82743
3. Khattak HA, Shah M, Khan S, Ali I (2019) Perception layer security in the Internet of Things.
Future Gener Comput Syst 100:28
4. Sengupta J, Ruj S, Bit SD (2020) A comprehensive survey on attacks, security issues and
blockchain solutions for IoT and IIoT. J Netw Comput Appl 149:102481
5. Jagriti, Lobiyal DK (2018) Energy consumption reduction in S-MAC protocol for wireless
sensor network. Procedia Comput Sci 143:757–764
6. Tahsien SM, Karimipour H, Spachos P (2020) Machine learning based solutions for security
of Internet of Things (IoT): a survey. J Netw Comput Appl 161:102630
7. Malik V, Singh S (2020) Evolutionary computing environments: implementing security risks
management and benchmarking. Procedia Comput Sci 167:1171–1180
8. Mihajlov B, Bogdanoski M (2014) Analysis of the WSN MAC protocols under jamming DoS
attack. Int J Netw Secur 16:516–524
9. Maitra T, Roy S (2016) A comparative study on popular MAC protocols for mixed wireless
sensor networks: from implementation viewpoint. Comput Sci Rev 22:107–134
10. López M, Peinado A, Ortiz A (2019) An extensive validation of a SIR epidemic model to study
the propagation of jamming attacks against IoT wireless networks. Comput Netw 165:106945
Simulation and Analysis of Jamming
Attack in IoT Networks
Imane Kerrakchou , Sara Chadli , Amina Kharbach ,
and Mohammed Saber

Abstract The Internet of Things (IoT) is one of the latest increasing technologies
that has largely been drawn to the industry and the academic world. Life sans IoT is
completely imperative. In order to dispel any concerns about its far-reaching accep-
tance, IoT needs both real and reliable responses to ensure essential security and
defense. This paper provides a brief overview of IoT and the various technologies
used, such as WSN, and discusses attacks targeting wireless sensor networks, such
as DoS style Jamming attacks. The S-MAC protocol is suggested among the existing
MAC protocols to evaluate the WSN performance in the event of a Jamming attack
utilizing the OMNeT++ test system. Distinctive scenarios have been tested. The main
parameters taken into account in our survey, to analyze network performance, are
the number of packets delivered, energy consumed, network lifetime, and the rate of
lost packets.

Keywords Internet of Things (IoT) · Attack · Jamming · DoS · S-MAC · Wireless

sensor network (WSN)

1 Introduction

Kevin Ashton proposed the term Internet of Things for the first time in 1999. It didn’t
get much consideration after that for about 10 years. However, massive interest was

I. Kerrakchou (B) · S. Chadli · A. Kharbach · M. Saber

Mohammed First University Oujda, ENSA Oujda, SmartICT Lab, Oujda, Morocco
e-mail: [email protected]
S. Chadli
e-mail: [email protected]
A. Kharbach
e-mail: [email protected]
M. Saber
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 323
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_30
324 I. Kerrakchou et al.

paid to IoT technologies in 2014 [1]. Any part of our lives could be made simpler
by the Internet of Things (IoT). Its development is motivated by the promise of a
more convenient everyday life, safer road traffic, a more environmentally sustainable
source of energy, and a healthy lifestyle. A broad number of areas of life [2] are
protected by future implementations. The key concept of IoT is to empower structured
communication between objects of normal life (things that have been granted data
identification, transfer, and handling capabilities) utilizing the Internet and to connect
a local object network using uniform protocols to a global network. Intelligent objects
(things) can be interacted utilizing different network technology, like Wireless Local
Area Networks (WLAN), Radio Frequency Identification (RFID), Wireless Sensor
Networks (WSN), and Cellular Services (3G, 4G, LTE, and 5G) [3]. IoT allows
things and people equipped to use any sort of media and service, to communicate
with anybody and anything, everywhere.
There are numerous dangers and vulnerabilities facing the IoT device [4],
including the communication canal, which helps the attacker to manipulate the trans-
mitted data. There are minimal processing power and poor storage space for the
detection device. Due to price specifications, sensor nodes are not furnished with
sealed facilities. If a large number of machine nodes, or a few simple nodes, are
targeted, the life of the network could be limited to a few days. The radio subsystem
is the primary source of energy waste in these sensor systems. For interacting with
the radio, the data link layer, in particular, the MAC (Media Access Control) protocol
is responsible. In this case, the MAC protocol can retain the radio in a non-powered
standby state. Much work in the field of sensor node energy conservation focuses
on MAC protocols. The sensor MAC protocol (S-MAC) will be investigated for this
purpose in this article. In order to compare network performance with and without
the Jamming attack, two cases are proposed.
The remainder of this paper is presented as followed. The security objective of
IoT and the classification of attacks are presented in Sect. 2. A description of the
S-MAC protocol functioning and the attack targeting this protocol are given in Sect. 3.
Through explaining their implementation on the simulator, Sect. 4 discusses the effect
of Jamming attacks against the MAC protocol used in this paper. Finally, a conclusion
is presented in Sect. 5.

2 IoT Security

2.1 IoT Security Objective

Mechanisms to ensure the security, availability, and confidentiality of services deliv-

ered by IoT must be provided by software, protocols, and infrastructures of this new
technology [5]. On the basis of the following requirements, each IoT application can
include facilities for protecting confidential information.
Simulation and Analysis of Jamming Attack in IoT Networks 325

– Authenticity: For communication between any two nodes, only allowed nodes
can be involved.
– Confidentiality: The leakage of any unauthorized recipient of classified infor-
mation shall be excluded.
– Integrity: When the information is transmitted to IoT devices, data integrity guar-
antees that the originality of the information is not rewritten, copied, or substituted
by the attacker.
– Privacy: In order to preserve privacy, the identity of an individual user should be
covered by a secure IoT device.
– Availability: Approved users can use different services provided by the protocols
of the IoT network to facilitate access to services.

2.2 IoT Attacks Classification

Different IoT systems are separated into different layers: Perception Layer, Network
Layer, and Application layer [6]. Every layer has a feature of its own. To form a full
IoT system, these layers work together. In this section (see Fig. 1), we briefly discuss
the functioning of the IoT layers with the most common security attacks targeting
each of them [7].
– Perception layer: the sensors are used in this layer to detect and acquire infor-
mation about the environment in which they are implemented. The protocols
related to communication between IoT network devices are included in this layer.
In terms of battery energy, storage resources, and computing capability, IoT
devices are resource-restricted. Among the attacks targeting this layer, we can
find Eavesdropping, Jamming attack, physical attacks, etc.

Fig. 1 Layered
classification of IoT attacks
326 I. Kerrakchou et al.

– Network layer: This layer is linked to the provision of network connectivity.

The data gathered from the perception layer is delivered through this layer to the
specific device for processing. It is the duty of the network layer protocols to
decide the best way to route data packets from one host to another on the network.
This layer is also susceptible to attacks such as Sybil attack, Wormhole attack,
Man In The Middle, etc.
– Application layer: this layer forms the applicable section of the whole network
with the implementation of the service provided for users. It allows the utilization
of different protocols to efficiently provide the users with the requested service.
The layer of the application acts as an interface between the users and the IoT
network. Therefore, it is vulnerable to threats that can impact nodes or application
systems such as attacks on stability, malicious code injection attacks, etc.

3 Jamming Attack on S-MAC Protocol

3.1 The Functioning of S-MAC Protocol

S-MAC is a protocol based on contention, where time is fragmented into large frames.
Two phases are composed of each frame (see Fig. 2):
– Active period: where the node can connect with neighboring nodes, uses control
packets like SYNC (synchronize), RTS (Request To Send), CTS (Clear To Send)
and, ACK (Acknowledgment) and sends DATA messages.
– Sleeping period: of which the node radio is switched off to prevent loss of energy.
S-MAC applies what is referred to as virtual clustering in order to achieve some
coordination between nodes, in which these last regularly transmit special SYNC
packets at the beginning of every frame to ensure that the node and its neighbors
wake up simultaneously. Virtual clusters are created by neighboring nodes and a
shared sleep schedule is set up [8].

Fig. 2 Frame for S-MAC protocol

Simulation and Analysis of Jamming Attack in IoT Networks 327

3.2 Jamming Attack

Jamming is a form of Denial-of-Service attack which may occur at the level of

the perception layer. This attack stops the channel from being used via nodes to
connect, by occupying it through injection and persistent data or control packages
transmission. The key objective of a DoS attack is to deplete node resources, like
bandwidth, battery life, and storage, and stop sensor data transmitted from reaching
the destination, thus, impacting its long term availability. In a WSN, a Jamming
attack is catastrophic as it does not require the realization of any specific hardware
device or software [9].

4 Attack Implementation and Results

4.1 Proposed System

The architecture of the system we have proposed consists of a base station (SINK),
which is used to collect and process data centrally, and a set of nodes that allows
packets to be sent. The system will present a Jamming type attack. In our case, it
is assumed that the attacker knows the protocol used in the network, which is the
SMAC. Once the malicious node is implemented in the network, the attacker analyzes
the traffic to determine the protocol used. Then, he will continuously transmit a large
number of RTS (Request To Send) control packets in the transmission canal. Here,
two situations will be studied. In the first situation, the attacker will target an ordinary
node other than the SINK. As soon as the malicious node is installed in the network, it
will look for the node closest to it and which belongs to this system. Once detected, the
Jammer node will start attacking this closest node by sending a lot of control packets
generating high traffic. In the second situation, the malicious node will attack only
the SINK. As a result, the flow of packets sent by the attacker will be transmitted to
the base station.

4.2 Simulation

The simulator used in this document is the OMNeT++ simulator. We use a set of 19
fixed nodes, where node 0 is the Sink. The simulation period is set to 200 s. We use
a star architecture in which all the nodes send their packets to the base station. The
initial energy of each node is 18720 J. The protocol used is SMAC. The transmission
power is fixed at 36.3 mW and the number of packets sent equals five packets per
second. The table below represents the simulation parameters (Table 1).
328 I. Kerrakchou et al.

Table 1 Simulation
Parameters Values
parameters.
Simulation time (s) 200
Simulation area (m) 60 × 60
Topology Star
No. of nodes 19
Mobility model No mobility
Transmit power (mW) 36.3
Packet rate (pps) 5
Data packet size (bytes) 100
End time End of simulation
Protocol S-MAC
SYNC packets size (bytes) 11
RTS packets size (bytes) 13
Frame time (ms) 610
Contention period (ms) 10

Fig. 3 Scenario 1—Normal

case

Fig. 4 Scenario 2– Network

with Jamming attack_NODE

In our simulation, three scenarios are proposed. The first scenario (see Fig. 3)
represents a network without attack, i.e. no malicious node is implemented in the
system. The nodes synchronize with each other and then transfer data packets to the
base station.
Simulation and Analysis of Jamming Attack in IoT Networks 329

Fig. 5 Scenario
3—Network with Jamming
attack_SINK

Table 2 Simulation
Parameters Values
parameters for Jammer node
No. of jammers 1
Transmit power (mW) 57.42
Trajectory Fixed
Packet rate (pps) 30
Protocol SMAC
SYNC packets size (bytes) 11
RTS packets size (bytes) 200
Frame time (ms) 610
Contention period (ms) 10
End time (s) End of simulation

The second scenario (see Fig. 4) represents the Jamming attack by implementing
a malicious node, which is node 19, in the network. In this scenario, we will simulate
the attack that targets an ordinary node, which is node 18, of the network as explained
in Sect. 4.1.
In our experiment, we added a third scenario (see Fig. 5) which also represents
the Jamming attack in the network with the difference that in this third scenario, the
attacker targets the base station.
The second and third scenarios have the same values of the parameters of the
attacker node (jammer node), in order to study the difference in the impact of each
of these attacks.
The Jamming used in our simulation is represented by a fixed node, which allows
the sending of a large number of control packets (RTS) with a sending rate equals to
30 packets per second. The transmission power of 57.42 mW and a control packet
size (RTS) equals to 200 bytes. The table below shows the parameters of the jammer
node (Table 2):
330 I. Kerrakchou et al.

Fig. 6 Number of packets sent per node

4.3 Results and Discussion

This section describes the performance analysis of the three proposed scenarios. The
objective of the simulation of scenario 1 (normal case) is to analyze the behavior
of the network in the absence of undesirable effects, and finally to compare it to a
network under attack. Then, we proposed in scenario 2, the attack of a network node,
and in scenario 3, the attack of the base station in order to compare the severity of
these attacks on the network. The results obtained are as follows:
Figure 6 shows the number of packets sent by each node, in the network, for
the three scenarios. In the normal case (scenario 1) all nodes send almost the same
number of packets. Therefore, the system is working properly. In scenario 2, where
one node in the network (node 18) was attacked, we notice that the whole network
works in the same way as the normal case, except the attacked node that sends only
a very small number of packets. The reason is that the attacker continuously sends
a set of false packets to the attacked node and prevents it from communicating with
the rest of the network. For scenario 3, we notice that the attack has infected the
whole network, which pushed all the nodes to send a very large number of packets.
The reason for this is that the attacker has prevented the sink from responding to the
other nodes, and when the other nodes do not receive any response from the base
station, they keep sending the packets back until they get a response or their batteries
run out.
Figure 7 shows the packet loss rate by each node, on the network, for the three
simulations. For Scenario 1, it can be seen that the packet loss rate is approximately
zero, meaning that all packets that have been sent by the network nodes reach the base
station. Therefore, the network is functioning properly. For Scenario 2, we notice
that the packet loss rate is a bit high compared to Scenario 1 due to the high traffic
generated by the attacker in the transmission canal, which prevented the routing of
some packets. For the attacked node (node 18), almost half of the sent packets were
not routed to the SINK because of the Jamming attack that interrupted the normal
operation of the packet transmission process with huge unnecessary packets. On the
Simulation and Analysis of Jamming Attack in IoT Networks 331

Fig. 7 Packet loss rate per node

Fig. 8 Energy consumption per node (J)

other hand, in scenario 3, almost all the packets sent by all the nodes did not reach
the base station because this attack left the sink occupied by the fake packets sent by
the attacker and made it unable to receive the packets sent by the other nodes.
The graph in Fig. 8 shows the battery energy consumption of each node in Joule.
It can be seen that the energy consumed in scenarios 2 and 3 is higher than that
consumed in the normal case for the entire network because of the attacker, which
generated high traffic in the transmission canal and led the nodes to waste their
energy in unnecessary retransmissions. We notice that the energy consumption of
the attacked node (node 18) for scenario 2, and the base station (node 0) for scenario 3,
is also very high because of the successive sending of large RTS messages generated
by the attacker.
The estimation of the energy consumed in the network, more specifically of the
nodes, also allows estimating the lifetime of the system, which is an essential param-
eter of WSN. Thus, the high energy consumption of each node automatically reduces
the life of its battery and then that of the network as shown in Fig. 9.
332 I. Kerrakchou et al.

Fig. 9 Estimated network

lifetime (days)

As mentioned above, the objective of the simulation of scenarios 2 and 3 is to

analyze the effect of each of these two attacks. According to the results obtained, we
notice that the attack that targets the base station is much more efficient compared to
the one that targets an ordinary node and has the greatest impact on the network. An
attack that targets the SINK affects almost all the nodes that make up the system in a
very significant way and degrades network performance, such as the very high energy
consumption of all the nodes, the loss of packets transmitted to the base station, etc.
The thing that destroys the whole system.

5 Conclusion

Most recent IoT security studies are concerned with knowledge on integrity and
confidentiality, generally without regard to the availability of data. Sensor networks
are vulnerable to a progression of possible attacks that concentrate on gradually
depleting the batteries of sensor nodes and making the network unusable in the case
of impossibility to protect the physical canal where contact takes place.
A summary of IoT, specifically WSN, with an emphasis on Jamming attacks
and the power efficiency of wireless systems is given in this article. For evaluating
network efficiency, four parameters were utilized: the number of packets delivered,
energy consumption, the lifetime of the network, and the loss rate of packets. Using
the OMNET++ simulator, the impacts on these performances and the effects of
Jamming attacks on wireless communication networks were analyzed using the S-
MAC protocol.
Future research will analyze a defense framework and find ways to adapt it to
currently existing detection devices in order to build unique methods to defend them
against this form of attack and to prepare new energy-efficient MAC protocols.
Simulation and Analysis of Jamming Attack in IoT Networks 333

References

1. Jha RK, Kour H, Kumar M, Jain S (2020) Layer Based Security in Narrow Band Internet of
Things (NB-IoT). Comput Netw 185:107592
2. Anand S, Sharma A (2020) Assessment of security threats on IoT based applications. Materials
today: proceedings
3. Khattak HA, Shah M, Khan S, Ali I (2019) Perception layer security in the Internet of Things.
Future Gener Comput Syst 100:28
4. Yugha R, Chithra S (2020) A survey on technologies and security protocols: reference for future
generation IoT. J Netw Comput Appl 169:102763
5. Tahsien SM, Karimipour H, Spachos P (2020) Machine learning based solutions for security of
Internet of Things (IoT): a survey. J Netw Comput Appl 161:102630
6. Malik V, Singh S (2020) Evolutionary computing environments: implementing security risks
management and benchmarking. Procedia Comput Sci 167:1171–1180
7. Hassija V, Chamola V, Saxena V, Jain D, Goyal P, Sikdar B (2019) A survey on IoT security:
application areas, security threats, and solution architectures. IEEE Access 7:82721–82743
8. Jagriti, Lobiyal DK (2018) Energy consumption reduction in S-MAC protocol for wireless sensor
network. Procedia Comput Sci 143:757–764
9. Osanaiye O, Alfa AS, Hancke GP (2018) A statistical approach to detect jamming attacks in
wireless sensor networks. Sensors 18:1691
A Brief Survey on Internet of Things
(IoT)

Fatima Zahra Fagroud, Lahbib Ajallouda, El Habib Ben Lahmar,

Hicham Toumi, Ahmed Zellou, and Sanaa El Filali

Abstract Today it be a necessity to setting off past the thought that IT as a cost
center and seeing it as a benefit center, on backing of the improvement and method
of the organization. Internet of things can be defined as a propelled standard which
make straightforward gadget a shrewdly gadgets which allow transferring simple
device an intelligent devices capable of exchanging data over a network without
human interaction. It represent a technology that interest research and industry, also
serve on transform the way to think, and work. This paper present an overview on
Internet of things (IoT), its application areas and its challenges in the aim to give a
guideline to future researchers.

Keywords IoT · Internet of Things · Application · Challenges

1 Introduction

Internet of Things is a new revolution of the Internet. Objects make themselves

recognizable and they obtain intelligence by making or enabling context related
decisions thanks to the fact that they can communicate information about themselves.
They can access information that has been aggregated by other things, or they can
be components of complex services. This transformation is concomitant with the
emergence of cloud computing capabilities and the transition of the Internet towards
IPv6 with an almost unlimited addressing capacity [1].

F. Z. Fagroud (B) · E. H. B. Lahmar · S. El Filali

Laboratory of Information Technology and Modeling, Faculty of Sciences Ben M’sick, Hassan II
University of Casablanca, BP 7955 Sidi Othman Casablanca, Morocco
H. Toumi
Higher School of Technology - Sidi Bennour, Chouaïb Doukkali University, El Jadida, Morocco
e-mail: [email protected]
L. Ajallouda · A. Zellou
ENSIAS Mohammed V University in Rabat, Rabat, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 335
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_31
336 F. Z. Fagroud et al.

The IoT infrastructure is in view of 3 fundamental components: park of connected

objects (fixed or mobile, geographically distributed), network (can be wired or wire-
less short or long-range, mobile or low-speed) and application (Most often developed
in a web technologies, which collects data from the object network to provide aggre-
gated and reprocessed information). The majority of the data gathered from the
connected objects permits making a decision, permanent follow, failure prediction.
Three significant improvements have permitted IoT to appear: Internet for its
communication capacity, increment of storage capacity limit that permit expanded
the amount of recorded data, lower access costs. According to study of GSMA the
number of IoT connections in world attained 25,2 billion connections and the global
market of IoT rise to 1100 billion dollars excluding hardware (devices, modules and
chips) by 2025.
Internet of things represent one of the emerging and evolving domains that interest
research and industry. By studying this topic several questions arise, among these
questions: In which area can be applied it? What are the IoT challenges? What is the
number published in this field?
The rest of this paper is structured as following: in Sect. 2 a presentation of
Internet of things. After we present its application areas. In Sect. 4 we discuss some
IoT challenges, afterward we give an overview on research publications in this topic.
At the latest a conclusion in Sect. 6.

2 Internet of Things

Internet of things have many definitions, until now there is no standardized definition.
IoT mean system that groups IT devices, mechanic and digital equipments, objects,
animals or people which are interconnected and can communicate data via a network
in a self-sufficient manner, in real time and without any human interaction. It repre-
sent a dynamic network infrastructure which include physical and virtual “things”
which possess unique identifiers, physical attributes, and virtual characteristics into
the information network. Auto-configuration, self-organization, transmit informa-
tion and data, share resources, react and act face to divers situations and changes in
the environment represent the IoT capabilities.
According to the International Telecommunication Union “The Internet of Things
is a global infrastructure for the information society, which provides advanced
services by interconnecting objects (physical or virtual) through interoperable infor-
mation and communication technologies existing or in evolution”. The concept of
“objects” is intended to pose no limit as to the feasible possibilities of implementing
technological tools in everyday elements.
The Cluster of European Internet of Things Research Projects (CERP-IoT) defines
the Internet of Things as: “a dynamic infrastructure of a global network. This global
network has auto-configuration capabilities based on standards and interoperable
A Brief Survey on Internet of Things (IoT) 337

communication protocols. In this network, physical and virtual objects have identi-
ties, physical attributes, virtual personalities, and intelligent interfaces, and they are
seamlessly integrated into the network.” [CLU 2010].
The development of a system where objects are identifiable in a unique
way, communicating with each other and capable of collecting data without any
human intervention is only possible through integrating different technologies.
Various technologies are deployed for the implementation of IoT including: RFID
(Radio Frequency Identification), Sigfox, WSN (Wireless Sensor Networks), Cloud
Computing, NFC (Near Field Communication), and Artificial intelligence.
IoT work in the following way:
• Collect of data: collect of data from the physical world by using sensors
• Communication: transfer of data collected through different types of networks
• Storage and analysis of data: based on many technologies for example Cloud
Computing to storage data collected
• Visualization: Present information to users in a comprehensible way.
IoT is characterized by:
• Heterogeneity: diversity of platforms, hardware and networks used for creating
of internet of things devices. These devices can interact with other devices or
service platforms by various communications technologies.
• Connectivity: allows accessibility and network compatibility. Accessibility is
over a network while compatibility furnish the common ability to consume and
generate data.
• Intelligence: allows objects to respond intelligently to a well-specified situation
and help them to carry out precise tasks
• Dynamic change: about number and state of device
• Interconnectivity: everything will be able to be interconnected with the global
information and communication infrastructure
• Huge scale: the number of devices to manage and communicate between them
will be high compared to devices connected to the Internet
• Detection: the information detected (information of the physical world) allows
the understanding of our complex world
• Security: must be guaranteed to protect the privacy of users.

3 Application Area

IoT is present in different area like health, energy and agriculture which affirm field
of the IoT application is almost unlimited.
Several IoT applications have been developed in different areas since its appear-
ance, for example: Ji-chun Zhao et al. [2] offer a remote greenhouse monitoring
system, which allows automatic control over the ambient temperature, moisture
factors. Tejaswinee A. Shinde and Dr. Jayashree R. Prasad have developed a system
based on IoT, which allows monitoring of animal health. The user can supervise
338 F. Z. Fagroud et al.

Fig. 1 IoT application areas

the health of the animal from a distance and be alerted by his mobile phone in the
event of any changes in the health status of the animal [3]. Tuan Nguyen Gia et al.
[4] present a system that permit remote control of glucose in real-time and contin-
uously. The presented system help doctors and caregivers to check their patients
anytime, anywhere per a browser or a mobile application. Security is one of the main
concerns, as all health data is considered private personal data. To mitigate the risks
of these applications, strong security mechanisms are needed. Jorge Gomez et al. [5]
have proposed a system which enable students to interact with physical objects in
their environments associated virtually with a learning subject. Khaleel Mershad and
Pilar Wakim [6] have developed a framework for a future LMS (Learning Manage-
ment System: tool for creation, distribution, monitoring and management of various
types of educational and training material) enhanced by IoT capabilities, several
LMS elements will be affected by IoT. Academics, researchers and students are in a
same position to carry on the exploration and advancement of IoT systems, devices,
applications, and services. The deployment of these solutions in our environment
will transform our environment into intelligent environment and conducive to entire
human activity.
A Brief Survey on Internet of Things (IoT) 339

4 IoT Challenges

4.1 Security

Security is an important aspect of IoT and must be guaranteed in the various layers
of the IoT architecture. They are several challenges related to security in internet
of things for example control access, data exchange security, protect the privacy.
Traditional security mechanisms cannot be applied in internet of things which make
necessary the creation of a news paradigms. To solve this issue we can use Machine
Learning, Blockchain, Gateway, Software Defined Networking, symmetric cryptog-
raphy, biometric modalities, wireless device radio fingerprinting to boost the security
and privacy of IoT systems.
Several research are proposed like: FLAUZAC Olivier et al. [7] propose a new
security architecture for IoT (Internet of Things) based on SDN (Software defined
Network), the latter allows to achieve network security in a more efficient and flexible
way. Hui Suo et al. [8] present an analysis of the characteristics and security require-
ments from the IoT architecture based on four layers (perceptual layer, network layer,
support layer and application layer). An Autonomic Agent Model Trust is a model
that serves to reduce security issues, raise reliability and credibility, and afford infor-
mation in IoT environments. It consists of installing Trustworthy Agent Execution
Chip (TAEC), which adduce a profitable software and hardware platform for secured
operation of the agent, on each sensor node, providing a hardware execution envi-
ronment for agents which is secured and autonomous [9]. The work of Markowsky
et al. [10] demonstrate that the search for vulnerable devices in the IoT is possible by
using a different methods: Shodan (find vulnerable Cayman DSL routers), Masscan
(quickly search a large address space for devices vulnerable to the Heartbleed Bug)
and Nmap with PFT (find and connect to vulnerable networked printers).

4.2 IoT Architecture

With the high expansion of internet of things it become a necessity to determine

a standard architecture of the generic IoT or by domain, several architectures have
been proposed since the appearance of this domain but the proposed architectures
is generally adapted to the need for a specific application. Among the proposed
architectures we present: IoT-Academia Architecture is an IoT architectural model
that covers almost every vital division of the academic forum [11]. Ivan Ganchev
et al. [12] proposed a high-level generic IoT architecture particularly appropriate to
the foundation of smart cities. Weigong LV et al. [13] defined a unified IoT system
architecture to determine a standard for building and implementing the IoT system
and diminish the hardness of building an IoT system. Ioan Ungurean et al. present
an architecture dedicated to IoT that is based on OPC.NET specifications able to
340 F. Z. Fagroud et al.

be deployed in industrial environments and in intelligent construction [14]. Michele

Nitti et al. developed an architecture for sustainable tourism application in a Smart
City scenario. It is based on a specific whole of exigencies and may be implemented
using a cloud-based IoT platform [15].

4.3 Interoperability in IoT

Most of IoT devices proposed until now are note interoperable which increase data
consolidation costs. There are 4 types of interoperability:
• Technical Interoperability: related to hardware/software components, systems
and platforms
• Syntactic Interoperability: structures and formats of data communicated
• Semantic Interoperability: same content are comprehended by different means
• Organizational Interoperability: organizations may communicate and transfer
data even they use a diversity of information systems on much different
infrastructures, perhaps at different geographic regions and cultures.
To solve the issue of interoperability in IoT many techniques and methodologies
can be used like: Ontology, RDF, SPARQL, IoT HUB. For example: Agarwal et al.
[16] propose a holistic and a light-weight ontology that aims to attain semantic inter-
operability between divers fragmented testbeds (that store data in their proprietary
format) in IoT domain.

4.4 Data Storage and Analysis

IoT devices yield a broad deal of data in real-time and unstructured, which requires
a decentralized structure to store and analyze this huge data. Integration of Cloud
Computing and IoT represent a valid approach that will serve to solve this issue.
Different IoT cloud platforms have been developed to solve various service areas such
as application development, administration of devices and systems, heterogeneity
management, data management, analysis tools and deployment [17]. Cloud4IoT is
a platform for self-acting deployment, arrangement and dynamic configuration of
software components for IoT support and data processing and analysis applications,
enabling plug-and-play integration of novel sensors and dynamic scalability [18].
CloudThings Architecture is a cloud-based IoT platform that supports IaaS, PaaS,
and SaaS to accelerate the application development and administration of the Internet
of Things [19]. This integration faces a set of challenges: Resource assignment,
Protocol Support, Identity Management, Energy effectiveness, Quality of Service
attainment, Data Storage Location, Security and Privacy.
A Brief Survey on Internet of Things (IoT) 341

5 Research Evolution

The opportunities of internet of things are diverse and unlimited which make then
an important research topic that occupies latterly a great importance by researchers.
The amount of research works interested in the IoT thematic increases from year to
year, to affirm this we present an overview on the number of publications included
in two scientific databases (IEEE and dblp). We carried out this study based on two
scenarios:
• First scenario: search for the number of publications including the chosen
keyword
– Keywords used: Internet of Things, IoT, Internet of Things challenge, IoT
challenge, internet of things application, IoT application
– Results:
• Figure 2: show the number of journal papers published in dblp and IEEE
and respond to our request
• Figure 3: show the number of conference and workshop papers published
in dblp and IEEE and respond to our request.

• Second scenario: search for the number of publications interested to a specific

challenge or application area and published in 2019
– Keywords used: Internet of Things architecture, IoT architecture, Internet of
Things security, IoT security, internet of things in education, IoT in education,
internet of things in healthcare, IoT in healthcare
– Results:
• Figure 4: show the number of works published in journals, conference and
workshop proceedings especially in IEEE database, in 2019 and respond to
our request

Fig. 2 Number of publication in IEEE and dblp databases: Journal Papers

342 F. Z. Fagroud et al.

Fig. 3 Number of publication in IEEE and dblp databases: Conference and Workshop Papers

Fig. 4 Number of publication in IEEE database published in 2019: Journal, Conference and
Workshop Papers

Fig. 5 Number of publication in dblp database published in 2019: Journal, Conference and
Workshop Papers
A Brief Survey on Internet of Things (IoT) 343

• Figure 5: show the number of conference, workshop and journal papers

published in dblp in 2019 and respond to our request.

6 Conclusion

Internet of Things present one of the trends in the field of new technologies and
innovation, which have a big behoove in research. It has several advantages for its
users (companies and civil users), despite these benefits many challenges are present
which leads to think about developing of a new solutions to solve these. The novelty
that IoT present is the use of internet and their protocols for daily objects. In this
article we have presented a brief study of Internet of Things, its fields of applications
and issues and a brief view on the number of research works, which will prove that
IoT is an area of great importance and multidisciplinary. In the future work we wish
to develop some internet of things applications/solutions and choice an issue of this
field in the aim to propose a new solution.

References

1. Madakam S, Ramaswamy R, Tripathi S (2015) Internet of Things (IoT): a literature review. J

Comput Commun 3(05):164
2. Zhao J, Zhang J, Feng Y, et al (2010) The study and application of the IOT technology in agri-
culture. In: 2010 3rd international conference on computer science and information technology.
IEEE, pp 462–465
3. Shinde TA, Prasad JR (2017) IoT based animal health monitoring and crop monitoring system
4. Gia TN, Ali M, Dhaou IB et al (2017) IoT-based continuous glucose monitoring system: a
feasibility study. Procedia Comput Sci 109:327–334
5. Gómez J, Huete JF, Hoyos O et al (2013) Interaction system based on internet of things as
support for education. Procedia Comput Sci 21:132–139
6. Mershad K, Wakim P (2018) A learning management system enhanced with Internet of Things
applications. J Educ Learn 7(3):23–40
7. Olivier F, Carlos G, Florent N (2015) New security architecture for IoT network. Procedia
Comput Sci 52:1028–1033
8. Suo H, Wan J, Zou C, et al (2012) Security in the Internet of Things: a review. In: 2012
international conference on computer science and electronics engineering. IEEE, pp 648–651
9. Xu X, Bessis N, Cao J (2013) An autonomic agent trust model for IoT systems. Procedia
Comput Sci 21:107–113
10. Markowsky L, Markowsky G (2015) Scanning for vulnerable devices in the Internet of Things.
In: 2015 IEEE 8th international conference on intelligent data acquisition and advanced
computing systems: technology and applications (IDAACS). IEEE, pp 463–467
11. Elyamany HF, Alkhairi AH (2015) IoT-academia architecture: a profound approach. In: 2015
IEEE/ACIS 16th international conference on software engineering, artificial intelligence,
networking and parallel/distributed computing (SNPD). IEEE, pp 1–5
12. Ganchev I, Ji Z, O’droma M (2014) A generic IoT architecture for smart cities
13. Lv W, Meng F, Zhang C, et al (2017) A general architecture of IoT system. In: 2017 IEEE inter-
national conference on computational science and engineering (CSE) and IEEE international
conference on embedded and ubiquitous computing (EUC). IEEE, pp 659–664
344 F. Z. Fagroud et al.

14. Ungurean I, Gaitan N, Gaitan VG (2014) An IoT architecture for things from industrial envi-
ronment. In: 2014 10th international conference on communications (COMM). IEEE, pp
1–4
15. Nitti M, Pilloni V, Giusto D et al (2017) IoT Architecture for a sustainable tourism application
in a smart city environment. Mob Inf Syst 2017:1–9
16. Agarwal R, Fernandez DG, Elsaleh T, et al (2016) Unified IoT ontology to enable interop-
erability and federation of testbeds. In: 2016 IEEE 3rd world forum on Internet of Things
(WF-IoT). IEEE, pp 70–75
17. Ray PP (2016) A survey of IoT cloud platforms. Future Comput Inform J 1(1–2):35–46
18. Pizzolli D, Cossu G, Santoro D, et al (2016) Cloud4iot: a heterogeneous, distributed and auto-
nomic cloud platform for the IoT. In: 2016 IEEE international conference on cloud computing
technology and science (CloudCom). IEEE, pp 476–479
19. Zhou J, Leppanen T, Harjula E, et al (2013) Cloudthings: a common architecture for inte-
grating the Internet of Things with cloud computing. In: Proceedings of the 2013 IEEE 17th
international conference on computer supported cooperative work in design (CSCWD). IEEE,
pp 651–657
20. Stergiou C, Psannis KE, Kim B et al (2018) Secure integration of IoT and cloud computing.
Future Gener Comput Syst 78:964–975
21. Fagroud F, Ben Lahmar E, El Filali S (2019) Internet Of Things: statistical study on research
evolution. Int J Adv Electron Comput Sci 6(5)
22. Arnaert M, Bertrand Y, Boudaoud K (2016) Modeling vulnerable Internet of Things on
SHODAN and CENSYS: an ontology for cyber security. In: Proceedings of the tenth inter-
national conference on emerging security information, systems and technologies (SECURE-
WARE 2016), Nice, France, pp 24–28
23. Aldowah H, Rehman SU, Ghazal S, et al (2017) Internet of Things in higher education: a study
on future learning. In: Journal of physics: conference series, p 012017
24. Xiao L, Wan X, Lu X et al (2018) IoT security techniques based on machine learning: how do
IoT devices use AI to enhance security? IEEE Signal Process Mag 35(5):41–49
25. Yasuda S, Miyazaki S (2017) Fatigue crack detection system based on IoT and statistical
analysis. Procedia CIRP 61:785–789
26. Prajapati H, Mishra AK (2018) A review on baggage tracing and handling system using sensor
networks and IoT
27. Fraga-Lamas P, Fernández-Caramés TM, Suárez-Albela M, Castedo L, González-López M
(2016) A review on Internet of Things for defense and public safety. Sensors 16(10):1644
28. Suresh P, Daniel JV, Parthasarathy V, et al (2014) A state of the art review on the Internet of
Things (IoT) history, technology and fields of deployment. In: 2014 international conference
on science engineering and management research (ICSEMR). IEEE, pp 1–8
29. Alansari Z, Soomro S, Belgaum MR, et al (2018) The rise of Internet of Things (IoT) in big
healthcare data: review and open research issues. In: Progress in advanced computing and
intelligent engineering. Springer, Singapore, pp 675–685
30. Anggorojati B, Prasad R (2018) Securing communication in the IoT-based health care systems.
Jurnal Ilmu Komputer dan Informasi 11(1):1–9
31. Banerjee M, Lee J, Choo KKR (2018) A blockchain future for Internet of Things security: a
position paper. Digit Commun Netw 4(3):149–160
32. Burhanuddin MA, Mohammed AAJ, Ismail R, Basiron H (2017) Internet of Things architecture:
current challenges and future direction of research. Int J Appl Eng Res 12(21):11055–11061
33. Zhou C, Damiano N, Whisner B, Reyes M (2017) Industrial Internet of Things: (IIoT)
applications in underground coal mines. Min Eng 69(12):50
34. Asim M, Iqbal W (2016) Iot operating systems and security challenges. Int J Comput Sci Inf
Secur 14(7):314
Design and Implementation of Smart
Irrigation System Based on the IoT
Architecture

Asmae Hafian, Mohammed Benbrahim, and Mohammed Nabil Kabbaj

Abstract Water is an indispensable resource in agriculture, and with the likely

increase in restrictions on the volumes of water allocated to agriculture, irrigation is
required to meet the needs of plants. In this sense, this article proposes the design
and implementation of an intelligent irrigation system based on the IoT architecture.
These systems are designed especially for urban areas where remote IoT devices don’t
have direct access to the internet or power grids and where wireless communications
are difficult.
The proposed system is composed of different technologies to exploit their advan-
tages to increase both the efficiency of water use and the efficiency of agricultural
yields.

Keywords Smart irrigation · IoT · ZigBee · LoRa · Wi-Fi · WSN

1 Introduction

Irrigation is the operation of artificially bringing water to cultivated plants [1]. It is

important to know that agriculture is by far the largest consumer of water whose
irrigation is a necessary practice in several agricultural cropping systems in arid and
semi-arid areas [13]. The development of a field distributed system based on irri-
gation sensors offers the potential to increase yield and quality while saving water
[2], in which wireless sensor network technology can provide significant support
that will enable accurate resource management due to their cost-effective nature

A. Hafian (B) · M. Benbrahim · M. N. Kabbaj

Engineering, Modeling and Analysis of Systems Laboratory, Sidi Mohamed Ben Abdellah
University, Fez, Morocco
e-mail: [email protected]
M. Benbrahim
e-mail: [email protected]
M. N. Kabbaj
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 345
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_32
346 A. Hafian et al.

and deployment flexibility. Furthermore, the wireless sensor network possesses low
cost and highly reliable sensors, and have a transmission radius, the latter measure
autonomously the humidity, temperature parameters and communicate via wireless
technology with a central data collection station. This allows the farmer to use this
technology that provides the opportunity for producers to access remotely via wire-
less radio communication to see the field conditions and the irrigation operation, at
home or the office.
The search for an innovative and adequate material architecture (cost and perfor-
mance) for the achievement of the objectives of our work led us to do a biblio-
graphic search at the level of existing systems and related fields. For example, [3]
which presented a network of wireless sensors controlled by the broker. The latter
is in constant communication with an online server that uses the Message Queuing
Telemetry Transport (MQTT) protocol via a Wi-Fi connection that communicates
with the broker using a peer-to-peer LoRa connection. Another interesting work is
proposed by [4], which presented the fundamental design and implementation of a
wireless sensor network featuring a Zigbee protocol. This work describes the use
of a high transmission power with a relatively low energy consumption based on
a Zigbee wireless sensor network for the monitoring system of the water irrigation
control.
Also, [5] represents the design and implementation of an intelligent irrigation
system based on LoRa technology. This article presents the energy consumption
results, which take into account the different operating modes of the terminal nodes.
In this sense, to ensure adequate access to digital agriculture, we have proposed an
intelligent irrigation system with a diverse architecture using ZigBee, LoRa, and Wi-
Fi protocols based on sensor networks and wireless connectivity to all levels of the
system. The use of these different protocols in our proposed architecture allowed us
to take advantage of the different types of communication such as WLAN (wireless
local area network), WPAN (wireless personal area network), LPWAN (Low Power
Wide Area Network), which offers the possibility of remote access and at any time
to our irrigation field.
The proposed architecture far exceeds existing architectures in terms of low cost
and sophisticated performance. Indeed, it is an architecture making it possible to
benefit, on the one hand, from IoT and Cloud Computing and, on the other hand,
from hardware or software evolutions at the level of each part of the system.

2 Design of the System

2.1 System Architecture

Figure 1 represents the proposed system architecture which consists of five parts:
IoT nodes, local gateways, main gateway, Cloud server, and application.
Design and Implementation of Smart Irrigation System … 347

IoT Nodes are responsible for collecting data from the sensor layer components.
They are responsible for exchanging information with the local gateway, and they
receive remote irrigation commands from the gateways or the cloud. IoT Nodes send
the information obtained via the ZigBee protocol. This information from the various
sensors will be processed locally by a local gateway.
Subsequently, the main gateway performs the essential functions of collecting
data from local gateways via the LoRa protocol. Also, third-party services such as
weather forecasting can be used when deciding on irrigation programs, as well as
programming, monitoring, and diagnosis of the main gateway [6].
Using Message Queuing Telemetry Transport (MQTT), the main gateway controls
the network, which is in constant communication with the Cloud via a Wi-Fi connec-
tion. This communication is responsible for the exchange of messages between the
Cloud and the multiple local gateways of the network, which is, in turn, communi-
cate with the broker via the Lora connection. The cloud server is mainly responsible
for storing data [15] and implementing the MQTT interface for the LoRa server and
applications.
For various applications can be offered: mobile applications that developed on
both Android or iOS platforms or web applications. These applications allow users
to get the status of IoT nodes as they can also control and supervise the irrigation
system by sending commands.

2.2 Wireless Communication Protocols

Devices in the IoT network use different communication protocols with each other.
The wireless communication protocol is a set of rules that enable different electronic
devices to communicate with each other wirelessly.
The communication standard chosen to have an innovative and adequate hardware
architecture in terms of cost and performance for our proposed system are three
protocols: ZigBee, LoRa, Wi-Fi.
This connectivity meets the following constraints:
• Wi-Fi (Wireless Fidelity): Local connectivity without taking into account energy
consumption (WLAN). It is the common popular IoT communication protocol for
wireless local area networks (WLAN), which uses the IEEE 802.11 standard [14]
via 2.4 GHz UHF and 5 GHz ISM frequencies. Wi-Fi provides Internet access
to devices located approximately 20 to 40 m from the source. It has a data rate
of 600 Mbps maximum, depending on the frequency of the channel used and the
number of antennas,
• ZigBee: Local connectivity with consideration of energy consumption. ZigBee
is a wireless standard based on IEEE 802.15.4 [7]. It is a technology for robust
Wireless Personal Area Network (WPAN) communication. This technology is
now used in many equipment, very generally embedded, which imposes a very
low consumption, and are enough of a very low data rate and it ranges up to
348 A. Hafian et al.

Fig. 1 Architecture of proposed system

Design and Implementation of Smart Irrigation System … 349

Table 1 Characteristics of
Feature ZigBee LoRa Wi-Fi
different wireless
communication protocols Data rate [kbps] 250 110 11 * 102
Range [m] 10–100 2000 1–100
Frequency [GHz] 2.4 0.868 2.4
Power consumption [mA] 1–10 1–10 100–350

100 m [8]. The Zigbee protocol has a large number of applications in machine
to machine & IoT technologies. It offers a range of approximately 10 to 100 m
maximum, and its data transfer rate is approximately 250 Kbps,
• LoRa: Remote connectivity with low energy consumption (LPWAN). LoRa wire-
less technology is a long-range, low-power platform that uses unlicensed radio
spectrum [3] widely used in IoT applications. Low energy consumption, it offers
LPWAN connectivity between wireless sensors and the Cloud [9]. To ensure reli-
able monitoring, LoRa-based sensors can be installed in smaller devices. It aims to
eliminate repeaters, reduce the cost of devices, increase the battery life of devices,
improve network capacity and support a large number of devices [3].
Table 1 presents the characteristics of different wireless communication protocols
used in our proposed architecture [3].

3 System Implementation

To design an intelligent irrigation system based on our IoT architecture proposed in

Fig. 1, this section presents the implementation of the different components of the
system.
The different sensors used:
• YL-69, a soil moisture sensor, a simple rupture, consisting of two buffers that
function as probes. It uses the conductivity between these two buffers to measure
moisture in soil and similar materials.
• DHT22, a basic digital sensor, at low cost making it possible to effectively measure
the temperature and humidity of the ambient air thanks to its two-in-one combi-
nation of a capacitive humidity sensor and a thermistor. The temperature reading
range is: −40 to +80 °C and for humidity is: 0 to 100%, with an accuracy of
0.5 °C and 2–5% for temperature and humidity respectively [3];
This data is collected using a microcontroller, which is represented by an Arduino
ADK. This latter is a microcontroller board based on the ATmega2560. It has a USB
host interface to connect with Android-based phones, 54 digital input/output pins (of
which 15 can be used as PWM outputs), 16 analog inputs, 4 UARTs (hardware serial
ports), 16 MHz crystal oscillator, USB connection, a power jack, ICSP header, and
a reset button. It contains everything needed to support the microcontroller; it can
350 A. Hafian et al.

be powered via the USB connection or with an external power supply. The power
source is selected automatically [10].
The expected output is then used to send the control signal via serial commu-
nication to Arduino to control the water pump to water the field accordingly. This
action can be commanded via the local gateway or remotely via applications. For the
wireless communication between IoT nodes and the local gateway, we used two of
XBee S2C Pro, which is using a 2.4 GHz wireless transceiver to communicate with
another XBee module.
Each local gateway is composed of two main elements: ESP32 (dual-core 32-
bit MCU with built-in Wi-Fi and BLE), with LoRa node chip SX1278; The LoRa
radio transceiver module capable of creating a multipoint network, with addresses
of individual nodes and encryption, within a range of about 2000 m [3].
About the main gateway we chose Raspberry Pi; it is an open-source, Linux-
based, credit card-sized computer board. It has USB ports to which you can connect
a keyboard and mouse, and a High-Definition Multimedia Interface (HDMI) video
output to which you can connect a TV or monitor. The Raspberry Pi is based on a
Broadcom BCM2835 chip. It does not feature a built-in hard disk or solid-state drive,
instead relying on an SD card for booting and long-term storage [11].
The main gateway uses the Wi-Fi connection to collect data via the LoRa protocol.
This layer also ensures that the collected signals are filtered and are ready to perform
the physical unit conversion.
For the cloud, we linked the raspberry to an MQTT broker, which will allow us
to control the irrigation system remotely. MQTT is a common protocol used in IoT
systems to connect low-level devices and sensors [12]. MQTT is used to pass short
messages to and from a broker.

4 Results and Discussion

The test of the different configurations was done through a miniaturized prototype
similar in terms of components to an agricultural gateway of real dimensions (Fig. 2).
We chose a site with an irrigation system to test our proposed solution. The site
is located at the Dhar ElMahraz Faculty of Sciences in Fez, Morocco (Fig. 3). The
miniaturized prototype comprises a network of 2 sensor nodes, plus the main gateway.
Every 5 min, the sensor data is acquired in the local gateway via the XBee protocol
(Fig. 4). This data will then be sent to the main gateway via the LoRa protocol.
Depending on the water needs of the crop and the climatic conditions, the main
gateway decides to act on the local gateway to operate the pump.
The loss of data is low in the case of the mobility of the nodes. The number of
retransmissions is similarly low.
For remote monitoring (application), we used the LabVIEW software to control
and supervise our system remotely, with a simplified operator interface dialoguing
via TCP/IP (Transmission Control Protocol/Internet Protocol) serial digital commu-
nication with the broker. TCP/IP is used for the transfer of data on the Internet, it
Design and Implementation of Smart Irrigation System … 351

Fig. 2 Miniaturized prototype of the proposed system

Fig. 3 Satellite image of the test site

represents in a certain way the set of communication rules on the Internet and is based
on the notion of IP addressing, that is to say, the fact of providing an IP address to
each machine of the network to be able to convey data packets.
As Fig. 5 shows, the monitoring system was used with a LabVIEW interface; it
represents the curves for soil moisture and temperature monitoring as well as the
unit of the pump.
352 A. Hafian et al.

Fig. 4 Sensor data acquired at the local gateway

Fig. 5 LabVIEW interface in the remote computer

Design and Implementation of Smart Irrigation System … 353

5 Conclusion and Future Work

In this article, we presented the design and implementation of a prototype of a new

generic wireless sensor network platform for an intelligent irrigation system. The
proposed design represented the combination of several protocols to ensure adequate
access to digital farming. The use of these protocols improves the reliability and
efficiency of the installation of soil humidity and temperature sensors and allowed us
to remotely control the parameter of the soil water content in real-time. Server-side
monitoring can present data logs from each sensor; Also, the power consumption of
wireless network devices is lower, and the system performs a long-term function.
To further test the system, we will, in the future, implement it in a real garden with
several irrigation areas. Also, an application for agriculture, whose objective is to
calculate not only when to bring the water, but also the amount of water that the
plants need, and automatically adjusts the operating time according to the stage of
growth and seasonal fluctuations.

References

1. Belkacem I (2014) L’irrigation intelligente. In: 15th annual global development conference,
Ghana
2. Kim Y, William M (2008) Remote sensing and control of an irrigation system using a distributed
wireless sensor network. IEEE Trans Instrum Meas 57(7):1379–1387
3. Gloria A, Dionisio C, Simoes G, Sebastiao P, Souto N (2019) WSN application for sustainable
water management in irrigation systems. In: 2019 IEEE 5th world forum on Internet of Things
(WF-IoT), Limerick, Ireland, pp 833–836
4. Rasin Z, Hamzah H, Mohd Aras MS (2009) Application and evaluation of high power
Zigbee based wireless sensor network in water irrigation control monitoring system. In: IEEE
symposium on industrial electronics applications, ISIEA 2009, vol 2
5. Zhao W, Lin S, Han J, Xu R, Hou L (2017) Design and implementation of smart irrigation
system based on LoRa. In: 2017 IEEE Globecom workshops (GC Wkshps), Singapore, pp 1–6
6. Lamas PF, Echarri MC, Azpilicueta L, Iturri PL, Falcone F, Caramés T (2020) Design and
empirical validation of a LoRaWAN IoT smart irrigation system. In: Proceedings, vol 42, p 62
7. IEEE 802.15.4TM (2011) IEEE standard for local and metropolitan area networks. Part 15.4:
low-rate wireless personnal area networks (LR-WPANs)
8. Jackson F, Ferial V, Jiamin P, Yan Xin P, Thierry V (2013) ZigBee, de la théorie à la pratique :
création d’un réseau ZigBee avec transmission de données. 3EI 71:1–18
9. Ayaz M, Ammad-Uddin M, Sharif Z, Mansour A, Aggoune EM (2019) Internet-of-Things
(IoT)-based smart agriculture: toward making the fields talk. IEEE Access 7:129551–129583
10. Arduino store. https://store.arduino.cc/
11. Kumar Sahu C, Behera P (2015) A low-cost smart irrigation control system. In: IEEE sponsored
2nd international conference on electronics and communication system
12. Kodali RK, Kuthada MS, Yogi Borra YK (2018) LoRa based smart irrigation system. In:
4th international conference on computing communication and automation (ICCCA), Greater
Noida, India, pp 1–5
13. Hafian A, Benbrahim M, Kabbaj MN, Bouazi A, Abouabdillah A (2020) A new approach for
optimal sizing of photovoltaic pumping systems for irrigation. In: 5th international conference
on renewable energies for developing countries, Marrakech, Morocco, pp 1–6
354 A. Hafian et al.

14. Chong C, Kumar S (2003) Sensor networks: evolution, opportunities and challenges. Proc IEEE
91(8):1247–1256
15. Roopaei M, Rad P, Choo K (2017) Cloud of things in smart agriculture: intelligent irrigation
monitoring by thermal imaging. IEEE Cloud Comput 4:10–15
Lightweight Key Exchange Algorithms
for MQTT Protocol in Its Environment

Anwar D. Alhejaili and Omar H. Alhazmi

Abstract One of the most important challenges in the use of the Message Queue
Telemetry Transport (MQTT) protocol in the Internet of Thing (IoT) environment is
data confidentiality and user authentication. Moreover, it is easy to manipulate the
default security provided by MQTT during user authentication through username
and password, and it does not provide any type of encryption of the data to ensure
confidentiality or integrity. In this paper, we propose a key exchange algorithm to
provide a secret key that can be used for data encryption using permuted congruential
generator PCG-Rand as well as to ensure data integrity using Spongent lightweight
hash functions.

Keywords IoT · MQTT · Key exchange algorithm · PCG-Rand · Spongent

light-weight hash functions

1 Introduction

In the last few centuries, the world has witnessed significant advancement in terms
of technology and communication. An individual’s daily life has become dependent
on the Internet of Thing (IoT), which includes most of the things or objects that are
connected to each other remotely. In IoT environments, everything becomes smarter
and is able to process data, manage and communicate with the other surrounding envi-
ronments. Moreover, it provides many services and has created a significant impact
in several fields, such as industry, smart homes, smart cities, healthcare, etc. All these
different fields collect various types of data and use a variety of techniques and proto-
cols. Currently, IoT security is one of the most popular topics in the field of research.
IoT security may include a number of different aspects, such as how to provide
confidentiality and maintain data integrity. Moreover, it provides authentication and

A. D. Alhejaili (B) · O. H. Alhazmi

Department of Computer Science, Taibah University, Madinah, Saudi Arabia
O. H. Alhazmi
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 355
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_33
356 A. D. Alhejaili and O. H. Alhazmi

authorization to access smart devices, which are increasingly becoming a part of our
everyday lives. However, there are different types of protocols that can be adapted
to the IoT environment, such as Advanced Message Queuing Protocol (AMQP),
Constrained Application Protocol (CoAP), Message Queue Telemetry Transport
(MQTT), and Data Distribution Service (DDS). Each of these protocols has its own
advantages and disadvantages, and the most commonly used protocol is the MQTT
protocol.
The Message Queue Telemetry Transport (MQTT) is one of the application layer
protocols that works on Transmission Control Protocol/Internet Protocol (TCP/IP),
and it is very stable for IoT devices with limitation. According to [1] MQTT is
extremely suitable for communication that contains various types of data-link layer
protocols such as Wi-Fi. MQTT is a lightweight protocol for IoT devices, and it is
known as a publisher/subscriber protocol. The MQTT protocol contains three main
layers: the Physical layer, the TCP/IP stacks, and the MQTT application layer [2].
Moreover, it contains the following components: the publisher, the subscribers, and
the MQTT Broker (or MQTT server). Every IoT device can be a publisher; it can
connect with the MQTT Broker and publish data on a specific topic. Every subscriber
in this topic will receive the message “publish by publisher”. Figure 1 illustrates a
simple scenario to explain the MQTT publisher/subscriber concept.
Indeed, an increase in the amount of data generated by IoT devices takes place
every day, and it then becomes big data. This data is prone to various types of
attacks, which aim to obtain user information, breach privacy, and lead to unwanted
communication from the nodes due to spamming [3] and lead to Distributed Denial
of Service (DDoS). According to Vaccari et al. in [4] the SlowITe attack, which
is a type of denial of service (DoS) attack, targets the MQTT protocol as well as
all protocols that can work over TCP. Furthermore, the MQTT protocol provides
the service of user authentication through username and password, and it does not
provide encryption to ensure data confidentiality. This paper focuses on the MQTT
protocol because it is the most popular one. The contribution of this paper is the
proposed lightweight algorithm that can provide a secret key between the publisher

Subscribers

Publisher
MQTT Broker
Subscribers

Fig. 1 Scenario that explains the MQTT publisher/subscriber concept

Lightweight Key Exchange Algorithms for MQTT Protocol in Its Environment 357

and subscriber over an insecure channel based on the MQTT protocol in order to
ensure and achieve data confidentiality and authentication.
The rest of this paper is organized as follows: Sect. 2 mentions some related work
on MQTT lightweight algorithms. Section 3 provides an overview of the exploited
technologies along with a discussion of the main concepts and technologies that we
used. Section 4 discusses the proposed key exchange algorithm. The last section
includes the conclusion and a discussion on future research.

2 Related Work

MQTT protocol is one of the most popular lightweight protocols that is used in IoT
devices. The main drawback of the MQTT protocol is that it does offer security func-
tions by default and it is based on SSL/TLS, which uses username and password for
authentication. Nowadays, there are various types of attacks that target the MQTT
protocols, such as the DoS and the SlowITe attack. This creates the need for robust
security algorithms to overcome most types of attacks. Indeed, there is a various
methods that can be used to make the MQTT protocol more secure, such as encryp-
tion/decryption (by symmetric or asymmetric techniques) or digital signatures. In
the following section, we have discussed certain resources and studies that proposed
a different way of fulfilling this objective.
Symmetric block cipher or secret key encryption is one of the techniques that can
be used to create lightweight algorithms. One of the most suitable encryption algo-
rithms for the publisher/subscriber concept is Attribute-Based Encryption (ABE); it
provides access control over the data and confidentiality. The authors in [5] proposed
that the hybrid security algorithm passes through two main phases: The first phase is
the setup phase where the key authority generates a public key and a master key by
running the Attribute Based Encryption (KP-ABE) algorithm, and it distributes the
public key among all users. The second phase is the encryption where the publisher
encrypts the message by using the Dynamic S-box AES algorithm, followed by
encryption by KP-ABE with both the S-box AES Dynamic and public key as input.
As a result, the proposed algorithm provides confidentiality and access control over
data. Additionally, the study by Iyer et al. [6] proposed the use of a symmetric cipher
with 64 bits and Spongnt hash function for achieving integrity and authenticity. The
authors tried different types of lightweight block cipher, such as Speck, Present, and
Simon; the best result was obtained by using the Speck block cipher. Another study
where Bashir and Hussain in [7] proposed an algorithm to prevent unauthorized users
from accessing the data. This algorithm is based on the following: generating ID for
each node, generating a set of random numbers as block seed, generate a random
integer number used as block selector, and generating a random integer for choosing
a block, and then using all the previous numbers with the ID for encryption through
the chosen random block. Additionally, the author in [8] proposed MQTTAuth as a
security solution for the MQTT protocol. It was based on the augmented password-
authenticated key exchange (AugPAKE) algorithm, which creates a secure session
358 A. D. Alhejaili and O. H. Alhazmi

key to provide authorization and ensure the confidentiality of the transmitted data.
On the other hand, in the case of asymmetric encryption, the Elliptic-curve cryptog-
raphy (ECC) algorithm is the most popular and efficient for IoT devices. Amnalou
et al. in [9] proposed a lightweight algorithm based on the use of Elliptic Curve Diffie
Hellman Ephemeral (ECDHE) and a pre-shared key.
The relevant previous work may offer methods that show better performance,
but they consume the power of the devices due to the computational cost or make
it difficult to adapt them to all devices. Therefore, we proposed a lightweight key
exchange algorithm that can be used with the different types of a cryptography
algorithm to make it suitable for IoT devices.

3 Overview of the Exploited Technologies

This section discusses the main technologies used in the proposed algorithms.

3.1 Key Exchange Algorithms

There are various types of key exchange algorithms, such as Diffie-Hellman, Elliptic
curve Diffie-Hellman (ECDH), and ElGamal encryption. All previous of algorithm
serve the same purpose, i.e., to exchange a secret key over an insecure channel.
However, the benefit of using an elliptic curve algorithm is that it can be used to mini-
mize the length of the key. The secret key can be used with various types of encryption.
According to some studies that make a comparison between Diffie-Hellman (DH),
Rivest–Shamir–Adleman (RSA) and ECDH [10], the ECDH algorithm showed the
best result.

3.2 Spongent Hash Function

Spongent is lightweight hash function. It has different sizes of input (such as 88, 128,
160, 224, and 256 bits). The sponge is composed of three stages. The first stage is
the initialization where the message is padded by 0 as the initial value and cut into
the r block. The second stage is the absorbing, and it implements the XOR operation
with the first r-bit and the permutation of b and so on to the last r block in the data.
The third stage is the squeezing, and it produces the output in bits based on the user’s
choice [11].
Lightweight Key Exchange Algorithms for MQTT Protocol in Its Environment 359

3.3 PCG-Rand Random Number Generator

The PCG stands for permuted congruential generator. It is an algorithm that is used to
generate random numbers in cryptographic algorithms. The PCG family for random
number generation is simple, easy to use, statistically efficient, and fast. Additionally,
it has a high rate of predictability of difficulty and a small code size. According to
O’Neill [12], the PCG-Rand algorithm is applicable to a range of 64 and 128-bit
streams that can provides 32-and 64-bit outputs.

4 The Proposed Key Exchange Algorithm

The following sections discuss the proposed generate/exchange secret key algo-
rithm. Our algorithm is related to the Elliptic-curve Diffie–Hellman (ECDH) concept
but with a reduced overhead of computation. It uses PCG Rand random number
and lightweight Spongent hash (ESK-PCG-Spongent) to ensure data integrity. Our
proposed algorithms are based on the assumption of the trusted system. The trusted
system can be a secure bootstrapping that generates a prime number (PN) for
publishers and subscriber, after making sure they are authorized to be enter according
to their ID. Moreover, the trusted system is trusted by every node in the environment.
However, to ensure security, the PN will be generated for each new session. The PN
is same for the publisher and the subscriber for every session.

4.1 Generate and Exchange Stage

In this stage and after the trusted system ensure publisher and subscriber are
authorized to access, will generate PN as bootstrapping.
1. Send the prime number that generated by trusted system to MQTT broker to be
distributed on publisher and subscriber.
2. The publisher and subscriber will have their own private key (PKp and PKs,
respectively). The private key is generated by using PCG Rand random number.
3. Calculate the shared key for the publisher SHp = (PKp * PN) and the key for
the subscriber, SHs = (PKs * PN).
4. Then, after exchanging the key, the secret key will be calculated, which will be
same value for both. The secret key for the publisher will be S = (SHs * PKp)
and that for subscriber S = (SHp * PKS).
All these symbols are summarized in Table 1. Figure 2 illustrates all these steps.
360 A. D. Alhejaili and O. H. Alhazmi

Table 1 Symbol used in the proposed algorithm and their meaning

Symbol Description
PN Prime number for publisher and subscriber
PKp PCG Rand random number as a publisher private key
PKs PCG Rand random number as subscriber private key
SHp The publisher shared key, which equal (PKp * PN)
SHs The subscriber shared key, which equal (PKs * PN)
S The secret key for publisher and subscriber

4.2 Encryption MQTT Payload Scenario Stage

Now, after generating and exchanging the secret key. The publisher and subscriber
connect with Broker through CONNECT and get CONNECTACK as response, and
subscriber subscribes to specific topic as illustrated in Fig. 2. We will use the secret
key to encrypt the MQTT payload. Then we will have the following scenario:
• Using the simple logical operation XOR to reduce the overhead and power
consumption. Also, it will use the Spongent hash to calculate the hash value
of the payload before encrypting for payload integrity. Figure 3. Illustrates XOR
encryption scenario.

4.3 Security Analysis

In this section will discuss some attack may effect on the proposed algorithms.
• Brute-force Attack: In this type of attack the attacker exploits the length of a key
and tries to find all possible values to find the key. We assume the secret key size
is equal to 128 bit. The number of possible values to get the secret key will be as
illustrates in Eq. (1).

Possible number of key = 2128 (1)

Possible number of key = 3.40 × 1038

In addition, the average of possible number of tries to get the correct key will be
less than or equal the possible number as illustrated in Eq. (2).

Average = 3.40 × 1038 /2 (2)

Lightweight Key Exchange Algorithms for MQTT Protocol in Its Environment 361

Trust System Publisher MQTT Broker Subscriber

ID
ID ACK
ID
ID ACK
Prime number: PN
PN PN

Shared key Shared key

SHp = PKp * PN SHs = PKs * PN

Publisher Subscriber
shared key shared key
Subscriber Publisher
shared key shared key

Secret key Secret key

S = SHs * PKp S = SHp * PKs

CONNE
CONNE
CONNECTAC
CONNECTAC

Subscriber

SUBACK

Fig. 2 The generate and exchange secret key phase

Fig. 3 The scenario

encryption by using XOR Publisher MQTT Broker

Hash value + [E (payload)] XOR

362 A. D. Alhejaili and O. H. Alhazmi

5 Discussion

This section will mention the improvements that have been added by our work into
the key exchange algorithm that can be used on MQTT protocol.
In our algorithm, the trusted system is the responsible for verifying the autho-
rization. The trusted system will perform the authorization using the publisher and
subscriber ID’s. Also, we reduce the overhead computation on the algorithm by using
one prime number for publisher and subscriber, which then sent out to MQTT broker
to distribute between publisher and subscriber. That means, we make the MQTT
broker is responsible for the process of distributing and exchanging keys, because
we might have more than MQTT broker on the system. Also, we use key length 128
bit, which takes a long time to be break. In addition, to ensure protection even in case
the prime number is discovered, it will be generated for every session. Furthermore,
we use a Spongent hash for the payload before encrypting it for integrity to be sent
with the payload encryption.

6 Conclusion

In the IoT environment, one of the most important factors is the security of the data.
There are various types of protocols that can be used in the IoT environment. The
MQTT protocol the most popular protocol for several reasons, and it follows the
publisher/subscriber concept. In this paper, we proposed an algorithm for creating a
secret key over an insecure channel. The proposed algorithm provides a secret key
by length 128 bit that can be used with various types of encryption algorithms. Addi-
tionally, we encrypt the payload by using XOR operation. For provide integrity, we
proposed the use of Spongent lightweight hash functions. Also, in security analysis
for the algorithm, we implemented brute-force attack on the key length and it take
many years to break it. Finally we mention to the most important improvements that
have been proposed on the algorithm.
In future work, we will implement the algorithm in the IoT environment to measure
its performance on the batteries of devices. Furthermore, we will test the performance
during different attack scenarios, such as MITM, spoofing attack, and SlowITe at-
tack. We will replace the XOR operation Scenario with other lightweight encryptions,
such as the HIGHT block cipher and the PRESENT block cipher to identify the best
result.

References

1. Mrabet H, Belguith S, Alhomoud A, Jemai A (2020) A survey of IoT security based on a

layered architecture of sensing and data analysis. Sensors 20(13):3625
Lightweight Key Exchange Algorithms for MQTT Protocol in Its Environment 363

2. Vaccari I, Maurizio A, Cambiaso E (2020) SlowITe, a novel denial of service attack affecting
MQTT. Sensors 20(10):2932
3. Ajmal M, Bag S, Tabassum S, Hao F (2017) Privy: privacy preserving collaboration across
multiple service providers to combat telecoms spam. IEEE Trans Emerg Top Comput 8:313–
327
4. Vaccari I, Aiello M, Cambiaso E (2020) SlowITe, a novel denial of service attack affecting
MQTT. Sensors 20(10):2932
5. Bisne L, Parmar M (April 2017) Composite secure MQTT for Internet of Things using ABE
and dynamic S-box AES. In: 2017 innovations in power and advanced computing technologies
(i-PACT), pp 1–5. IEEE
6. Iyer S, Bansod GV, Naidu P, Garg S (December 2018) Implementation and evaluation of
lightweight ciphers in MQTT environment. In: 2018 international conference on electrical,
electronics, communication, computer, and optimization techniques
7. Bashir A, Hussain Mir A (2018) Securing communication in MQTT enabled Internet of Things
with lightweight security protocol. EAI Endorsed Trans Internet Things 3(12):154390
8. Calabretta M, Pecori R, Vecchio M, Veltri L (2018) MQTT-Auth: a token-based solution
to endow MQTT with authentication and authorization capabilities. J Commun Softw Syst
14:320–331
9. Amnalou S, Bakar KAA (2019) Lightweight security mechanism over MQTT protocol for IoT
devices
10. Goyal TK, Sahula V (September 2016) Lightweight security algorithm for low power IoT
devices. In: 2016 international conference on advances in computing, communications and
informatics (ICACCI), pp 1725–1729. IEEE
11. Bogdanov A, Knežević M, Leander G, Toz D, Varıcı K, Verbauwhede I (September 2011)
SPONGENT: a lightweight hash function. In: International workshop on cryptographic
hardware and embedded systems. Springer, Heidelberg, pp 312–325
12. O’Neill ME (2014) PCG: a family of simple fast space-efficient statistically good algorithms
for random number generation. ACM Trans Math Softw
13. Hong D, Sung J, Hong S, Lim J, Lee S, Koo B-S, Lee C, Chang D, Lee J, Jeong K, Kim H, Kim J,
Chee S (October 2006) HIGHT: a new block cipher suitable for low-resource device. In: Inter-
national workshop on cryptographic hardware and embedded systems. Springer, Heidelberg,
pp 46–59
IoT Design and Realization
of a Supervision Device for Photovoltaic
Panels Using an Approach Based
on Radiofrequency Technology

Zaidan Didi and Ikram El Azami

Abstract In this paper, we present an innovative method that allows to measure,

calculate and transmit the parameters of a photovoltaic panel based on radiofrequency
communication. Our method focuses on the integration of a current sensor and a
voltage divider to determine the power as well as the energy, these quantities are
transferred in real time by a radio frequency-based approach. Note that our method
can operate in real time and the quantities received are displayed on the serial monitor
and on an LCD screen.
In comparison with other approaches which point to the same objectives, our
method is proven to work reliably even in the absence of internet because it is based
on the 433 MHz radio frequency for data transfer.
Our realization has been implemented and tested, the results show that our project
is successful in calculating, measuring and transmitting the characteristics of a
photovoltaic panel in real time.

Keywords Radiofrequency · Photovoltaic panel · Solar energy · Current sensor ·

IoT

1 Introduction

Photovoltaic energy is now experiencing vigorous development and is now becoming

a grid-connected energy source, hence the latest PSC power conversion efficiency
report (PCE, the ratio of the energy of incident solar photons to the electrical energy)
exceeded 25% [1, 2]. But like any industrial evolution, this system must be competent
to notice the anomalies of a photovoltaic installation and to process the measured
parameters and, this is called a monitoring system [3]. A photovoltaic system can
be subjected during its operation to various anomalies leading to a reduction in
performance of the system [4].

Z. Didi (B) · I. El Azami

Computer Science Research Laboratory (LaRI), Faculty of Sciences, Ibn Tofail University,
Kenitra, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 365
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_34
366 Z. Didi and I. El Azami

Many interesting studies point to the measurement of different parameters of

photovoltaic panels based on the Arduino hardware platform [5], other projects use
the ACS712 current sensor based on the hall effect [6].
In the same design, other achievements have integrated the Wi-Fi module to send
the characteristics of a photovoltaic panel [7].
In this paper we propose a design and construction project for a monitoring system
for the current, voltage and energy quantities of solar panels in real time. This IoT
project will be a menu of connectivity allowing the user to transfer the parameters
of photovoltaic solar panels by an RF radio frequency link.

2 Materials and Methods

The hardware design uses an Arduino Uno board and esp32 microcontroller, a voltage
divider to measure the voltage across the photovoltaic panel and an ACS712-30A
sensor to measure the current, a 433 MHz (FS1000A) RF transmitter and the peer
receiver to perform the radio frequency communication, at the end two 16 × 02 LCD
displays for the transmitter and receiver.

2.1 Measure Voltage and Current

Measuring electrical current and voltage and very useful for calculating energy and
power in electrical systems. In this paper, we used the current sensor ACS712-
30A (see Fig. 1) which is based on the Hall Effect magnetic field to measure the
electric current produced by the PVs, this sensor produces at output a direct voltage
proportional to the current at a rate of 0.066 V/A [8].
As you can see, this sensor is absolutely simple and contains only a few parts,
including the ASC712, passive components and connectors.
There are three variants of the ACS712 sensor depending on the current
detection interval ±5 A, ±20 A and ±30 A [8], the output sensitivity varies as
follows (Table 1):

Fig. 1 The ACS712 current

sensor
IoT, Design and Realization of a Supervision Device for Photovoltaic Panels … 367

Table 1 The sensitivity of

Model ACS712 Optimized current Output sensitivity
the current sensor ASC712
ACS712 ELC-05 ±5 A 185 mV/A
ACS712 ELC-20 ±20 A 100 mV/A
ACS712 ELC-30 ±30 A 66 mV/A

Fig. 2 Hall effect of the

ASC712-30A sensor

As mentioned before, the ASC 712 is based on the Hall effect (see Fig. 2). A
copper strip joins the pins IP+ and IP−. The generated magnetic field is detected by
the sensor to transform it into an adequate voltage.
Then we are going to make a voltage divider in order to measure the voltage at the
terminals of the photovoltaic panel so as not to damage the Esp32 microcontroller.
To do this, we will integrate a voltage divider because the voltage across the Esp32
must not exceed 5 V. We take R1 = 1 K and R2 = 12 K, the output voltage is
measured at the terminals of R2 when a voltage from the photovoltaic panel is applied
to the input of these two resistors (see Fig. 3).
Based on mesh law and then Ohm’s law with the voltage at pin34 of esp32, we can
retain the authentic value of the voltage at the terminals of the photovoltaic panel.
N.B: we can easily calculate the power P consumed as well as the energy dissipated
because we have measured the two quantities, the current (I) and the voltage (V).

2.2 Radio Frequency Communication

In this part we are going to carry out a radio frequency communication of 433 MHZ
to transmit the values Voltage, Current, Energy and Power previously determined.
We integrate the FS1000A transmitter with an Esp32 card and the proprietary
receiver with an Arduino Uno card.

Hardware Description and Specifications of the 433 MHz Module [9]

• One-way wireless (RF) transmitter and receiver.
• Transmission distance: 3 m (without antenna) to 40 m (Convenient)
368 Z. Didi and I. El Azami

Fig. 3 Circuit diagram (current sensor and voltage divider)

• Modulation technique: ASK (amplitude modulation)

• Transmitter operating current: 9 mA to 40 mA.
• Operating frequency: 433 MHz
• Data transmission speed: 10 Kbps.

The FS1000A RF Transmitter

The FS1000A transmitter includes three pins, Vcc, Data, and GND, (see Fig. 4).
The Vcc pin has an input voltage of 3 V to 12 V. GND is ground (0 V). The DATA
pin is the data pin to which the signal to be transmitted is sent. After this signal is
modulated by ASK (Amplitude Shift Keying) [10] then transferred to the waves with
a frequency of 433 MHz.
The wiring of the transmitter is very simple, it has only three pins. The VCC pin
(5 V), the GND (0 V) on the ESP32. The Data-In pin is connected to pin n° 12 of
esp32. Figure 5 shows the wiring of the transmitter.

Fig. 4 Image of the

FS1000A transmitter
IoT, Design and Realization of a Supervision Device for Photovoltaic Panels … 369

Fig. 5 Connection of the FS1000A transmitter with the Esp32

The RF Receiver
The receiver associated with the FS1000A transmitter has four pins: Vcc, DATA and
the two middle pins are attached together, and the GND, (see Fig. 6).
Unlike the transmitter, the Vcc pin, the receiver must be powered by a regulated
5 V supply. GND is a ground pin (0 V).
Wiring the receiver is as easy as the transmitter. Noted that the output pin of the
receiver should be connected to pin 11 of the Arduino (see Fig. 7).

Organization Charts
• Radiofrequency transmitter.
Figure 8 shows the block diagram of the FS1000A transmitter.
• Radiofrequency Receiver

Figure 9 represents the block diagram of a radiofrequency receiver.

For the code, we must install the RadioHead library in our Arduino IDE, to ensure
data transfer between the transmitter and the receiver.

Fig. 6 The Radiofrequency

receiver
370 Z. Didi and I. El Azami

Fig. 7 Arduino UNO assembly and receiver

Fig. 8 Schematic
diagram—RF transmitter

Fig. 9 Schematic
diagram—RF receiver

Then we generate an ASK object called “rf_driver”. In the loop, we define a buffer
for the size of the transmitted message. You have to normalize this parameter to match
the message size, you have to include spaces and punctuation marks because they
evaluate as characters. We then examine whether we have obtained a valid package.
Finally, we use the strtok() function to capture the values: Current, Voltage, Power
and Energy [11].

Main Organization Chart

Figure 10 shows the main flowchart of operation of the assembly.

3 Results and Discussion

Figures 11, 12, 13 and 14 show the real-time results of our project following a
433 MHZ radio frequency link to transmit the two measured and calculated quantities
IoT, Design and Realization of a Supervision Device for Photovoltaic Panels … 371

Fig. 10 Main flowchart

(Voltage, Current, Power and Energy) while using the FS1000A transmitter with a
card Esp32 and the receiver associated with an Arduino Uno board. It should be
noted that in this experiment we chose the simplest solution, we soldered an antenna
on the ANT pin, with this trick we increased the distance between the transmitter
and the receiver to 100 m. unfortunately this type of antenna is omnidirectional.
This method presents an inexpensive solution for the transmission of data in real
time based on radio frequency, particularly in isolated “no internet” environments,
so it is a realization of a remote supervision arrangement of photovoltaic panels.
The images which represent our project are given as follows, Fig. 11 represents
the box of the RF transmitter and Fig. 12 represents the box of the RF receiver.
372 Z. Didi and I. El Azami

Fig. 11 Photo of the RF transmitter box

Figure 13 shows the data captured on the Arduino Uno serial monitor at the
receiver following a radio frequency link with a length of 30 m between transmitter
and receiver.
These data represent current, voltage, power and total energy (Current, Ts, Power,
Energy/Min and Energy/T).

Code Test
We can finally display the total energy variation per unit of time (see Fig. 14) on the
free platform ThingSpeak (Open-Source application) for the Internet of Things, it is
a service carried by MathWorks allowing to store and collect data from.
IoT, Design and Realization of a Supervision Device for Photovoltaic Panels … 373

Fig. 12 Photo of the RF receiver box

Fig. 13 Testing the code on the Arduino serial monitor

374 Z. Didi and I. El Azami

Fig. 14 Variations in total energy

4 Conclusion

As you can see, adding a wireless link to Arduino implementations is very easy
and inexpensive to send/receive data one way, with 433 MHz transmit and receive
modules.
These RF modules are absolutely matched to the job at hand and the price is
affordable.
For an efficient and robust operation of the photovoltaic solar energy system, a
vigilance system is mandatory to identify difficulties and system failures before any
major failure.
By taking advantage of the different measured data, we can simply draw particular
data over time intervals, which leads to a commonly traceable pattern for detecting
operating failures. Unfortunately, this type of antenna is omnidirectional so there is
no point in radiating in the opposite direction and to the sides, which is not useful it
is necessary to better guide the radiation in a desired direction.
In addition, the development of the protection system in the solar photovoltaic
energy system increases the reliability of the system and ensures its proper func-
tioning. An effective monitoring system can help the user of that energy system keep
the system running smoothly for years to come.

References

1. NREL: Best research-cell efficiency chart. https://www.nrel.gov/pv/assets/pdfs/best-research-

cell-efficiencies.20190802.pdf. Accessed Sep 2019
IoT, Design and Realization of a Supervision Device for Photovoltaic Panels … 375

2. Chen B, Yu Z, Liu K, Zheng X, Liu Y, Shi J, Spronk D, Rudd PN, Holman Z, Huang J (2019)
Grain engineering for perovskite/silicon monolithic tandem solar cells with efficiency of 25.4%.
Joule 3:177–190. https://doi.org/10.1016/j.joule.2018.10.003
3. Daliento S, Chouder A, Guerriero P, Pavan AM, Mellit A, Moeini R, Tricoli P (2017) Moni-
toring, diagnosis, and power forecasting for photovoltaic fields: a review. Int J Photoenergy.
https://doi.org/10.1155/2017/1356851
4. Chaibi Y, Malvoni M, Chouder A, Boussetta M, Salhi M (2019) Simple and efficient approach
to detect and diagnose electrical faults and partial shading in photovoltaic systems. Energy
Convers Manag 196:330–343. https://doi.org/10.1016/j.enconman.2019.05.086
5. Vestenicky M, Matuska S, Hudec R (2019) Simple method of photovoltaic panel power charac-
teristic measurement based on Arduino hardware platform. Transp Res Procedia 40:579–585.
https://doi.org/10.1016/j.trpro.2019.07.083
6. Yatchev I, Sen M, Balabozov I, Kostov I (2018) Modelling of a hall effect-based current
sensor with an open core magnetic concentrator. Sensors 18(4):1260. https://doi.org/10.3390/
s18041260
7. de Arquer Fernandez P, Fernandez MAF, Cand JLC, Arboleya PA (2021) An IoT open source
platform for photovoltaic plants supervision. Int J Electr Power Energy Syst 125:106540.
https://doi.org/10.1016/j.ijepes.2020.106540
8. Allegro Microsystem. ACS712: fully integrated, hall effect-based linear current sensor with 2.1
kVRMS voltage isolation and a low-resistance current conductor (2017), ACS712-DS, Rev.
7. Northeast Cutoff Worcester, Massachusetts 01615-0036 U.S.A, https://www.sparkfun.com/
datasheets/BreakoutBoards/0712.pdf. Accessed 25 Aug 2017
9. Smith DW (2013) Chapter 12 - radio transmitters and receivers. PIC projects and applications
using C. A project-based approach, 3rd edn., pp 129–134. https://doi.org/10.1016/B978-0-08-
097151-3.00012-X
10. Binh LN (2008) Multi-amplitude minimum shift keying modulation format for optical commu-
nications. Opt Commun 281:4245–4253. https://doi.org/10.1016/j.optcom.2008.04.041
11. Attaway S (2016) Matlab, 4th edn. (Cover date: 2017), pp 237–266. https://doi.org/10.1016/
B978-0-12-815479-3.00007-6
ESP8266 Wireless Indoor Positioning
System using Fingerprinting
and Trilateration Techniques

Safae El Abkari, Jamal El Mhamdi, Abdelilah Jilbab,

and El Hassan El Abkari

Abstract Location-based systems have come under the spotlight in the last decade.
They became significant with the bloom of applications that have tracking services.
Indoors, wireless technologies can be an alternative for precise positioning in small or
large areas where GPS signals are too weak and inaccurate. In this paper, we present
a Wi-Fi Received Signal Strength based system using fingerprinting and trilateration
techniques. We constructed radio- maps by collecting signals with ESP8266 modules
in the offline phase. Then, we looked for the high number of matches in real-time
by comparing collected signals to those stored in database. Simultaneously, we use
trilateration to find positions with three nearby APs in the online phase. We achieved
an accuracy of 98% by combining fingerprinting and trilateration locating techniques.
We also analyzed the error probability for position calculation at different time of
the day and demonstrated high stability of our system with a low standard deviation.

Keywords Wireless · Indoor · Position · Localization · RSS · Fingerprinting ·

Trilateration · ESP8266

1 Introduction

The wide use of location services has increased with the growth and the popu-
larity of locating computation applications in indoor environments, such as airports,
schools, malls… etc. [1]. Consequently, tracking and positioning system industries
have emerged and developed smart locating technologies. Because of the consistent
emergence of location-based systems, [2] estimates that the general annual revenues
of location-based system will approximately reach 25$ billion by 2022. In fact, Indoor
localization offers an improvement of healthcare monitoring, network management,
security and personal navigation [3, 4]. Multiple indoor locating technologies such
as RFID [5], Wi-Fi [6] and Bluetooth [7] have been utilized for indoor locating
systems. Most of those technologies require massive hardware and infrastructure

S. El Abkari (B) · J. El Mhamdi · A. Jilbab · E. H. El Abkari

E2SN, ENSAM, ST2I, Mohammed V University in Rabat, Rabat, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 377
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_35
378 S. El Abkari et al.

deployment, which leads to high complexity and cost maintenance. Even though the
main functionality of Wi-Fi is not locating service, Wi-Fi seems to be a promising
technology and potentially a counterpart of Global Positioning System (GPS) [8]
in indoor environment. Therefore, Wi-Fi localization using fingerprinting technique
has been adopted by many researchers because of its availability in most indoor
environments (no extra hardware needed).
Wi-Fi based locating system using fingerprinting is employed to overcome the
path-loss model issue which consists of variation of the received signal strength
(RSS) intensity at certain positions caused by multi-path fading effect in indoor
environments [9]. Thus, the knowledge of the received signal properties such as
the RSS distribution, maximum, minimum, median, mean and standard deviation
are important for the correct modeling of fingerprinting locating system. In fact,
the knowledge of necessary data number to achieve high accuracy and number of
necessary APs for precise location identification can be provided by studying the
nature of RSS and how it is affected by the environment.
In this paper, we present a Wi-Fi based positioning system using fingerprinting
and trilateration techniques. In the first stage which is the offline phase, we collect the
RSS data to construct radio-maps. In the second stage, we look for the high number of
matches while we calculate the position using trilateration. Results demonstrate high
positioning accuracy in presence of indoor constraints (fluctuations of RSS signals).

2 Wireless Positioning Systems

2.1 Related Works

In this section, we discuss the existing literature about the wireless localization
systems (Table 1).

2.2 Our Contribution

We propose a Wi-Fi based locating system using fingerprinting and trilateration

techniques. In the first stage which is the offline phase, we collected RSS values from
APs to construct positioning radio-maps. In the second stage, we started the location
process by reading RSS value in real-time. We determined positions by looking for
the high number of matches while calculating positions using trilateration. We also
studied the error probability for position calculations at different time of the day and
analyzed the statistical values of error in different experiment scenarios. Our system
can be adapted to various applications and easily be deployed because of the Wi-Fi
technology popularity.
ESP8266 Wireless Indoor Positioning System ... 379

Table 1 Related works

Ref. Description Pros and cons
[10] They presented variable trilateration Six positions which is a small number were
algorithms for indoor navigation. They tested on a single floor
created an android-based application that User configurations are not discussed in the
measures RSS of nearby Wi-Fi APs. The paper
server calculates positions when RSS data is
received and positions are returned to a
custom-built radio-map. They reached an
average margin of error of 1.99 m
[11] They provided an indoor positioning system Accuracy is not disclosed
using Wi-Fi trilateration Stopped providing their positioning system
when they rebranded to MazeMap
[12] They proposed a commercial indoor User configurations are not fully discussed
positioning system [12]. They combine
Wi-Fi and BLE to determine positions
[13] It provided a positioning system that Methodology is not dicussed
estimates positions of the user with an error Complex process
of two to three meters. They combine
accelerometer, compass and Wi-Fi and BLE
fingerprinting to achieve this accuracy

2.3 Positioning Wireless Technologies

This section presents a survey of the most three prominent Radio Frequency
technologies used for indoor positioning as illustrated in table below.
• RFID: is a process of automatic identification using Radio Frequency communi-
cation between RFID readers and tags for Indoor locating. It determines the posi-
tion and the orientation of people or assets carrying the tag. It permits Real-Time
positioning, with less complex programming and a very high response time.
• Wi-Fi: is a wireless technology that works at high frequency waves to ensure
connectivity and communication between the network nodes in indoor locating
environment. It is suitable for locating in presence of multipath. Thus, there is
no need of extra hardware due to the availability and the popularity of Wi-Fi
infrastructure.
• Bluetooth: is a technology that uses the client-server architecture. By con-trolling
Bluetooth devices specifications such as communication parameters and emis-
sion power, we can enable the communication between the nodes of the network
and start the positioning process. It permits high data speeds but has mobility
limitations.
380 S. El Abkari et al.

Table 2 Review and comparison of indoor radio frequency technologies

Technology Technique Algorithm Complexity Accuracy Real-time
positioning
RFID Fingerprinting RSS Medium Low Yes
Wi-Fi Trilateration RSS Medium Medium Yes
fingerprinting TDOA
Bluetooth Trilateration RSS Medium Low – Yes
fingerprinting TDOA Medium

2.4 Positioning Methods

• Received Signal Strength: this method is only possible for radio signals [14]. It
estimates the distance of an unknown tag to a reference access point by using the
attenuation of emitted signals strength. Basically, the smaller the RSS values the
farthest the tag is from the reference access point (AP).
• Time Of Arrival (TOA): It is a method based on the arrival time synchronization
of an emitted and received signal [14]. We calculate distance between the emitter
and the receiver using the following equation:

distance = time ∗ speed o f the signal (1)

2.5 Positioning Techniques

Multiple techniques (Table 3) are utilized in wireless based locating system. In this
section, we present two techniques used for Wi-Fi locating which are:
• Fingerprinting: it is a method consisting of two phases [15]:

– The Offline phase: which is also called calibration or training phase. At this
phase, we set up radio-maps by empirical measurements with the definition of a
reference node. The radio-maps are databases of radio signal characteristics such
as the SSID/ RSS and coordinates.
– The Online phase: wherein the process of locating is operational. The tag carried
by the user measures SSID and RSS values of the signals at an unknown position.
This value is compared then to those in the radio-map to find the higher number
of matches.

The main disadvantages of this technique are the time consumption in the offline
phase and its sensitivity to changes in the locating environments. Therefore, the
process of calibration must be repeated from time to time to guarantee a good
accuracy.
ESP8266 Wireless Indoor Positioning System ... 381

Table 3 Comparison between trilateration and fingerprinting techniques

Technique Type of measurement Affected by Coverage Accuracy
multipath
Trilateration Difference of arrival time Yes Good (in case of High
line-of-sight signals)
Fingerprinting Received signal No Good High
Strength

Table 4 ESP8266 hardware

Module
specifications

Type Single board microcontroller

CPU ESP8266
Storage 4MBytes
Memory 128kBytes
Power supply 3V, 5V
IDE Arduino IDE

• Trilateration: it is a method consisting of the calculation of positions using

measurements of distances [16]. This technique is based on the Received Signal
Strength (RSS) or the propagation time (TOA…). Position of a person or an asset
is determined from collected measurements using at least three nearby access
points (Table 3).

2.6 ESP8266 Module

In this work, we used ESP8266 NodeMCU module as a locating node and as a mobile
tag. ESP8266 is Wi-Fi microchip with microcontroller capability and TCP/IP stack. It
is cost effective and highly integrated for various IoT applications. Table 4 illustrates
the hardware specifications of ESP8266 module.

3 Experimental Work and Results

Our experimental work is divided into two main parts:

• The first one uses fingerprinting technique wherein we constructed radio-maps
(offline phase) in an indoor environment (8x8 meters) divided into 12 grid cells
of same sizes. We then defined a calibration point (CP) at the center of each cell.
382 S. El Abkari et al.

We collected RSS values using an ESP8266 Module [17] at the 12 grids at a

fixed orientation. We obtained different RSS samples at a unique cell vector and
different sampling times. We constructed two radio-maps, with 100 samples for
the first one and 300 samples for the second one. Using a laptop connected to an
ESP8266 module, we retrieved RSS/SSID information and we stored all values
in dBm.
• The second experimental work uses trilateration to find positions in 2D using three
access points (APs). We placed APs in a way that minimizes the effect of the multi-
path effects. We divided then the area of location into coordinates (x, y).
The access points used for the experiment are ESP8266 modules [18, 19] that
supports 802.11 n band operating at 2.4 GHz only. At full strength, ESP8266 nodes
transmit at a maximum power of 25 mW which corresponds to 14 dBm.
Collected RSS values from same APs fluctuate and vary because of walls, door…
etc. We handle this issue and reduce it by correcting RSS signals. When part effects
of its fluctuations are eliminated, collected RSS signal become more stable.

3.1 Locating Using Fingerprinting

3.1.1 Radio Map

Radio-map is constructed using collected signals in an area of 8 × 8 m using algo-

rithm below. We divided the locating area into 12 grid cells with equal surfaces. We
placed the calibration point in each cell grid center. We use 3 APs (ESP8266 Module)
to collect RSS values. To achieve a suitable radio-map and a better experiment perfor-
mance, we constructed two radio-maps with 100 and 300 samples of RSS. The laptop
is then used to process received data and retrieve SSID/RSS information.
Note that corrected RSS signals guarantee a best performance of both collection
and matching processes.

3.1.2 RSS Measurements in Offline and Online Phases

• Mean and standard deviation of collected signals

Devices working under 802.11 standards such as ESP8266 use Clear Channel Assess-
ment protocol to prevent collision between two signals transmitted on the same
channel [20]. This method checks if any other SSID than the one emitting is transmit-
ting on the channel before sending data (frames). Table below shows the statistical
values of our three APs (ESP8266 modules) at two different locations of the user
tag (used for collection). One of the positions is near to APs while the second one is
far from the APs. This table shows that RSS mean values and standard deviation of
each ESP are similar at both position P1 and P2 (Table 5).
ESP8266 Wireless Indoor Positioning System ... 383

Table 5 Statistical values of RSS at two positions

Position Position P1 Position P2
SSID ESP1 ESP2 ESP3 ESP1 ESP2 ESP3
RSS mean value (dBm) −32 −33 −34 −74 −72 −69
Standard deviation 4.5 5.4 5.1 2.2 2.5 2.3

Table 6 Distance
SSID P1 = 1m P2 = 4m P3 = 8m
measurement at three
positions ESP1 1.72 2.56 1.84
ESP2 1.9 2.14 2.23
ESP3 1.51 2.97 1.96

Table 7 Statistical values of

Method Average (m) Standard Precision
RSS at two positions
deviation
Fingerprinting only 1.124 1.658 75%
Fingerprinting + 0.697 1.433 89%
trilateration

• Path-loss of ESP8266 RSS-signals

According to the path-loss model [21], RSS values decrease with the increasing
of the distance. Table 6 shows that when distance increases, the standard deviation
decreases. Thus, measurements of ESP mean standard deviation are most comparable
to the real path-loss model than the individual ones.

3.2 Locating Using Fingerprinting/Trilateration

To guarantee a qualitative performance of the positioning process, we place the

APs (EPS8266 modules) at defined positions to minimize geometrical errors that
maximize position errors [22–24] when locating using trilateration.
Our next experiment is carried out to evaluate performances of our proposed
method which demonstrates a high positioning quality. In fact, with our proposed
method we reach 89% of precision while 75% when using only fingerprinting for
positioning at P = 0.6 m (Table 7).
The presence of the experimental environment constraints such as doors, walls,
presence of other signals… etc., positioning error increases with the increasing of
distance. Therefore, we study the performance of our method by comparing the
median, mean and maximum errors in different locating environment scenarios. Thus,
our method demonstrates high stability with a low standard deviation compared to
traditional location techniques (Table 8).
384 S. El Abkari et al.

Table 8 Statistical values of error in different experiment scenarios

Position Median Mean Maximum
error (m) error (m) error (m)
RSS Positioning 1.4971 1.7847 4.1052
RSS positioning + Wi-Fi APs 2.4563 2.7645 6.4882
RSS positioning + fingerprinting 0.86 0.4677 2.79
+ other Wi-Fi APs

Table 9 Comparison of existing methods and our proposed locating technique

Method Description Accuracy
RedPin [25] Indoor map Positioning by the user 90% (Room Level)
FreeLoc [26] Comparison of relative RSS <3m
Manifold learning [27] Special correlation of RSS Path loss 2.4 m (4 APs)
model
Indoor map
ARIADNE [28] Path loss model Indoor map 2.5 m – 3 m
Simulated annealing algorithm
Forward Interpolation [29] RSS RMS 2.82 m (4 APs)
Cubic spline
Our proposed method RSS + Filtration Fingerprinting and 0.4677 m (3APs, room level)
trilateration Combination

3.3 Comparison of Our Contribution to Existing Methods

Author in [25] presented the RedPin locating system that reached 90% of accuracy
in a room level. Author in [26] presented the FreeLoc that reached an accuracy less
than three meters. Manifold learning system was presented in [27]. It used four APs
and achieved an accuracy of 2.4 m. In paper [28], the author presented ADRIANE
system which achieved 2.5 m of accuracy. Forward Interpolation was introduced in
[29] and reached 2.82 m of accuracy with four APs while our system used three APs
and reached 0.46 m of accuracy (Table 9).

4 Conclusion

In this paper, we presented a Wi-Fi technology based indoor locating system using
fingerprinting and trilateration techniques. We implemented the positioning proce-
dure in ESP8266 modules and we carried our experiment in an 8 × 8 m area. In the
first stage which is the offline phase, we constructed RSS radio-maps by collecting
RSS fingerprints at certain positions. In the online phase, the mobile user carried an
ESP8266 module as a tag to read RSS values, the fingerprints of Wi-Fi and radio-
ESP8266 Wireless Indoor Positioning System ... 385

maps. We determined positions by finding the high number of fingerprints’ matches

while combining it with trilateration technique to increase accuracy. Results show
that despite of the presence of indoor environment constraints, we can determine
precise positions. In fact, our proposed system reaches a mean error of 0.4677 m
which is sufficient in our study case. We also analyzed the error probability for posi-
tion calculation at different time of the day and demonstrated high stability of our
system with a low standard deviation.
Besides the good performance of our proposed method, the positioning error is
considerably affected by the number of access points (number of ESP modules used
in the experiment). Therefore, our future will focus on the enhancement of our results
by expanding the number of AP within the locating network.

References

1. Yang Z, Wu C, Liu Y (2014) Location-based computing: localization and localizability of

wireless networks. Tsinghua University Press
2. Indoor Positioning and Navigation System Market - Forecast 2022 | IPIN System
Market. https://www.marketresearchfuture.com/reports/indoor-positioningnavigation-system-
market-1775. Accessed 26 Jun 2019
3. Mhamdi J, El Abkari S (2015) Contriving an RFID system for Alzheimer patients tracking.
In 3rd international workshop on RFID and adaptive wireless sensor networks, pp 23–28,
7173273, Agadir
4. Bandyopadhyay A, Hakim D, Funk BE, Kohn EA, Teolis CA, Blankenship G (2014) System
and method for locating, tracking, and/or monitoring the status of personnel and/or assets both
indoors and outdoors. US8712686B2
5. Giuliano R, Mazzenga F, Petracca M, Vari M (2013) Indoor localization system for first
responders in emergency scenario. In 9th international wireless communications and mobile
computing conference (IWCMC), pp 1821–1826
6. El Abkari S, Jilbab A, El Mhamdi J (2018) Real time positioning over WSN and RFID net-
work integration. In 4th international conference on advanced technologies for signal and image
processing (ATSIP), pp 1–5. IEEE, Sousse
7. Biswas J, Veloso M (2010) WiFi localization and navigation for autonomous indoor mobile
robots. In IEEE international conference on robotics and automation, pp 4379–4384. IEEE,
Anchorage
8. Forouzan AB (2006) Data communications & networking (sie). Tata McGraw-Hill Education
9. Kaplan ED, Hegarty C (2006) Understanding GPS: Principles and Applications, 2nd edn.
Norwood, MA, USA, Artech House
10. Pathak O, Palaskar P, Palkar R, Tawari M (2014) Wi-Fi indoor positioning system based on
RSSI measurements from wi-fi access points—a trilateration approach. Int J Sci Eng Res
5:1234–1238
11. CampusGuiden. https://www.campusguiden.no/. Accessed 11 May 2017
12. Accuware. https://web.archive.org/web/20170511084748/https://www.accuware.com/.
Accessed 10 May 2019
13. Accuware: accuracy. https://web.archive.org/web/20170511084748/https://www.accuware.
com/. Accessed 10 May 2019
14. Insiteo. https://web.archive.org/web/20170511084949/https://www.insiteo.com/joom-la/ind
ex.php/en/. Accessed 21 Jul 2020
15. Multi-SSID Deployment Considerations. https://documentation.meraki.com/@api/de-ki/pag
es/1985/pdf/Multi-SSID/. Accessed 03 May 2020
386 S. El Abkari et al.

16. Gezici S (2008) A survey on wireless position estimation. Wirel Pers Commun 44(3):263–282
17. Gu Y, Lo A, Niemegeers I (2009) A survey of indoor positioning systems for wireless personal
networks. IEEE Commun Surv Tutorials 11(1):13–32
18. Zhang D, Xia F, Yang Z, Yao L, Zhao W (2010) Localization technologies for indoor human
tracking. In 5th international conference on future information technology, pp 1–6. IEEE
19. Ismail MB, Fathi A, Boud A, Nurdiana W, Ibrahim W (2008) Implementation of location
determination in a wireless local area network (WLAN) environment. In 10th international
conference on advanced communication technology, vol 2, pp 894–899
20. El Abkari S, Jilbab A, El Mhamdi J (2020) RSS-based indoor positioning using convolutional
neural network. Int J Online Biomed Eng 16(12):82–93
21. Barai S, Biswas D, Sau B (2017) Estimate distance measurement using Node MCU ESP8266
based on RSSI technique. In conference on antenna measurements & applications (CAMA),
pp 170–173. IEEE
22. ESP8266 nodemcu. https://www.nodemcu.com/. Accessed 27 Sep 2019
23. Laitinen E, Lohan ES (2015) Are all the access points necessary in WLAN-based indoor
positioning. In: international conference on location and GNSS (ICL-GNSS), pp 1–6. IEEE
24. Baala O, Caminada A (2006) WLAN-based indoor positioning system: experimental results
for stationary and tracking MS. In International Conference on Communication Technology,
pp 1–4
25. El Abkari S, Jilbab A, El Mhamdi A (2020) Wireless Indoor localization using fingerprinting
and Trilateration. J Adv Res Dyn Control Syst 12(7):2597–2602
26. Bolliger P (2008) Redpin adaptive, zero configuration indoor localization through user collabo-
ration. In Proceedings of the first ACM international workshop on Mobile entity localization
and tracking in GPS-less environments - MELT 2008, pp 55–60, San Francisco, CA, USA
27. Yang S, Dessai P, Verma M, Gerla M (2013) FreeLoc: calibration free crowdsourced indoor
localization. In Proceedings IEEE INFOCOM, pp 2481–2489, Turin, Italy
28. Sorour S, Lostanlen Y, Valaee S, Majeed K (2015) Joint indoor localization and radio-map
construction with limited deployment load. IEEE Trans Mob Comput 14(5):1031–1043
29. Ji Y, Biaz S, Pandey S, Agrawal P (2006) ARIADNE: a dynamic indoor signal map construc-
tion and localization system. In Proceedings of the 4th international conference on Mobile
systems, applications and services - MobiSys 2006, pp 151–164, Uppsala, Sweden
30. Peng L, Han J, Meng W, Liu J (2010) Research on radio map construction in indoor
WLAN positioning system. In: First International Conference on Pervasive Computing, Signal
Processing and Applications, pp. 1073–1077, Harbin, China
High Performance Predictive Control
for Permanent Magnet Synchronous
Machine Drive: FPGA-Based
Implementation

Badre Bossoufi and Ahmed Lagrioui

Abstract The use of PMSM motors is very important in industry today; they require
great robustness in terms of performance (speed, torque, flux, etc.). In this work we
present a new control strategy to improve the performance of the machine given its
non-linear model. This technique based on a new non-linear predictive controller,
which is based on the expansion of the Taylor series. The minimum variance of the
nonlinear predictive controller (CPNLVM) aims to minimize the speed error and the
robustness of the systems given the parametric variation. Another contribution in this
work is the real-time implementation on the FPGA card which will be connected to
an experimental Benchmark. First, the operating principle of the Predictive control
is developed, and then it is validated on Matlab & Simulink. Subsequently, an exper-
imental validation on a Benchmark based on PMSM and FPGA in order to test the
proposed technique. The results obtained (simulation and experimental) show the
validation of the proposed model and also a great improvement in terms of system
performance.

Keywords Predictive control · FPGA · PMSM · Robustness · Adaptive control

1 Introduction

The PMSM Synchronous Permanent Magnet Machine is a very popular candidate

in high performance applications such as dams or machine tools, also in robotics….
For this, the performance of my machine must be very robust and stable (Speed,
Torque).
The PMSM synchronous machines are renowned for their torque compared to
other types of machines; the use of permanent magnets reduces their losses and

B. Bossoufi (B)
LIMAS Laboratory, Faculty of Sciences Dhar El Mahraz, Sidi Mohamed Ben Abdellah
University, Fez, Morocco
A. Lagrioui
Department of Electrical and Computer Engineering The Higher National School of Arts and
Trades, Moulay Ismail University, Meknes, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 387
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_36
388 B. Bossoufi and A. Lagrioui

improves their performance. Recently, the start-up of the PMSM without speed
sensor has attracted great interest in the industry where the use of position sensors is
limited. This advantage reduces the cost of training and improves the robustness of
the machine. The quality of the performance of the PMSM machine is mainly based
on knowledge of the position and speed of the rotor. The control strategies for esti-
mating the position of the sensorless rotor are mainly based on the non-linear PMSM
motor model. Most of the control strategies are based on the motor model detecting
the FEM, which is directly related to the position and speed of the rotor, using esti-
mation observers. Other strategies are based on flux estimation in the machine model
[2]. In addition, adaptive estimate observers were used to estimate the stator current,
speed and position of the rotor.
The predictive controller used in this work is a nonlinear technique of minimal
variance (CPNLVM) which is based on the expansion of the Taylor series; its objec-
tive is to minimize the error between the output and the reference applied to the
machine.
Currently, the use of FPGA technology has become very important in various fields
(signal processing, Telecom, Control systems, Calculators, Robotics….) thanks to
its simplicity of operation and its memory capacity and the number of ‘I/O and its
price compared to other types of cards (DSP, microcontrollers).
The field of electric machine control has known a great evolution in the industrial
field thanks to this type of technology, several techniques have been successfully
implemented on the FPGA target (PWM, FOC, Backstepping, DTC,….). In this
work we are interested in the implementation of a new control technique (predctive)
on the FPGA card in order to improve the performance of the systems in terms of
speed and robustness.
Figure 1 shows the general structure of the control of a PMSM motor:

Fig. 1 Architecture of the control

High Performance Predictive Control ... 389

2 Non-linear Model of PMSM Motor

The model of my Synchronous permanent magnet machine in the park domain in

state representation, taking account of disturbances is presented by the following
equations [4]:

ẋ = f (x) + g1 u + g2 b
(1)
y = h(x)
⎡ ⎤
  i sd      
u sd ⎣ ⎦ y1 i sd fd
u= ; x = i sq ; y = = ; b= (2)
u sq y2  f

⎡ 1 ⎤ ⎡ ⎤
0 − 1
0

⎢ sd
L
⎥
⎢ L sd ⎥
g1 = gd gq = ⎣ 0 L1sq ⎦; g2 = bd bq = ⎣ 0 0 ⎦ (3)
0 0 0 − 1J

With:
⎡Rs L sq ⎤
⎡ ⎤ ⎢ −.i + p .i .
L sd
sd
L sd
sq
⎥
f 1 (x) ⎢ ⎥
⎢ ⎥ ⎢⎢ R s L sd  f ⎥
⎥
f (x) = ⎣ f 2 (x)⎦ = ⎢ − .i sq − p .i sd . − p . ⎥ (4)
⎢ L sq L sq L sq ⎥
f 3 (x) ⎣ ⎦
f 3p 3p
− . + (L sd − L sq ).i sd .i sq +  f .i sq
J 2J 2J
With b represents the disturbance which influences the regulation of the outputs,
its dynamics are slow compared to the state variables. We have:

db
=0 (5)
dt

3 Predictive Controller Observer Minimum Variance

The cost function proposed to develop the predictive control technique for PMSM
is written as follows [5]:

T1 T2
1 1
= (ed (t + τ )) dτ +
2
(e (t + τ ))2 dτ (6)
2 2
0 0
390 B. Bossoufi and A. Lagrioui

ed (t + τ ) = y1r (t + τ ) − y1 (t + τ )
(7)
e (t + τ ) = y2r (t + τ ) − y2 (t + τ )

Where: τ is the prediction time.

yi (t + τ ) is the prediction of the output to τ time.
yir (t + τ ) is the reference path to the time future.
After using this cost function, the machine equations become: [6]:
⎧

⎨ y1 (t + T1 ) y2 (t + T2 ) T = i sd (t + T1 ) (t + T2 ) T
⎩ y (t + T ) y (t + T )
T = i
T (8)
1r 1 2r 2 sdr e f (t + T1 ) r e f (t + T2 )

The objective of this control technique is to find the control quantities which
minimize the performance error defined by (7).
The exits and their future instructions are calculated approximately by the expan-
sion of the Taylor series. The error between the output yi and the reference yir is
repeated ρi time to reveal the regulation, hence [8]:
⎧

⎨ y1 (t + τ ) y2 (t + τ ) T = (τ )Y (t)
⎩ y (t + τ ) y (t + τ )
T = (τ )Y (t)
(9)
1r 2r r

With

Y (t) = [y1 y2 ẏ1 ẏ2 ÿ2 ]T
(10)
Yr (t) = [y1r y2r ẏ1r ẏ2r ÿ2r ]T
   
1 (τ ) 10τ 0 0
(τ ) = = (11)
1 1 0 τ τ2
2
2 (τ )

Using (9) in Eq. (6), the quadratic performance criterion will be:

1
= (Y (t) − Yr (t))T (T1 , T2 )(Y (t) − Yr (t)) (12)
2
T1 T2
(T1 , T2 ) = (1 (τ )) 1 (τ )dτ +
T
(2 (τ ))T 2 (τ )dτ (13)
0 0

We get after development:

High Performance Predictive Control ... 391

T2
⎡ ⎤
T1 0 21 0 0
⎢ 2 3 ⎥
⎢ 0 T2 0 T22 T62 ⎥
⎢ T2 ⎥
(T1 , T2 ) = ⎢ T13
⎢ 2 02 3 03 04 ⎥
1 ⎥ (14)
⎢ T2 T2 T2 ⎥
⎣ 0 2 0 3 8 ⎦
T3 T4 T5
0 62 0 82 202

After the combination of (5) and (10), we will have:

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
y1 (t) − y1r (t) 01×2 01×2
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ y2 (t) − y2r (t) ⎥ ⎢ 01×2 ⎥ ⎢ 01×2 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢
Y (t) − Yr (t) = ⎢ L f h 1 (x) − ẏ1r (t) ⎥ + ⎢ L g1 h 1 (x) ⎥u(t) + ⎢ L g2 h 1 (x) ⎥ ⎥b(t) (15)
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ L f h 2 (x) − ẏ2r (t) ⎦ ⎣ 01×2 ⎦ ⎣ L g2 h 2 (x) ⎦
2
L f h 2 (x) − ÿ2r (t) L g1 L f h 2 (x) L g2 L f h 2 (x)

In the absence of disturbance, the law of control minimizing the cost function is
obtained from the resolution of (16) [9]:

d
=0 (16)
du
This leads us to solve the following equations:
 T
d d
= (Y (t) − Yr (t) (T1 , T2 )(Y (t) − Yr (t)) = 0 (17)
du du
⎛⎡ ⎤⎞T ⎛⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎞
01×2 y1 (t) − y1r (t) 01×2 01×2
⎜⎢ ⎥⎟ ⎜⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎟
⎜⎢ 01×2 ⎥⎟ ⎜⎢ y2 (t) − y2r (t) ⎥ ⎢ 01×2 ⎥ ⎢ 01×2 ⎥ ⎟
⎜⎢ ⎥⎟ ⎜⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎟
⎜⎢ L g h (x) ⎥⎟ (T )⎜⎢ L f h 1 (x) − ẏ1r (t) ⎥ + ⎢ L g h (x) ⎥u(t) + ⎢ L g h (x) ⎥b(t)⎟ = 0 (18)
⎜⎢ 1 1 ⎥⎟ ⎜⎢ ⎥ ⎢ 1 1 ⎥ ⎢ 2 1 ⎥ ⎟
⎜⎢ ⎥⎟ ⎜⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎟
⎝⎣ 01×2 ⎦⎠ ⎝⎣ L f h 2 (x) − ẏ2r (t) ⎦ ⎣ 01×2 ⎦ ⎣ L g2 h 2 (x) ⎦ ⎠
L g1 L f h 2 (x) L 2f h 2 (x) − ÿ2r (t) L g1 L f h 2 (x) L g2 L f h 2 (x)

The optimal solution for this system is:

⎛⎡  ⎤ ⎞

1
(i)
⎜⎢ i=0 K i1 y1r (t) − L if h 1 (x) ⎥ ⎟
u(t) = G 1 (x)−1 ⎜ ⎢
⎝⎣ 2  ⎥ ⎟
⎦ − G 2 (x)b(t)⎠ (19)
(i)
K i2 y2r (t) − L if h 2 (x)
i=0
⎧
⎪ 3
⎪
⎨ K 0 = 2T ; K 1 = 1
1 1
1
(20)
⎪
⎪ 10 5
⎩ K 02 = ; K2 = ; K2 = 1
3T22 1 2T2 2

The expression developed in the predictive control law (CPNLVM) is that of the
matrix G1 (x) defined by (6). For this we define a generalized predictive command
matrix G2 (x) is given by:
392 B. Bossoufi and A. Lagrioui

   − L1sd 0

L g2 h 1 (x)  
G 2 (x) = = 3 p(L sd −L sq )i sq
L g2 L f h 2 (x) + K 12 L g2 h 2 (x) − 2J L sd
1
J
−K 2
1 + f
J
(21)

The dynamic error taking into account the disturbance can be written under:

K 01 ed (t) + K 11 ėd (t) = 0
(22)
K 02 e (t) + K 12 ė (t) + K 22 ë (t) = 0

Equation leads to the following roots:

⎧
⎪ 3
⎪
⎪ s1d = −
⎪
⎪ 2T1
⎪
⎨ 1.25 1, 3307
s1 = − −i (23)
⎪
⎪ T T2
⎪
⎪
2
⎪
⎪ 1.25 1, 3307
⎩ s2 = − +i
T2 T2

From Eq. (23), the system is stable. Otherwise, the prediction time is very small
which shows a great speed of the system. This is to the detriment of torque control
[10].
In cases where the perturbation constraint is not affected, its estimation will be
used. In the same way as in the case of predictive control (CPNLVM), the equation
of the disturbance estimator is written as follows:
⎡  t
 ⎤
⎢ μd K 1
0 ed (τ )dτ + ed (t) ⎥
b̂(t) = ⎢
⎣
 t
0 ⎥
⎦ (24)
μ K 02 e (τ )dτ + K 12 e (t) + ė (t)
0

We also show that the stability of the controller (24) is ensured by:
 2 
μd K1 f
< 0 ; μ − 2 <0 (25)
L sd J J

Once the observer blocks are saturated, the law of control takes the following
form:
   ∗   
u dr e f (t) u d (t) u d (t)
u(t) = = + = u ∗ (t) + u(t) (26)
u qr e f (t) u q∗ (t) u q (t)
High Performance Predictive Control ... 393
⎛⎡  ⎤ ⎞

1
(i)
⎜⎢ i=0 K i1 y1r (t) − L if h 1 (x) ⎥ ⎟
u ∗ (t) = G 1 (x)−1 ⎜ ⎢
⎝⎣ 2  ⎥
∗ ⎟
⎦ + G 2 (x)b̂ (t)⎠ (27)
(i)
K i2 y2r (t) − L if h 2 (x)
i=0
⎛ ⎞
t
u(t) = −⎝G 1 (x)−1 G 2 (x)μ G 1 (x)(u(τ ) − sat (u(τ )))dτ ⎠ (28)
0

The limit values of the control variables are identical to the variables developed
for predictive control (CPNLVM).

4 Results and Simulations

The results obtained by applying a predictive observer on the PMSM, are given for
the following parameters:
• the sample time Tsc = 50 μs,
• the prediction time to control the current T1 = 0.5 ms
• prediction time to control the speed T2 = 5 ms.
• The constant of the controller: μd = –0.1 and μ = 0.00001.
In order to evaluate the performance of the predictive command applied to the
PMSM machine, we performed certain simulations on Simulink:
• Pursuit test or trajectory follow-up: to check the continuation of the speed.
• Robustness test: in order to guarantee correct operation despite the parametric
variation of the machine. The test was carried out for a change of the machine
parameters (Rs = 0,5Rsn, Lsq = 0,5Lsqn, Lsd = 0,5Lsdn, f = 0,5fn, J = 0,5Jn).
Another change in flux with an increase of 20%.
The Fig. 2 shows the general structure of the predictive controller based on the
equations developed in part 3:

4.1 Pursuit Test

394 B. Bossoufi and A. Lagrioui

Fig. 2 Diagram of the CPNLVM applied to PMSM

400
5 10
La vitesse de rotation et sa rÈfÈrence

is1(A)
Les courants direct et quadratique

is2(A)
Les courants statoriques
300 4
wref (Rad/s) is1(A)
5
w (Rad/s) 3
200
isd(A)
2 isq(A)
100 0

1
0
0 -5

-100
-1
-10
-200 0 1 2 3 4 5 6 7 8 9 10
0 1 2 3 4 5 6 7 8 9 10 -2
0 1 2 3 4 5 6 7 8 9 10 t/Te 4
t/Te x 10
4 x 10
t/Te 4
x 10

Fig. 3 Pursuit performance of the trajectory of the controlled system CPNLVM

Le courant statorique de la phase1(is1(A))
la vitesse de rotation et sa rÈfÈrence

300 12 10
Ce (N.m)
Les e coules de charge Cr
et ÈlectromagnÈtique Ce

200 10 Cr (N.m)
5
wref (Rad/s) 8
100
w (Rad/s)
6 0
0
4

-100 -5
2

-200 0
0 1 2 3 4 5 6 7 8 9 10 -10
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
4
t/Te x 10 t/Te 4 4
x 10 t/Te x 10

Fig. 4 Disturbance rejection performance of the system with a controller CPNLVM

4.2 Robustness Test

4.3 Parametric Uncertainties

The results obtained in this test (Fig. 3) show better tracking for speed with zero
dynamic error.
High Performance Predictive Control ... 395

Vitesse de rotation et sa rÈfÈrence

250 0.1 2.5

les courants direct et quadratique

200 0.08
2

Erreur de Vitesse
150 wref (Rad/s) 0.06

w (Rad/s)
100 0.04 1.5
isq(A)
isd(A)
50 0.02
1

0 0

0.5
-50 -0.02

-100 -0.04
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0
0 1 2 3 4 5 6 7 8 9 10
t/Te 4
x 10
t/Te 4
x 10 t/Te 4
x 10

La vitesse de rotation et sa rÈfÈrence

3 250
Courant de phase statorique is1(A)

0.08

2 200

0.06

Erreur de vitesse
1 150
wref (Rad/s)

w (Rad/s)
0 100 0.04

-1 50
0.02
-2 0

-3 -50 0

-4 -100
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
-0.02
t/Te 4 t/Te x 10
4
0 1 2 3 4 5 6 7 8 9 10
x 10
t/Te 4
x 10

Fig. 5 Performance of the controlled system CPNLVM in case of a fast dynamic trajectory with
anti saturation scheme

Regarding robustness (parametric variation), the results obtained (Fig. 4) show

the effect of the use of compensation in the predictive controller, this cancels the
static speed error with a very short response time. In Fig. 4 we observe that the
performance of the systems is of very good quality in terms of monitoring, stability
and speed.

5 FPGA-Based Implementation of an Robust Predictive

Control

FPGA technology (Actel, Xilinx and Altera…) offers a very interesting reconfig-
urable structure for the implementation of control strategies. It offers great flexibility
in programming and design. Currently, the FPGA offers the use of Blocks dedicated
to the field of control (PWM, PI).
The FPGA architecture was designed using CAD tools. The description of the
program is done graphically (ia blocks) or via an HDL (Hardware Description
Language) VHDL or Verilog programming language. These two languages are stan-
dardized and provide description at different levels, and above all the advantage of
being portable and compatible with all the FPGA technologies previously introduced.
Figure 5 summarizes the different steps for programming an FPGA card. A
programming tool (System Generator: merger of Xilinx and Matlab), allows you
to generate the command part program (already done on Simulink) in VHDL code
through routing steps, Timming, and I / O allocation. Thereafter the SG generates
a VHDL file to be downloaded into the memory of the FPGA. This file is called
bitstream.
396 B. Bossoufi and A. Lagrioui

Fig. 6 The steps of implementation in FPGA

Fig. 7 Prototyping platform control

In order to test the validation of the proposed control model, a test bench was
produced in the laboratory based on an FPGA card and PMSM machine (Fig. 6). For
this work, we use a FPGA card type XC3S500E Spartan3E from Xilinx.
1. Experimental results
Figure 7 presents the scheme of predictive control on Simulink-Xilink (System
Genertaor). The sampling frequency for this implementation is 20 kHz.
The functional model for predictive Controller from SYSTEM GENERATOR is
composed the different blocks:
• IC coder interface block to adapt the input of the FPGA with the position sensor.
High Performance Predictive Control ... 397

Fig. 8 Functional model for predictive controller from SYSTEM GENERATOR

Fig. 9 Stator current, abc-axis current in the PMSM, d-axis and q-axis current in the PMSM

• ADC analog analog printer interface block to adapt the input of the FPGA and the
converter (ADCS7476MSPS 12 bits A/D) which is linked to the current sensors
(Hall effect)
• Transformation blocks of Park (abs-dq) and Park Inverse (dq-abc),
• SVW block which is the most important, to provide switching signals to the IGBTs
of the inverter in the power section.
• Predictive control block to control the speed and stator currents of the PMSM,
• “Timing” block to manage the period of use of each bloc,
• RS232 block for signal synchronization and signal acquisition.
The implementation of the predictive control on FPGA card is characterized by a
reduced operation time.
398 B. Bossoufi and A. Lagrioui

Figure 8 presents the experimental results obtained from PMSM by applying a

predictive control with the FPGA platform. All the results were extracted from the
FPGA by the ChipScope tool from Xilinx (Fig. 9).
The experimental results obtained show the improvement in performance (in terms
of tracking and robustness) of the PMSM machine, using the predictive control
strategy.

6 Conclusion

This work was dedicated to the application of a new predictive control technique with
estimators on a PMSM-based system. As well as the real-time implementation of the
control algorithm on the FPGA target. The proposed model has been successfully
validated (pursuit and robustness) by a simulation on Simulink and also by practice
on a benchmark based on FPGA.

References

1. Rodic M, Jezernik K (2002) Speed-sensorless sliding-mode torque control of an induction

motor. IEEE Trans Ind Electr 49(1):87–95. https://doi.org/10.1109/41.982252
2. Monmasson E, Cirstea M (2007) FPGA design methodology for industrial control systems – a
review. IEEE Trans Ind Electron 54(4):1824–1842
3. Bossoufi B, Karim M, Ionita S, Lagrioui A (2012) Low-speed sensorless control of pmsm
motor drive using a nonlinear approach backstepping control: FPGA-based implementation. J
Theor Appl Inf Technol JATIT 36(1):154–166
4. Bossoufi B, Karim M, Ionita S, Lagrioui A (2012) Nonlinear non adaptive backstepping with
sliding-mode torque control approach for PMSM motor. J Electr Syst JES 8(2):236–248
5. Lee TC, Song KT, Lee CH, Teng CC (2001) Tracking control of unicycle-modeled mobile
robots using a saturation feedback controller. IEEE Trans Control Syst Technol 9(2):305–318
6. Bossoufi B, Karim M, Lagrioui A, Taoussi M, Derouich A (2015) Observer backstepping
control of DFIG-generators for wind turbines variable-speed: FPGA-based implementation.
Renew Energy 81:903–917. https://doi.org/10.1016/j.renene.2015.04.013
7. Lagrioui A, Mahmoudi H, Bossoufi B (2011) Discrete linear predictive control of permanent
magnet synchronous motor (PMSM). J Theor Appl Inf Technol JATIT 31(1):21–28
8. Sghaier N, Ramzi T, Gdaim S, Bossoufi B, Mimouni MF (2013) Vector control for induction
motor with rotor flux oriented using Xilinx System Generator. J Theor Appl Inf Technol JATIT
54(2):259–268
9. Bossoufi B, Karim M, Ionita S, Lagrioui A (2012) Nonlinear non adaptive backstepping with
sliding-mode torque control approach for PMSM motor. J Electr Syst JES 8(2):236–248
10. Jung CF, Chen JS (2006) Water bath temperature control by a recurrent fuzzy controller and
its FPGA implementation. IEEE Trans Ind Electron 53(3):941–949
High-Level Synthesis Implementation
of Monocular SLAM on Low-Cost
Parallel Platforms

Ayoub Mamri, Mohamed Abouzahir, Mustapha Ramzi, and Mohamed Sbihi

Abstract Simultaneous Localization And Mapping (SLAM) algorithm allows the

robot to map the environment while locating it in the space. SLAM algorithm is the
more efficient and more accredited system by autonomous vehicle navigation and
robotic application in current research. However, it has not yet adopted a complete
end-to-end hardware implementation. Our work is aimed at hardware/software opti-
mization of a time-consuming and expensive functional block of a SLAM-based
autonomous navigation application, precisely monocular ORB-SLAM. Thus, the
proposed optimization is achieved through the implementation of a heterogeneous
FPGA-based embedded architecture. A comparative study is then conducted with
parallel architectures such as automotive embedded architecture TX1 and a high-
end machine. This work shows that heterogeneous FPGA-based embedded archi-
tectures are appealing embedded platforms for the heavy computational algorithm,
achieving a 16% improvement of the original, whereas high-end machines achieve
45% improvement.

Keywords SLAM · Optimization · Heterogeneous architecture · FPGA ·

Embedded system

1 Introduction

SLAM [1] is one of the ongoing research topics in the autonomous driving and robotic
area. However, SLAM is a heavy computational and complex algorithm that contains
compute-intensive must be optimized and implemented in suitable embedded archi-
tecture. The compute-intensive task is often managed by CPU, but it remains limited
according to characteristics and number of cores. Rather, GPU is usually used for this
purpose even though it enforces some limitations on accelerated algorithms. Thus,
Field Programmable Gate Array (FPGA) [2] is an attractive accelerator to bridge this

A. Mamri (B) · M. Abouzahir · M. Ramzi · M. Sbihi

Laboratory of Systems Analysis, Information Processing and Industrial Management,
Higher School of Technology, Mohamed V University of Rabat, Sale, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 399
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_37
400 A. Mamri et al.

gap, which provides low power, low cost, high speed, and significant gains. VHDL
and Verilog are the recommended hardware description languages for FPGA, except
they are often more difficult in the case of complex algorithms. The High-Level
Synthesis (HLS) [3] is then developed to convert C/C++ programming language to
the hardware description, taking into account of low usage of FPGA resources. In
the ongoing research framework, OpenCL [4] HLS based aims for targeting a new
trend of heterogeneous FPGA-based embedded architectures. Moreover, OpenCL
is a portable language framework that provides the capability to explore multiple
embedded accelerators, including FPGAs, GPUs, and DSPs. Besides, CUDA [5]
is a high-level language developed by NVIDIA to handle its specific GPU-based
heterogeneous systems.
The ultimate goal of this work is to improve performance and embed ORB-SLAM
[6, 7], Visual Odometry (VO) [8], and Visual Inertial Odometry (VIO) through effi-
cient hardware/software optimization of a time-consuming functional block. Besides,
the proposed optimization provides a parallel implementation that handles sequential
parts using high-level OpenCL synthesis for FPGA-based heterogeneous embedded
architectures and CUDA for heterogeneous NVIDIA GPU-based architectures. The
comparative study is then conducted between the original system and the proposal
with parallel architectures and languages. The reminder of this paper is structured
as follows. The related work in Sect. 2. The targeted functional block is presented
in Sect. 3. The Sect. 4 presents the OpenCL software optimization and the hardware
implementation of the proposed optimization. The heterogenous embedded hardware
is described in Sect. 5, with results. Finally, a conclusion in Sect. 6.

2 Related Work

Recent works based on the ORB-SLAM algorithm target the feature tracking process
to deal with the computational time using heterogeneous CPU-GPU architecture.
Nguyen et al. [9] developed a novel VSLAM system named HOOFR SLAM based
on a FAST detector and amended FREAK descriptor. Afterward, they adopt the
OpenMP multithreading language to implement HOOFR extractor on CPU due to the
machine learning optimization, whereas the HOOFR descriptor is ported into GPU
accelerator using OpenCL. As a result, their proposal achieved remarkable gains
and effective speedup compared to stereo ORB-SLAM2 [10]. Jincheng Li et al. [11]
presented an improved MultiCol-SLAM using embedded heterogeneous CPU-GPU
architecture. Thus, they attempted to improve and parallelize ORB feature extraction
using CUDA and then implement it on GPU accelerator. Besides, they proposed a
CPU-based multithreading pipelining method to solve CPU and GPU load imbalance.
As a result, their proposal improved the frame rate and achieved a higher speedup and
low resource utilization than the CPU implementation. There are similar works of
ORB-SLAM2 with CPU-GPU implementation in [12, 13]. However, these works are
unfinished and require a low power consumption for embedded applications, which
is guaranteed by heterogeneous CPU-FPGA architecture.
High-Level Synthesis Implementation of Monocular SLAM ... 401

Abouzahir et al. [14] attained a complete end-to-end implementation of Fast-

SLAM2.0 in embedded CPU-FPGA architecture. They adopted a parallel implemen-
tation on FPGA and GPU to optimize all block of the system except the feature extrac-
tion FAST detector [15] that ported into one-core of CPU, optimized by machine
learning. In addition, they gave a comparative study between FPGA and GPU accel-
erators, which shows a significant improvement of the SLAM algorithm using FPGA-
based heterogeneous architecture. FastSLAM2.0, however, is not preferable for large
scale indoor/outdoor environments because it suffers from the depletion of samples
and exhaustion of particle problems.
Xuesong Shi et al. [16] presented Heterogeneous Extensible Robot Open (HERO),
an appealing integrated computing platform that combined an Intel Core i5 CPU and
a high-performance scalable compute FPGA accelerator Intel Arria 10. This plat-
form is attractive for the research area that deals with compute-intensive tasks, algo-
rithm acceleration, and heterogeneous computing. Within the scope of this paper, the
authors presented a heterogeneous implementation of HectorSLAM [17] by acceler-
ating the scan matching process on their proposed platform. As a result, the evalua-
tion performed 4-times speedup than software implementation with CPU and 3-times
faster than HDL implementation with only Arria 10 SoC.
Meanwhile, the ORB-SLAM2 [10] is the purest and reliable Visual SLAM system,
as well as Visual Inertial ORB-SLAM3 [7], that operate in large and small environ-
ments and in real-time as well. However, they are a heavy computational system, and
they have not yet a hardware implementation and embedded application. The imple-
mentation in embedded architecture is notoriously difficult due to the computational
complexity and compute-intensive tasks. For this purpose, the SLAM systems get
trends toward heterogeneous implementation to benefit from powerful accelerators
such as GPU and FPGA. In contrast, this work is devoted to improving the perfor-
mance of the ORB-SLAM system and orienting it towards heterogeneous embedded
implementations based on FPGA-based heterogeneous architecture.

3 Algorithm Study

3.1 ORB-SLAM System

ORB-SLAM contains three concurrent threads; tracking, local mapping, and loop
closing. The tracking thread localizes the camera at each acquired frame and decides
to add it to the system. For the local mapping thread, the system is based on the
added keyframes to execute high-cost Bundle Adjustment (BA) [18] to reconstruct
an optimal map. At the end of the process, a loop closing thread is achieved to correct
the estimated localization based on the detected potential loops at each received
keyframe.
The proposed optimization concerns the monocular ORB SLAM2, the monoc-
ular ORB-SLAM3, VO, and VIO. The monocular ORB-SLAM2 is selected to
402 A. Mamri et al.

Table 1 ORB-SLAM2 performance (ms)

Thread Functional block Intel PC Embedded TX1
Tracking ORB extraction 19.58 62.32
Initial pose estimation 26.03 67.46
Track local map/Keyframe decision 6.42 22.15
Local New keyframe processing 16.87 42.10
mapping MapPoint culling/New points creation 103.31 297.45
Local BA 156.45 469.53
Keyframe culling 6.97 18.51
Loop Loop detection 3.80 5.67
closing Loop correction

investigate the time-consuming blocks, then evaluated to lead toward the targeted
block. Table 1, shows the evaluation by the CPU of the targeted platforms using
TUM1 dataset [19].
The evaluation of the three threads shows that MapPoint culling and Local BA
are the most time-consuming blocks that require high dependency with other blocks.
Hence, the proposed optimization aims to parallelize the tracking thread, especially
the third time-consuming block called Initial pose estimation.

3.2 Map Initialization

Map initialization is a part of the Initial pose estimation functional block (FB2) that
handles the monocular initialization of the relative camera pose between two frames
using two geometrical models (fundamental matrix F [20] for the non-planar scene
and a homography H [20] for the planar scene). However, it performs 30% of the
tracking thread and 67% of FB2. Practically, CPU carries out the two geometrical
models F and H using C++ multi-threading API. For the sake of accuracy, each
model count on the normalization method that handles 2001–2010 keypoints. Mean-
while, the normalization
 method
 (Normalize) is executed consecutively
  twice for the
reference image u ri xir , yir ∈ Fr and the current image u ic xic , yic ∈ Fc into each
model. Therefore, the main idea of this optimization is to introduce the first step to
implement the initial pose estimation functional block by accelerating the iterative
normalization method in one execution available to H and F into FPGA and GPU
accelerators. Normalize function computes at first the mean xr,c , yr,c of Nr,c corner
coordinates in both current and reference frame with the following equation:

1 i=N r,c 1 i=N r,c

x r,c = x , yr,c = y (1)
Nr,c i=0 i Nr,c i=0 i

And then, calculates the deviation according to the (x,y) coordinates:

High-Level Synthesis Implementation of Monocular SLAM ... 403

i=N  r,c  i=N  r,c 

x r,c  x − x r,c  yr,c  y − yr,c 
dev = i=0 i , dev = i=0i (2)

At the end, the following equation gives the normalized points:

r,c    
xi,n = xir,c − xr,c sxr,c , r,c
yi,n = yir,c − yr,c s r,c
y (3)

With:
1 1
x =
sr,c , y =
sr,c (4)
xr,c
dev yr,c
dev

The normalized matrix is:

⎛ ⎞
sxr,c 0 −xr,c sxr,c
Tr,c = ⎝ 0 s r,c r,c r,c ⎠
y −y s y (5)
0 0 1

4 Heterogeneous Implementation

This section describes the accelerated Normalize version step by step. The descrip-
tion is based on OpenCL for FPGA, whereas the CUDA version could be inferred
easily for GPU. Normalize function contains equations that are parallel in nature,
while others are not obvious to parallelize and require a special modification of the
algorithm to adjust to the FPGA kernel. To bridge this gap, we propose OpenCL
optimization for Normalize in order to parallelize the sequential parts and efficient
implementation of FPGA-based heterogeneous architecture with efficient usage of
FPGA resources.

4.1 Accelerated Normalize Version

For the sake of the proposal clarity, a good understanding of the software and hard-
ware is required. Algorithm 1, provides insight into the software, where the host
c
adjusts the input/output by resizing the current keypoints kpsx,y and the reference
r
keypoints kpsx,y to the closest required value by the architecture (i.e., 2010 to 2048)
in line 3, and then merges them in one single vector kps in line 4. It has to be noted
r c
that kpsx,y and kpsx,y are vectors combined of their coordinates (x, y). Instead of the
CUDA optimization, the coordinates of each keypoints are buffered as arguments to
the GPU without fusion.
404 A. Mamri et al.

The host allocates kps as inout using clCreateBuffer with

CL_MEM_READ_WRITE flag, and then buffers it into the constant memory of
the FPGA using clEnqueueWriteBuffer Afterward, the host launches
Normalize.cl kernel usingclEnqueueNDRangeKernel. The kernel adopts
NDRange kernel optimizations and a special approach to deal with sequential
computation, explaining the execution of line 5 to 15 in the next section.

4.2 OpenCL Hardware Implementation

The HLS creates more efficient hardware structures for the kernel and schedules
the buffer kps in the constant memory for a reason to process each kps coordinate
in their reserved work group. As depicted in Fig. 1, the HLS provides a hardware
description of the proposal using 4 work groups with (6) and 512 work items on each
work group with (7).

kps size
nbr work group = (6)
one coordonate of kps
kps size
nbr work item = (7)
nbr work group × nbr of vectorization

Thus, the data flow on each work group based on:

High-Level Synthesis Implementation of Monocular SLAM ... 405

Fig. 1 The proposed OpenCL software and the data flow on each workgroup with kps (blue),
localSum (gray) and T (green)

• NDRange kernel optimizations are a set of optimizations offered by Altera

SDK for OpenCL [21] dedicated to FPGA kernel to improve data processing
and memory access efficiently. Thus, the host buffer kps to the FPGA in on-
chip embedded caches load-store units (LSUs) that have low latency than global
memory. The vectorization (VEC) is then applied to coalesce memory by vector-
izing kps to a single float4 vector for optimal widening of load/stores. After-
ward, the vectorized kps will be preloaded in the local memory (LM) to apply the
SumReduction approach and reserve the privacy of each keypoints in their work
group. In the end, kps and T are buffered to the host, while T is rebuilt using the
Eigen library to infer the normalized matrices Tr and Tc .
406 A. Mamri et al.

• SumReduction approach is a strategy adopted to deal with the sequential parts

of (1) and (2) that require reloading of kps to LM., where each work item in the
workgroup accumulates two particles separated by an offset of work_group_size
divided by two. This operation is then repeated, dividing the offset by two at
each iteration until it becomes less than 1. In the end, the first thread gives the
result. For the sake of the save operation, the CLK_LOCAL_MEM_FENCE barrier
synchronizes all work-items of each workgroup in order to manage and control
memory accesses accurately.

5 Architectures Description and Evaluation

The implementation of the proposed kernel is performed on three heterogeneous

architectures in order to evaluate the kernel performance and give a comparative
study.

5.1 Architectures Description

DE1-SoC board, the embedded FPGA-based heterogeneous architecture, see

Fig. 2(a). Rather, it is low-power embedded hardware produced by the Altera System-
on-Chip (SoC) FPGA. The board comprises an ARM Cortex A9 MPCore dual-core
processor clocked at 800 MHz, 1 GB of DDR3 SDRAM, a single hard memory
controller, and a large set of peripherals. The chip also incorporates a Cyclone V
FPGA that provides quad 50 MHz clock sources from the clock generator and 64
MB SDRAM. In addition, the chip contains several embedded resources, including
4,450 Kbits of embedded memory, 87 DSP blocks, 85K logic elements, a single hard
memory controller, and 6 fractional PLLs.
NVIDIA Jetson TX1, the automotive embedded GPU-based heterogeneous
architecture, see Fig. 2(b), is dedicated to low power embedded application. The
board is equipped with ARM Cortex A57 quad-core operates at 1.9 GHz with 4 GB
of memory. The chip also integrates a powerful embedded GPGPU, which contains
256 CUDA cores (Maxwell architecture) operating at 1 GHz, 4 GB of global memory,
and has around 13 MHz of memory clock rate.
Intel PC offers a Core i5 4300U CPU with 8 cores running at 1.9−2.5 GHz
and provides 8 GB DDR3 memory. Furthermore, it integrates a high-end NVIDIA
GeForce 920 MX GPU that contains 256 CUDA core (Maxwell architecture) operate
at 980 MHz, 2 GB of global memory and 2505 MHz of the memory clock rate.
High-Level Synthesis Implementation of Monocular SLAM ... 407

5.2 Results and Discussion

The proposed implementation is achieved after several attempts to develop a

supported kernel by DE1-SoC board, taking into account several criteria such as
resource utilization, throughput, and memory latency. In Table 2, the CUDA version,
which is considered as a naive OpenCL kernel for FPGA with simple modifications,
was not supported by the DE1-SoC board due to the resource limitation at the level of
the FPGA. In contrast, the optimized kernel holds low resource usage. The optimized
kernel usage is provided by the HLS report, which uses 60% of logic elements, 28%
of registers, 71% of memory blocks, and 9% of DSP blocks.
In Table 3, the benchmarking study of the proposal is performed on the targeted
platforms. Although the CUDA kernel and platform specifications were almost
similar for NVIDIA platforms, the proposal achieved 1,8-times faster than the orig-
inal [10] with the NVIDIA GeForce 920 MX GPU, whereas TX1 GPU could not
tackle the processing time. The difference lies in the frequent and huge preloading
of data, besides the heavy memory clock rate for TX1 GPU.
As is obvious for embedded boards, DE1-SoC FPGA achieved almost an
appealing result than TX1 GPU about 1,25 speedup, moreover, 1,2 faster than its
CPU. The reason is due to the memory conception and the applied optimizations.
On the one hand, the applied optimization for FPGA as vectorization increases the

(a ) (b )

Fig. 2 Heterogeneous embedded hardware

Table 2 Resource utilization

Naive kernel Optimized
of FPGA
kernel
Utilization (%) Logic 92 60
elements
Registers 42 28
Memory 159 71
blocks
DSP blocks 29 9
408 A. Mamri et al.

Table 3 Kernel performance (ms).

Platform High-end machine Embedded
Intel PC TX1 DE1-SoC
CPU GPU CPU GPU FPGA
Original using C++ multithreading 0.133 - 0.365 - -
[10]
Proposal OpenCL/CUDA - 0.073 - 0.384 0.308

throughput and reduces memory accesses. On the other hand, the memory concep-
tion of FPGA provides LSU embedded cache that maintains high bandwidth and
short latency. Instead of GPU, that has a heavy memory clock rate, which requires a
Direct Memory Access (DMA) system to overcome the memory bank conflict.

6 Conclusion

This work offers an efficient strategy to tackle the computational time of the compute-
intensive parts using heterogeneous architectures. Furthermore, the main idea is to
present FPGA-based heterogeneous architecture capability to handle the computa-
tional complexity and improve the performance of the ORB-SLAM system, instead
of using only GPU-based heterogeneous architecture. Thus, our further work will
adopt a high-performance scalable compute FPGA accelerator as ARRIA 10 to attain
a complete end-to-end embedded SLAM system.

References

1. Thrun S (2007) Simultaneous localization and mapping. Springer, Berlin, Heidelberg

2. National instrument, introduction to fpga technology. https://www.ni.com/en-lb/innovations/
white-papers/08/fpga-fundamentals.html
3. Cong J, Liu B, Neuendorffer S, Noguera J, Vissers K, Zhang Z (2011) High-level synthesis
for FPGAs: From prototyping to deployment. IEEE Trans Comput-Aided Des Integr Circuits
Syst 30(4):473–491. https://doi.org/10.1109/TCAD.2011.2110592
4. Xu JY (2008) OpenCL – the open standard for parallel programming of heterogeneous systems
5. Li H, Yu D, Kumar A, Tu YC (2014) Performance modeling in CUDA streams—a means for
high-throughput data processing. In 2014 IEEE international conference on big data (big data),
pp 301–310
6. Mur-Artal R, Montiel JMM, Tardos JD (2015) ORB-SLAM: a versatile and accurate monocular
SLAM system. IEEE Trans Robot 31(5):1147–1163
7. Campos C, Elvira R, Rodríguez JJG, Montiel JM, Tardós JD (2020) ORB-SLAM3: an accurate
open-source library for visual, visual-inertial and multi-map SLAM. arXiv preprint arXiv:2007.
11898
8. Scaramuzza D, Fraundorfer F (2011) Visual odometry [tutorial]. IEEE Robot Autom Mag
18(4):80–92
High-Level Synthesis Implementation of Monocular SLAM ... 409

9. Nguyen D-D, Elouardi A, Florez SAR, Bouaziz S (2018) HOOFR SLAM system: an
embedded vision SLAM algorithm and its hardware-software mapping-based intelligent
vehicles applications. IEEE Trans Intell Trans Syst 20(11):4103–4118
10. Mur-Artal R, Tardos JD (2017) ORB-SLAM2: An open-source SLAM system for monocular,
stereo, and rgb-d cameras. IEEE Transac Robotics 33(5):1255–1262. https://doi.org/10.1109/
TRO.2017.2705103
11. Li J, Deng G, Zhang W, Zhang C, Wang F, Liu Y (2020) Realization of CUDA-based real-time
multi-camera visual SLAM in embedded systems. J Real-Time Image Process 17(3):713–727.
https://doi.org/10.1007/s11554-019-00924-4
12. Aldegheri S, Bombieri N, Bloisi DD, et al (2019) Data flow orb-slam for real-time performance
on embedded gpu boards. In 2019 IEEE/RSJ International Conference on Intelligent Robots
and Systems (IROS), pp 5370–5375. IEEE
13. Bourque D (2017) Cuda-accelerated orb-slam for uavs
14. Abouzahir M, Elouardi A, Latif R, Bouaziz S, Tajer A (2018) Embedding SLAM algorithms:
has it come of age? Robot Auton Syst 100:14–26
15. Rosten E, Porter R, Drummond T (2010) Faster and better: A machine learning approach to
corner detection. IEEE Trans Pattern Anal Mach Intell 32(1):105–119. https://doi.org/10.1109/
TPAMI.2008.275
16. Shi X, Cao L, Wang D, Liu L, You G, Liu S, Wang C (2018) HERO: accelerating autonomous
robotic tasks with FPGA. In 2018 IEEE/RSJ International Conference on Intelligent Robots
and Systems, pp 7766–7772
17. Kohlbrecher S, von Stryk O, Meyer J, Klingauf U (2011) A flexible and scalable SLAM system
with full 3D motion estimation. In 2011 IEEE International Symposium on Safety, Security,
and Rescue Robotics (SSRR), pp 155–160
18. Triggs B, McLauchlan PF, Hartley RI, Fitzgibbon AW (2000) Bundle adjustment a modern
synthesis. In Vision Algorithms: Theory and Practice. New York, NY, USA, pp 298–372.
Springer
19. TUM1 dataset, https://vision.in.tum.de/data/datasets/rgbd-dataset/download
20. Hartley R, Zisserman A (2003) Multiple view geometry in computer vision. 2nd ed., Cambridge,
U.K.: Cambridge Univ. Press
21. Altera sdk for OpenCL, Best Practices Guide. https://www.intel.com/content/dam/www/pro
grammable/us/en/pdfs/literature/hb/opencl-sdk/aocl-best-practices-guide.pdf
Improving Multispectral Image
Processing for Real-Time Agricultural
Indices Tracking Using an Embedded
System

Amine Saddik, Rachid Latif, and Abdelhafid El Ouardi

Abstract Recently, embedded systems have been used in precision agriculture for
application control, tracking, and agricultural field monitoring. The integration of
embedded systems in precision agriculture will help much better in making deci-
sion. This needs the study of such systems in terms of complex algorithms processing
and results from consistency. This study requires software optimizations and hard-
ware requirements to speed up processing times while keeping the results reli-
able. This work presents the evaluation of the most known vital indices in agricul-
ture precision algorithms on a heterogeneous embedded system using the OpenCL
parallel programming language. The algorithm evaluated in this work is dedicated
to agricultural field tracking.
Two implementations are compared: the OpenMP optimization on a multi-CPU
architecture and the OpenCL on a CPU-GPU heterogeneous architecture. In this last
case, results show improvements of 45x in speedup, enhancing image processing-
based agriculture field tracking.

Keywords Embedded system · Precision agriculture · Heterogeneous systems ·

CPU-GPU · Vital indices · OpenCL · OpenMP

1 Introduction

The evolution of precision agriculture requires integrating embedded systems to help

users of different technologies make rapid and correct decisions. The progress of
embedded systems in the case of algorithms processing showed its proofs in various
systems and tools that can be used in agriculture applications [1]. Precision farming
applications are based on several instruments. Among these tools, we find soil robots,

A. Saddik (B) · R. Latif

LISTI, ENSA Ibn Zohr University Agadir, 80000 Agadir, Morocco
e-mail: [email protected]
A. El Ouardi
Université Paris-Saclay, ENS Paris-Saclay, CNRS, Systèmes et Applications des Technologies de
l‘Information et de l‘Energie, 91405 Orsay, France

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 411
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_38
412 A. Saddik et al.

satellites, Unmanned Aerial Vehicles (UAV), or Unmanned Ground Vehicles (UGV)

[2–4]. Before the appearance of these tools, agriculture monitoring was based on
The Internet of Things (IoT). The problem of using techniques based on IoT is the
high cost of sensors used in precision farming, especially when we are talking about
large agricultural soils. The permanent monitoring of sensors used in precision agri-
culture generates additional costs and requires other parameters that play a very
important role in soil monitoring. With the approach based on robots, large agricul-
tural fields’ monitoring time is reduced, which is not the case with the method based
on the Wireless Sensor Network (WSN) [5]. The use of embedded systems in agri-
cultural applications that need real-time processing requires studying the hardware-
software co-design methodology. This approach will allow us to implement different
algorithms on specific architectures [6].
On the other hand, we recently find the strong use of different learning techniques,
namely deep learning (DL) and machine learning (ML). We can find the use of these
techniques in various precision agriculture applications, for example, plant counting,
weed detection, and 3D field monitoring [7–9]. But the problem with using ML or
DL is the real-time constraint. The algorithms require data training, which increases
the processing times [10]. Agricultural field mapping is one of the most commonly
used applications in precision agriculture. The work of S. Li et al., 2020, presents a
real-time Crop Row recognition technique based on a System-on-a-Chip integrating
a Field Programmable Gate Array. The evaluation of the algorithm proposed in
this work was based on Xilinx Zynq UltraScale + MPSoC ZCU102 SoC-FPGA.
The database contains images with a resolution of 1920 × 1080 pixels. The high
accuracy of the algorithm (89%) and the low execution time (210 ms) showed that the
technique proposed in this paper had shown satisfactory results in terms of accuracy
and execution time compared to other approaches of state-of-the-art [11]. M. Imperoli
et al., 2018 present a technique based on the combination of different data in order to
build a 3D agricultural field map. The evaluation of the proposed system showed that
an improvement had been achieved compared to the GPS-only approach. The authors
showed that the accuracy increased from 37% to 76%. The proposed algorithm was
evaluated on a Core-i7 @ 2.7 GHz laptop [12]. I. Vidović et al., 2016, proposed a
new technique for crop line detection in agricultural fields. The algorithm proposed
in this paper is based on three steps: the first step consists of detecting vegetation in
crop lines. The second step aims at the detection of regular patterns and, finally, the
detection of optimal sequences. The proposed approach is based on 281 images; it
has been compared to the Hough transform [13].
This work presents an algorithm implementation on a heterogeneous CPU-GPU
system to track the agricultural fields based on multispectral images. The embedded
system used in this work is the Odroid XU4 board. We exploit the performance of
CPU and GPU to evaluate software optimizations based on parallel programming
languages: OpenCL and OpenMP. In this evaluation, we used two types of images,
one collected with a UAV and the other acquired with a ground robot. The evaluation
results showed a processing time of 47 ms with OpenMP optimization while this
processing time decreases to 1.107 ms with OpenCL parallel implementation.
Improving Multispectral Image Processing ... 413

2 Embedded Systems in Precision Agriculture

The monitoring of agricultural field parameters is necessary to obtain a good quality

of agricultural products. Generally, the algorithms developed for indexes monitoring
are based on 3 phases. The first one is the data acquisition using different tools either
by embedded cameras on UAVs, ground robots, or satellite images. After image
acquisition, it is necessary to verify if the image bands require a separation. Phase 2
consists of indexes calculation according to the application. For example, if we want
to monitor vegetation, we have to focus on the R, G, and NIR bands to calculate the
different vegetation indexes. On the other side, we find the humidity and water indexes
that require the SWIR, NIR, and G bands [14]. After the extraction of the different
indices, it is important to determine the application used later; this identification
will allow us to make decisions. P. Horstrand et al. 2019 have elaborated on real-
time surveillance of agricultural fields in this context. The work presents a real-time
monitoring system based on the calculation of the NDVI, MSAVI, and MCARI
(Modified Chlorophyll Absorption in Reflectance Index) indexes. The authors used
a Matrix-600 drone. A CAN bus is implemented to manage different sensors data
and a heterogeneous embedded CPU-GPU system (Jetson TK1 with an NVIDIA
GK20, 326 GFLOPS, and ARM CPU Cortex-A15 quadra-core @2.32 GHz) for
index calculation and drone control. A hyperspectral camera (FX10) with 224 bands
and an RGB camera were used [15]. To detect weeds, I. Sa et al., 2018, based their
studies on conventional CNN neural networks. The hardware used is a Jetson TX2
GPU and Titan X @ 1.3 GHz boards [4]. In the case of using the satellite-based
approach, we find the work of C. Coakley et al. 2019, based on the vegetation index
to monitor agricultural fields that are dedicated to Khmer Rouge production to control
the irrigation system [16]. For greenhouse farms, the work of M. Dongdong et al.
2019, presented a new greenhouse management system to increase plant yield. This
system is based on software and was evaluated on a LENOVO desktop with 16 GB
RAM and an Intel Xeon E1270 CPU @ 3.70 Ghz. Then, Matlab R2016a software was
used to process the collected images. [17]. Databases in precision agriculture are very
important tools to evaluate algorithms. A database that contains a large number of
images with different resolutions is recommended in order to evaluate the algorithm
functionality and the processing times. Generally, 3 types of databases exist. The first
one is the satellite images for monitoring large agricultural fields with low accuracy.
The second one is the medium agricultural field images based on Unmanned Aerial
Vehicles. The third type is small agricultural fields and greenhouses images given by
soil robots [18–20].

3 Evaluation and Results

The evaluation of the algorithm was based on the OpenCL parallel programming
language. OpenCL represents a combination of the world of GPU, CPU, and FPGA.
414 A. Saddik et al.

It is dedicated to the programming of heterogeneous systems such as CPU-GPU,

CPU-FPGA. This will allow us to exploit the resources of different systems (CPU,
GPU, FPGA) in order to reduce the processing time. This language’s use requires
an appropriate study of various blocks of the proposed algorithms to choose which
one(s) can be parallelized. The difference between the programming language of
OpenCL and CUDA is very clear on the choice of architecture; for example, the
CUDA language is dedicated only to the Nvidia CPU-GPU architecture, limiting the
use of this language. OpenCL is for other architectures. This evaluation’s embedded
system is based on the Odroid XU4 board with 2 CPUs and MALI GPUs. Table 1
shows the characteristics of the board used, and Fig. 1 shows it’s the architecture.
The algorithm proposed in this evaluation consists of three blocks; the first block
aims to create a sequence of images based on the database used. In this context, we
evaluated two databases, one collected by a ground robot and the other based on
a UAV. The first database offers images collected with the TetraCam camera. This
camera includes an integrated Incident Light Sensor (ILS) measurement system for
downlink radiation that can be used to generate accurate reflectance data for the NIR,
R, and G bands. The resolution given by this camera is 1296 × 966 pixels producing
one frame per second [22]. A multispectral Parrot Sequoia camera collects the second
database. This camera offers separate images with different spectral bands. This

Table 1 Odroid XU4

Type XU4
specifications
Processor Exynos 5422 big. LITTLE
Cores Octa
CPU ARM A15/A7
GPU Advanced Mali
Language C/C++ /OpenCL/OpenGL
Frequency 2 Ghz/1.4 Ghz
Dimensions 82 × 58 × 22 mm

Fig. 1 Architecture of Odroid XU4 [21]

Improving Multispectral Image Processing ... 415

camera’s strong point is its low weight (72 g) compared to the RedEdge camera (250
g). This camera offers four separate bands and one RGB image with a resolution of
4608 × 3456 pixels and the other bands with a resolution of 1280 × 960 pixels [23].
The second block is dedicated to processing images emitted by the camera. This
processing is done in real-time to follow the number of time-laps delivered each
second. This real-time processing is based on the calculation of vital indices, namely
the normalized difference vegetation index (NDVI) and the normalized difference
water index (NDWI). The third block is designed to apply the threshold operation
to classify agricultural fields and the storage of classified data. The index evaluation
was based on a heterogeneous CPU-GPU system, using both OpenCL and OpenMP
languages. Figure 2 shows the algorithm used in this work.
The complexity of this algorithm is of O = k × i × j with K the number of images
desired i and j is the size of the image. The processing of the algorithm is based on
the calculation of NDVI and NDWI as mentioned in the Eq. 1 and 2.

N I R − RE D
N DV I = (1)
N I R + RE D

Fig. 2 Algorithm overview

Fig. 3 Algorithm overview

416 A. Saddik et al.

G − NIR
N DV I = (2)
NIR +G

The evaluation results showed a great benefit compared to the published work
[14]. Our approach showed that we can achieve improvement with 90x of speedup.
This will allow us to monitor transfer of images collected by the cameras used for
the database. Table 1 shows the processing time obtained by different tools. The time
obtained with OpenCL has shown a great optimization of the processing time that
keeps cameras’ acquisition frequency. This optimization will also allow adding other
blocks for image pre-processing and compression afterward. The architecture of the
implementation on the XU4 board is shown in Fig. 3. After evaluating each block’s
algorithm, we concluded that the second block consumes more time than the other
block. For this reason, we opted to send the second block to the GPU for acceleration,
as shown in Fig. 3.
The following figure shows the NDVI and NDWI evaluation results. The left
image shows the NDWI values using an image collected by a ground robot, the
bluest regions are close to the value 1. Similarly, for the image on the right using a
UAV, the reddest regions have a higher NDVI index.
Figure 4 shows two images resulting from the NDVI and NDWI based assessment
with a data matrix, which helps us have a different map containing the index values.
This map allows farmers to determine which areas lack water or vegetation and make
decisions manually. In case we need an autonomous system that automatically makes
the decision, we need a thresholding operation that will separate the different regions.
We can perform thresholding techniques as performed in Otsu [24], but the use of
these techniques has increased the algorithm complexity as well as the processing
time. For this reason, we used a simple binary thresholding operation. Figure 5 shows
the NDVI and NDWI evaluation results. The middle image shows the NDWI values

Table 2 Processing times

Tools Desktop XU4 Desktop XU4 (OpenMP) XU4
(OpenMP) (C/C + +) (OpenCL) (OpenCL)
Total time (ms) 15 91 [14] 0,39 47 [26] 1,10

Fig. 4 Evaluation result of NDVI and NDWI indices

Improving Multispectral Image Processing ... 417

Fig. 5 NDVI using thresholding operation with SRed = 0.5

using an image collected by a ground robot; the bluest regions are close to the value
1. Similarly, for the image on the right using a UAV, the reddest regions have a higher
NDVI index.

4 Discussion

The thresholding operation is very important in our case for the classification of
agricultural fields. In the image on the left of Fig. 5 we used a thresholding operation
with an index of NDVI = 0.5. With this value, we can select only regions in white
with a higher value of 0.5. These regions will be in black on the other case (NDVI
lower than 0.5). The right image is based on a threshold value of 0.3, which shows
more regions with an index of more than 0.3. In the middle image, we used the red
color to select the desired regions. The use of colors in the built maps allows us to
have a single map with different NDVI or NDWI values. The figure shows the NDVI
thresholding operation.
After the operation is proceeded by the third block, it is necessary to generate
a map with different colors. In Fig. 6 the regions that have a high index close to

Fig. 6 NDVI using

thresholding operation
418 A. Saddik et al.

Fig. 7 Evaluation of 80
images with different tools

Fig. 8 Evaluation based

OpenCL on XU4

1 will be colored red. Otherwise, the regions that have an index close to 0 will be
yellow. This operation allows the classification based on the generated map. In this
context, we can find software that allows us to do the classification. In this context,
authors in [25] used the eCognition software to classify satellite images. The limit of
these techniques is their processing time that requires offline processing and hence
causes problems to applications that need fast processing in order to give decisions
in real-time.
The logarithmic curve in Fig. 7 shows the processing time for different imple-
mentations. The green curve presents the processing time based on XU4 and C/C++
language. In this case, we have a time that varies between 80 ms and 100 ms with an
average of 91 ms, the blue curve gives the C/C++ implementation on the desktop. The
curve in purple and red show the processing times using OpenCL and OpenMP imple-
mentations. From this figure, we can conclude that heterogeneous systems, based on
CPU-GPU, have a very important impact on the processing times. In this case, an
advanced study of the Hardware-Software Co-Design approach is recommended
to propose more optimized implementations in terms of memory consumption and
execution times.
Figure 8 shows the processing time using an OpenCL implementation on the XU4
architecture. Using OpenCL allows us to exploit the GPU part (Mali-T628 MP6 @
600 MHz) of the XU4 and the CPU cores (Quadra-core ARM A57 @ 1.9 Ghz,
Quadra-core ARM A7@ 1.2 Ghz). The variation in execution time depends on the
amount of information in the images. Generally, the time varies between 1.5 ms and
0.85 ms, with an average of 1.1 ms. This meets the constraint of real-time processing
in embedded systems.
The processing time obtained shows the heterogeneous system’s strength, which
can accelerate algorithms in the GPU part. The heterogeneous programming tool
allows us to use both features of the XU4 board, despite the use of OpenMP, which
Improving Multispectral Image Processing ... 419

allows using both processors of the CPU part of the XU4 board, but it remains limited.
This limitation is shown in Table 2, which presents a processing time of 47 ms. This
time obtained is minimized compared to the time proposed in C/C++; it only gives
21 fps processing compared to C/C++, which gives just 10 fps. On the other hand,
the use of OpenCL in the GPU part of the board benefits ×45. this optimization, we
denounce an idea on the use of this type of system.

5 Conclusion

The use of embedded systems in precision agriculture has become a strong point for
decisional systems. This becomes possible thanks to the new generation of embedded
architectures’ performances, either homogeneous (CPU, GPU, FPGA) or heteroge-
neous (CPU-GPU, CPU-FPGA). In this work, we have presented an implementation
of our algorithm for tracking different vital indices on a homogeneous and hetero-
geneous architecture based embedded system. The results show that the use of the
OpenCL-based CPU-GPU implementation has a benefit of 90x of speedup than that
given in [14], and 50x of speedup tan the OpenMP implementation, which reduces
the processing time for real-time decisions. As future work, we aim to validate the
algorithm implementation on the Odroid XU4 platform using the HIL (Hardware In
the Loop) evaluation and a dataset acquired with a UAV operating in a referenced
environment.

Acknowledgment We owe a debt of gratitude to the National Centre for Scientific and Technical
Research of Morocco (CNRST) for their financial support and for their supervision (grant number:
19 UIZ2020).

References

1. Berni JAJ, Zarco-Tejada PJ, Suarez L, Fereres E (2009) Thermal and narrowband multispectral
remote sensing for vegetation monitoring from an unmanned aerial vehicle. IEEE Trans Geosci
Remote Sens 47(3):722–738
2. Baek J-Y, Jo Y-H, Kim W, Lee J-S, Jung D, Kim D-W, Nam J (2019) A new algorithm to
estimate chlorophyll-a concentrations in turbid yellow sea water using a multispectral sensor
in a low-altitude remote sensing system. Remote Sens 11:2257
3. Guan K et al (2018) Mapping paddy rice area and yields over thai binh province in viet nam from
MODIS, landsat, and ALOS-2/PALSAR-2. IEEE J Selected Topics Appl Earth Observations
Remote Sens 11(7):2238–2252
4. Sa I et al (2018) weedNet: dense semantic weed classification using multispectral images and
MAV for smart farming. IEEE Robot Autom Lett 3(1):588–595
5. Hu Z et al (2019) Application of non-orthogonal multiple access in wireless sensor networks
for smart agriculture. IEEE Access 7:87582–87592
6. Maxime M (2019) High level approach for accelerating algorithms on CPU/GPU/FPGA hetero-
geneous architectures. Application to the qualification of radars and electromagnetic listening
systems. Parallel, distributed and shared computing. Paris-Saclay University, French
420 A. Saddik et al.

7. Liu X et al (2019) Monocular camera based fruit counting and mapping with semantic data
association. IEEE Robot Autom Lett 4(3):2296–2303
8. Riekert M, Klein A, Adrion F, Hoffmann C, Gallmann E (2020) automatically detecting
pig position and posture by 2D camera imaging and deep learning. Comput Electr Agric
174:105391. https://doi.org/10.1016/j.compag.2020.105391
9. Aichen W, Wen Z, Xinhua W (2019) A review on weed detection using ground-based machine
vision and image processing techniques. Comput Electr Agric 158:226–240 ISSN 0168–1699
10. Andreas K, Peta-Boldú FX (2018) Deep learning in agriculture: a survey, Comput Electr Agric
147:70–90 ISSN 0168–1699
11. Li S, Zhang Z, Du F, He Y (2020) A new automatic real-time crop row recognition based on
SoC-FPGA. IEEE Access 8:37440–37452
12. Imperoli M, Potena C, Nardi D, Grisetti G, Pretto A (2018) An effective multi-cue positioning
system for agricultural robotics. IEEE Robot Autom Lett 3(4):3685–3692
13. Vidović I, Cupec R, Hocenski Ž (2016) Crop row detection by global energy minimization.
Pattern Recogn 55:68–86 ISSN 0031–3203
14. Latif R, Saddik A, Elouardi A (2019) Evaluation of agricultural precision algorithms on
UAV images. In 2019 International Conference of Computer Science and Renewable Energies
(ICCSRE), Agadir, Morocco, pp 1-4
15. Horstrand P, Guerra R, Rodríguez A, Díaz M, López S, López JF (2019) A UAV platform
based on a hyperspectral sensor for image capturing and on-board processing. IEEE Access
7:66919–66938
16. Coakley C, Munro-Stasiuk M, Tyner JA, Kimsroy S, Chhay C, Rice S (2019) Extracting khmer
rouge irrigation networks from pre-landsat 4 satellite imagery using vegetation indices. Remote
Sens 11:2397
17. Ma D, Carpenter N, Amatya S, Maki H, Wang L, Zhang L, Neeno S, Tuinstra MR, Jin
J (2019) Removal of greenhouse microclimate heterogeneity with conveyor system for
indoor phenotyping. Comput Electr Agric 166:104979. https://doi.org/10.1016/j.compag.2019.
104979
18. Yang C (2020) Remote sensing and precision agriculture technologies for crop disease detection
and management with a practical application example. Engineering 6(5):528–532
19. Radoglou-Grammatikis P, Sarigiannidis P, Lagkas T, Moscholios I (2020) A compilation of
UAV applications for precision agriculture. Comput Netw 172:107148. https://doi.org/10.1016/
j.comnet.2020.107148
20. Tian H, Wang T, Liu Y, Qiao X, Li Y (2020) Computer vision technology in agricultural
automation —a review. Inf Process Agric 7(1):1–19
21. Roy R (2018) «ODROID-XU4 Manual». ODROID Magazine. https://magazine.odroid.com/
odroid-xu4
22. https://www.sensefly.com/education/datasets/?dataset=5632&industries%5B%5D=2
23. https://www.ipb.uni-bonn.de/data/uav-sugarbeets-2015-16/
24. Xue J, Titterington DM (2011) t -Tests, F -tests and otsu’s methods for image thresholding.
IEEE Trans Image Process 20(8):2392–2396
25. Blaschke T (2010) Object based image analysis for remote sensing. ISPRS J Photogrammetry
Remote Sens 65(1):2–16
26. Saddik A, Latif R, Elhoseny M, El Ouardi A (2021) Real-time evaluation of different indexes
in precision agriculture using a heterogeneous embedded system. Sustain Comput Inf Syst
30:100506. https://doi.org/10.1016/j.suscom.2020.100506
Design and Implementation of an Open
Source and Low-Cost Nanosatellite
Platform

Ahmed Hanafi , Aziz Derouich , Mohammed Karim ,

and Assiya Lemmassi

Abstract In recent years, nanosatellites and CubeSats market have grown consider-
ably, making it possible to carry out, at low cost and in a short time, a wide range of
spacecraft missions (industrial or scientific) in Low Earth Orbit (LEO). In this paper,
we propose the use of the open-source movement to space technology by designing a
low-cost hardware and software platform for LEO nanosatellite. The hardware design
is based upon open source and flexible Arduino technology. The idea is to propose a
clear, simple, ease of programming, rapid prototyping, and publicly available hard-
ware platform. However, the standard Arduino API is restricted to the capabilities
found on a core set of library function, lacking support for multithreaded programs or
specification of real time requirements. So, to take maximum advantage of the hard-
ware platform performance, with a more memory-efficient programming, we have
implemented a FreeRTOS based software platform. The development environment
used was Atmel Studio and the AVR toolchain, which made it possible to directly
program the Arduino AVR Atmega microcontroller in pure C language.

Keywords CubeSat · Opensource · Nanosatellite · Arduino · Platform · FreeRTOS

A. Hanafi (B) · A. Derouich · A. Lemmassi

EST, Sidi Mohamed Ben Abdellah University, Fes, Morocco
e-mail: [email protected]
A. Derouich
e-mail: [email protected]
A. Lemmassi
e-mail: [email protected]
M. Karim
FSDM, Sidi Mohamed Ben Abdellah University, Fes, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 421
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_39
422 A. Hanafi et al.

1 Introduction

Triggered by the success of the CubeSat standard, the past ten years have seen the
nanosatellite segment grow by a factor of 10x, from as few as 20 satellites in 2011
to nearly 200 in 2019, and it is estimated that 1800 to 2400 nano/microsatellites
will need to be launched over the next 5 years [1]. Most of the recent CubeSats
were launched by commercial operator. However, the academic world takes a largest
share among CubeSat developer [2]. From this, it is evident that academia has a
strong impact on the future of the space industry by enabling the CubeSat program
not only to be an ideal tool for teaching hands-on space technology, but also a great
chance to open access to space technology to a much wider audience, including
students from emerging and developing countries.
The work presented in this paper is the start of an open-source initiative for
CubeSat technologies with the intention to encourage and support Moroccan graduate
students to design, build, test, and operate satellites within their academic curriculum.
The motivation of this initiative comes from the at times frustrating experience of
having built the Masat1 University 1U CubeSat in an academic environment. Our
perception was that collaboration within the CubeSat community is at a minimum,
with the most projects that show design and test results but omit all specific details.
As a result, any new university team deciding to start a CubeSat project is facing the
challenge of starting their development from scratch or buying expensive CubeSat
products on the market.
In the following, we first explain in Sect. 2 the benefits of an “open-source design”.
Then, Sect. 3 describes the adopted modular architecture for the hardware platform
while Sect. 4 present the main features of the detailed design and implementation of
the real-time software platform. Finally, Sects. 5 and 6 outline the main results of
the software testing and conclusions with future work to improve and complete the
proposed open-source architecture.

2 Open-Source Design Approach

An open-source design refers to a combination of open-source software and open-

source hardware. According to the Open-Source Initiative (OSI), recognized author-
ities for open-source software regulation, Open-source software is software that can
be freely used, changed, and shared (in modified or unmodified form) by anyone [3].
However, the open-source hardware is a relatively new concept. The Open-Source
Hardware Association (OSHWA), founded in 2012, provides the following defini-
tion [4]: Open-source hardware is a physical item whose design and source files have
been openly shared and can be freely studied, copied, or modified.
The use of an open-source model in the design of a nanosatellite platform will
allow designers to:
Design and Implementation of an Open Source and Low-Cost Nanosatellite Platform 423

Fig. 1 Nanosatellite platform design approach

• Improve reliability with freely available source code/design for inspection. Design
bugs and sources of previous CubeSat mission’s failure could be quickly detected,
and users can even suggest solutions to identified problems.
• Customize the open-source solutions to the specifics of a CubeSat mission since
it is easy to modify the shared software and hardware design files as desired.
• potentially reduce costs and design times, as well as the human and technological
resources required to successfully carry out a CubeSat mission.
Our initiative aims to provide an open-source reference hardware and software
platform for generic CubeSat mission. Partially based on the guidance from the
NASA System Engineering Manual [5], a top-down design approach was adopted to
translate mission objectives and requirements into subsystems, hardware compo-
nents, interfaces, and software tasks. Figure 1 summarizes the platform design
approach which is divided into four interdependent, highly iterative, and recursive
processes, resulting in a validated set of requirements and a validated design solution
that satisfies a set of stakeholder expectations.

3 Open-Source Hardware Platform Design

3.1 Modularity of the Nanosatellite Platform

As a reference framework for our work, we partially used the ECSS (European Coop-
eration on Space Standardization) standards, which freely provide [6] the necessary
reference documents and tools for the management of the design and its system
engineering. For example, in the analysis phase of the project, we applied the project
planning and implementation standard ECSS-M-ST-10C which encompasses the
424 A. Hanafi et al.

Fig. 2 Project tree of the nanosatellite platform

entire process to be carried out to plan and execute a space project in a coordi-
nated, efficient, and structured manner [7]. This process makes it possible to identify
the Project breakdown structures which constitute a common and unique repository
making it possible to identify the tasks and responsibilities of each actor.
Thus, the product tree, shown in Fig. 2, conceptually decompose a typical CubeSat
space segment into platform and payload, with the platform (subject of our work)
composed of the traditional subsystems, such as on-board computer, power system,
communication system and attitude control system.

3.2 Open-Source Arduino Technology

The initial purpose of the concept for the Arduino open-source hardware
was to build a simple and low-cost microcontroller board that students could
easily use for projects. The great advantage of the Arduino platform is that the
hardware and software resources of the different boards and shields are completely
open source [8].
However, the use of a Commercial off-the-shelf (COST) open-source module
like Arduino in the space applications, is a challenge. Indeed, even if some CubeSat
missions have already successfully integrated Arduino modules in their design [9,
10], they do not share the open-source spirit of our initiative and indicate the
integration and implementation procedures of these modules.
Considering constraints related to the CubeSat dimensions, weight, and energy
consumption, especially for 1U formats, the main component of the platform
subsystem design will be the Arduino Nano module. The Arduino Nano is a board
of dimensions 45 × 18 mm2 , noticeably light (5 g) and easily weldable on electronic
Design and Implementation of an Open Source and Low-Cost Nanosatellite Platform 425

boards meeting the CubeSat 1U standard namely 10 × 10 cm2 . It works with 5 V

and has enough capacity and interfaces to control and communicate with the other
nanosatellite modules:
• Frequency: 16 MHz
• Memory: Flash 32 kb - EEPROM 1 kb - SRAM memory (internal) 2 kb
• Analog input pins: 6 (A0–A5)
• Digital I/O pins: 14 (out of which 6 provide PWM output)
• Communication: IIC, SPI, USART.

3.3 Hardware Platform Architecture

In this sub-section, we will discuss the adopted reference design solutions for
the nanosatellite hardware platform, more specifically the architecture and the
communication bus.
As shown in Fig. 3, we have opted for a distributed architecture where each
subsystem is equipped with a local microcontroller that will manage the subsystem
specific tasks as well as communication with the rest of the platform. This choice
offers the following advantages:
• The on-board computer is no longer a singular point of failure since its failure will
not necessarily lead to the failure of the entire system and/or the loss of contact
with the space segment.
• Independent design and testability of the subsystems which reinforces the open-
source spirit of the project. Indeed, the hardware interfacing of the subsystems
being standard, they can be developed independently by different work teams.
Regarding the choice of the communication bus, we had to choose between I2C
(Inter Integrated Circuit) and CAN (Controller Area Network). For the selection
criteria, energy consumption was favoured over robustness and reliability, as shown
in Table 1:
The choice of the I2C bus has the advantage of low consumption, ease of imple-
mentation and an open-source license. But it is a low-level protocol that must
be improved by implementing a high-level connection-oriented protocol based on
addressing and error checking.
426 A. Hanafi et al.

Fig. 3 Hardware platform architecture

Table 1 I2C and CAN serial communication bus comparison

Criteria Reference Weight I2C CAN
Consumption Active 25 mW/Standby 5 mW 5 3 1
Robustness Differential/CRC 4 1 3
Implementation Hard/Soft/Tools Arduino 3 3 1
Speed 1 Mbps 2 2 2
Total 32 24
Design and Implementation of an Open Source and Low-Cost Nanosatellite Platform 427

Fig. 4 AVR source code to uploading it to Arduino board

4 Open-Source Software Platform Design

4.1 Preparation of the Free Development Environment Atmel

Studio 7

Programming the Arduino platform by using the “pure” C language, instead of

Arduino’s Integrated Development Environment (IDE), allows the creation of
smaller, more optimized and more memory-efficient programs. We then must adapt
the Atmel Studio development environment to follow the same AVR toolchain used
by the Arduino IDE as shown in Fig. 4.
The following points summarize the preparation of the Atmel studio environment
to be able to program the Arduino Nano board via the bootloader without Arduino
IDE or programmer:
• Creation of a new “GCC C Executable Project” by selection the AVR microcon-
troller ATMega328P as device.
• In the toolchain properties of the project, make sure that “avr-gcc” is the used
compiler, and add the defined Symbols “F_CPU = 16000000UL”.
• Adding support for the Arduino programming mechanism, which requires the use
of the avrdude utility that is included in the Arduino IDE software installation. This
important part will start by identifying the COM port to which the Arduino board
is connected (is our case COM3). Then we will add an external programming
tools in the tools menu as shown in Fig. 5. The command and the arguments used
are deduced from the Arduino IDE verbose outputs:
– Command: C:\Program Files (x86)\Arduino\hardware\tools\avr/bin/avrdude.exe
– Arguments: -C “C:\Program Files (x86)\Arduino\hardware\tools\avr/etc.
/avrdude.conf” -v -patmega328p -carduino -PCOM3 -b115200 -D -
Uflash:w:"$(ProjectDir)Debug\$(TargetName).hex":i

4.2 Implementation of FreeRTOS Operating System

The onboard flight software of our nanosatellite platform presents the characteristics
of a real-time system, as it must manage various external events and perform tasks
with time constraints. The use of a real-time operating system such as FreeRTOS
makes it possible to meet this requirement by giving the possibility of having a
428 A. Hanafi et al.

Fig. 5 Atmel studio 7 -

external tools dialog

multi-tasking, modular and more reliable code to ensure the smooth operation of the
platform.
FreeRTOS is a small footprint real-time kernel for microcontroller. It is released
under the free GNU General Public License (GPL) and it offers different ports,
notably for ATmega323, but not for ATmega328. We then use the open-source repos-
itory of files “AVR ATmega port of freeRTOS” which can be flexibly configured uses
timers available on AVR ATmega devices to work with Arduino device [11].
For the port of the most recent version 10.4.1 of freeRTOS:
• The “freeRTOSBoardDefs.h” file contains most of the necessary variables to
change hardware platform. We define the variable allowing the use of the 328p
Timer0:

#elif defined(__AVR_ATmega328P__) || defined(__AVR_ATmega168__)

…..
#define portUSE_TIMER0 // to use 8 bit Timer0 for xTaskIncrementTick
// #define portUSE_TIMER1 // to use 16 bit Timer1 for xTaskIncrementTick

• The configuration file “FreeRTOSConfig.h” tailors the RTOS kernel to the devel-
oped software platform. We especially distinguish the clock frequency of the
microcontroller which is indicated by the F_CPU constant which was defined by
the Atmel Studio environment:

#define configCPU_CLOCK_HZ ( ( uint32_t ) F_CPU )

Design and Implementation of an Open Source and Low-Cost Nanosatellite Platform 429

Table 2 Main functions of the on-board computer library

Function Description
uint8_t clock_get_DateTime(tm * Getting date and time from the on-board RTC:
timeDate); DS3231
uint8_t clock_set_DateTime (tm * Setting date and time of the on-board RTC: DS3231
timeDateSet);
uint8_t setRegister(uint8_t reg, uint8_t Set the value of an I2C device internal register
value);
uint8_t readRegister(uint8_t reg); Set the value of an I2C device internal register. Used
to read the temperature sensor integrated in the RTC
DS3231

• There are some relevant and often used libraries added to the basic freeRTOS
capabilities, especially the “lib_io” that contains the tools to use the I2C bus.
The file “i2cMultiMaster.c” contains integrated interrupt driven master and slave
routines.

4.3 On-Board Computer Library

As a first step in the customizing of our open-source software platform, we have

added to the freeRTOS repository files a new library named “obc” and containing
drivers and the API functions for on-board computer I2C peripherals and sensors,
namely: the real time clock (RTC), the temperature sensor and the gyroscope.
This library contains the definition of the used data structures, the internal registers
of the I2C peripherals and several functions, the main ones of which are summarized
in Table 2:

4.4 Implementation of a Communication Protocol Over I2C

Bus

A connection-oriented protocol was designed on top of the I2C bus to allow commu-
nication between the different subsystems of the nanosatellite platform. This protocol
uses a transmit method separates between three types of packets as shown in Fig. 6:
Each type of frame begins with the destination sub-system address followed by
an identifier field which indicates the type of the frame:
• Write Data Frame: Used when transmitting data. The 2 most significant bits of
the identifier are set to “00” and the remaining 5 bits indicate the device module
number.
430 A. Hanafi et al.

Fig. 6 Communication protocol packet over I2C Bus

• Read Data Frame: Used when requesting data from the receiver. The 2 most
significant bits of the identifier are set to “01” and the remaining 5 bits indicate
the device module number.
• Request Frame: Used when requesting the receiver for doing some commands.
The 2 most significant bits of the identifier are set to “10” and the remaining 5
bits indicate the command to be executed followed if necessary, by parameters.
This approach allows software designers to choose the commands and data formats
relevant to the tasks of each subsystem.

5 Implementation, Test and Results

A real-time freeRTOS based application was created in the Atmel Studio develop-
ment environment to test the port of the operating system, the different functions of
our on-board computer library, and communication via the I2C bus.
The application has been implemented and tested in the Arduino nano module of
the on-board computer of the hardware platform shown in Fig. 7:
The total flash footprint came up to approximately 18 Kbyte which includes Free
RTOS, application and library code. A simple debug console interface provides a set
of text commands to test the implementation as shown in Fig. 8:
Design and Implementation of an Open Source and Low-Cost Nanosatellite Platform 431

Fig. 7 Experimental test setup

Fig. 8 Console interface showing test running of the application.

6 Conclusions and Future Work

This work is the beginning of an ongoing project to use open design to make space
technology accessible. Our team was thus able to develop in a noticeably short time
(3 months), the basis of an open-source hardware and software architecture that could
be used as a platform for CubeSats projects.
This was possible by using Commercial off-the-shelf products (Arduino modules)
and by adapting and enriching existing open-source software libraries. The encour-
aging results of this work have allowed us to set the following steps as future
work:
• Design a MPPT and Arduino Nano based smart electrical power system for our
nanosatellite platform.
• Design of the on-boards computer and power supply PCB (Printed Circuit Board)
by respecting the mechanical interfacing imposed by the PC104 standard and the
dimensional constraints (10 × 10 cm2 ) imposed by the CubeSat standard.
432 A. Hanafi et al.

• Improve the debugging shell interface used in this work and implement CRC
check in the communication protocol over I2C Bus, to detect accidental changes
to raw data during the transfer.
• Provide a complete repository containing the history of our project source files.
It will be hosted in a single Git repository like GitHub.

References

1. Williams C, DelPozzo S (2020) Nano microsatellite market forecast. 10th Edition, SpaceWorks
Annual Nano/Microsatellite Market Assessment
2. Erik Kulu, Nanosatellite Database. https://www.nanosats.eu/index.html#database. Accessed 5
Oct 2020
3. Open-source initiative. https://opensource.org/. Accessed 10 Oct 2020
4. Open-source hardware association. https://www.oshwa.org/. Accessed 10 Oct 2020
5. National Aeronautics and Space Administration NASA Headquarters (2007) “Systems
Engineering Handbook”. NASA/SP-2007–6105 Rev1
6. European Cooperation for Space Standardization. https://www.ecss.nl/. Accessed 10 Oct 2020
7. ECSS-M-ST-10C, (2009) Project Planning and Implementation. ESA-ESTEC
8. Arduino Open-source initiative. https://arduino.cc/. Accessed 10 Oct 2020
9. Geeroms D, Bertho S, De Roeve M, Lempens R, Ordies M, Prooth J (2015) ARDUSAT, an
Arduino-based cubesat providing students with the opportunity to create their own satellite
experiment and collect real-world space data. In 22nd ESA Symposium on European Rocket
and Balloon Programs and Related Research, vol 730, p 643
10. Roman-Gonzalez A, Quiroz-Olivares AE, Vargas-Cuentas NI (2020) Advances in the UCHSat-
1 nanosatellite: design and simulation. Adv Astronaut Sci Technol 3:65–74. https://doi.org/10.
1007/s42423-020-00054-1
11. FreeRTOS & libraries for AVR ATmega. https://github.com/feilipu/avrfreertos. Accessed 20
Oct 2020
Design and Realization of Fire Safety
System for Controlling and Monitoring
a Siren Using Arduino Uno

Abdennabi Morchid, Mustapha El Alaoui, Rachid El Alami, Hassan Qjidaa,

Karim El Khadiri, and Youness Mehdaoui

Abstract In this work, the design and realization of fire safety system for controlling
and monitoring a siren using Arduino Uno are proposed for decreases the number
of risks in the evacuation process to protect the medical team and the patients living
in the Moroccan University Hospital Centers. To solve the problem of controlling
and monitoring the siren in the existence of fire, we have proposed a new system
to control the siren during a fire and transmit its status to the Control and Signaling
Equipment, this new system will be placed between the ATCAV control system and
the siren. We used the Arduino UNO, voltage sensor, H-bridge, voltage regulator,
Bipolar Transistor NPN TIP120, Optocoupler, LCD and siren in the hardware tools.
And also we used the Arduino IDE and EasyEDA in the software tools.

Keywords Fire safety system · Arduino Uno · H-bridge · Siren · Sensor voltage

1 Introduction

As we all know the fire appears since the existence of human beings on earth, it
is among the materials necessary to nourish the needs and daily requirements of
humanity. Moreover, fire also has a bad history with humanity because of the fires
caused lately in the different parts of our world. For that, it has always been a threat
to human life.

A. Morchid (B) · M. El Alaoui · R. El Alami · H. Qjidaa

Laboratory of Computer Science, Signals, Automation and Cognitivism (LISAC),
Department of Physics, Faculty of Sciences Dhar El Mahraz, Sidi Mohamed Ben Abdellah
University, Fez, Morocco
e-mail: [email protected]
K. El Khadiri · Y. Mehdaoui
Laboratory of Computer Science and Interdisciplinary Physics, Sidi Mohamed Ben Abdellah
University, ENS-Fez, Fez, Morocco
Y. Mehdaoui
Research Team in Electronics, Instrumentation and Measurements, USMS, Beni Mellal, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 433
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_40
434 A. Morchid et al.

The first intervention of humanity to control and extinguish the fire is being by
the use of water, but this solution is less effective in some cases. For this reason the
humanity looked for other solutions that depend on the technological development
to detect and control the risk of fire automatically.
The risk of fire can be very significant and cause loss of life. The only solution
to relieve these losses is to respond quickly to this emergency. In such situations,
the early detection of fire and the provision of a safety alert will protect loss of life
and property through the use of a fire alarm system that can be monitored locally or
remotely, as appropriate [1].
However, to be able to use the technology currently available, it will have to be
beneficial to the lives of others, which will be a great contribution to the community
[2, 3]. Although there are several well-developed systems available for use in practical
scenarios [4, 5] and [6], but also these systems are not cost-effective, unreliable in
automatic response, and not easy to implement. Today, in the market, there are
different types of fire alarm systems that are very complex in design, which requires
preventive maintenance to ensure its operation. In the meantime, if the system in
use requires maintenance to be performed each time, the cost of the use of it could
increase [1]. There are many researchers who have already worked on the fire safety
system to reduce the risks resulting from fire hazards [7, 8].
The fire safety system installed at the level of the Moroccan University Hospital
Centers (CHU) includes an ATCAV control system (Addressable Controllable Tech-
nical Alarm of the Vega range), this latter can only control the siren during the fire
but does not have the monitoring system for the siren.
In addition, if the fire exists and the siren fails at the same time at the university
hospital, in this situation the Control and Signaling Equipment (ECS) cannot trigger
and locate the fault, which increases the number of risks in the evacuation process
to protect the medical team and the patients living at the university hospital.
To solve this problem, we have proposed a new Fire Safety System to control
the siren during a fire and transmit its status (working, corrupted) to the Control and
Signaling Equipment (ECS), this new system will be placed between the ATCAV
control system and the siren.
We presented this paper as the following: in Sect. 1 an Introduction; in Sect. 2 a
Description of the proposed Fire Safety System; in Sect. 3 a Description of the tools
used; in Sect. 4 a Simulation and hardware realization of the proposed Fire Safety
system; and in Sect. 5 a conclusion.

2 Description of the Proposed Fire Safety System

In case of fire a signal is sent from the Fire Detector to ATCAV which will switch the
relay armature of the ATCAV to the power supply pin, consequently a command is
sent to the Fire Safety System which will switch on the siren. The principal functions
of the proposed Fire Safety System are illustrated in Fig. 1.
Design and Realization of Fire Safety System … 435

Fig. 1 Principal functions of

the proposed fire safety
system

Fig. 2 Connecting pin of the

ATCAV

The ATCAV is connected to the siren by the NO pin of its relay and the common
(COM) pin is connected to a 24 V power supply. In case of fire, a signal sent by the
control unit allows the relay armature switched to the 24 V pin, which operates the
siren.
The ATCAV installed in Moroccan CHU can only control the siren during the fire,
it never has the monitoring part of the siren, that is to say that the contact between
pins M and J of the ATCAV will always remain closed as illustrated in Fig. 2.
On the part of monitoring the siren, our proposed Fire safety System is based
on the measurement of the voltage at the terminal of a resistor with the siren and
their polarity resistance in the middle of H-bridge, the purpose of using an H-bridge
for reverse polarity direction. In addition, the part of controlling is assured by an
Arduino Uno card that has been programmed to control the siren.
The siren is powered by 24 V, when the fire is declared by the Fire Detector block,
the normal polarity is applied. That means the positive terminal of the voltage source
in the middle of the H-bridge is connected with the positive terminal of the siren and
the negative terminal is connecting to the negative terminal, this polarity allows to
measure a voltage to determine the state of monitoring as shown in Fig. 3.
436 A. Morchid et al.

Fig. 3 Normal polarity

Fig. 4 Reverse polarity

On the other hand, when the fire is not declared by the Fire Detector block, the
reverse polarity is applied. That is means the positive terminal is connected to the
negative terminal and the negative terminal is connected to the positive terminal of
the siren as shown in Fig. 4.

3 Description of the Tools Used

The hardware tools used in our proposed Fire Safety system are Arduino Uno, voltage
sensor, H-bridge, voltage regulator, Bipolar Transistor NPN TIP120, Optocoupler,
LCD and siren. On the other hand the software tools used are Arduino IDE and
EasyEDA.

3.1 Arduino Uno Card

The Arduino system is a tool for making small computers that can control more things
in the hardware world than your desktop computer [9]. It is an open source platform
of programs, electronics that are based on a microcontroller card and software, a true
integrated development environment, to write, compile and transfer the program to
the card.
Among all the cards of the Arduino family, we chose the UNO version because
of its characteristics and it is from this base that we developed the whole project.
Design and Realization of Fire Safety System … 437

Fig. 5 Arduino Uno card

The Uno model from Arduino is an electronic card whose heart is an ATMEL
microcontroller of the ATMega328 reference. The ATMega328 microcontroller is an
8-bit microcontroller from the AVR family, which can be programmed in C language.
The Arduino Uno card as shown in the Fig. 5 above is a microcontroller card based
on the ATmega328. It has: 14 digital input/output pins (6 of which can be used as
PWM (pulse width modulated) outputs), 6 analogue inputs (which can also be used
as digital input/output pins), a 16 Mhz quartz crystal, a USB connection, a power
jack connector, an ICSP connector (“in-circuit” programming) and a reset button.

3.2 Voltage Sensor V < 25 V and Photo-Coupler PC817

Figure 6 shows the voltage sensor V < 25 V. This module is based on a resistive
design divider principle, can reduce the voltage of the input terminal connected five
times, Arduino analog input voltage up to 5 V. The main characteristics of it are: an
input voltage range (DC 0–25 V), a voltage detection range (DC 0.02445–25 V), an

Fig. 6 The voltage sensor V

< 25 V

Fig. 7 The photo-coupler

PC817
438 A. Morchid et al.

analogue voltage resolution 0.00489 V, a DC input (positive terminal connected to

VCC), GND negative pole and an output interface (“+” then 5/3.3 V, “−” then GND,
“s” then the AD Arduino pin) [10].
An Optocoupler consists of an infrared LED and a phototransistor or photo-
diode as shown in Fig. 7. There is an electrical isolation between the two. The
electrical characteristics are: a maximum current in the LED, a maximum current in
the phototransistor (CTR: current transfer ratio), a leakage current in the phototran-
sistor (dark current), a power that the assembly can dissipate and insulation voltage
of the Optocoupler [11].
An Optocoupler is based on an LED and a phototransistor or photodiode. When a
current is passed through the LED, it glows (emits infrared) in a light-tight housing.
The light emitted by the LED is picked up by the phototransistor, which then becomes
conductive. It is therefore possible to transmit an electric current while at the same
time providing electrical insulation. In its principle, the Optocoupler makes the
successive conversions: an electric current, an infrared light and an electric current.

3.3 H-Bridge

The H-bridge allows a current to flow through a load in both directions as illustrated
in Fig. 8. It is used to run a DC motor in both directions and stop it (motor controlled),
to create an AC voltage (inverter) and to create an adjustable DC voltage (chopper)
[12].
A H-bridge is composed of 2 parts, a power part composed of 4 SW-controlled
switches (relay, bipolar transistor or MOS…) and a control part which controls the
opening (O) or closing (F) of the switches. This control is most often done like PWM
(pulse width modulation).
The objective of using an H-bridge based on NPN TIP120 bipolar transistors
in our project to reverse the polarity direction from the idea of reversing the two
directions of a load (Siren).

Fig. 8 H-bridge with an

active branch
Design and Realization of Fire Safety System … 439

Fig. 9 The 16 * 2 LCD

display

Fig. 10 The bipolar

transistor TIP120 of NPN
type

3.4 16 * 2 LCD Display and Bipolar Transistor TIP120 NPN

An LCD display is capable of displaying all the usual alphanumeric characters and
some additional Symbols as shown in Fig. 9. For some displays it is even possible
to create your own characters [13].
The Bipolar Transistor TIP120 of NPN type is shown in Fig. 10. It is used to switch
high current loads (up to 5A), as a medium power switch with high amplification is
required [14]. And also it is used in speed control of motors and inverter and other
rectifier circuits.

3.5 Voltage Regulator LM317 and AVS 200SIP Siren

The voltage regulator LM317 is shown in Fig. 11. The LM317 is a linear voltage
regulator. This family of integrated circuits allows, for any input voltage, to provide
a fixed output voltage from which current can be drawn. It is a 3-leg serial regulator
capable of providing 1.5 A on an output voltage ranging from 1.25 to 37 V. It is
exceptionally easy to use and requires only 2 external resistors to fix its output
voltage. In addition, the power supply rejection and insensitivity to load impacts
make it an advantageous regulator [15].

Fig. 11 The voltage

regulator LM317
440 A. Morchid et al.

Fig. 12 The Siren AVS2000-SIP

The AVS 2000 sound diffuser is a sound evacuation device for fire. It can be
integrated into the Alarm Equipment (EA) provided for in the NFS 61 936 standard
and can be controlled by an Alarm Management Unit (AGU). The Siren AVS2000-
SIP is shown in Fig. 12.
It is designed around piezoelectric technology, AVS 2000 guarantees very low
power consumption. Its circular design and its side vents allow it to broadcast sound
in several planes, ensuring perfect hearing of the alarm throughout the volume [16].

4 Simulation and Hardware Realization of the Proposed

Fire Safety System

4.1 Simulation

Figure 13 shows the schematic of our proposed Fire Safety System in EasyEDA
software. And Fig. 14 shows the schematic of the power Supplies system, Voltage
Sensor and input pins in same software.
We have used the LM317T regulator in schematic to adapt the input power supply
voltage that equal 24 V to an adapted voltage at 9 V which used to power the Arduino
Uno Card. And also the voltage sensor used to measure the voltage at the resistor
terminal.
In the power part, the siren and the polarity resistor located in the middle of the
H-bridge as illustrated in Fig. 15.
In the control part, the Arduino Uno Card with the connection connectors and
all the components as well as the output pin and also the part who declares to the
proposed Fire Safety System if there is a fire and the status of the siren, are illustrated
in Fig. 16.
Design and Realization of Fire Safety System … 441

Fig. 13 The schematic of the proposed fire safety system in EasyEDA software

Fig. 14 The schematic of

the power supplies system,
voltage sensor and input pins
in EasyEDA software

Fig. 15 The schematic of an

H-bridge in EasyEDA
software
442 A. Morchid et al.

Fig. 16 The schematic of

the control part in EasyEDA
software

4.2 Hardware Realization

We have used a 24 V DC power supply for the siren which powered during a fire.
When there is no fire the voltage sensor will measure a zero voltage which means a
disconnection for the proposed Fare Safety system.
The practical realization of the voltage regulator is illustrated in Fig. 17, we have
used a multimeter to show the actual output value by using two resistors (R1 = 220 
and R2 = 1.5 ) that regulate the output value at 9.5 V and also two capacitors to
suppress parasites (noise) in the output.
If the fire existing and the siren on we will have that the push button is pressed,
the normal polarity on the H-bridge is +24 V, Either the polarity resistance exists
or not, the measured voltage is 8.64 V when R is ON and 7.42 when R is OFF. line
ok = working and green LED on, and no change in the contact between the M and j
pins of the ATCAV as there is no fault as illustrated in Fig. 18. If else the fire existing
and the siren off we will have that the push button is pressed, the normal polarity on

Fig. 17 The practical

realization of the voltage
regulator

Fig. 18 The practical

realization of the proposed
fire safety system in the case
of the fire existing and the
siren working
Design and Realization of Fire Safety System … 443

the H-bridge is +24 V, the line OFF = default, the red LED lights up, indicating that
a disconnection has occurred, open the contact between pins M and J of the ATCAV,
then the ECS will declare the existence of a fault accompanied by the address of
the ATCAV, and the value of the voltage 0.02 V with the type of fault at the siren is
displayed on the LCD screen as illustrated in Fig. 19.
If the fire not existing and the siren on we will have that the push button is not
pressed, the application of reverse polarity on the H bridge is −24 V, Either the
polarity resistance exists or not, the measured voltage is 7.20 V when R is ON and
5.59 V when R is OFF, the line ok = working and green LED lit and no change in
the contact between pins M and j of the ATCAV since there is no disconnection as
illustrated in Fig. 20. If else the fire not existing and the siren off we will have that

Fig. 19 The practical

realization of the proposed
fire safety system in the case
of the fire existing and the
siren corrupted

Fig. 20 The practical

realization of the proposed
fire safety system in the case
of fire not existing and the
siren working

Fig. 21 The practical

realization of the proposed
fire safety system in the case
of fire not existing and the
siren corrupted
444 A. Morchid et al.

the push button is not pressed, the application of reverse polarity on the H bridge is
−24 V, the red LED lights up to indicate disconnection, the line OFF = fault and the
measured voltage is 0.04V, open the contact between pins M and J of the ATCAV
and then the ECS will declare the existence of a fault accompanied by the address of
the ATCAV as illustrated in Fig. 21.

5 Conclusion

The proposed Fire Safety System have been successfully designed and realized to
controlling and monitoring a Siren using Arduino Uno. Simulation, analysis and
practical realization are all included in this study. It significantly solve the problem
of controlling and monitoring the siren (transmit its status) in the existence of fire. The
proposed Fire Safety System implemented in Arduino Uno by using a voltage sensor,
an H-bridge, a voltage regulator, a Bipolar Transistor NPN TIP120, an Optocoupler,
LCD and a siren. The new system simulated in Arduino IDE and EasyEDA software
tools.

References

1. Suresh S, Yuthika S, Vardhini GA (2016) Home based fire monitoring and warning system. In:
2016 international conference on ICT in business industry & government (ICTBIG), Indore,
pp 1–6. https://doi.org/10.1109/ICTBIG.2016.7892664
2. Zeebaree SR, Yasin HM (2014) Arduino based remote controlling for home: power saving,
security and protection. Int J Sci Eng Res 5(8):266–272
3. Piyare R, Tazil M (2011) Bluetooth based home automation system using cell phone. In:
15th international symposium on consumer electronics (ISCE), Singapore. IEEE, pp 192–195.
https://doi.org/10.1109/ISCE.2011.5973811
4. Jinghong L, Xiaohui Z, Lu W (2012) The design and implementation of fire smoke detection
system based on FPGA. In: 24th Chinese control and decision conference (CCDC), Taiyuan,
pp 3919–3922. https://doi.org/10.1109/CCDC.2012.6244626
5. Celik T, Demirel H, Ozkaramanli H, Uyguroglu M (2006) Fire detection in video sequences
using statistical color model. In: IEEE international conference on acoustics speech and signal
processing proceedings, Toulouse, p II. https://doi.org/10.1109/ICASSP.2006.1660317
6. Angayarkkani N, Radhakrishnan N (2009) Efficient forest fire detection system: a spatial data
mining and image processing based approach. Int J Comput Sci Netw Secur 9(3):100–107
7. Song WS, Hong SH (2007) A reference model of fire detection and monitoring system using
BACnet. J Build Environ 42(2):1000–1010. https://doi.org/10.1016/j.buildenv.2005.10.028
8. Sharma S, Chand K, Sharma D, Guha P (2018) Development of an early detection system for
fire using wireless sensor networks and Arduino. In: International conference on sustainable
energy, electronics, and computing systems (SEEMS), Greater Noida, India, pp 1–5. https://
doi.org/10.1109/SEEMS.2018.8687333
9. https://components101.com/microcontrollers/arduino-uno
10. https://components101.com/sensors/voltage-sensor-module
11. https://learnabout-electronics.org/Downloads/PC817%20optocoupler.pdf
12. https://www.build-electronic-circuits.com/h-bridge/
Design and Realization of Fire Safety System … 445

13. https://components101.com/16x2-lcd-pinout-datasheet
14. https://components101.com/transistors/tip120-pinout-datasheet-equivalent
15. https://components101.com/lm317-pinout-equivalent-datasheet
16. https://studylibfr.com/doc/901746/fiche-tech-avs2000-mise-en-page-1.qxd
Water-Purifying Distiller with a Cooling
System Controlled by a Photovoltaic
Panel

Kamilia Mounich, Aicha Wahabi, and Mohammed Chafi

Abstract Globally, the demand for good quality drinking water is growing. Indeed,
the population is increasing rapidly and the water needs of industry and agriculture are
higher and higher. This paper presents the theoretical study and practical realization
of a system which consist to desalinate and purify sea water, equipped with a cooling
system powered with an automatic photovoltaic panel sun tracker. It’s a cheaper
design, simple and robust, adapted to the needs and specificities of our vast country.
This equipment has been tested during the months of December and October 2020.
The daily production recorded in distilled water varies from 600 to 645 ml under a
daily radiation received on the plan and vary within a range of 5.0 to 5.3 kWh/m2
under an ambient temperature which varies from 21.6 to 33.5 °C.

Keywords Sea water distillation · Servo drive · PIC · Arduino

1 Introduction

By 2022, water consumption [1] is expected to increase by 40% and half the planet
will lack clean water [2]. Today, no less than 900 million people do not have access
to drinking water. Each year, eight million people, including 1.8 million children, die
from pathologies linked to the consumption of unsanitary water, diarrhea, cholera,
typhoid, malaria, etc. [3].
For these facts, we have been encouraged to think about a solution: That is the
sea water desalination system [4]. To explain the technique, seawater distillation is
boiling followed by liquefaction. This process is therefore based on a phase change:
The seawater solution is introduced into the heated flask. When it boils, it turns into
vapor and rises in the air column. The water molecules vaporize, thus separating
from the chloride ions Cl− and sodium Na+ [5]. The water then flows through a
cooling tube to return to a liquid state and falls as droplets into the beaker, forming
a pure water distillate [6]. Thus, through this process, the water is separated from

K. Mounich (B) · A. Wahabi · M. Chafi

LIPE, Higher School of Technology, Hassan II University of Casablanca, Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 447
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_41
448 K. Mounich et al.

the salt and other microbes that remained in the initial solution. Ahmed Khedim [7]
studied in this context a seawater distiller with multiple stages and heat recovery. It
is therefore a question of finding an operating mode, allowing the distiller to obtain
the best efficiency, when thermal energy is recovered several times. This method of
recovery and use of several times the same amount of solar energy captured has for
aim to significantly reduce energy consumption. Experience shows that the system
provides a good return. The idea developed by the authors is to achieve using simple
technical means based on solar energy and with a minimum cost, a modest supply of
drinking water. This method used, involves recovering and using the same amount of
thermal energy several times. It has for aim to lower energy consumption so that solar
energy becomes an alternative and competitive source among all water desalination
technology techniques.
Ali A. Badran [8] presented a purely experimental study on a solar still connected
to a solar collector to study the effect of increases in settings on the still under local
conditions. The results show that productivity increases significantly compared to
that of the distiller alone.
Rajesh Tripathi and G.N. Tiwari presented an experiment to determine the convec-
tive heat transfer [9] coefficient for the active solar distillation system. They found
that the coefficient of convective heat transfer between the water and the air surface.
The interior of the condensation cover depends significantly on the water depth.
Also, they observed that the yield obtained is important over the hours of sunshine
compared to the daily production for water depths in the solar still (0.10 m and
0.15 m), due to the storage effect. The active still is more efficient than the passive
still but their cost is very high and their use is more complex. In the experiment
we presented, the heat source is provided by a butane cylinder, however, it must
be replaced by a solar heat source to minimize transformation cost and avoid the
complexity of the process.

2 Experimental Part

2.1 Equipment and Materials

Our study was conducted on a laboratory-scale distiller. It is consisting of a heat

source and a 750 ml container of sea water and a 6 mm diameter copper tube formed
into 7 turns, a cooling fan powered through a 3.7 V lithium battery * 2 which in
turn is charged through a 5 V, 12 W photovoltaic panel controlled by an SG09 servo
motor, 4 LDR and an Arduino Nano Atmel Atmega 328p board. The fresh water
passes through an activated carbon filter using gravity phenomenon and then to the
final container.
Water-Purifying Distiller with a Cooling System Controlled ... 449

Fig. 1 Sea water distiller, purifying schema

2.2 Experiences

The experience is spread over a period of 2 months (September, October). The water
of the process comes from sea water. The stainless-steel container is filled with 700 ml
of water, 2/3 of the volume, the salinity of the water is 28.7 g [10]. The increase in
temperature due to the thermal effect which heat the saline water which evaporates.
This evaporation capacity increases as the temperature increases to boiling. The
state of water changes from liquid to gas. The water vapor in the hot air is cooled
through the passage to the copper solenoid; the vapor condenses inside the copper
tube (Fig. 1). There is a formation of water droplets which flow through the spiral
which leads it to an activated carbon filter [11].
The experimental study will focus on the measurement of the following param-
eters: Global solar radiation on an oblique plane is measured using a pyranometer.
The different temperatures are measured using type K thermocouples (Chromel -
Alumel). These climatic parameters of the atmosphere are taken at the end of each
hour by an instrument called Testo 925. The water salinity measurement is done
using the HI98319 Salinity Tester, the water purity measurement is done using the
TDS meter (Total Dissolved Solid) (Table 1).
450 K. Mounich et al.

Table 1 Taste of water with

TDS level (milligrams/liters Evaluation
different concentrations of
TDS [12] Less than 300 Excellent
300–600 Good
600–900 Acceptable
900–1200 Bad
More than 1200 Unacceptable

Fig. 2 Electronic schema

2.3 Electronic components

The electronic part is the key milestone of this project success. In fact, the electric
power is ensured by the photovoltaic panel to all electronic items. The type of photo-
voltaic panel is monocrystalline 5 V 12 W. Below are both electronic schema and
real prototype picture (Fig. 2).

2.4 Grouping of photovoltaic cells

We have coupled the two photovoltaic panels in parallel. A parallel mounting of two
cells increases the generator output current thus created. In a group of identical cells
Water-Purifying Distiller with a Cooling System Controlled ... 451

Fig. 3 Real project

Fig. 4 Characteristics resulting from a grouping of Np cells in parallel

connected in parallel, the cells are subjected to the same voltage and the characteristic
resulting from the grouping is obtained by adding the currents, Fig. 3. Equation (1)
in turn summarizes the electrical characteristics of a parallel association of NP cells.

Icc = N ∗ Icc (1)

2.5 Electrical Characteristics of a Photovoltaic Cell:

Figure 4 shows the equivalent diagram of a photovoltaic cell under lighting. It corre-
sponds to a current generator mounted in parallel with a diode. Two parasitic resis-
tances are introduced in this diagram. These resistances have a certain influence
452 K. Mounich et al.

Fig. 5 Equivalent electrical diagram of a PV cell

on the characteristic I = f (V) of the cell: The series resistance (Rs) is the internal
resistance of the cell; it mainly depends on the resistance of the semiconductor used.
The shunt resistance (Rsh) is due to a leakage current at the junction; it depends
on how it was mounted (Fig. 5).
The mathematical model for the current-voltage characteristic of a PV cell is given
by the formula (2) [13].
    
q* Vpv + Ipv * Rs Ipv *Rs + Vpv
Ipv = Iph − Isat exp + −1 − (2)
a*K*T*N Rsh

Where:
Iph : light generated current of PV module in Amperes
Isat : reverses saturation current in Amperes
Rs : series resistance in Ohms
Rsh : parallel resistance in Ohms a Diode ideality factor
K: Boltzmann constant (1.38 × 10−23 J/K)
q: electron charge (1.6 × 10−19 C)
N: number of series connected cells in a module
T: module temperature in Kelvin
Ipv : output current of PV module in Amperes
V: output voltage of PV module in Volts

3 Results and Discussion

We chose the day of 09/10/2020 which is characterized by strong sunshine throughout

the day with a maximum temperature of 32 °C, specify that the time of the first drop
Water-Purifying Distiller with a Cooling System Controlled ... 453

was detected at 10:30 a.m. and the total amount of distillate collected until 12 p.m.
is equal to 645 ml (Figs. 6 and 7).

DisƟlled water quanƟty per Ɵme 09/10/2020

700
600
500
400
300 quanƟty

200
100
0
10:40 10:50 11:00 11:10 11:20 11:30 11:40 11:50 12:00

Fig. 6 Quantity of distillate by time

Fig. 7 Solar energy potential in Morocco

454 K. Mounich et al.

PPM of salt Salt measurement in ppm

35000
30000
25000
20000 Series1
15000
10000
5000
0
Saline water Fresh water
09/10/2020

Fig. 8 Salt measurement TDS

Among the results we did the physico-chemical test of the fresh water coming
from the filter with the TDS device and we had the following results (Fig. 8):

4 Conclusion

The role of this project is to give a clear vision of the importance of using the
renewable energies (solar energy) as a key milestone to solve the problem of the
scarcity of drinking water in certain critical areas, and in the rest of the world in
general. Indeed, the use of solar energy offers a gigantic pure source and reliable
with competitive cost.
The problem of producing fresh water by solar means is important to two
Viewpoints: need to save energy and concern for new water resources.
This paper presents the experimental evaluation of the main parameters influ-
encing the operating characteristics of a solar distiller with photovoltaic panels.
The results obtained showed that the production, the overall efficiency of the Solar
still increases with solar irradiation provided by the servo motor and photo resistors
(Sun follower) and the efficiency of the copper tube cooling system.
Following the various experiments carried out on our solar distiller with photo-
voltaic panels, we found that to improve its performance and make it efficient, it will
take:
– Heat gain can be provided by preheating the feed water from a solar thermal
source.
Water-Purifying Distiller with a Cooling System Controlled ... 455

– Add a cooling module (Peltier) to accelerate the condensation.

– Reduce droplet losses through the filter by using suitable materials.
In addition, we must highlight the urgency to take the benefits that nature offers
such as the use of the activated carbon to purifying reused water.
This project is under improvements to ensure a max of performance.

References

1. Christopoulosa K, Pospotikisa N, Kostopoulosa Em, Kondilib E, Kaldellis JK (2018) Experi-

mental analysis of the water salinity impact on the energy consumption of small desalination
plants. Procedia 2(5):171–178
2. Anderson TL, Scarborough B, Watson LR (2013) Water crises, water rights, and water markets.
In: Encyclopedia of energy, natural resource, and environmental economics, pp 248–254
3. Mallick R, Mandal S, Chouhan P (2020) Impact of sanitation and clean drinking water on
the prevalence of diarrhea among the under-five children in India. Child Youth Serv Rev
118:105478
4. Jeffrey G, Hui T (2016) Thermodynamic perspective for the specific energy consumption of
seawater desalination. Desalination 386:13–18
5. Gradziel S (2019) Analysis of thermal and flow phenomena in natural circulation boiler
evaporator. Energy 172:881–891
6. Lito MGH, Camões MF, Viçoso CM (2006) Catarina effect of condensation phenomena on
potentiometric measurements. Anal Chim Acta 575:45–51
7. Khedim A, Schwarzer K, Faber C, Müller C (2004) Production décentralisée de l’eau potable
à l’énergie solaire. Desalination 168:13–20
8. Badran AA, Assaf LM, Kayed KS, Ghaith FA, Hammash MI (2004) Simulation and
experimental study for an inverted trickle solar still. Desalination 164:77–85
9. Zhanga M, Lia R, Wu Y, Wang L, Song G, Li J (2020) Numerical study of the convective heat
transfer coefficient of the hand and the effect of wind. Build Environ 188:107482
10. Lerm. https://doc.lerm.fr/salinite-leau-mer/
11. Sbardella L, Comas J, Fenu A, Rodriguez-Roda I, Weemaes M (2018) Advanced biological acti-
vated carbon filter for removing pharmaceutically active compounds from treated wastewater.
Sci Total Environ 636:519–529
12. WHO (2008) Guidelines for drinking-water quality. Total dissolved solids in drinking-water,
3rd edn, vol 1. Health criteria and other supporting information. World Health Organiza-
tion, Geneva. https://www.who.int/water_sanitation_health/dwq/chemicals/tds.pdf. Accessed
10 Oct 2020. ISBN 978 92 4 154761 1
13. Jena D, Ramana VV (2015) Modeling of photovoltaic system for uniform and non-uniform
irradiance: a critical review. Renew Sustain Energy Rev 52:400–417
Security at the Application Layer
and the Physical Layer in 5G

Hafida Amgoune and Tomader Mazri

Abstract While applying the magnificent benefits of storage methods to meet 5G

service demands, security, trust issues in Cloud, and fog Computing are essential
and exciting topics. There are various types of attacks in different logical planes (i.e.
control plane and data plane), for example, the physical is particularly vulnerable
to attacks such as spy attacks, scrambling attacks, identity theft attacks, primary
user emulation (PUEA) attacks, and threats through wireless channels. The main
attacks in the control plane are the attacks of the protocol of the MAC layer (Media
Access Control) and common threats from control channels such as MAC spoofing,
attack by extended denial of service, attack by jamming, etc. The 5G Service-Based
Architecture (SBA) undergoes a lot of processing because of its importance, which
is to make it the most targeted by attacks. Among the attacks that can be found in
SBA: virtualization attacks, threats to user privacy, and other security threats.

Keywords 5G · Security · Physical layer · Application layer · eMBB · mMTC ·

URLLC

1 Introduction

The researchers found that the security architecture utilized in 4G did not meet the
needs of 5G. Although, some researchers have proposed a sort of inflexible unified
network security architecture, the security architecture must always be adapted
to the needs of the 5G eMBB (enhanced Mobile Broadband), mMTC (massive
Machine Type Communications), and URLLC (Ultra-Reliable and Low Latency
Communications) application scenarios. They also proposed a physical layer authen-
tication scheme dependent on the hash technique. Despite the fact that this tech-
nique is light, the application scenarios are limited. Whereas, others are conducting
cross-authentication studies.

H. Amgoune (B) · T. Mazri

IBN TOFAIL University, Kenitra, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 457
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_42
458 H. Amgoune and T. Mazri

Nowadays, the security between user equipment (UE), base stations, and core
networks is highly needed, furthermore, between the UE and the base station, inter-
layer security measures are appropriate, whereas, the physical layer security measures
are perfect between the base station and the main network. In the case of UEs (D2D
(Device-to-device) pairs or UE), the inter-layer security measures and the traditional
security measures are adopted to the circumstances. Moreover, physical layer security
measures work best in low-speed cases, thus, in high-speed cases, they have to face
inter-layer security measures. Many questions deserve to be studied and explored in
the field of security in 5G. In particular, can we use cryptographic algorithms on all
devices in5G services (eMBB, mMTC, and URLLC)? It should be noted that many
of these devices cannot have additional circuitry, power consumption and code space
required to process data for cryptographic methods due to the limited resources of
some 5G devices. The rest of this paper is organized as a following. In Sect. 2, we
cover security at the application layer, in Sect. 3, we present security methods at the
physical layer. In Sect. 4, we present the discussion. The conclusion is formulated in
Sect. 5.

2 Security at the Application Layer

Ensuring security between IoT devices and MEC (Multi-Access Edge Computing)
servers is a challenge because we are talking about both physical and superior secu-
rity. Security mechanisms must be applied to protect the offload process from espi-
onage attacks when IoT devices delegate computational tasks to MEC servers. Partic-
ularly complicated, especially for IoT devices that have limited resources. To do this,
security mechanisms must be integrated into the MEC architecture. The cooperative
calculation and transition procedure consists of two stages:
• Safe execution of account tasks.
• Secure transfer of account results.
Appear IoT applications such as wearables i.e. electronic devices which have the
possibility to store and process user data. These devices are integrated into the clothes
or accessories of the person wearing them. These devices store and manage sensitive
information about humans and intelligent healthcare. The security drawbacks of these
wearables are numerous, as an example, the leakage of personal information is cited,
which will impact the privacy of the person.
Previously, IoT security was enhanced by data encryption at the upper layer of
the communication protocol stack. However, generating, distributing and managing
secret keys can be complicated and difficult to implement large-scale IoT networks
especially since there are many complex nodes like sensors.
According to the researchers [1], the 5G network should support differentiated
scenarios and various wireless services. Indeed, different types of services have very
different security requirements. For example, online payment requires a much higher
level of security than usual. However, encryption-based methods can only provide
Security at the Application Layer and the Physical Layer in 5G 459

“binary” levels of security. In other words, the information is protected if the secret
key is exchanged securely.
According to the same researchers, IoT applications presented by machine type
communications (MTC) must be supported by 5G, where huge MTC devices run out
of energy, storage and computing capacity, and the complex encryption/decryption
algorithms or protocols cannot be applied [1].
For the reasons below these researchers have proposed to use physical layer secu-
rity solutions (PLS) which take full advantage of the characteristics of wireless
channels to degrade the qualities of the signal received by malicious users and to
realize transmission. Secure without keys via signal design and signal processing
techniques. PLS avoids the difficulties of distributing and managing secret keys
and offers flexible security levels thanks to the design of an adaptive transmission
protocol. In addition, PLS techniques correspond well to the characteristics of 5G
networks. The most promising solution to avoid security threats at the 5G network
level is the application of PLS [1].

3 Security at the Physical Layer

The physical layer provides the best wireless network security [2]. There are many
security methods at the physical layer level, we mention:

3.1 The First Method: Using the Authentication Method

EAP/AKA

In the article [3], the researchers proposed a more flexible security method that can
be adapted to the needs of the 5G application scenarios, namely eMBB, mMTC
and URLLC. They proposed to introduce a light authentication scheme between
layers below the proposed scheme. Channel state information and the cross-layer
authentication based on EAP/AKA protocol proposed by [3] (Fig. 1).
Extensible Authentication Protocol or EAP is a network communication protocol
embedding multiple methods of partner authentication, i.e., it is designed to provide
an authentication mechanism to authorize or prohibit the establishment of the
connection. EAP-AKA (Authentication and Key Agreement) is an EAP method for
customers of mobile telephone networks. This method allows mutual authentication
of the user equipment and the network. This protocol is mainly used in the following
settings:
• Authentication when attaching to a mobile phone network.
• Reuse of authentication elements from the mobile network when attaching to a
wireless network which belongs to the operator.
460 H. Amgoune and T. Mazri

Fig. 1 Channel state

information and the
cross-layer authentication
based on EAP/AKA protocol
proposed by [3]

3.2 The Second Method: Physical Layer Security Assisted

Encryption Scheme in F-OFDM System

A new waveform frame was proposed by Huawei called filtered OFDM (f-OFDM) to
enhance spectrum usage in eMBB and enabling flexible waveform for 5G. The main
task of f-OFDM is to divide the bandwidth into multiple sub-bands so that different
services can be hosted in different sub-bands.
The F-OFDM inherited all the benefits of the OFDM, such as high spectral effi-
ciency and MIMO adaptability, but it has also solved some of its flaws by improving
flexibility and spectral efficiency. It becomes a key technology to cut the 5G air
interface [4].
In [3], the authors proposed a new encryption scheme assisted by physical layer
security which is applied to the f-OFDM system (Fig. 2).

3.3 The Third Method: Quantum Cryptography

Quantum cryptography, specifically quantum key distribution (QKD), is an example

of physical layer security. This method uses the quantum physical properties of
Security at the Application Layer and the Physical Layer in 5G 461

Fig. 2 The f-OFDM transceiver diagram

optical signals while ensuring high security. QKD is based on the principles of
quantum physics and is not based on mathematical or computational assumptions
[5].
This quantum cryptography makes it possible to distribute a secret encryption key
between two distant interlocutors, while ensuring the security of the transmission
thanks to the laws of quantum physics and information theory. This secret key can
be used afterwards in a symmetric encryption algorithm, to encrypt and decrypt
confidential data.
We use the quantum cryptography system to transmit a key, and not the message
itself for two essential reasons:
• The bits of information communicated by the mechanisms of quantum cryptog-
raphy can only be random. This is not suitable for a message, but perfectly suited
for a secret key, which must be random.
• Even if the mechanism of quantum cryptography guarantees that spying on
communication will always be detected, it is possible that the spy intercepts
bits of information before it is detected. This is unacceptable for a message,
but unimportant for a random key, which can be simply discarded in the event of
interception.
However, a smooth integration and wide adoption of QKD systems in telecom-
munications networks is an important step towards a practical and economical use of
quantum cryptography, which is still difficult due to technical and economic obstacles
[5]. Indeed, its physical implementation has been a problem for scientists. Several
462 H. Amgoune and T. Mazri

types of listening methods have been introduced, each with its own advantages and
limitations [6].

3.4 The Fourth Method

The broadcast nature of wireless media makes transmission vulnerable to eavesdrop-

ping. To ensure that information is transmitted securely, encryption techniques are
often used in the upper layers. We quote here, the symmetric cryptography methods
(for example, Advanced Encryption Standard) this method use a common private
key pre-shared between the source and destination ie the same key for encrypted and
decrypted messages, Unlike the symmetric method such as Public Key Cryptosys-
tems (PKC), asymmetric encryption methods also use public and private keys. The
public key can be communicated to everyone without any risk, this key is used for
encryption. Unlike the public key, the private key should never be communicated,
this key is used for decryption.
Symmetric cryptography has been used in current mobile communication, due to
its low computational cost compared to asymmetric cryptography (PKC). For new
generation wireless networks, such as 5G wireless networks, the key management
process (generation of keys and secure exchange key) will become even more impor-
tant as the number of nodes increases on a massive scale and the nodes become more
heterogeneous in their computing capacities. Also, in this article, a new method is
proposed to exchange a secret key between two legitimate users using the physical
layer.
The method proposed for these researchers is based on the following steps:
1. The transmitter transmits a training sequence to the receiver for the estimation
channel.
2. The transmitter estimates the channel and determines the transmission filter
which reverses the channel. There is now a “perfect” channel between the
transmitter and the receiver.
3. The transmitter sends the secret key to the receiver after having passed it via an
LDPC encoder.
4. The receiver receives the pre-equalized signal (containing the session key). The
signal then passes through an LDPC decoder.
5. Thus, the secure key sequence is transmitted from the transmitter to the receiver.
6. The receiver sends an ACK/NACK to the transmitter based on the syndrome
code. If the syndrome has errors, a NACK is sent to the transmitter and steps 1
to 5 are repeated. Otherwise, an ACK signal sent indicating that the session key
is established and the key exchange procedure is complete.
7. The exchange of keys for a secure transmission is now established between the
transmitter and the receiver. To transmit data securely, the generated secret key
is entered into random number generators (RNG).
Security at the Application Layer and the Physical Layer in 5G 463

3.5 The Fifth Method: Authentication

with Kolmogorov-Smirnov (K-S)

The future fifth generation (5G) wireless network is envisioned to be very heteroge-
neous and capable of supporting future applications requiring transparent coverage,
low latency and high speed. These requirements pose strict challenges in the design of
transfer authentication. In article [7], the researchers presented physical layer authen-
tication. They proposed to use transfer authentication with Kolmogorov-Smirnov (K-
S). To summarize, this work presents an authentication scheme based on the phys-
ical layer for SDN-compatible 5G HetNets. The proposed authentication scheme
consumes less computing and storage resources compared to the GLRT methods
according to the researchers [7]. Simulations have shown that the proposed K-S test
based on the scheme can provide reliable security performance for the purpose of
additional authentication [7].
On the other hand, researchers [8] proposed the Kolmogorov-Smirnov test for
user identification and an authorization procedure based on the typing dynamics
of users. In the method proposed by these researchers, the underlying time series
is preprocessed using an adaptive filter. They showed that the proposed method
performs much better than the basic classifier (1 NN DTW).
The researchers in the paper [9] proposed the digital radiofrequency (RF) tattoo
which is a new physical layer authentication technology, the carrier of which is
not traditional multimedia signals but digital signals. A blind watermark detection
scheme with Kolmogorov-Smirnov (KS) test is used.

3.6 The Sixth Method

Security-based power control and resource allocation technologies can adjust the
transmitter parameters to the instantaneous CIS, so that the SNR received from
legitimate users can be improved or maintained, while that of listening varies
randomly.
The perfect CSI of the legitimate channel [10] is not available on the transmitter
side due to a delay or an estimation error of the channel. To solve this problem,
researchers [10] used the knowledge contained in obsolete CSIs to decide whether
to transmit.
A channel amplitude quantification method that can use a single threshold value to
efficiently quantify the channel response is proposed in paper [11] in order to extract
a random manipulation sequence with good secrecy properties. More precisely, a
time division duplex (TDD) wireless system is considered to be independent iden-
tical distributed (iid) distribu Rayleigh fast discoloration channel (Rayleigh is a mode
of diffusion of waves, for example electromagnetic or sound, whose length wave is
much larger than the size of the diffusing particles. We speak of elastic scattering,
because this is done without variation of energy, in other words the wave keeps the
464 H. Amgoune and T. Mazri

same wavelength), where potential passive eavesdropping (Eves) can only estimate
their own channel and have no knowledge of the CSI between the legitimate commu-
nication parties. The transmitter (Alice) only knows the CSI of the legitimate user
(Bob). In particular, the security technique proposed by these researchers [11] takes
the bits of the transmitted data packets and manipulates them with a logical vector
which characterizes the randomness of the channel as a function of the estimated
CSI gain. The manipulation process is implemented at the bit level using an XOR
operation exactly before the modulation process. The same XOR operation is imple-
mented after the demodulation process on the detected bits to extract the hidden bits.
According to the researchers, the simulation results obtained show that the use of
such a mechanism can guarantee the confidentiality of the data.

3.7 The Seventh Method: Two Path Successive Relaying

Relay is one of the useful techniques for improving the security of the wireless phys-
ical layer. The existing literature shows that the use of a full duplex relay instead of a
conventional half-duplex relay improves the capacity of secrecy and the probability
of interruption of the secrecy.
In this article, the successive two-way relay (TPSR) was proposed in [12] to
improve the security of wireless transmission. We compare the achievable ergodic
secret rate, the probability of interception and the secret probability of interruption of
the proposed TPSR compared to half-duplex relay (HDR), full-duplex relay (FDR)
and full-duplex interference (FDJ) existing. The numerical results presented in this
paper reveal that the proposed TPSR achieves the same achievable ergodic confiden-
tiality rate as the FDR. The TPSR also offers the lowest probability of interception
and interruption of secrecy compared to the other schemes due to the efficiency of
the bandwidth in full duplex and the two mutually independent links assisted by the
relay pair R1 and R2.. The probabilities of interception of the TPSR and the FDR
are analyzed and verified simultaneously. The analysis shows that the probability
of interception of the TPSR is quadratic lower than that of the FDR. In short, with
the proposed TPSR protocol, secure wireless transmission can be achieved using
conventional half-duplex relays without the use of sophisticated interference and/or
interference cancellation techniques.

4 Discussion

The 5G network is expected to support massive user connections and an exponen-

tial increase in wireless services, making network security a new topic to study.
Unlike traditional security assurance techniques which rely heavily on cryptographic
approaches at the upper layer level. Security at the physical layer aims to strengthen
the confidentiality of data in communication systems. To prevent the decoding of
Security at the Application Layer and the Physical Layer in 5G 465

messages by unauthorized persons. In this paper we have cited several methods,

namely: authentication method EAP/AKA, physical layer security assisted encryp-
tion scheme in F-OFDM system, quantum cryptography (QKD: quantum key distri-
bution), LDPC encoder/LDPC decoder, transfert authentication with Kolmogov-
Smirnov (KS), CSI, and two paths successive relaying. The method which is based
on F-OFDM was proposed by Huawei. The concept of this technology is to divide
the bandwidth into several sub-bands so that different services can be hosted in
different sub-bands. On the other hand, other researchers [13] have proposed to
change the positions of the subcarriers on each OFDM symbol by following a prob-
ability distribution known only to legitimate transmitters and receivers. We find that
the method proposed by Huawei is more effective, especially since the researchers
[13] mentioned in their paper that the proposal would not work for applications that
canwork with low SNR values as voice transmission. There are also other techniques
like NOMA, which allows multiple users to share the same time and frequency band-
using overlay coding at the transmitter and successive interference cancellation (SIC)
at the receiver. In addition, NOMA can also lead to better performance in terms of
ergodic sum rate, equity, performance in the event of a breakdown, and energy effi-
ciency [14]. From the above discussions, we can see that research on the security of
the physical layer generated a vast literature, with subjects ranging from theoretical
studies of information to the design of practical plans. However, it does not remain
easy to develop a theory and innovative PLS transmission methods that correspond
well to the unique characteristics of 5G networks. The majority of existing PLS solu-
tions have drawbacks, which (prohibit) their applications to 5G networks. Thus, the
following method of securing the proposed physical layer must be respected: Rapid
user authentication, good reliability, various data rates, low latency and massive
connectivity to be effective for all 5G services, the MBB service which needs a high
speed, the URLL service which is sensitive to latency and finally the mMTC service
for IoT.

5 Conclusion

Nowadays, researchers pay Attention to 5G New Radio (NR) Physical Layer Security
Problems. This due to the introduction of all IP technology, the growing number of
subscribers, the interconnection of most networks to the Internet, and the needs of
high confidentiality into certain applications. The security of today’s smart devices
has become vulnerable to malware and other malicious software attacks, as these
devices are used in sensitive processes, such as banking, location, or even online
health applications.
Thus, further improvements and strengthening of security standards for 5G mobile
communications are needed. In recent years, there has been an increase in research
interest in the potential of the physical radio channel properties for providing secure
communications.
466 H. Amgoune and T. Mazri

These research efforts include discoloration, interference, and the diversity of

paths to develop safety techniques for implementation in the new 5G radio (NR).
Hence, we can say that the security of the physical layer has been considered as
a promising method to ensure a secure transmission of the wireless channels. The
main advantage of this method compared to cryptographic techniques is to exploit the
resources of the system, mainly the characteristics of the channel, such as multipath
propagation, without greatly increasing the complexity, compared to the security of
the upper layer. In addition, it can be used in conjunction with top-layer security
schemes and as a supplement to traditional cryptographic techniques.

References

1. Sun L, Du Q (2017) Physical layer security with its applications in 5G networks: a review.
China Commun 14:14
2. Mazin A, Davaslioglu K, Gitlin RD (2017) Secure key management for 5G physical layer
security. In: 2017 IEEE 18th wireless and microwave technology conference (WAMICON),
Cocoa Beach, FL, USA, April 2017, pp 1–5. https://doi.org/10.1109/WAMICON.2017.793
0246
3. Pan F, Jiang Y, Wen H, Liao R, Xu A (2017) Physical layer security assisted 5G network security.
In: 2017 IEEE 86th vehicular technology conference (VTC-Fall), Toronto, ON, September
2017, pp 1–5. https://doi.org/10.1109/VTCFall.2017.8288343
4. Up in the air with 5G - Huawei Publications. https://www.huawei.com/en/about-huawei/pub
lications/communicate/80/up-in-the-air-with-5g. Accessed 25 May 2020
5. Aleksic S (2015) Towards fifth-generation (5G) optical transport networks. In: 2015 17th
international conference on transparent optical networks (ICTON), Budapest, July 2015, pp
1–4. https://doi.org/10.1109/ICTON.2015.7193532
6. Gupta S, Sau K, Pramanick J, Pyne S, Ahamed R, Biswas R (2017) Quantum computation
of perfect time-eavesdropping in position-based quantum cryptography: quantum computing
and eavesdropping over perfect key distribution. In: 2017 8th annual industrial automation and
electromechanical engineering conference (IEMECON), Bangkok, Thailand, August 2017, pp
162–167. https://doi.org/10.1109/IEMECON.2017.8079582
7. Ma T, Hu F, Ma M (2017) Fast and efficient physical layer authentication for 5G HetNet
handover. In: 2017 27th international telecommunication networks and applications conference
(ITNAC), Melbourne, VIC, November 2017, pp 1–3. https://doi.org/10.1109/ATNAC.2017.
8215379
8. Ceffer A, Levendovszky J (2016) Kolmogorov-Smirnov test for keystroke dynamics based user
authentication. In: 2016 IEEE 17th international symposium on computational intelligence and
informatics (CINTI), Budapest, Hungary, November 2016, pp 000105–000110. https://doi.org/
10.1109/CINTI.2016.7846387
9. Xie X, Xu Z (2020) Blind watermark detection based on K-S test in radio-frequency signals.
Electron Lett 56(1):30–32. https://doi.org/10.1049/el.2019.2478
10. Hu J, Yang W, Yang N, Zhou X, Cai Y (2016) On–off-based secure transmission design with
outdated channel state information. IEEE Trans Veh Technol 65(8):6075–6088. https://doi.org/
10.1109/TVT.2015.2477427
11. Hamamreh JM, Furqan HM, Ali Z, Sidhu GAS (2017) An efficient security method based on
exploiting channel state information (CSI). In: 2017 international conference on frontiers of
information technology (FIT), Islamabad, Pakistan, December 2017, pp 288–293. https://doi.
org/10.1109/FIT.2017.00058
12. Liau Q, Leow C, Ding Z (2016) Physical layer security using two-path successive relaying.
Sensors 16(6):846. https://doi.org/10.3390/s16060846
Security at the Application Layer and the Physical Layer in 5G 467

13. Efstathiou D (2019) A collaborative physical layer security scheme. Int J Electr Comput Eng
IJECE 9(3):1924. https://doi.org/10.11591/ijece.v9i3.pp1924-1934
14. Hojeij MR (2018) Resource allocation techniques for non-orthogonal multiple access systems,
p 190
Blockchain Based Security and Privacy
in VANETs

Sana Iqbal, Narmeen Zakaria Bawany, and Ayesha Zulfiqar

Abstract Vehicular Ad-hoc Network (VANET) is one of the emerging network

technologies leading to intelligent transportation systems. Numerous applications
are developed for VANETs to provide safety and comfort to drivers. VANET is a
special type of Mobile Ad Hoc Network (MANET) in which a temporary network
is created using vehicles’ own communication and processing equipment. The data
can be exchanged directly from vehicle to vehicle, if there is no communication
infrastructure. Consequently, this give rise to various security and privacy threats
jeopardizing the availability, confidentiality and integrity of data. Therefore, secu-
rity and privacy remain the foremost concern in VANETs applications. Recently,
blockchain technology is being explored to address security and privacy issues in
VANETs. This paper presents the security and privacy requirements for different
VANET applications. Moreover, various existing blockchain and non-blockchain
based solutions for security and privacy in VANETs are comprehensively reviewed.

Keywords Blockchain · Vehicular ad-hoc networks · Privacy · Security

1 Introduction

VANET is the leading vehicular network used for communication because its assis-
tances in trouble-free driving as it can reduce traffic congestion and recurrence of
accidents. Traditionally, data regarding road traffic can only be collected through
cameras and surveys. On the contrary, VANET facilitates real-time data collection.
This data can be combined to get the real-time image of the road. Numerous applica-
tions have been developed through VANETs but they are lacking security and privacy
[1]. Therefore, there is a need to improve the privacy and security of VANETs.

S. Iqbal (B) · N. Z. Bawany · A. Zulfiqar

Center for Computing Research, Department of Computer Science and Software Engineering,
Jinnah University for Women, Karachi, Pakistan
N. Z. Bawany
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 469
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_43
470 S. Iqbal et al.

Recently, blockchain has been adopted in various domains which can address
the basic security and privacy problems, especially for information reliability and
integrity as it is an immutable, transparent and irreversible database. It handles
program applications with the help of transaction trace reliably and conveniently.
Blockchain is a prominent and widely used application which is extensively
applicable in digital assets [2], distributed storage [3] and smart contracts [4].
Many surveys exist on blockchain [5, 6] that are addressing security and privacy
requirements in blockchain-based VANETs. These surveys emphasize on the appli-
cations and design. However, this research is based on the investigation of how
blockchain can be used to address various privacy and security requirements in
VANET.
Comparative study of various security and privacy requirements for VANET appli-
cations and blockchain based solutions are discussed. It emphasizes on constraints
and breaches of existing solutions and their impact on scalable scheme of blockchain
in VANETs.
This research work is arranged in 8 concise sections. Background of VANET and
applications are discussed in Sect. 2. Security and privacy requirements of VANET
are presented in Sect. 3. Section 4 discusses the constraints of existing VANET
solutions and their requirements. Section 5 explains the basic concepts of blockchain
and its usage. Section 6 discusses blockchain-based solutions addressing privacy and
security requirements of VANETs. Comparison of blockchain and non-blockchain
solutions is also presented in Sect. 6. Lastly, in Sect. 7 the paper is concluded.

2 Background

In this section emphases are made on the general concept of VANET and its appli-
cations while the next section discusses the fundamentals of blockchain and its
usage.

2.1 What is VANET?

VANET is a special type of Mobile Ad Hoc Network (MANET) which allows

digital communication between vehicles and roadside unit (RSU) to facilitate safety
management, traffic management and internet services. RSU is an access point
utilized along with the vehicles to permit data dispersal within the roads. A vehicle
is considered a smart node that is capable of communicating with its neighbours and
different vehicles in the networks. Various types of communication occur in VANET.
The exchange of data between vehicles is known as vehicle to vehicle (V2V) commu-
nication. V2V is widely used for supportive driving. The communication between
vehicles and fixed roadside infrastructure is known as (V2I) vehicle to infrastructure
Blockchain Based Security and Privacy in VANETs 471

communication. V2I provide updates related to real time traffic and weather condi-
tion. Many applications are developed for VANETs which can be divided into below
mentioned three categories:

Safety Applications. Safety applications include road monitoring to solve problems

such as traffic congestion and emergency conditions. It is handled through crash
control and hazard-control alerts so that other vehicles can change their ways. these
applications can store real-time traffic information on road side unit that is accessible
to the vehicles at any point. This plays an imperative part in tackling the issues such
as emergency in case of accidents; congestions and traffic jam [7].

Commercial Applications. Commercial applications include internet streaming,

web access and entertainment services. these applications help drivers to download
maps of the route while travelling to the new area for guidance. It also notifies
about restaurants and petrol pump services within their communication range. Such
applications can help particularly those suppliers who want to attract customers
towards their stores. It can work even offline [7].

Convenience Applications. These applications deal with route divergence facility

in case of road congestion, e-toll payment, notifies about available parking, aids
in prediction of the road topography [7]. The next section explains the various
requirements that need to be addressed while designing a VANET architecture.

3 Privacy and Security Requirements in VANET

Applications

Vehicles in VANETs can create self-organizing systems without any potential infor-
mation of other vehicle. Consequently, the security and privacy requirements of
VANETs ought to be fulfilled before its deployment. Table 1 summarizes the VANET
applications with respect to these requirements that are discussed below:

3.1 Privacy Requirements

The privacy requirements are thoroughly considered while creating a privacy

conserving design for VANETs. There are number of privacy requirements [8]
which are taken into account to fulfil the applications’ requirement. These privacy
requirements include Anonymity, Confidentiality, and Conditional Privacy, which
are defined as follows:

Anonymity. Anonymity is an essential characteristic where a vehicle Node and a

RSU can be connected without specifically stating vehicle’s genuine identity. A node
472 S. Iqbal et al.

sharing information should be anonymous due to privacy reasons. VANET needs to

hide the identity of the sender. Therefore, anonymity requirement is top most priority.

Confidentiality. Protecting the secrecy of data implies preventing unauthorized

access so the user’s privacy can remain intact. Essential information such as driver’s
name, license plate number, vehicle sensor data, speed and travelling route of the
passenger should be confidential [9].

Conditional Privacy. A system is required to be developed which allows anony-

mous messaging to the generic nodes. However, support is also required for identifi-
cation by the central authorities in rare cases such as road accidents [10]. Therefore,
the system must incorporate conditional privacy settings.

3.2 Security Requirements

Several security requirements need to be considered while developing applications

for VANET. These security requirements are critical for adaption of applications.
Some of the typical security requirements [11] are defined below:

Authentication. Authentication is one of the important features in VANET as it

provides the validity of the information communicated through the channel. In
VANETs, authentication must be implemented in a manner so that the system
can distinguish between authorized and unauthorized nodes as vehicles act upon
the information coming from other vehicles. Therefore, malicious information can
cause traffic congestion and accidents. Moreover, the authentication requirement will
remain a high priority in all applications.

Data Integrity. Keeping up the integrity of data implies that only authorized parties
can control the data legally. Attacks over the network can cause false modification
in information shared between nodes and it must be protected. This requirement will
also remain high in VANET applications.

Non-repudiation. It is the assurance of the data origin and integrity and does not
allow someone to deny from their act. Therefore, it is essential that information
transmission between nodes should be prevented from repudiation.

Availability. Resources are said to be available if they are accessible by autho-

rized parties under all conditions. Information received with delay becomes useless
therefore its availability should be ensured to legitimate nodes. In case of safety
applications, the delayed information would be meaningless. Therefore, data avail-
ability requirement will be high for safety applications while in commercial and
convenience applications delay can be tolerable but it will consume vehicle fuel and
time of drivers.
Blockchain Based Security and Privacy in VANETs 473

Table 1 VANET applications requirements

Requirements Applications
Safety Convenience Commercial
Privacy Anonymity High High High
Confidentiality High High High
Conditional privacy Medium Medium Medium
Security Authentication High High High
Data integrity High High High
Data verification High High High
Non-repudiation High High High
Availability High High-Medium High-Medium
Mobility High High High
Key management Medium Medium Medium

Data Verification. The data accumulated should be verified to guarantee that infor-
mation that is assembled is as precise as possible. Therefore, information must be
verified in VANET on regular basis to eliminate the fake data. Fake information
would mislead drivers that can lead to severe accidents or wastage of time and fuel
[12].

Mobility. Mobility is the capability to move freely. highly mobile nodes in VANET
entail less execution time of security protocols therefore learning based method
should be developed so that nodes can evolve to recognize each other’s behavior.
Mobility will be highly required in all applications so that vehicle can easily leave
or join the network for getting timely information [13].

Key Management. Key management is the set of operations that are fundamental
to create, maintain, secure, and control the usage of cryptographic keys. In VANETs,
key creation for allocation should be managed especially [13].

In the next Section various approaches are discussed that are dealing with the
above-mentioned requirements.

4 Non-blockchain Based Solutions

Different approaches for VANET security and privacy are presented below:
474 S. Iqbal et al.

4.1 Pseudonym/ID Based Authentication Schemes

S. J. Horng et al. [14–16] proposed a pseudonym-based scheme known as an identity-

based scheme (IBS) which was incorporated with a batch verification process. To
attain conditional privacy, nodes in communication use pseudo identities that are
retrievable to authorities. Moreover, IBS is less efficient than traditional signature
schemes. R. Hussain et al. [17] offered a solution using pseudonyms mechanism
to preserve privacy in a vehicle infrastructure that is dependent on the centralized
system for key management. In such a system, identities can be associated with
background information as pseudonyms do not give absolute privacy protection.
Tzeng et al. [18] presented a solution using ID authentication schemes based on
bilinear pairing, that guarantees and fulfils the privacy requirements of VANET.
Although, performance of those schemes is not as good as bilinear operation. It
lacks authentication, which should be implemented in VANETs. Manvi and Tangade
[19] presented various authentication schemes which make efforts in resolving the
secure authentication schemes which make efforts in resolving the security issues
in VANET, although they have some limitations too. Mostly pseudonyms schemes
are based on public key infrastructure (PKI) that causes delay in communication and
signature verification due to huge computational overhead.

4.2 Group Signature Scheme

Han et al. [20] proposed a protocol that handles efficiency and security problems.
It improves the efficiency of certification by using group based key broadcasting
technique without certification authority (CA) which an authorized vehicle from
being attacked using a deniable technique known as group key contract. However,
it is not a scalable scheme as it requires each vehicle to verify others throughout
the key contract resulting in a delay in verification. Conditional privacy scheme was
proposed by Xiaoyan Zhu [21] which is based on group signature. This scheme is
aimed at attaining conditional privacy through a hash chain technique which reduces
computational overhead. Group manager retains the key and every participant signs
the message on the behalf of the group. In this case, group signature verification is
more costly than traditional signature. Moreover, canceling compromised individuals
appropriately is still an issue. Kanchan and Chaudhari [22] presented a method based
on group signature combined with re- encryption method. In this method, a message is
broadcasted and is available to all the individuals of the group by using re-encryption
method. Nonetheless, privacy can be compromised in such method.
Blockchain Based Security and Privacy in VANETs 475

4.3 Cryptography Based Schemes

P. Asuquo et al. [23] discussed possible solutions for retaining location privacy based
on the use of cryptography for retaining the users. The approach is based on central-
ized method by which a central authority is accountable for forcing the secrecy
strategies. This authority can be the key dispersion framework inside a cryptographic
framework that is accountable for limiting the framework to the authentic clients.

4.4 Message Authentication Code Schemes

C. Lyu. [24] proposed an efficient message authentication and verification scheme

based on MAC message authentication code. However, message authentication
required the RSU support as vehicle cannot authenticate the received message
independently.

4.5 Broadcasting Schemes

Receiver Consensus (ReC) based broadcasting scheme have been proposed by J.

Liu [25] for dissemination of cautionary messages to nodes in VANET. The vehi-
cles autonomously are assigned a priority to broadcast the message. However, it
causes delay in communication because of contention in medium access. W. Zeng
[26] focused on smart bus gateways and in-vehicle communication security through
observing data transmission between vehicles. They have not given security mech-
anisms for vehicle connecting to the Internet. These all schemes are summarized in
Table 2.
Next section contains domain knowledge about blockchain technology and its
integration with VANET.

5 Blockchain and Its Applications

Blockchain is a chain of records, called blocks, linked through cryptography. For each
block, a hash is generated with timestamp with transaction data of preceding block
[27]. It controls a dispersed record to permit transactions among nodes efficiently in
the network. It does not allow data modification and guarantees transaction integrity
[28]. It is a smart method of data transfer from A to B in completely safe and
mechanized way. Transaction process get started by one party that generates a block.
The block gets proved by millions of systems spread around the internet and is later
added to the chain that is kept across the network. This generates a record with unique
476 S. Iqbal et al.

Table 2 Existing solutions for VANETs security and privacy requirements

Schemes References Addressed security Addressed privacy
requirements requirements
Group signature scheme [20] Authentication None
[21] Authentication None
[22] Authentication None
Pseudonym /ID based [14, 15, 16] None Conditional privacy
authentication schemes [17] None Anonymity
[18] None Anonymity
[19] Authentication, None
Non-Repudiation
Data integrity
Broadcasting schemes [25] Availability None
[26] Authentication None
Message authentication [24] Authentication None
code schemes
Cryptography based [23] No Anonymity
schemes

history. Therefore, forging a record would mean that forging the whole chain, which
is impossible in such Scenario. The possible use cases of blockchain technology are
given below [29].

5.1 Currency

The first application that was introduced is known as Bitcoin. It is a digital currency
that allows safe transactions without central banking. Blockchain technology is
expected to use decentralization to advantage industries and developments where
trust level is required.

5.2 Finance

The utmost extensively known blockchain application lies in finance field, it guar-
antees much valued transparency among the transactional parties. Each transaction
in community, stocks, bonds or derivatives could be recorded in blocks and later be
confirmed by the authority for its legality. It’s simpler to identify money laundering
or extortion cases through stock trade moves.
Blockchain Based Security and Privacy in VANETs 477

5.3 Voting

Voting systems remain in several states a questionable matter, due to occurrence of

appointive pressure (invalid or wrong vote, different enrollment) and the enormous
rate of abstention often shape the ultimate result. With blockchain, each vote that is
tallied has to be entirely and precisely confirmed. This implies that authorities can
affirm the voting outcomes after they are ended.

5.4 Smart Contracts

Smart contracts are introduced with self-execution program. In the real-world

scenario, a mediator guarantees that all parties carry through all the stated terms and
conditions. The blockchain not only delay the requirement for third parties, but also
guarantees that all record contributors know the agreement elements which legally
bind terms and execute naturally once conditions are met. For all sort of conditions,
smart contracts can be used such as insurance payments, property regulation and
economic products.

5.5 Identity

The blockchain protects a person’s identity by encrypting it and securing it from

intruders. A person carries range of IDs which includes license, system password,
ID cards etc. Blockchain ID changes all these forms of identification. It’s vital to
note that for the blockchain to work, the node-to-node system must be encouraged
and agree to work under ethical standards. Once these guidelines are followed, the
blockchain may get to be a capable tool for making strides commerce, conducting
reasonable exchange, democratizing the universal economy and serving more fair
societies. Blockchain based solutions are recently being developed to address the
security and privacy issues in for various applications. Researchers have explored
blockchain based solutions for VANETs security and privacy. Next sections briefly
describe such solutions.

6 Blockchain Based Solutions

Centralized schemes for VANETs cannot handle the increasing complexity of intel-
ligent transportation system (ITS) structures. The speedy evolution of Internet of
Vehicles (IoV) has provided large demanding situations for huge information storage,
478 S. Iqbal et al.

asset management, and information and transaction security. Recently, security and
privacy issues of VANETs are addressed through variety of blockchain mechanisms
within the literature few of them are discussed in this section. Z. Lu [30] proposed
a system based on public key structure to create a trust-based privacy preserving
model for VANETs. This study is focused on the methods based on privacy and
summarized three trust models to establish trust management in VANETs. N. Malik
[31] proposed a frame- work based on authentication using blockchain in VANETs
which covered only privacy concerns and ultimately failed to deal with security
of nodes communication. R. Shrestha et al. [32, 33] proposed a solution for nodes
communication security based on crypto trust point. However, they failed to address
privacy issues. A trust-based framework has been proposed by M. Singh [34] using
blockchain technology. It provides the trust environment based on proof of driving to
support information sharing and secure communication among vehicles. Lilli et al.
[35] provided an incentive announcement network based on blockchain technology.
It was proposed to guard vehicles’ privacy and encourage users to share traffic statis-
tics. However, location privacy is not considered in this model and the network
idleness to deliver data and to inform the RSI isn’t assessed. A blockchain-based
framework was proposed by A. Dorri [36] to report privacy and security of smart
vehicles. The research presents several cases related to remote program updates as
well as adaptable automotive insurance plans. They have neither given technical
details of their framework nor address the issues. X. Zhang [37] Emphasized the
extent of statistics generated via vehicular adhoc network and strain the significance
of Mobile Edge Computing (MEC). It was designed to compensate utilization of
resources in VANET based on blockchain. It helps in lowering the computational
overhead of blockchain. However, the MEC does no longer make it decentralized.
A secure system proposed by X. Zhang and X. Chen [38] for storage and data
sharing based on the blockchain consortium in VANETs. It can address the identity,
validity and message reliability issues but in result it generates additional overhead.
A secure key management framework based on blockchain has been proposed by A.
Lei [39] for the heterogeneous ITS systems. It gives good security in communica-
tion between vehicles and infrastructure. But it fails to conserve the vehicle’s privacy
which if exposed, can locate and put it at possible risk. A blockchain based decentral-
ized trust management system have been proposed by Z. Yang [40] in the vehicular
networks. In this, vehicle allot a positive vote to the neighbouring trusted vehicles
and upload the rating to the Roadside Units which are responsible for calculating
the trust score and set it in trusted blockchain. It has not provided vehicles’ vote
privacy mechanism and doesn’t consider the trustworthy weight of partaking vehi-
cles. A point-to-point trans- action model presented by J. Kang [41] is grounded on
the blockchain consortium that resolves the issue of scalability. It doesn’t completely
ensure the data exchange procedure. Trust point mechanism have been presented by
Singh and Kim [42] to secure the communication among vehicles. For data sharing
such methods couldn’t work in vehicle-based edge computing. It is because launching
a blockchain in resource limited vehicles is costly. Table 3 gives the summarize view
of the above discussed blockchain solutions with respect to VANET security and
privacy.
Blockchain Based Security and Privacy in VANETs 479

Table 3 Existing blockchain based solutions for security and privacy requirements
Year Paper reference Addressed security Addressed privacy requirements
requirements
2018 [30] No Anonymity
2017 [36] Data integrity Anonymity, Confidentiality
2018 [35] No Anonymity, Conditional privacy
2018 [31] Authentication No
2018 [32] Data verification, No
Authentication
2018 [33] Authentication, Data integrity No
2018 [34] Authenticity Anonymity
2018 [37] Data integrity No
2019 [38] Authentication, Data integrity, Anonymity
Availability, Data verification,
Non-repudiation
2017 [39] Key management No
2018 [40] Authentication, Data integrity No
2017 [41] Authentication, Data Conditional privacy
verification Data integrity, Key
management
2018 [42] Authentication, Availability No

Table 4 compares the blockchain and non-blockchain based solutions of VANET.

It is clear that existing non-blockchain solutions are ensuring either privacy or secu-
rity. Whereas, on the other hand, some blockchain based solutions do exist that are
achieving both security and privacy simultaneously. However, security and privacy
in VANETs remains an open research area and there is lot to be done to provide
pragmatic security and privacy.

Table 4 Comparison of solutions according to requirements

Solutions Privacy Security
Non-Blockchain [19–22, 24–26] Low High
[14–18, 23] High Low
Blockchain [30, 35] High Low
[34, 36, 38, 41] High High
[31–33, 37, 39, 40, 42] Low High
480 S. Iqbal et al.

7 Conclusion

This research has concluded that blockchain is one of the most innovative tech-
nologies that can address the various security and privacy requirements in VANET.
Initially, VANETs security and privacy requirements with respect to various applica-
tions are discussed. Secondly, existing security and privacy solutions for VANET are
presented. Further, this paper explained the integration and working of blockchain
with VANET and highlights how blockchain based solutions can address the secu-
rity and privacy requirements of VANET. Blockchain gives superior adaptability in
getting to the information and it gives better security and privacy in communication
between nodes. Lastly, this paper will be helpful in bringing up the studies related to
blockchain in VANET. This paper will provide fundamental thought to understand
the necessity of blockchain in VANET.

References

1. Abdelatif S, Derdour M, Ghoualmi-Zine N, Marzak B (2020) VANET: a novel service for

predicting and disseminating vehicle traffic information. Int J Commun Syst 4288
2. Dorri A, Kanhere SS, Jurdak R (2017) Towards an optimized blockchain for IoT. In: IEEE/ACM
second international conference on internet-of-things design and implementation (IoTDI),
Pittsburgh, pp 173–178
3. Betts B (2016) Blockchain: the missing link in cloud security. Computer Weekly
4. Christidis K, Devetsikiotis M (2016) Blockchains and smart contracts for the Internet of Things.
IEEE Access 4:2292–2303
5. Tschorsch F, Scheuermann B (2016) Bitcoin and beyond: a technical survey on decentralized
digital currencies. IEEE Commun Surv Tutor 18(3):2084–2123
6. Crosby M, Pattanayak P, Verma S (2016) BlockChain technology: beyond bitcoin. applied
innovation review. Pantas and Ting Sutardja Center for Entrepeneurship Technology, Berkeley
Engineering, UC Berkeley, p 2
7. Kumar V, Mishra S, Chand N (2013) Applications of VANETs: present future. Commun Netw
5
8. Mansour MB, Salama C, Mohamed HK, Hammad SA (2018) VANET security and privacy-an
overview. Int J Netw Secur Appl 10:13–34
9. Xie L, Ding Y, Yang H, Wang X (2019) Blockchain-based secure and trustworthy internet of
things in SDN-enabled 5G-VANETs. IEEE Access 1
10. Guerrero-Ibáñez J, Zeadally S, Contreras-Castillo J (2018) Sensor technologies for intelligent
transportation systems. Sensors 18:1212
11. Shringar RR, Kumar M, Singh N (2013) Security challenges, issues and their solutions for
Vanet. Int J Netw Secur Appl 5:95–105
12. Chaubey NK (2016) Security analysis of vehicular ad hoc networks (VANETs): a comprehen-
sive study. Int J Secur Appl 10(5):261–274
13. Zaidi K, Rajarajan M (2015) Vehicular internet: Security privacy challenges and opportunities.
Future Internet 7:257–275
14. Horng SJ, Tzeng SF, Pan Y, Fan P, Wang X, Li T, Khan MK (2013) B-SPECS+: Batch veri-
fication for secure pseudonymous authentication in VANET. IEEE Trans Inf Forensics Secur
8(11):1860–1875
15. Liu J, Yu Y, Zhao Y, Jia J, Wang S (2018) An efficient privacy preserving batch authentication
scheme with deterable function for VANETs. In: 12th international conference on network and
system security, Hong Kong, China, pp 288–303
Blockchain Based Security and Privacy in VANETs 481

16. He D, Zeadally S, Xu B, Huang X (2015) An efficient identity-based conditional privacy-

preserving authentication scheme for vehicular ad hoc networks. IEEE Trans Inf Forensics
Secur 10:1
17. Hussain R, Kim D, Son J, Lee J, Kerrache CA, Benslimane A, Oh H (2018) Secure and
privacy-aware incentives-based witness service in social internet of vehicles clouds. IEEE
Internet Things J 5(4):2441–2448
18. Tzeng SF, Horng SJ, Li T, Wang X, Huang PH, Khan MK (2017) Enhancing security and
privacy for identity-based batch verification scheme in VANETs. IEEE Trans Veh Technol
66(4):3235–3248
19. Manvi SS, Tangade S (2017) A survey on authentication schemes in VANETs for secured
communication. Vehicular Commun 9:19–30
20. Han M, Hua L, Ma S (2016) A self-authentication and deniable efficient group key agreement
protocol for VANET. KSII Trans Internet Inf Syst 11
21. Zhu X, Jiang S, Wang L, Li H, Zhang W, Li Z (2013) Privacy-preserving authentication based
on group signature for VANETs. In: IEEE Global communications conference (GLOBECOM),
Atlanta, GA, pp 4609–4614
22. Kanchan S, Chaudhari NS (2016) Integrating group signature scheme with non-transitive proxy
re-encryption in VANET. In: International conference on computing, analytics and security
trends (CAST), Pune, pp 227–231
23. Asuquo P, Cruickshank H, Morley J, Ogah CPA, Lei A, Hathal W, Sun Z (2018) Security
and privacy in location-based services for vehicular and mobile communications: an overview,
challenges, and countermeasures. IEEE Internet Things J 5(6):4778–4802
24. Lyu C, Gu D, Zeng Y, Mohapatra P (2015) PBA: Prediction-based authentication for vehicle-
to-vehicle communications. IEEE Trans Dependable Secure Comput 13:1
25. Liu J, Yang Z, Stojmenovic I (2013) Receiver consensus: on-time warning delivery for vehicular
adhoc networks. IEEE Trans Emerg Top Comput 1(1):57–68
26. Zeng W, Khalid MAS, Chowdhury S (2016) In-vehicle networks outlook: achievements and
challenges. IEEE Commun Surv Tutor 18(3):1552–1571
27. Leible S, Schlager S, Schubotz M, Gipp B (2019) A review on blockchain technology and
blockchain projects fostering open science. Front. Blockchain
28. Peng C, Wu C, Gao L, Zhang J, Yau KLA, Ji Y (2020). Blockchain for vehicular internet of
things: recent advances and open issues. Sensors (Switzerland)
29. Tomar R, Prateek M, Sastry GH (2016) Vehicular Adhoc network (VANET)-an introduction.
Int J Control Theory Appl 9(18). International Science Press
30. Lu Z, Liu W, Wang Q, Qu G, Liu Z (2018) A privacy-preserving trust model based on blockchain
for VANETs. IEEE Access 6:45655–45664
31. Malik N, Nanda P, Arora A, He X, Puthal D (2018) blockchain based secured identity authen-
tication and expeditious revocation framework for vehicular networks. In: 17th IEEE interna-
tional conference on trust, security and privacy in computing and communications and 12th
IEEE international conference on big data science and engineering (Trustcom/BigDataSE),
New York, pp 674–679
32. Shrestha R, Bajracharya R, Nam SY (2018) Blockchain-based message dissemination in
VANET. In: IEEE 3rd international conference on computing, communication and security
(ICCCS), Kathmandu, pp 161–166
33. Singh M, Kim S (2018) Crypto trust point (cTp) for secure data sharing among intelligent
vehicles. In: International conference on electronics, information and communication (ICEIC),
pp 1–4
34. Singh M, Kim S (2018) Branch based blockchain technology in intelligent vehicle. Comput
Netw 145:219–231
35. Li L, Liu J, Cheng L, Qiu S, Wang W, Zhang X, Zhang Z (2018) CreditCoin: a privacy-
preserving blockchain-based incentive announcement network for com- munications of smart
vehicles. IEEE Trans Intell Transp Syst 19(7):2204–2220
36. Dorri A, Steger M, Kanhere SS, Jurdak R (2017) BlockChain: a distributed solution to
automotive security and privacy. IEEE Commun Mag 55(12):119–125
482 S. Iqbal et al.

37. Zhang X, Li R, Cui B (2018) A security architecture of VANET based on blockchain and mobile
edge computing. In: 1st IEEE international conference on hot information-centric networking
(HotICN), Shenzhen, pp 258–259
38. Zhang X, Chen X (2019) Data security sharing and storage based on a consortium blockchain
in a vehicular adhoc network. IEEE Access 7:58241–58254
39. Lei A, Cruickshank H, Cao Y, Asuquo P, Ogah CPA, Sun Z (2017) Blockchainbased dynamic
key management for heterogeneous intelligent transportation systems. IEEE Internet Things J
4(6):1832–1843
40. Yang Z, Yang K, Lei L, Zheng K, Leung VCM (2018) Blockchain-based decentralized trust
management in vehicular networks. IEEE Internet Things J 6(2):1495–1505
41. Kang J, Yu R, Huang X, Maharjan S, Zhang Y, Hossain E (2017) Enabling localized peer-to-
peer electricity trading among plug-in hybrid electric vehicles using consortium blockchains.
IEEE Trans Industr Inf 13(6):3154–3164
42. Singh M, Kim S (2018) Trust Bit: reward-based intelligent vehicle commination using
blockchain paper. In: IEEE World Forum on Internet of Things (WF-IoT), Singapore, pp 62–67
Dynamic Data Protection for Mobile
Agents: XACML/ABAC Policies

Hassan Razouki

Abstract A mobile agent can move from one platform to another during runtime
to access data or resources. He moves with his own code and data. Agents take on
two roles in the runtime environment. First, agents are treated as subjects, because
agents access the different resources of a platform. Second, they are considered as
active resources in the execution environment. The protection of platform resources
against the attack of malicious mobile agents is already addressed in the literature.
However, the protection of code and agent data is an open and difficult issue due to
the fact that the execution environment has full control over the mobile agent.
A malicious platform may attempt to attack a mobile agent in order to analyze or
modify the data collected (Dynamic Data) by this agent while it is running on another
platform. To solve this problem, we have developed a security mechanism to manage
access to the agent’s dynamic data between the platforms visited by this agent. This
mechanism is implemented by the SAS (Security Adaptation System) which allows
defining the access control rules to the dynamic data of the mobile agent. The SAS
estimates the security level of each agent component (agent code in the form of
software components) and the degree of confidence of the platforms visited by this
agent. It then uses the multi-level model to define these rules. We adapt the ABAC
model and the XACML language to create and publish access control policies for
dynamic mobile agent data based on these rules.

Keywords Mobile agents · Security · Security policy · XACML language ·

Attribute-Based Access Control (ABAC)

1 Introduction

Most applications on the Internet require interaction between different entities across
the network to exchange data and distribute tasks. This increases the need for
autonomous entities capable of solving complex problems related to distributed

H. Razouki (B)
Faculty of Sciences and Technologies, University Sultan Moulay Slimane, Beni Mellal, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 483
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_44
484 H. Razouki

systems. In this context, researchers are increasingly interested in the use of mobile
agent technology as it brings significant gains for several application areas, such
as Wireless Sensor Networks (WSN) [1], Cloud Computing [2], e-commerce [3],
information search [4], telecommunications [5], web services [6], transport system
optimization [7] and many other areas.
In particular, the safety issue is a barrier to the expansion of this technology [8,
9]. When an agent arrives on an execution platform, it exposes its code and data.
As a result, this platform will have the ability to modify them. It can modify the
agent’s data or the results obtained by the mobile agent within previous platforms
(dynamic data). The owner of an agent does not want an untrusted platform to be
able to modify the agent or the results he has obtained on other platforms [9]. In this
case, it is necessary to encrypt sensitive data in a way that only authorized platforms
can read or modify the data. The problem we encountered is how to secure access to
the agent’s encrypted dynamic data between the platforms visited by the agent.
To solve this problem, we have developed a mobile agent model based on a
software component [10]. Then, we associated with each agent a security policy
defined by the agent creator [11]. This policy contains a set of security requirements
that must be met by the visited platforms such as encryption and data signing. To
protect the sensitive dynamic data of the mobile agent, we thought to encrypt these
data by the visited platform (the encryption algorithm used is determined by the
agent’s security policy) [11], the encryption key is generated and stored by the KDC
server (Key Distribution Center).
To manage access to the decryption key between platforms, we have devel-
oped a Security Adaptation System (SAS) to create access control policies for the
agent’s encrypted dynamic data (it manages access to the decryption key). To create
these policies, the SAS estimates the security level of the mobile agent components
according to the security policy defined by the agent creator [11]. Then, it uses
access control constraints to create access control rules for agent data based on the
Bell and La Pudula multilevel model [12]. These rules are then transformed into
XACML/ABAC access control policies. These policies manage access to dynamic
data between the platforms visited by the agent.
The remainder of this document is organized as follows. In Sect. 2, we present the
architecture of our protection system. In Sect. 3, we present the estimated security
levels of the software components of a mobile agent and the confidence levels of
the platforms visited by the mobile agent to create the access control rules for the
dynamic data of this agent. Section 4, presents how to protect the agent’s dynamic
data during its execution. Section 5, presents the performance of this solution against
different types of attacks from a malicious platform. Finally, Sect. 6, concludes the
document and proposes directions for future work.
Dynamic Data Protection for Mobile Agents … 485

2 General System Architecture

Our architecture is based on three phases: creation, publication, and use of access
control policies to protect mobile agent data. The interaction between the different
entities is illustrated in Fig. 1.
Our system consists of four components:
Server PAP (Policy Administration Point) allows us to create different XACML
policies for a mobile agent. The PAP stores the policies and makes them available
to the PDP. These policies represent the complete management policy that controls
the decisions made by the PDP.
PDP Server (Policy Decision Point) is the entity that manages authorization
requests from the KDC server, it is in charge of determining the rules and poli-
cies applicable to a given request. After evaluating the targets and conditions of
the ruleset, the PDP returns the response to the PDP. The PDP server relies on the
Certification Authority (CA) to verify the validity of the certificates of the different
entities.
Server LCA (Local Certificate Authority) is a component of PKI (Public Key
Infrastructure) that allows us to create and publish signed digital certificates
containing the public key and identity of each entity in our infrastructure. Digital
certificates are used to authenticate different entities in the system.

Fig. 1 Interaction between the different entities

486 H. Razouki

KDC (Key Distribution Center) is used by our system to protect agent data during
its execution. Therefore, only platforms authorized by the PDP server can access
encrypted mobile agent data. The KDC contains two components:
• Policy Enforcement Point (PEP) is a component that makes authorization requests
to the PDP, it enforces the authorization decision made by the PDP. It is the PEP
that technically performs the assignment or denial of symmetric keys to platforms
to encrypt its data or decrypt data from other platforms.
• Key Distribution Service (KDS) allows to intercept requests and verify the identity
of key requesters. The KDS uses a database to manage symmetric keys distributed
to platforms. The KDS component performs three functions: key creation, key
distribution, and key access control. It uses PEP to retrieve the authorization
decision from PDP.

3 Creation of Agent Data Access Control Policies

In this section, we propose a new technique for creating access control rules for
mobile agent data access control using the Security Adaptation System (SAS). To
do so, we have estimated the security levels of the agent’s software components and
the confidence levels of the platforms visited by the agent.

3.1 Estimate the Security Level of a Component (SLC)

We have adapted the multi-level model of Bell & La Padula (Bell and La Padula) to
create policies for controlling access to agent data. This security model is based on
the lattice concept that formalizes the security objectives and authorization scheme
of their policy, which can be used to demonstrate consistency.
Within the framework of a multi-level settlement, we have associated security
levels to the different components of the agent according to its security policy, and
degrees of confidence to the platforms running the agent. We defined the security
level of the components as follows:
Security Level of a Component (SLC) is the quadruplet (SR1, SR2, SR3, SR4)
of Security Requirements defined by the agent creator in the form of a security
policy. These requirements are classified according to four ordered security levels
(top secret, secret, confidential, non-confidential). The SLC reflects the sensitivity
of code and data produced by this component in a platform.
The principle of our approach is to assign a classification to each Security Require-
ment (SR). These requirements are derived from the agent’s security policy to protect
a particular component. There are four hierarchical security levels in order: top secret
(TS = 3), secret (S = 2), confidential (C = 1), and non-confidential (NC = 0).
The highest value of a requirement is considered the most secure. The four secu-
rity requirements are supported in the process of estimating the security level of a
Dynamic Data Protection for Mobile Agents … 487

functional component of the mobile agent. Table 1 below gives the classification of
security requirements for estimating the security level of a component.
Formally, we present the security level of a component as the set of security
requirements described in Table 1. These requirements are classified according to
the security policy defined by the agent creator. It is assumed that a mobile agent in

Table 1 Classification of security requirements

Non-confidential Confidential C = 1 Secret S = 2 Top secret TS = 3
NC = 0
Security The agent creator The base-level This description This description
requirement 1 does not specify a description also contains also describes the
base level specified by the information about data structure to
description for agent creator the functionality be added in the
this component describes only the of this component agent after the
services provided execution of this
and required by component by
this component platforms
Security The agent’s Agent security The agent’s policy The agent’s policy
requirement 2 security policy policy requires a requires strong requires strong
does not require single sign-on to authentication authentication
any authenticate with a digital with a digital
authentication to runtime platforms certificate certificate issued
run the (username and or validated by
component password) the local
certification
authority (LCA)
Security The security The agent’s The component is The component is
requirement 3 policy does not security policy encrypted with a encrypted with
require any requires symmetric key the public key of
encryption for the symmetrical (3DES, AES …). the platform
component encryption of this The symmetric running the
component key is enveloped component, in
(3DES, AES, …). with the public this case, only
The key is key of the one platform can
distributed platforms running decrypt and run
between the the component the component
platforms running
the component
Security The security The security The security The security
requirement 4 policy does not policy requires policy requires policy requires
require any symmetrical symmetrical the encryption of
encryption of data encryption of data encryption and collected data
collected by the collected by the signature of the with the agent
component component. The data collected by owner’s public
symmetric key is the component key. Only the
distributed to the agent owner can
platforms decrypt this data
with his private
key
488 H. Razouki

the adaptation phase contains n functional components. The following quadruplet is

used to estimate the security level of the components (Eq. 1):

∀i ∈ n SLC i = (S R1i , S R2i , S R3i , S R4i ) (1)

• SLC: represents the Security Level of Component i.

• SR1i , SR2i , SR3i and SR4i : represent the Security Requirements classification of
a component i.

After the SAS has estimated the safety levels of the components, the next step is
to determine the degrees of confidence of the platforms visited by the agent.

3.2 Estimate the Degree of Confidence in the Platforms

Our mobile agent consists of a set of functional components and a security policy
to be followed during its execution. Each component runs on one or more platforms
depending on the agent itinerary. To manage access control to agent data between
platforms, the SAS defines a set of access control rules based on the security level
of the components and the degree of confidence of the visited platforms. We defined
the degree of confidence of a platform as follows:
Degree of confidence of a platform (DCP) is the quadruplet (DA1, DA2, DA3,
DA4) of Degrees of Authorization (DAi) estimated from one or more components
executed by that platform. Each Degree of Authorization (DAi) takes the maximum
value of the security requirement classification of the components executed by that
platform.
To maintain the confidentiality of data collected from execution platforms. An
agent contains n components (C) and is executed by m platforms (P). The SAS
analyses the agent’s security policy to estimate the security level of each component
(SLC). Then, it estimates the degree of confidence of each platform (DCP) from the
components executed by this platform. The DCP is expressed by Eq. 2:
⎧
⎪ ∀ j ∈ [1, m] ∃ A j ⊂ [1, n]/∀i ∈ A j the component Ci is run in P j
⎪
⎪
⎪
⎪ SLC i = (S R1i , S R2i , S R3i , S R4i )
⎪
⎨
DA1 j = max(SR1i ); DA2 j = max(SR2i )
i∈A j i∈A j (2)
⎪
⎪
⎪
⎪ DA3 = max (SR3 ); DA4 = max (SR4i )
⎪
⎪
j i j
⎩ i∈A j i∈A j

DCP j = DA1 j , DA2 j , DA3 j , DA4 j

At the end of this step, SAS estimates the degree of confidence of the different plat-
forms and then uses the multilevel model to create the rules for controlling access to
agent data. The security levels and degrees of confidence form a lattice that is partially
ordered by a dominance relationship denoted “≤”. The protection of data confiden-
tiality is expressed by the fact that a platform P has a DCP degree of confidence.
Dynamic Data Protection for Mobile Agents … 489

This platform can only access data collected by component C in another platform if
the degree of confidence of the platform (DCP) dominates the security level of this
component (SLC). There are three constraints for creating rules to control access to
its data:
Constraint 1: If a Pj platform is allowed to run a Ci component, then the Pj platform
has the right to add encrypted data to the agent after running this component.
Constraint 2: A Pj platform wants to read the data collected by a Ck component
only if the two following conditions are verified:
• The sum of all security requirements of component Ck must be less than or equal
to the sum of all platform authorization degrees Pj .
• The security requirement relates to the confidentiality of the data produced by the
component Ck must not be classified as “Top secret = 3”.
This constraint is expressed by Eq. 3:
⎧ 4
⎨ 
4
SRi (Ck) ≤ DAi (Pj)
(3)
⎩ i=1 i=1
SR4 (Ck) < 3

Constraint 3: A Pj platform wants to modify the data collected by a Ck component

only if the following three conditions are verified:
• The sum of all security requirements of component Ck must be strictly less than
the sum of all platform authorization degrees Pj .
• The degree of confidence of the platform Pj dominates the security level of the
component Ck (Each component’s SRi security requirement must be less than or
equal to the corresponding platform’s authorization degree).
• The security requirement relates to the confidentiality of the data produced by the
component Ck must be classified as “non-confidential = 0” or “confidential = 1”.
This constraint is expressed by Eq. 4:
⎧ 4
⎪  
4
⎪
⎨ SRi (Ck) < DAi (Pj)
i=1 i=1
(4)
⎪
⎪ ∀i ∈ {1, 2, 3, 4} SRi (Ck) ≤ DAi (Pj)
⎩
SR4 (Ck) < 2

Finally, the SAS defines the rules for controlling access to agent data, based
on the degrees of confidence of the platforms and the security levels of the agent
components. Then, the SAS asks the PAP to create policies corresponding to these
rules. The PAP ensures that the access rules conform to XACML syntax and then
creates and stores the access control policies.
490 H. Razouki

4 Mobile Agent Data Protection

4.1 Protection of Data Collected on a Platform

Suppose that the first platform visited by the agent is PL-1, the agent route determines
the component that must be executed by this platform, as well as the security policy
used by the component [11]. After the agent authentication. Suppose the platform
wants to add data to the agent. In this case, the platform uses its Data Storage
Component (DSC) to add its data (XML format) to the running agent.
Assuming that the agent creator requires in its security policy the encryption of
the data produced by this component, the PL-1 platform uses the Access Control
Component (ACC) to send a request for a symmetric key to the KDS (The KDS is an
internal KDC server component). This request contains the identity of the platform,
the identity of the agent and the running component, the label of the data to be
encrypted, and the “Create” action. In our approach, we use three values for the
action attribute:
– Create: to request the creation of a new symmetric key to the KDS,
– Update and Read: are used to retrieve symmetric keys from KDS.
The platform’s request is signed by its private key. The KDS receives the request
and then authenticates the requestor of the key. Our solution is based on the access
control rules defined by the SAS at the time of agent adaptation. In this case, only
authorized platforms can add data to the agent. The KDS uses the PEP component to
send a SAML authorization request to the PDP, this request contains the necessary
attributes for decision making. The PDP determines the rules and policies applicable
to the PEP application. After evaluating the targets and conditions of the set of rules,
the PDP returns the authorization decision to the PDP. If the response is favorable
“Permit”, the PDP gives authorization to the KDS to generate a new symmetric key
specific to the PL-1 platform. The KDS stores the key in an array indexed by platform
identity. After registering the key, the KDS sends the generated key to the ACC, the
Data Storage Component (DSC) uses the key to encrypt this data. The used key is
then destroyed before the agent migrates to a new platform. Note that the key is later
used by the KDS to provide keys to authorized platforms to decrypt this data.

4.2 Access to Encrypted Data by Other Platforms

When the agent migrates from one platform to another, it moves with it the data of
its execution in the previous platforms, we take the example illustrated in Fig. 2.
The platform (PL-2) wishes to read or update the data created and encrypted by the
platform (PL-1).
After the migration of the agent from platform PL-1 to platform PL-2 (step 1).
The PL-2 platform authenticates the agent via AS and LCA servers (step 2). Then
Dynamic Data Protection for Mobile Agents … 491

Fig. 2 Encryption/decryption process of mobile agent data

it requests the symmetric key from KDS to decrypt the data encrypted by PL-1
(steps 3 and 4). This request contains the identity of the platform, the agent and
the component that produces the data in the PL-1 platform, the data label and the
action to be performed “Update” or “Read”. The KDS verifies the authenticity and
integrity of the request (step 5) and then redirects the request to the PDP. The PDP
then sends a request for SAML authorization to the PDP with the necessary attributes
for decision making (step 6.1). The PDP evaluates the rules and policies applicable
to the application (Step 6.2) and sends its decision to the PDP (Step 6.3). If the
decision is favorable, the PDP asks the KDS to search and send the symmetric key
to the platform (steps 7.1 and 7.2). The CCA retrieves the symmetric key and then
decrypts the data encrypted by the PL-1 platform (step 8). This illustration shows how
one platform decrypts data encrypted by other platforms, as well as the interaction
between the different entities to recover the symmetric key.

5 Protection of Dynamic Mobile Agent Data Against

Different Types of Attacks

In this section, we will discuss the security mechanisms used against different types
of attacks on dynamic mobile agent data.
• Dynamic data spying: Agent security policy requires encryption of data collected
by a component during its execution. In this case, the execution platform sends an
asymmetric key request to the KDC. If the platform is allowed to add the encrypted
data to the agent, the KDC generates a new key specific to that platform. The
platform uses this key to encrypt its data with the XML-Encryption specification.
Note that we used XML to store this data in the mobile agent.
492 H. Razouki

• Alteration of dynamic data: The mobile agent requires in its security policy the
signature of the data produced by certain components. In this case, platforms use
their public key certificates to sign this data. This signature is done through the
XML-Signature specification.
• Dynamic data access control of the mobile agent: the mobile agent migrates
from one platform to another with encrypted data. To manage access control
to this data between platforms, we have created XACML/ABAC access control
policies. These policies are based on the rules defined by SAS at the time of agent
adaptation. With this security mechanism, only platforms authorized by the PDP
can retrieve the symmetric key from KDC to access this encrypted data.

6 Conclusion

We used XACML/ABAC access control policies to protect access to mobile agent

data. In the adaptation phase, the SAS is timing the security levels of the components
and the degree of confidence of the platforms visited by this agent. It then uses access
control constraints to create the rules for controlling access to agent data. These rules
are then transformed into XACML policies managed by the PAP and PDP entities.
These policies manage access to agent data across platforms.
We discussed the reliability of our protection against different types of attacks. In
the future, we are working on how to integrate this solution with the JADE platform.

References

1. Mehak, Khandnor P (2019) Energy efficient data aggregation using multiple mobile agents
in wireless sensor network. In: Advances in intelligent systems and computing, vol 669, pp
279–287
2. Aditya Sai Srinivas T, Ramasubbareddy S, Govinda K (2019) Discovery of web services using
mobile agents in cloud environment. In: Lecture notes in networks and systems, vol 74, pp
465–471
3. Roopa GM, Nirmala CR (2019) Intelligent mobile agent framework for searching popular
e-advertisements. In: Advances in intelligent systems and computing, vol 714, pp 177–189
4. Mwiti JI, Kahonge M, Abande E (2019) Mobile agents in distributed environment for
information retrieval. Int J Appl Inf Syst 12(20):8–10
5. Kang H, Kim H, Kwon YM (2019) RECEN: resilient MANET based centralized multi
robot system using mobile agent system. In: 2019 IEEE symposium series on computational
intelligence (SSCI), pp 1952–1958
6. Seghier NB, Kazar O (2020) Mobile agent and ontology approach for web service discovery
using QoS. Int J Reason-Based Intell Syst 12(1):17–33
7. Shehada D, Yeun CY, Jamal Zemerly M, Al-Qutayri M, Hammadi YAl (2017) Secure mobile
agent protocol for vehicular communication systems in smart cities. In: Information innovation
technology in smart cities, pp 251–271
8. Razouki H, Hair A (2014) Self-adaptive security for mobiles agents. Int J Comput Appl 94(13)
9. Razouki H, Hair A (2015) Security for mobile agents: trust estimate for platforms. TELKOM-
NIKA Indonesian J Electr Eng 15(2):381–389
Dynamic Data Protection for Mobile Agents … 493

10. Razouki H, Hair A (2013) Components-based software architecture for secure mobile agents
via two strategies of adaptation. In: 2013 8th international conference on intelligent systems:
theories and applications, SITA 2013
11. Razouki H (2019) Security policy modelling in the mobile agent system. Int J Comput Netw
Inf Secur (IJCNIS) 11(10):26–36
12. Bell DE, La Padula LJ (1976) Secure computer system: unified exposition and multics
interpretation
Securing the Storage of Passwords Based
on the MD5 HASH Transformation

Hamza Touil, Nabil El Akkad, and Khalid Satori

Abstract The hash function is designed to fold an input array of any size into a bit
string. For the MD5, this is a very convenient algorithm for comparison and storage.
For example, Unix systems use it to store user passwords in a 128-bit encrypted
format. It is widely used to check the integrity of files. The length of the output string
is 128 bits. When it first appeared, it was impossible to break it, but today this is
no longer the case. Many programs select the source word based on the hash. The
vast majority of them make an enumeration in the dictionary. However, Rainbow
Crack methods are based on generating a set of hashes from a set of characters to
search for a hash using the resulting base. Besides, the MD5, like any hash function,
has a concept such as collisions - it is about getting the same hashes for different
characters’ source strings. This paper will propose a method capable of increasing
password storage security based on the random generation of an MD5 HASH and
then use it to transform the original HASH according to pre-defined rules before
storing it in the database.

Keywords Hash function · Password · MD5 · Storage · Transformation

1 Introduction

Cryptographic hash functions are an indispensable and ubiquitous tool for

various tasks, including authentication, data integrity checking, file protection, and
even malware detection. Many hash algorithms differ in cryptographic strength,

H. Touil (B) · N. El Akkad (B) · K. Satori (B)

LISAC, Faculty of Sciences, Dhar-Mahraz (FSDM), Sidi Mohamed Ben Abdellah University,
Fez, Morocco
e-mail: [email protected]
N. El Akkad
e-mail: [email protected]
N. El Akkad
Laboratory of Engineering, Systems and Applications (LISA), National School of Applied
Sciences (ENSA), Sidi Mohamed Ben Abdellah University, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 495
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_45
496 H. Touil et al.

complexity, bit depth, and other properties. Today, hash has acquired many new
properties and is used in various fields of information technology. The main require-
ment for hash functions is the uniform distribution of their values for a random
selection of argument values. For cryptographic hash functions, it is also crucial
that the slightest change in the argument changes the function a lot. It is called the
avalanche effect [1–4].
The impossibility of fabrication requires that it is complicated to find a message
with the correct bending value. The second is the incredible complexity of matching
a given message with a known bend value with another message with the right bend
value [5–7].
Unidirectionality is understood as the incredible complexity of finding a message
by a given convolution value. It should be noted that there are currently no hash
functions used with proven unidirectionality.
Collision resistance refers to the difficulty of finding a pair of messages with
the same convolution values. Usually, the discovery by cryptanalysts of a collision
construction method serves as a first signal that the algorithm is obsolete and the
need to replace it soon [8–14].
The resistance to finding the second pre-image is understood as the complexity
of finding a second message with the same convolution value for a given message
with a known convolution value. There are two critical types of cryptographic hash
functions: key and keyless. Key hash functions are called message authentication
codes. They can be used, without additional means, to ensure both the data source’s
accuracy and data integrity in systems with trusted users. Keyless hash functions are
called error detection codes. They enable additional means (such as encryption) to
ensure data integrity. These hash functions can be used in systems with trusted and
untrusted users [15–21]. In the rest of this paper, we will see the works in relation.
Then we will dissect our method ending with the experiments.

2 Related Works

A set of studies has been proposed to ensure password storage security: the technique
[22] capable of hiding user data and making them abstract and can also support all
forgery attacks, using a secure remote authentication scheme based on smart cards.
Kumar [23] et al. Presents a new advanced security architecture for user identifica-
tion, which includes two authentication factors, encryption, and decryption of data
uploaded to the cloud using an AES system, verification and locking of users by
the administrator, recovery of user IP data, distributed database storage, i.e., data is
stored in tiers, which means that user login data is stored in one database and encryp-
tion/decryption data (downloaded files, key) is stored in different databases. Using
the hash technique, Matsunaga [24] et al. uses a message authentication protocol
using a homo-morphic hash function without full key security in a WANET envi-
ronment with limited resources for each node. Bebe [25] et al. invent a mechanism
under the name "OSA-SHSDS" to store user data in the cloud with a higher privacy
Securing the Storage of Passwords Based on the MD5 HASH Transformation 497

rate and less complicated space. In order to store the data in the server, the cloud
user’s data must be recorded. After registration, the cloud server generates the ID
and password for each registered cloud user. Each time the user wants to store the
data, he has to log in to the cloud server with his ID and password to authenticate.

3 Proposed Method

3.1 Explanation of the Method

Our method consists of two different phases. In the first phase, the “subscription
phase,” where the registration page is invoked, the user fills in and sends the form.
Then the server extracts the submitted information, username, password, and other
additional information. The most interesting for us is the password that will be
hashed by the MD5 hash algorithm. The server will then generate a second hash
(MD5) randomly, as explained (Fig. 1).
The next step is to build the new hash according to the algorithm (Fig. 3). The
new HASH will be stored in the database, and then the end-user will receive the two
OUTPUTs first and then size so that they can use it during the next authentication
“authentication phase” (Fig. 2).

Step 1: New registration

Subscription Form: Step 2: send the

Login data to the server
Password for processing
further information….

Step 3: Generate password

hash + random HASH

Step 6: communicate to the

user the numbers he must use
with his login and password in Step 4: Apply the algorithm
the form (first, size) described in figure 3

Step 5: Store the new HASH in

the database.

Database

Fig. 1 Operating principle of the method - “subscription phase.”

498 H. Touil et al.

4
Step 5: verification phase, ap-
ply the reverse process and
make a comparison between the
Step 1: Authentication
generated HASH and the initial
process
HASH.

Step 2: Send a request

Authentication
Form:
Login
Password
First Step 6 : accept or refuse
Size

Step 3: Call information Step 4: Response contains the

(Login, HASH) from Database requested information

Database

Fig. 2 Operating principle of the method - “authentication phase”.

3.2 Resistance Against Attacks

A. Pre-image
A pre-image attack on cryptographic hash functions attempts to find a message that
has a specific hash value. A cryptographic hash function must resist attacks on its
pre-image (set of possible entries).
With an ordinary MD5 hash, one can identify:
L: hash size estimated in 128bits.
N: Number of possible hash such that N = 2L.
P: the probability of finding the right HASH such that: P = 1/N.
Then the number of trials will be identified as follows:
∞
E(N) = i*(1 − p)(i−1) * P
i= 0

With our method, the complexity of falling on the HASH corresponds to the
password is almost impossible, given the significant number of tests the hacker must
perform.

B. Collision attack √
If a hash function produces N output bits, an attacker that only computes 2N hash
operations on a random input is likely to find two matching outputs. If there is a more
straightforward method than this brute force attack, it is generally considered a flaw
in the MD5 hash function.
Using our method, this hypothesis will not be beneficial to detect the password
given the change in hash size plus the reasonably large number of iterations.
Securing the Storage of Passwords Based on the MD5 HASH Transformation 499

Fig. 3 Algorithm of constriction of the new HASH.

4 Experimentation

4.1 Generating HASH

We wish to apply our method on the password: “Storage”.

1) Generate the original HASH (the hash of our “Storage” password) in MD5 of
size 128 bits as follows:
500 H. Touil et al.

2) Generate a 128 bits MD5 HASH randomly, based on a random integer, for
example: “1425”.

3) As described in the algorithm (Fig. 3), the two parameters first and size must
be generated as follows:

Example 1: First = 0, size = 4 (4bits = 1/2 byte).

In this case, we will work directly on the hash in hexadecimal and will not need
to convert it to binary.

The goal is to build a new 256-bit HASH from the above two HASHs by alternating
construction starting with the Original HASH.

Example 2: First 1, size = 16bits (16bits = 2 bytes).

In this case, we will work directly on the hash in hexadecimal and will not need
to convert it to binary.

The goal is to build a new 256-bit HASH from the two HASHes above by
alternating construction starting with the random HASH.

Later this hash will be stored in the database with a size of 256 bits.
We can notice through these two examples that the generated HASHs are different.
Securing the Storage of Passwords Based on the MD5 HASH Transformation 501

We can generate 14 different hashes for this password.

4.2 Recover the original HASH

The reverse operation capable of distinguishing the original HASH from the random
one, taking the HASH generated on the second example.

We already know the parameters: First 1, size = 16 bits = 2 bytes.

Just detect the random hash and then delete it.

5 Conclusion

We have demonstrated a new authentication security specialist method that will make
pre-image and collision attacks almost impossible, given the new HASH size. Our
method offers a reasonably high level of security. Even if a hacker manages to access
our database and decrypt the HASH, he will inevitably come across a value that
has nothing to do with the password stored with a good user experience and very
low delays. The security of passwords has always been an essential concern in the
deployment of platforms, applications. However, it is almost impossible for a hacker
to find the secret being used with our approach, given the number of possibilities we
have and the new size of the hash. Moreover, even if we reuse the same password in
multiple websites, this will pose any security problems since each apparent hash is
unique.
502 H. Touil et al.

References

1. Touil H, El Akkad N, Satori K (2020) Text Encryption: Hybrid cryptographic method using
Vigenere and Hill Ciphers. In: 2020 International Conference on Intelligent Systems and
Computer Vision (ISCV), Fez, Morocco, pp. 1–6
2. Touil H, El Akkad N, Satori K (2020) H-Rotation: Secure storage and retrieval of passphrases
on the authentication process. Int J Safety Security Eng 10(6):785–796
3. El Akkad N, El Hazzat S, Saaidi A, Satori K (2016) Reconstruction of 3D scenes by camera
self-calibration and using genetic algorithms. 3D Research, 6(7): 1–17
4. Elazzaby F, El Akkad N, Kabbaj S (2019) A new encryption approach based on four squares
and Zigzag. The 1st International Conference on Embedded Systems and Artificial Intelligence,
ESAI
5. Es-sabry M, El akkad N, Merras M, Saaidi A, Satori K (2019) A new color image encryption
using random numbers generation and linear functions. In: The 1st International Conference
on Embedded Systems and Artificial Intelligence, ESAI
6. El hazzat S, Merras M, El akkad N, Saaidi A, Satori K (2019) Enhancement of sparse 3D
reconstruction using a modified match propagation based on particle swarm optimization.
Multimedia Tools and Applications (Springer), 78, pp. 14251–14276
7. Cheng R, Jin Y (2015) A competitive swarm optimizer for large scale optimization. IEEE Trans
Cybern 45(2), art. no. 6819057, pp. 191–204
8. Zhang W-X, Chen W-N, Zhang JA (2016) Dynamic competitive swarm optimizer based-on
entropy for large scale optimization. In: Proceedings of the 8th International Conference on
Advanced Computational Intelligence, ICACI 2016, art. no. 7449853, pp 365–371
9. Al-Tashi Q, Abdul Kadir SJ, Rais HM, Mirjalili S, Alhussian H (2019) Binary Optimization
Using Hybrid Grey Wolf Optimization for Feature Selection (Open Access) (2019) IEEE
Access, 7, art. no. 8672550, pp. 39496–39508
10. Singh H, Sharma M, Sharma S, Anand P (2019) A New Hybrid Algorithm Based on Grey
Wolf Optimization and Crow Search Algorithm for Unconstrained Function Optimization and
Feature Selection (Open Access) IEEE Access, 7, art. no. 8643348, pp 26343–26361
11. Mohapatra P, Nath Das K, Roy S (2017) A modified competitive swarm optimizer for large
scale optimization problems. Appl Soft Comput J 59:340–362
12. Das S, De S, Kumar R (2016) Implementation and comparative analysis of RSA and MD5
algorithm. Int J Comput Appl (0975–8887), 141(9)
13. Hossain A, Islam’S K, Kumar S, Nashiry M (2012) Cryptanalyzing of message digest
algorithms MD4 and MD5. Int J Cryptography and Inf Security, 2(1)
14. Niharika M, Ethiraj R (2015) Design and implementation of hybrid RC4 and Sha-2 algorithms
for WIFI applications. Int J Emerg Eng Res Technol
15. El Akkad N, Merras M, Baataoui A, Saaidi A, Satori K (2018) Camera self-calibration having
the varying parameters and based on homography of the plane at infinity. Multimedia Tools
Appl 77(11):14055–14075
16. El Akkad NE, Merras M, Saaidi A, Satori K (2013) Robust method for self-calibration of
cameras having the varying intrinsic parameters. J Theor Appl Inf Technol 50(1):57–67
17. El Akkad NE, Merras M, Saaidi A, Satori K (2013) Camera self-calibration with varying
parameters from two views. WSEAS Trans Inf Sci Appl 10(11):356–367
18. El Akkad N, Saaidi A, Satori K (2012) Self-calibration based on a circle of the cameras
having the varying intrinsic parameters. In: Proceedings of 2012 International Conference on
Multimedia Computing and Systems, ICMCS, pp 161–166
19. Merras M, El Akkad N, Saaidi A, Nazih AG, Satori K (2014) Camera calibration with varying
parameters based on improved genetic algorithm. WSEAS Trans Comput 13:129–137
20. El Hazzat S, Merras M, El Akkad N, Saaidi A, Satori K (2018) 3D reconstruction system based
on incremental structure from motion using a camera with varying parameters. Visual Comput
34(10):1443–1460
21. Merras M, Saaidi A, El Akkad N, Satori K (2018) Multi-view 3D reconstruction and modeling
of the unknown 3D scenes using genetic algorithms. Soft Comput 22(19):6271–6289
Securing the Storage of Passwords Based on the MD5 HASH Transformation 503

22. Mukesh S, Tejas P, Anuj J (2020) Security analysis on remote user authentication methods:
Lecture Notes on Data Engineering and Communications Technologies, vol 31, pp 506–513
23. Kumar S, Jafri SAA, Nigam N, Gupta N, Gupta G, Singh SK (2019) A new user identity based
authentication, using security and distributed for cloud computing: IOP Conference Series:
Materials Science and Engineering, vol 748, Issue 1, Article number 012026. In: International
Conference on Mechanical and Energy Technologies 2019, ICMET 2019; Galgotias College
of Engineering and Technology, Greater Noida, Uttar Pradesh, India
24. Matsunaga A (2019) Message authentication scheme for ad hoc Networks with homomorphic
hash function. In: Proceedings of 2019 2nd World Symposium on Communication Engineering,
WSCE 2019 December 2019, Article number 9040951, pp 138–141
25. Bebe PC, Akila D (2019) Orchini similarity user authentication based streebog hash function
for secured data storage in cloud. In: Proceedings of 2019 International Conference on Compu-
tational Intelligence and Knowledge Economy, ICCIKE 2019, December 2019, Article number
9004393, pp 461–466
Is Blockchain Technology Applicable
in Small and Medium-Sized Enterprises?

Zakariya Chabani , Salima Hamouche , and Raed Said

Abstract The purpose of this research is to investigate the applicability of

blockchain technology in Small and Medium Enterprises (SMEs). We have exam-
ined the advantages and the disadvantages of the use of this emerging technology in
SMEs, based on the existing literature on this research topic.
Applying blockchain in SMEs has many advantages: promoting transparency in
the organization, facilitating market analysis for better decision-making, eliminating
reconciliation costs, and making auditing easier. However, it has some disadvantages,
such as; the reshuffling of job positions and the absence of clarity related to the
industrial standards.
Therefore, we proposed some recommendations in order to adopt this technology
successfully. Firstly, the SMEs should design the distributed ledger technology (DLT)
in a way to have flexible governance and regulations so that it will be easily upgrade-
able; additionally, it is essential that the adopted DLT design by an SME can be
migrated to a different blockchain whenever the need arises. Also, DLT designers
should ensure that measures are taken to limit the collection of personal information.
The SMEs should also establish trust and interoperability to realize the full benefits
of DLT. Finally, SMEs should invest in research that will ensure more knowledge
about the perfection of blockchain technology.

Keywords SMEs · Blockchain technology · HRM · Strategic management ·

Finance

Z. Chabani (B) · S. Hamouche · R. Said

Canadian University of Dubai, Dubai, UAE
e-mail: [email protected]
S. Hamouche
e-mail: [email protected]
R. Said
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 505
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_46
506 Z. Chabani et al.

1 Introduction

According to ([20], p. 6): “Blockchains are tamper-evident and tamper-resistant

digital ledgers implemented in a distributed fashion (i.e. without a central repository)
and usually without a central authority (i.e. a bank, company, or government). At
their basic level, they enable a community of users to record transactions in a shared
ledger within that community, such that under normal operation of the blockchain
network, no transaction can be changed once published.”
The application of blockchain technology in business management is an emerging
research field. Certain authors have outlined the general concepts of DLTs, their
associated challenges, and the future implications in the management of enterprises
[1, 5, 10]. Despite the benefits linked to the adoption of blockchain technology,
these authors point out that DLT’s may not be the most appropriate tool for SMEs,
that are in a bid to improve the efficiency of their operations [5, 10, 15]. As such,
before considering an investment in blockchain, decision-makers in SMEs ought
to thoroughly investigate the feasibility of blockchain’s innovative features in their
current business processes. Failure to scrutinize this feasibility can invalidate the
investment and be a misfortune for small businesses through wastage of their financial
and human resources. Accordingly, this research’s main objectives are to examine
the advantages and disadvantages of using blockchain in SMEs and advise them on
using DLTs better effectively and efficiently based on the existing literature on this
research topic.
The rest of this research will be organized as follows; the second, third, and fourth
sections analyze the adaptability of blockchain technology from a human resource
management perspective, strategic management perspective, and finance perspective.
In the fifth section, we discuss the results, and we present the main advantages and
disadvantages of using blockchain technology in SMEs. Finally, the conclusion of
the paper is provided in the sixth section.

2 Blockchain Adaptability in SMEs from a Human

Resource Management Perspective

Managing human resources in SMEs is challenging, considering their size and finan-
cial capacity. Compared to large-sized companies, SMEs cannot afford to have a
large team of human resource specialists. Generally, small firms tend to rely on
informal human resource management processes [6]. They do not have formal Human
Resource Management structures and policies. In addition, they often do not have
formal training for managers in the implementation of HRM practices.
According to some authors, blockchain technology helps companies become
stronger competitors when applied to different organizational functions. HRM is one
of these organizational functions that can be revolutionized and optimized, thanks
Is Blockchain Technology Applicable … 507

to blockchain technology. In this context, it may be relevant for SMEs to use this
technology in order to optimize the efficiency of their HRM.
The decentralized nature of blockchain should help automate information veri-
fication, enhancing the accuracy of staffing, managing, and reducing costs. It helps
organizations optimize the human resource department’s administrative task, record,
search, and validate employees’ data without the intervention of a mediator, upon
the condition of having the employee’s agreement [18].
In addition to eliminating the involvement of a third party, the blockchain
technology’s main contributions are privacy, security and transparency. The smart
contract built-in blockchain technology helps organizations store encrypted data in
a decentralized location, increasing the level of privacy and security. Furthermore,
blockchain technology does not limit the access of information to managers; it is
accessible to the company’s employees [18]. In this context, they can access the
information related to HRM functions within the organization, like training, perfor-
mance appraisal, and recruitment. This open access to employees fosters transparency
in the organization, increases the perception of fairness, improves the company’s
credibility, and develops a healthy and positive atmosphere in the organization [18].
Blockchain technology may result in more accurate methods of hiring employees.
It can be a source of significant potential solutions to eliminate recruitment issues by
supporting decision-making related to recruitment. Blockchain technology can help
organizations better recruit specialized talents by obtaining better information about
applicants, which cannot get with the traditional recruitment methods. Organizations
can obtain trustworthy information no matter the country or region the applicant
has obtained his degree or employment experience. With the employee’s consent,
employers can have access and can check the veracity of the provided information.
Besides staffing, blockchain technology should have a positive impact on compen-
sation, mainly employees’ payroll system. It makes it easier, shorter and faster to
process without the mediating role of banks. Thanks to smart contracts, automatic
payment is possible, based on a specific code and a specific persistent condition
related to the payroll system, which reduces significatively the risk of fraud.
Blockchain technology can also have a positive impact on employees’ training and
development as well as performance management; it helps to store information about
training and employees’ movements and promotions within the organization, with
access that is not limited to managers, which will create transparency considering
that the information is also available to employees [18].
Therefore, blockchain technology may optimize the HRM functions in SMEs.
However, some authors argue that the implementation of blockchain technology
should be adapted to SMEs’ business needs and should follow a tailored approach
when it is adopted due to some disadvantages. For instance; The DLT can store
tamper-proof information and other details about an employee, which is otherwise not
needed in the company, and exposes the information to the peers. Using blockchain
presents privacy and confidentiality challenges on the personal information of specific
employees [13]. In some cases, personal information is required for privacy protec-
tion and then hacked, thus, exposing the employees or senior managers to chal-
lenging privacy issues and potential fraud. The other challenge concerns an increase
508 Z. Chabani et al.

in human error since blockchain is a new concept to most SMEs. Implementation

without thorough training may result in more human error.
For the past few decades, there has been a debate that new technologies and
digitalization of the work environment may result in human labor loss. Blockchain
and especially the DLT technology will result in a machine doing some of the work
traditionally done by humans. The implication is the loss of jobs, which affects the
company’s human resource [4]. There is also the challenge that with the introduction
of blockchain, job roles within the company may change. Specifically, there may be
a reshuffling of job positions; some employees may have reduced tasks while others’
task is increased. It becomes a challenge to re-evaluate each of the job designs and
to review the remuneration structure [11].

3 Blockchain Adaptability in SMEs from a Strategic

Management Perspective

Strategic management represents a significant challenge for organizations regard-

less of their size. However, it might be more challenging and demanding for SMEs
compared to large enterprises. Their size paired with their restricted resource and
the uncertainty of their external environment make strategic planning and manage-
ment a complex exercise. The main concerns might be how to be competitive in an
environment characterized by uncertainty and rapid changes? How to align mission,
vision, and organizational strategy? Furthermore, how to ensure the achievement of
the company’s strategic goals? In this context, it is essential to identify blockchain
technology’s contribution from a strategic management perspective.
Since blockchain eliminates loops for fraud, the SMEs will be protected [2].
They will be able to adjust their short term and long term goals according to their
situational analysis. When a company defines smart goals that are feasible, it will
motivate employees and boost their morale [1]. Many companies fail to grow because
they do not know how to set-up goals properly. Very high goals demoralize efforts put
by SMEs, while meager goals underutilize scarce resources. A company will be able
to match its financial capabilities with those of its competitors. Therefore, instead of
copying and pasting their goals, SMEs will make the necessary adjustment to their
objectives, which will promote their survival and sustainability [7]. If many people
can access this feature about an enterprise, other enterprises will want to continue
with the same trend of creating unique services and products to fill in the market’s
current gaps.
Adopting the blockchain seems to be beneficial for SMEs; however, some
researchers stated some disadvantages. For instance, the blockchain eliminates
centralized infrastructure in favor of decentralized governance, which may not be
valid for the entire SME’s infrastructure. In most cases, even for companies that have
successfully implemented the DLT system such as Bitcoin and Ethereum, arriving
Is Blockchain Technology Applicable … 509

at decisions on managerial changes for the DLT structure is often difficult. Tradi-
tionally, the regulators of the financial sectors have relied on effective centralized
management for their decision. As indicated by Gatteschi et al. [5], the administra-
tors may be subject to specific DLT arrangements that may be difficult to enforce.
Thus, SMEs need to be aware of this potential difficulty.
Transition to using blockchain also brings about legal and regulatory challenges.
The developments of standards at the industrial level are necessary, yet they are still
at the incubation stage [10]. This implies that for a single SME to adopt blockchain
when industrial standards are not set will bring complexity in business operations
[3]. In addition, legal clarity regarding the jurisdiction and ownership of the DLT
system may be challenging. The primary reason is that there is no legal entity that
controls distributed ledger in the blockchain. The fact that the distributed ledgers in
blockchain are immutable creates a gap regarding how transaction disputes can be
resolved as they arise [15].

4 Blockchain Adaptability in SMEs from a Finance

Perspective

Financial resources represent another challenge for SMEs, often struggling with how
to optimize their financial resources, capacity, and management.
Based on blockchain technology, the DLT is known for having the capacity to
enable direct transmission of token or digital values between two parties and a
decentralized record-keeping. Resultantly, there is no need for a central authority
that controls the ledger. The results are; better scalability; they reach the market
faster and at a reduced cost. The other advantage is that DLT enhances transparency
and makes auditing easier because all the network members retain a full copy of
the ledger (Encrypted so as not to make changes). Resultantly, given that all ledgers
are encrypted, and the absence of central authority makes it hard to commit fraud.
At the same time, it eliminates reconciliation costs. The DLT transaction approach
features transactional exchanges on a peer to peer basis with conditions (such as the
date of release and rules for the failure of loan repayment) that are embedded and
predetermined [12]. Consequently, there is no need to set up additional business rules
for collaterals between transacting parties.
Besides, the blockchain can offer an audit trail that is immutable and verifiable
for any form of digital and physical assets. The DLT also removes friction in trans-
actions, thus enhancing efficiency and speed. Besides, DLT ensures that friction in
the clearing of settlements and the transactions is removed, increasing efficiency in
the accounting sector. The process is fully automated, and the intermediaries are
eliminated to ensure that the settlement process is cleared within the shortest time
[12]. The other advantage is that when well secured, the DLT offers more resilience
against cyber-attacks than the traditional system. The reason is that with DLT, a single
510 Z. Chabani et al.

point of security is eliminated and replaced with multiple security points; hence it
will be challenging for attackers to eliminate all points.
One of the disadvantages pertains to blockchain’s scalability, both with regard to
the speed of verification and the volume of transactions [5]. The transaction speed of
blockchains that do not require permissions is limited. For example, in Bitcoin, the
DLT has been designed to only allow 4–7 transactions per second due to the block
size limit to a single megabyte. This disadvantage can be resolved by increasing the
transaction throughput, as is the case with Ethereum [9]. Several factors must be
considered when appropriating the blockchains to SMEs, given that one model may
be appropriate for a particular business and, at the same time, disadvantageous to the
other.
The integration of DLT in financial infrastructure can be very complicated. Specif-
ically, the integration requires coordination from the entire industry. Coordination
and collaboration at an industry level may require additional expenses, especially
at the initial implementation stages. The other disadvantage is that when there is a
single weakness in the blockchain, the attackers can capitalize on committing fraud.
The success of DLT thus relies on the nature of blockchain and the measures taken
to ensure that the algorithms determining the core values of the DLT system are
not compromised. However, it is possible to maintain integrity and complexities
that make it difficult for attackers to find any fault line. For instance, the Bitcoin
blockchain has never been attacked since it adopted the DLT system in 2009.

5 Results and Discussion

This literature review has shown that blockchain provides SMEs an opportunity to
eliminate the intermediaries within the company. Specifically, blockchain thinking
gives a business a chance to experience accountability, the Internet of Things (IoT),
decentralized decision-making, and sustainability within an industry. All the advan-
tages and disadvantages are summarized in Table 1 to prepare different managerial
functions, including strategic management, HRM, and finance, to understand the
positive and negative effects of adopting DLT.

5.1 The Performance Issues of DTL Design

Scalability of the DLT System. It is essential for the DLT to have an appropriate
scale that will ensure it is not vulnerable to cyber-attacks. The DLT should have
flexible governance and regulations so that it is easily upgradeable. This will also
ensure that a hard fork’s chances are significantly lowered because the regulations
are easy to follow. The blockchain should also be designed to ensure interoperability
in applications of the IoT and smart grid. The blockchain market should not just be
efficient but also transparent to build trust for all stakeholders in the chain. The DLT
Is Blockchain Technology Applicable … 511

Table 1 The advantages and disadvantages of adopting the blockchain technology in the SMEs.
The perspective Advantages Disadvantages
Human resource management To promote transparency in the Privacy and confidentiality
organization problems
To increase the perception of
fairness within the
organization
To improve the company’s Blockchain adoption may
credibility cause a reshuffling of job
To develop a healthy and positions
positive atmosphere in the
organization
To improve compensation
management, mainly
employees’ payroll system
To optimize the management
of training, employees’
movements and promotions
Strategic management To facilitate the market Applying blockchain may
analysis for better necessitate changing the whole
decision-making structure of the organization to
Proper valuation of the fit the new technology
company’s assets
The activities run faster than in Standards and regulations at
the traditional forms of the industrial level are not clear
organizations
Planning goals and achieving
them is more comfortable than
in the traditional form of
organizations
Finance To Enhance transparency and The initial stage of adopting
making auditing easier blockchain requires a
considerable amount of capital
To eliminate reconciliation Cyber-attacks, if successful,
costs can lead to substantial financial
losses

is fast evolving to meet these objectives and ensure that ledgers and data are securely
recorded across multiple sources.

Interoperability of DLT. It is essential that the DLT design that has been adopted
by an SME can be migrated to a different blockchain whenever the need arises.
For instance, the SME may purchase a new subsidiary and transfer the current DLT
to the new premise without experiencing complications of any kind. Besides, in
consideration of the DLT interoperability, it is crucial to ensure that the migration
of the DLT is completed at a minimal cost. There is a shift in focus from a single
512 Z. Chabani et al.

blockchain, such as Bitcoin, to multiple chains that interoperate together and offer
niche services to all the users.

Privacy by Design. The issue of privacy was predominant in most literature included
in this study. It emerged that blockchain systems that have not been designed to
protect users’ data are more susceptible to unauthorized access and cyber-attacks.
DLT designers should ensure that measures are taken to limit the collection of
personal information. In cases where it is necessary to collect confidential infor-
mation, it is necessary to protect it. Findings from a study done by Islam and Hasan
[8], revealed that the DLTs in energy training is fond of using pseudonyms to protect
personal data. However, the pseudonyms can be misleading since it can only be valid
as long as the pseudonym has not been linked to a person. However, the moment a
connection is linked between a person and the pseudonym reveals all the confiden-
tial information. The other shortcoming of the use of pseudonyms is that it prevents
enterprises from sharing propriety information in a secure manner [19]. SMEs can
adopt privacy-enhancing information at the conception stage of the DLT system.

6 Conclusion

DLT blockchain technology is a novel approach to record sharing information across

several ledgers and among the network participants in a wide range of industries.
Although the technology is mostly used in the finance sector, it can also result in
the efficiency of strategic management and human resource sectors. In finance, it
enhances transparency and accountability, makes auditing easier, and decentralizes
transactions. In management, DLT blockchain ensures faster and holistic decision-
making through blockchain algorithms that mimic human intelligence, enhance
business universality, and make management efficient. In the HRM, blockchain
technology may revolutionize staffing, compensation, and employees’ training and
administrative management of human resources. It also comes in handy for the
reorganization of work design and HRM efficiency.
The recommendation for SMEs is to invest in research that will ensure more
knowledge on the perfection of the blockchain technology is garnered since DLT is
still in its incubation stage [9]. The SMEs should also establish trust and interop-
erability to realize the full benefits of DLT. Additionally, the SMEs should design
the DLT to have flexible governance and regulations so that it is easily upgradeable.
It is also essential that the DLT design adopted by an SME can be migrated to a
different blockchain whenever the need arises. Finally, DLT designers should ensure
that measures are taken to limit the collection of personal information. In cases
where it is necessary to collect confidential information, it is necessary to ensure its
protection.
Is Blockchain Technology Applicable … 513

References

1. Andoni M, Robu V, Flynn D, Abram S, Geach D, Jenkins D, McCallum P, Peacock A (2019)

Blockchain technology in the energy sector: a systematic review of challenges and opportu-
nities. Renewable Sustainable Energy Rev 100:143–174. https://doi.org/10.1016/j.rser.2018.
10.014
2. Ballabio G (2017) Prime-partitioned blockchain: a scalable and efficient blockchain imple-
mentation. SSRN Electron J. https://orcid.org/0000-0002-5460-5940
3. De Kruijff J, Weigand H (2017) Understanding the blockchain using enterprise ontology. In:
International conference on advanced information systems engineering. Springer, Cham, pp
29–43. https://doi.org/10.3390/fi11120258
4. Ehnert I, Harry W, Zink KJ (2014) Sustainability and human resource management: developing
sustainable business organizations. Springer, New York. https://doi.org/10.1177/097417392
0130413
5. Gatteschi V, Lamberti F, Demartini C, Pranteda C, Santamaria V (2018) To blockchain or not
to blockchain: that is the question. IT Prof 20(2):62–74. https://doi.org/10.3390/fi10020020
6. Heilmann P, Forsten-Astikainen R, Kultalahti S (2018) Agile HRM practices of SMEs. J Small
Bus Manag 1–16. https://doi.org/10.1111/jsbm.12483
7. Hunter M (2015) Strategic utilization of information systems in small businesses. IGI Global.
https://doi.org/10.4018/978-1-4666-8708-0
8. Islam K, Hasan F (2019) An efficient algorithm for peer-to-peer energy trading using blockchain
in microgrid energy markets. Eur J Electrical Eng Comput Sci 3(3). https://doi.org/10.24018/
ejece.2019.3.3.80
9. Lehmann C, Schock C, Uhlemann TJ, Küfner T (2017) Requirements for blockchain applica-
tions in manufacturing small and medium sized enterprises. DEStech Trans Eng Technol Res
(ICPR). https://doi.org/10.12783/dtetr/icpr2017/17601
10. Makridakis S, Christodoulou K (2019) Blockchain: current challenges and future
prospects/applications. Future Internet 11(12):258. https://doi.org/10.3390/fi11120258
11. Met I, Uysal E, Ozkaya K, Orc E (2020) Key success factorss for strategic management in
digital business. In: Hacioglu U (ed) Digital business strategies in blockchain ecosystems.
Springer, Cham, pp 283–304. https://doi.org/10.1007/978-3-030-29739-8
12. Min H (2019) Blockchain technology for enhancing supply chain resilience. Bus Horiz
62(1):35–45. https://doi.org/10.1016/j.bushor.2018.08.012
13. Morabito V (2019) Business innovation through blockchain: the B3 perspective. Springer,
Cham. https://doi.org/10.1007/978-3-319-48478-5
14. Natarajan H, Krause S, Gradstein H (2017) Distributed ledger technology and blockchain.
World Bank. https://openknowledge.worldbank.org/handle/10986/32436
15. Nehaï Z, Guerard G (2017) Integration of the blockchain in a smart grid model. In: Proceedings
of the 14th international conference of young scientists on energy issues (CYSENI 2017),
Kaunas, Lithuania, pp 25–26. https://doi.org/10.3390/s18082575
16. Onik M, Miraz MH, Kim C-S (2018) A recruitment and human resource management technique
using blockchain technology for Industry 4.0. https://doi.org/10.1049/cp.2018.1371
17. Qiao Y, Ma C, Lan Q, Zhou Z (2019) Inventory financing model based on blockchain tech-
nology. In: Fourth international conference on economic and business management (FEBM
2019). Atlantis Press. https://doi.org/10.2991/febm-19.2019.7
18. Wang X, Feng L, Zhang H, Lyu C, Wang L, You Y (2017) Human resource information manage-
ment model based on blockchain technology. Paper presented at the 2017 IEEE symposium on
service-oriented system engineering (SOSE). https://doi.org/10.1109/SOSE.2017.34
514 Z. Chabani et al.

19. Wu H, Li Z, King B, Ben Miled Z, Wassick J, Tazelaar J (2017) A distributed ledger for
supply chain physical distribution visibility. Information 8(4):137. https://doi.org/10.3390/inf
o8040137
20. Yaga D, Mell P, Roby N, Scarfone K (2018) Blockchain technology overview. National Institute
of Standards and Technology, Gaithersburg, p 6. https://doi.org/10.6028/NIST.IR.8202
Anomaly Detection from Log Files Using
Multidimensional Analysis Model

Yassine Azizi, Mostafa Azizi, and Mohamed Elboukhari

Abstract Software applications provide a variety of mechanisms for logging every-

thing that can happen within a system. The analysis of these log files in order to
extract some knowledge about security issues is a real challenge.
In this paper, we propose a methodology for analyzing log files based on the
development of decision support system. The goal of this methodology is to collect
log file events from different machines and save the necessary data to a data ware-
house in order to extract security knowledge, monitoring, reporting, get operational
information, and make better decisions.
As illustration, we have considered a concrete case study on web server log files
based on data warehousing techniques to diagnose the state of the system security, to
track key performance indicators, and to identify abnormal behavior and detect some
malicious events within computer systems in the hope of taking counter-measures
against them. The obtained results are promising; we can easily visualize servers
health and monitor problems.

Keywords Data warehouse · Security · Log files · BI · OLAP · MDX

1 Introduction

The global volume of digital data production is growing at a fast pace [1, 2], and
IT security has become increasingly important for businesses and organizations that
constantly store and transfer a huge pile of critical data across their local or wide
area networks [3, 4].

Y. Azizi (B) · M. Azizi · M. Elboukhari

Lab. MATSI, ESTO, University Mohammed 1st, Oujda, Morocco
M. Azizi
e-mail: [email protected]
M. Elboukhari
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 515
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_47
516 Y. Azizi et al.

Servers, firewalls, and other IT equipment keep updated log files where are record
all events and transactions. These files retrace the history of events occurring on a
server or application. The exploitation of logs represents a relevant operation to
improve security measures [5–7].
In this paper, we present our decision-making approach for log file exploitation
based on data warehousing techniques. We collect data from log files, which represent
a mine of information to exploit, we save them in a data warehouse to facilitate navi-
gation and analysis to produce a useful knowledge. This methodology can provide
a reliable view of the security level of our information system. In addition, it aims
to detect potential computer threats and enables the most appropriate correction or
reinforcement measures to be adopted accordingly.
This paper is organized as follows: Section 2 presents main related works that
have exploited log files for extracting useful information. Then, Section 3 illustrates
our methodology that we use to process log files to extract data on attack attempts,
intrusions and performance indicators. Before concluding, in the last section, we
show a case study on web server logs.

2 Related Work

In order to guarantee the security of the systems, it has become vital to develop new
monitoring and anomaly detection solutions, in this context log files have become
a very useful source of data. Anomaly detection through logs is a problem that has
already been covered in the literature [8–10], using a wide variety of analysis and
exploitation models based on different approaches. In our previous work [11, 12],
we have provided a methodology of security analysis that aims to apply Big Data
techniques, such as MapReduce, over several system log files in order to locate and
extract data probably related to attacks made by malicious users whose intends to
compromise a system.
In [13], Authors present the methodology used in pre-processing of high volume
web log files, studying the statics of website and learning the user behavior using
Apache Flume and Pig.
In [14], the researchers propose an automated tool to support developer investi-
gation of performance issues in distributed systems. They aim to leverage the vast
log data available from large scale systems, while reducing the level of knowledge
required for a developer to use their tool uses machine learning techniques. In [15],
authors exploit data mining techniques to identify students’ online SRL. They exam-
ined both self-reported self-regulation surveys and log files to predict online students’
achievements.
Anomaly Detection from Log Files … 517

3 Business Intelligence Approach

Business Intelligence (BI) encompasses all the means, methods, tools for collecting,
consolidating, modeling, and restoring data from internal systems and external
sources, in order to offer decision support to professionals and decisions-makers.
In the literature, several recent works are based on the use of BI approaches and
data warehousing techniques, in [16], authors propose a data warehouse model for
climate big data enables the identification of weather patterns that would be useful
for agriculture fields, climatic change studies and contingency plans over weather
extreme conditions, in [17], they used data warehouse model for the analysis of field
and laboratory data from multiple heterogeneous photovoltaic test sites.
The objective of Business intelligence is to improve the timeliness and quality of
information, and enable managers to be able to better understand the actual state to
deliver valuable information for making a decision. BI is an area of Decision Support
System (DSS), the purpose of which is to gather and analyze past, present and future
data in order to facilitate monitoring and detailed understanding of the environment,
results, internal and external actors, to optimize strategic decisions and the conduct
of operations.
The architecture of BI can be containing three essential phases:
• Data Capture/Acquisition: operating data must be collected and recovered from
the various available sources, filtered and adapted in view of a decision-making
use,
• Data Storage: the data is structured, centralized and made available in the data
warehouse
• Data Access & Analysis: the end user must be able to analyze the information
and draw conclusions

4 Proposed Methodology

We have been witnessing for a few years now a quantitative explosion of data that
has illustrated three major characteristics of data: volume, variety and velocity. This
has exponentially increased the complexity of analysis and processing.
In this sense, we propose a methodology based on the decision-making infor-
mation system using data warehouse techniques for integrating, coordinating and
analysis this immense diversity of log files (Fig. 1).
Our proposed methodology includes the following layers:
• Collection layer: The collection is done from log files (error logs, access logs …).
These files can be in different formats and represent different systems. These files
are therefore generally heterogeneous, which is why we will have to go through
a so-called integration phase to be able to manipulate them before storing them
in our decision support system.
518 Y. Azizi et al.

Fig. 1 Proposed approach

• ETL layer: It is at this level that the first software layer of the decision-making
environment, namely the ETL (Extract, transform, loading). This layer offers
functions for extracting all events from different log file sources, transforming
these data (homogenization, filtering, calculation) and loading them into the DW
(data warehouse). It guarantees the relocation of the computing load and a better
availability of sources.
• Data Warehouse Layer: allows to store the data in a Data warehouse, the main
features of this data warehouse are subject-oriented, integrate, time-variant and
non-volatile.
• Data mart layer: a Datamart is a sub-element of the data warehouse, collects a set
of organized, targeted, aggregated and aggregated data to meet specific needs.
• OLAP cube Layer: multidimensional analysis is the ability to analyze data that
has been aggregated into several dimensions. So we wanted to access data already
aggregated according to the needs of the user, quickly and efficiently. For this we
use OLAP hypercube, the data are represented in n-dimensional hypercubes, and
structured along several analysis axes (dimensions).
• Reporting/analysis Layer: The last phase concerns the restitution of the results

5 Result and Discussion

The main objective of our proposed methodology is the exploitation and control of
log files to extract knowledge related to system security and also to determine and
control key performance indicators to make better decisions.
In this case study, we exploited log files of web servers as a source of data, these
files can contain verbose details about the inner workings of our servers.
There are two types of log: access log and error log [18]:
• Access Log records all requests that were made of this server including the client
IP address, URL, response code, response size etc.
• Error Log records all the details such as Timestamp, Severity, Application name,
Error message ID, Error message details (Fig. 2).
Anomaly Detection from Log Files … 519

Fig. 2 Access log format

5.1 Multidimensional Design

The data warehouse modeling can follow one of the multidimensional data building
models, namely the star model, the snowflake model or the constellation model.
These models involve a table of facts as well as tables of dimensions.
We opted for the star model, which is the classic model. It is considered to be the
simplest model and offers great ease of navigation (Fig. 3).

Fig. 3 Star schema architecture

520 Y. Azizi et al.

Fig. 4 ETL process

5.2 Extraction, Transformation and Loading Phase

The ETL phase is the first phase of the realization of our decision-making system,
its purpose is to build the data warehouse. In this respect, we approached “Pentaho
Data Integration” (PDI) [19] which is a complete tool with advanced features of ETL
processing, it allows to create processes composed of data import and export, and
different operations such as conversions, joins, applying filters, or even executing
JavaScript functions. PDI has many connectors, both read and write, allowing it to
access a large number of databases and all types of files. It is by means of this tool
that we realized this phase ETL.
We applied, to the data extracted from the access and error log, the necessary
transformations and operations such as sorts and joins before loading them into the
data warehouse. We started by loading, first, the dimension tables before loading the
fact table in order to respect the integrity constraints when loading the foreign keys
in the fact table (Fig. 4).

5.3 Creating the OLAP Cube Phase

After building and feeding our data warehouse, comes the OLAP cube development
step. A cube is a multidimensional structure that contains dimensions and measures,
the dimensions define the structure of the cube while the measures provide the numer-
ical values that interest us. This step is essential in any decision-making system since
it allows us, thereafter, to perform the analyzes on the performance indicators.
For this purpose, we used the Workbench Schema [20] that provides a graphical
interface to perform this task.
Anomaly Detection from Log Files … 521

Fig. 5 OLAP cube

The creation of an OLAP cube consists in the determination of the tables

which will enter the analysis, then the determination of the performance indicators
(DDOS_attempt, SQLi_attempt, serious_error_count, code_response_error) then the
construction of the dimensions, their attributes and their hierarchies (Fig. 5).

5.4 Multidimensional Analysis

Setting up a Pentaho BI Server will allow advanced analysis, Pentaho [21] is a

business intelligence (BI) software that provides data integration, OLAP services,
reporting, information dashboards, data mining and ETL capabilitie. In order to
analyze our cube we used MDX Queries, MDX (Multidimensional Expressions) is
a multi-dimensional database query language that allows us to explore OLAP cubes
(Fig. 6).
Based on reporting/analysis tools (Pivot4j analytics) and complex MDX queries,
multidimensional analysis allowed us to obtain desirable results.
For our case study, we are interested in the advanced analysis of the data stored in
the data warehouse in order to extract knowledge about the indicators that characterize
some attacks and also about the attempts and the sources that generate DDOS and
SQLI attacks.
This multidimensional analysis makes it easy for us to track, monitor log events,
to extract useful information about a DDOS, SQLI attack attempts and a certain
number of performance keys from a web server.
It is possible to visualize clearly the evolution of the attempts of DDOS and
SQLI attacks by the time, the IP which generate a mass of queries in a shorter time
interval, the source machines most used to attack a server and also we can deduce
522 Y. Azizi et al.

Fig. 6 Multidimensional analysis

the malicious users who attempt to inject dangerous characters into server variables
(Fig. 7).
On the other hand, this proposal gave us the opportunity to follow the web server
performance indicators, namely the most serious errors, monitoring server error
reasons, identifying malicious client IPs…
Through this analysis, we’ll be able to understand the HTTP errors (by the isolation
of access logs that contain 400- and 500-level status codes), and on what pages they
occurred so we can fix the problem.

Fig. 7 SQL frequency result

Anomaly Detection from Log Files … 523

6 Conclusion

In this paper, we present a methodology for analyzing large volumes of log files
based on data warehousing techniques. This multidimensional analysis technique
allowed us to collect the events from several log files, saved and exploited them
to extract useful security knowledge, such as attempted attacks (DDOS, SQLI),
and error detection. It makes easier performing some system tasks like monitoring,
diagnosis, and creating reports about the current state of the system. The obtained
results are encouraging, our methodology allows to keep the data collected and
integrated from different log files and stores a large number of records necessary for
long-term analysis, however there are some difficulties concerning the identification
and the determinations of performance indicators and the attacks through the log
data. Our future work will focus on use of machine learning techniques on log data
analysis.

References

1. Oussous A, Benjelloun F-Z, Lahcen AA et al (2018) Big data technologies: a survey. J King
Saud Univ-Comput Inf Sci 30(4):431–448
2. Lee I (2017) Big data: dimensions, evolution, impacts, and challenges. Bus Horiz 60(3):293–
303
3. Zhang D (2018) Big data security and privacy protection. In: 8th international conference on
management and computer science (ICMCS 2018). Atlantis Press
4. Chandra S, Ray S, Goswami RT (2017) Big data security: survey on frameworks and algorithms.
In: 2017 IEEE 7th international advance computing conference (IACC). IEEE, pp 48–54
5. Mavridis I, Karatza H (2017) Performance evaluation of cloud-based log file analysis with
apache hadoop and apache spark. J Syst Softw 125:133–151
6. Kandpal N, Bandil DK, Shekhawat MS (2020) Analysis of web server logs of websites using
different web mining tools. Solid State Technol 1492–1506
7. Rao RS, Arora J (2017) A survey on methods used in web usage mining. Int Res J Eng Technol
(IRJET) 4(5):2627–2631
8. Amar M, Lemoudden M, El Ouahidi B (2016) Log file’s centralization to improve cloud
security. In: 2nd International conference on cloud computing technologies and applications
(CloudTech). IEEE, pp 178–183
9. Moh M, Pininti S, Doddapaneni S, et al (2016) Detecting web attacks using multi-stage log
analysis. In: IEEE 6th international conference on advanced computing (IACC). IEEE, pp
733–738
10. Salama SE, Marie MI, El Fangary, LM et al (2011) Web server logs preprocessing for web
intrusion detection. Comput Inf Sci 4(4):123–133
11. Azizi Y, Azizi M, Elboukhari M (2018) Tracking attacks data through log files using MapRe-
duce. In: International conference Europe Middle East & North Africa information systems
and technologies to support learning. Springer, Cham, pp 331–336
12. Azizi Y, Azizi M, Elboukhari M (2019) Log files analysis using MapReduce to improve security.
Procedia Comput Sci 148:37–44
13. Nagdive AS, Tugnayat RM, Regulwar GB et al (2019) Web server log analysis for unstructured
data using apache flume and pig
14. Nagaraj K, Killian C, Neville J (2012) Structured comparative analysis of systems logs to
diagnose performance problems. In: Proceedings of the 9th USENIX conference on networked
systems design and implementation. USENIX Association, p 26
524 Y. Azizi et al.

15. Cho M-H, Yoo JS (2017) Exploring online students’ self-regulated learning with self-reported
surveys and log files: a data mining approach. Interact Learn Environ 25(8):970–982
16. Gad I, Manjunatha BR et al (2017) Hybrid data warehouse model for climate big data analysis.
In: 2017 international conference on circuit, power and computing technologies (ICCPCT).
IEEE, pp 1–9
17. Hu Y, Gunapati VY, Zhao P et al (2016) A nonrelational data warehouse for the analysis of
field and laboratory data from multiple heterogeneous photovoltaic test sites. IEEE J Photovolt
7(1):230–236
18. Souza LT, Girish UR (2015) Error log analytics using big data and MapReduce. System 3:4
19. PDI Homepage. https://community.hitachivantara.com/docs/DOC-1009855-data-integration-
kettle. Accessed 15 June 2019
20. Workbench Homepage. https://help.pentaho.com/Documentation/8.2/Products/Schema_Wor
kbench. Accessed 15 June 2019
21. Pentaho BI server Homepage. https://help.pentaho.com/Documentation/7.0/0H0. Accessed 15
June 2019
Web-Based Techniques and E-Learning
A Review on Content Based
Recommender Systems in Tourism

Asmae Bentaleb, Younès El Bouzekri El Idrissi, and Ayoub Ait Lahcen

Abstract Recommender systems (RSs) have been used worldwide in several fields
to facilitate the tourists’ planning activities. Tourism is one of the fields that uses RSs
to reduce the overload of information to end users. Accordingly, tourists get recom-
mendations that are most suitable to their profiles. This paper presents a detailed
overview of the tourism recommender systems that were developed since 2008. It
focuses mainly on the content based systems and their applications in the tourism
field.

Keywords Recommendations · Content · Tourism · Review

1 Introduction

Due to the huge amount of information presented online, end users find the task
of planning their touristic activities very tedious. It is because they might not have
the required knowledge or time to go over all search results and select the more
suitable ones for them. RSs have come up with the best solution to this issue to help
end users deal with the information overload. These systems provide end users with
recommendations that suit their registered profiles and are aligned with the behavior
of the rest of the registered community [1].
All over the literature, during the past decades, significant advances in recommen-
dation technology have been achieved. Therefore, RSs have become widely used in
several fields like e-commerce. Tourism is one of these fields where RSs offer services

A. Bentaleb (B) · Y. El Bouzekri El Idrissi · A. Ait Lahcen

Laboratoire Sciences de L’Ingénieur, ENSA, Ibn Tofail University, Kenitra, Morocco
e-mail: [email protected]
Y. El Bouzekri El Idrissi
e-mail: [email protected]
A. Ait Lahcen
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 527
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_48
528 A. Bentaleb et al.

to end users like lists of points of interests, trip advisories and tourists’ packages that
much their profiles and preferences [2].
As far as the added value of this paper is concerned, it has been written for the
following purposes: Since in the literature there is not a state of art that is recent and
handles the methods we will need in our contribution, we have accomplished this
survey that responds to our needs. Furthermore, our work will facilitate the work
of other researchers, instead of redoing the same work of examining the literature;
they can start directly by working on their contribution. Also, this survey compares
articles that have proposed contributions to the field of tourism recommendation
especially in Sect. 4.
This survey is organized as follows: it begins by introducing recommender systems
in general, along with their architecture and most known approaches, later it describes
in a detailed manner the content-based approach for recommendation. Then, in the
related works section, it cites some surveys and determines the differences and
common points with them. This is to position the current survey among the already
existing ones found in the literature and that have discussed a similar topic. Finally,
in the applications’ section, it presents some of the well known RSs tourism applica-
tions in the literature that used the content-based approach merged with some other
techniques.

2 Background

2.1 Definition of Recommender Systems

A recommender system (RS) [3] is a mechanism that filters information to end

users based on their interests. The backbone of the recommendation process is the
concept of “personalization” [4]. These systems infer the information that are most
suitable to end users based on their profiles provided initially, and by considering
other aspects like the behavior of other similar users. Therefore, RSs are like a
filter that customizes the information available online to suit end users’ interests
and preferences. The resulted information is a set of elements presented in the way
specified in the interface. In fact, the personalization technique is composed of three
stages [5]:
– Compiling user information: in order to better analyze users’ preferences;
information about them is compiled and stored in their profiles.
– Recommending items: filters items based on the knowledge acquired in the
previous step and presents them to end users.
– Satisfaction degree: presents the end users’ feedback on the recommended items.
A Review on Content Based Recommender Systems in Tourism 529

2.2 Classification of Recommender Systems

Most papers that were published in that sense have stated three approaches of Recom-
mendation: Content-based Filtering, Collaborative Filtering, and Hybrid Filtering as
shown in Fig. 1.
Content Based Approaches Found in the Literature
Definition: Content-based recommendation systems suggests items to end users
based on their interests represented in their profiles and the items’ description.
Users can build their profiles through filling forms, but mainly, this is done through
collecting feedback of these users on items. In the next section, a general architecture
of content-based systems will be presented along with its main components.
The General Architecture of the content based recommender systems: Content- based
recommender systems encapsulate adequate methods and techniques to ensure good
representations of items, generate user profile and consequently compare this latter
with the items. Figure 2 illustrates this architecture. According to this architecture,
the process of recommendation follows these three steps:
– Content Analyzer: the main objective of this step is to generate adequate repre-
sentation of the content items to be usable for the next step in the process. This
pre-processing step is required because sometimes items data has no structure.
The result of this step is the input of the profile learner and filtering component.
– Profile Learner: this component gathers elements representing the users’ prefer-
ences and generalizes them to build the user’ profile. This is done thanks to the

Fig. 1 Recommendation approaches

530 A. Bentaleb et al.

Fig. 2 Content-based architecture [10]

machine learning techniques [11] that can infer the user model based on the items
they like or dislike in previous interactions with the system.
– Filtering Component: this module suggests adequate items to users through
matching the user’s profile with the items’ representation. This is performed using
the similarity metrics [12]. For instance, the matching can be done using the cosine
similarity between the vectors of the prototype and items.
Items can be found in two main representations: either structured or unstructured
[13]. In the first case, items are represented in a database table where each record
stores values for attributes that have a known set of values. Also, according to [10]],
the representation of items can be called structured when they are labelled by the
same set of attributes that can take a known set of values. In such situation, several
ML Algorithms can help in deducing the profile of the user [10]. In the second case,
the descriptions of items are in the form of text deduced from emails, websites,
articles or descriptions of products [10]. Unstructured data does not have attributes
with defined values. Also, the fact that a word can have synonyms and be polysemous
(having different meanings depending on the context) might increase the complexity
of natural language and create enormous issues when learning the profile of a user.
Hence, the old method which is the keyword based profile capturing is not fully able
to deduce the end users’ preferences as they are first managed by operation of string
matching [10].
To overcome these limitations, many methods were suggested throughout the liter-
ature: knowledge bases (ontologies or lexicons) in order to provide a representation
of items and profiles to produce a semantic interpretation for users’ needs.
A Review on Content Based Recommender Systems in Tourism 531

In the following sections, a description of the basic keyword-based approach for

document representation will be provided, and an overview of semantic analysis
techniques that based on knowledge of ontologies will be presented.
– Keyword-based Vector Space Model: This method corresponds to the vector
space model (VSM) with basic TF-IDF weighting [10]. It is an illustration in
an n-dimensional space of text documents using a vector. Each dimension refers
to a term in the vocabulary of the document. Accordingly, a vector of term
weights illustrates the degree of relationship relating the document in question
and term describing a document. The term weighting scheme used within this
above method is the Term Frequency Inverse Document Frequency (TF-IDF)
weighting. This process is performed after applying some pre-processing on the
data including mainly stemming, tokenizing the data and stopping words removal.
Within the content-based approach for recommendation, systems that adopt the
VSM method, weighted term vectors are used to represent users’ profiles and
items. Items’ recommendations are derived through computing the cosine simi-
larity. One of the main drawbacks of this method is that it does not take into
consideration the context in which words are used.
– Semantic Analysis based on Ontologies: This approach is mainly about learning
users’ profiles in an accurate way as they include references to concepts already
defined in external knowledge base. This method has dealt with the challenges of
providing recommender systems that respect the linguistic and cultural knowledge
and interprets documents with natural language [10]
– Semantic Analysis by using Encyclopedic Knowledge Sources: Many knowledge
bases are available for use in learning users’ profiles like Open Directory Project,
Yahoo and Wikipedia. Explicit Semantic Analysis [10] is an approach that uses
Wikipedia to generate a fine semantic representation of natural language texts in
high dimensional space.
– Learning user model: is an illustration of classification learning as the end users’
interests are deduced from their history. The classification learner’s training data
is classified into binary categories: “items that interest end users” and “items end
users do not like”. These data are collected either explicitly where users rate items
using some interfaces or implicitly through observing how the users interact with
items. As a matter of fact, there is a tradeoff in this situation as the implicit way can
generate a huge amount of information but with a certain uncertainty concerning
the fact that whether the user actually prefers the item or not. Whereas, for the
explicit methods, they are more certain and the training data has no noise, however,
end users provide explicit feedback on few items [14].
There are plenty of algorithms performing classification learning. These algo-
rithms are the pillars of content-based recommendation systems as they serve at
modeling users’ interest based on the function learned during the training data phase.
Such function aims at predicting whether a given user would be interested in a given
item. This function is basically created by the classification learning algorithm; it
outputs the probability that a user might like an unseen item. Therefore, the final list
532 A. Bentaleb et al.

of recommendations is sorted based on their probabilities. Otherwise, other algo-

rithms are designed to produce a numeric value reflecting the degree of interest. For
the traditional algorithms, they are developed to work on already structured data. In
case of unstructured such as free text, data are first converted to be structured before
proceeding to algorithms processing. However, there exist other algorithms that can
work in high dimensional spaces and do not need feature’s preprocessing [14].
Collaborative Filtering Based Approach
Within this approach, recommendations are based on the similarity of terms between
users. Therefore, the active user is provided with items that his/her neighbors are
interested in. To build a suitable collaborative recommender system that produces
accurate recommendations, a proper algorithm of collaborative filtering should be
adopted. In fact, there are two types: algorithms based on the memory/user and the
ones based on models/items: [5].
• User based approaches: perform the prediction of QOS values based on the simi-
larities between users. It is about filtering users who have similar rating behavior
to the user in question, and consider their ratings on other items to generate what
might the current user actually like. [15]This mechanism uses a rating matrix,
similarity function, and a method that generates predictions based on the simi-
larities and ratings [16]. Some of the main advantages of this method is that it
is simple to implement and the resulting recommendations are easy to interpret.
Although this approach was broadly used, it turned out to be not easily scalable
and sometimes inaccurate. [17]
• Item based memory: this approach has been invented to overcome the shortcom-
ings of the user-user collaborative filtering method, which is mainly scalability.
This later is caused by the huge growth of the user bases and the operation of
searching for similar users is an O(|U|) or worst [15]. This is why scholars have
come up with a more scalable approach that can be extended to larger user bases
that is the item-item based collaborative filtering. Therefore, it has become one of
the most used techniques nowadays. To predict preferences, this method considers
the similarities between the items’ ratings’ patterns. For two items to be similar,
they need to have the same users like or dislike them. Therefore, end users tend
to have similar preferences for similar items [16].

Hybrid Approaches
Definition: The hybrid approach combines the collaborative filtering and content
based method in order to produce recommendations with higher accuracy. This
method has overcome the shortcomings of the collaborative filtering and content
based filtering when used individually. Nevertheless, the task of combining these
two methods in one recommendation process has turned out to be complex and time
consuming [9].
The mixture of these methods can be done in the following manner: Implementing
the algorithms separately and combining the outcome; Using the method of content
A Review on Content Based Recommender Systems in Tourism 533

based filtering in combination with the collaborative one; Using the method of collab-
orative based filtering in combination with the content based one; Building a unique
frameworks that uses both approaches [6].
Techniques of Hybridization
Weighted hybridization: This method generates recommendations through
combining results from several recommenders. It merges the scores of each of the
used techniques through a linear formula. P-tango is an example of the weighted
hybridized method. This system is made of content based and collaborative recom-
menders. Initially, they are assigned equal weights, that later are adjusted as
predictions are confirmed or not. The advantages of the weighted hybridization
method is that all strengths of the used recommender system are used during the
recommendation process in a smooth way [18].
Switching hybridization: According to this approach, a heuristic algorithm is
applied to swap to the most suitable recommendation techniques to generate a better
rating. This technique has been proposed to overcome the shortcomings specific
to one recommendation method such as the “new user problem” of content-based
recommendation, in this case, it switches to a collaborative recommendation system.
The main benefit of this switching hybrid method is its full awareness and sensitivity
to the strengths and weaknesses of its recommending components. However, the
main disadvantage of this approach is that it adds more complexity to the recom-
mendation process by integrating the switching criterion among the parameters of
the recommendation system [19].
Cascade hybridization: Within this approach, the recommended items are ordered
and refined through an iterative process. The recommendation list output by a tech-
nique are refined by another recommendation approach. Usually, the first list is
coarse. This hybridization technique is efficient and acceptant to the noise created
by the iteration process [20].
Mixed hybridization: This method combines recommendations generated from
several techniques instead of assigning one recommendation per item. Recommen-
dations associated to items are generated from different recommendation techniques.
Within this approach, individual performances do not affect the general picture [19].
Feature combination: The features generated by a certain recommendation tech-
nique are given as inputs to another recommendation technique. For example, in
collaborative filtering, the feature that is the ratings of users’ similarity is consid-
ered, in a case-based reasoning recommendation technique, among the features used
to generate the similarity between items [6].
Feature augmentation: The inputs of this method are the ratings and informa-
tion resulted from the previous recommender. The feature augmentation hybrids are
superior to the methods of feature combination as they add a few number of features
to the initial recommender.
Meta level the internal model: This method is generated by a recommendation
technique and used as input for another one. The generated model is richer in data
than single ratings. This type of hybrids had solved the issue of sparsity faced with
534 A. Bentaleb et al.

collaborative filtering techniques. This is because it uses the entire model learned by
the previous recommendation technique as input for next one [6].

3 Related Works

Throughout the literature, several authors have published surveys concerning the
systems of recommendation in the tourism field.
“Personalizing recommendations for tourists” is a survey proposing a state of
art of the tourism recommender systems (until early 2009). The presented systems
are grouped under several criteria including: the used personalization approaches
including mainly (Bayensian networks method and tools of decision making); the
chosen approach of recommendation (content, collaborative or hybrid approach); the
hardware devices, or the items used in for recommendations (restaurants, flights or
hotels). As a matter of fact, this survey presents these various systems per categories
but in a general way. It did not provide a detailed technical explanation on how they
have been built, therefore, it did not relate to the latest advances of technology lately
(exploration of social networks, advanced tools of geolocalization.) [2].
This survey presents an examination of the tourism recommender systems from
the perspective of sociologies, not from the innovative point of view. This paper
claims that smart frameworks are fundamental in the tourism industry as it includes
several aspects to be dealt with: the fact that data sources are distributed, the new
environments are full of risk and uncertainty; the tourists are mobile most of the
time. Within this survey, the author presents some systems that have dealt with these
issues. It also, provided an evaluation as far as the user interaction, the social side,
the context, and decision making mechanisms are concerned, to provide the best
services to end users. Nevertheless, this overview did not tackle the technical side of
those systems [21].
“The planning of cycle trips in the province of East Flanders” is another review
that presents some frameworks that have been introduced from the years 2001 to
2011 serving at building plans of trips. This survey’s authors compared the cited
systems regarding their various functionalities that are mainly planning trips. They
have used the following criteria for that concern: budget, weather circumstances, and
mandatory points of interests, hotel selections and routing. This paper tackles the
orienteering issue and how its extensions can be utilized to demonstrate trip planning
functionalities [22].
“Intelligent tourism recommender systems: A survey “ is another overview of the
tourism recommender frameworks. This paper has presented the most known smart
e-tourism recommender systems that have been published in the conferences and
Artificial Intelligence journals since 2008. Within this survey, the authors considered
different criteria including: the functionalities provided by these systems, the used
artificial intelligence approaches, the various recommendations algorithms used, and
different kinds of interfaces. As far as the details of building these recommendation
systems is concerned, this survey has presented some guidelines in that sense, it also
A Review on Content Based Recommender Systems in Tourism 535

mentioned that there might be some interesting future works in that fields for the
coming years [23].
Here is another survey that has presented a review of the state of art in that
field especially the mobile tourism recommendation systems. It provided insights
concerning the services they offer, the challenges they encounter along with the
promising areas of further research that can be conducted in the fields of mobile
recommender systems in tourism [24].
“A Short Survey of Recommendation Technologies in Travel and Tourism” is
another survey found in the literature tackling the same topic as this. It starts by
presenting the different paradigms of recommender systems used in e-commerce
applications. There are three approaches: the collaborative filtering, content based
and the hybrid paradigm, along with their advantages and disadvantages. Applica-
tions and examples have been used to better explain the paradigms. Then, the authors
moved to discussing the different applications of recommender systems, especially in
the tourism field. They have chosen that field since it is very important and dynamic,
and lately, it contributes to more than 50% of the global B2C turnover. They discussed
the shortcomings of existing applications. Compared to our survey, this one lacks the
detailed aspects. This has been noticed in the background part while presenting the
paradigms, it gave a short overview of them, also, when discussing the applications
of tourism, they used few ones, while there exist so many systems in the literature
that have implemented the recommender systems [25].
“A survey on mobile tourism Recommender Systems” is another survey found
in the literature and has tackled a similar topic, which is tourism recommenda-
tion mobile systems. It has presented a review of state of art of these systems,
and suggested a classification of mobile tourism recommendation systems. Later,
it discussed the possible research directions and the encountered challenges when
dealing with tourism mobile recommender systems. Notably, the accuracy of context
aware mobile RSs, the measurements of the quality of recommendations and consid-
eration of new aspects when planning tours/ routes for users. Compared to our survey,
this one focuses on mobile applications mainly [26].
“Recommender System Application Developments: A Survey” claims the impor-
tance of recommender systems as they reduce the overload caused by selecting the
most relevant services to end users from a tremendous amount of data based on their
preferences and profiles. It presents a review of the existing applications of recom-
mender systems in several categories: e-tourism, e-government, e-business. also, it
tackled the main approaches and techniques used in recommender systems. At last,
the authors identified new topics and directions for researchers to take. For the arti-
cles, this survey used two types: the first type is articles dedicated to explaining
the techniques and approaches used in recommendations systems while the second
type presented the different applications and case studies concerning recommender
system applications. The filtering and selection process of these articles have been
fulfilled in four steps. Compared to our survey, this one has tackled the topic of
recommender systems from a broad aspect as it included applications from several
domains. While our survey focuses in the tourism fields, this one handled the topic
of RSs in a general manner [27].
536 A. Bentaleb et al.

“A survey of food recommenders” has covered another kind of recommenda-

tion systems. These systems aim at assisting end users to choose what and where
to eat. This includes food, recipes and restaurant recommendations. Such applica-
tions/services are very useful as nowadays, there are plenty of food choices and
varieties. This article states the types of these systems depending on their purposes,
the different models and technical approaches, the datasets used to train models and
the challenges faced by these systems. Compared to our survey, this one is more
specific as it handles one specific aspect of tourism recommendation systems which
is food and restaurant recommendations [28].

4 Content-Based Recommender Systems for Tourism

Before going on a trip, end users prepare their plans before travelling. They search
for details of possible destinations, transportations, hotels, book reservation. As it
seems, it is very tedious to examine, estimate and compare all these aspects of
planning a tour. Here comes, the need of assistance to overcome the complexity of
this time consuming task; there have been various tourism recommender systems in
the literature that facilitate the operation of planning such activities [29].

4.1 Planning Recommendations

Here are some of the most known applications of recommender systems in the domain
of tourism and more specifically in the activity of planning:
“A hybrid recommendation approach for a tourism system” has presented a recom-
mender system for tourism using a methodology where classification based on associ-
ation is used. It is called Associative Classification Methods; it is about using concepts
from both association and classification to be able to employ association rules in a
prediction context. The associative classification could provide more reliable recom-
mendations on sparse data contexts whereas fuzzy rules allowed classifying a user
into multiple groups. Throughout case studies, it has been proven that such method-
ology could enhance the quality of generated recommendations and overcome the
shortcomings of already existing recommendations approaches [30].
“Sightseeing planning system (PSiS)” is another system that has been developed
to help users in planning what should be done in a certain place. It is based on several
aspects such: the location and the purpose of the visit, the time details and the weather
forecast. This recommender system is useful as it provides end users with refined
suggestions while they are planning for a tour or in real time when they are actually
on a trip [31].
“eCompass” is another application, that has been designed, to help users through
the public transit. It has been presented in both mobile and web platforms. It is a
service oriented architecture approach where outputs are given by end users in order
A Review on Content Based Recommender Systems in Tourism 537

to suggest a tour. A service for multimodal planning has been used in addition to
yahoo weather forecast services in order to plan trips. The resulted recommendations
for the tour is presented in either a map or a list view [32].
“Filter first; tour second” is another framework that has been proposed to produce
adequate recommendations. Its processing is based on online data sources and infor-
mation from social media. For the points of interest, there are some predefined ones
based on users’ preferences, while the rest is identified through the collaborative
filtering approach. This is done to maximize the satisfaction of end users. This
framework is based on two modules: 1) collaborative filtering engine to produce
the mandatory points of interests along with the optional ones, based on the popu-
larity of this POIs on social media and online reviews. 2) The iterated Tabu search
is used to generate the daily tours [33].

4.2 Travel Package Recommendations

Some of the applications that offer travel packages recommendations are: “Per-
sonalTour” is a multi agent recommender system that has been built to produce
adequate travel packages based on users’ preferences. This framework uses the AI
paradigm. The knowledge bases of this system store the recommendations of the
users along with their reviews of the offered travel services. The PersonalTour agent
transforms these reviews to confidence degrees given to the various services (flight,
hotel, attraction services) [34].
Another travel package recommendation framework is the “Object oriented RS”,
which generates personalized travel packages. This framework has proposed a new
way to represent travel packages where the object is the collection of feature value
pairs. In order to mine the implicit similarities between objects, two models are used:
The Objected Oriented Topic Model (OTM) and the Objected Oriented Bayesian
Network (OBN). Later on, the recommendations are generated based on these models
and using the collaborative filtering approach [35].
This application provides recommendations of personalized travel package for
several POIs using the crowd sourced user footprints gathered from the Jie Pang
system. The latter platform is used very widely in china to “check in” and sharing
locations, events and feedback about services. This application offers travel packages
with various types of locations with different sequences. The model of the data (user
preferences and POIs) is done through the spatiotemporal features used in check-in
records. Therefore, recommendations are generated based on user’s demands and by
joining the users’ preferences along with the spatiotemporal constraints [36].
538 A. Bentaleb et al.

4.3 Group Recommendations

In case of having a group of end users who would like to get recommendations before
starting a touristic activity, the following applications have brought some pertinent
solutions to help them in that:
“e-Tourism” is a web based application to provide tourists with recommenda-
tions concerning tours. This platform has been implemented specifically for the city
of Valencia and can be used by a single or a group of users. It is designed as a
set of modules. The main one is the “Generalist Recommender System Kernel”
(GRSK) which is responsible of generating recommendations in case of single user.
The system uses a hybrid approach that combines (demographic, content-based and
general likes filtering) to generate recommendations based on users’ preferences. In
the case of a group of users, recommendations are computed by the application of
the intersection and aggregation mechanisms on the individual preferences [37].
“A group recommender system for tourist activities” is another application that
has presented a promising method for generating adequate recommendations for a
group of users. During the process of generating recommendations, it combines the
following elements: group tastes, the users demographic classification and places
they visited in previous trips. The generated items are computed from the indi-
vidual personal recommendations. This platform combines the use of the following
technique: intersection, aggregation, and incremental intersection [38].
“A multi agent architecture for single user and group recommendation in the
tourism domain”, this paper presents a Multi Agent System used for planning touristic
activities for single or group of users at a given region. The agents interact to learn
dynamically the users’ profiles to generate a list of proposed activities for the end
users or groups based on the interaction of users with similar ones and the system
[39].
“Social group recommendation in the tourism domain”, this work suggests
a promising approach for generating recommendations as it includes not only
users/groups preferences, but considers the social relationship among them. Profiles
are built based on a combined approach that includes the demographic information
of users, the feedback of the community (collaborative filtering) and the descriptions
of the items to be recommended (the content based filtering). This hybrid profile
is used to generate ratings of unknown items. In the case of groups, the individual
ratings are weighted based on the social relationships between the group members
and the preferences of the rest of the team. Then, the estimations of the group are
obtained by combining the ones of the group members [40].

4.4 Point of Interest Recommendations

To help end users determine the most interesting points of interests, the following
applications have been designed for that purpose:
A Review on Content Based Recommender Systems in Tourism 539

“Smart personal assistant for historical tourism”, presents an application that

offers recommendation services for historical POIs based on knowledge bases on
internet. This application has been designed according to a smart space approach.
It used a semantic network to represent the preferred POIs input by the user, and
provide ranking for available POIs over the ones with personalization. Therefore, the
used smart space approach has brought noticeable advantages to the development of
such applications which are mainly: the data is represented using the most advanced
technologies offered by semantic web; the way end users interact with services has
been enhanced in a smarter way to overcome the shortcomings of traditional web
services [41].
“Ontological model of multi source smart space content for use in cultural heritage
trip planning” is another application that has used the smart space approach. It is based
on an ontological model with a combined approach that used several attributes:
the description of cultural heritage, service specific recommended areas, climate
time. Five Knowledge processors (KPIs) have been used as data sources: user’s
preferences, personal data, position, weather, geological information, event [42].
As known, one of the main challenges in designing a content-based recommender
system is to create a rich knowledge base of POIs along with their description and
their relevance of the topic of interest. Cicero is a content-based recommender system
for traveling, where its engine is a content analyzer. It built its knowledge base of
POIs description for a given geographic area, automatically by extracting it from
public data sources (OpenStreetMap and Wikipedia) [43].

5 Conclusion

Throughout this survey, an overview of recommender systems in general has been

presented. The main approaches of recommendations has been defined including:
content, hybrid and collaborative filtering. Later, the focus was on the content-based
recommendation systems architecture, stages, items representations and learning
users’ models. Then, in the related works section, a positioning of this survey has
been done compared to previous existing surveys. Finally, some of the well known
applications of contents based recommenders in the domain of tourism have been
cited.
After analyzing the applications cited in the last section “content based recom-
mender systems in tourism”, enhancements of the application Cicerco can be
proposed like using other data sources besides Wikipedia, including an automatic
process for itineraries computations, and adding the possibility to access them via
context aware mobile applications.
540 A. Bentaleb et al.

References

1. Adomavicius G, Tuzhilin A (2005) Toward the next generation of recommender systems: a

survey of the state of the art and possible extensions. IEEE Trans Knowl Data Eng 17:734–749
2. Kabassi K (2010) Personalizing recommendations for tourists. Telemat Inform 27(1):51–66
3. Schafer JB, Konstan JA, Riedl J (2001) E-commerce recommendation applications. Data
Mining Knowl Discov 5(1/2):115–153
4. Adomavicius G, Tuzhilin A (2005) Personalization technologies: a process-oriented perspec-
tive. Commun ACM 48(10):83–90
5. Roger Ciurana Simó E (2012) Development of a tourism recommender system. https://upc
ommons.upc.edu/bitstream/handle/2099.1/16444/Emili%20Ciurana%20Simo%20MIA.pdf.
Accessed 30 Sept 2020
6. Isinkayea FO, Folajimib YO, Ojokohc BA (2015) Recommendation systems: principles,
methods and evaluation. Egypt Inform J 16(3):261–273
7. Chen L, Wu J, Jian H, Deng H, Wu Z (2014) Instant recommendation for web services
composition. IEEE Trans Serv Comput 7(4):586–598
8. Han SM, Hassan MM, Yoon CW, Huh EN (2009) Efficient service recommendation system
for cloud computing market. In: Proceedings of the 2nd international conference on interaction
sciences: information technology, culture and human. ACM, pp 839–845
9. Mezni H, Abdeljaoued T (2018) A cloud services recommendation system based on fuzzy
formal concept analysis. Data Knowl Eng 116:100–123
10. Lops P, de Gemmis M, Semeraro G (1970) Content-based recommender systems: state of the
art and trends
11. Mitchell T (1997) Machine learning. McGraw-Hill, New York
12. Herlocker L, Konstan JA, Terveen LG, Riedl JT (2004) Evaluating collaborative filtering
recommender systems. ACM Trans Inf Syst 22(1):5–53
13. Zisopoulos C, Karagiannidis S, Demirtsoglou G, Antaris S (2008) Content based recommen-
dation systems
14. Pazzani MJ, Billsus D (2007) Content-based recommendation systems. In: The adaptive web.
Lecture notes in computer science, pp 325–341
15. Ray S, Sharma A (2011) A collaborative filtering based approach for recommending elective
courses. Communications in computer and information science
16. Ekstrand MD, Riedl JT, Konstan JA (2014) Collaborative filtering recommender systems. Now
Publishers, Hanover
17. Sallam R, Hussein M, Mousa H (2020) An enhanced collaborative filtering-based approach for
recommender systems. Int J Comput Appl 176:9–15. https://doi.org/10.5120/ijca2020920531
18. Suriati S, Dwiastuti M, Tulus T (2017) Weighted hybrid technique for recommender system.
J Phys: Conf Ser 930:012050. https://doi.org/10.1088/1742-6596/930/1/012050
19. Çano E (2017) Hybrid recommender systems: a systematic literature review. Intell Data Anal
21:1487–1524. https://doi.org/10.3233/IDA-163209
20. Burke R (2002) Hybrid recommender systems: survey and experiments. User Model User-Adap
Interact 12. https://doi.org/10.1023/A:1021240730564
21. Gretzel U (2011) Intelligent systems in tourism: a social science perspective. Ann Tourism Res
38(3):757–779
22. Souffriau W, Vansteenwegen P, Vanden Berghe G, Van Oudheusden D (2011) The planning of
cycle trips in the province of East Flanders. Omega 39(209–213):13
23. Borràs J, Moreno A, Valls A (2014) Intelligent tourism recommender systems: a survey. Expert
Syst Appl 41(16):7370–7389
24. Gavalas D, Konstantopoulos C, Mastakas K, Pantziou G (2013) Mobile recommender systems
in tourism. J Netw Comput Appl 39:319–333
25. Felfernig A, Gordea S, Jannach D, Teppan E, Zanker M (2007) A short survey of recommenda-
tion technologies in travel and tourism. OGAI J (Oesterreichische Gesellschaft fuer Artificial
Intelligence). 25:17–22
A Review on Content Based Recommender Systems in Tourism 541

26. Gavalas D, Kasapakis V, Konstantopoulos C, Mastakas K (2013) A survey on mobile tourism

recommender systems. In: 2013 3rd international conference on communications and infor-
mation technology, ICCIT 2013, pp 131–135. https://doi.org/10.1109/ICCITechnology.2013.
6579536
27. Lu J, Wu D, Mao M, Wang W (2015) Recommender system application developments: a survey.
Decis Support Syst 74:12–32. https://doi.org/10.1016/j.dss.2015.03.008
28. Anderson C (2018) A survey of food recommenders. arXiv, abs/1809.02862
29. Chaudhari K, Thakkar A (2019) A comprehensive survey on travel recommender systems.
Arch Comput Methods Eng
30. Lucas JP, Luz N, Moreno MN, Anacleto R, Figueiredo AA, Martins C (2013) A hybrid
recommendation approach for a tourism system. Expert Syst Appl 40(9):3532
31. Coelho B, Martins C, Almeida A (2009) Adaptive tourism modeling and socialization system.
In: 2009 international conference on computational science and engineering. IEEE, pp 645–652
32. Gavalas D, Kasapakis V, Konstantopoulos C, Pantziou G, Vathis N, Zaroliagis C (2015) The
eCOMPASS multimodal tourist tour planner. Expert Syst Appl 42(21):7303
33. Kotiloglu S, Lappas T, Pelechrinis K, Repoussis P (2017) Personalized multiperiod tour
recommendations. Tour Manag 62:76
34. Lorenzi F, Loh S, Abel M (2011) PersonalTour: a recommender system for travel packages.
In: 2011 IEEE/WIC/ACM international conference on Web intelligence and intelligent agent
technology (WI-IAT), vol 2. IEEE, pp 333–336
35. Tan C, Liu Q, Chen E, Xiong H, Wu X (2014) Object-oriented travel package recommendation.
ACM Trans Intell Syst Technol (TIST) 5(3):43
36. Yu Z, Xu H, Yang Z, Guo B (2016) Personalized travel package with multi point of interest
recommendation based on crowdsourced user footprints. IEEE Trans Hum Mach Syst 46(1):151
37. Garcia I, Sebastia L, Onaindia E (2011) On the design of individual and group recommender
systems for tourism. Expert Syst Appl 38(6):7683–7692 ISSN 0957-4174
38. Garcia I, Sebastia L, Onaindia E, Guzman C (2009) A group recommender system for tourist
activities. In: Proceedings of the 10th international conference on e-commerce and web
technologies, EC-web 2009. Springer, Berlin, pp 26–37
39. Sebastia L, Giret A, Garcia I (2011) A multi agent architecture for single user and group
recommendation in the tourism domain. Int J Artif Intell 6(11):161–182
40. Christensen I, Schiaffino S, Armentano M (2016) Social group recommendation in the tourism
domain. J Intell Inf Syst 47. https://doi.org/10.1007/s10844-016-0400-0
41. Varfolomeyev A, Korzun D, Ivanovs A, Petrina O, Arapatsakos C, Razeghi M, Gekas V (2014)
Smart personal assistant for historical tourism. In: Proceedings of 2nd international conference
on environment, energy, ecosystems and development (EEEAD 2014), pp 9–15
42. Kulakov K, Petrina O (2015) Ontological model of multi-source smart space content for use
in cultural heritage trip planning. In: 2015 17th conference of open innovations association
(FRUCT). IEEE, pp 96–103
43. Binucci C, De Luca F, Di Giacomo E, Liotta G, Montecchiani F (2017) Designing the content
analyzer of a travel recommender system. Expert Syst Appl 87:199
Towards Semantic Web Services Density
Clustering Technique

Naoufal El Allali, Mourad Fariss, Hakima Asaidi, and Mohamed Bellouki

Abstract With the rapid evolution of SOA (Service Oriented Architecture) and
wide techniques of web services, the web service semantic has attracted the most
attention during recent years. The WS (Web Service) has become a promising tech-
nology for the development, deployment, and integration of Internet applications. The
large-scale deployment of web services brings great value to web service customers;
however, the large-scale, dynamic, and heterogeneous web services make it difficult
to discover and compose WS. Therefore, as one of the challenges to interpreting rela-
tionships of semantics is to understand the functional associations between ontology
concepts. Several approaches have been described to extract relationships between
ontology concepts using term-matching methods. However, it is necessary to have
flexible and efficient methods to ensure functional relationships. To facilitate the
discovery the relevance of a service to a query, there are many solutions based on the
use of Semantic Web technologies to develop homogeneous service descriptions to
reason on and to support precise and flexible discovery. In this paper, we proposed
an approach to measure semantic similarity by combining Latent Semantic Analysis
Similarity (LSA) and IO-MATCHING semantic matching. The density clustering
services method is treated as a pre-processing method for semantic matching between
the web service and the user requirements. The results of the simulation show a high
accuracy to cluster semantically web services.

Keywords Latent semantic analysis · Density clustering · Ontology · Semantic

similarity measures · Semantic web services

1 Introduction

With the growing utilization of web services, the research community has attracted
a lot of attention in the area of service discovery and composition [1, 2]. However, it
remains difficult to accurately obtain the necessary information and knowledge from

N. El Allali (B) · M. Fariss · H. Asaidi · M. Bellouki

Laboratory MASI, FP, University Mohamed I, Selouane, B.P. 300, Nador, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 543
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_49
544 N. El Allali et al.

the data. WS registry search provides limited results for conventional approaches for
WS discovery due to the general unavailability of semantic descriptions of services.
Besides, with a growing number of Web services, these approaches generally result
in low productivity with marginal performance. For that reason, the majority of
the community is oriented to Semantic Web technologies (Fig. 1) [3], for example
RDF (Resource Description Framework) [4], RDFS (Resource Description Frame-
work Schemas) [5], OWL (Ontology Web Language) [6], to enhance Web services
as semantic descriptions. The combination of Web services and Semantic Web
technologies leads to the Semantic Web Services (SWC) concept.
Embedding Web Services in the Semantic Web domain is becoming ever more
interesting [7]. Using the concepts of the Semantic Web in the Web Services domain
provides a tool for using the meaning and significance of the data exchanged to
improve the feasibility of different tasks such as discovery, composition, etc. SWC
can be defined as Web services with semantic descriptions. This semantic is provided
through the ontologies, which is one of the most important technologies of the
Semantic [8]. It aims to describe concepts and the relationships that link them to
each other, and with rules of deduction to make them more understandable and
relevant.
The Latest solutions [9, 10] that aim at minimizing the discovery duration and
enhancing the precision ratio are becoming more important by clustering to reduce
the large-scale of the web services and to facilitate service discovery. Although
clustering is one of the most successful approaches of data mining [11], analysts
neglect sufficient resources for rapid, iterative clustering analysis, which is important
for the creation of conclusions and data rationale. A precondition for clustering
algorithms is the correct definition of a set of measures which exactly evaluate the
semantic similarity between services. under these conditions, Web service similarity
measurement and clustering has become a challenging issue. Therefore, to overcome
the problem of inaccurate service discovery in a fast and efficient way. We define a
combination of semantic similarity by subsumption reasoning and LSA, taking into
consideration the service clustering based on the DBSCAN algorithm [12] to reduce
the time of semantic matching between the requested service and massive service.
The remainder of this paper is organized as follows: The next section gives an
overview of related work. Section 3 introduces our proposed approach. Section 4
presents the experimental results. Section 5 points out the conclusion with the future
works.

2 Related Works

This section presents the different works that aimed at minimizing the discovery
duration and improvement the accuracy ratio for having an efficient system, with the
techniques of data-mining, in particular clustering analysis [13], can be used to group
similar services, it allows to reduce the matching time between services and to accu-
rately classify the web services in the repository. A large share of the studies focused
Towards Semantic Web Services Density Clustering Technique 545

Fig. 1

XML-S
Standard technologies for OWL
semantic web services
RDF-S

XML
RDF

HTTP URI

on this area develop new ontological methods using different Semantic similarity
algorithms [14, 15]. This paper is closely related to the semantic similarity measure
with the discovery service method by clustering. For that reason, we only present
related works in this direction:
In [16], the authors presented a new web services discovery mechanism based
on semantic matching, and the clustering of services is brought forward for efficient
and appropriate web services discovery, which combines functional similarity with
process similarity. Their proposed approach relies on the information available from
the semantic description model, based on the improvement of Lin’s [17] semantic
similarity measure to include opposition or degree of contrast as defined in [18].
Their objective is to develop the efficient WS discovery method by pre-clustering
that allows them to perform semantic and scalable WS discovery in a minimum of
time and thus reduce the search space. According to their results are acceptable in
terms of time compared to the results obtained by [19], but it remains a difficult
task to depend on semantic dictionary such as WordNet to obtain the oppositeness
relationships between concepts, which allows for not finding adequate services by
simple methods, and thus the algorithm is not accurate to describe semantically due
to the ignorance of the antisense relationships between two concepts.
In the work of [20] focuses on the web services clustering before discovering a WS
based on the OWL-S description language. They use two aspects to measure service
similarity (functional and process) for the aim of clustering services. A weighted
domain ontology approach has been used to compute functional similarity by using
I/O parameters with domain ontology developed with the help of WordNet and
available web ontologies. According to their experimental results showed for the
clustering algorithm based on the matching method of web service I/O [21] as an
efficient way to discover a WS.
The authors [22] proposed a hierarchical clustering approach for semantic WS
discovery. The main objective is to enhance the efficiency and accuracy of traditional
Web services discovery with the use of the vector space model. The Web services
are converted into a standard vector format via WS descriptions. Then, a semantic
analysis is used to reduce the size of the vector term with the help of WordNet. It
has performed the semantic expansion to respond to the user service need.
The authors [23] proposed a WordNet-VSM model to measure feature similarity
by generating service name, process and message vectors. Using unsupervised neural
546 N. El Allali et al.

networks based on kernel cosine similarity estimation rather than conventional clus-
tering algorithms. However, this approach has some challenges with clustering, such
as the output node linking to the weight and requiring a lot of processing.
In [24] proposed a service clustering approach inspired by the actions of ants.
Using a semantic service description related to a domain ontology that includes
concepts was used to annotate service input and output parameters. A set of metrics
was defined to compute the degree of match between services [21]. This approach has
been checked and validated on all the SWC available in the SAWSDL-TC benchmark
collection. The experimental results showed that there are many parameters that affect
the behavior of the service clustering algorithm.
As presented in [25], a new WS discovery approach is proposed, where common
topic groups are extracted from the service topic distribution matrix created by
topic modelling methods, and these topic groups support the process of matching
user requests with appropriate web services. The proposed approach can preserve
service discovery performance at a high level by considerably reducing the number
of candidate Web Services, consequently reducing response time.
The work proposed by [26] is to measure similarity based on WordNet in order to
determine the similarity between the service name and the text description using input
and output parameters based on the ontology of the domain. They use the OWL-S
files to extract the semantic information and the service similarity computed through
it. They are used as a K-medoids clustering algorithm.
In [27] proposed a method to facilitate cluster discovery of web services using
WSDL document meta tags. The efficiency of clustering was maintained by inco-
herent distribution and classification of noise tags. Moreover, the authors proposed
automated clustering of web services by tagging web services on a domain search
engine or UDDI or ontology. However, syntax-based approaches that have a lower
performance disadvantage due to the complexity of natural language, a variety of
semantic approaches are also suggested.

3 Proposed Approach Based on Density Clustering

As the availability of services increases and their diversity of types, so it becomes

more difficult to identify the most relevant services from the service sources. Besides,
the deployment of services is based on the needs of the user but is inhomogeneous.
Some regions have more dense services and some regions have dispersed services.
Therefore, this section deals with the issue of the service discovery accuracy by
DBSCAN algorithm with the use of a combination semantic similarity measurement
method.
Figure 2 represents the overall model for the proposed approach. In the first
level, is focused on the ontology description of the requested service as the OWL-S
formulation as the WS, therefore retrieve a more relevant service to the requirements
of the user. In the second layer, first, retrieve the desired headers of the cluster which
most closely matching the query, then searching for the nearest cluster and sorting by
Towards Semantic Web Services Density Clustering Technique 547

Extracting of target Searching for rele-

parameters and values. vant clusters

Retrieve the most

Query User (q) relevant service

Calculate the distance between the q and each cluster head, select the closest cluster

Cluster-1 Cluster-2 Cluster-3

Web Services

Cluster Head Cluster extent

Density Clustering Web service similarity Web service descrip- OWL-TC

Technique measurement tion handling
Dataset

Fig. 2 Overview of the proposed system

similarity to obtain the search results. At the third level, DBSCAN pre-clustering of
services to limit the search space and thus enable rapid semantic matching in large-
scale service pools. Generally, the performance of Web services discovery could be
improved by semantic matching and the rate of discovery could be improved by
pre-clustering.

3.1 Semantic Service Similarity Measurement

Semantic similarity measures of Web services offer many advantages and the majority
of the WS discovery methods depend on measuring the similarity of the wanted
service with each of the advertised services. The community of Semantic Web
Services has evolved many models for semantically representing global web services
such as Ontology Web Language for Services (OWL-S) and WS Modelling Ontology
(WSMO). However, it remains a challenge to evaluate the similarity between two
concepts in order to make an accurate selection of the most appropriate service. The
efficiency of clustering relies on the way similarity measurement has been performed
between data. In addition, there are solutions [16–26] based on semantic dictionaries
such as WordNet to interpret the relationships existing in the ontology between
two concepts, but it’ is limited to precision and automatization to describe services
semantically. With the description language OWL-S provides more flexibility and
efficiency semantic matching compared to WordNet. For another part, most of the
measures based on distance are based on counting the number of arcs that separate
two directions in a taxonomy such as the work of [28, 29], which just depend on
548 N. El Allali et al.

distance without describing the semantic relationships. To have an efficient clus-

tering of web services with accurate matching in different groups, we develop a new
semantic similarity measure combining IO-MATCHING [21] and LSA [30]. To
support the relationships existing in the domain ontology hierarchy, we will highlight
the different concepts to solve the problem mentioned previously:

Definition 1 (Query): A query is characterized as Q = I n q |Outq |PD B SC AN  where

I n q is the set of required input parameters, Out q is the set of required output
parameters and PD B SC AN is the set of the DBSCAN parameters.

Definition 2 (Web Service): Web Services are described by the OWL-S ontology. It
may be realized as a 4-tuple: W S = S N ame |ST ext Des |I n S |Out S  Where S N ame is the
service name, I n S and Out S are the input and output concepts respectively, ST ext Des
is a summary of what the service is providing.

Definition 3 (Latent Semantic Analysis (LSA)): the LSA [30] is based on a multi-
dimensional representation of the meaning of words in language. The idea behind
it is that words in a text corpus appearing on the same linguistic contexts have
similar meanings. It uses the algebra technique The Singular Value Decomposition
(SVD) which deals the transformation and decomposition of a matrix. Formally, the
word-service matrix decomposition of matrix A can be defined as:


A=U VT (1)

Where,
A is Input matrix (m × n)
is Words × extracted concepts (n × n)
U
Scaling values, diagonal descending matrix (n × n)
V T Documents × extracted concepts (n × n)
Exploiting OWL-S textual elements such as “<profile:textDescription>” and
“<profile:serviceName>” to provide us with more information for semantic
matching. To evaluate the similarity between two services, with a massive volume
of data, using Latent Semantic Analysis to reduce the dimensionality of our service
vectors instead of directly comparing the cosine similarity of the bag-of-word vectors.
After having a decomposed matrix, requires to use cosine similarity between services
either at the text description name or service name level. The measurement of cosine
similarity is defined as follows:

  Si · S j
Sim Si , S j =    (2)
 Si  S j 

where S i and S j represent services.

Towards Semantic Web Services Density Clustering Technique 549

Definition 4 (Similarity-Textual): The textual similarity between two services is

expressed through the service description and the service name in order to evaluate
the textual similarity between both services:

   
j j
Sim T extual Si , S j = ρ × Sim Descri ption STi ext Des , ST ext Des + σ Sim N ame S iN ame , S N ame (ρ, σ ) ∈ [0, 1]
(3)

where Si m Descr i pt i on and Si m N ame is calculated by the cosine similarity

respectively.

Definition 5 (Semantic Similarity Degree): The semantic degree matching is deter-

mined through the service semantic data of the service parameters: inputs: I n S and
Out S . The evaluation is based on the IO-MATCHING model which uses four types
of matching score: Exact, Plugin, Subsume and Fail introduced in [21], then to
calculate the matching between two I n S /Out S concepts is as follows:

(4)

where C1 and C2 are the services concepts respectively. to calculate the average
similarity of service concepts pair I/O is done as follows:
m 
j
  i
i
Match SC1 , SC2
Sim Match Si , S j = (5)
m
where m is number of concepts.

Definition 6 (Semantic similarity functionality): Semantic similarity measurement

is achieved by combining textual similarity and Semantic Similarity Degree:

     
Fun_sim Si , S j = α × Sim T extual Si , S j + β × Sim Match Si , S j (α, β) ∈ [0, 1]
(6)

where α is the textual similarity weight and β is the Semantic Similarity Degre
weight.

3.2 Semantic WS Based DBSCAN Clustering

As mentioned above, the deployment of services depends on the requirements of

the user. Therefore, it’s hard to separate similar services, the task becomes much
more difficult when the data includes noise and outliers. To handle a massive spatial
550 N. El Allali et al.

database, the DBSCAN algorithm allows us to detect any form in the search space,
unlike other clustering algorithms based on distance. The fundamental concept of
the DBSCAN is to identify high-density regions which are separated from each other
by low-density regions. The main parameters of the algorithm are eps (Maximum
radius of the neighborhood) and the MinPts (minimum number of points). The
concept of the eps neighborhood of Sj is given by:
 
 
Neps S j = S j ∈ D|dist Si , S j ≤ eps} (7)
 
To calculate the service similarity is done by the distance d i st S i , S j between
two services S i and S j :
   
dist Si , S j = 1 − Fun_sim Si , S j (8)

A point can be considered as a core point if the number of samples in the eps-
neighborhood is more than or equal to the MinPts density threshold. When there is
a core point in the neighborhood of a non-core point, the sample is referred to as a
boundary point. Noise is a data point that is neither a core nor a boundary point. The
core point set is referred to as d c is expressed by:

 
dc = S j ∈ D|Neps S j ≥ Min Pts} (9)

The set of boundary points is denoted as d b :

 
db = S j ∈ D\dc |Neps S j ∩ dc = ∅} (10)

The noise is the set of items in D that make no relation to any Ci cluster, this set
denoted as dn :

/ Ci }, i = 1, 2, . . . , l
dn = {Sk ∈ D\dc | ∀i : Sk ∈ (11)

If Si is directly reachable from a point q when it belongs to the neighborhood of

Sj . While Si is reachable from a point Sj if there is a sequence of points S1 , …, Sm
such that Si = S1 and Sn = Sj and that Si+1 is directly reachable from Si . A point Si is
density-connected to point Sj , whether there is a point Sk ∈ D such that both Si and

Table 1 The parameters of

Parameter Value
the proposed approach.
ρ 0.75
σ 0.25
MinPts 45
α 0.51
β 0.49
Towards Semantic Web Services Density Clustering Technique 551

Sj are Density-reachable from Sk . A cluster of DBSCAN is the maximum connected

density set resulting from a density-reachable relationship.

4 Implementation and Performance

We used the OWLS-TC version 4.0 dataset [31] to evaluate our proposed approach.
The dataset includes of 1083 indexed OWL-S services, that have all been collected
from the public IBM UDDI databases and semi-automatically converted from WSDL
to OWL-S. The experiments have been tested on a PC with an Intel ® Core (TM)
i7-4770 CPU and 8 GB main memory running Windows 10 operating system. Use
Java language programming to implement the DBSCAN Semantic Web Service
Clustering. The Parameters of the proposed method is defined as follows in Table 1:
To evaluate the performance of the proposed method, we have compared our
results with other work of [16].These results have been analyzed in terms of preci-
sion, recall, and F-Measure to prove the effectiveness of clustering approaches with
a suitable semantic similarity measurement method.
The results in the Fig. 3 shows that the proposed similarity method is highly
performant in terms of precision when the threshold of “eps-neighborhood” is equal
to 0.45, that means the performance of clustering web services will be more efficient
and more similar when the precision value is equal to 84.25% and the recall is equal
to 82.15%.
From the analysis of the results, it is possible to conclude the similarity of the web
services in this article is more efficient to support in either the discovery process or
in the composition phase. As compared to other approaches relying on the semantic
dictionary like WordNet as a solution but with less accuracy to find web services as
mentioned in Sect. 3.

0.8 Precision
Recall
0.6 F-Score

0.4

0.2

0 eps
0.25 0.35 0.45 0.55 0.65

Fig. 3 Evaluation of the proposed method based on semantic density clustering

552 N. El Allali et al.

5 Conclusion

In this work, we proposed a similarity method to group semantic web services effi-
ciently and accurately, as it enables us to quickly find web services either in the
discovery or in the composition process. Experimental results were more encour-
aging in terms of precision, recall, and f-score. As a future step in our work, it
consists of evaluating the effectiveness of this proposed approach during the compo-
sition process, since discovery is a more important phase to quickly find relevant
services, on the other side focusses mainly on enriching our model with semantic
descriptions by other functional aspects such as pre-conditions and post-conditions
and thus making the matching scheme more intelligent.

References

1. Garriga M, Flores A, Cechich A, Zunino A (2015) Web services composition mechanisms: a

review. IETE Tech Rev 32(5):376–383. https://doi.org/10.1080/02564602.2015.1019942
2. Lemos AL, Daniel F, Benatallah B (2015) Web service composition: a survey of techniques
and tools. ACM Comput Surv 48(3):1–41
3. Cabral L, Domingue J, Motta E, Payne T, Hakimpour F (2004) Approaches to semantic web
services: an overview and comparisons. In: European semantic web symposium, pp 225–239
4. Miller E (1998) An introduction to the resource description framework. Bull Am Soc Inf Sci
Technol 25(1):15–19
5. Tawfeq JF, Mohammed SM (2015) Resource description framework schemas for e-library. J
Madenat Alelem Univ Coll 7(2):26–35
6. Antoniou G, Van Harmelen F (2004) Web ontology language: owl. In: Handbook on ontologies.
Springer, Heidelberg, pp 67–92
7. Fensel D et al (2006) Enabling semantic web services: the web service modeling ontology.
Springer, Heidelberg (2006)
8. Fariss M, El Allali N, Asaidi H (2019) Review of ontology based approaches for web service
discovery. Springer, Cham
9. Bhardwaj KC, Sharma RK (2016) ontologies: a review of web service discovery techniques.
Int J Energy Inf Commun 7(5):1–12
10. Sagayaraj S, Santhoshkumar M (2017) A survey on clustering methods in web service
discovery. In: 2017 4th international conference on electronics and communication systems
(ICECS), pp 189–194
11. Han J, Pei J, Kamber M (2011) Data mining: concepts and techniques. Elsevier
12. Ester M, Kriegel H-P, Sander J, Xu X (1996) A density-based algorithm for discovering clusters
in large spatial databases with noise. In: KDD, vol 96, no 34, pp 226–231
13. Dong G, Bailey J (2012) Contrast data mining: concepts, algorithms, and applications. CRC
Press
14. Jiao H, Zhang J, Li JH, Shi J (2017) Research on cloud manufacturing service discovery based
on latent semantic preference about OWL-S. Int J Comput Integr Manuf 30(4–5):433–441
15. Chen F, Lu C, Wu H, Li M (2017) A semantic similarity measure integrating multiple conceptual
relationships for web service discovery. Expert Syst Appl 67:19–31
16. Chen F, Li M, Wu H, Xie L (2017) Web service discovery among large service pools utilising
semantic similarity and clustering. Enterp Inf Syst 11(3):452–469
17. Lin D (1998) An information-theoretic definition of similarity. In: ICML, vol 98, no 1998, pp
296–304
Towards Semantic Web Services Density Clustering Technique 553

18. Mohammed SM, Dorr BJ, Hirst G, Turney PD (2011) Measuring degrees of semantic
opposition. Technical report
19. Kokash N (2006) A comparison of web service interface similarity measures. In: STAIRS, pp
220–231
20. Xie L, Chen F, Kou J (2011) Ontology-based semantic web services clustering. In: 2011
IEEE 18th international conference on industrial engineering and engineering management,
pp 2075–2079
21. Paolucci M, Kawamura T, Payne TR, Sycara K (2002) Semantic matching of web services
capabilities. In: International Semantic Web Conference, pp 333–347
22. Gao H, Wang S, Sun L, Nian F (2014) Hierarchical clustering based web service discovery. In:
International conference on informatics and semiotics in organisations, pp 281–291
23. Chen L, Yang G, Zhang Y, Chen Z (2010) Web services clustering using SOM based on kernel
cosine similarity measure. In: The 2nd international conference on information science and
engineering, pp 846–850
24. Chifu VR, Pop CB, Salomie I, Dinsoreanu M, Acretoaie V, David T (2010) An ant-inspired
approach for semantic web service clustering. In: 9th RoEduNet IEEE international conference,
pp 145–150
25. Wang J, Gao P, Ma Y, He K, Hung PCK (2017) A web service discovery approach based on
common topic groups extraction. IEEE Access 5:10193–10208
26. Wen T, Sheng G, Li Y, Guo Q (2011) Research on web service discovery with semantics
and clustering. In: 2011 6th IEEE joint international information technology and artificial
intelligence conference, vol 1, pp 62–67
27. Wu J, Chen L, Zheng Z, Lyu MR, Wu Z (2014) Clustering web services to facilitate service
discovery. Knowl Inf Syst 38(1):207–229
28. Rada R, Mili H, Bicknell E, Blettner M (1989) Development and application of a metric on
semantic nets. IEEE Trans Syst Man Cybern 19(1):17–30
29. Wu Z, Palmer M (1994) Verb semantics and lexical selection, arXiv preprint
30. Dumais ST (2004) Latent semantic analysis. Annu Rev Inf Sci Technol 38(1):188–230
31. Klusch M, Khalid MA, Kapahnke P, Fries B, Vasileski M (2010) OWLS-TC OWL-S service
retrieval test collection
KSTAR: A Knowledge Based Approach
for Socially Relevant Term Aggregation
for Web Page Recommendation

Deepak Surya, Gerard Deepak, and A. Santhanavijayan

Abstract Recommendation System is a filtering technology used to compute an

item’s priority or rating to the user. The recommender system is used to generate audio
or video playlist for users in media streaming applications, recommend products in
e-commerce platforms, recommend posts on social media platforms, etc. The web
usage mining technique is incorporated to extract topical words from the user’s
historical web content. The data extracted is synonymized using WordNet, and a
standard knowledge base is created. Classification algorithms such as Random Forest
is employed on the knowledge base to classify the standard dataset, and the Ant
Colony Optimization has been encompassed to determine the shortest route through
graphs. This knowledge is fed to machine learning algorithms to classify the standard
dataset.

Keywords Knowledge base · Query language · Random forest classification ·

Synonymization

1 Introduction

Social tagging is a technique to identify and retrieve the contents available on the
internet based on their common attributes. In this process, a keyword is associ-
ated with any content on the internet, such as website content, social bookmarks,
photos, videos, blogs, etc. It is widely used in social networks, media service appli-
cations, and traditional websites. The users create tags, and there are not any rules
in naming them. Tags are named based on the features or characteristics of the web
content according to the user’s perspective. This paper introduces a novel approach
to generate recommendation tags based on the Random Forest Classification and Ant
Colony Optimization strategies.

D. Surya · G. Deepak (B) · A. Santhanavijayan

Department of Computer Science and Engineering, National Institute of Technology,
Tiruchirappalli, Tiruchirappalli, India
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 555
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_50
556 D. Surya et al.

Motivation: In this paper, the Web usage Mining technique is employed to extract
users’ profile data. Web usage Mining, a subset of Data Mining, extracts data from
the user’s historical data content on the internet, such as search logs, transaction logs,
social network activity, etc.

Contribution: In this paper, an approach to generate tags that accurately capture

the user’s intention is proposed. The classification algorithm is incorporated with
an optimization algorithm to find the highly relevant tags associated with a partic-
ular user tag. The results obtained from the Random Forest Classification and Ant
Colony Optimization have more accuracy and precision than other baseline models.
Experimental evidence supports the results obtained.

Organization: The remainder of the paper is organized as follows. Section 2 of

the paper consists of Related Works. The Proposed Methodology is described in
Sect. 3. Section 4 discusses the Implementation. Section 5 describes the Results and
Performance Evaluation. Finally, the paper is concluded in Sect. 6.

2 Related Work

De Caro et al. [1] have put forth a highly scalable tag-based recommender system that
can simultaneously suggest mixed types of content. It also provides suggestions to
enhance Software’s performance and reusability. Dorigo et al. [2] review the develop-
ments and the recent trends in the Ant Colony Optimization research. Dhandi et al. [3]
provide an overview of web mining techniques required to maintain and customize a
web application. Anitha et al. [4] a method to predict users’ behavior by employing
web usage mining techniques on users’ web server log files. It presents an overview
of the different algorithms and methods used to analyze the behavior of the user.
Wei et al. [5] suggested two deep learning-based models that can work effectively
by solving the cold start. The content of the features is extracted using a deep neural
network called SADE. The proposed solution enhances the past experience of the
user and the confidence of recommendation systems. Hameed et al. [6] categorized a
collaborative filtering recommendation system and the required criteria for selecting
the data set. It also discusses the measures to evaluate the recommendation system’s
performance, the uses, and the challenges faced by the collaborative filtering recom-
mendation system. Subramaniyaswamy et al. [7] proposed a novel model that can
generate many movie recommendations compared to other models using filtering
and clustering techniques. Jiang et al. [8] present a reliable novel multi-deployable
slope one algorithm developed by merging reliable data and user similarity approach.
The algorithm used in their work has higher accuracy compared to the conventional
slope one algorithm. Fu et al. [9] developed a novel deep learning-based method used
as a Collaborative Filtering recommendation model. The proposed algorithm gains
the semantic information revealing the correlation between user-user and item-item.
The algorithm also learns the low dimensional user and item vectors. Jain et al. [10]
KSTAR: A Knowledge Based Approach … 557

present an overview of movie recommendation systems. It proposes a model that uses

collaborative and content-based filtering techniques to generate movie suggestions.
Kawasaki et al. [11] proposed a recommendation system to suggest products for users
on e-commerce websites. A Matrix comprising of user items and the constraints in
the form of links are built using the purchase and browsing data. A semi-supervised
learning technique is employed on the constructed matrices for the generation of
a derived matrix comprising of user items. Pearson similarity is applied to yield
and compute the correlation between user and items, thereby to apply collaborative
filtering techniques. Davagdorj et al. [12] proposed an information filtering system
for rating prediction.

3 Proposed Methodology

3.1 Random Forest Classification

Random Forest is supervised learning technique used to sort the input data items
into predefined categories. It creates a decision tree based on different data samples
available in the data source and predicts it. There are so many trees of such is created
that it resembles a forest. A decision tree is a tree-shaped structure in which each
node performs a test on the attribute, and each branch of the tree represents a predic-
tion outcome of decision analysis. It can handle multi-output problems, i.e., both
numerical and categorical data. The final decision is based on merging each tree’s
prediction.

  
Q
Txky = (1)
k
Lk

Txky (2)
k

Tx y ← (1 − p)Tx y + Txky (3)
k
   β
Txαy ∗ ηx y
pxk y =    β (4)
α
z∈allowedx Tx y ∗ ηx y

In Eq. (1), Txy, gives the pheromone secreted by the kth ant moving from x to
y, Q is a constant, and Lk is the cost for the ant’s path. Equation (2) gives the total
amount of pheromone found along the path xy, ignoring the vaporization effect.
Equation (3) gives the total amount of pheromone found along the path xy consid-
ering the vaporization effect: p is the pheromone evaporation coefficient. Txy is the
558 D. Surya et al.

amount of pheromone deposited when the ant moves along xy. Equation (4) simulates
the probability of the ant to choose a path. Txy signifies the quantity of deposited
pheromone when the ant moves alongside the edge xy, nxy = 1/dxy (d is the distance
between x and y). The influence of Txy and the influence of (η)xy is controlled by
parameters α and β respectively.

3.2 Data Preparation

The user interests are captured and represented to predict the user’s intention. The
user social web data is examined to collect the keywords associated with the user. The
collected terms undergo feature population and synonymization. WordNet 3.0 is used
to perform synonymization. The architecture of the proposed KSTAR framework is
depicted in Fig. 1.

Fig. 1 Proposed architecture for KSTAR framework

KSTAR: A Knowledge Based Approach … 559

3.3 Classification of terms using Random Forest

The random forest algorithm is employed on the knowledge base to classify the
content. Each decision tree is constructed based on the CART (Classification and
regression tree) algorithm. The entire training dataset is fed as the input to the root
node of the decision tree. The root node performs the first classification and creates
two child nodes. The input data set is divided into two groups, and each group is
rendered as the input for each node. The child nodes further classify their corre-
sponding input data set into two groups. At each level, the input data set is classified
to obtain the purest possible distribution. This recursive division process repeats to
get a data set with zero uncertainty, i.e., only one label type. The random forest has
been chosen for various reasons. As the final prediction is based on all the decision
tree predictions, it reduces the risk of overfitting, thus increasing the model’s perfor-
mance. It reduces the training data time and can run efficiently on large datasets with
high dimensionality.

3.4 Optimizing the Classified Content Using ACO

Ant Colony Optimization algorithm is employed on the classified contents. It is used

to find the terms that accurately capture user intentions in the created dataset. It
outperforms other algorithms such as neural net, genetic algorithms, and simulated
annealing. It is chosen over other optimization algorithms as it works well in dynamic
applications and avoids early convergence. Finally, based on the semantic relatedness,
recommendation tags are generated.

4 Implementation

This paper presents an approach to suggest tags that are highly relevant to the user’s
intention. The algorithm’s input is the user’s intention, web usage data, and has
access to their social web data. Incipiently, the topics related to the web usage data
are extracted. The terms generated based on user intention from the user’s web usage
data undergoes the feature population. The terms obtained from the previous step
undergoes synonymization using WordNet 3.0. WordNet is a lexical database of
words, mainly English, linked by their semantic relationships.. Synonymization is a
technique to furnish the synonyms related to a word. This technique has been used
to generate semantically related words associated with the topical words extracted
using the web usage data. WordNet is used to determine the words that are semanti-
cally related to one another. This step is known as feature enrichment. The keywords
generated in the feature enrichment step are used to construct a standard knowl-
edge base like Dbpedia. It has been exploited to extract structured information from
560 D. Surya et al.

the information available in Wikipedia. The information from the database can be
modified and retrieved by a query language like SPARQL. SPARQL is an RDF
(Research Description Framework), a framework originally designed to represent
information in the Web, query language (i.e., semantic query language) for databases
The Real-world knowledge base that has been incorporated in this approach is the
DBpedia. The Random Forest algorithm is employed to classify the contents of the
dataset. Finally, Ant Colony Optimization is applied to each classified content. Upon
successful implementation, the algorithm generates tags that capture the user’s inten-
tion efficiently as the output. Algorithm 1 depicts the proposed Tag Recommendation
algorithm of the KSTAR Framework.

Algorithm 1: The proposed Tag Recommendation Algorithm of the KSTAR

Framework
KSTAR: A Knowledge Based Approach … 561

5 Results and Performance Evaluation

The dataset on which the experimentations for this research work was conducted is
the NASA web access log from 01/07/1995 to 06/07/1995 (six days). The dataset
consists of 4,49,911 records upon filtering and preprocessing. The proposed model is
baselined with VOMMRS [23] which employs Fuzzy C-Means Clustering with Vari-
able Order Markov Model for Recommendation. Fuzzy C-Means clustering works
reasonably well with Variable Order Markov’s Model, but the approach does lag
since they used only Web log data without using auxiliary information. There is a
need for extensive Web log information or the users’ historical information, which
makes this method entirely user centric. The fuzzy C-Means algorithm also has some
disadvantages associated with it. The computational time is relatively high, and it is
susceptible to noise (outlier points). The main limitation of Variable Order Markov
mode is that it cannot express the relations or dependencies between the hidden
states. Table 1 depicts the performance evaluation for the proposed KSTAR and its
comparison with other baseline models like GAAHC, IFWIAR, and VOMMRS.
IFWIAR [21] used Fuzzy Weighted Iterative Association Rule-Based Ontology
Model for query recommendation. A static domain ontology model restricted to a
specific domain, namely the Healthcare domain, has been proposed. The model
performs well, but the fuzzy weighted iterative association rule mining can be
improved based on the amalgamation of the recent algorithms. GAAHC [22] uses the
Group Average Agglomerative Hierarchical Clustering. It performs well over a static
dataset but does not when dynamic knowledge has been supplied to the approach.
The approach focuses on the ranking webpages based on time and frequency. The
modified Levenshtein distance has also been incorporated, and the personalization
is the focus of the approach. However, there is a need for the web search to be made
knowledge-centric and user-relevant. Thus, a newer paradigm is needed.
KSTAR employs Random Forest Classification and Ant Colony Optimization. It
performs much better than the baseline approaches because it captures users’ inten-
tions and extracts socially relevant terms from the Social Web APIs. The topic extrac-
tion is initially extracted from the users’ web usage information. The model focuses
on the feature population based on synonymization and the extraction of trending

Table 1 Evaluation parameters comparison with other baseline models.

Parameters Proposed KSTAR GAAHC IFWIAR VOMMRS
Precision (%) 89.86 84.33 83.17 71.63
Recall (%) 93.72 86.17 87.33 78.81
Accuracy (%) 91.79 85.25 85.25 75.22
F-Measure (%) 91.74 84.78 85.19 74.37
FDR 0.11 0.16 0.17 0.29
Filtration Rate (%) 80.12 68.12 63.17 61.45
Computational time (ms) 41 62 51 63
562 D. Surya et al.

terms from the Social Web. Apart from this, the DBpedia knowledgebase populates
the real-world knowledge into the model. The terms and knowledge aggregated based
on the user intentions are used to classify the datasets using the Random Forest, which
is quite efficient when the knowledge density is relatively high. However, ACO is
incorporated to optimize the recommendation sets based on the classified content
by prioritizing users’ web usage data. There is a precise balance between the users’
past information and the available real-world knowledge, which makes KSTAR quite
efficient than the baseline models.
Figure 2 depicts the distribution of accuracy in percentage against the number
of recommendations generated. It is evident from the graph that the suggested
approach has better accuracy than any other baseline models. The precision in the
proposed model is more remarkable than GAAHC by 5.53%, IFWIAR by 6.79%, and
VOMMRS by 18.33%. The recall percentage of the proposed model is higher than
GAAHC by 7.02%, IFWIAR by 7.55%, and VOMMRS by 14.91%. The accuracy
percentage in the proposed approach is greater than GAAHC by 6.54%, IFWIAR by
6.54%, and VOMMRS by 16.57%. The F-Measure percentage of KSTAR is more
significant than GAAHC by 6.96%, IFWIAR by 6.55%, and VOMMRS by 17.37%.
The FDR of the proposed technique is lesser than GAAHC by 0.05, IFWIAR by
0.06, and VOMMRS by 0.18. The filtration rate percentage of KSTAR model is
greater than GAAHC by 12%, IFWIAR by 16.95%, and VOMMRS by 18.67%. The
computational time of KSTAR is lesser than GAAHC by 21 ms, IFWIAR by 10 ms,
and VOMMRS by 22 ms.

Fig. 2 Accuracy distribution

graph
KSTAR: A Knowledge Based Approach … 563

6 Conclusions

It is very evident from the results that the proposed method, KSTAR operates much
better than the other baseline models. The terms generated from the user’s web
usage data undergoes the feature population and synonymization. This approach also
considers the user’s social web data to generate terms that match the user’s intent.
A real-world knowledge base is constructed while other approaches use domain
ontology. The Random Forest Algorithm classifies the content, and ACO is applied
to the classified content to select the most relevant tag in response to a particular user
tag. Experimental results reveal an average accuracy and F-measure of 91.79% and
91.74%, respectively, higher than the other baseline approaches.

References

1. De Caro P, Pini MS, Sambo F (2016) A tag-based recommender system. In: Intelligent
autonomous systems 13. Springer, Cham, pp 1049–1061
2. Dorigo M, Stützle T (2019) Ant colony optimization: overview and recent advances. In:
Handbook of metaheuristics. Springer, Cham, pp 311–351
3. Dhandi M, Chakrawarti RK (2016) A comprehensive study of web usage mining. In: 2016
symposium on colossal data analysis and networking (CDAN). IEEE, pp 1–5
4. Anitha V, Isakki P (2016) A survey on predicting user behavior based on web server log files in a
web usage mining. In: 2016 international conference on computing technologies and intelligent
data engineering (ICCTIDE 2016). IEEE, pp 1–4
5. Wei J, He J, Chen K, Zhou Y, Tang Z (2017) Collaborative filtering and deep learning based
recommendation system for cold start items. Expert Syst Appl 69:29–39
6. Hameed MA, Al Jadaan O, Ramachandram S (2012) Collaborative filtering based recommen-
dation system: a survey. Int J Comput Sci Eng 4(5):859
7. Subramaniyaswamy V, Logesh R, Chandrashekhar M, Challa A, Vijayakumar V (2017) A
personalised movie recommendation system based on collaborative filtering. Int J High Perform
Comput Networking 10(1–2):54–63
8. Jiang L, Cheng Y, Yang L, Li J, Yan H, Wang X (2019) A trust-based collaborative filtering
algorithm for E-commerce recommendation system. J Amb Intell Hum Comput 10(8):3023–
3034
9. Fu M, Qu H, Yi Z, Lu L, Liu Y (2018) A novel deep learning-based collaborative filtering
model for recommendation system. IEEE Trans Cybern 49(3):1084–1096
10. Jain KN, Kumar V, Kumar P, Choudhury T (2018) Movie recommendation system: hybrid
information filtering system. In: Intelligent computing and information and communication.
Springer, Singapore, pp 677–686
11. Kawasaki M, Hasuike T (2017) A recommendation system by collaborative filtering including
information and characteristics on users and items. In: 2017 IEEE symposium series on
computational intelligence (SSCI). IEEE, pp 1–8
12. Davagdorj K, Park KH, Ryu KH (2020) A collaborative filtering recommendation system
for rating prediction. In: Advances in intelligent information hiding and multimedia signal
processing, pp. 265–271. Springer, Singapore
13. Deepak G, Santhanavijayan A (2020) OntoBestFit: a best-fit occurrence estimation strategy
for RDF driven faceted semantic search. Comput Commun 160:284–298
564 D. Surya et al.

14. Pushpa CN, Deepak G, Kumar A, Thriveni J, Venugopal KR (2020) OntoDisco: improving web
service discovery by hybridization of ontology focused concept clustering and interface seman-
tics. In: 2020 IEEE international conference on electronics, computing and communication
technologies (CONECCT). IEEE, pp 1–5
15. Kumar A, Deepak G, Santhanavijayan A (2020) HeTOnto: a novel approach for concep-
tualization, modeling, visualization, and formalization of domain centric ontologies for heat
transfer. In: 2020 IEEE international conference on electronics, computing and communication
technologies (CONECCT). IEEE, pp 1–6
16. Deepak G, Kasaraneni D (2019) OntoCommerce: an ontology focused semantic framework
for personalised product recommendation for user targeted e-commerce. Int J Comput Aided
Eng Technol 11(4–5):449–466
17. Gulzar Z, Leema AA, Deepak G (2018) PCRS: personalized course recommender system based
on hybrid approach. Procedia Comput Sci 125:518–524
18. Deepak G, Teja V, Santhanavijayan A (2020) A novel firefly driven scheme for resume parsing
and matching based on entity linking paradigm. J Discret Math Sci Cryptogr 23(1):157–165
19. Haribabu S, Sai Kumar PS, Padhy S, Deepak G, Santhanavijayan A, Kumar N (2019) A
novel approach for ontology focused inter-domain personalized search based on semantic set
expansion. In: 2019 fifteenth international conference on information processing (ICINPRO),
Bengaluru, India, pp 1–5. https://doi.org/10.1109/ICInPro47689.2019.9092155
20. Deepak G, Kumar N, VSN Sai Yashaswea Bharadwaj G, Santhanavijayan A (2019) OntoQuest:
an ontological strategy for automatic question generation for e-assessment using static and
dynamic knowledge. In: 2019 fifteenth international conference on information processing
(ICINPRO). IEEE, pp 1–6
21. Santhanavijayan A, Naresh Kumar D, Deepak G (2021) A semantic-aware strategy for auto-
matic speech recognition incorporating deep learning models. In: Intelligent system design.
Springer, Singapore, pp 247–254
22. Deepak G et al (2019) Design and evaluation of conceptual ontologies for electrochemistry as
a domain. In: 2019 IEEE international WIE conference on electrical and computer engineering
(WIECON-ECE). IEEE
23. Deepak G, Priyadarshini JS (2018) Personalized and enhanced hybridized semantic algorithm
for web image retrieval incorporating ontology classification, strategic query expansion, and
content-based analysis. Comput Electr Eng 72:14–25
24. Sumathi G, Akilandeswari J (2020) Improved fuzzy weighted-iterative association rule based
ontology postprocessing in data mining for query recommendation applications. Comput Intell
36(2):773–782
25. Bt HK, Vibha L, Venugopal KR Personalized recommendation of web pages using group
average agglomerative hierarchical clustering (GAAHC)
26. Gopalakrishnan T, Sengottuvelan P, Bharathi A, Lokeshkumar R (2018) An approach to
webpage prediction method using variable order Markov model in recommendation systems.
J Internet Technol 19(2):415–424
MKSTD: Multiple Knowledge Sourcing
Approach for Topic Directories
Construction and Linking Web Pages
Using Fusion Semantics

Gerard Deepak and A. Santhanavijayan

Abstract Mapping of Knowledge sources is one of the preliminary strategies for

making Semantic Search foundational. The Web is flooded with information but
making it meaningful requires the integration of useful resources and attenuated
semantics. In this paper, an approach for integrating knowledge from several hetero-
geneous sources such as the knowledge bases, crowdsourced knowledge, and tradi-
tional web directories using an integration of semantic similarity and Simpson’s
Diversity Index has been proposed for the construction of hierarchical topic directo-
ries, and yield topic-specific Linked Open Data. The proposed MKSTD framework
is a knowledge sourcing and integration scheme for Linked Open Data generation
and further leveraging the web pages to the generated topics based on the index
words and the content words of the web pages for facilizing the Semantic Search.
The MKSTD is evaluated on the Bookshelf dataset from the Gutenberg Project cate-
gorizations and an average F-Measure of 95.98% has been achieved with an nDCG
score of 0.91 and a very low FDR of 0.06 for leveraging the indexed web pages with
the topic hierarchies.

Keywords Knowledge bases · Topic directory · Linking web pages · Multiple

knowledge sourcing · Simpson’s diversity index

1 Introduction

The existing structure of the Web is tending towards “Web of Knowledge” and Web
2.0 is converging into the Semantic Web 3.0. The Web contains a lot of hierarchical
data that is organized into a classified structure, but it is highly disintegrated in
several pockets and not structured together. This creates a distinction between the
integrative nature of the Web information. There is always a necessity to integrate
several relevant contents in the World Wide Web based on topical hierarchy and
link the entities in the Web together to facilitate Semantic Search. The structure of

G. Deepak (B) · A. Santhanavijayan

Department of Computer Science and Engineering, National Institute of Technology,
Tiruchirappalli, India

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 565
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_51
566 G. Deepak and A. Santhanavijayan

the Web is beyond one’s imagination as the entire structure of the Web cannot be
visualized together owing to the millions of zettabytes of data. The size of the indexed
Web alone is approximated to 5.59 billion pages. The only strategy to align the web
pages with the underlying semantic layer of the web is to generate topic directories
and facilitate a link among the topics in the Web to furnish Open Linked Data for
individual topics and anchor the indexed web pages to these topics based on the nature
and the structure of the World Wide Web. Dynamic Generation of Topical Knowledge
from several highly significant data repositories like the Knowledge Bases, Crowd-
Sourced Intelligence, and the traditional Web Directories for facilitating linking of
the web pages is the key towards knowledge-centric semantic search. Most of the
existing approaches either use static Ontological Models as knowledge or use one of
the standard Knowledge bases or a combination of the two. A few other approaches
did not make use of Ontological Models but depended on Web Directories, and a
static knowledge thesaurus. Labels and Embeddings [1] have also played a vital role
in semantic and language processing applications. However, since the present Web
has a facility of Generating Dynamic Knowledge by aggregation of these sources of
Knowledge, a new paradigm for semantic search with increased knowledge density
and a higher degree of relevance is the key towards a Knowledge Fusion Paradigm
in facilitating Semantic Search.
Motivation: Due to the density of the Semantic Web, there is a presence of a large
number of topics that are related. There is always a need for generating topic hier-
archies and leveraging the existing web pages to these topic hierarchies. This can
facilitate Semantic Search by accelerating the quality of recommendations from the
linked data standards of the Semantic Web 3.0. The existing structure of the Web
is also organized to a certain extent, but useful data resources are either separately
stored in Knowledge Bases or are distributed as a collective collection in domain-
based websites or blogs or the traditional and the forgotten Web Directories. An
approach to integrate all these contents to generate Topic Hierarchies and anchor
web pages to them is the need of the hour.

Contribution: An effective MKSTD framework for mapping of web pages to varied

categories has been achieved based on fusion semantics that combines a Semantic
Similarity Model with a Diversity Index. The approach follows the harvesting of
dynamic knowledge from several real-world classes of knowledge sources which
comprise knowledge bases, crowd intelligence, Domain Ontologies, and traditional
Web directories. The MKSTD generates topic-specific Linked Open Data based
on the topic categories with the focus of relevance and diversification and further
leverage web pages to the topic hierarchies of Linked Open Data. The MKSTD
is one of the first frameworks to link several classes of knowledge using several
knowledge sources using SPARQL Endpoints. A fusion semantics by combining the
Jaccard Similarity and Simpson’s Diversity Index has been encompassed to achieve
the best-in-class performance for the Gutenberg Project Categorizations.

Organization: This paper is organized as follows. Section 2 provides an overview

of the related research work. Section 3 presents the Proposed System Architecture.
MKSTD: Multiple Knowledge Sourcing Approach for Topic Directories … 567

Implementation is discussed in Sect. 4. Performance Evaluation and Results are

discussed in Sect. 5. This paper is concluded in Sect. 6.

2 Related Work

Sun et al. [1] have proposed a scheme for leveraging label and word embedding
for the semantically driven discovery of web services in a sparse environment. The
approach used label based neutral topic model for web services retrieval based on
categories of web services. Qi et al. [2] have put forth a keyword-driven strategy for
test model generation for web applications by discovering relevant states and transi-
tions. Delpeuch et al. [3] have proposed a simple scheme for Named Entity linking for
aggregating contents from Wikidata to facilitate lightweight Entity linking. Carlos
et al. [4] have introduced an LDA with a probabilistic topic model for Topics
Discovery in Semantic Markup Literature. Sahiti et al. [5] have proposed a strategy
to link lecture slides to facilitate navigation such that a Web of Slides using the simi-
larity of slides as a core concept. Patricia et al. [6] proposed a four-stage approach for
linguistic Linked Open Data generation for the Legal Domain based on identification,
generation of language resources and further converting into RDF, and finally linking
the resulting terminological resources. El Vaigh et al. [7] have put forth a strategy of
linking Entities based on the semantics of the knowledge base. The interrelationship
semantics based on weighting the relations based on the RDF KB structure has been
proposed. Jang et al. [8] have proposed an approach for the automatic construction
of Taxonomy from the Text-Rich Network based on an instance level selection and
term embedding based on the semantics and node granularity. The approach also
integrates clustering to yield subtopics for the taxonomy node. Venugopal et al. [9]
have proposed a Levenshtein Similarity Weight computation scheme for mapping
web pages with topic directories by also incorporating a hashing scheme for linking
web pages with topic hierarchies. Leena et al. [10] have proposed the MA algorithm
for the generation of related clustered hierarchy for the Semantic Web for Keyword
Search which takes the level information and the entity value for fetching the clusters
which match the input keyword where a concise summary tree is constructed. Deepak
et al. [11] have proposed the OntoBestFit which predominantly uses the Simpson’s
Diversity Index with the EnAPMI measure in an environment of static and dynamic
Ontologies for yielding web pages based on detailed ontology semantics. Giri et al.
[12] have proposed the OntoYield which is a semantic similarity-driven contextual
method for the generation of ontologies where the structure of the relationships is
preserved. In [13–16] several Ontology-driven approaches in several applications
have been proposed.
568 G. Deepak and A. Santhanavijayan

3 Proposed System Architecture

The architecture of the web pages to topic mapper has been proposed which is a
dynamic web page to a topic mapper that mainly focuses on linking the real-world
knowledge with the existing web pages based on the indices of the web pages which
ensures the formulation of a reasonable Semantic Web. Since the World Wide Web
is transfiguring into a Web of Data, there is a need for organizing the contents on
the World Wide Web into topic hierarchies such that newer links can be formu-
lated among the existing knowledge bases. The easiest way to do this is to build a
dynamic agent for aggregating terms from various sources and intelligently collect
the crowd intelligence based on the topic-based directory for browsing the data on
the World Wide Web. The users’ intelligence must be integrated into the existing
topic directories to create newer links or even loops on the contents of the World
Wide Web.
The architecture of the proposed Semantic Mapper is based on the Diversification
of the Terms by integrating real-world knowledge into the proposed approach. In
the proposed approach the Linked Open Data is generated, and the web pages are
mapped based on their indexing and contents. The hierarchical organization of the
topics is checked pertaining to seven-stage levels of relevance when looked on with
the parent term. Also, the anchorage of web pages with the terms that have been
discovered is evaluated for its relevance. The architecture of the proposed MKSTD is
depicted in Fig. 1 where an array of knowledge sources is integrated for a hierarchical
topic generation. The integration of static local Domain Ontologies with real world
knowledge bases namely the DBPedia, YAGO, NELL, Wikidata, DebatePedia, and
even quite others has been common. However, in the Semantic Web Mapper, apart
from these knowledge bases and the Domain Ontologies modeled, the sessional data
for every user query from several profiles of web usage data has been considered.
Apart from the individual user-click information of the selective users, certain old-
fashioned web directories are crawled and are made use for experimentation, apart
from the index words or labels in the Web dataset.
The Web Directories are a useful source of information and serve as a massive
knowledge base. The experimentation queries were used to crawl the contents from
these old-fashioned directories which serve as a ready knowledge base that has been
forgotten to be an intrinsic part of the web search engines. However, for populating
the individuals to the entities for the facts in the knowledge bases, the Web Directo-
ries are quite essential. Moreover, when the entities from Web Directories have been
anchored to the ontologies and knowledge bases, the density of information over a
spear of relevance increases, thereby the probability of knowledge for a specific user
query or a topic increases enhancing the diversity of results that are recommended
for a specific topic or a user query. The Web Directories that have been encompassed
are the Business Bureau, BOTW, Yelp, Foursquare, and HotFrog. The interface of
the user-intent and query words is integrated through an Agent into these Web Direc-
tories and their web repositories are crawled using a localized crawler. Owing to
MKSTD: Multiple Knowledge Sourcing Approach for Topic Directories … 569

Fig. 1 Architecture of the proposed MKSTD framework

the incorporation of these directories for the user intents and queries, the amount of
information density of facts over a specific topic is quite high.
A topic-based agent is employed for crawling domain adherent knowledge from
Knowledge Bases and stored as a Subject, Predicate, and Object from the series of
Web Directories. However, SPARQL driven agents have been employed to query the
Knowledge Bases based on the user intents and user query topics. Apart from this,
the experimentation topics were given to several users’ and their browsing history
was collected, and a topic mapping based on the relevance of terms among the
user intents have been mapped as an open topic directory and are mainly used for
leveraging the web pages onto these topics. The relevant items are selected based
on the semantic similarity computation which is achieved using the Jaccard Index
[17, 18] and the Simpson’s Diversity Index [19] to yield the diversified entities. The
570 G. Deepak and A. Santhanavijayan

reason for using the Jaccard Index is mainly owing to its ease of use and its efficient
operability between the sample sets. The Simpson’s Diversity Index works well
especially in an environment where there is a high degree of diversity in entities that
are relevant. In this case, owing to the heterogeneity of Knowledge Resources such
as the Web Directories, Static Domain Ontologies, Real-World Knowledge Bases,
Crowdsourced Topic Hierarchies in both the concept selection and the selection of
individuals.

4 Implementation

The proposed system is implemented in JAVA and the SPARQL Endpoints are
embedded in a JADE driven multiagent. The Domain Ontologies are modified using
Web Protégé and OntoCollab. The knowledge is modeled using the RDF which
is expressed in N-Triple format for storing the crowdsourced knowledge as Facts.
The Web directories are also accessed using the Multiagent JADE system. Stanford
CoreNLP has been encompassed for pre-processing and for all language processing
tasks. The proposed MKSTD algorithm has been depicted in Table 1 for Linked
Open Data Generation and linking of web pages with the generated topic hierarchy.

5 Results and Performance Evaluation

The experimentations were carried out for the Bookshelf from the Gutenberg Project
categorizations (https://www.gutenberg.org/ebooks/bookshelf/). The categorizations
in the Gutenberg Project Bookshelf were used as a topic directory for generating topic
hierarchies and thereby anchoring web pages to the topic directories. Standard metrics
such as the Precision, Recall, Accuracy, F-Measure, nDGC (Normalized Discounted
Cumulative Gain), and the False Discovery Rate (FDR) was used for evaluating the
performance of the proposed approach. The nDGC computes the diversity in the
furnished results. However, the other metrics compute the relevance of the results
yielded by the proposed approach.
It is evident from Table 2 that the proposed MKSTD yields an average Precision
of 94.89%, an average Recall of 97.04%, an average Accuracy of 95.97% with
a very low FDR of 0.06, and an nDGC score of 0.91. The proposed MKSTD is
baselined with two approaches namely the MPC-LSW [9] and the MA Algorithm
[10] and several variations by excluding the integrated approaches are used as a
standard for comparing the efficacy of the approaches integrated into the MKSTD.
The proposed MKSTD has a 6.75% of higher precision, 6.9% of higher recall, than
the MPC-LSW, and 6.83% of accuracy higher than that of MPC-LSW. MKSTD has
achieved an increase in the Precision, Recall, and Accuracy of 11.71%, 11.16%, and
11.84% respectively when compared to the precision, recall, and accuracy of the MA
Algorithm. When the Web Directories were removed from the proposed MKSTD,
MKSTD: Multiple Knowledge Sourcing Approach for Topic Directories … 571

Table 1 Proposed MKSTD algorithm for linked open data generation and mapping of web pages
Input: The query Q (Query Words can be changed to terms) which is input as a
keyword or a set of multi-keyword, Query Word Based Classified Knowledge
Resources namely the Directories, Ontologies, Real-World Knowledge Bases, and
Crowdsourced Topic Hierarchies.
Output:
Begin
Step 1: The Query Q if multiword is preprocessed by lemmatization, stemming, and
Stop word removal.
Step 2: for each qw in Q
Set Onto← Derive EntitiesKnowledge from the Static Domain Ontologies
end for
Step 3: while (Onto.next()!=NULL)
Jd←JaccardSimilarity(Onto(i), SPARQLEndpoints(Kb)
Sd← SimpsonsDiversityIndex(Onto(i),Kb)
i++;
return Fact←Jd U Sd
end while
Step 4: while (Fact.next()!=NULL)
Jd←JaccardSimilarity(Fact(i), CrowdEmbeddings)
Sd← SimpsonsDiversityIndex(Fact(i), CrowdEmbeddings)
i++;
return EnrichedFact←Jd ∩ Sd
end while
Step 5: Using an agent, query the entities in the EnrichedFact to Web Directories and
populate the individuals from the list of Web Directories to Generate OLData.
Step 6: Based on the Computation of the Jaccard Similarity and the Simpsons
Diversity Index, link the web page URLs with the OLData entities, based on
the index words and the content words of the page.
Step 7: Recommend the OpenLinkedData OLData, the Mappings of the Topic
Hierarchies to the user.
end

Table 2 Performance analysis of the proposed MKSTD framework

Model/Variants Precision % Recall % Accuracy % FDR nDCG
MPC-LSW [9] 88.14 90.14 89.14 0.12 0.81
MA algorithm [10] 83.18 85.08 84.13 0.17 0.78
Without web directories in the 87.16 89.02 88.09 0.13 0.82
proposed approach
Eliminating the Simpson’s diversity 89.12 91.63 90.38 0.11 0.76
index in the proposed approach
Single source knowledge base 87.63 90.78 89.21 0.13 0.82
Proposed approach with multiple 94.89 97.04 95.97 0.06 0.91
knowledge bases, web directories, and
diversity index (MKSTD)
572 G. Deepak and A. Santhanavijayan

the accuracy receded by 7.88%. Similarly, when the Diversity Index was removed
from MKSTD, the accuracy decreased from 95.97 to 90.37%. When there was only
a single knowledge base in the proposed approach, the overall accuracy decreased by
6.76%. Since the Accuracy is derived from Precision, and Recall, the direct decrease
in the accuracy signifies a considerable decrease in the percentages of Precision, and
Recall respectively, and a corresponding increase in the FDR value.
A powerful approach of a lethal combination of a standard Semantic Similarity
measure and a Diversity Index, for generating the Linked Open Data as topic hier-
archies, and leveraging the web pages to the topic that happens based on the index
words or the content words in the web pages. Apart from this, there is a large number
of real-world entities injected into the approach from several heterogeneous sources
namely the DBPedia, YAGO, NELL, Wikidata, and the DebatePedia. The Knowl-
edge Base resources form a major part of the TBOX of the Ontologies. The array of
Web Directories populates the individuals to the conceptual schema, thereby trans-
figuring this approach into a perfect knowledge-driven approach for generating topic
hierarchies and leveraging web pages based on the topic. Moreover, the usage of the
Diversity Index discovers the terminologies that are relevant and novel. The usage of
Jaccard Similarity ensures that there is no divergence with respect to the relevance
of the web pages. The F-Measure percentage Vs the Number of Recommendations
is depicted in Fig. 2.

Fig. 2 F-measure vs no. of recommendations

MKSTD: Multiple Knowledge Sourcing Approach for Topic Directories … 573

The proposed MKSTD has the lowest FDR of 0.06 and the highest nDCG score of
0.91 when compared with the baseline approaches and other variants when certain
important methodologies were eliminated. The MKSTD has an FDR of 0.06 less
than MPC-LSW and 0.1 nDCG more than the MPC-LSW. The MKSTD has FDR of
0.11 less than that of the MA Algorithm and has an nDCG of 0.13 higher than the
MA Algorithm. Since all the other variants of the strategies in MKSTD has one of
the important techniques eliminated from its hybridization, all the other variations
have lower values of Precision, Recall, Accuracy and nDCG when baselined with
the MKSTD. The FDR of MKSTD is much lower than the other variants. The reason
for the high values of the nDCG in the MKSTD is mainly due to the incorporation of
a large range of heterogenous knowledge sources ranging from standard knowledge
bases that hold facts to the Web directories, and the infusion of the Simpson’s Diver-
sity Index. The FDR is quite low mainly due to the regulation of using a combined
semantic similarity measure, the Jaccard Similarity with the Simpson’s Diversity
Index. Owing to the same reason, the F-Measure Vs Number of Recommendations
for the MKSTD is higher than the baseline approaches and other variants of the
MKSTD when one of the important hybridizations has been eliminated.

6 Conclusions

The MKSTD framework for mapping and integrating knowledge from several hetero-
geneous sources for topic hierarchy generation and linking of web pages has been
proposed. The approach integrates knowledge from a series of knowledge bases,
crowd intelligence, and a number of traditional Web directories. The Linked Open
Data is generated by linking the topic hierarchies using a combination of the Jaccard
Similarity and Simpson’s Diversity Index. Also, the anchorage of the web pages with
the topic hierarchies in the Linked Open Data is achieved by the combination of the
Jaccard Similarity with the Simpson’s Diversity Index. The MKSTD is one of the
first systems to have integrated the standard Knowledge Bases with the traditional
web directories to harvest knowledge and incorporate the same in an environment of
Semantic Search. The MKSTD has achieved an average accuracy of 95.97% with an
nDCG of 0.91 and a very low FDR of 0.06 for the Gutenberg Project categorizations
which proves the efficacy of relevant knowledge integration into an environment of
semantic search.

References

1. Sun C, Lv L, Tian G, Wang Q, Zhang X, Guo L (2020) Leverage label and word embedding
for semantic sparse web service discovery. Math Probl Eng
2. Qi XF, Hua YL, Wang P, Wang ZY (2019) Leveraging keyword-guided exploration to build
test models for web applications. Inf Softw Technol 111:110–119
574 G. Deepak and A. Santhanavijayan

3. Delpeuch A (2019) Opentapioca: lightweight entity linking for wikidata. arXiv preprint arXiv:
1904.09131
4. Montenegro C, Navarrete R (January 2020) Improving core topics discovery in semantic
markup literature: a combined approach. In: Proceedings of the 3rd International Conference
on Software Engineering and Information Management, pp 237–243
5. Labhishetty S, Pei K, Boughoula A, Zhai C (2019) Web of slides: automatic linking of
lecture slides to facilitate navigation. In: Proceedings of the Sixth (2019) ACM Conference on
Learning@ Scale, pp 1–4
6. Martín-Chozas P, Montiel-Ponsoda E, Rodríguez-Doncel V (2019) Language resources as
linked data for the legal domain. Knowl Law Big Data Age 317:170
7. El Vaigh CB, Goasdoué F, Gravier G, Sébillot P (September 2019) Using knowledge base
semantics in context-aware entity linking. In: Proceedings of the ACM Symposium on
Document Engineering 2019, pp 1–10
8. Shang J, Zhang X, Liu L, Li S, Han J (April 2020) Nettaxo: automated topic taxonomy
construction from text-rich network. In: Proceedings of the Web Conference 2020, pp
1908–1919
9. Venugopal KR, Srikantaiah KC (2020) Construction of topic directories using Levenshtein
similarity weight. In: Web Recommendations Systems. Springer, Singapore, pp 67–86
10. Giri LG, Mathur A, Manjula SH, Venugopal KR, Patnaik LM (2014) MA Algorithm to generate
semantic web related clustered hierarchy for keyword search. Int J Recent Trends Eng Technol
11(1):222
11. Deepak G, Santhanavijayan A (2020) OntoBestFit: a best-fit occurrence estimation strategy
for RDF driven faceted semantic search. Comput Commun 160:284–298
12. Giri GL, Deepak G, Manjula SH, Venugopal KR (2018) OntoYield: a semantic approach
for context-based ontology recommendation based on structure preservation. In: Proceedings
of International Conference on Computational Intelligence and Data Engineering. Springer,
Singapore, pp 265–275
13. Kumar A, Deepak G, Santhanavijayan A (July 2020) HeTOnto: a novel approach for
conceptualization, modeling, visualization, and formalization of domain centric ontologies
for heat transfer. In: 2020 IEEE International Conference on Electronics, Computing and
Communication Technologies (CONECCT). IEEE, pp 1–6
14. Deepak G, Priyadarshini S (2016) A hybrid framework for social tag recommendation using
context driven social information. Int J Soc Comput Cyber-Phys Syst 1(4):312–325
15. Deepak G, Teja V, Santhanavijayan A (2020) A novel firefly driven scheme for resume parsing
and matching based on entity linking paradigm. J Discrete Math Sci Cryptogr 23(1):157–165
16. Haribabu S, Sai Kumar PS, Padhy S, Deepak G, Santhanavijayan A, Kumar DN (2019) A
novel approach for ontology focused inter-domain personalized search based on semantic set
expansion. In: 2019 Fifteenth International Conference on Information Processing (ICINPRO),
Bengaluru, India, pp 1–5. https://doi.org/10.1109/icinpro47689.2019.9092155
17. Jaccard P (1901) Étude comparative de la distribution florale dans une portion des Alpes et des
Jura. Bulletin de la Société vaudoise des sciences naturelles 37:547–579
18. Jaccard P (1912) The distribution of the flora in the alpine zone. New Phytol. 11(2):37–50.
https://doi.org/10.1111/j.1469-8137.1912.tb05611.x
19. Simpson EH (1949) Measurement of diversity. Nature 163(4148):688
A Fuzzy Near Neighbors Approach
for Arabic Text Categorization Based
on Web Mining Technique

Mounir Gouiouez

Abstract Nowadays, the quantity of textual content available online has experienced
such a colossal increase. Hence, the need for a system to investigate this content data
is mandatory. In this concern, Text Categorization (TC) highlights many performance
methods and techniques to analyze, explore and classify various types of documents.
This study consists of two main steps. First, we extract terms from text documents
using Fuzzy Near Neighbors (FNN) with web-based mining techniques algorithm.
Second, we identify documents according to a particular form of similarity based on
combining all Arabic encyclopedic dictionaries using clustering algorithms. In this
article, Fuzzy C-Means (FCM) as a clustering algorithm is used to perform the preci-
sion of documents’ classification. This work suggests Arabic TC based on a multi-
lingual encyclopedic dictionary (Arabic WordNet, OMW, Wikipedia, OmegaWiki,
Wictionary, and Wikidata). To evaluate the efficacy of TC approach with FNN and
FCM, an experimental study using a real-world dataset is carried out. The results
of the present study indicate that proposed approach outperforms the traditional one
and produces good results.

Keywords Text categorization · Fuzzy near neighbors algorithm · Web-based

mining techniques · Fuzzy C-means clustering algorithm · Machine learning

1 Introduction

The advent of technological revolution and the vast accumulation of information in

digital media promote the exchange of documents between different societies and
nations, and establish new conditions for users to access this wealth of data. In this
regard, text categorization (TC) promotes an important role for users to classify the
most relevant information. However, the vast majority of the current TC approaches
are structured to cover the effectiveness of documents published in English and not for

M. Gouiouez (B)
Psychological, Social and Cultural Studies Laboratory, Sidi Mohamed Ben Abdellah University
(USMBA), Fez, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 575
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_52
576 M. Gouiouez

texts written in the Arabic language. Developing TC systems for Arabic documents
is a difficult activity due to the complicated and rich nature of the Arabic language.
Text Classification as described by Baraa et al. [1] in 2014 is a series of supervised
learning tasks undertaken to structure a collection of documents using a variety of
different methods, including distance-based, tree-based decision-making, Bayesian
classification, SVM … etc. TC techniques are used in many disciplines, including
mail filtering, irrelevant filtering messages, document archives, and automatic data
indexing of scientific articles. In theory, the standard text categorization process
consists of four stages, each of which includes many steps.
These stages provide as first the pre-processing phase of the document. This stage
involves several actions to prepare the content of documents that are seen by the
classification model. It tokenizes the text, removes stop-word, and weights terms.
The second stage of TC is the extraction of terms, follow by the last step of mapping
and disambiguation.
This study proposes a generic approach for extracting words using Fuzzy
near neighbors with web-based mining techniques. This approach identify docu-
ments according to a particular form of similarity based on combining all Arabic
encyclopedic dictionaries using clustering algorithms.
Fuzzy C-Means, as a clustering algorithm, is used to perform the approach of this
research. This study adapts an experimental design using real-world dataset.
The objective of this article is twofold. First, it seeks to investigate the ability
of constructing a fuzzy logic theoretical model for Arabic text categorization that
provides appropriate results in relation with other different classifiers. Second, it
aims to suggest an enhanced version of the easy reduce way to create various high
reduction based on all Arabic encyclopedic dictionaries.
This paper is organized in this manner. Section 2 presents a thorough definition
of the key concepts used in this article. Section 3 gives an overview of Arabic Text
Categorization. The design of the approach used in this study is provided in Sect. 4
with its different steps. Section 5 presents the experimental results. Finally, Sect. 6
provides further directions to the TC problem using the Fuzzy logic.

2 Preliminaries

To review the literature of the current study, it is of paramount importance to

define some key concepts that are the core matter of Arabic text classification
approaches. These concepts are the following: 1) web-based mining (WUM), 2)
Machine Learning (ML), and 3) algorithm of a fuzzy c-mean (FCM).
A Fuzzy Near Neighbors Approach for Arabic Text Categorization ... 577

2.1 A Web Data Mining

According to [2], the web data mining process consists of three steps: pre-processing
data, discovering patterns, and analyzing patterns. The first step, pre-processing, is
concerned with converting web data into an easily evaluated data. The second stage
of web data mining deals with the extraction of special knowledge and patterns
from labeled data using mining and machine learning algorithms [3]. The last step,
patterns’ analysis, copes with collecting similar patterns from the appropriate ones.
Web data Mining is the process of data mining methods for useful information
from the web and its trends of use. Web mining helps to increase the control of the web
search by categorizing web documents and recognizing web content. It is used for
search engine, to predict behavior of users, and for a specific website encyclopedic
dictionary. Web Data Mining and knowledge exploration are continually growing
for researchers [4]. Web data mining generally contains two categories: web services
Mining and Web Structure Mining. Web services mining is the request of extracting
beneficial data from the content of web documents, while web structure mining is
an application for the discovery of organizational features from web content [5].

2.2 Fuzzy C-means Clustering Algorithm

FCM [6] is an unsupervised learning algorithm introduced by Dunn in 1973 [7] and
modified by Bezdek in 1987 [8]. It is a clustering method that split one corpus of data
to their correspond clusters (two or more). This methodology is utilized for selecting
adequate information. This FCM is founded on optimizing the following function:
2

N 
C
Dinj |xi − cj| (1)
i=1 j=1 1≤m≤∞

Where, n > 1, Dij dependence level to xi in cluster j, xi reflects observed d-

dimensional data, cj presents the center of cluster, and || || expresses normal similarity
between these observed information and the center Fuzzy segmentation. In theory
uij expressed by iteration function:


N
Dinj xi
1 i=1
Di j = c   2 cj = (2)
 |xi −c j | n−1 N
xi −ck Dinj
k=1 i=1
578 M. Gouiouez
 
If maxi j |Di(k+1)
j − Di(k)
j | ≤ ε, where ε is a condition criterion between 0 and 1,
the iteration loop stopped. K number of iteration steps. The algorithm is composed
of the following steps:

Step 1: Initialize D = Di j matri x, D(0) 
Step 2: estimate the center vectors C(k) = c j with D(k) by:


N
Dimj xi
i=1
cj =
N
Dimj
i=1

Step 3: Update D(k), D(k + 1)

1
Di j = c   2
 |xi −c j | m−1
xi −ck
k=1

If ||D(k + 1) − D(k)|| < ε then STOP; then coming back to step 2.

3 Related Work for Arabic Text Categorization

During the last years, numerous studies have been conducted on the importance of
Arabic TC. These studies need are briefly discussed in order to provide empirical
evidence for the ongoing research. For instance, in 2006, many research studies used
the K-nearest neighbor technique to categorize data in Arabic TC [9]. Duwairi et al.,
in [10] listed the main classifiers in their article that have shown great effectiveness in
Arabic text processing- Naïve Bayes (NB), K-nearest neighbor (K-NN) and distance-
based algorithms. Then, Mesleh [11] suggests a method that use support vector
machine (SVM).
In 2011, several studies on Arabic text categorization were conducted using
classical methods used before but for different data resource and with different
pre-processing methods [12].
In 2012, the authors of [13] introduced a neural network-based approach with the
learning vector quantization (LVQ) algorithm and the Kohonen self-organizing map
(SOM) to discovery similarities in text for categorization. In 2015, a comparative
study of several approaches of Arabic TC presented in [14]. Similarly, Al-Radaideh
and Al-Khateeb used a hybrid classifier to classify Arabic documents. Table 1 shows
the most popular contributions in Arabic Text Categorization.
However, almost all works have taken into account the specific resource based
on a particular model and studies the relationship between term to documents or
documents to categories. In fact, in our model we present a design that calculates
A Fuzzy Near Neighbors Approach for Arabic Text Categorization ... 579

Table 1 Overview on best article of Arabic text classification

Year Authors with most cited Articles cited more than 80 Contribution
indexation times
1998 Al-Shalabi et al. Cited 160 times Uses different algorithms
to find the roots and
pattern
2002 Larkey et al. Cited 489 times Developed several light
stemmers based on
heuristics method
2005 Duwairi Cited 489 times Machine learning for
Arabic text
2006 Khreisat et al. Cited 180 times Uses N-gram frequency
2007 Mesleh et al. Cited 250 times -Uses Chi square feature
extraction and SVM
2015 El-Halees et al. Cited 160 times Arabic text classification
using maximum entropy
2017 Al-Anzi Cited 86 times Uses cosine similarity and
latent semantic indexing

the semantic similarity between two terms based on multiple resources before the
extracting phase and organizing the datasets in cluster prior to classification.

4 Our Approach

One of the characteristics of human reasoning is that it is based on incorrect or

incomplete data, while the computer is based on accurate data. Therefore, there
is a need to develop a new method that takes into account uncertainty in decision
making. On this concern, it seems that fuzzy logic is the best solution that meets
this need. Contrary to classical logic, fuzzy logic makes it possible to manipulate
other values of truth than the true and false absolutes. To highlight this concept, a
commonly used example is that of the size of an individual (see Fig. 1) for which
three classes are defined, “small” “medium” and “large” in the intervals “small =
[40, 160]”, “medium = [160, 180]” and “large = [180, 250]”. Conventional logic is
that a person × 159.9 cm is qualified as small and not medium size, although there
is a small difference.

4.1 Methodology

The current study is based on the idea that semantically similar terms are used in
similar contexts in documents. We also assume that the set of words that co-occur in
580 M. Gouiouez

Fig. 1 Illustration of our design with fuzzy near neighbors and clustering algorithm

a snippet with a target term to constitute a common meaning for the occurrence of
the target word with K Fuzzy near neighbors.
The approach suggested in this research is composed of two stages as indicated in
(Fig. 1). The first stage relates to the learning phase. It consists of generating a new
text representation based on mapping terms with their associated concept using K
Fuzzy near neighbors. The second stage relates to the classification phase. It consists
in weighting categories, and calculating the profile of the document to be classified
using Clustering algorithm for the creation of a multi classifier system.
With this approach, we aim for two things: the increase of similarity of extract
terms via the web mining and the reduction of the size of documents through the
Clustering algorithm.
A Fuzzy Near Neighbors Approach for Arabic Text Categorization ... 581

The importance of web-based mining in this work gives us the opportunity to

exploit all resources (Arabic WordNet, OMW, Wikipedia, OmegaWiki, Wictionary,
and. Wikidata) as a single database.

K Fuzzy Near Neighbors. This type of classifier takes advantage of fuzzy logic in
distance calculation used for the selection of the K nearest neighboring. In [15], the
authors define this fuzzy measure as follows:

  Fe −1.0
d(x, R)
u(x) = 1 + (3)
Fd

With: d(x, R) the distance between the unknown term x and the sample R of resource.
Fd and Fe are two positive constants used to control Fuzzy term in the previous
calculation.

Clustering Algorithm for the Creation of a Multi-classifier System. The creation

of the multi-classification system is based on multi-classifier clustering. This method
makes it possible to reduce the complexity of recognition by the distribution of the
problem on a set of specialist classifiers in different areas of the problem. This
technique works according to the premise of dividing documents to subsets. For
this purpose, a clustering algorithm (Fig. 2) is used to divide the pseudo-document
database in several subsets or clusters of different classes whose form is near. Each
of these subsets is then used to the learning and testing of one of the classifiers.

Fig. 2 Scalable clustering method

582 M. Gouiouez

Fig. 3 Dictionary size for Number of Documents

CCA datasets, the 2,00,000 Number of Words
BBC-Arab datasets, and the
EASC datasets 1,00,000

DicƟonnary
0

5 Evaluation and Discussion

5.1 Corpora Summary

Different companies’ data has been selected to conduct the present study. The data
selected include small and large corpus with few and many categories. The companies
from which the data is collected include the CCA datasets, the BBC-Arab datasets,
and the EASC datasets (Fig. 3).

5.2 Experimental Procedure and Performance Measure

The assessment of the performance of the classification algorithm for classifying data
into the appropriate category shall be carried out using a number of mathematical
rules, such as recall, precision and F-measure, which are described as follows:
In this section, we introduce and evaluate experimental data using the suggested
method (Text Categorization algorithms with SVM and NB techniques). In so doing,
each corpus is divided into two segments (80% for training and 20% for testing). For
testing, a computer with 8 GB RA processor is used.
The assessment of the performance of the classification algorithm for classifying
data into the appropriate category is carried out using a number of mathematical
rules, such as recall, precision and F-measure, which are described as follows.

TP TP
Precision = And Recall = (4)
TP + FP TP + FN

TP is the number of data that is correctly classified to the category. TN is the

number of observations that are incorrectly allocated. FP is the number of data,
which the system wrongly assigned. FN is the number of documents that belonged
to the category but are not assigned to the category. F1-measure is calculated as
follows:
2.Precison.Recall
F1 - measure = (5)
precision + Recall
A Fuzzy Near Neighbors Approach for Arabic Text Categorization ... 583

Table 2 Comparison
DataSet EASC CCA BBC-Arabic
between SVM and NB Using
FNN and FC Methods SVM 92 94,5 89,4
NB 84,2 83,25 87,2
SVM + Fuzzy logic clustering 92,07 95,52 89,89
NB + Fuzzy logic clustering 83,37 85,01 85,34

5.3 Dimensionality Reduction

In this paper, all the methods discussed organize the features according to their
importance to the category. The top ranking features from each category are then
chosen and represented to the classification algorithm. Consequently, we obtain a
new form of document with a new dimension.
The results showed in Table 2 disclose that Fuzzy logic clustering almost always
performs text classification than any of the other learning algorithms in different
corpora. In regard to the results of all classifiers, Table 2 has revealed that SVM
combined by Fuzzy logic clustering method for feature reduction are typically better
than NB.

6 Conclusion

This study has approached the problem of text representation for Arabic text docu-
ments. It has introduced a novel approach using Fuzzy Near Neighbors (FNN) with
web-based mining techniques algorithm. Similarly, this paper has identified docu-
ments according to a particular from of similarity based on the combination of all
Arabic encyclopedic dictionaries using clustering algorithms. The results of this
study have shown that the suggested method increase the classification of Arabic
documents better than the classical techniques.

References

1. Baraa S, Nazlia O, Zeyad S (2014) An automated Arabic text categorization based on the
frequency ratio accumulation. Int Arab J Inform Technol 11(2):213–221
2. Kumar GD, Gosul M (2011) Web mining research and future directions. In: Wyld DC,
Wozniak M, Chaki N, Meghanathan N, Nagamalai D (eds.) Advances in Network Security
and Applications. CNSA
3. Tharwat A (2019) Parameter investigation of support vector machine classifier with kernel
functions. Knowl Inf Syst 61:1269–1302
4. Ali N, Neagu D, Trundle P (2019) Evaluation of k-nearest neighbour classifier performance
for heterogeneous data sets. SN Appl Sci 1:1559
5. Chen J, Dai Z, Duan J et al (2019) Improved Naive Bayes with optimal correlation factor for
text classification. SN Appl Sci 1:1129
584 M. Gouiouez

6. Gavade AB, Rajpurohit VS (2020) Sparse-FCM and deep learning for effective classification
of land area in multi-spectral satellite images. Evol Intel
7. Schroder HW, Welling H, Wellegehausen B (1973) Appl Phys 1:343–348
8. Bezdek JC (1987) Analysis of Fuzzy Information, vol 1, 3. CRC Press, Boca Raton
9. Al-Radaideh QA, Al-Abrat MA (2019) An Arabic text categorization approach using term
weighting and multiple reducts. Soft Comput 23:5849–5863
10. Chantar H, Mafarja M, Alsawalqah H, et al (2019) Feature selection using binary grey wolf
optimizer with elite-based crossover for Arabic text classification. Neural Comput Appl
11. Mesleh A, Ghassan G (2008) Support vector machine text classification system: using ant
colony optimization based feature subset selection. Comput Eng Syst
12. Razavi AR, Gill H, Åhlfeldt H, Shahsavar N (2005) A Data Pre-Processing Method To Increase
Efficiency And Accuracy In Data Mining. In: Miksch S, Hunter J, Keravnou ET (eds.) Artificial
Intelligence in Medicine. AIME 2005. Lecture Notes in Computer Science, vol 3581. Springer,
Heidelberg
13. Biehl M (2012) Admire LVQ—adaptive distance measures in relevance learning vector
quantization. Künstl Intell 26:391–395
14. Al-Radaideh Q, Al-Khateeb S (2015) An associative rule-based classifier for Arabic medical
text. Int J Knowl Eng Data Min 3(3–4):255–273
15. Kramer O (2013) K-nearest neighbors. In: Dimensionality Reduction with Unsupervised
Nearest Neighbors. In: Intelligent Systems Reference Library, vol 51. Springer, Heidelberg
Recommendation System Based
on Semantic Web Approach

Sara Belattar, Otman Abdoun, and Haimoudi El Khatir

Abstract Recently, most countries suffer a decrease the crop productivity due to
weather conditions, so many kinds of research are suggested for growth and enhance
crop productivity with different methodologies and techniques. The Semantic Web
(SW) approach is proposed in this work to build a recommendation system for helping
the farmers to take an adequate decision of recommended fertilizer in fast time
by specific parameters (e.g., temperature, humidity, moisture, soil type, crop type,
nitrogen, potassium, and phosphorus). This proposed approach is based on some
essential technologies such as Ontology, Web Ontology Language (OWL), Resource
Description Framework (RDF), and RDF Query Language (SPARQL). In this work,
our goal is to present another technique to improve the quality of service for farmers
with more intelligence in order to increase their crop productivity. The results showed
that the capability of the recommendation system based on SW was effective.

Keywords Semantic Web (SW) · Ontology · Resource Description Framework

(RDF) · RDF Query Language (SPARQL) · Recommendation system ·
Recommended fertilizer

1 Introduction

The Semantic Web (Web 3.0) is the World Wide Web extension, standardized by
the World Wide Web Consortium (WC3) (Berners-Lee et al. 2001) [1]. This term
represents a field of research as other researches domains (e.g., artificial intelligence,
machine learning, etc.). The purpose of SW is to improve the current World Wide
Web by machine-understandable semantic documents and data on the web with the
fast response of user’s questions thanks to its technologies used. So, the Semantic web
has come to solve the most problems facing the current Web, due to the increasing
the data and information day by day that why it will be more difficult to facilitate its
management in the WWW and can’t interpret its semantic. There are some essential

S. Belattar (B) · O. Abdoun · H. E. Khatir

Computer Science Department, Advanced Science and Technologies Laboratory, Polydisciplinary
Faculty, University of Abdelmalek Essaadi (UAE), Larache, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 585
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_53
586 S. Belattar et al.

technologies used for creating the SW such as Resource Description Framework

(RDF) is a language used to structure the data for making it readable by a machine
(Manola et al. 2004) [2], SPARQL Query language for obtaining the information
from RDF (Hommeaux and Seaborne 2008) [3] and ontologies are a knowledge
base for facilitating the representation of data (Gruber 1993) [4]. The Semantic Web
approach was found in many fields of research including medicine, a system capable
of taking the appropriate decision was proposed for Tuberculosis patients based on
the Semantic Web approach to improving the quality of health services (Lima et al.
2019 & Pellison et al. 2019) [5, 6], a model of a health observatory conceptual was
developed using the Semantic Web to implement the health observatory processes
(Yoshiura et al. 2018) [7]. In E-Learning, the state-of-the-art Massive Open Online
Courses (MOOC) platforms using the Semantic Web field was described to facilitate
personalization in modern MOOC (Kiselev and Yakutenko 2020) [8], the E-Learning
Computer Science course system was created by Semantic Web for providing effi-
cient personalized learning (Gomathi and Suguna, 2020) [9]. In industry, the ontology
was created to represent the industrial megaproject that facilitates the processing of
data using Linked Data and Semantic Web (Zangeneh and McCabe 2020) [10]. After
this literature review cited, we can see the power of the Semantic Web approach in
various fields especially in unstructured data issue for that in this paper, we suggested
developing a recommendation system based on SW in the agriculture field since this
technique presented successful works in this domain such as; (Marwaha 2016) [11]
building an ontology-based an expert system by using Semantic Web approach for
taking the suitable decision-making by farmers in a fast time, the agriOpenLink
platform based on Semantic Web was implemented for precision agriculture (Tomic
et al. 2015) [12], additionally, a recommendation system is a useful tool for filtering
the information in order to give the best information to users, and it was also applied
in the agriculture field with different methods [13–15]. The main objective of our
study is to create a recommendation system for recommending a suitable fertilizer
related to specific parameters (e.g., temperature, humidity, moisture, soil type, crop
type, nitrogen, potassium, and phosphorus) using the Semantic Web approach. Due
to the increase in the data and the documents in the agriculture field, the SW repre-
sents a suitable tool in this domain (Drury et al. 2019) [16], and it can updates itself
dynamically from the ontologies used in a specific domain. Moreover, it can solve
the issues of traditional mechanisms used in the agriculture field for helping the
farmers to improve their crop productivity, reduce the production risks, and answer
their questions in fast time. In other ways, the SW has a high performance to work
with non-quantitative data i.e., description the information on categories and groups
rather than numbers compared to the Artificial Neural Network (ANN) accept only
the quantitative data i.e., numerical data as inputs. In order to develop our proposed
approach, we structured the data of fertilizer in the form of ontology for facilitating
its processing and representation. These data used in this work, exist on the web.
As a result, it was demonstrated that the Semantic Web approach gave the desired
results.
The rest of this paper is organized as follows: in the next section, some related
works are presented for describing the different tools and methodologies used to
Recommendation System Based on Semantic Web Approach 587

implement the recommendation systems in the agriculture field. In the third section,
we describe our methodologies proposed for the Semantic Web and its technologies.
In the fourth section, the results achieved by applying our approach. Finally, we give
a conclusion and perspectives of our next work.

2 Related Work

During the literature review, many works developed the recommendation systems
in agriculture with different methods and techniques such as; a system of recom-
mendation was implemented to recommend the appropriate fertilizer related to
specific soil fertility, climate, and crop by using the Nutrient Expert (NE) model
is a computer-based decision support system, they calibrated Nutrient Expert for
maize agro-ecological areas in Nigeria, Ethiopia, and Tanzania by generating the
data from 735 trials on-farm between 2015 and 2018. In this study, their objective
was to reduce the crop productivity risk and efficiency of using fertilizer to help the
farmers to increase their crop productivity. The results showed that their proposed NE
model was effective from cost-effective for recommending the appropriate fertilizer
to farms by offering another solution for soil testing (Rurinda et al. 2020) [13]. The
agricultural recommendation system was proposed to overcome the pests affecting
a crop by identifying a pest and giving a suitable treatment. For developing this
system, they based on “Pest in Crop and their Treatment” Ontology (PCT-O) in
order to facilitate the usability of their data for making an intelligent system (Lacasta
et al. 2018) [14]. A crop recommendation system was developed to improve the crop
productivity based on ensembling techniques (e.g., Random Forest, Naive Bayes,
and Linear SVM) of machine learning for recommending the right crop related to
soil specific type. Their experimental results showed that the average accuracy of
classification using ensembling techniques was 99.91% (Kulkarni et al. 2018) [15].

3 Materials and Methods

In this paper, we propose a Semantic Web approach to an agricultural recommenda-

tion system for answering the farmer’s questions about the suitable fertilizer in order
to enhance their crop productivity. In recent years, precision agriculture is based
upon the data, and with the growth of these data, precision agriculture is becoming
difficult to processing (Naiqian et al. 2002) [17]. The Semantic Web approach has
an important role to respond to these issues by providing a good representation and
structuration of the data rather than the current Web. In the following sections, we
describe the SW approach and the technologies that are necessary to build it.
588 S. Belattar et al.

3.1 Semantic Web

The Semantic web (SW) represents an evolution of the World Wide Web. This term of
SW involves specific techniques recommended by the World Wide Web Consortium
(W3C) that improve the current Web and its uses. These techniques provide a meaning
of the data without worrying about its representation (Robu et al. 2006) [18] and it
can understand by both humans and machines. Additionally, the objective of SW is to
create a global database of connected data on the Web (Laufer 2015) [19]. Thus, the
Semantic Web technologies are Ontology, Resource Description Framework (RDF),
Web Ontology Language (OWL), and SPARQL Query language.

Ontology
Firstly, the concept of ontology is originating from philosophy and it represents a
knowledge base (in the symbolic artificial intelligence sense) (Hitzler 2020) [20]. The
main role of ontology is to define the knowledge representation in a specific domain
(Craig 1998) [21], in order to share, exchange, discover the idea or data and improve
the communication between different agents. The ontology contained different
elements (e.g., individuals, classes, attributes, semantic relations, and events) for
improving the Web search. A simple example of crop ontology representation is
shown in (Fig. 1).

Web Ontology Language (OWL)

Web Ontology Language (OWL) allows users to represent and write their ontologies
in a specific domain, it is built on the top of RDF (McGuinness and Van Harmelen
2004) [22]. For creating these ontologies, we can use a free open-source Protégé
ontology editor, this platform is popular in the Semantic Web field and it was devel-
oped in JAVA. With this editor, the user could create and manipulate the ontologies
in various representation formats.

Fig. 1 Crop ontology

Recommendation System Based on Semantic Web Approach 589

Fig. 2 The RDF triple

Resource Description Framework (RDF)

Resource Description Framework (RDF) is a model to describe the resource on the
web and its metadata or semantics in order to interchange the data on the web. The
RDF consists of three concepts (Gandon et al. 2011) [23] such as;
Resource: this term refers to a Web page, image, person, video, fragment page,
which means anything that has a URI can be considered a resource.
Description: for describing the resource in order to be more understandable, the
description has attributes, features, and specific relation between resources.
Framework: for providing the language syntaxes of these descriptions.
The information on the web is structured in the form of the RDF triples (Cyganiak
and Wood 2014) [24] (see Fig. 2) that computers can process this information easily.
The RDF triples are used for graph description, and it represented by XML language
(or syntax).

SPARQL Protocol and RDF Query Language (SPARQL)

RDF Query Language (SPARQL) is a protocol used for querying and updating the
RDF documents (Group WSW 2013) [25]. The purpose of using the SPARQL Query
language is that the RDF is based on the XML language, and for making the commu-
nication and exchanging the information, we need the SPARQL query. The query
language must understand the syntax of RDF and also must understand the data
model and semantic vocabulary of RDF (Che 2006) [26].

3.2 Design Flowchart of Our Proposed Work

In order to develop our recommendation system based on the Semantic Web

approach, we used all Semantic Web technologies as described above. Thus, all steps
followed for developing our system to recommend a suitable fertilizer for farmers
are depicted in the flowchart below (see Fig. 3).

4 Results and Discussion

In order to get the results, we based on some technologies tools for building our
recommendation system.
In the development of the graphical interface of our recommendation system
based on Semantic Web (Frontend development), HTML5, JavaScript, CSS, and
technique to dynamically create HTML code (JSP) “Java Server Pages” were used
590 S. Belattar et al.

Fig. 3 Recommendation system design flowchart based on the semantic web approach

to implement the web page. Additionally, the Bootstrap framework was used for
designing the graphical interface.
In the development of our recommendation system based on Semantic Web
(backend development), Eclipse IDE for Enterprise Java Developers and Apache Jena
API framework (for creating the Semantic Web Application) were used to implement
all classes and methods to work with fertilizer ontology created and semantic data
contained in the RDF. For creating our fertilizer ontology we based on data that
already exist on the web; these data consist of the specific parameters for recom-
mending the suitable fertilizer. For adding all our classes we used the OWLClasses
tab (see Fig. 4).
For fertilizer type (e.g., DAP and Urea) we add it by clicking on the individuals tab,
and each type of fertilizer related to specific parameters (e.g., temperature, humidity,
moisture, soil type, crop type, nitrogen, potassium, and phosphorus), for adding
these parameters we used the properties tab by clicking on add data type property.
Besides this, each crop can need 2 types of fertilizer according to the parameters we
have as depicted in the figure above (see Fig. 4), the sugarcane crop needs the DAP
and Urea fertilizer for increasing its productivity but with different parameters. By
applying all previous steps for developing our proposed approach, the interface of
our recommendation system based on the SW is ready to use by farmers (see Fig. 5).
In order to get the recommended fertilizer, farmers must insert all parameters of
temperature, humidity, moisture, nitrogen, potassium, phosphorus, crop type, and
soil type, after that the farmers click on the submit SPARQL Query button (see
Fig. 5).
Our ontology is written in RDF/XML, for that we need the SPARQL Query
language for getting the results. When we click on Submit SPARQL Query, the
query used all parameters inserted by the farmer (see Fig. 6). The SPARQL Query
Recommendation System Based on Semantic Web Approach 591

Fig. 4 Our fertilizer ontology created using Protégé 3.3.1

Answering of recommended fertilizer based on these parameters, i.e., temperature

= 29, humidity = 52, moisture = 45, nitrogen = 12, potassium = 0, phosphorus =
36, crop type = Sugarcane, and soil type = Loamy is DAP.
Where, PREFIX ns represent the namespace of ontology, PREFIX rdf represent
the namespace of RDF, rdf: type is the property, and ns: Recommended_fertilizer is
the object.
This paper proposed a recommendation system based on the SW approach that
provides facilities to present a suitable fertilizer using the specific parameters for
the farmers by answering a machine-understandable form and in the fast time range.
Additionally, in the research field, the data represent an important tool for devel-
oping intelligent systems and taking decision-making in various domains. However,
sometimes these systems face problems of using some types of data, due to its repre-
sentation. In the majority, there are two types of data e.g., quantitative data and
non-quantitative data (or qualitative data). The first one is based on the numbers
form and the second one is focusing on describing the information with categories
rather than numbers. In artificial intelligence especially, Artificial Neural Networks
(ANN) it very difficult to work with non-quantitative data as inputs for that reason,
we need to transform or map it into numerical value (quantitative data) to make well
understood from ANN (e.g., in order to recommend the suitable fertilizer we must
enter the crop type and the soil type, in ANN we must transform these parameters into
an integer for making the learning). Therefore, the Semantic web has been proposed
to solve these data representation issues through its capability to work with non-
quantitative data thanks to its technologies used, These capabilities of our proposed
592 S. Belattar et al.

Fig. 5 Recommendation system based on SW - question and suitable fertilizer answer of a farmer

work help the farmers to take suitable decisions of specific fertilizer type in order to
improve their crop productivity.
Recommendation System Based on Semantic Web Approach 593

Fig. 6 SPARQL query based on parameters giving by a farmer

5 Conclusion

This paper proposes another technique called the Semantic Web approach for devel-
oping the agricultural recommendation system that could improve crop productivity,
this system recommended the appropriate fertilizer for farmers by specific parame-
ters i.e., temperature, humidity, moisture, soil type, crop type, nitrogen, potassium,
and phosphorus. The Semantic Web represents a solution to work with massive data
and different types of data among them non-quantitative data compared to Artificial
Neural Network (ANN) works only with quantitative data as inputs. In this study, our
objective was to help the farmers by improving the Web search, enhancing efficiency,
and giving the answer in fast time. Additionally, this system facilitates the exchanging
the information of both human and machine thanks to Semantic Web Technologies
and it can updates itself dynamically from the ontologies used that why it named the
intelligent web. The results showed that the recommendation system based on the
SW approach was effective. Our perspective is to improve this system in future work
by creating an overall crop ontology that solves several issues in the agriculture field.
594 S. Belattar et al.

References

1. Berners-Lee T, Hendler J, Lassila O (2001) The semantic web. a new form of web content that
is meaningful to computers will unleash a revolution of new possibilities. Sci Am 284(5):34–43
2. Manola F, Miller E, McBride B (February 2004) RDF primer. W3C recommendation, World
Wide Web Consortium, pp 1–107
3. Hommeaux EP, Seaborne A (January 2008) SPARQL query language for RDF. W3C
recommendation, World Wide Web Consortium
4. Gruber TR (1993) A translation approach to portable ontology specifications. Knowl Acquis
5(2):199–220
5. Lima V, Pellison F, Bernardi F, Carvalho I, Rijo R, Alves D (2019) Proposal of an inte-
grated decision support system for Tuberculosis based on semantic web. Procedia Comput Sci
164:552–558
6. Pellison FC, Lima VC, Lopes Rijo RPC, Alves D (2019) Integrating Tuberculosis data in state
of São Paulo over semantic web: a proof of concept. Procedia Comput Sci 164:686–691
7. Yoshiura VT, Yamada DB, Pellison FC, De Lima IB, Damian IPM, Rijo RPCL, De Azevedo
Marques JM, Alves D (2019) Towards a health observatory conceptual model based on the
semantic web. Procedia Comput Sci 138:131–136
8. Kiselev B, Yakutenko V (2020) An overview of massive open online course platforms: person-
alization and semantic web technologies and standards. Procedia Comput Sci 169:373–379
9. Gomathi B, Suguna S (2020) Semantic web application in e-learning using protege based on
information retrieval. J Adv Res Dyn Control Syst 12(7 Special Issue):1847–1855
10. Zangeneh P, McCabe B (2020) Ontology-based knowledge representation for industrial
megaprojects analytics using linked data and the semantic web. Adv Eng Inform 46(June):
11. Marwaha S (20 January 2016) Ontology Based Expert System. Springer
12. Tomic D, Drenjanac D, Wöber W, Hörmann S, Auer W (2015) Enabling semantic web for preci-
sion agriculture: a showcase of the project agriOpenLink. In: CEUR Workshop Proceedings,
vol 1481, pp 26–29
13. Rurinda J, Zingore S, Jibrin JM, Balemi T, Masuki K, Andersson JA, Pampolino MF,
Mohammed I, Mutegi J, Kamara AY, Vanlauwe B, Craufurd PQ (2020) Science-based deci-
sion support for formulating crop fertilizer recommendations in sub-Saharan Africa. Agric Syst
180(July 2019):
14. Lacasta J, Lopez-Pellicer FJ, Espejo-García B, Nogueras-Iso J, Zarazaga-Soria FJ (2018) Agri-
cultural recommendation system for crop protection. Comput Electron Agric 152(June):82–89
15. Kulkarni NH, Srinivasan GN, Sagar BM, Cauvery NK (2018) Improving crop productivity
through a crop recommendation system using ensembling technique. In: 3rd International
Conference on Computational Systems and Information Technology for Sustainable Solutions,
pp 114–119
16. Drury B, Fernandes R, Moura MF, de Andrade Lopes A (2019) A survey of semantic web
technology for agriculture. Inf Process Agric 6(4):487–501
17. Zhang N, Wang M, Wang N (2002) Precision agriculture a worldwide overview. Comput
Electron Agric 36(2–3):113–132
18. Robu I, Robu V, Thirion B (2006) An introduction to the semantic web for health sciences
librarians. J Med Libr Assoc 94(2):198–205
19. Laufer C (2015) Guia_Web_Semantica, p 133
20. Hitzler P (2020) Semantic web: a review of the field. In: Proceedings of ACM, New York, NY,
USA, p 7
21. Craig E (1998) Routledge Encyclopedia of Philosophy. Routledge, New York
22. McGuinness DL, Van Harmelen F (2004) OWL web ontology language overview. W3C
Recomm 10:2004
23. Gandon F, Krummenacher R, Han S-K, Toma I (2011) The resource description framework
and its schema. Handbook of Semantic Web Technologies (Issue January)
24. Cyganiak R, Wood D, Lanthaler M (2014) RDF 1.1 concepts and abstract syntax. W3C Recomm
25
Recommendation System Based on Semantic Web Approach 595

25. Group WSW (2013) SPARQL 1.1 overview. W3C Recomm W3C
26. Che H (2006) A semantic web primer. J Am Soc Inf Sci Technol 27(8)
Sentiment Analysis of Moroccan Tweets
Using Text Mining

Moncef Garouani, Hanae Chrita, and Jamal Kharroubi

Abstract This paper introduces a framework system that analyzes the data extracted
from the social network Twitter. In our study, we deal with Moroccan tweets written
in Standard Arabic (SA) and Moroccan Dialect (MD). This system allows collecting,
processing and visualizing sentiments that could be extracted from Moroccan user’s
communication. As a training set, 36114 tweets were collected but only 13,550
tweets were found to be valid. Firstly, the tweets are classified into 2 categories: SA
and MD. The tweets belonging to each category are then classified into 4 classes:
positive, negative, neutral or mixed. We have performed several experiments to eval-
uate all the possible combinations of the following weight schemes: N-grams and
stemming terms techniques. The experimental results provide the best scenario for
each classifier and indicate that LSTM classifier using word embedding without
removing Stopwords outperforms CNN classifier as well as the CNN-LSTM classi-
fier (with the same features), and it outperforms SVM with term frequency–inverse
document frequency (TF-IDF) weighting scheme with stemming through Bigrams
feature. Moreover, the obtained results outperformed other comparable related work
results.

Keywords Opinion detection · Sentiment analysis · Polarity · TF-IDF · MD · SA

1 Introduction

With the advance of the web 2.0, many social media platforms emerged, allowing
people to express their feelings and give their opinions regarding a particular

M. Garouani (B) · H. Chrita · J. Kharroubi

Faculty of Sciences and Techniques, LSIA Laboratory, Fez, Morocco
e-mail: [email protected]
H. Chrita
e-mail: [email protected]
J. Kharroubi
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 597
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_54
598 M. Garouani et al.

subject such as commercial products, services (e.g. travel agencies, internet service
providers) or events (e.g. elections, new laws) [1]. Their interaction on these plat-
forms generates a huge amount of unstructured data. This latter is a valuable resource
that could be used to improve the decision-making process and acquiring knowledge
of people trends.
For this reason, it is useful to analyze the social network platforms (Facebook
or Twitter) content that have gained great popularity around the world [2]. Given
the number of users and their high level of use, these platforms reflect the mood of
the public as well as their trends regarding the current events. Most of the studies
performed in this field have been conducted on foreign languages especially the
English and the Japanese ones [3]. Nevertheless, researches analyzing the sentiment
of written Arabic language are very limited especially on Moroccan dialect.
This can be explained by three factors: Firstly, processing this language is too
complicated [4]. Secondly, there is a lack of web content regarding this language and
thirdly, no extra resources are available for this language [5]. The lack of resources
is considered as the main reason that impedes the development of natural language
processing tools for Arabic dialects in general and the Moroccan dialect (MD) in
particular. There exist no annotated (labeled) corpora for extracting the sentiments
from Moroccan posts. This is the main reason that motivated us to generate an opinion
corpus for the Moroccan dialect.
Indeed, our work aims to develop a system that analyzes the data extracted from the
social network Twitter. The proposed system allows collecting, process and classify
the polarity of Moroccan tweets according to the category they belong to (Standard
Arabic or Moroccan Dialect Arabic), and according to the class they belong to (posi-
tive, negative, neutral, mixed) and ultimately provide statistics for all the regions of
Morocco as well as their dominant sentiments.
In our work we conduct a study to discover the performance and the efficiency
of the latest techniques for automatic text classification: Support Vector Machine
(SVM), Long Short Term Memory (LSTM), Convolutional Neural Networks (CNN)
and the combination of them CNN-LSTM as well as evaluating their relevance
using different pre-processing strategies such as: stop-words removal, stemming, N-
grams and different weighting schemes (TF- IDF, word embedding) to automatically
analyze tweets written in Moroccan Arabic.
The remainder of the paper is structured as follows. Section 2 exposes the
promising fields that benefit from sentiment analysis as well as the works that
have already performed it. Section 3, describes our methodology in building the
tweets corpus. Section 4, presents the used methods and experimentations. Section 5,
discusses the obtained results and finally, main conclusions are drawn and potential
horizons are highlighted.
Sentiment Analysis of Moroccan Tweets Using Text Mining 599

2 Background and Literature Review

2.1 Application Areas of Sentiment Analysis

– Marketing: Actually, it is agreed that social media platforms are a wide environ-
ment for advertising products. To do so, companies seek to target our emotions
not our rationality. Therefore, using sentiment analysis besides the techniques
they are already using would be of great benefit. They will be able to personalize
their offers so that they suit well their customers and as a result, the development
of marketing will be easy with sentiment analysis.
– Politics: Acquiring knowledge of social media sentiments is an efficient method
that could be useful for politicians and political organizations. Nowadays, the
feedback of a population on social media is used by leaders in order to be aware
of any potential menaces or conflicts that might arise within their organizations.
– Finance: Thanks to sentiment analysis, investors can spot the promising oppor-
tunities and avoid gambling. They are able to move easily behind their preferable
companies with the intention of guiding their sentiment data in real-time.

2.2 Related Works

In the last few years, considerable attention has been paid to mining sentiment
and opinions from text, due to the available amount of documents and messages
expressing personal opinions on the Internet [1]. Shoukry et al. in [4] Compiled
1,000 tweets about several news subjects. They built two lists: the first contained
negative sentiment words, while the second comprised positive ones. They applied
the unigrams and bigrams features and extracted all Arabic stop words in the training
datasets. The SVM algorithm was applied in their experiment using Weka Suite soft-
ware for the classification process. The accuracy of the SVM classifier was 78.8%.
However, their dataset was relatively of small size.
The work reported by [5] compared the performances of machine- learning and
lexicon-based approaches for the Arabic sentiment analysis. They used maximum
entropy, k-nearest neighbors (KNN), Naive Bayes, and support vector machine
(SVM), for the lexicon-based approach, they constructed a dictionary of Arabic
subjective words, and Their MaxEnt classifier achieved an accuracy of 63% while
for the SVM one, it was of 53.76%. The best accuracy (63.58%) was achieved with
KNN.
Almas et al. (2007) [2] proposed a particular methodology of sentiment analysis
for classifying Arabic and English languages. They used specific features extraction
components that proper for the Arabic language. The resulted accuracy was good
but their training set was limited only to extremist words.
Another work worth mentioning is the project done Kumar et al. (2019) [6]. They
proposed a Bidirectional Soft Attention-Based LSTM (sAtt-BLSTM) approach with
600 M. Garouani et al.

CNN for Sarcasm Detection. The approach was investigated using the SemEval-2015
balanced benchmark and an unbalanced one (consist of 15000 sarcastic and 25000
non-sarcastic ones). They achieved an accuracy of 98%.

2.3 Challenges of Analyzing Arabic Language

Arabic language is a morphologically rich language. This can be described by the

following points:
– Depending on the context, many forms can be extracted from the same root
( ).
– Many users do not employ the SA but an Arabic that is proper to their countries
( ). Moreover, dialect language can be divided into a group of subdialects
according to regions of the country (accents).
– Repeating some letters in order to insist on the feeling’s (a common style found in
informal channels). For example, the word ‘ ’, means thank you very much.
– Arabic language has several diacritics that go with the word. The meaning of the
word changes based on their presence or absence. To illustrate this, the words
“ ” (a sort of Sunday school where children learn about religion) and “ ”
(a book), will be the same word when written without diacritics.
– Depending on its position in the sentence, some words can have different mean-
ings, “ ” a month in the Islamic calendar and “ ” mean this is
forbidden.

3 Methodology

The proposed system is built upon the NLP concepts. Therefore, it contains all the
necessary tools to handle the linguistic diversity that characterizes the communication
of Moroccan users on social network Twitter. We followed the proposed framework
of sentiment analysis bellow as a general methodology, where the sentiment anal-
ysis process is divided into three main phases mainly: data collection, cleaning and
annotating, data preprocessing, classification and evaluation (Fig. 1).

3.1 Dataset Collection

We have collected over 36,114 Moroccan tweets written using the three stages:
– In the first stage, we used the streaming API that allows getting the tweets in
real-time. We have selected tweets written in Arabic that meet the ‘Morocco’
Sentiment Analysis of Moroccan Tweets Using Text Mining 601

Fig. 1. The proposed framework for Moroccan tweets sentiment analysis.

geolocation requirement. We were able to extract 7,578 valid tweets from 23,743
tweets collected between January 21, 2020, and February 24, 2020.
– In the second stage, we used SocialBakers to determine the most active Moroccan
accounts. That gave us a list of 29 names. Then, based on this list, we were able to
extract 5,537 valid tweets from 10,371 tweets collected using the user_timeline
API.
– In the third stage, we resorted to Trendsmap to determine the most popular hash
tags in Morocco, we got about 950 distinct hash tags. This collected data are then
used to download tweets using the search API. We were able to extract 2,345 valid
tweets after filtering the Arabic tweets.

A. Final Corpus
Our final corpus consists of Moroccan tweets that were collected between January
21st and February 24th, 2020. Among the 36,114 collected tweets, the 13550 that
were selected as valid tweets were divided as follows: 8750 were published by men
602 M. Garouani et al.

Table 1 Statistics about the final corpus.

Number of collected tweets 36 114
Number of valid tweets 13 550
Number of distinct users 3602

while 4100 by women and for 700 tweets, the gender of the authors could not be
identified because of the lack of information (Table 1).

B. Data Cleaning
A major issue that faces us when we were dealing with the text posted on Twitter
is the unstructured, noisy, and the informal nature of the posts due to the spelling
mistakes, nonstandard words, lack of standardization, and repetitions. The majority
of the collected tweets are written informally, do not follow any grammatical rule,
and contain many abbreviations and errors. Hence, it is essential to undergo the
preprocessing process to minimize the effect of this informality on our classification.
To improve the accuracy of our classifiers, we detected the misspelled words and we
corrected them. Below, we describe some of the most common methods used during
the preprocessing phase (Table 2):
• Delete numbers, punctuation, web links, emails address, emoticons, Hashtags.
• Remove diacritics (harakat, )
• Eliminate duplicated characters that users use to affirm the meaning ( )

C. Dataset Annotation
Until now, there are no annotated linguistic resources for tweets written by Arabic
Dialect speakers (Moroccan, Tunisian, Algerian…). Remember that each country
has its different dialect. In this work, we focused on tweets written by Moroccan
users. Due to the lack of annotated resources, we decided to build our corpus.

D. Statistics
The dataset has four classes of sentiments, Positive, Negative, Neutral and Mixed
(positive + negative), and two language classes used MD and SA. The distribution
of the data according to their classes is illustrated in the following figures (Fig. 2).

Table 2 Example of tweets before and after preprocessing.

Sentiment Analysis of Moroccan Tweets Using Text Mining 603

Fig. 2. Distribution of polarities in SA and MD corpuses.

Table 3 Normalization Rules.

4 Dataset Experiments

The classifiers: SVM, CNN, LSTM, CNN-LSTM, NB and LR form the algorithms
we have chosen to carry out this study. To train the algorithms, a 3-fold stratified
cross-validation method was used. For the features representation, we used the TF-
IDF (Term Frequency-Inverse Document Frequency) and word embedding represen-
tations based on the n-grams features (unigrams, bigrams and trigrams). Unigrams
are the simplest features extraction method, while bigrams and trigrams provide the
ability to capture negation or sentiment expression patterns [7].

4.1 Dataset Preparation

One of the main purposes of this work is to investigate how preprocessing the tweets
written in Moroccan Arabic (standard or dialect) could enhance the results of the
sentiment analysis of these tweets.

A. Normalizing the Annotated Tweets

We have created a normalization function that unifies the orthography of alifs,
multiple alifs, ettae, daal… Table 3 defines the normalization rules.

B. Moroccan Dialect Stop Words

Due to the absence of any stop words list for the Moroccan dialect, we had to build
our list. It consists of 120 + words. In Table 4, we present some examples of these
stop-words and their meanings.
604 M. Garouani et al.

Table 4 Some dialectal SW with their meanings

Table 5 Some dialectal prefixes and suffixes with their meanings.

C. Stemming the Normalized Tweets

Many efforts on different levels of complexity were done in order to address stemming
on the Arabic language. However, the majority of these works were conducted for
Standard Arabic (SA) and they didn’t handle the different dialect-specific rules such
as Moroccan dialect.
An instance of that, the SA stemmer (ISRI) will treat the word “ ” (tired) as
it’s “ ” (eye) as it ends with ( ); since the suffix “ ” reflects duality. In fact, this
word should not be stemmed originally. It was the reason that pushed us to create a
customized stemmer. Principally, our stemmer aims to reduce the input word to its
shortest form with no changes in its meaning.
To implement the created stemmer, we have constructed two lists: the first treats
irregular terms that should not be stemmed (terms that originally start or end by any
of the prefixes or suffixes), the second one is for irregular plurals, and their singular
forms. The principle of the proposed stemmer consists of three levels: (1) Prefix
removal (2) Suffix removal, (3) Infix removal, which applies the rules for broken
plurals.
Generally, the first achieved level is the prefix removal, followed by the suffix
removal level, and then the Infix elimination level comes at the end. In order to
decide whether to keep on stemming or give up, a verification in the dictionary of the
converted word is done after every level. Table 5 shows part of the sets of prefixes
and suffixes proposed for the Moroccan dialect.

D. 3-Fold Stratified Cross-validation

A 3-fold stratified cross-validation method was used; this implies that the system
automatically split the corpus into three more or less equal parts. We proposed two
tweets analysis tasks that focused on analyzing the sentiment and detect the language
used.
Task 1: Given a tweet, this task consists to identify the language used (SA or
MD).
Sentiment Analysis of Moroccan Tweets Using Text Mining 605

Task 2: Given a tweet written in SA or MD, this task consists of classifying it

according to the emotion expressed by its author (positive, negative, neutral or
mixed).

5 Dataset Experiments

Tables 6 and 7 shows the different scenarios settings that we have evaluated. The first
column represents the used classifier name. The second column indicates the used
features and whether a Stopwords filter was used (1) or not (0). The third column
shows the accuracy of each corpus used (only SA, only MD and the fusion of the two
corpora). In the fourth column, we have trained the classifiers in the whole corpora
and we carried out separate tests (a test on SA corpus and MD corpus).
The conclusions drawn from the experimental evaluation results of the combina-
tion of classical classifiers with a TF-IDF weighting system presented in Table 6 can
be summarized as follows:

Table 6 Results of ML classifiers + TF-IDF

Model Features Accuracy Accuracy all
AS DM AS_DM AS DM
SVM TF-IDF Uni-gram 0 83.50 67.01 78.70 82.06 70.08
1 82.30 66.75 78.77 82.78 68.07
Bi-gram 0 84.75 67.80 80.05 83.84 70.00
LR 1 83.47 68.68 79.41 82.54 71.40
Uni-gram 0 82.23 64.38 78.18 82.03 69.82
1 82.07 65.78 78.55 82.44 68.59
Bi-gram 0 81.27 62.36 77.52 80.93 68.77
1 81.14 60.88 78.08 81.14 70.26

Table 7 Results of deep learning classifiers using word

Model Features Accuracy Accuracy all
AS DM AS_DM AS DM
CNN Word 0 91.78 84.17 89.56 91.62 85.37
embeddings 1 91.39 83.82 89.08 91.12 84.78
LSTM 0 92.09 83.36 89.60 91.80 84.69
1 91.64 82.87 89.49 91.36 85.04
LSTM-CNN 0 91.83 81.75 89.46 91.50 85.55
1 91.67 82.00 88.93 91.46 85.26
606 M. Garouani et al.

(1) The SVM classifier using the different grams (Unigrams and Bigrams) achieved
a very good accuracy (84.75%) in the first corpus SA using bigrams without
eliminating stop words, and an accuracy of 80.05% in the third corpus MSDA.
(2) Similarly, the Logistic Regression classifier achieved 82.23% of accuracy for
the corpus SA, using the unigrams without eliminating the stops words, and
78.55% for the corpus (MSDA) by eliminating the Stop Words.
Concerning Table 7, the results of the combination of the Deep Learning classifiers
with the Word Embedding representation model were presented as follows:
(1) The CNN classifier achieved accuracy results of 91.78% keeping the stop
words, and 91.39% eliminating them in the first corpus (SA), and attractive
results in the dialectal corpus with an accuracy of 85.37%, and 89.56% in the
combination of the two corpora.
(2) The combination of the two classifiers CNN- LSTM without elimination of
the stop words obtained a good result with an accuracy of 91.83% in the first
corpus (SA), and 85.55% in the corpus MD.

Task 1:
Table 8 presents the results of classification according to the accuracy for the second
task. The best results achieved in term of the accuracy metric are 88.86% using
LSTM without elimination of Stopwords and 87.95% by filtering them.
From Tables 8 and 9, we can make the following observations:
1. In the experiments, removing stop words decreased all the classifiers perfor-
mance by almost 0.6%, which showed their good effect on the semantic meaning
of the text.
2. In general, the Bigrams function and the inclusion of Stopwords have obtained
the best accuracy results in all classifiers.
3. According to the results obtained, it is clear that deep learning algorithms
(LSTM, CNN, CNN-LSTM) outperform other classical algorithms such as LR
and SVM.
The methods, size of corpora and performances of some state-of-the art papers
are summarized in Table 9.

Table 8 Results of
Model Features SW filter Accuracy
Identification of the used
language LSTM Word embeddings 0 88.86
1 87.95
CNN-LSTM Word embeddings 0 88.43
1 87.06
SVM TF-IDF 0 88.39
1 87.25
LR TF-IDF 0 88.17
1 87.02
Sentiment Analysis of Moroccan Tweets Using Text Mining 607

Table 9 Summary of some works on sentiment analysis of Arabic tweets

Paper Corpus size Arabic variant Result
Moncef and Hanae 13.550 Marocain (Standard & Dialectal) 92.09%
Heikal et al. (2018) [8] 3.315 Arabe (Standard & Dialectal) 64.30%
Al-Twairesh et al. (2017) [9] 17.573 Saoudite (Standard & Dialectal) 62.27%
Nabil et al. (2015) [10] 10.000 Egyptien (Standard & Dialectal) 69.1%
Abdulla et al. (2015) [11] 2.000 Jordanian (Standard & Dialectal) 87.2%

6 Conclusion

In this works, we addressed sentiment analysis in Moroccan Arabic tweets. We had

collected about 36000 tweets and we had manually labeled 13550+ tweets. For each
single tweet, we assigned a category (SA or MD) and a Positive, Negative, Neutral
or Mixed label.
In our experiments, using 13550 tweets distributed on 3-fold stratified cross-
validation, our models have achieved great results. We note that the results of the
deep learning algorithms were better than those of classical techniques such as SVM,
and LR. Because of the increasing usage of diverse dialects in Morocco, our system
still needs some improvements. Moreover, the Moroccan dialect is not stable which
can might alter the sense of some words. Hence, improving the analysis of the
Moroccan dialect will require more future improvements. In the future, we plan to
further extend and evaluate our framework by applying it to different kinds of use
cases. This may also require increasing the size of the dataset, especially the MD
dataset, adding support for further parameters, features, and ML algorithms.

References

1. Pang B, Lee L (2008) Opinion mining and sentiment analysis. Found Trends Inf Retrieval
2(1–2):1–135
2. Yousif A, Ahmad K (2007) A note on extracting sentiments in financial news in English, Arabic
& Urdu. In: The Second Workshop on Computation, al Approaches to Arabic Script-Based
Languages, pp 21–22
3. Mohamed E, Elfeky M (2010) Mining Arabic business reviews. In: IEEE International
Conference on Data Mining Workshops (ICDMW), 13 December 2010, pp 1108–1113
4. Amira S, Rafea A (2012) Preprocessing Egyptian dialect tweets for sentiment mining. In: The
4th Workshop on Computational Approaches to Arabic, AMTA 2012
5. El-Halees A (2011) Arabic opinion mining using combined classification approach. In:
Proceeding of the International Arab Conference on Information Technology
6. Kumar A, Sangwan SR, Arora A, Nayyar A, Abdel-Basset M (2019) Sarcasm detection using
soft attention-based bidirectional long short-term memory model with convolution network.
IEEE Access 7:23319–23328
7. Khreisata L (2009) A machine learning approach for Arabic text classification using N-gram
frequency statistics. J Inform 3(1):72–77
608 M. Garouani et al.

8. Heikal M, Torki M, El-Makky N (2018) Sentiment analysis of Arabic tweets using deep
learning. Procedia Comput Sci 142:114–122
9. Al-Twairesh N, Al-Khalifa H, Al-Salman A, Al-Ohali Y (2017) AraSenTi-tweet: a corpus for
arabic sentiment analysis of Saudi tweets. Procedia Comput Sci 117:63–72
10. Nabil M, Aly M, Atiya A (September 2015) ASTD: Arabic sentiment tweets dataset, pp 2515–
2519
11. Nawaf AA, et al (2015) Towards improving the Lexicon-based approach for Arabic sentiment
analysis. Int J Inf Technol Web Eng
Contribution of Digital Collaboration
and E-Learning to the Implementation
of Smart Mobility in Morocco

Mohammed Mouhcine Maaroufi, Laila Stour, and Ali Agoumi

Abstract This article attempts to highlight the need to rethink, more intelligently in
Morocco, how to manage mobility, a major pillar of economic competitiveness and
social well-being, by first initiating a broad reflection on the very needs of displace-
ment, using as much as possible digital collaboration, e-learning and distance training
and then the massive use of Intelligent Transport Systems in order to ensure optimal
exploitation of the existing infrastructure before to launch cumbersome investment
for irreversible infrastructure projects. Our analysis has shown that with the compre-
hensive diagnoses made and the clear and consistent strategies that have been adopted
in Morocco, there is only a tiny link missing to put digital and new technologies to the
benefit of human and sustainable development. Indeed, it only takes a national and
unified mobilization to move from the stage of digital emergence, to the transition
and then to the transformational phase. A digital ecosystem is emerging, bringing
innovative solutions through artificial intelligence, virtual reality, and augmented
reality, which need to be deployed to measure their reach and their limits and try to
evolve them to meet increasingly diverse and constraining collaboration and learning
needs.

Keywords Smart mobility · Sustainable mobility · Digital collaboration ·

E-learning · Distance training · Virtual reality · Virtual platform

1 Introduction

Digital technologies around the world are bringing structural changes in industries,
economies and social relationships. The agility they offer allows for cost optimization
and the resolution of the most complex problems.

M. M. Maaroufi (B) · L. Stour

Faculty of Science and Technology of Mohammedia, Hassan II University of Casablanca,
Casablanca, Morocco
A. Agoumi
Civil Engineering, Hydraulics, Environment and Climate Laboratory, Hassania School of Public
Works, Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 609
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_55
610 M. M. Maaroufi et al.

Since the approval of the new constitution in 2011, Morocco has taken steps to
achieve a digital transformation in the service of the administration, the economy
and its cities.
The Royal speech, at the opening of the legislature on 13/10/2016, outlined the
strategic directions inviting the executive to use new technologies for an upgrade of
the administration in favor of the citizen in Morocco. His Majesty the King said on
October 13, 2017 that Moroccans will need a quality education that allows them easy
access to the labor market.
The king’s new development model will be a boon for a shared, integrated and
comprehensive vision for the next ten or fifteen years and will provide a framework
where humans are at the center of development.
It is therefore obvious that in order to make Morocco a country of solidarity and
prosperity, we must avoid the urban/rural digital divide by opting for smart cities but
also for an attractive and connected rural world [1].
After describing the problem, we will first review the initiatives and strategies
concerning mobility, training and digital in Morocco and in order to encourage inno-
vation within Moroccan digital startups, we will then present a prototype virtual
platform developed by a Moroccan startup that can answer many questions using
Virtual Reality and Increased Reality. Finally, we will conclude with a few recom-
mendations. At no time do we claim the completeness of the findings, analyses or
solutions, these are only strands that can develop further and continuously.

2 Research Problem

Smart cities can be identified by six closely linked levers: smart economy, smart
mobility, smart environment, smart people, smart lifestyle and a smart administra-
tion [2]. So, we can’t call it a “smart city” without talking about “smart mobility.”
Indeed, the 4th industrial revolution and the digital transition have led to a rethinking
of how to manage mobility: Designing and building road infrastructure and imag-
ining new services for “connected mobility” without excessive waiting or congestion
at interconnections. Provide agile, functional and cost-effective solutions for opera-
tions and maintenance (smart sensors, user information solutions, infrastructure and
intelligent transport systems, etc.).
While the term “smart” refers to better use of technology, intelligent mobility is
also defined as “a set of coordinated actions to improve the efficiency and environ-
mental sustainability of cities. [3].” The term “mobility” highlights the preponderance
of humans in relation to infrastructure and vehicles.
In the face of traffic congestion, the solutions currently adopted, which consist of
addressing the problem by developing and expanding infrastructures, only encourage
increased travel and postpone the problem with more criticality. It is true that beyond
a certain density of traffic, even ITS, are out of date. Hence the need to act on
the problem at the root level by optimizing the very need for displacement. Such
action can only be successful through the provision of innovative solutions to serve,
Contribution of Digital Collaboration and E-Learning … 611

collaborate and train remotely. At this point, dynamic “mobility” technologies will
be able to manage and smooth traffic efficiency. It is only by combining agility,
efficiency and sustainability that we can speak of “smart mobility”.
Therefore, the questions that need to be asked are: what “smart mobility” to adopt
in Moroccan cities? Is it just “connected mobility”? Or, first, a “sustainable mobility”
that contributes to a humane, inclusive and resilient city and which is, above all, a
thoughtful, optimal and coherent mobility? And how can digital collaboration and
distance learning help build “smart mobility”?

3 Background on Mobility and Training

and the Challenges of Digital Transformation in Morocco

3.1 Mobilize Your City Initiative1

100 metropolises have committed to the “Mobilize Your City (MYC)” initiative
aimed at a proactive transition and a profound cultural and organizational transfor-
mation of people’s movements and related logistics activities in order to make cities
more inclusive and reduce their Green House Gas (GHG) emissions. This initiative
builds on the Networks’ Collective Intelligence and existing planning practices and
incorporates the principles of integration, participation and evaluation.
Morocco is one of the first countries to be engaged in the MYC initiative at COP22
in Marrakech (2016). The desire to include mobility planning in 26 municipalities, in
the Sustainable Development approach, resulting from the MYC project, was aimed
at:
– Improving the quality of life and economic attractiveness of the city;
– Offer alternatives to individual vehicles;
– Put in place a Sustainable Urban Mobility Plan (SUMP).

3.2 Roadmap for Sustainable Mobility in Morocco

Morocco has seen significant progress and reforms in the areas of the environment,
Sustainable Development (SD) and the fight against Climate Change (CC) over the
past decade. Several sectoral strategies, including transport and logistics, integrate
the environmental dimension. The economic stakes are important since the cost of
pollution in the area in Morocco represents more than 10 Md DH (1% of GDP) [4].
The National Sustainable Development Strategy (SNDD 2030) identified the
transport sector as the third largest energy consumer in Morocco in 2014. It accounts

1 https://mobiliseyourcity.net/.
612 M. M. Maaroufi et al.

for 16% of total emissions and 28% of emissions from the energy module. Sustain-
able mobility is defined as “a transport policy that seeks to reconcile accessibility,
economic progress and the reduction of the environmental impacts of selected
transport” [5].
In addition, Morocco is the first country to have undertaken an adaptation of the
global macro-roadmap for the transformation of transport from the Paris Process on
Mobility and Climate (PPMC). Morocco’s 2018 roadmap, which aims to support
national strategies and in particular the SNDD, recommends less travel for more
productivity and better quality of life. Indeed, time spent in transport limits the
productivity of the economy and the quality of life of citizens. Less travel means more
time and resources for value-added activities and social life. The roadmap also calls
on governments to move forward with dialogue between employers’ and employees’
organizations in favor of remote work in order to reduce commuter kilometers.

3.3 National Vocational Training Strategy (SNFP 2021)

SNFP 2021 aims for quality vocational training everywhere, for all and throughout
life, at the service of the development and enhancement of human capital and for a
better competitiveness of the company [6]. This strategic vision is oriented in three
directions: The citizen, the enterprise and the state/territorial communities.
In order to have a flexible, responsive and professionally rooted vocational training
system, SNFP 2021 has five strategic objectives:
– Ensuring the right to vocational training: social inclusion and territorial extent
through the training of 10 million citizens;
– Improving the company’s competitiveness as an actor and a privileged training
space: Training, from 2021, of 20% per year of employees and 50% of trainees
in professional settings;
– Maximizing professional integration through continuous improvement in the
quality of training: 75% of the winners’ integration rate by 2021;
– Have on-demand training and ensure a better attractiveness of vocational training
to enable young people to express their vocations;
– Renovate and strengthen the governance of public policy of vocational training
inducing performance and synergy of the system.
As part of the ‘Digital Inclusion Vector Distance Training’ project, the ‘Devel-
opment of Distance Training as a New Training’ project aims to dedicate and insti-
tutionalize distance learning as a form of training and graduation and certification
in addition to strengthening digital inclusion. The specific objectives of this project
are:
– Improving access to vocational training (young people, workers, people with
special needs, housewives, inmates, etc.)
– Optimize human and financial resources dedicated to vocational training.
Contribution of Digital Collaboration and E-Learning … 613

3.4 Strategic Vision for the 2015–2030 Reform of Éducation,

Training and Scientific Research

The strategic vision for the 2015–2030 reform, which has been translated into
Framework Act 51.17 on the education, training and scientific research system,
aims to establish a new school with three main foundations: equity and equality
of opportunity, quality for all and the promotion of the individual and society [7].
From this perspective, the first objective of change is the transition from the logic
of the linear transmission of knowledge and memorization to a logic of learning, the
development of a sense of criticism, the construction of a personal project, the acqui-
sition of languages, knowledge and skills, values and digital technologies. Thus, the
new school will promote the transition from a knowledge-consumption society to a
society that produces and disseminates it, through the mastery of digital technolo-
gies and the development of scientific research and the culture of innovation and
excellence.

3.5 Maroc Digital 20202

Since 1996, Morocco has been part of the restructuring of the field of informa-
tion technology by the adoption of Law 24–96, which has enabled the launch of
a first step in the liberalization of the telecommunications sector. In this context,
Morocco has adopted multiple digital strategies: the 1999–2003 five-year plan, e-
Morocco 2010 covering the period 2005–2010 and the MN2013 strategy, which is
being implemented over the period 2009–2013 [1].
In order to further boost the development of the digital economy sector, a new
vision for digital technology by 2020 has been developed. The first of the three pillars
of the strategy aims to transform the national economy digitally through the reduction
of the digital divide for the benefit of citizens (connectivity for all, primary, secondary
and higher education programs, public access to WIFI) and businesses in particular
small and medium companies/very small companies (regions disadvantaged in terms
of connectivity).
The Agency of Digital Development (ADD), created in 2017 with the main
mission of implementing the state’s digital development strategy, has drawn up a
note in 2019 to propose to the government guidelines for the development of digital
by 2025.
In December 2019, the prospects for digital development in Morocco were defined
with four main challenges:
a. A digital administration (Smart Government) at the service of citizens and busi-
nesses in order to reduce the rate of dissatisfaction of citizens/companies from

2 https://www.add.gov.ma/.
614 M. M. Maaroufi et al.

85–90% to less than 20% through the transformation of interactions with public
administration through end-to-end digitalization of citizens/business paths;
b. A competitive economy thanks to the performance gains brought by a digital
ecosystem and innovation through the Start-up Program, which consists of the
installation of a fabric of more than 2,500 startups, over the next five years,
Digital park, Smart factory, Digital SME: a complete device for digitalization
of Small and Medium-sized Enterprises (SMEs), and Digital self-entrepreneur.
Thus, innovation and digital can be a lever that creates wealth;
c. An inclusive society through digital technology. Digitalization can open up
new opportunities by giving less advantaged populations access to information,
social benefits, care services, and education. It offers, for example, digital classes
in Moroccan schools located in remote areas.
d. An environment and digital trust. This component mainly concerns the estab-
lishment of a regulatory framework, a national programme of higher educa-
tion, vocational training and continuing training for digital professions, called
‘Digital Generation’ and its support ‘Digital Infrastructure’ for the establish-
ment of an advanced infrastructure (very high-speed, 5G, government cloud,
regional data centers…) necessary for the digital transformation of the country.
It is also planned to contribute to the launch of an ‘AI Ecosystem’ program on
artificial intelligence thematic research, based on cases of sectoral use, with the aim
of coordinating around the targeted research topic the prime contractors or end-
customers, innovative startups and public and private training bodies. This project
aims to set up an ecosystem dedicated to artificial intelligence on the natural language
processing and natural language generation components for digital inclusion with a
conversational AI engine.

4 Analysis of the Situation of Mobility, Training and Digital

Transformation in Morocco

4.1 Analysis of the Mobility Situation in Morocco

With 4.3 million vehicles, the car fleet more than doubled in Morocco between
2002 and 2018 [8]. Travel trends and needs are also increasing, leading to increased
demand for people and goods.
The management of mobility in cities, undergoing an acceleration of the
phenomenon of urbanization and rural exodus, is increasingly becoming a thorny
constraint. Failure to do so will inevitably prevent their transformation into smart
cities.
Road transport infrastructure is under pressure. The solutions adopted to date
are always on infrastructure with large-scale, extremely expensive and definitive
actions (cable-stayed bridges, hoppers, road widening, etc.). Improved infrastruc-
ture generates additional induced traffic and existing diverted traffic, causing further
Contribution of Digital Collaboration and E-Learning … 615

congestion. To remedy this situation, managers are using Intelligent Transportation

Systems with time-based processing to optimize infrastructure capacity as there is
no longer the possibility of acting on space. Finally, they seek to encourage modal
deferral and limit non-essential travel.
The challenge is therefore to reverse practices and change cultures, starting with
the use of AI and NTIC to reduce “not-intelligent mobility” in order to streamline
the construction of new training centers, new administrations and transport infras-
tructure, as well as to promote the inclusion of several categories of geographically
or physically marginalized citizens.
In a health or security crisis, on days of transport strikes, and more broadly in
view of the increasing difficulties of travel in large urban areas or due to remoteness,
collaboration and distance training are positioned as a promising alternative.

4.2 Analysis of the Professional Training Situation

in Morocco

With regard to e-learning, SNFP 2021 has only planned the design and development
of an e-Learning platform [6]. Virtual Reality and Augmented Reality based on
Artificial Intelligence, essential for the practical assimilation of several disciplines,
have not been programmed. At the same time, dealing with near-reality situations
through augmented reality can also energize face-to-face training in an innovative
Blended Learning approach.
Indeed, transmitting knowledge to learners through Immersive Learning is the best
way to enable practice and move into Agile Mode with Virtual Training, experiential
learning, immersive simulations and digital training modules.
SNFP 2021 has planned several actions scheduled for 2017–2018 in order to offer
a demand-driven training offer. However, it has not put in place a process for the
retraining of young graduates, self-entrepreneurs and workers who will be brought
into line with the new tools and concepts of the labor market.

4.3 SWOT Analysis of the Situation of Digital

Transformation in Morocco

The SWOT analysis, conducted in this work, summarizes the diagnostic elements
of the situation of digital transformation in Morocco. It sees the strengths of the
digital ecosystem, consolidated through the various strategies developed and raises
the shortcomings, which remain to be perfected. This analysis is also an opportunity
to recall the opportunities that can boost the implementation of digital in the Kingdom
and outlines some risks that may hinder its implementation (see Table 1).
616 M. M. Maaroufi et al.

Table 1 SWOT analysis of the digital transformation situation in Morocco

Strengths Weaknesses
Bet on Artificial Intelligence Inadequate broadband infrastructure
Prioritization of 16 projects according to the 4 Lack of interest in Intelligent Transportation
strategic axes defined by the Digital Strategy Systems
Creation of the ADD for the implementation of Robotics and AI training courses preparing for
the Digital Strategy jobs, not real careers
There are national champions in the public and Modest banking of the population and
private sectors transactions over the internet
Recognized export performance (customer Inadequate communication and provision of
relationship, offshoring, means of payment) data around the Digital Strategy
Moroccan researchers and engineers’ capacity Delay in the institution of collaboration and
for innovation distance training as a complement to work and
Encouraging self-entrepreneurship to promote face-to-face training
flexibility in work and skills
Opportunities Threats
Digital is a source of innovation and the Competition from resource-rich, low-cost
development of new services and jobs in countries such as Tunisia, Turkey and South
Morocco Africa
Digital is an opportunity to make a switch Massive flight of qualified executives abroad
towards a new mindset
Digital technology is a fundamental In the absence of transparency and
transformation of government and territories, commitment, digital technology will always be
serving citizens and businesses perceived as a threat by some large public and
private companies
Digital technology enhances the agility of The non-inclusion of all citizens and territories
businesses and administrations (collaborative in the digital transformation risks exacerbating
and dematerialization platforms) the urban/rural gap
Digital is a source of soft power and allows the Digital risks call for the development of a
development of specialized offerings (Smart platform approach by Moroccan
factory, Smart city, Smart health, Additive 3D administrations and companies
Manufacturing, Cloud computing, Big data, Securing funding sources to implement this
Internet of Things (IoT), Blockchains, Strategy
Cybersecurity, conversational AI, Maching
Learning)

5 Prototype of Virtual Platform for Collaboration

and E-Learning

5.1 Description of the Prototype

At a time when dematerialization is booming, the virtual platform of collaboration

and training constitutes a new channel of exchange and sharing combining innova-
tion, efficiency and profitability, which complements the levers of the creation of
Contribution of Digital Collaboration and E-Learning … 617

the ecosystem dedicated to artificial intelligence and allows to promote a dynamic

image and position Morocco of tomorrow as a leader on an international scale., in
the interactive virtual exhibits and e-learning industry.
The virtual platform, 100% eco-friendly, also contributes to the democratization
of exhibition and training opportunities for the benefit of self-entrepreneurs, SMEs
and students. With state-of-the-art tools in an immersive, user-friendly and interactive
environment, it offers wide geographic coverage.
The virtual platform of collaboration and training, global trend, uses virtual reality
to position itself as a new form of exchange and learning face-to-face. It is a playful,
attractive and intuitive WEB environment, replicating the atmosphere and context of
a physical event online.

5.2 Components of the Prototype

Provided in Platform as a Service mode (PaaS), this digital solution is accessible from
any connected device. It is fully customizable and adapts to any type of event or exhi-
bition, public or private service, training or conference, meeting or networking space
or open houses. This virtual meeting place, with innovative features (live streaming,
machine and instant translation, Artificial Intelligence, Smart Data), offers an acces-
sible platform, dematerialized logistics and online management tools for admin-
istrations, universities, public facilities, marketplace-B2B, museum, media library,
conference and seminar center, exhibition halls, showroom.
A true 3D virtual replica of the “hard” environment, only the tools change! 3D
customization allows you to choose the layout and layout of virtual spaces (walls,
floor, visual and display colors, animation videos) to the identity of the exhibitor or
speaker (see Sect. 5.4 Models of the prototype).

5.3 Benefits of the Virtual Platform

No more stress of late printing, orders for furniture leading to chaotic deliveries. The
brochures, and training materials become downloadable files and the virtual screen
allows you to view all presentations.
The virtual platform is a matching space, matching a dynamic offering to a diverse
and unpredictable demand via an intelligent preference system that adapts and capi-
talizes to submit categorized and qualified suggestions. For example, reception areas
in the form of information desks, thematic halls with virtual booths for a permanent
or one-off exhibition of products and services, conference rooms and workshops,
and training rooms at Augmented Reality allow for an ideal responsiveness in Agile
Mode. The platform includes:
– An essential complement and an extension of physical events;
618 M. M. Maaroufi et al.

– A true Eco-responsible initiative preserving the environment (less travel, less air
pollution and noise pollution, less waste and materials used to build hard spaces);
– A time-saving opportunity and a contribution to road safety;
– Unlimited accessibility space, including People with Reduced Mobility;
– A capitalization assured through 100% rise;
– Insurance of no risk of cancellation in the event of strikes, epidemics or natural
disasters;
– An alternative to making automated webinars with the possibility of ques-
tion/answer interactivity;
– Immersive and intuitive navigation in real time.

5.4 Models of the Prototype

The virtual platform, developed by a Moroccan startup, was able to succeed several
events in different editions. Thus, the 2nd edition of the Virtual Building and Public
Works Show BTPEXPO3 in 2016, was the first virtual exhibition in the world to
receive the COP22 label and saw the participation of 200 national and international
exhibitors and more than 20,000 qualified visitors, as well as the broadcast of 9
thematic conferences over 3 days.
On the occasion of the fiftieth anniversary of its creation, the National Docu-
mentation Centre organized in 2016, using this platform, the first Moroccan Virtual
Interactive Show of the standby and the documentation DOCEXPO4 .
Other successful events have been organized such as the Virtual Forum of Self-
Entrepreneurs AEXPO5 and the Virtual Immovable Show IMMOEXPO6 .

6 Conclusions and Recommendations

It appears that digital collaboration and e-learning is not yet widely used in Morocco.
The integration of Virtual Reality and Augmented Reality as in the prototype
presented in this article will be able to solve certain problems related, for example to
practical work in teaching and the non-friendly and static environment in traditional
collaboration and service platforms.
The scope of possibilities for remote work offered by virtual reality is consid-
erable. Indeed, new ways of working together from a distance are emerging. More
immersive, more realistic and replicating our professional environment. But the issue
of telework is broader than that because working at home requires skills such as

3 https://www.youtube.com/watch?v=kRv7LdWx5-0.
4 https://www.youtube.com/watch?v=IDoPO9oAJL0.
5 https://www.youtube.com/watch?v=V0kDckSk7ko.
6 https://youtu.be/7S0h2NQfJ8M.
Contribution of Digital Collaboration and E-Learning … 619

autonomy, rigor, concentration, but also accept to be sometimes a little isolated. To

avoid this it is important as for a mixed formation to find the right balance between
the virtual and the real.
In any case, mobility will only benefit from the reduction in non-essential travel.
ITS will then be able to ensure optimized and efficient traffic management.
In addition, a barometer to take stock of the dynamics of digital collaboration
and e-learning in Morocco, the e-learning strategies and projects developed or under
development in the Kingdom and their real impacts on the reduction of displacement
will be able to chart the roadmap to follow. Starting from the point of view of
the various stakeholders, this barometer will have to explain how the concept of
immersive learning using Augmented Reality sees mixed reality is perceived, the
prerequisites necessary for its implementation as well as the priority themes to be
developed. In addition, it will need to identify the means of funding available or to be
activated and initiatives to strengthen the dynamics of the actors (public and private
actors as well as citizens). Finally, this barometer should indicate the steps taken
and the obstacles faced by startups in the implementation and development of new
technologies and online learning management systems.

References

1. CC (Court of Auditors) (2014) Summary of the evaluation of the Morocco Digital Strategy 2013,
13 p
2. Giffinger R, Fertner C, Kramar H, Kalasek R, Pichler-Milanovic N, Meijers E (2007) Smart
Cities: Ranking of European Medium-Sized Cities. Centre of Regional Science, Vienna, 12 p
3. Benevolo C, Dameri RP, D’Auria B (2016) Smart mobility in smart city. In: Torre T, Braccini
A, Spinelli R (eds) Empowering Organizations. Reading Notes in Information Systems and
Organisation, pp 13–28
4. METLE (Ministry of Equipment, Transport, Logistics and Water) (2018) Roadmap for
Sustainable Mobility in Morocco, 19 p
5. SEDD (Secretary of State to the Minister of Energy, Mines and Sustainable Development,
responsible for Sustainable Development) (2017) National Sustainable Development Strategy
2030. Executive Summary, 60 p
6. SEFP (State Secretariat for Vocational Training) (2015) National Professional Training Strategy
2021, 89 p
7. CSEFRS (Higher Council for Education, Training and Scientific Research) (2014) Strategic
vision for the 2015–2030 reform, 27 p
8. METLE (Ministry of Equipment, Transport, Logistics and Water) (2019) The METLE in
numbers 2018, 37 p
MOOCs Semantic Interoperability:
Towards Unified and Pedagogically
Enriched Model for Building a Linked
Data Repository

Hanane Sebbaq and Nour-eddine El Faddouli

Abstract The proliferation of MOOCs has led to an issue of heterogeneity of data

schemas and a problem of interoperability between different MOOC providers. The
goal of this research paper is to propose a unified data model that serves as a basis
for the design and implementation of a Linked Data repository. In order to achieve
this goal, on the one hand, we rely on semantic web technologies: Linked Data
and Ontology. On the other hand, we enrich this model with metadata related to the
pedagogical aspect of MOOCs using Bloom’s Taxonomy as a theoretical foundation.

Keywords Semantic interoperability · MOOC ontology · Linked data · Bloom’s

Taxonomy · RDF · Jena Fuseki Server · SPARQL

1 Introduction

At the end of 2019 and during 2020, the spread of COVID-19 caused the shutdown of
all educational institutions around the world. In order to ensure the continuity of the
learning process, many universities have turned to E-learning and especially MOOCs.
The number of users of MOOCs has doubled compared to 2019 [1] before the spread
of the virus. Many instructors adopt MOOCs to provide their academic courses.
However, there are challenges to overcome, such as the design of the course. This
work is part of a general framework of recommendation system [2], which plans to
benefit from the existing models, reduce the workload and limit the expense of making
MOOCs. The aim here is to help the MOOC makers find suitable and personalized
content from different providers. However, each provider handles different formats
or schemas of metadata or vocabulary; this infers an issue of heterogeneity.
Our purpose is to gather those different pieces of MOOCs in a new MOOC, and
this in spite of the basic heterogeneity between resources. This is the premise of the
idea of interoperability. The interoperability of educational resources has been the
subject of several studies, both at the standardization level, and at the architectural

H. Sebbaq (B) · N. El Faddouli

RIME Team, MASI Laboratory, E3S Research Center, Mohammadia School of Engineers,
Mohammed V University in Rabat, Rabat, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 621
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_56
622 H. Sebbaq and N. E. Faddouli

level of an integrated repository. At the standardization level, few research papers

have focused on MOOCs. While some of them proposed MOOC ontology, they
do not cover the concepts necessary for our research, such as pedagogical model or
learning objectives. In this sense we propose a MOOC ontology as a reference model
that will serve as a standard to unify the representation of MOOCS and facilitate
interoperability between MOOCs platforms. We enrich this ontology with metadata
about learning objectives classified according to the Bloom’s Taxonomy. At the
architectural level of an integrated repository, most of related researches do not
go as far as proposing architecture for a data repository and they do not rely on
Linked Data. In our work, we base our method on the standards of Linked Data to
propose Framework of a MOOCs data repository. This Linked Data repository gives
semantically rich descriptive metadata extracted from different MOOC providers
that is accessible through a SPARQL endpoint.
The content of this paper is structured as follows: Sect. 2 provides a background on
existing standards that allow the collection of metadata from educational resources
and shows the limits of these standards in relation to our research objective. Further-
more, it introduces how the use of ontology, semantic web and Linked Data can
enable us to overcome these limits. This section also gives a review on related work
about, on the one hand, a unified model to resolve the issue of diversity of standards,
and on the other hand, an architecture for the integration of heterogeneous metadata
of MOOCs. Section 3 describes our proposed approach. Section 4 gives an outline of
the stages of implementation of the various modules. Finally, in Sect. 5 a conclusion
of this paper is drawn and further research work is proposed.

2 Background and Literature Review

2.1 Semantic Web to Enhance Educational Resources

Interoperability

Interoperability is essentially the integration of heterogeneous systems to enable

human users or machines to discover resources. Educational resources [3, 4] should
be consistently defined and labeled by a standard schema in order to facilitate the
indexing, storing, searching and retrieving of these resources by any agent (human or
software). Metadata is called the data used for this reason. Such metadata, however,
must obey very particular standards and norms.One of the most frequently used
standards on the Web of Data is the Dublin Core metadata schema [5, 6]. Dublin
Core is a simple and generic descriptive format that is presented in XML. Another
standard is ‘IEEE 1484.12.1–2002’ (LOM) [7] which is an acronym for “Metadata
from Learning Objects”. Although the use of LOM has been important, problematic
issues have arisen. The semantics of the content are poorly described because they
are mostly limited to keywords [10]. A new ISO 19788 (MLR) (Metadata for Tools of
Learning) standard was established in 2011. One of MLR’s objectives is to harmonize
MOOCs Semantic Interoperability … 623

the metadata of LOM and Dublin Core, as it seeks to allow both the aspects of LOM’s
“learning object” and the Abstract Dublin Core Model [8]. For this purpose, MLR
can be the solution to the interoperability problem between educational resources.
In 2013, the Dublin Core initiative and several major players in the IT industry,
including Google, Bing and Yahoo search engines, support the development of the
LRMI standard [9]. Their primary purpose is to incorporate elements of educational
metadata. Based on these elements, the LRMI will extend the interoperability to the
worldwide search engine level and will be able to allow specialized research [9].
Semantic interoperability is one of the challenges facing the world of E-learning
at present. Different learning repositories freely exchange a large number of online
content and data, and these repositories are based on various implementation
concepts, making the integration process expensive. Our research work aims to
integrate existing MOOCS providers and is focused on current learning systems
comprising very rich and varied content. Standards such as LRMI could be an
adequate solution, but because it is only recommended for systems under develop-
ment, it cannot be adopted. Our work therefore focuses on the integration of MOOCs
in particular, while LOM and Dublin Core do not cover the MOOC field. The LRMI
integrates such MOOC specific fields, including the type of learning resource and the
time needed, but does not cover other areas such as learning objectives or pedagogical
model necessary for our work.
The above restrictions push us to propose a unified and specific model for MOOCs.
Our aim is to ensure semantic interoperability between the various MOOC providers
and to provide a shared understanding of the terms that will be exchanged. In order to
address the problems of semantic heterogeneity and interoperability of MOOCs data,
semantic web technologies based on ontologies and linked data appear as a solution.
The strength of ontology lies in its ability to describe information closely connected
by syntactic or semantic relationships [11]. Those ontologies are used by Linked
Data to link and integrate data from different sources and with different formats. In
the field of education, linked data is relatively unexplored; we have already checked
this in our previous work [2].

2.2 Related Work

Several studies have discussed the problem of the interoperability of various

educational resource repositories. This issue was dealt with in two ways:
• The proposal for a unified model to overcome the problem of standards diversity.
• The proposal for an architecture of a Linked Data repository to integrate
heterogeneous metadata of MOOCs.
MOOC Ontology. We looked for studies dealing with the development of MOOC
Ontology. In this regard, “https://schema.org” offers a MOOC vocabulary but does
not cover the entire MOOC domain. The “MOOCLink” project [12] proposes
a semantic data model for the integration of MOOC data by extending the
624 H. Sebbaq and N. E. Faddouli

“schema.rdfs.org” ontology. However, the “schema.rdfs.org” site is no longer avail-

able. In addition, it does not cover all MOOC concepts, including the pedagog-
ical component or learning objectives. Similarly, the work in the paper [13] builds
a MOOC ontology but this ontology is not available on the WEB, beside this,
it only covers the general concepts of MOOCs and does not go deeper. In [14],
researchers rely on the Edx ontology and higher-level ontologies for the design of
their MOOC ontology. This ontology does not suit our need because it is limited to
the Edx provider. For their part, the authors of [15] develop an ontology based on
the “Singapore System for Dublin Core Application Profiles” metadata profile. This
ontology represents the teaching material for computer science courses in primary
and secondary education in Greece. Moreover, it does not go as far as suggesting
a data repository, and does not cover MOOCs. In the same sense, authors in the
paper [16] propose the “DCOntoRep” ontology based on the metadata involved in
the Dublin Core standard. They suggest an ontology-based model that promotes
semantic interoperability between OERs, independently of the metadata standard to
be adopted. However, they do not go to its implementation. In [17], authors suggest a
generic ontology as part of a case-based recommendation framework for the indexing
of OERs and MOOCs, but they do not define the specific features of MOOCs. In
addition, they are limited to the modeling level with Ontology without moving deeper
into the modeling and implementation of a repository.

Linked Data Repository. The authors of [18] provide a semantic Web-based solu-
tion with an ontology of concepts and pedagogical skills, a repository of semantic
educational resource annotations, and a Web service based on SPARQL. They high-
light the implementation of the services provided by the current solution and others.
Their approach covers both phases: the proposal for ontology and the creation of a
Linked Data repository. However, it does not cover MOOCs. Similarly, in order to
meet the needs of learners or teachers belonging to various universities, the authors of
[19] propose a distributed warehouse model that enables the indexing of educational
objects based on the principle of ontology. They propose an architecture, based on
a classic relational database, of a distributed educational warehouse. They are inter-
ested in university-level educational services (digital or not). This approach is costly
because of its focus on a process of indexing and semantic annotation at each univer-
sity’s level. Likewise, researchers in [20] provide a data repository of books based
on LOM, Dublin Core and IMS, which are cut up and then annotated. For their part,
authors in the paper [21] propose a model of data repository for OERs but without
ontology or Linked Data being used. Similarly, the authors of the paper [22] do not
take use of ontology or Linked data to offer their data warehouses. Their work is about
classifying MOOCs by directly attacking the data extracted through web scraping
according to the learning style. In [23], a Data Warehouse is also suggested. The
automatic generation of SPARQL Build queries is used, with unified target vocabu-
lary terms defined in the Build clause, which are then sent to run on various SPARQL
endpoint sources. RDF graphs with a single vocabulary are the retrieved data.
There are several limitations to the works above. First, some of them do not cover
the MOOC domain. Second, while some propose a MOOC ontology, they do not
MOOCs Semantic Interoperability … 625

cover the concepts necessary for our research, such as pedagogical model or learning
objectives. Third, most of them do not go as far as proposing an architecture for a
data repository. Fourth, in setting up the data repository, they do not rely on Linked
Data.

3 MOOCs Semantic Integration

3.1 Proposed Approach

We propose an ontology designed to capture the broad scope of MOOCs and unify
their representations. This ontology will allow us to build a MOOC metadata repos-
itory based on linked data principles (see Fig. 1). This repository will serve as the
basis for our recommendation system [2] and may be used by other researches.
The contribution of this approach lies on three folds. First, it relies on the Bloom’s
Taxonomy to propose an ontology enriched with metadata for the characterization
of the pedagogical dimension; second, our ontology will serve as a standard to unify
the representation of MOOCS and can facilitate interoperability between MOOCs
platforms. Third, it provides an automated collection of metadata from MOOCs and
builds a repository accessible through a SPARQL endpoint that can be used for other
purposes.
We start with the search for a theoretical framework that will serve as the basis
for our approach. Later, we propose an ontology designed to capture the wide scope
of MOOCs and to unify their representations. In order to clarify and concretize the
epistemological position of the pedagogical dimension of MOOCs, we perform a
mapping between the learning objectives and our theoretical framework.

3.2 Theoretical Foundation for the Pedagogical

Classification of MOOCs

There is a multitude of theoretical frameworks that are applied in evaluation and

description of MOOCs. Conole [24] links the quality of learning to “Good Learn-
ing”. In his opinion, in order to achieve effective learning, it is essential to meet the
characteristics of good learning [24]. Conole uses these principles for the develop-
ment of the 12-dimensional Assessment Framework [25], as well as the “7Cs for
Learning Design” Framework [26]. For its part, Merril [27] reveals that “Merrill’s
First Principles of Instruction” are the basis of all contemporary pedagogical models
and theories. Merill offers five principles. Margaryan [18] builds on Merill’s “First
Principles of Instruction” for the learning activities part and adds five other princi-
ples for the learning resources part. These two Theoretical Frameworks discussed
626 H. Sebbaq and N. E. Faddouli

Fig. 1 The linked data repository framework

above guided the research of Conole and Magrayan in the development of Evalua-
tion Frameworks. This evaluation is based on open questions and often requires the
intervention of an expert. For their part, [29] adopts the “web-based online instruc-
tion” theory, which is more general and integrates three global dimensions of design
elements: information, instruction and learning.
The aim of our research is the large-scale evaluation and classification of the
pedagogical dimension of MOOC on the basis of its learning objectives. Bloom’s
Taxonomy is the most appropriate one for our context since it covers the different
levels of cognitive learning and allows classifying the learning objectives on its six
hierarchical levels [31]. The initial purpose of Bloom’s taxonomy is to assist teachers
in the development of rubrics and in measuring achievement of their learning goals
by providing guidelines [30, 31]. The research paper [31] proposes a revision of
the original taxonomy. It defines a two-dimensional framework: Knowledge and
cognitive processes. The first dimension takes the subcategories of the first level of
the original taxonomy; the second dimension takes the six levels and renames them
in verb: remembering, understanding, applying, analyzing, evaluating, and creating.
For our research, we take into account the dimension of the cognitive learning
process but we reduce it to three levels. We used a modification of Bloom’s Taxonomy
adapted from [32]. The following Table 1 gives a summary of this adaptation.
MOOCs Semantic Interoperability … 627

Table 1 Adaptation of Bloom’s Taxonomy: the cognitive process dimension

Levels of thinking Process verbs Level description
Level 1 Remembering-Understanding Explain, list, Focus on recall or
describe, recall, memory or reflect a
define, identify, show, literal understanding
restate, summarize,
list, demonstrate,
illustrate, explain
Level 2 Applying-Analyzing Organize, classify, Enabling learners to use
relate, prioritize, what they have learned
discuss, group, in a different way or to
model, apply, break down knowledge
compare, contrast, into its rudimentary
distinguish, component
categorize, take apart,
combine
Level 3 Evaluating-Creating Extend, design, Combine elements of
reconstruct, learning into a new
reorganize, create, whole that assess the
develop, speculate, value of particular ideas
propose, predict, or solutions
generate, interpret,
judge, justify,
critique, evaluate, use
criteria, dispute

4 Implementation

The first module of the architecture of our Linked Data Repository is “Ontology
Design and Modeling (see Fig. 1). This module consists of the modeling and devel-
opment of a semantic ontology. The two following modules are “Data Extraction”
and “Data Integration”. The first module seeks to extract raw data from MOOC
providers, while the second one aims to convert and integrate this raw data according
to our ontology. Finally, we use the Apache Jena Fuseki Server to build and set up a
SPARQL Endpoint, and execute SPARQL queries on the created RDF data.

4.1 Ontology Design and Modeling

For the design of ontologies, there is a variability of methodologies. After a review

of these methodologies, we chose ontology Development 101 [33], due to its consis-
tency and acceptance in most R&D projects. For the design, we used the “Pro-
tégé” tool (see Fig. 2). In order to model a semantic data structure that can cover as
much data as possible from different data sources; we conducted a study on MOOCs
metadata from various providers.
628 H. Sebbaq and N. E. Faddouli

Fig. 2 MOOC Ontology representation with OntoGraf

4.2 Data Extraction and Integration

The data is collected from the two providers edX and Coursera. Coursera provides a
RESTful API for accessing its catalog via HTTP GET. This API exposes all Coursera
courses. It is publicly available and does not require authentication. The data is
retrieved in JSON format. Coursera also provides various optional parameters that
can be included in the API request. For its part, edX does not provide an API for
accessing its catalog. To extract data from edX, we opt for Web Scraping. Several
types of libraries and Framework allow us to do Web Scraping. In particular Scrapy
[34], Selenium [35], BeautifulSoup [36], to name only the most famous. For our
work, we use Scrapy. Scrapy is an open source and collaborative framework for
performing requests, scraping and saving data on the web, it is sufficient on its own
to build a web scraping project. Scrapy’s functionality can be extended by importing
other libraries. For our project, we use also splash [37] for the extraction of pages,
which contains JavaScript, and selenium for navigation between web pages.
The “Data Integration” module takes JSON files as input and converts them to an
equivalent Java object using the “gson” library. We use after the JENA Framework
[38] to perform an RDF mapping.

4.3 SPARQL Endpoint

After we convert the raw data to the appropriate RDF format, the data resides in
a stable location where SPARQL queries [39] will be made. For this purpose we
use the Apache Jena Fuseki server which is an RDF store as well as a SPARQL.
There are several modes in which the Apache Fuseki server can operate, either as an
operating system service, as a Java web application, or even as a stand-alone server
[40]. Apache Jena Fuseki Server provides an intuitive graphical user interface in
which the user can upload multiple files to which SPARQL queries are made (see
Fig. 3).
MOOCs Semantic Interoperability … 629

Fig. 3 SPARQL Endpoint to request RDF repository

5 Conclusion

In this research paper, we tackle the problem of the interoperability of various educa-
tional resource repositories, especially in the domain of MOOCs. To resolve this issue
we propose, in one hand, an ontology that unifies the representation of MOOCs and
allows the interoperability between the different providers despite the heterogeneity
of their data schema. On the other hand, we implement a Linked Data repository
based on the ontology constructed as well as the semantic web technologies. This
repository will serve as the basis for our recommendation framework.
In order to concretize the epistemological position of the pedagogical dimension
of MOOC, we carried out a mapping between MOOC’s Learning Objectives and
Bloom’s Taxonomy levels. However, one of the limitations of our current research
work is that the MOOC classification remains manual, which is tedious given its large
scale aspect. In a future work, our proposal will be the automation of the pedagogical
classification of MOOCs.

References

1. The report by class central. https://www.classcentral.com/report/mooc-stats-pandemic/

2. Sebbaq H, El Faddouli N (2020) Recommender system to support MOOCs teachers: framework
based on ontology and linked data. In: SITA conference (SITA 2020). https://doi.org/10.1145/
3419604.3419619
3. IEEE. Institute of Electrical and Electronics Engineers. https://www.ieee.org/portal/site
4. Bourda Y (2004) Pourquoi indexer les ressources pédagogiques numériques?, p 4
630 H. Sebbaq and N. E. Faddouli

5. The Dublin Core Metadata Initiative (DCMI). https://dublincore.org/

6. ISO 15836:2009 Information and documentation - The Dublin Core metadata element set
7. https://standards.ieee.org/standard/1484_12_1-2002.html
8. Bourda Y, Gauthier G (2010) Métadonnées pour ressources d’apprentissage (MLR): nouvelle
norme ISO de description de ressources pédagogiques, p 12
9. Parent M-A (n.d.) Arrimage avec Learning Resource Metadata Initiative (LRMI), p 7
10. Barker P, Campbell LM (2010) Metadata for learning materials: an overview of existing
standards and current developments. Technol Instr Cognition Learn 7(3–4):225–243
11. Bennett M, Baclawski K (2017) The role of ontologies in linked data, big data and semantic
web applications. Appl Ontol 12(3–4):189–194
12. Kagemann S, Bansal S (2015) MOOCLink: building and utilizing linked data from massive
open online courses. In: Proceedings of the 2015 IEEE 9th international conference on semantic
computing (IEEE ICSC 2015), Anaheim, CA, USA, February 2015, pp 373–380. https://doi.
org/10.1109/ICOSC.2015.7050836
13. Rabahallah K, Mahdaoui L, Azouaou F (2018) MOOCs recommender system using ontology
and memory-based collaborative filtering. In: Proceedings of the 20th international conference
on enterprise information systems, Funchal, Madeira, Portugal, pp 635–641. https://doi.org/
10.5220/0006786006350641
14. Volchek D, Romanov A, Mouromtsev D (2017) Towards the semantic MOOC: extracting,
enriching and interlinking E-learning data in open edX platform. In: Różewski P, Lange C
(eds) Knowledge engineering and semantic web, vol 786. Springer, Cham, pp 295–305
15. Poulakakis Y, Vassilakis K, Kalogiannakis M, Panagiotakis S (2017) Ontological modeling
of educational resources: a proposed implementation for Greek schools. Educ Inf Technol
22(4):1737–1755. https://doi.org/10.1007/s10639-016-9511-z
16. Veron S, Celeste V, Alejandra AM, Maria de los Milagros G (2016) An interoperability model
based on ontologies for learning object repositories. In: 2016 international symposium on
computers in education (SIIE), Salamanca, September 2016, pp 1–6. https://doi.org/10.1109/
SIIE.2016.7751843
17. Faqihi B, Daoudi N, Hilal I, Ajhoun R (2020) Pedagogical resources indexation based
on ontology in intelligent recommendation system for contents production in d-learning
environment. J Comput Sci 16(7):936–949. https://doi.org/10.3844/jcssp.2020.936.949
18. Pelap GF, Zucker CF, Gandon F, Polese L (2018) Web semantic technologies in web based
educational system integration. In: International conference on web information systems and
technologies. Springer, Cham, pp 170–194
19. Korchi A (2021) Ontology for distributed architecture of pedagogical warehouse in the field
of machine learning. Educ Inf Technol 12
20. Buffa M, Dehors S, Faron-Zucker C, Sander P (2005) Towards a corporate semantic web
approach in designing learning systems. Review of the TRIAL SOLUTION project, p 5
21. Nahhas S, Bamasag O, Khemakhem M, Bajnaid N (2018) Linked data approach to mutually
enrich traditional education resources with global open education, p 8
22. Aryal S, Porawagama AS, Hasith MGS, Thoradeniya SC, Kodagoda N, Suriyawansa K (2019)
MoocRec: learning styles-oriented MOOC recommender and search engine. In: 2019 IEEE
global engineering education conference (EDUCON), Dubai, United Arab Emirates, pp 1167–
1172. https://doi.org/10.1109/EDUCON.2019.8725079
23. Moaawad MR, Mokhtar HMO (2017) On-the-fly academic linked data integration, p 10
24. Conole G (2016) MOOCs as disruptive technologies: strategies for enhancing the learner
experience and quality of MOOCs. RED, no 50. https://doi.org/10.6018/red/50/2
25. Doherty I, Harbutt D, Sharma N (2015) Designing and developing a MOOC. Med Sci Educ
25(2):177–181. https://doi.org/10.1007/s40670-015-0123-9
26. Conole G (2014) The 7Cs of learning design – a new approach to rethinking design practice,
p8
27. Merrill MD (2012) First principles of instruction. Wiley, Hoboken
28. Collis B, Margaryan A (2014) Design criteria for work-based learning: Merrill’s First M. Bali.
MOOC pedagogy: gleaning good practice from existing MOOCs, vol 10, no 1, p 13
MOOCs Semantic Interoperability … 631

29. Principles of Instruction expanded (2005) vol 36, no 5, p 14

30. Xing W (2018) Exploring the influences of MOOC design features on student performance
and persistence, p 17
31. Bloom BS et al (1956) Taxonomy of educational objectives. Vol. 1: cognitive domain, vol 20.
McKay, New York, p 24
32. Krathwohl DR (2002) A revision of bloom’s taxonomy: an overview. Theory Into Pract
41(4):212–218. https://doi.org/10.1207/s15430421tip4104_2
33. Noy NF, McGuinness DL (2001) Ontology development 101: a guide to creating your first
ontology, p 25
34. Scrapy Homepage. https://scrapy.org/
35. Selenium Homepage. https://www.selenium.dev/
36. BeautifulSoup. https://www.crummy.com/software/BeautifulSoup/bs4/doc/
37. Splash. https://splash.readthedocs.io/en/stable/faq.html
38. Apache Jena Framework. https://jena.apache.org/
39. SPARQL Query Language for RDF. https://www.w3.org/TR/rdf-sparql-query/
40. Apache Jena Fuseki. https://jena.apache.org/documentation/fuseki2/
IBN TOFAIL UNIVERSITY: From
Classical University to Smart University

Khawla Mazwa and Tomader Mazri

Abstract The combination of Internet of Things (IoT), Wifi network and other
advanced technologies have enabled universities to transform from classical envi-
ronment into smart, digital and connected universities. Modern universities create
new way for students to learn, simplify operations for university administrators,
provide a safer and a secure environment, improve environmental sustainability, and
more. This research describes smart university as a new concept and focuses on how
IBN TOFAIL University can become a smart environment.

Keywords The Internet of Things · Wifi network · Advanced technologies ·

Connected universities · Smart environment

1 Introduction

The emerging technologies like: Internet of Things, Big Data and the Cloud
Computing changed the way we live our lives as the information is always available
anytime and everywhere. Also this advancement has many impacts on the world
around us.
University can be as large as cities. The main idea is to transform a classical
university into a smart university based on the concepts and experiences of smart
cities. In fact, smart university is similar to a smaller version of smart city, it can be
a physical or digital environment in which humans and technologies interact with
each other to enable and create new experiences or services for students, professors,
and employees also to improve operational efficiency.
Recently the number of smart universities is growing faster than expected which
can provide a better academic environment for the students. Around the world,
different universities have earned the ‘smart’ tag. In fact, Stanford University, Curtin
University, Arizona State University and the University of San Francisco are just a
few examples of universities which adopted smart university technologies to remain

K. Mazwa (B) · T. Mazri

Advanced Systems Engineering, Ibn Tofail Science University, Kenitra, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 633
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_57
634 K. Mazwa and T. Mazri

competitive in the higher education landscape [1]. The purpose of this research is to
transform IBN TOFAIL University from a traditional into smart university based on
the smart city concept.
In this regard, this paper is divided to seven sections. In Sect. 2 and 3, we summa-
rized the main concepts of smart city and smart university. Section 4, describes our
vision to transform the University of IBN TOFAIL to a smart university. Section 5,
presents the methodology used in our vision. For Sect. 6, we identified the main
systems could be involved in our smart university. In Sect. 7, we discussed the chal-
lenges and issues that could face our smart university. Finally, the conclusion is in
Sect. 8.

2 Smart Cities Concept

Smart city has emerged as an important concept at the end of the year 2000. It
has gained huge amount of attention from professionals, academics, and researchers
from multiple disciplines. A global phenomenon whose success is indisputable, the
smart city remains enigmatic without precise definition. For Smart city concept,
the researchers propose many definitions and multitude forms for example “Digital
City”. Four important axes are included in those definitions [2]:
• Managing the city’s features by using the Information and Communication
Technologies (ICT).
• Creating for the residents a high quality of life.
• The central role of citizens by making cities based on citizens needs.
• Smart cities develop new efficient services in numerous areas, for example
informatics, urbanization or sociology.
Actually, the smart city and the sustainability are now side by side. Based on the
analysis and according to the International Telecommunication Union Focus Group
on Smart Sustainable Cities, the most suitable definition for a smart sustainable city
is:
“A smart sustainable city is an innovative city that uses information and commu-
nication technologies (ICTs) and other means to improve quality of life, efficiency
of urban operation and services, and competitiveness, while ensuring that it meets
the needs of present and future generations with respect to economic, social and
environmental aspects” [3].
This definition highlights the interdisciplinary areas of the smart cities concept,
which present the city as a set of services to provide better solutions while improving
quality of life.
IBN TOFAIL UNIVERSITY: From Classical University … 635

3 Smart University Concept

A smart environment is an environment that is able to acquire and apply knowledge

about the environment and its inhabitants in order to improve their experience in that
environment [4].
In general, every university is connected to the internet and many objects can
be converted into smart objects. For example, doors, windows, projectors, etc. as
simple objects or complex objects like library, classrooms, parking, etc. They can
be attached to sensors by giving a significant level of intelligence using the data
generated from the real world to allow operation of actors to make decisions and
even to change the state of the environment.
In addition, one of the most important aspects in the development of the smart
university is the infrastructure. The infrastructure determines the capacity of univer-
sity to handle a vast range of connected devices. To implement a smart univer-
sity, it needs to design a digital infrastructure that provides beneficial services for
surrounding people.
With the combination of all these elements, universities focus on three main keys:
• Intuitive: today’s students are used to advanced technologies and to the intuitive
platforms. By building a smart university where they can interact with a system
that is intuitive and simple to use.
• Flexible: by implementing the most effective technology that has the greatest
chance of improving the smart university which provides multiple services
depending on the area or the user request.
• Scalable: a smart university can easily modernize while remaining sustainable and
it should allow for global scalability by leveraging digital tools and technologies
to provide a connected student experience.

4 UIT Smart University

The University of IBN TOFAIL is a generalist university that was established in 1989.
It includes a multiple institutions and several branches in different areas such as liter-
ature, humanities sciences, law, economics, sciences, education, communications,
business, management, computer science and telecommunications. The university
has a variety of faculties and engineering schools. Its employees exceed 200, 490
academic staff and it has over 33000 students [5].
As part of the preservation of the environment, the university is strongly involved
in environmental issues and among its objectives are to support the learning perfor-
mance of every student and to promote the research and innovation field. Our univer-
sity has already many applications in place that require being equipped through IoT
devices and effective systems that can reach the goals of smart university. Smart
universities are designed by combining different technologies like IoT, RFID…etc.,
in order to deliver teaching techniques and research strategy.
636 K. Mazwa and T. Mazri

Fig. 1 Important areas in UIT Smart University

In this context, the main goal is to implement Smart City concepts to the univer-
sity in order to transform the University of IBN TOFAIL into a reference in envi-
ronmental sustainability. Three important areas are required to reach this goal (see
Fig. 1): Conservation and construction, sustainability and application of advanced
technologies. These three keys create a smart, sustainable, social, and comfortable
university [6].

5 Methodology

Similar to smart cities, many application categories must be interpreted in our smart
university but the main key of this work is ICTs (Information and Communications
Technologies), without substantial advances in ICT our university will continue to
be inefficient and perform poorly [7].
Through the power of ICTs, applying technology is an inclusive mechanism to
control and manage each and every category with various aspects of action:
• Include the measuring equipments like surveillance cameras, sensors…etc. in
every part of the university.
• Connecting the data with the cloud computing applications by using telecommu-
nication networks and protocols.
• Analyzing and monitoring all the collected data by artificial intelligence and
big-data techniques.
• Usage of actuators which have the task of carrying out physical actions using the
AI algorithms.
Supporting every task of the UIT smart university must be provided with the
application of ICTs.
IBN TOFAIL UNIVERSITY: From Classical University … 637

6 Systems Involved in UIT Smart University

Every Smart University integrates electronic networks and communication infras-

tructure with connectivity services by taking advantage of new technologies such
as: ICT, big data, artificial intelligence and IoT. Every person and object whether it
is application or device on the university must share the same digital infrastructure
where they can depend on each other to enable experiences that weren’t possible
before.
There are plenty systems and smart applications can be involved in UIT where
IoT can be utilized to dramatically change our university to a smart one such as (see
Fig. 2):

6.1 Smart Classrooms

Smart classroom can provide a better learning and teaching opportunities with the
adoption of new high-end digital technology and smart systems. It can be comprised
with [8, 9]:

ID Cards: Those cards can identify students, professors, administration, staff and
visitors using RFID tags to control and to monitor who is inside the building at all
times.

Temperature and Environment Sensors: It can monitor temperatures and send

alerts to maintenance if the temperature gets too high or low.

Fig. 2 Systems involved in UIT Smart

638 K. Mazwa and T. Mazri

Smart HVAC Systems (Heating, Ventilation, and Air Conditioning): It can

provide better comfort and air quality and also it can help reduce the energy use
by functioning only when needed.

Smart Attendance (Automated Attendance Tracking Systems): It can monitor

student attendance and activity during testing and class work.

Smart Lighting: The system can help staffs control and personalize light settings by
using a Smartphone application. Or by using occupancy sensors that detect people’s
presence indoor within a specific area to turn lights on or off automatically to save
energy and improve safety.

Interactive Whiteboards: It can help teachers deliver more effective lectures and
it can optimize learning and engage more students into the education process. Those
smart boards can also record everything written.

Smart Door Locks: IoT devices can be installed on door locks to control who has
access to the room.

Smart Windows: It can offer a high level of control over light by monitoring the
amount of light that enters the classroom.

Cameras: By deploying smart cameras, it can control the security systems that can
track individuals and facial recognition applications.

6.2 Smart Library

The smart library system is based on the Internet of Things (IoT) and Radio Frequency
Identification (RFID). The books and other assets from the library are adopted with
tag, RFID can improve the efficiency of tracking and monitoring of the books like
book circulation and management of reservations…etc. [10].
The smart library system can be combined with an Android application that is
created and is installed in all the student and staffs phone. This application can inform
the students about the books available in the library, question papers and other books,
magazine information or the status about the borrowed books [11].

6.3 Fire Detection Systems

Fire detecting system is needed in UIB to protect lives by identifying a developing fire
automatically to alert the university responsible and fire emergency organizations.
Some of the components used for fire detection systems are [12]:
IBN TOFAIL UNIVERSITY: From Classical University … 639

Flame Sensor: is a sensor which is used to recognize the presence of fire or flame.
It will generate flame signal when fire is caused.

Gas Sensor: is a sensor which will detect different types of gas. They are used to
identify gases like carbon dioxide, carbon monoxide and many more.

Temperature Sensor: is a sensor which will measure the temperature of its

environment and send the data to the processing unit.

Microcontrollers: Microcontroller system has middleware that sends the details to

system and check if a fire is generated or not.

6.4 Smart Parking

Smart parking management provides information related to the available parking

lot, the usage, restrictions and data regarding vehicles in the parking zone. This
smart parking system designed to install sensors and using a Smartphone application.
It includes data regarding the occupation condition of parking lots, directions and
distance produced through parking sensors. Each feature in the mobile application
provide a real time information [13, 14].

6.5 Smart Waste Management System

A university produces a lot of waste that should be managed properly at regular inter-
vals so that the university can be clean and hygienic. The Dustbins in the smart waste
management system can comprise sensors measuring fill level and a communication
system that transfers this data to the garbage trucks or waste disposal unit with the
location and status of each dustbin. Then, a request is automatically generated for
waste collection [15].

7 Challenges and Issues

The smart universities in many regards are considered as small cities. In this paper,
we identified many kinds of systems and technologies that can transform UIT from
classical to smart university. In facts, our UIT is in the process of adopting innovative
techniques for creating a smart environment.
With the development of technologies and within a modern society, our smart
university will confront many challenges among them the technical obstacle.
640 K. Mazwa and T. Mazri

Many of the technologies discussed above are not necessarily new but they will
just be used in new ways. However, the security and the privacy are still big issues.
Below, we list some challenges and issues that can face our UIT [16]:

Security Attacks: The data from every connected object and device in our smart
university are vulnerable to security threats such as gossip attack, automated invasion
attack…etc. However, strong security mechanism is needed that help defend from
cyber attacks.

Network Congestion: Due to the important number of connected objects, defi-

nitely network congestion is occurred in data transmission. So, proper solutions and
measures must be considered to improve the circulation of information and remove
congestion.

Security for Hardware: Certainly, there is a high probability of external intrusion

against the connected devices and sensors. So, every object connected in our smart
university need to be protected properly.

Quality of Service: In order to transfer a huge amount of data for various services,
it is possible that we discover lack of quality of services. Therefore, quality of
service features must be implemented to ensure that even in the critical situations
the applications can still function to provide better services to our smart university.

8 Conclusion

Nowadays several concepts are proposed such as Big Data, IoT (Internet of Things),
and the Cloud Computing. All of them gained greater emphasis and our University
of IBN TOFAIL is a great place where all of these areas could be implemented.
In this paper, we investigated the different systems that could be involved in our
UIT which allows for greater security, more interactivity and better learning. In
addition, we give insights on the challenges that could face our UIT smart university
from becoming reality.

References

1. Nuzzaci A, La Vecchia L (2012) A smart university for a smart city. IJDLDC 3(4):16–32.
https://doi.org/10.4018/jdldc.2012100102
2. Dameri RP, Cocchia A (2013) Smart city and digital city: twenty years of terminology evolution,
p8
3. Ahvenniemi H, Huovila A, Pinto-Seppä I, Airaksinen M (2017) What are the differences
between sustainable and smart cities? Cities 60:234–245. https://doi.org/10.1016/j.cities.2016.
09.009
IBN TOFAIL UNIVERSITY: From Classical University … 641

4. Cook DJ, Das SK (2007) How smart are our environments? An updated look at the state of the
art. Pervasive Mob Comput 3(2):53–73. https://doi.org/10.1016/j.pmcj.2006.12.001
5. www.uit.ac.ma. https://www.uit.ac.ma/. Accessed 15 Nov 2020
6. Min-Allah N, Alrashed S (2020) Smart campus—a sketch. Sustain Cities Soc 59:102231.
https://doi.org/10.1016/j.scs.2020.102231
7. ur Rahman M, Himanshi, Deep V, Rahman S (2016) ICT and internet of things for creating smart
learning environment for students at education institutes in India. In: 2016 6th international
conference - cloud system and big data engineering (confluence), Noida, India, January 2016,
pp 701–704. https://doi.org/10.1109/CONFLUENCE.2016.7508209
8. Temkar R, Gupte M, Kalgaonkar S (2016) Internet of things for smart classrooms. Int Res J
Eng Technol 3(7):203–207
9. Bautista G, Borges F (2013) Smart classrooms: innovation in formal learning spaces to
transform learning experiences. Bull IEEE Tech Committee Learn Technol 15(3):18–21
10. Brian ALA, Arockiam L, Malarchelvi PDSK (2014) An IoT based secured smart library
system with NFC based book tracking. Int J Emerg Technol Comput Sci Electron (IJETCSE)
11(5):18–21
11. Pandey J, Kazmi SIA, Hayat MS, Ahmed I (2017) A study on implementation of smart library
systems using IoT. In: 2017 international conference on infocom technologies and unmanned
systems (trends and future directions) (ICTUS), Dubai, December 2017, pp 193–197. https://
doi.org/10.1109/ICTUS.2017.8286003
12. Sowah RA, Ofoli AR, Krakani SN, Fiawoo SY (2017) Hardware design and web-based
communication modules of a real-time multisensor fire detection and notification system using
fuzzy logic. IEEE Trans Ind Applicat 53(1):559–566. https://doi.org/10.1109/TIA.2016.261
3075
13. Vieira A et al (2019) Smart campus parking – parking made easy. In: Rodrigues JMF, Cardoso
PJS, Monteiro J, Lam R, Krzhizhanovskaya VV, Lees MH, Dongarra JJ, Sloot PMA (eds)
Computational science – ICCS 2019, vol 11540. Springer, Cham, pp 70–83
14. Bagula A, Castelli L, Zennaro M (2015) On the design of smart parking networks in the smart
cities: an optimal sensor placement model. Sensors 15(7):15443–15467. https://doi.org/10.
3390/s150715443
15. Popescu DE, Bungau C, Prada M, Domuta C, Bungau S, Tit DM (2016) Waste management
strategy at a public university in smart city context. J Environ Prot Ecol 17(3):1011–1020
16. Visvizi A, Lytras MD (eds) (2019) Smart cities: issues and challenges: mapping political, social
and economic risks and threats. Elsevier, Amsterdam; Oxford; Cambridge
Collaborative Metacognitive Technique
in Smart Tutoring System

Hibbi Fatima-Zohra , Abdoun Otman , and Haimoudi Elkhatir

Abstract The smart tutoring system is a system that helps learners to solve their exer-
cises. To improve this system, an integration of collaborative technique is required.
This tool, aid learners to understand very well their courses and resolve the difficul-
ties of case studies that are based on team work. The aim of this paper is to fix the
problem of the implicit tutor in the smart tutoring system and propose the integra-
tion of collaborative metacognitive technique as a solution. The process starts with a
collaborative design phase and finished with the implementation phase using activate
storyline 360 environment.

Keywords Smart tutoring system · Collaborative metacognitive · Learning ·

Learning process

1 Introduction

Online courses can meet many needs such as group work, solving problem skills, etc.
The advantage of this type of training is that it can train as many learners as possible
in the entire world at their own place, to promote collaboration. A lot of researchers
suggest that the collaborative learning can improve the teaching [1]. For example,
researchers had detected that collaborative learning and the quality of learning of the
experimental group was considerably higher than that of the monitoring group [2].
Nevertheless, some learners don’t want to rely on to any group or contribute their
works [3]. Others may feel unjustly exploited by members of the group and may give
up on their efforts to them because of the sucking effect [4].
These problems are due to the fact that there is no appropriate tool to supervise and
quantify the learner’s process in collaborative learning. The researchers suggested
different contributions to promote the engagement of learners from different perspec-
tives [5]. Some researchers are concentrated on the feedback model of the smart
tutoring system using analytical mechanisms to take into consideration the level and

H. Fatima-Zohra (B) · A. Otman · H. Elkhatir

Laboratory of Advanced Sciences and Technologies Polydisciplinary Faculty, Abdelmalek
Essaadi, Larache, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 643
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_58
644 H. Fatima-Zohra et al.

previous knowledge of the learners [6]. Other researchers had focused on the engage-
ment of learners in teaching groups [7]. To augment the efficiency of the learning
process, they [8] proposed an adaptation of learning sequence using a genetic algo-
rithm to optimize a list of exercises in a training course. However, with the emergence
of these new teaching and learning technologies due to the pandemic of coronavirus;
the changes brought by new digital techniques are particularly relevant with regard
of the content and the quality. In this regard, the educational mediatization raises
many issues [9]. The objective of this paper is to describe a conceptual solution for
the computer modelling of educational knowledge, which should be according to the
current learning. The goal is to improve collaborative learning using metacognitive
techniques into the intelligent tutoring system. The paper is described as follows:
after presenting the background of the topic, the authors will define the concept
of collaborative learning and its main benefits. After that, they will describe the
proposed process of integrating Collaborative metacognition technology in the intel-
ligent tutoring system. Finally, they will present the implementation phase of the
proposed process.

2 Background

In this section, the authors will present: some techniques related to this work; the
collaborative design phase and inquiry based learning.

2.1 Collaborative Design Phase

The collaborative design phase is a collaborative content design process, similar to

the software engineering. The design is based on a life cycle with several advantages
such as: promotes traceability, improves visibility over the entire life cycle. In a view
of its advantages, the choice was made for the model known as “V-Cycle”. The Fig. 1
illustrates the steps of the collaborative content design process.

2.2 Inquiry-Based Learning

Inquiry-based learning (IBL) is designed as a problem-solving approach, using

several procedures inherent in this strategy. According to [10] and [11]; the research
step is when curiosity transforms into action, to answer research questions or to eval-
uate research results assumption. It was divided into three sub-phases: data explo-
ration, experimentation and interpretation. IBL is designed as a problem-solving
approach, using several procedures inherent in this strategy. According to [10] and
[11].
Collaborative Metacognitive Technique in Smart Tutoring System 645

Fig. 1 Life cycle of collaborative design process [10]

Students explore, observe, plan and perform different experiments, modify vari-
able values, make predictions and interpret results. Exploration is a systematic
method for conducting a search with the intention of finding the relationship between
the variables that are involved.
Interpretation of data aims to find the meaning of data collected, by synthesizing
new knowledge. The final result of the research phase is the following interpretation
of the data (discovery of the relationship between the variables) that answers the
main problem.

3 Collaborative Meta-cognitive Techniques

3.1 The Concept

Collaborative learning is a relationship among learners that involves positive inter-

action and personal engagement (a sense of immersion or swimming together). The
collaborative learning is a teaching method in which learners of various degrees can
work together in a small group to achieve a common objective. Learners are in charge
of their learning and their own learning. For instance, a learner achievement helps
other students to success [12]. The main objective of the E-learning platform courses
in these environments was for students to learn from each other and share information.
There are various advantages presented in [13] for collaborative learning.
646 H. Fatima-Zohra et al.

The Metacognition make it possible for people to take care of their own learning.
It is about increasing their sensitivity to how they learn, measuring their learning
needs, building strategies to meet those needs, and then applying those strategies.

3.2 Meta-cognitive Strategies

Meta-cognitive techniques facilitate learning process. The authors can appropriate

these strategies, into e-learning courses, social learning experiences, training activ-
ities and other formal or informal learning experiences. These techniques are listed
as below:
• Ask Question;
• Encourage self-reflection;
• Foster self-questioning;
• Promote self-directed learning;
• Establish access to supervisors;
• Resolve case studies with a team member;
• Personal explanation;
• Offering the chance to make errors.
To summarize, meta-cognition is a collection of skills that make learners to be
aware of how they learn, evaluate and adapt these abilities to become more and more
efficient in their learning. In a world that involves lifelong learning, offering people
new and better meta-cognitive strategies is a donation that can last forever.

4 Description of the Contribution

The existing Intelligent Tutoring System contains three modules: Student module,
Expert module and Instructional module. This later contains two types of tutors
which are: Explicit and implicit tutor. Our goal in this contribution is to improve the
implicit tutor using a specific type of indirect intervention which is: Collaborative
metacognitive techniques; this method involves the conscious monitoring of cogni-
tive strategies to accomplish particular goals. The Fig. 2 present the description of
the contribution.
Collaborative Metacognitive Technique in Smart Tutoring System 647

Fig. 2 The integration of metacognitive strategies into the implicit tutor

4.1 The Process of Integrating Collaborative Meta-cognitive

Technique into Smart Tutoring System

In this section, the authors describes the process of integrating collaborative metacog-
nitive methods into smart tutoring system, and how the inquiry based learning and
conceptual map can be used to improve the interactivity in online collaborative space.
The process illustrated in the Fig. 3 present the indirect interventions used by the
implicit tutor in the instructional module of smart tutoring system. The steps that
were presented in the picture are described as below: The learners connected to an
E-learning platform. After that, they subscribe into the online collaborative space;
the learners who passed a pre-Test, a list of activities are enabled such as: resolving
the study case with their team, group activities, etc. The implicit tutor intervene
using the Collaborative Metacognitive Technique, including inquiry based learning
method and conceptual maps tool.
The Fig. 4 presents the effectiveness of meta-cognitive technique. N indicates
the total gross number of statements in each category. The line is corresponding to
a statement that meets the criteria. Collaborative meta-cognition is marked with an
asterisk [14].
648 H. Fatima-Zohra et al.

Fig. 3 The process of integrating collaborative meta-cognitive technique into smart tutoring system

Fig. 4 The effectiveness of meta-cognitive technique in a collaborative talk type

4.2 Collaborative Design Phase

The collaborative design phase is a process of setting up a pedagogical design. It

makes it possible to design courses or learning systems. It contains four essential
phases: the Pedagogical phase which is the constitution of a meta-cognitive tech-
niques, this later is transformed into basic tools like conceptual Maps and Inquiry
Collaborative Metacognitive Technique in Smart Tutoring System 649

Fig. 5 Collaborative design phase

based learning as described in the didactical phase. After that, the authors build a
storyboard of the proposed approach and then, they develop a solution using LMS
to create the E-Learning platform.
The Collaborative design phase is related to the proposed approach of this contri-
bution and the simulation phase. The essential phases of this collaborative design
illustrated in the Fig. 5 and described on details as below:
• Pedagogical phase: present the appropriate approach used in the contribution;
since the authors will use meta-cognitive strategies, a collaborative workspace is
required.
• Didactical phase: contains a list of tools to apply the pedagogical phase which is
the uses of the meta-cognitive techniques.
• Mediatic phase: In this phase the authors refer to the storyboard to carry out the
pedagogical activities, exercises, graphics, etc. it allows them to test the module
to identify and address any ergonomic problems.
• Technological phase: the aim of this phase is to develop interactive interfaces.

5 Implementation Phase

5.1 Environment: Storyline

Articulate storyline is an authoring tool that allows the creation of interactive courses
and E-learning modules without the necessity of using programming language.
Indeed, this software allows different participant in training to easily produce inter-
active content. This tool contains four versions, but we will describe the three recent
versions that are presented as below:
650 H. Fatima-Zohra et al.

• The first one is: Articulate storyline360: which contain a list of product, some
of them will be downloaded and others will be installed, also it requires a yearly
subscription;
• The second one is: storyline3: it has different states that allow us to modify the
appearance of an object according to the learners’ actions. Storyline3 also provide
an enlarge button when the learner hovers over it or add a highlight effect, and
when he/she click on it. States also make it possible to modify the expression and
attitude of a character;
• The last one is: Storyline2: it offers an improved user interface and interactive
learning element;
All the articulate storyline versions are required Microsoft.Net Framework, Visual
C++ and Adobe Flash Player (10.3). HTML5 and JavaScript are needed also. The
articulate storyline versions contain the same options. However, there is some distinct
point that makes us easy to choose the appropriate version. The Fig. 6 represents a
comparative study for the three versions of articulate storyline to implement our
pro-posed process:

Fig. 6 Comparative study of the three versions (Articulate Storyline)

Collaborative Metacognitive Technique in Smart Tutoring System 651

5.2 Simulation of the Proposed Approach

To implement our contribution, the software architecture used is based on an advanced

technological selection applied in building interactive content. The environment used
in this implementation is articulate storyline360: this tool provides the creation of a
learning content such as: software exhibition, software simulation, a trendy scenario
and generate activities and quizzes. The Fig. 7 present the interface of this tool:

Fig. 7 Interface of Articulate Storyline

Fig. 8 Interactive collaborative space

652 H. Fatima-Zohra et al.

Fig. 9 A case study activity

The learners will subscribed and build a team called “Our Team”. Then, they
will forward to an interactive collaborative space which they should check an option
from the list illustrated in the Fig. 8. After that, if for example the learner selects the
option: group activities; the case study will appear in Fig. 9. This activity contains 26
min, if they could not resolve, or they had a problem. The tutor will intervene with
two metacognitive techniques: inquiry based learning and mind mapping which can
help them to sort out the case study.

6 Conclusion

This paper presents the issues of classical E-Learning platform. The authors under-
took the collaborative learning: definition and their advantages. Then, they presented
the related works, after that, they described the contribution, presented the problems
of the implicit tutor in the smart tutoring system and proposed the solution which is
the integration of the collaborative metacognitive techniques.
The proposed solution can’t be realized without an essential process which is
the collaborative design phase. After the pedagogical design, a simulation of the
proposed approach is presented in this paper. The integration of a smart technique in
collaborative learning in a corporate context will be done with an efficiency result.
However, the practical implementation of these generic recommendations and the
validation of their effectiveness are deferred to future work. These experiments will
be carried out on the e-learning platform of the Larache Polydisciplinary Faculty,
which is part of the Abdelmalek Essaadi University in Morocco.
Collaborative Metacognitive Technique in Smart Tutoring System 653

References

1. Berland M, Davis D, Smith C (2015) AMOEBA: designing for collaboration in computer

science classrooms through live learning analytics. Int J Comput-Supp Collab Learn 10(4):425–
447
2. Chen B, Hwang GH, Lin TS (2020) Impacts of a dynamic grouping strategy on students’
learning effectiveness and experience value in an item bank-based collaborative practice system.
Br J Educ Technol 51:36–52
3. Hall D, Buzwell S (2013) The problem of free-riding in group projects: looking beyond social
loafing as reason for non-contribution. Active Learn Higher Educ 14(1):37–49. https://doi.org/
10.1177/1469787412467123
4. Hutter M, Diehl M (2011) Motivation losses in teamwork: The effects of team diversity and
equity sensitivity on reactions to free-riding. Group Process Intergroup Relat 14:845–856
5. Zepke N, Leach L (2010) Improving student engagement: ten proposals for action. Active
Learn Higher Educ 11(3):167–177
6. Hibbi F-Z, Abdoun O, Haimoudi EK (2019) Exploration of analytical mechanisms in the feed-
back model. Procedia Comput Sci 148:201–207. https://doi.org/10.1016/j.procs.2019.01.037
ISSN 1877-0509
7. Fatima-Zohra H, Otman A, El Khatir H (2019) Smart detector of learner style. In: 5th interna-
tional conference on optimization and applications (ICOA), Kenitra, Morocco, pp 1–4. https://
doi.org/10.1109/ICOA.2019.8727671
8. Popescu E (2014) Providing collaborative learning support with social media in an integrated
environment. World Wide Web 17(2):199–212
9. Hibbi F-Z, Abdoun O, El Khatir H (2021) Coronavirus pandemic in Morocco: measuring the
impact of containment and improving the learning process in higher education. Int J Inf Educ
Technol 11(1):30–34
10. El Mhouti A, Nasseh A, Erradi M (2014) Design and implementation of a socio constructivist
model of collaborative learning design (SMC-LD) dedicated to distance learning. Int J Comput
Sci Eng 2(8):2347–2693
11. Keselman A (2003) Supporting inquiry learning by promoting normative understanding of
multivariable causality. J Res Sci Teach 40(9):898–921
12. Dewey J (2007) How we think. Howwe-Think-Johndewey
13. Banchi H, Bell R (2008) The many levels of inquiry. Sci Child 46(2):26–29
14. Gokhale AA (2011) Collaborative learning enhances critical thinking. J Technol Educ (2011)
Enhancing Education System
with a Q&A Chatbot: A Case Based
on Open edX Platform

Rachid Karra and Abdelali Lasfar

Abstract The emergence of MOOCs (Massive Open Online Course) increased the
need for personalized student support. Question and Answering solution can reduce
the dropout rate and improve learning. In this paper, we explore the potential of using
chatbots to enhance the learning experience for students and how to take advantage
of Deep Learning. A review of chatbots models suggests that task-oriented ones
are more convenient in the educational context. To address this issue, we propose
an integrated chatbot in Open edX platform. We study architecture and proposed
technical options. The model is developed to perform a question-answering task. The
trainings were carried out using TensorFlow and NTLK. The paper also presents the
various steps in the conception and implementation. The chatbot performs well and
contributes to the improvement of the e-learning systems.

Keywords Chatbot · Deep learning · Education system · Human-machine

interchange · NLP

1 Introduction

Morocco has made significant efforts to generalize access to the education system.
Unfortunately, this system suffered from school dropout. According to UNICEF, the
school enrollment rate in rural areas does not exceed 32.4%. This rate drops to 30.6%
for access to high school and does not exceed 21.9% for girls [1]. The reason for
this loss is the distribution of villages over vast areas and the geographic difficulty.
UNICEF recommends the use of innovative technologies in parallel with traditional
training to fight school dropout. Indeed, the rural environment can benefit from the
new e-learning tools.
With the Sars-Cov-2 pandemic, the educational offer is becoming increasingly
restricted. Indeed, the constraints of distance oblige us to adopt new learning styles.

R. Karra (B) · A. Lasfar

LASTIMI Laboratory, Mohammadia School of Engineers, Mohammed V University in Rabat,
Rabat, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 655
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_59
656 R. Karra and A. Lasfar

The classroom is, by definition, a closed space and a rapid transmission vector of
Sars-Cov-2. Adaptation necessarily requires the adoption of modern technologies.
Blended-Learning is one of the effective solutions in the administration of this
pandemic. Shifting between face-to-face in small groups and remote study seems
like a good compromise. This allows one group of students to study in class and the
other at a distance. This alternation yields the possibility of taking advantage of both
methods. Since several universities have put their courses online, the flipped class-
room becomes a desirable choice. The knowledge acquired by the student during the
distance study can be deepened or corrected with the teacher. All these instructional
strategies need improved distance learning tools.
Artificial intelligence is redefining e-learning. From preventing student’s dropout
to aiding and providing information about the university and the courses, there is a
lot of ways how AI can help [2]. Universal Design for Learning (UDL) is a scientific
framework for education. It aims to improve and optimize methods in education. It
is based on how students interact with educational systems and derive from research
in education and neuroscience. It recommends offering more than one format for
learning and multiple ways to interact with content [3].
It is difficult to monitor a hundred students at the same time and offer support
to each of them. One of the options we explored is adding a chat agent that allows
students to get automatic answers.
Whereas many economic sectors have benefited from chatbots capabilities.
The education sector has lagged. Chatbots can supply opportunities to improve
engagement, proximity, and assistance between student and teacher.
Making a chatbot can be done in two ways, the rule-based approach or the base-
ruled approach. The second one makes the system learn from data. One of the
best known is the ALICE system, which is based on Artificial Intelligence Markup
Language (AIML). In recent years, several chatbots have emerged, especially those
based on a natural language process engine (NLP) and hosted in the cloud such as
Cortana from Microsoft or Alexia from Amazon [4].
NLP uses neural networks to understand what humans intend to tell (the inquiry).
Several types of chatbots exist including the question answering bot. Q&A bots are
trained on a knowledge base. However, the knowledge base must be quite substantial
to give reliable results. For NLP to perform well and give satisfactory results the
knowledge base must be large and complete [5]. To implement a chatbot successfully,
we must identify the target audience, its behavior, and the use cases. Knowing the
target audience gives us the decision on how the chatbot responds to question either
in a friendly or professional way. Ideally, the chatbot should be able to engage in a
human discussion according to the chosen topic.
A Q&A system is a combination of NLP model and database knowledge. Q&A
bots make it possible to deal with scenarios never considered and generate responses.
In these chatbots, the synergy between the system and the student is typically a one-
to-one sentence. Some students do not adapt to the transition from high school to
university. Indeed, they do not manage to follow a system of self-education [5]. The
chatbot (Q&A) finds its importance in lectures where the number of students easily
exceeds a hundred and the teacher cannot answer all the questions [6].
Enhancing Education System with a Q&A Chatbot … 657

In Sect. 2, we describe the sequence to sequence model and the encoder-decoder.

In Sect. 3, we describe the integration of the chatbot in our e-learning platform. In
Sect. 4, we discuss the results and how to enhance the chatbot limits.

2 Background

Deep Neural Networks (DNNs) is one of the most efficient machine learning models.
He achieves excellent results in problems as difficult as diverse like the conception
of new drugs or facial recognition [7]. Despite these advantages, DNNs only apply to
issues where inputs and outputs can be encoded to fixed sizes [7]. They accept a vector
of fixed size (question) and give a result of fixed size (answer). This limitation restricts
the use of DNNs, especially with the problems of variable sequences (dimension and
length).
Recurrent Neural Network (RNNs) is a generalization of neural networks with
sequential characteristics. They find nonlinear relationships between sequential data
of input and output. For an input sequence (I 1 , I 2 … I m ) the RNN calculates the
sequences of hidden layers (h1, h2 … hm ) and the output sequence (O1 , O2 … Om )
by iterating the function:
 
h t = tanh W hx It + W hh h t−1 (1)

Ot = Woh h t (2)

With tanh nonlinear function that act as a coordinate-wise and W weight matrix. In
practice, RNNs are confronted to the instability problem due to gradient descent. This
causes the problem of vanishing or exploding gradients. The difficulty of training
RNNs for problems with long-range temporal dependencies almost hampered their
development. As shown in (Fig. 1), Sequence-to-Sequence model (Seq2Seq) is based

W X Y Z <eos>

Encoder

<eos> W X Y Z

A B C Decoder

Fig. 1 Sequence-to-Sequence model

658 R. Karra and A. Lasfar

on encoder-decoder architecture. It is used to translate, resume text, or dialogue

generation. Both the encoder and decoder are RNNs [7].
To overcome these drawbacks, researchers have developed a technique that allows
long-range learning. This method is called the Long-Short Term Memory (LSTM).
The connection weight value of the output of each neuron in the network is returned
as input. In this way, it receives a combination of the added information from other
neurons plus the older version of the output. This gives this type of neuron a memory.
LSTM can have an input length different from the output length [8].
Another property of LSTM is the representation of the input sentence as a vector
of fixed size. It encourages LSTM to find the meaning of the question even if it is
paraphrased [8]. The association of words with the same meaning or usually used
together, helps give a meaningful answer. By its characteristics, the LSTM is more
suited to Natural Language Processing. In no time, it became the state of the art in
the NLP.
The chatbot uses the NLP model and tries to highlight the entities, the utterances,
and the intents. Deep learning adds intelligence to the dialogue between the student
and the chatbot. Nevertheless, the system must preprocess the message by tokenizing
and removing punctuation and stop words. To know the intent in a message we used
the vectorizer [9]. It consists in giving the same representation to the same word.
TF-IDF (Term Frequency—Inverse Document Frequency) and word2vec are among
the most widely used. TF-IDF is an indicator of the weight of the word in a sentence,
typically used in information retrieval and text mining. Each term has its own couple
(TF, IDF) score. The importance of the word is calculated from the standardized
frequency of its appearance in the document [7].

3 Architecture and Description

The Q&A chatbot simulates a conversation between the student and a virtual
instructor. Chatbots operate on a one-to-one logical sequence [10]. The informa-
tion flow is a question followed by an answer. A student enters a question and waits
for the chatbot to give a correct answer.
Open edX project is a web application based on the Django framework for
producing MOOCs. Django powers edx.org and many other leading online education
sites. It is based on two main components, the studio, and the learning management
system. The studio can be considered as a CMS (Content Management System) for
course authoring, scheduling, and grading policy [11]. The LMS is the foreground
of the Open edX project. Learners use it to take courses and exams. The Open
edX system is based on a modular architecture called XBlocks. It makes it possible
to develop new functionalities for educators and learners. The XBlock technology
allows the use of several components from the Open edX platform. The different
applications are grouped so that they can interact with each other and with the main
Enhancing Education System with a Q&A Chatbot … 659

system. XBlocks can be used to represent a single module, web component, multi-
media elements, social learning experiences, or interactive content. Furthermore,
sections or entire courses can be constituted with the XBlocks.
Open edX uses many technologies. A basic XBlock has fields, a student View,
handlers for web services, and the web pages where the results are rendered.
MongoDB is used for course persistence and data analytics [11]. MySQL database
to store user data, settings, and their progress. (Fig. 2) shows a general description
of the Q&A chatbot:
Our Chatbot extension has been included as a new plug-in inside the LMS.
XBlock’s front-end use Bootstrap. It makes our Q&A compatible with mobile
devices. As part of the entire system, we have implemented a Q&A XBlock equipped
with an NLP model capable to retrieve answers and interact with students. This model
uses different tools principally TensorFlow 1.13.1 and NLTK 3.5 for the DNLP capa-
bility. The Q&A solution was trained with GPU support using Nvidia Quadro P3000
6GB. There are different training strategies. [5] trained the chatbot on a first dataset
to give it the language capabilities and on another dataset about insurance so it has
domain-specific knowledge. Our approach is to separate trained models. Each course

Learners

ChatBot

Incoming Activity Turn Context Send Activity Librairies

Deep Natural
Processing and rules parser
XBlock

Language Processing
TensorFlow

Tokenization
Numpy
NLTK

Seq2Seq
Text processing
Encoder Decoder
Entity extraction
RNN-LSTM
Intent classifier

Knowledge Base
Data Source

MongoDB

Fig. 2 Proposed Chatbot architecture and used tools

660 R. Karra and A. Lasfar

training_resource question_answer discussions

_id: Object ID _id: Object ID _id: Object ID
course_id: Integer course_id: Integer text_disc: String
question: String user_id: Integer type_disc: String
answer: String discussions: Array date_disc: Date

Fig. 3 Persistence data model for Q&A chatbot

has its own model [12]. With this choice, we avoid the catastrophic forgetting effect.
The loss of previously gotten knowledge when we try to train our model on more
data. In fact, [4] have shown that the forgetting phenomenon is not reserved for DNN
but also LSTM. Their model was able to perform an accurate output only on the
newly trained sequences. The base knowledge is persisted in MongoDB (already
provided by Open edX platform) as collections of BSON documents. (Fig. 3) shows
the proposed data model for discussions:
Course resources must be formatted following the ‘training_resource’ collection
of four fields _id, course_id, question, answer. The history of discussions between the
bot and the student are stored in the collections question_answering(_id, course_id,
user_id, discussions) and discussions(_id, text_disc, type_disc, date_disc).
To develop the chatbot we used the XBlock-SDK. It is a set of features and rules
to follow. The developed subsystem is compatible with Open edX. The Xblock-SDK
has the same runtime as the whole Open edX solution [10]. These specifications make
it very convenient for development and test purposes. In a production environment,
we can change the XBlock default database from SQLite to MongoDB.

4 Results

Integrating the chatbot with the LMS appears to be a promising solution for improving
the quality and interactivity of the e-learning system. During the tests, we have
asked several questions to our ChatXBlock. The responses were generally satisfying.
The user interface is designed in an intuitive way. It also respects the Open edX
graphic design. Simplicity was one of the pillars of our design. The student gets the
information intuitively.
(Figure 4) shows the process of dialog between the chatbot and a student enrolled
in a course:
However, we found some limitations in the Q&A chatbot. If the student asks badly
or makes a mistake in the question, the chatbot gives an incoherent answer. The
second limitation is how to guide the student to ask the right questions. Especially
if we assume the student course ignorance. The chatbot is useful only in the case of
oversights or revisions. Indeed, it does not help in the first phases of learning where
the student begins to discover the subject, and even worse can be misleading.
Enhancing Education System with a Q&A Chatbot … 661

Fig. 4 Example of question answering of a student enrolled in a course

Human Takeover was added to the system. It means the taking over of the conver-
sation by the instructor if the chatbot does not respond. This possibility is available
when the response of the chatbot does not satisfy the student. He can trigger the
intervention of the instructor [6]. This allows the student to come back to the teacher
if necessary and verify the chatbot answer. The mixture of machine and human
interaction enhances collaborative teaching and gives diversity in learning styles.

5 Conclusion

Several disciplines overlap for the enhancement of the student experience. Students
need feedback and guidance. In this study, we presented the challenges of recent
technologies for education. We have developed an excellent Q&A chatbot by using
deep learning and natural language processing. It fills the lack of proximity and
availability of the instructor in MOOCs.
It remains appropriate to focus our studies on knowledge transfer. It allows adding
blocks of knowledge to a pre-trained model without going through new training each
time. This way is closer to the natural functioning of the human brain and reduces
the duration and cost of each new training. [13] has introduced two parameters to
662 R. Karra and A. Lasfar

control the user’s recommendations according to their preferences and find a balance
between questions on one side and recommendations and their relevance on the other
side. Extending our chatbot with recommendations possibility is another issue that
worth considering. The chatbot can recommend a module or a chapter that matches
a user’s gap based on the questions the student posed.

References

1. UNICEF (2014) Out of school children in Morocco

2. Huang R, Spector JM, Yang J (2019) Educational technology: a primer for the 21st century.
Springer, Singapore
3. Gregori P, Martínez V, Moyano-Fernández JJ (2018) Basic actions to reduce dropout rates in
distance learning. Eval Program Plan 66:48–52
4. Tetko IV, Kůrková V, Karpov P, Theis F (eds) (2019) Artificial neural networks and
machine learning. In: ICANN 2019: deep learning: 28th international conference on artifi-
cial neural networks, Munich, Germany, 17–19 September 2019, proceedings, part ii. Springer
International Publishing, Cham
5. Nuruzzaman M, Hussain OK (2020) IntelliBot: a dialogue-based chatbot for the insurance
industry. Knowl-Based Syst 196:105810
6. Mckie IAS, Narayan B (2019) Enhancing the academic library experience with chatbots: an
exploration of research and implications for practice. J Aust Libr Inf Assoc 68:268–277
7. Sutskever I, Vinyals O, Le QV (2014) Sequence to sequence learning with neural networks.
arXiv:1409.3215 [cs]
8. Sutskever I, Martens J, Hinton G. Generating text with recurrent neural networks, p 8
9. Vinyals O, Le Q (2015) A neural conversational model. arXiv:1506.05869 [cs]
10. Luong M-T, Pham H, Manning CD (2015) Effective approaches to attention-based neural
machine translation. arXiv:1508.04025 [cs]
11. EdX: Overview of EdX Insights
12. Terekhov AV, Montone G, O’Regan JK (2015) Knowledge transfer in deep block-modular
neural networks. In: Wilson SP, Verschure PFMJ, Mura A, Prescott TJ (eds) Biomimetic and
biohybrid systems. Springer International Publishing, Cham, pp 268–279
13. Ikemoto Y, Asawavetvutt V, Kuwabara K, Huang H-H (2019) Tuning a conversation strategy
for interactive recommendations in a chatbot setting. J Inf Telecommun 3:180–195
A Multi-agent and Content-Based
Course Recommender System
for University E-learning Platforms

Meryem Amane, Karima Aissaoui, and Mohammed Berrada

Abstract E-learning has recently become widespread all over the world. All coun-
tries are working more and more to integrate information and communication tech-
nology in their educational systems. In this concern, recommender systems have
played a crucial role for suggesting items that best match user’s preferences in
different area. Currently, in order to prevent the spread of coronavirus pandemic and
to ensure continuity of education during the confinement period likewise our country
Morocco, distance learning has become an imperative necessity. In this regard, we
are going to delve into the role of e learning in achieving quality education and the
difficulties of its implementation in Moroccan context. Several digital courses have
been published on e-learning platforms. However, learners, who do not have enough
experience to choose the courses they need, often face many challenges. In this paper,
we present Multi-agent system as a helpful tool in recommendation, then we review
the current recommendation methods. Moreover, we implement a recommendation
system for e-learning platforms that recommends university courses for students.

Keywords E-learning · Recommender systems · Collaborative filtering ·

Content-based filtering · Hybrid recommender system

1 Introduction

The implementation of e-learning in education has become a necessity because of

its significance in facilitating the process of learning and teaching. Technology has
converted the world into a small global village where we has easy access to all
what is going on over the world including culture, values, news and so forth. In
the education, there has been developed a lot of online programs and platforms
whereby we can study synchronously or asynchronously and wherever we are. Since
its inception, many countries including Morocco, have resorted to distance learning
as a helpful strategy in order to maintain the process of teaching and learning, and

M. Amane (B) · K. Aissaoui · M. Berrada

Artificial Intelligence, Data Science and Emergent Systems Lab Sidi, Mohammed Ben Abdellah
University, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 663
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_60
664 M. Amane et al.

at the same time, they can manage this crisis successfully without blocking the
educational system. Nevertheless, it has so many challenges; particularly, in the time
consuming for searching a good support. Recommendation systems can be one of the
best solution to this problem for learners, it can help them to find their favorite content
from the huge amount of information [1]. Most of the choices that individuals select
are constructed on recommendations or suggestions from others [2]. For example, in
e-commerce, clients need to examine item surveys before buying. Netflix.com uses
the recommendation to assist individuals discover movies and videos that interest
them using collaborative filtering; Youtube.com creates a personalized home page
which displays the videos recommended for each user according to his access history
[3]. In the same way, in the education field, a recommendation system is a great tool
that can help to improve learning. For instance, in university, RS could help students
to find suitable courses quickly as well as to reduce the time to explore the courses
they will take [3]. The learner can have an overview of the course according to the
vote, trends and content of the course, so the analysis of the best course can be carried
out easily. Many online learning platforms have been designed to improve access to
high quality educational content for students connected to Internet [4]. To provide
suggestions to learners who do not have enough experience to choose the courses they
need, researchers have tried to develop various approaches to recommend relevant
learning objects to learners in order to improve the effectiveness of studies. In fact, it
has become ultimately a pre-requisite to implement our recommendation e-learning
system in intelligent way, thanks to the significant role that recommender systems
play in improving teaching and learning quality. In this concern, and in order to
fulfil communication and interaction between users, we are going to use multi-agent
systems to improve the quality of the courses recommendation list, by extracting
positives and negatives comments of each course, and then reclassify the final list.
This method can enhance and recommend the best courses to learners.
This paper is oriented to the use of recommendation systems in online learning
systems, it is organized as follows: In sect. 2, we provide details about the common
used recommendation methods. In Sect. 3, we explain the proposed approaches
by researchers in e-learning recommender system, as well as the recommendation
techniques used in each article. In Sect. 4, we present our proposed course recom-
mender system, using the content-based approach and the multi-agents system, the
description of course and the similarity.

2 Classification of Recommendation Methods

2.1 Collaborative Filtering (CF)

It is one of recommendation methods widely used in several fields it depends on the

presumption that “similar clients have the equivalent preferences” [5]. For instance,
Amazon, the well-known e-commerce website, utilizes its own recommendation
A Multi-agent and Content-Based Course Recommender System … 665

Fig. 1 Course content-based

recommendation

system. At the point when you select a product to purchase, Amazon suggests other
products that different clients bought dependent on that original product [6]. In the
context of e-learning, in a course recommendation system for example, the system
can suggest a course to a learner without needing to know the content of the course,
but only based on the ratings of similar learners.

2.2 Content-Based Recommendation (CBF)

This technique recommends an element to a user based on the description of the

element. In this type, a user is recommended objects similar to those he has favored
in the past. For example, we consider a recommender system that use content-based
method to recommend courses to learners. If the course is about Big Data technologies
and a learner named “karim” enroll on it, the recommender system will recommend
another course to karim, since these courses will include much more vocabulary
related to this topic (for example, “Hadoop”, “HDFS”, “MapReduce”) than articles
on other topics [7] (Fig. 1).

2.3 Hybrid Recommendation Methods

This is another classification of recommendation system, which try to overcome the

constraints of different methodologies talked about above. It is a mix of at least two
diverse recommendation strategies. The most used hybrid approaches are that of
content-based and collaborative filtering. It utilizes both the content of the element
and the ratings of all users [8].
For example, Fig. 3 shows hybrid course recommendation system using the
combination of content-based methods and collaborative filtering (Fig. 2).
666 M. Amane et al.

Fig. 2 A Hybrid course recommender system

3 Related Works

At present, online learning systems represent an interesting area of research in educa-

tion; these systems use different methods to recommend different courses to online
learners, based on their preferences and learning style of other students with similar
profile [9]. E-learning recommender system (ERS) have become the most basic tool
for giving customized recommendations to students in response of the issue of data
overload in web based learning situations. It can suggest a better combination of
subjects that interest learners, the suggestions are created utilizing a lot of rules and
strategies to propose a list of learning materials using the learner’s data concerning
his professional interest and his objectives [10]. A number of online learning recom-
mendation systems (ERS) based on different approaches have been introduced to help
learners to find their specific learning needs. They automatically recommend learning
materials personalized to each learner profile based on their interests (Table 1).

4 Discussion

Most of the approaches proposed above use collaborative filtering as part of their
recommendation systems. This technique can be used to create recommendation
systems that give suggestions to a student based on the interests and preferences of
similar learners, one of the advantages of collaborative filtering is that they do not
need or do not require any knowledge of database elements, the user ratings matrix
is the only entry required. However, it has various drawbacks, such as:
• Cold start problem: It occurs when the recommendation system is unable to
make decisions about users or items, on which enough information has not been
collected. The first time a new user visits a specific recommendation system,
for example, none of the items have been rated. Therefore, the recommendation
system does not know what are the likes and dislikes of this user. The same problem
occurs when a new item is introduced to the recommendation framework. Since
nobody has ever evaluated this element, the recommendation system cannot know
A Multi-agent and Content-Based Course Recommender System … 667

Table 1 An overview of recent papers on E-learning recommender systems

References Year Collaborative filtering Content based Proposed approach
[11] 2010 ✓ Recommendation of learning
materials using good learners’
ratings and content-based
filtering
[12] 2010 ✓ Recommendation system that
provides learners with the most
appropriate answers and clues
[13] 2011 ✓ Recommendation of elective
courses
[14] 2013 ✓ Automatic Academic Advisor
[15] 2015 ✓ Recommendation system to
recommend faculty of an
educational institution to
students
[16] 2016 ✓ ✓ A content-based
recommendation algorithm for
learning resources
[17] 2017 ✓ Recommendation system for
optional university courses
[18] 2017 ✓ ✓ Recommendation system for
learning materials
[19] 2017 ✓ ✓ Hybrid recommendation
system for university courses
[20] 2018 ✓ ✓ HAR-SI: A novel hybrid article
recommendation approach
[21] 2019 ✓ Book recommendation system
[22] 2020 ✓ An improved recommendation
system for online learning
environments

what other elements that are similar. For example, in a course recommendation
system, new courses cannot be recommended until certain number of learners
evaluate it [23].
• Sparsity: is also a significant problem in collaborative filtering, it occurs when the
number of learners who have rated courses is too small compared to the number
of available courses. In practice, it is difficult to find similar neighbors for the
learners to generate appropriate recommendation. This is one of the most obvious
drawbacks of recommendation systems based on collaborative filtering [23].
To deal with these challenges, we propose to use content-based approach with
multi-agent system for our course recommender system, because this method can
overcome the cold start problem, thus, it only has to analyze the items and a single
student’s profile for the recommendation.
668 M. Amane et al.

5 Methodology

The number of students using e-learning platforms is explosively growing day by

day, as well as the number of online courses posted. So, it becomes essential to utilize
a recommender system that help them to find the optimal and appropriate courses
for learners.
In this section, we present our proposed architecture of a recommendation system
in an e-learning platform with multi-agent system, the system is based on content-
based method to recommend relevant university course to a student, this method
uses only information about the description and features of the courses that users has
previously viewed. We compute similarity between courses, to obtain the list of top N
courses, and then, we use two agents, the first for positive comments and the second
for negative comments (messages received from chat, forum and labels of every
platforms). These agents will improve the Top N recommendation list according to
the values extracted from comments. The figure below (Fig. 3) shows the structural
architecture of our proposed recommender system for e-learning, students view and
rate the documents; they can also interact by likes or comments. The output of the
system is the Top-N similar courses to the active learner. Below we describe the
workflow steps of the course recommender system:
1. Importing data of learners and courses from the database.
2. Data preprocessing incorporates data integration, collection, and preparation.
3. Data weighting by using the TF-IDF measure.
4. Calculating similarity using cosine similarity and generating the Top N
recommendation list.

Fig. 3 Proposed E-learning recommendation system architecture with Multi-agent system

A Multi-agent and Content-Based Course Recommender System … 669

5. Extracting positives and negatives comments of each course using multi-agent

system composed of two agents, the first one for positives comments, and the
second for the negatives comments.
6. Reclassify the top N list considering the values obtained from the multi-agents
system.
7. Exporting list of recommended courses to the active learner.

5.1 Recommendation Filtering Techniques

Vector Space Model (VSM). Content-based recommendation systems use the

vector space model (VSM), it is a representation that is often used for text docu-
ments. A document is represented as an m-dimensional vector, where each weight
vector indicates the degree of association between the document and the term.

Two primary stages of calculation are involved in the vector space model: weight
calculation and cosine similarity calculation [11].
• One of the most used CBF techniques is the TF-IDF measure [24]. The weight
wi,j is calculated (TF-IDF) with normalized frequency as follow in Eq. 1:
 
f i, j D
Wi, j = × log (1)
max t f t, j di

where:
– fi,j: number of term i appear in text j,
– maxt ft,j : is the Max frequency by the whole of all t keywords that occur in d j,
– D is the sum of recommended documents to learners, and di is the number of
documents that comprise term i.
• After the weighting step, a measure of vector similarity must be defined. This
similarity aims to determine the proximity between two documents. There are
many similarity measures, but the most widely used measure in the literature is
cosine similarity [11]:
Where:

→ − −
→− →
Cos − → = wc . ws
wc , w (2)
wc .−
−
→ →
s
w s
670 M. Amane et al.

5.2 Experimental Results

To test the recommendation systems, the datasets are available on MovieLens [11],
these data are less suitable for testing on the recommendation systems in e-learning
because the datasets are based on the film rankings. For the experiment, we use
a sample of dataset of university courses, which contain information (Course Id,
Course Title, Description…) about 8012 courses as shown in this capture (Fig. 4).
After preprocessing the data and choosing the relevant features for our problem,
which are Course Title and Description, we start calculating TF-IDF scores, and
we compute the Cosine Similarity between courses. The next step is to returns the
most similar courses based on the cosine similarity score. After developing and
implementing our recommendation system, we have as result the list of recommended
courses.
For example, the top ten similar courses of “ASP.NET 3.5 Fundamentals” are
(Table 2):

Fig. 4 The head of courses’ Data Frame

Table 2 Top ten similar

Course Id Course Title Score
courses of “ASP.NET 3.5
Fundamentals” 8 ASP.NET 3.5 Working 0.7615954282461982
With Data
6 ASP.NET Ajax 0.7615945517647128
Fundamentals
6174 ASP.NET Identity 2 0.7615945441987076
Fundamentals
565 Introduction to ASP.NET 0.7615945263051523
Web Services
1882 ASP.NET 5: First Look 0.761594512596039
10 ASP.NET MVC 0.761594479737011
Fundamentals
791 One ASP.NET From 0.761594476106794
Scratch
A Multi-agent and Content-Based Course Recommender System … 671

6 Conclusion and Future Work

Recommendation systems play an important role in improving e-learning systems;

they try to recommend intelligently beneficial courses to a learner according to their
preferences and needs. This article gives an analysis of the recommendation systems
used in the context of e-learning, first, we classify the recommendation systems in
three classes: content-based filtering, collaborative filtering and hybrid method. We
explained the principle of how each technique works, as well as their use in e-learning
systems. According to our analysis, we find that collaborative filtering method is the
most used technique in course recommender system. However, it has some drawbacks
like cold start issue and sparsity of data, for this reason, our proposed approach is
based on content-based filtering and multi-agent system, this method can deal with
these challenges, the experiment results show that the courses recommended are very
similar to the user’ preferences.
In the future, we will concentrate on creating a hybrid e-learning platform recom-
mendation system based on several recommendation techniques in various ways, in
order to benefit from their advantages and integrates intelligent methods to improve
the performance of recommendations for online learners.

References

1. Jiang L, Cheng Y, Yang L, Li J, Yan H, Wang X (2019) A trust-based collaborative filtering

algorithm for E-commerce recommendation system. J Ambient Intell Hum Comput 10:3023–
3034
2. Obeidat R, Duwairi R, Al-Aiad A (2013) A collaborative recommendation system for online
courses recommendations. In: 2019 international conference on deep learning and machine
learning in emerging applications (Deep-ML). IEEE, Istanbul, pp 49–54
3. Ma B, Taniguchi Y, Konomi S (2019) Design a course recommendation system based on asso-
ciation rule for hybrid learning environments. Information processing society of Japan, p 7
4. Singh H, Miah SJ (2020) Smart education literature: a theoretical analysis. Educ Inf Technol
25:3299–3328
5. Wu Y, Zhao Y, Wei S (2020) Collaborative filtering recommendation algorithm based on
interval-valued fuzzy numbers. Appl Intell (2020)
6. Melville P, Sindhwani V (2010) Recommender systems. In: Sammut C, Webb GI (eds) Ency-
clopedia of machine learning. Springer, Boston, pp 829–838
7. Alharbi H, Jayawardena A, Kwan P (2014) Social recommender system for predicting the needs
of students/instructors: review and proposed framework. In: 2014 international conference on
future internet of things and cloud. IEEE, Barcelona, pp 534–540
8. Prasad R (2012) A categorical review of recommender systems. IJDPS. 3:73–83
9. Al-Badarenah A, Alsakran J (2016) An automated recommender system for course selection.
IJACSA 7:166–175
10. Upendran D, Chatterjee S, Sindhumol S, Bijlani K (2016) Application of predictive analytics in
intelligent course recommendation. Procedia Comput Sci 93:917–923
11. Ghauth KI, Abdullah NA (2010) Learning materials recommendation using good learners’
ratings and content-based filtering. Educ Tech Res Dev 58:711–727
12. Souali K, El Afia A, Faizi R, Chiheb R (2011) A new recommender system for e-learning envi-
ronments. In: 2011 International Conference on Multimedia Computing and Systems. IEEE,
Ouarzazate, pp 1–4
672 M. Amane et al.

13. Ray S, Sharma A (2011) A collaborative filtering based approach for recommending elective
courses. In: Dua S, Sahni S, Goyal DP (eds) Information intelligence, systems, technology and
management. Springer, Heidelberg, pp 330–339
14. Taha K (2012) Automatic academic advisor. In: Proceedings of the 8th IEEE international
conference on collaborative computing: networking, applications and worksharing. IEEE, Pitts-
burgh
15. Nagpal D, Kaur S, Gujral S, Singh A (2015) FR: a recommender for finding faculty based on
CF technique. Procedia Comput Sci 70:499–507
16. Shu J, Shen X, Liu H, Yi B, Zhang Z (2018) A content-based recommendation algorithm for
learning resources. Multimed Syst 24:163–173
17. Bhumichitr K, Channarukul S, Saejiem N, Jiamthapthaksin R, Nongpong K (2017) Recom-
mender systems for university elective course recommendation. In: 2017 14th international
joint conference on computer science and software engineering (JCSSE). IEEE, Nakhon Si
Thammarat, pp 1–5
18. Turnip R, Nurjanah D, Kusumo DS (2017) Hybrid recommender system for learning mate-
rial using content-based filtering and collaborative filtering with good learners’ rating. In: 2017
IEEE conference on e-learning, e-management and e-Services (IC3e). IEEE, Miri, Sarawak, pp
61–66
19. Khoja Z, Shetty S (2017) Hybrid recommender system for college courses. In: 2017 interna-
tional conference on computational science and computational intelligence (CSCI). IEEE, Las
Vegas, pp 1167–1171
20. Wang G, He X, Ishuga CI (2018) HAR-SI: a novel hybrid article recommendation approach
integrating with social information in scientific social network. Knowl-Based Syst 148:85–99
21. Bhagirathi YN, Kiran P (2019) Book recommendation system using KNN algorithm. 2:4
22. Sunil, Doja MN (2020) An improved recommender system for e-learning environments to
enhance learning capabilities of learners. In: Singh PK, Panigrahi BK, Suryadevara NK, Sharma
SK, Singh AP (eds.) Proceedings of ICETIT 2019. Springer International Publishing, Cham, pp
604–612
23. Jain S, Grover A, Thakur PS, Choudhary SK (2015) Trends, problems and solutions of recom-
mender system. In: International conference on computing, communication & automation.
IEEE, Greater Noida, pp 955–958
24. Sielis GA, Tzanavari A, Papadopoulos GA (2015) Recommender systems review: types, tech-
niques and applications. 12
Image and Information Processing
Sentiment Classification for Moroccan
Arabic Using the Lexicon-Based
Approach

Chorouk Ferroud and Abdeljalil Elouardighi

Abstract Social networks users contribute in producing huge quantity of data due
to their increasing online activity. However, extracting useful knowledge from social
networks can be a difficult process, especially when conducting sentiment analysis
on Arabic texts. The challenging aspect of Arabic language resides primarily in the
mutual use of standard and dialectal forms in expressing opinions. This work focuses
on sentiment analysis using the semantic orientation approach that rely on a lexicon
of opinion words and a rule-based classifier to achieve sentiment classification. In
this methodology, we build a large sentiment analysis lexicon that covers modern
standard Arabic and Moroccan dialectal Arabic. Then, we have developed several
rules and tested them in the classification process. The proposed methodology showed
promising results.

Keywords Sentiment classification · Semantic orientation · Moroccan dialect ·

Arabic lexicon

1 Introduction

Today, we are in the era of massive information sharing, accentuated by mobile

technologies, since almost anytime and anywhere, it is possible to share information,
comment, or opinion using social networks (SN).
The Arabic opinions shared via SN provides valuable information for the decision
makers. However analyzing such unstructured data is a challenging process since it
require to deal with a massive amount of comments and messages that are written
in informal styles and often with dialects [1]. In addition to the complexity of the
language and the little research work dealing with its particular characteristics and
morphologies, it is difficult to extract the semantic orientation or polarity of these
opinions without using manual annotation which can be costly and time consuming
[2].

C. Ferroud · A. Elouardighi (B)

LM2CE Laboratory, FSJES, Hassan 1st University, Settat, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 675
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_61
676 C. Ferroud and A. Elouardighi

In sentiment analysis (SA), we intend to find out attitudes, knowledge and feelings
declared in a shared text or comment, in order to extract its polarity: positive, negative
or neutral. Two approaches have been developed for SA. One based on the machine
learning (ML) and the other one on semantic orientation (SO) [3].
In this work, we focus on the SO approach, an unsupervised sentiment classifica-
tion based on a lexicon and rule-based algorithm, in order to classify Arabic opinions
that are expressed in modern standard Arabic (MSA) and Moroccan dialectal Arabic
(MDA). Thus in the first step of our methodology we build a lexicon that covers
both languages, and then we develop a set of rules that enable to calculate the SO.
Moreover, we want to test the effect of different polarity rating scales on the SO.
Firstly, the binary rating scale; where each word in the lexicon is rated with +1 if it
is positive or −1 if it is negative. Secondly, the ordinal rating scale; where each word
is rated in a scale from +1 to +5 if it is positive and from −1 to −5 if it is negative.
We start our article by presenting the most relevant work of the SA employing
the social networks, then we describe our SA approach mainly the construction of
a SA lexicon in MSA and MDA as well as the design and implementation the rule-
based classifier. The main obtained results are presented in Sect. 4, and we close our
contribution with a conclusion and some possible perspectives for this work.

2 Related Work

Over the past decade, there has been an explosion of research works interested in
SA such as analyzing attitudes and emotions of individuals towards entities such as
products, services, organizations etc. We start our paper by presenting some research
work in this area and close to our contribution.
In [1] Abdulla et al. proposes two approaches for SA: one based on the Arabic
corpus and the other on the lexicon. As there are few publicly available Arabic
datasets and lexicons for SA, they created by hand an annotated dataset and they
presented the steps of building their lexicon. They found that for their first approach,
the SVM method used for classification offers the highest precision.
In addition, they noted that the increase of the lexicon allows an improvement in
the performance of their second approach.
Hatzivassiloglou and McKeown [2] used a method that recovers automatically
the SO information positive or negative by collecting indirect information from a
large corpus of the conjoined adjectives. To predict whether they have the same or
different SO, they used a log-linear regression model. The achieved quality of the
developed model was 82%. They classified several sets of adjectives based on the
inferred links and labelled them as positive or negative. They achieved 92% accuracy
for the classification task and 100% for the labelling one.
Duwairi [4] used ML techniques for SA to define the polarity of Arabic tweets
with dialects. One of his dares is that stemming algorithms do not work well with
dialects words with an extended set of stop words. He proposed a Framework that
is able to perform SA on tweets written in MSA or in Jordanian dialectal Arabic.
Sentiment Classification for Moroccan Arabic … 677

The basis of this Framework is a dialect lexicon that maps dialect words to their
corresponding MSA words.
Hu et al. [5] used a SA approach that depends on the correlation and indication
of emotional signals and was tested on two sets of twitter data. The obtained results
show the effectiveness of their approach and the significant of including the different
emoticons in the SA.
From a large data set collected from Arabic tweets, Twairesh et al. [6] have
developed two lexicons for SA. The importance of their lexicons is that their capacity
to capture the idiosyncratic nature of social media text. In addition, the possibility
of using them in different kinds of issues such as product reviews.
Ibrahim et al. [7] used a characteristic-based sentence-level approach for Arabic
SA. The approach developed is based, on the one hand on the lexicon of Arabic
idioms/sayings to improve the detection of the sentiment polarity in Arabic sentences.
On the other hand, syntactic features were used for the conflicting sentences allowing
the improvement of the performance of the sentiment classification. They introduced
a wide lexicon of Arabic sentiment words and automatically expandable. The lexicon
building process started by a manual step that consisted of collecting and annotating
gold-standard sentiment words, then it was expanded automatically through word
research in an online Arabic dictionary. The data set focuses on tweets and microblogs
written in modern standard Arabic and in Egyptian dialectal. The experimental results
indicated high levels of performance, with a preciseness over than 95%.
Abdul-Mageed and Diab [8] presents SANA, a huge Arabic lexicon that is a large-
scale multi-gender lexicon for Arabic SSA (subjectivity and sentiment analysis). This
lexicon covering also, the Egyptian Dialect and the Levantine dialect Arabic, and
was developed using manual and semi-automatic techniques.
Al-Ayyoub and Bani Essa [9] developed a sentiment lexicon with nearly 120,000
Arabic terms and a tool based on calculating predicates for AS using the lexicon
approach, which was one of the most commonly used technique. The obtained results
were better in terms of accuracy in comparison with that based on keywords approach.
Khaliq and Carroll [10] used an unsupervised methodology to learn non- concate-
native morphology, which they apply to induce a standard Arabic lexicon of roots
and pattern models. Their approach was based on the concept that roots and patterns
can be affirmed by jointly recursive notation based on hypothesized patterns and root
frequencies. Their morphological analysis with the induced lexicon carry out a root
identification precision greater than 94%.
Shoukry and Rafea [11] developed a sentiment classification model for the Egyp-
tian dialect using Arabic tweets that are analyzed to identify their polarity (positive or
negative). Their model finds its importance in the region of the Middle East that uses
largely the Arabic in the twitter network. Overall, this SA works has dealt with Arabic
language based on two approaches, the ML approach that uses supervised classifiers
and the SO approach that requires the construction of a lexicon. In the SO approach,
the authors followed different methods to construct their lexicons, including manual
and automatic methods. In addition, some works have been limited to treating the
standard Arabic only, while others have extended their lexicons by integrating one
or more Arabic dialects. In our work, we focus on SA to deal with comments written
678 C. Ferroud and A. Elouardighi

in MSA and in MDA, indeed this dialect is still not fully addressed in the context
of the SO approach. This is why the construction of a new lexicon devoted to both
standard Arabic and the Moroccan dialect constitutes an essential element of our
contribution.

3 Lexicon Based Approach

Supervised leaning models built for SA require a labelled data for learning. However,
the manual sentiment annotation can be very costly in time and effort. In addition,
the constructed model may be biased by the domain [12]. We devote this section
to present our methodology for polarity classification that address the sentiments
expressed in standard Arabic and Moroccan dialect.

3.1 Arabic Text Pre-processing

Sentiment classification of Arabic text is challenging process because the Arabic

language is characterized by an inflectional and morphological complexity and many
letters can takes different shapes according to their position in the word.
In addition, user shared post may contain a combination of MSA and MDA,
which makes text processing an extremely difficult job that extends to all steps of
the pre-processing and classification of Arabic texts [1].
We have applied several pre-processing tasks, starting with data cleaning by
deleting foreign characters, symbols, numbers, hash tags, punctuation marks and
symbols such as “? ! @ $ #”. The normalization was applied to unify Arabic char-
acters and removes all diacritics, anomalies and repetitions as (a lot). The
tokenization task splits the text into distinctive tokens (words) which enable to apply
more processing at word level. We have deleted some stop words like conjunctions,
particles and pronouns that do not provide any information for the SA.
For stemming we have used an Arabic light stemmer that treats both MSA and
MDA [13]. It is a process of removing a small set of prefixes and/or suffixes, without
trying to process the infixes, or recognizing patterns and finding roots. It reduces less
the dimension of the features, and preserves more the meaning of the words [14, 15].
We show in Table 1 the pre-processing task applied on Arabic text.

3.2 Lexicon Rules and Construction

The first step towards building our Arabic lexicon for SA is the collection of seed
words, which have been the subject of several extensions using synonyms of words
taken from an online dictionary [16]. For each word, about two to three synonyms
Sentiment Classification for Moroccan Arabic … 679

Table 1 The main pre-processing tasks on Arabic text

was added by inheriting the same polarity as their original seed words, and we have
translated the standard Arabic lexicon into the Moroccan dialectal.
In the end, the lexicon contained 5942 words, with 2127 positives word and 3815
negatives. To obtain the SO of an Arabic comment, we have built a tool that associate
each word of the lexicon to a weight identifying its polarity. The word does not exist
in the lexicon its polarity is null. For feature construction, we have used the n-gram
method that each word is considered as one token regardless to the surrounding
words. To calculate the SO of the text, it suffices to aggregate the polarity score of
each word constituting the text from the lexicon.
The obtained score represents the polarity of the whole text. In order to enhance
the performance of our lexicon, we have implemented some rules for negation,
intensification and a third rule for conflicting phrases.
A negative value is assigned to words that inverse the polarity of the word [15, 17].
The following list presents some negation words in MSA and MDA:
For example in (I don’t like noise) the verb (I like) is positive
word in our lexicon with polarity equals to 1. However, due to the presence of the
negation word‫( ال‬don’t) which precedes it, switch the polarity to −1. In another
situation, negation word reverse completely the polarity of any word to the opposite.
In addition, we have enriched the list by inserting some negation words from the
Moroccan dialects such as .
On the other hand, intensification increases the intensity of words polarity to make
it even stronger [17, 18]. Some words were used for intensification:
. Our approach deals with such cases by adding +1 to a positive word and
−1 to a negative one. For instance, the sentence (This book is very
cool), has a polarity equals to 2 since has +1 in our lexicon, and the presence of
(very) accentuate the polarity by append another +1.
Conflicting phrases consist of two words with opposite polarity; noun followed
by adjective, one of them is positive and the other is negative [7]. For example, the
pattern (Mysterious vision) consists of two sentiment words, the noun
(vision) which express positive polarity and the adjective (Mysterious)
which express negative polarity. Using the lexicon only, their polarity is 0, but our
classifier labelled it as negative polarity. As we observed in most cases, if there exists
a pattern of two opposite sentiment words, it yields a negative sentiment polarity
such as (Temporary release), (fake love), (misbehavior)
680 C. Ferroud and A. Elouardighi

Fig. 1 Sentiment classification process of Arabic comments using the lexicon-based approach

Table 2 Example of polarity score of Arabic comments with binary and ordinal scales

etc. We summarize in Fig. 1 the process of sentiment classification based on our

lexicon approach.
The following table (Table 2) shows an example of polarity scores obtained with
the binary and the ordinal scales.

4 Experimental Results

We dedicate this section to present the main results of the experiments obtained as
well as their comparisons with those obtained in other similar works. Two dataset are
used for testing the efficiency of our methodology. The first one is a balanced dataset
that contain 5000 labelled comments (2500 positive comments and 2500 negative
ones) and concern various domains including politics, culture, sports etc. The second
dataset is unbalanced since it contains 3544 positive comments and 6400 negative
ones with a total of 9944 comment [19]. Both datasets contain comments written in
MSA and MDA.
Sentiment Classification for Moroccan Arabic … 681

We evaluate the performance of the lexicon depending on two rating scales,

namely the binary scale and the ordinal scale. In the binary scale, the positive and
negative words have the same weight (+1) and (−1) respectively. The ordinal scale,
which is inspired from Likert scale, aim to consider the polarity strength of each
word in the lexicon.
Our main results obtained in the classification process, with the two datasets, are
presented in Tables 3 and 4. To evaluate the lexicon performance in the classification
step we have used the Accuracy and the F-score, the most performance measures
used in the classification. During these experiments, we have compared the scores of
the binary scale and the ordinal scale and we have tested the effect of the proposed
rules. Overall, these results show, with both datasets, that the combination of the
ordinal-scaled lexicon and the proposed rules has brought a slight improvement in
the sentiment classification. This confirms the importance of including these rules in
the implementation of our lexicon. Identical results were obtained with other Arabic
dialects for the SA. In [11] Shoukry and Rafea implemented sentiment classification
for Arabic tweets. The F-score obtained was 65.4%. Abdulla et al. [1] have applied
two approaches for Arabic SA: corpus-based and lexicon-based, the F-score obtained
was equal to 59.6%. The presentation of these results was not intended primarily to
compare them with our results but also to have approximately an order of magni-
tude of the performance that it can be achieved. The results obtained are not very
exceptional and our approach requires improvements. Among the possible tracks,
we cite the detection of sarcastic comments that can reduce the performance of our
classification process and the improvement of our lexicon by enriching it with other
Arabic words for both polarities (positive or negative).

Table 3 Classification
Dataset: Facebook 5000
results: dataset 1
Steps Accuracy F-score
1- Binary scale lexicon 69,48% 69,47%
2- Ordinal scale lexicon 72,41% 72,41%
3- Ordinal scale lexicon + rules 73,05% 73,04%

Table 4 Classification
Dataset: elec2016
results: dataset 2
Steps Accuracy F-score
1- Binary scale lexicon 67,04% 62,34%
2- Ordinal scale lexicon 68,31% 65,41%
3- Ordinal scale lexicon + rules 69,09% 66,04%
682 C. Ferroud and A. Elouardighi

5 Conclusion

This paper addressed Arabic Sentiment Analysis through the unsupervised approach.
The contribution consisted of building a large Arabic lexicon that contain opinion
words in Standard Arabic and Moroccan dialect. Several rules were developed in
order to achieve polarity classification of social web data. Thus, several experiments
were conducted through the various steps of this process in order to observe the
amelioration obtained on the performances. The results showed that the extended
rating scale of the lexicon (1–5) alongside with the proposed rules have returned the
best scores.
We devote the future work to deal with sarcastic comments, since detecting
sarcasm is still one of the significant issues faced in the context of SA.

References

1. Abdulla NA, Ahmed NA, Shehab MA, Al-Ayyoub M (2013) Arabic sentiment analysis:
Lexicon-based and corpus-based. In: 2013 IEEE Jordan conference on applied electrical
engineering and computing technologies (AEECT), pp 1–6. IEEE
2. Hatzivassiloglou V, McKeown K (1997) Predicting the semantic orientation of adjectives. In:
35th annual meeting of the association for computational linguistics and 8th conference of the
European chapter of the association for computational linguistics, pp 174–181
3. Liu B (2012) Sentiment analysis and opinion mining. Synth Lect Hum Lang Technol 5(1):1–167
4. Duwairi RM (2015) Sentiment analysis for dialectical arabic. In: 2015 6th international
conference on information and communication systems (ICICS), pp 166–170. IEEE
5. Hu X, Tang J, Gao H, Liu H (2013) Unsupervised sentiment analysis with emotional signals.
In: Proceedings of the 22nd international conference on World Wide Web, pp 607–618
6. Al-Twairesh N, Al-Khalifa H, Al-Salman A (2016) Arasenti: large-scale twitter- specific
arabic sentiment lexicons. In: Proceedings of the 54th annual meeting of the association for
computational linguistics (Volume 1: Long Papers), pp 697–705
7. Ibrahim HS, Abdou SM, Gheith M (2015) Sentiment analysis for modern standard arabic and
colloquial. arXiv preprint arXiv:1505.03105
8. Abdul-Mageed M, Diab MT (2014) Sana: a large scale multi-genre, multi-dialect Lexicon for
Arabic subjectivity and sentiment analysis. In: LREC, pp 1162–1169
9. Al-Ayyoub M, Essa SB, Alsmadi I (2015) Lexicon-based sentiment analysis of arabic tweets.
Int J Soc Netw Min 2(2):101–114
10. Khaliq B, Carroll J (2013) Induction of root and pattern lexicon for unsupervised morphological
analysis of arabic
11. Shoukry A, Rafea A (2012) Sentence-level arabic sentiment analysis. In: 2012 international
conference on collaboration technologies and systems (CTS), pp 546–550. IEEE
12. Read J, Carroll J (2009) Weakly supervised techniques for domain-independent sentiment
classification. In: Proceedings of the 1st international CIKM workshop on topic-sentiment
analysis for mass opinion, pp 45–52
13. Maghfour M, Elouardighi A (2018) Standard and dialectal arabic text classification for senti-
ment analysis. In: International conference on model and data engineering, pp 282–291.
Springer
14. Larkey LS, Ballesteros L, Connell ME (2007) Light stemming for arabic information retrieval.
In: Arabic computational morphology, pp 221–243. Springer
15. Saif H, Fernandez M, He Y, Alani H (2013) Evaluation datasets for Twitter sentiment analysis:
a survey and a new dataset, the STS-gold
Sentiment Classification for Moroccan Arabic … 683

16. Almaany online dictionary (2020). https://www.almaany.com/. Accessed 05 Feb 2020

17. Taboada M, Brooke J, To_loski M, Voll K, Stede M (2011) Lexicon-based methods for
sentiment analysis. Comput Linguist. 37(2):267–307
18. Farghaly A, Shaalan K (2009) Arabic natural language processing: challenges and solutions.
ACM Trans Asian Lang Inf Process (TALIP) 8(4):1–22
19. Elouardighi A, Maghfour M, Hammia H (2017) Collecting and processing Arabic Facebook
comments for sentiment analysis. In: International conference on model and data engineering,
pp 262–274. Springer
Forecasting Tourism Demand
in Marrakech with SQD-PCA-SVR

Houria Laaroussi, Fatima Guerouate, and Mohamed Sbihi

Abstract The previous studies have used search query data to predict the numbers
of tourists expected to visit a particular destination. The SQD-PCA-SVR model
was established in the aim of achieving the accurate forecasts of tourism demand.
This study used Search Query Data (SQD) to supplement traditional data in order
to predict tourist arrivals in Marrakech. The Principal Component Analysis (PCA)
is applied to extract the low-dimensional and efficient feature information. Support
Vector Regression (SVR) can approximate a non-linear system. In our approach,
we used SVR with GridSearchCv and ten-fold Cross-Validation to choose optimal
hyper parameters. The SQD-PCA-SVR model is a new combination, consisting of
a search engine data, dimensional reduction algorithm, and machine learning. The
results indicate that models with SQD-PCA-SVR provide more accurate forecasts
than other models.

Keywords SQD · PCA · SVR · GridSearchCv · Ten-fold cross-validation ·

Tourism demand

1 Introduction

Search query data has become an important data source for the prediction of tourism
demand [1]. Tourists are accustomed to use search engines to make different decisions
related to all details of their travels, such as hotels, restaurants and transportations.
The tourists’ queries reflect various kinds of information needs, these information
could be captured online in different time periods on those search engines.
The use of search query data in predicting tourism demand necessitates the
careful selection of keywords that have considerable correlations with the prediction
purposes. Also, high dimensions of variables may induce multicollinearity problems
[2]. To overcome the problem of higher dimension, Principal Component Anal-
ysis (PCA) is employed prior to the application of any machine learning algorithm.

H. Laaroussi (B) · F. Guerouate · M. Sbihi

LASTIMI- EST Salé - Morocco, Mohammed V University in Rabat, Rabat, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 685
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_62
686 H. Laaroussi et al.

PCA, a dimensionality reduction technique, decreases the number of variables while

preserving data information.
The tourism demands, such as the tourist arrivals, have a complex nonlinear char-
acteristic. Support vector regression (SVR) has been utilized, on a large scale, to
deal with non-linear time series forecasting problems; yet it suffers from the critical
parameters [3]. This method is not only unable to automatically draw out information,
but also it tends to falls into the local optimum and overfitting problems [4].
In this work, we proposed a new model to forecast tourism demand, incorporating
search query data, Feature Extraction and machine learning. The forecasted demand
variable was the tourist arrivals in Marrakech. We developed a way for automatically
crawling Google Trends with keyword suggestions. An SVR model was built. As
the performance of SVR depends on the efficient selection of its hyper-parameters,
it was imperative to select the kernel function type and the values of the penalty
parameter. Linear regression model, which only uses the historical information of
observed tourist arrivals, was used for comparison in order to assess the forecasting
performance of our model.
The plan of the paper is organized as follows: Sect. 2 presents the Litera-
ture Review. Section 3 presents the proposed methodology. Section 4 provides the
Empirical study. Section 5 highlights the conclusion of the study.

2 Literature Review

2.1 Forecasting Tourism Demand

The search query data is largely used in studies to predict the tourist arrivals. For
instance, [5] used the Baidu Index and Google Trends data to forecast tourist numbers
in China by using autoregressive moving average models. [6] made use of Google
search index to predict Caribbean Islands’ tourist flows, and found that Google Trends
does in fact help to forecast tourism demand. [7] used Internet search indexes and
machine learning to forecast tourist arrivals expected in popular destinations in China.
To extract principal components from Search Query Data, PCA can be used, as it
can eliminate the problems of multi-collinearity and over-fitting. For example, [8],
by using PCA, managed to decrease the dimension of the 15 Baidu indexes that are
employed to forecast tourist arrivals of Beijing and succeeded in obtaining the first six
principal components. [2] used principal component analysis (PCA) and generalized
dynamic factor model (GDFM) to create a composite search index based on Google
trends data.
Tourism demand forecasting predominantly uses SVR, which is suitable for
nonlinear prediction. For examples, [9] combined SVR with the Bat algorithm (BA)
with the aim of forecasting tourist flows by incorporating search engine data.
Forecasting Tourism Demand in Marrakech … 687

2.2 Support Vector Regression

The basic idea of SVR is as follows: Suppose training data pairs (xi , yi ) are given
where xi ∈ Rn , yi ∈ R, i = 1; 2; 3; . . . xi is the input value and y is the target value.
The SVR function can be inferred from following formula:

f (x) = ωϕ(xi ) + b. (1)

Where ω is the n-dimensional weight factor, b is the offset, and ϕ(x) is the map
function of mapping xi into the high l-dimensional feature space.
The final support vector regression model can be constructed using Lagrangian
function:


n
   
f (x) = αi − αi∗ K xi , x j + b. (2)
j=1

 
Where K xi , x j is the kernel function and αi , αi∗ is the Lagrangian multiplier

2.3 Principal Component Analysis

PCA is a frequently used data dimensionality reduction method, The PCA is used
to generate the principal components through recombination of tourism demand
influential factors. Basic steps of PCA algorithm as shown in Fig. 1.

Fig. 1 Principal Component

Analysis steps Data Standardizaon

Covariance matrix

Eigen Vectors and Eigen Values

Compung the principal components

Reducing the dimensions of the data

688 H. Laaroussi et al.

3 Methodology

This work puts forward a forecasting framework to construct the tourism demand
prediction model using search query data, as shown in Fig. 2. The detailed steps are
as follows:
(1) Selection of the search engine platforms: In this empirical study, the majority of
travelers will opt for a search engine to obtain information, and later make the
convenient plans concerning a particular destination. In this paper, we decided
to use Google search engine data.
(2) Selection of keywords: First, we determine the seed search keywords related
to the tourist destination. Then, we procure the search volume data of each
keyword from Google Trend, and we extract the related keywords to portray
all the facets of tourists’ concerns on the destination, by means of Google
Trends related queries.
(3) Process Data: Remove irrelevant keywords using Pearson coefficient correla-
tion.
(4) Feature extraction using PCA: PCA seek to represent the variation contained
within the dataset using a small number of factors.
(5) Forecasting model: SQD-PCA-SVR model is constructed and evaluated. In
this study, a Linear Regression Model was established, as a benchmark model.
(6) Model evaluation: We used three criteria to compare the forecasting perfor-
mance of the models: Mean Absolute Error (MAE), Root Mean Square Error
(RMSE), and Mean Absolute Percentage Error (MAPE). The best forecasting
model is the one that records the lowest values on the three criteria.

1-Selecon of the search engine plaorms:

Google
Split Data
2-Selecon of keywords: tourism, traffic,
lodging, dinning, shopping, and
Entertainment. Train
Test
Data
Data
3-Process Data: Clean Data ,Correlaon
Analysis SVR parameter
opmizaon by
Grid search and
ten-fold cross
4-Feature Extracon using PCA validaon

SVR
SVR
5-Forecasng model Forecasng Training

6-Model evaluaon

Fig. 2 The proposed framework

Forecasting Tourism Demand in Marrakech … 689

4 Empirical Study

4.1 Data Description

Tourism Demand Data

Marrakech plays a crucial role in Morocco’s economic development, and it is an
enormously popular tourist destination. It was selected as an application case in this
paper to forecast tourism demand. The Tourism Observatory website (http://www.
observatoiredutourisme.ma/) provided monthly tourist arrivals to Marrakech, with
data series covering the period from January 2012 to August 2019. Figure 3 shows
tourist arrivals in Marrakech.

Choosing of Search Queries Data

We implement the following procedure to choose keywords and collect the potential
keywords representing travelers’ interests and search behavior by using the function
of “related searches” on Google Trends [2, 5]:
1. The related literature [2, 5] indicates that there are six main elements of tourism
planning that reveal the travelers’ demand: tourism, traffic, lodging, dinning,
shopping, and Entertainment.
2. For each category, we designated the basic search queries. The original list of
search keywords is illustrated in Table 1. We used PyTrends library to collect
data on the search volume of these keywords.
3. We searched the queries that are related to the primary keywords in loops.
We found 298 keywords related to the initial list search keywords by using
Google Trends related-queries section. The monthly search volume data of each
keyword (from January 2012 to August 2019) was crawled using PyTrends
library. After discarding duplications and missing data extremely low volume,
we acquired a list of 243 related Google search keywords.
4. We calculated the Pearson correlation coefficient between dependent variables
(e.g. Marrakech tourist arrivals) and candidate queries. To forecast tourism
demand, we set 0.5 as a threshold; consequently, all poorly correlated keywords

Fig. 3 Tourist arrivals in

Marrakech
690 H. Laaroussi et al.

Table 1 Category of tourism and the corresponding keywords

Category Keywords
Tourism Marrakech tourism, Marrakech weather, Marrakech maps, Marrakech Agency,
guides
Traffic Marrakech airlines, Marrakech flights, Marrakech bus, Train Tickets
Lodging Marrakech hotels, Marrakech accommodation
Dining Marrakech food, Marrakech restaurant
Entertainment Marrakech night life, Marrakech recreation, Marrakech bars, Marrakech shows
Shopping Marrakech shopping, Marrakech Center

Table 2 The Pearson

Index Correlation coefficient
correlation between this 17
keywords and tourism Ctm marrakech 0.66416
demand Booking marrakech 0.638827
Hotel marrakech booking 0.629227
Hotel marrakech 0.6341
Marrakech hotel 0.596743
Essaouira 0.581919
Restaurant Marrakech 0.568054
Marrakech restaurant 0.568054
Gueliz marrakech 0.567331
Hotel spa marrakech 0.540885
Marrakech restaurant 0.538466
Vol marrakech 0.530394
Maroc 0.529573
Marrakech essaouira 0.51739
Marrakech airport food 0.516077
Point bar 0.513514
Marrakech bar 0.5135

below this threshold are excluded. Eventually, 17 keywords were chosen. The
Pearson correlation between these 17 keywords and tourism demand is shown
in Table 2.

4.2 Feature Extraction Using PCA

We standardized the data-set using standardscaler() method to normalize the data

within a given range. Then, PCA was used to reduce the number of variables.
Forecasting Tourism Demand in Marrakech … 691

Table 3 The first six

Principal component Contribution rate Cumulative
principal components
contribution rate/%
PCA1 58.8 58.8
PCA2 16.3 75.1
PCA3 7.8 82.9
PCA4 4 86.9
PCA5 2.9 89.8
PCA6 2.4 92.2

Fig. 4 The cumulative

contribution rate

In this paper, the PCA method is used in the Scikit-learn machine learning library
in the Python programming language. The module sklearn decomposition provides
the PCA object, which can convert the data into principal components.
PCA is used, and 17 factors are reduced to six principal components with a
cumulative contribution rate of over 92%. The contribution rate and cumulative
contribution rate of six principal components are presented in Table 3. The cumulative
contribution rate is presented in Fig. 4.

4.3 Model Implementation

The data procured by PCA was divided into two sections: 80% a training set and
20% a testing set. SVR model was implemented using Python and pipeline library
of Sklearn. Parameter tuning was executed by the inner-function GridSearchCV()
and ten-fold cross validation method. Grid search was applied to find the hyper-
parameter which has the best score. The best combination of parameter values is set
as follows: the penalty parameter is set to 1; the parameter gamma reaches 0.001; the
kernel function selects the radial basis function kernel; the parameter of degree is at
7; and the parameter epsilon is at 0.01. A comparison between the actual number of
tourists and the forecasted number of tourists using the SQD-PCA-SVR is illustrated
in Fig. 5. The predicted graph followed the ground truth in broad outlines.
The SQD-PCA-SVR model is compared with linear regression, for evaluation
of the two models, MAE (mean absolute error), RMSE (root mean square error)
692 H. Laaroussi et al.

Fig. 5 SQD-PCA-SVR model in test data

Table 4 Forecasting
MAE RMSE MAPE
evaluation
SQD-PCA-SVR model 1005.35 1323.54 24.43
Linear regression model 1524.68 1976.87 30.76

and MAPE (average absolute percentage error) and were calculated using sklearn
metrics. The results are reported in Table 4. The best forecasting model is the one that
has the lowest values in these measures. Finally, the capability and efficiency of the
SQD-ACP-SVR model to estimate the tourism demand in Marrakech are validated
by the results.

5 Conclusions

In this paper, an efficient approach SQD-PCA–SVR is introduced to forecast

Marrakech tourist arrivals, integrating the search query data (SQD), principal
component analysis (PCA) and support vector regression (SVR).
For many decades, it was very challenging to make a precise forecasting of tourism
demand, due to its non-linear nature. The SVR has the advantage of its capability to
learn non-linear models. Yet, it can be undermined by the disadvantages of tuning
a number of hyper parameters such as the penalty, kernel function, gamma, degree
and epsilon. Grid search was implemented so as to tune the hyper parameters of our
model, and Ten-Fold Cross-Validation was used for inspecting the model’s accuracies
for different partition of test and train data.
Forecasting Tourism Demand in Marrakech … 693

The precision of tourist volume forecasting can be considerably enhanced by the

use of web search data. Keywords were selected according to Pearson correlation
coefficient. The dimensionality of data was reduced by principal component analysis.
According to the variance and cumulative contribution rate, the first six principal
components were combined with SVR model. SQD-PCA–SVR was compared with
Linear regression models. Our model has the lowest MAE, MAPE and RMSE.

References

1. Li X, Qi W, Peng G, Lv B (2016) Tourism forecasting by search engine data with noise-

processing. Afr J Bus Manag 10(6):114–130. https://doi.org/10.5897/ajbm2015.7945
2. LiLi X, Pan B, Law R, Huang X (2017) Forecasting tourism demand with composite search
index. In: Tour Manag 59(April):57–66. https://doi.org/10.1016/j.tourman.2016.07.005
3. Chen R, Liang CY, Hong WC, Gu DX (2015) Forecasting Holiday daily tourist flow based on
seasonal support vector regression with adaptive genetic algorithm. Appl Soft Comput J. https://
doi.org/10.1016/j.asoc.2014.10.022
4. Zhang B, Pu Y, Wang Y, Li J (2019) Forecasting hotel accommodation demand based on LSTM
model incorporating internet search index. Sustainability (Switzerland) 11(17). https://doi.org/
10.3390/su11174708
5. Yang X, Pan B, Evans JA, Lv B (2015) Forecasting Chinese tourist volume with search engine
data. Tour Manag 46(February):386–397. https://doi.org/10.1016/j.tourman.2014.07.019
6. Bangwayo-Skeete PF, Skeete RW (2015) Can Google data improve the forecasting performance
of tourist arrivals? Mixed-data sampling approach. Tour Manag 46:454–464. https://doi.org/10.
1016/j.tourman.2014.07.014
7. Sun S, Wei Y, Tsui KL, Wang S (2019) Forecasting tourist arrivals with machine learning and
internet search index. Tour Manag 70(February):1–10. https://doi.org/10.1016/j.tourman.2018.
07.010
8. Li S, Chen T, Wang L, Ming C (2018) Effective tourist volume forecasting supported by PCA
and improved BPNN using Baidu index. Tour Manag 68(October):116–126. https://doi.org/10.
1016/j.tourman.2018.03.006
9. Zhang B, Huang X, Li N, Law R (2017) A novel hybrid model for tourist volume forecasting
incorporating search engine data. Asia Pac J Tour Res 22(3):245–254. https://doi.org/10.1080/
10941665.2016.1232742
Toward Improving Arabic Text
Preprocessing in Sentiment Analysis

Mohcine Maghfour and Abdeljalil Elouardighi

Abstract The quantity of data produced by social networks, achieves every year
unprecedented records, because these platforms became the main sources where
subjective opinions are expressed. Processing opinions data from the raw state to
extract useful information can be a very difficult task, especially when it comes to
Arabic text that includes standard and dialectal forms. In this paper, we focus on
sentiment analysis based on Facebook comments shared in Modern Standard Arabic
and in Moroccan Dialectal Arabic. Thus, we put the text preprocessing under the
scope by testing several schemes and by inspecting their effects on the supervised
classification of sentiments.

Keywords Sentiment analysis · Natural language processing · Facebook

comments · Arabic language · Supervised classification

1 Introduction

The rise and growth of social networks (SN) and microblogging has revolutionized
the social relationships and behaviors. With platforms like Facebook and twitter,
users can express their sentiments and opinions about different topics. The increasing
amount of contents generated by such platforms provide an enormous resource of
textual data and present an excellent opportunity to comprehend the user’s opinion.
Sentiment analysis (SA) is a domain of study that aim to analyze opinions and senti-
ments shared with subjective text in web platforms, such as blogs, forums and espe-
cially social networks, in order to extract valuable knowledge that can be exploited
in decision making process [1].
Because SA is based on text in natural language, the linguistic aspects reveals
the main challenges of this analysis. Particularly, the Arabic language possess a rich

M. Maghfour (B) · A. Elouardighi

LM2CE Laboratory, FSJES, Hassan 1st University, Settat, Morocco
A. Elouardighi
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 695
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_63
696 M. Maghfour and A. Elouardighi

vocabulary and a compact morphology, which presents a major preprocessing chal-

lenge. Otherwise, the interest addressed toward SA based on Arabic language has
surely increased; however, the progress made in this area is still narrow compared
with the other languages. Moreover, two forms of Arabic are used jointly in social
networks. On one hand, the modern standard Arabic (MSA) is an official form
commonly known in all Arab countries. On the other hand, the dialectal Arabic,
which is an informal and vernacular language spoken in daily life within a country
or a region. In SA, text processing contribute highly to reduce the noise of textual data,
yet its effect on the classification performance may differ depending on the executed
processing tasks. Furthermore, Arabic texts that are expressed via social network
evolve several issues and challenges such as the informal writing style and dialectal
dominance. Thus, it is important to investigate the effect of each preprocessing task
on the polarity classification.
In this paper, we focus on the effect of preprocessing on Arabic Facebook
comments that are expressed in MSA and in MDA for SA, Therefore, for the aims
of our study:
• We highlight some important challenges and issues in preprocessing text in MSA
and in MDA for SA;
• We propose concise and relevant preprocessing tasks to face and overcome these
challenges; therefore, we present a new stemmer to deal with the Moroccan
dialectal Arabic;
• We investigate the effect of the proposed tasks on real Arabic Facebook comments
by testing different text representations and various supervised classifiers.
The rest of this paper is organized as follows: in Sect. 2, we introduce some
related works and present their proposed approaches and results. In Sect. 3, we
outline the major text issues in both MSA and MDA, and then we describe our
methodology of collecting and processing Facebook comments with more focus on
the preprocessing tasks. In Sect. 3, we present the tested preprocessing schemes,
alongside with the text representations and the machine learning classifiers. The
obtained results are presented and discussed in Sect. 4. We conclude this work by
highlighting the important points and the perspectives.

2 Related Work

Several studies have addressed sentiment analysis based on Arabic text; nevertheless,
the linguistic aspects still one of the major issues in SA. Thus, researchers were
intrigued by idea of investigating the effects of text preprocessing, so they have tried
to test, select and improve the preprocessing tasks in order to overcome these issues.
In their work, Shoukry and Rafea [2] carried out a sentiment analysis study based
on tweets that were published in modern standard Arabic and in Egyptian dialect;
the substantial difficulties that they have faced were about the morphology of the
Toward Improving Arabic Text Preprocessing … 697

studied dialect. For this reason, the authors have developed their own stemming
method, which is suitable for the preprocessing of texts in the Egyptian dialect.
Duwairi and El-Orfali [3] tested several preprocessing strategies for SA on two
Arabic datasets; the first concerns politics, while the second is composed from movie
reviews. Their strategies included stemming, light stemming, stop words removal,
and feature reduction. The results showed major improvement when they have applied
root stemming and feature reduction technique with every classifier: SVM, NB or
KNN.
Moreover, in [4], Duwari et al. experimented the contribution of their adopted
processing tasks in SA, namely negation detection, stemming, stop words filtering
and dialect conversion to MSA. At first, their results showed insignificant improve-
ment with stemming and stop words filtering for every classifier: SVM, Naive Bayes
or KNN. In addition, despite the expected enhancement of classification performance
following the use of negation detection and dialectal conversion, their results were
counter-intuitive since they obtained similar performance to the baseline.
Ghelli et al. [5] presented an approach to construct an annotated corpus automat-
ically for Algerian SA in Arabic and Arabizi. A corpus composed from Facebook
comments was collected. After an Algerian SA lexicon was compiled, they have
applied several ML algorithms in the sentiment classification step, such as random
forest and SVM. The preprocessing was focused on cleaning of the messages, agglu-
tination treatment and negation treatment. The obtained results showed better results
with Arabic than Arabizi and with internal dataset than the external one. This work is
more focused on testing the combined effect of feature selection with ML classifiers
and text representation schemes. The results showed the outperformance of SVM
compared to Decisions trees and Radom forest by using the same preprocessing
tasks with every classifier.
In the previous related works the sources of the collected texts are twitter, movie
reviews websites and Facebook. The use of Arabic language on these platforms is
usually associated with one dialect at least; therefore, text-preprocessing tasks are
expected to have an impact on the SA performance.

3 Role of Preprocessing in Sentiment Analysis Methodology

3.1 Challenges of Arabic Language in Social Networks

Sentiment analysis is carried out through a process of extracting information retained

in subjective texts (comments, tweets, and reviews) in order to deduce the polarity
of sentiments expressed in them. However, such a process is restrained with consid-
erable challenges. From one hand, the nature of textual data is unstructured and
noisy; On the other hand, social network users adopt an informal writing style. In this
698 M. Maghfour and A. Elouardighi

context, the presence of dialects is common and spelling mistakes are frequent. There-
fore, the role of preprocessing is to reduce noise and prepare the feature extraction
(words, sequences of words) that will be used in the classification step.
Although the standard and the dialectal Arabic are commonly used in social
networks, the morphological differences between this forms of Arabic indicates
that each form require its own preprocessing. Arabic language has many complex
particularities such as vocabulary richness and morphological compactness. Most
words in Arabic are derived from a root with three or four characters [6]. For instance
with the root ( - to know), new words are created:

The compact aspect of Arabic results from the attachment and concatenation
capabilities for verbs, nouns and even particles. Besides the MSA, the MDA is used in
social networks frequently. The major aspect of MDA is spelling variation, therefore
with the lack of uniform rules in writing various styles can be used to express the
same phrase.

3.2 Sentiment Analysis Methodology

Our SA study is based on machine learning approach. Thus at first we describe

in Fig. 1 the proposed methodology of sentiment analysis. In data collection step,
Facebook Graph API was used to collect public comments that were expressed in
standard Arabic and Moroccan dialect. Text preprocessing plays an essential role in

Fig. 1 Text preprocessing tasks in sentiment analysis methodology

Toward Improving Arabic Text Preprocessing … 699

this methodology; it aims to provide clean data with less noise, so that the supervised
classification step achieve good performance [7].
Otherwise, Features construction is an intermediate step from the unstructured
textual data towards the structured feature vectors. In this study, we use two types
of configurations for features construction: the n-grams extraction scheme [7], and
feature-weighting scheme.
Polarity annotation of Facebook comments is a difficult task, since these comments
lack of polarity indicator like those used in product reviews (rating with stars or
a score). Furthermore, the sentiment expressed in Facebook comments may covers
serval topics besides the topic of interest [8]. Polarity annotation is even more compli-
cated when dealing with Moroccan Arabic comments because lexical resources dedi-
cated for this language are still insufficient. Thus, the use a human annotation was
indispensable.
The last step in the proposed methodology is the supervised classification; at this
level, the dataset is split randomly into two subsets, the learning subset is allocated to
train the supervised models, while the second subset is retained in order to evaluate
the performance.

3.3 Text Preprocessing Tasks in Depth

Basic Tasks: Cleaning, Normalizing and Tokenizing. Text expressed in social

networks may include irregularities that are irrelevant in determining sentiment,
such as hashtags and symbols, hyperlinks, punctuation, etc. Therefore filtering these
elements will help to reduce the noise. Within the words of a text, anomalies are also
produced, notably with the misspelling, the duplication of the letters, which are used
to emphasize a word. In Arabic language, diacritics represent short vowels; despite
their rareness in Facebook, some comments may contain them. Moreover, Arabic
tatweel (lengthening) tend to appear as well, the sentence: (the
book is useful) is composed from two words, diacritics were used in the first one
, while the second one is marked by tatweel. Thus, cleaning and
normalizing this sentence should return . Through the tokenization task, the
text is split into small entities known as tokens that are usually words or sequences of
words. since Arabic is a morphologically rich language, the tokenization produces
complex entities, for instance, a subject or an object can be concatenated with the
main verb of a sentence.

Stemming. The task of stemming aims to group the different varieties of a word into
a single stem. Two categories of stemmers are developed for the Arabic language;
the root stemmer tries to extract the base root of the word, which is composed from
three to four characters in Arabic. [9]. Otherwise, the light stemmer remove prefixes
and suffixes of the word and retain the stem only [10]. In this work, we try to test
three different Arabic stemmers, Root stemmer, light stemmer and a dialectal light
stemmer. For root stemming we use ISRI [11], this root-extraction stemmer for Arabic
700 M. Maghfour and A. Elouardighi

is similar to the Khoja stemmer [12] but without a root dictionary. For Arabic light
stemming we try Motaz light stemmer [13] that actually modifies the ISRI Arabic
Stemmer to perform light stemming for Arabic words. For dialectal stemming, we
have developed and implemented our own new stemmer through which we have tried
to deal with the Moroccan dialectal Arabic complexities. Therefore, the dialectal
stemmer provides two stemming options: MSA or MDA, where the MDA stemming
option incorporates almost all the MSA operations, moreover it include also the
specific tasks to process the Moroccan dialect.

Stop Words Removal. Stop words designs conjunctions, pronouns, prepositions,

etc. Their filtering is supposed, from a theoretical perspective, to reduce the
noise implicated with their low discriminative power on sentiment classification.
Depending on the language, wither the MSA or MDA, two lists of stop words were
developed and filtered from the corpus, although some stop words in MDA have the
quality of being attached with other words, thereby we have integrated their filtering
in the dialectal stemmer.

Negation. Negation words are essential to consider in SA because they change

partially or totally the expressed meaning. The Arabic negation words are divided
into two groups, negation words for nouns and negation words for verbs. In MSA
nouns negation we usually use and sometimes
In MDA, this negation is expressed with Verb negation in MSA is expressed
with various conjunctions depending on the verb tense, the most known of them are
for the past (did not), γ for the present (do not) and for the future (will not).
Verb negation is more complex in MDA, because besides the non-standardized
styles of writing; verbs incorporate simultaneously a prefix and a suffix. For inte-
grating negation in the SA, we have unified the different negation words within the
same category: nominal or verbal. Due the concatenation of negation words with the
verb in MDA, we have included its processing task as a feature in our stemmer.

Synonym Dictionary. Since the Moroccan dialectal language is a derived vernacular

language from the MSA, it has inherited most its vocabulary from standard Arabic,
yet it possess particular and independent vocabulary. In SA based on a machine
learning approach, feature extraction from comments written in both MSA and MDA
is vulnerable to redundancy. Otherwise, Arabic Light stemming participate efficiently
to merge relative group of words into the same stem with less over stemming errors,
which conserve the meaning. However, it is more sensitive to under stemming error
[14]. These two reasons led us to implement a corrective preprocessing task by
making a synonym dictionary.
At first, we apply the basic preprocessing task and stemming to the comments of
our corpus, then, we extract the features using the unigram scheme. At this stage we
apply the chi2 feature selection technique that allow to rank the feature by their impor-
tance [15]. We use the most important feature as seeds (200 feature) that we looked
for their synonyms, lemmas, and equivalents. The next step consists of replacing the
found synonyms by their seeds at the level of the corpus.
Toward Improving Arabic Text Preprocessing … 701

4 Implementing the Preprocessing Schemes

4.1 Dataset Description

In this study, we have collected a corpus composed from about 10,000 Facebook
comments that were published in standard Arabic and Moroccan Arabic. The topics
of these comments were various, since they concerned economics, politics, sports
etc. A group of experts was in charge of the polarity labeling; this task resulted in
7106 negative comments and 2939 positive comments.

4.2 Text Preprocessing Schemes

Text preprocessing, is an important step to prepare the text for classification,

because of the contribution in reducing noise. However empirical studies have shown
contrasted results about the contribution of each task in the classification perfor-
mance [3, 16]. In this work, we aim to design, test and select the best preprocessing
scheme for Facebook comments in MSA and in MDA, in other words, we will iden-
tify the tasks that will be included in the optimized text processing pipeline. The
first scheme represents the baseline of this analysis, where only the basic tasks are
applied including cleaning, normalizing, tokenizing. In scheme 2, stemming was
applied in addition to the basic tasks of the first, scheme at this stage; we have
tested three different Arabic stemmers: ISRI Root stemmer, Motaz light stemmer
and our own dialectal light stemmer. The best stemmer was selected for the next
steps. In scheme 3, 4 and 5, we will identify respectively, the contribution of stop
words removal, negation, and synonyms replacements using a dictionary. At the end,
scheme 6 corresponds to experimenting all the proposed tasks.

4.3 Machine Learning Approach

Several elements affect the performance in the sentiment analysis, especially the
language of the studied comments, their length and their topics [8]. For this raison,
it is important to carry out many experimentations on feature construction schemes
so that to achieve the best performance.
We have tested four text representations by combining feature-weighing tech-
niques: TF (term frequency) and TF-IDF (term frequency-inverse document
frequency) with feature extraction techniques: unigram and unigrams + bigrams.
Then, we have applied three classifiers that are already known for their robustness
in sentiment analysis [17–19], namely SVM (Support vector machines) with a linear
kernel, naive bayes and maximum entropy. For each of the four feature construction
schemes we divide the dataset into two subsets, 75% of the observations for the
702 M. Maghfour and A. Elouardighi

training set and 25% for the test set. We have applied a 10 fold cross validation
alongside with the parameter optimization to train the predictive models. The test
set, which is composed from unseen data, served us to evaluate the quality of these
models. The F-score was selected as an evaluation metric. We have used python
programming language for text preprocessing and scikit-learn library for polarity
classification.

5 Results and Discussion

Before presenting the results of this work, we briefly describe the steps of the iterative
approach followed for this purpose. Based on Facebook Arabic comments, four
features construction schemes were executed, and then at the classification level three
classifiers were used: Maximum entropy, Naive Bayes and SVM. For evaluation, the
F-score metric was calculated. The obtained results are presented Table 1. This results
show that the dialectal stemmer that we have developed outperforms both the root
stemmer and the light stemmer, regardless of the classifiers or the feature construction
scheme. Furthermore with this stemmer the best contribution was obtained with Naive
bayes and TF-Bigram + bigram scheme (F-score = 84.10%).
In scheme 3, the effect of stop words removal was tested. The results show signif-
icant effect of this task. Capturing negation in a sentiment analysis study was very
important because negation words have the capability of changing the expressed
meaning entirely.
The contribution of this task was observed through the results of preprocessing
scheme 4. The effect of negation task was positive in most cases with ME an NB
(expect the TF/unigram + bigram configuration). Synonym dictionary was built
in order to deal with redundancy and reduce further the sparsity of the features
vector. The results obtained with preprocessing scheme 5 were relevant. The best
performance obtained with this task concern ME with F-score = 84.94%, indeed this
result was the best in overall experimentations. Scheme 6 represents the contribution
of all proposed processing tasks; stemming, stop word, negation and dictionary.
The combined scheme enhances both NB and SVM to their maximum perfor-
mance. On one hand, the best F-score of NB (F-score = 84.53%) was achieved
under this scheme with TF/Unigram. On the other hand the best result of SVM (F-
score = 84.01%) was obtained with the configuration TF-IDF/Unigram + bigram.
Our study reveals similar results because the contribution of the Arabic root stemmer
that we have tested outperform the light stemmer, however, we have outlined that
by considering the dialectal issues in light stemming we have obtained even higher
score. Shoukry et rafea [2] carried out a SA study on Arabic tweets in Egyptian
dialect, they have tested their own stemmer that addresses Egyptian dialect issues
against an MSA light stemmer, they found better results with their implemented
dialectal stemmer. Furthermore, with stop words removal their study performance
increased by only 0.1%, which was similar to our findings. Similar works, based on
testing the effect of preprocessing on sentiment analysis, show similarities to our
Toward Improving Arabic Text Preprocessing … 703

Table 1 The obtained results

Maximum entropy TF TF TF-IDF TF-IDF
1g 1g+2g 1g 1g+2g
Base 80.20 80.20 80.55 80.68
Base + Root stemmer 82.90 83.37 83.36 83.50
Base + Light stemmer 82.03 82.10 82.24 82.40
Base + Dialectal stemmer 83.53 83.85 83.87 83.98
Base + Stem + Stop words 83.52 83.73 83.77 83.98
Base + Stem + Negation 83.77 83.86 83.97 84.39
Base + Stem + Dictionary 83.95 84.63 84.37 84.94
All tasks 84.11 84.37 84.21 84.56
Naive Bayes
Base 81.33 81.33 78.63 78.90
Base + Root stemmer 83.64 83.92 80.16 81.33
Base + Light stemmer 82.92 82.48 80.83 81.10
Base + Dialectal stemmer 83.94 84.10 82.04 82.42
Base + Stem + Stop words 84.29 84.15 81.98 82.63
Base + Stem + Negation 84.19 84.01 82.28 82.63
Base + Stem + Dictionary 84.26 84.39 81.25 81.92
All tasks 84.53 84.15 81.23 81.78
SVM
Base 79.15 79.15 80.07 80.36
Base + Root stemmer 82.31 82.46 82.63 82.84
Base + Light stemmer 81.17 81.27 81.50 81.53
Base + Dialectal stemmer 82.89 83.13 83.20 83.37
Base + Stem + Stop words 82.55 82.79 83.46 83.70
Base + Stem + Negation 82.65 83.07 83.59 83.37
Base + Stem + Dictionary 83.48 83.45 83.52 83.92
All tasks 83.30 83.57 83.70 84.01

study in term of results. In [3] duawri and El-Orfali have tested the effect of Arabic
light and root stemming on SA of politics dataset, they found that root stemming was
more helpful than light stemming.

6 Conclusion

This work was focused on the effect of processing Arabic text on SA. At first, we
have described the methodology of collecting and preprocessing Arabic Facebook
comments to perform supervised classification on sentiments. Then we have tested
704 M. Maghfour and A. Elouardighi

several preprocessing schemes in order to determine the contribution of each task.

Various feature representation were implemented in order to look for the highest
performance of each classification model. For the supervised classification task,
we have applied three algorithms: Maximum Entropy, Naive Bayes and SVM. The
obtained results showed contrasted effects of the preprocessing tasks, while Stem-
ming and synonym dictionary have participated in enhancing the classification perfor-
mance, the contributions of stop words removal and negation were less significant.
Additional and more advanced research is required to complete and improve this
work. The perspectives of our research are many.
Firstly, we aim to explore the lexicon-based approach, because this approach is
unsupervised and it will allow us to avoid the expensive task of manual annotation
of sentiments. In addition, the features of lexicon methods can be used to assist and
improve the learning of the supervised classification algorithms.

References

1. Pozzi F, Fersini E, Messina E, Liu B (2016) Sentiment analysis in social networks. Elsevier
Science
2. Shoukry A, Rafea A (2012) Preprocessing egyptian dialect tweets for sentiment mining. In:
The fourth workshop on computational approaches to arabic script-based languages, pp 47–56
3. Duwairi R, El-Orfali M (2014) A study of the effects of preprocessing strategies on sentiment
analysis for arabic text. J Inf Sci 40(4):501–513
4. Duwairi RM, Marji R, Sha’ban N, Rushaidat S (2014) Sentiment analysis in arabic tweets.
In: 2014 5th international conference on information and communication systems (ICICS), pp
1–6. IEEE
5. Guellil I, Adeel A, Azouaou F, Hussain A (2018) Sentialg: automated corpus annotation for
algerian sentiment analysis. In: International conference on brain inspired cognitive systems,
pp 557–567. Springer
6. Habash NY (2010) Introduction to arabic natural language processing. Synth Lect Hum Lang
Technol 3(1):1–187
7. Liu B (2012) Sentiment analysis and opinion mining. Synth Lect Hum Lang Technol 5(1):1–167
8. Pang B, Lee L et al (2008) Opinion mining and sentiment analysis. Found Trends Inf Retrieval
2(1–2):1–135
9. Duwairi R, Al-Refai M, Khasawneh N (2007) Stemming versus light stemming as feature
selection techniques for arabic text categorization. In: 2007 4th international conference on
Innovations in information technology, IIT 2007, pp 446–450. IEEE
10. Al-Anzi FS, AbuZeina D (2015) Stemming impact on arabic text categorization performance:
a survey. In: 2015 5th international conference on information & communication technology
and accessibility (ICTA), pp 1–7. IEEE
11. Taghva K, Elkhoury R, Coombs J (2005) Arabic stemming without a root dictionary. In:
International conference on information technology: coding and computing, ITCC 2005, vol
1, pp 152–157. IEEE
12. Khoja S, Garside R (1999) Stemming arabic text. Lancaster, UK, Computing Department,
Lancaster University
13. Saad MK, Ashour W (2010) Arabic morphological tools for text mining. Corpora 18:19
14. Larkey LS, Ballesteros L, Connell ME (2007) Light stemming for arabic information retrieval.
In: Arabic computational morphology, pp 221–243. Springer
Toward Improving Arabic Text Preprocessing … 705

15. Liu H, Setiono R (1995) Chi2: feature selection and discretization of numeric at tributes. In:
Proceedings, seventh international conference on tools with artificial intelligence, pp 388–391.
IEEE
16. Alwakid G, Osman T, Hughes-Roberts T (2017) Challenges in sentiment analysis for arabic
social networks. Procedia Comput Sci 117:89–100
17. Pang B, Lee L, Vaithyanathan S (2002) Thumbs up?: sentiment classification using machine
learning techniques. In: Proceedings of the ACL-02 conference on Empirical methods in natural
language processing, vol 10, pp 79–86. Association for Computational Linguistics
18. Dave K, Lawrence S, Pennock DM (2003) Mining the peanut gallery: opinion extraction and
semantic classification of product reviews. In: Proceedings of the 12th international conference
on World Wide Web, pp 519–528. ACM
19. Go A, Bhayani R, Huang L (2009) Twitter sentiment classification using distant supervision.
CS224N Project Report, Stanford 1(12)
IIMDR: Intelligence Integration Model
for Document Retrieval

S. Aditya, P. Muhil Aditya, Gerard Deepak, and A. Santhanavijayan

Abstract With the rise of digital documents on the internet, the demand for accurate
search results has been a challenging task and many notable algorithms have been
used to cater to such queries. However, with the increasingly fast-paced discoveries
in machine learning and artificial intelligence, many of the algorithms have been
outdated and no longer used. It is necessary to display only the relevant content that
the user queried for. In this paper, IIMDR: Intelligence Integration Model for Docu-
ment Retrieval is proposed. Jaccard Similarity is used to calculate word similarity
in search result documents. Experiments using the proposed model on RCV dataset
show that the proposed framework results in improved performance by providing
more relevant documents to the user query. An accuracy of 86.78% on the RCV1
dataset has been achieved with this method.

Keywords Document retrieval · Jaccard similarity · Random forest · SPARQL

endpoint

1 Introduction

The accessibility of advanced records over the web is expanded and there is a require-
ment for compelling and snappy techniques to recover and sort out the computerized
archives. Since information is scattered universally and is sloppy, it is a provoking
assignment to create viable strategies that can produce top-notch highlights in these
archives. It is important to diminish the hole between clients’ inquiry and recovered
outcomes. In this paper, IIMDR, an Intelligence Integration Model for Document
Retrieval is proposed. Word similarity is computed using Jaccard Similarity present
in search results documents. With the rise of digital documents on the internet, the

S. Aditya · P. Muhil Aditya

Department of Electronics and Communication Engineering, National Institute of Technology,
Tiruchirappalli, Tamil Nadu, India
G. Deepak (B) · A. Santhanavijayan
Department of Computer Science and Engineering, National Institute of Technology,
Tiruchirappalli, Tamil Nadu, India

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 707
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_64
708 S. Aditya et al.

demand for accurate search results has been a challenging task and many notable
algorithms have been used to cater to such queries. However, with the increasingly
fast-paced discoveries in machine learning and artificial intelligence, many of the
algorithms have been outdated and no longer used.

Motivation: The need for an improved and modern algorithm to identify the most
relevant documents. Modern algorithms such as Latent Semantic Analysis and
random forest classification were used in this paper which has improved the effi-
ciency and speed of the result drastically. The use of Jaccard Similarity has also
proved to produce efficient results.

Contribution: In this paper, an Intelligence Integration Model for Document

Retrieval to rank the documents has been proposed. Here the initial classification
has proved to produce very efficient results as we are reducing the load that goes
into the pre-processing stage. Similarity plays a vital role in determining the relevant
documents that the user searched for. The query is preprocessed using tokenization,
stemming, and stop word removal. Jaccard similarity Computation is performed
between query words and class labels of the classified set of documents. Here we use
random forest classification in the preprocessing stage. Ranking of the documents
is done by picking the top K classes of the class labels which helps to get the final
relevant documents that the user needs.

Organization: The remainder of the paper is organized as follows. Section 2

describes the Related Work, and the Proposed Architecture is depicted in Sect. 3.
Sections 4 and 5 describe the Recommender System Framework and Data collection
and Preprocessing respectively. Section 6 describes the Performance Evaluation, and
the paper is concluded in Sect. 7.

2 Related Work

Landauer et al. [1] demonstrated the use of LSA and its usefulness in the anal-
ysis. Foltz [2] described the behavior of LSA and explained how semantic similarity
for text-based research works. Kostoff et al. [3] used to discover text using phrase
frequency analysis and phrase co-occurrence analysis. However, this technique didn’t
prove to be useful in the data set that was chosen for this paper as it didn’t deliver the
results that were expected. Ramya et al. [4] used Latent Semantic Analysis which
delivered very promising results. They used a DRDLC framework that worked on
the dataset RCV1 where the documents were ranked and accessed. Grineva et al. [5]
presented a novel method wherein key term extraction from text documents is done
using the properties of graphs. They found that the terms related to main topics tend
to bunch up as densely interconnected subgraphs. Pushpa et al. [6] proposed in their
paper how to adequately find applicable web administrations. They suggested that
word-level semantics is applied utilizing grouping dependent on the NPMI esteem.
IIMDR: Intelligence Integration Model … 709

K-implies grouping strategy is utilized to the disclosure of different web administra-

tions for mining of basic semantics interfaces. Niwattanaku et al. [7] showed how a
search keyword is computed within acceptance criteria. In [8–18], several ontology-
based methods and their applications have been described which demonstrate the
effectiveness of the proposed approach by giving supporting literature.

3 Proposed Architecture

The need for classifying documents is because in Recommender system compares

all the documents with the user query. This leads to unnecessary computational time
because it compares documents with an irrelevant class label with the user query.
To avoid the overhead time, we classify the documents with relevant labels using
random forest algorithm and further reduce the size of the relevant documents by
concept similarity. After assigning a similarity score to each classified document
with the query, arrange it in descending order and choose the top thirty percent of
documents. These documents are used as a dataset for the Recommender system
algorithm to find out the relevant document for the given input query. Figure 1

Fig. 1 Architecture of the Proposed IIMDR

710 S. Aditya et al.

depicts the architecture of the proposed IIMDR framework. The steps involved are
pre-processing for removing unwanted data and reduce time complexity, Synoni-
mization and SPARQL Endpoint based enrichment to enhance the class labels of the
user query for better classification, classification of the documents to avoid comparing
irrelevant documents, and concept similarity computation to further reduce the clas-
sified documents. Reuters Corpus Volume-1 (RCV-1) is an unlabeled dataset, to
increase the performance of the recommender system it has been labeled using a
real-world domain ontology-based labeling.

3.1 Pre-processing

The input query given by the user will contain words that do not help in finding the
documents and do not help in class labeling of the user query. There are three steps
in this process. First is tokenization, since the query is in sentences, it is split into
words called tokens. The second step is Stop Word removal, words that are present
in the stop words dictionary was removed as they are used to give the sentence a
structure not meaning. The third step is stemming from this reduces the token size
to their respective root word by dropping unnecessary characters usually a suffix.

3.2 Synonimization and SPARQL Endpoint Based

Enrichment

Various words in English have the same meaning, so we found the synonyms of
the tokens from the preprocessed user query. These words will be class labels for
the query which was used for the classification of the dataset using random forest
classification. To obtain the class labels of the query, WordNet 3.0 was used. Then
SPARQL Endpoint based enrichment was done to further increase the number of
class labels for document classification.

3.3 Classification of Dataset

Dataset required to find the relevant document was further reduced using random
forest classification [20]. Random forest was used because it does not need any
feature scaling and provided high accuracy when comparing with other algorithms.


s
Gini = 1 − (Pi)2 (1)
i=1
IIMDR: Intelligence Integration Model … 711

In Eq. (1), Pi represents the frequency of the class observed in the dataset and s
represents the number of classes present in the dataset. Gini impurity was computed
by the random forest classifier at each branch accordingly and the classification was
done.

3.4 Concept Similarity Computation

The classified dataset was further reduced using Concept similarity [19]. A similarity
score was assigned to each document. After scoring the documents were arranged in
descending order of the score as they are documents similar to the user query and in
these set of documents top thirty percent were taken and sent into the recommender
system.

4 Recommender System Framework

The main aim is to retrieve the relevant document from the curated documents that
the user wants by taking the query from them. The architecture is split into three
phases Data Pre-processing, Pattern identification, and concept extraction.

4.1 Data Pre-processing

The Reuter Corpus Volume 1 (RCV1) dataset utilized in this paper contains news
story website pages in XML design. These XML contains extra information such as
webpage author information, time and place of the report, and other metadata infor-
mation which can increase the computation time as they are unnecessary information
that does not contribute to the document search and increase space complexity to
store this information. Steps in Data preprocessing include XML creation (custom)
and Stop word removal and Unique term identification

4.2 XML Creation (Custom)

Since RCV1 contains stories and news in XML design there is a necessity to prepro-
cess this because it contains unnecessary data which would lead to an increase in
time and space complexity while retrieving the relevant documents. To avoid this and
to make the framework exceptionally responsive, data that are not useful in search
activities are removed to make the XML data light. These custom XML data is then
stored in a data repository.
712 S. Aditya et al.

4.2.1 Stop Word Removal and Unique Word Identification

The common words from the dataset that does not contribute to the concept of the
question and not useful in finding the relevant documents. We take each word from
the document and check if it’s a part of the Stop word dictionary or not. Similarly,
we check whether the word is part of the unique words or not. Perform this for every
single document.

4.3 Pattern Identification and Concept Extraction

(i) TF-IDF Vectorizing

With the preprocessed data, TF-IDF [21] operation is performed to find the important
term in each file, by doing this we found the pattern for each file. Generally, the
relevant document that the user needs have the same pattern as the user’s query, so
the resultant patterns of the documents obtained is stored in a repository.

(ii) LSA (Latent semantic analysis) For concept extraction, LSA (Latent Semantic
Analysis) [22] was used, LSA performs SVD on the TF-IDF vectors for dimen-
sionality reduction. LSA itself is an unsupervised way of uncovering synonyms in a
collection of documents. LSA was used instead of other algorithms because to fill in
the “Semantic Gap” in the dimensionality reduction, by performing this we obtained
the topics of each document and stored them in the repository.

4.4 Jaccard Similarity

Patterns extracted from the previous phase was used to find the most similar document
to that of the user query. To find the most similar document pattern to that of the
query, Jaccard similarity [23] was implemented. The documents were ranked based
on the similarity score and the document with the highest score was the relevant
document for the given user query.
Jaccard similarity formula:

|A ∩ B|
J (A, B) = (2)
|A ∪ B|

In Eq. (2), A and B are two documents whose Jaccard similarity is to be found.
We find the ratio of common terms in A and B to the union of terms in A and B.
IIMDR: Intelligence Integration Model … 713

5 Data Collection and Preprocessing

RCV1 (Reuters Corpus Volume 1) is a repository of categorized newswire and stories.

It contains more than 800,000 of these newswires and stories. It is present in an XML
design format and widely used in Text Mining research. It contains a lot of metadata
which is unnecessary in this research. So, preprocessing is done to create a custom
dataset from RCV1 by removing XML labels, URLs, metadata, stop words and
further simplified by tokenization and stemming.

5.1 Implementation

The higher configuration systems may provide better results. User Query was
obtained first, then it was preprocessed where it has gone through tokenization, stop
word removal, and stemming. Next to find the documents with similar class labels
synonymizing was done to get the class labels of the query. After obtaining class
labels random forest classification was done to obtain the relevant documents from
which the top 30% of the documents were taken by assigning the documents simi-
larity score by concept similarity. This curated dataset was preprocessed to remove
the XML tags and stop words. TF-IDF was performed on the pre-processed dataset
and then LSA was performed where the concept of each document was extracted.
By doing so we have obtained the pattern of the documents. The user query’s pattern
is also obtained, and the relevant document was also found out by Jaccard similarity
the document with the highest score was chosen. Table 1 depicts the algorithm for
document retrieval in the proposed approach.

6 Performance Evaluation

LSA can be easily implemented and used. Many practical applications use LSA. LSA
is capable of assuring consistent results. It works well on diverse datasets. Further,
LSA seems to give promising results, especially in the document retrieval domain.
SPARQL was used as the query language in this experiment. As we know that
SPARQL is not constrained within one database it has been proved to be very effective
in the experiment we have performed. Cosine similarity and Jaccard similarity are
a measure of lexical similarity and the performance varies based on various factors.
However, the dataset used Jaccard similarity proved to be more efficient and gave a
more precise result.
Table 2 shows the various comparisons of various frameworks. DRDLC though
proved to be one of the precise frameworks IIMDR had performed better with the
dataset used in this experiment. Note how that LDA and LSA has performed in
combination with Jaccard. LSA has also performed more efficiently when combined
with Trigram as compared to LDA. Figure 2 depicts the FDR vs the number of
714 S. Aditya et al.

Table 1 Algorithm for Document Retrieval from the given input user query
Input: Reuters Corpus Volume 1, User query.
Output: Relevant document that user wants.
Begin
Step 1: Let each document in the data set be Di= (Ti , Ci) where Ti is text of the
document and the Ci is class labels.Let user query be Ur.
Step 2: The objective function is F(D ,Ur)=Di .Where D is total documents and Di
relevant document to the user query.
Step 3: Pre-processing of the user query.
Step 3.1: Tokenization
Step 3.2: stop word removal
Step 3.3: stemming
Step 4: Synonmization using WordNet3.0 and SPARQL Endpoint based enrichment
Step 5: Classification of the dataset using random forest algorithm to find the set of
documents similar to the class label
Step 6: Pick the top 30 percent of the documents based on the descending order of
their similarity score using concept similarity
Step 7: The documents chosen is then pre-processed
Step 7.1: creating custom XML
Step 7.2: stop word removal and unique term identification
Step 8: Pattern Identification and concept Extraction of the curated documents.
Step 8.1: TF-IDF vectorizing
Step 8.2: LSA
Step 9: The relevant document is obtained by arranging the Jaccard similarity score of
the documents with user query in descending order and choosing the first one.
End

Table 2 Comparison of various models to F measure and accuracy

Model Precision% Recall% F-Measure% Accuracy%
LDA + Jaccard 75.12 78.14 76.60 76.63
LSI + Jaccard 76.81 79.32 78.04 78.06
LSI + Jaccard + Knowledge Enrichment 80.17 82.87 81.49 81.52
LDA + Trigram 56.48 59.02 57.72 57.75
LSI + Trigram 59.14 64.12 61.52 61.63
RFD 59.89 61.17 60.52 60.53
DRDLC 74.86 77.32 76.07 76.09
Proposed IIMDR 84.78 88.79 86.73 86.78

recommendations of various frameworks and it shows how the proposed IIMDR

outperforms the other frameworks. We can also notice why LSA is the preferred
LDA from the inferences in Fig. 2.
Figure 3 shows the average nDCG vs values of various frameworks. It can be
interpreted that IIMDR has the highest nDCG value compared to all other proposed
models to yield the best result. LSI + Jaccard + Knowledge and DRDLC model also
have yielded good results.
IIMDR: Intelligence Integration Model … 715

Fig. 2 Comparison of FDR value against Number of recommendations for various frameworks

Fig. 3 Comparison of average nDCG against values for various frameworks

7 Conclusions

The proposed IIMDR has been modeled for recommending documents using a
hybridized approach. The approach successfully amalgamates LSI, Jaccard simi-
larity and Random Forest has been incorporated for classification. The approach uses
SPARQL endpoints for knowledge enrichment which makes it a knowledge-driven
paradigm. The hybridization enhances the outcome of the proposed IIMDR and
makes it better than the baseline models. IIMDR yields an overall accuracy of 86.78%
which makes IIMDR the best in class approach for recommending documents.
716 S. Aditya et al.

References

1. Landauer TK, Foltz PW, Laham D (2009) An introduction to latent semantic analysis
2. Foltz PW (1996) Latent semantic analysis for text-based research. Behav Res Methods Instrum
Comput 28:197–202
3. Kostoff RN, Stump JA, Johnson D, Murday JS, Lau CG, Tolles WM (2006) The structure and
infrastructure of the global nanotechnology literature. J Nanopart Res 8:301321
4. Ramya RS, Ganesh Singh T, Sejal D, Venugopal KR (2008) DRDLC: discovering relevant
documents using latent dirichlet allocation and cosine similarity. In: Conference: the 2018 VII
international conference
5. Grineva M, Grinev M, Lizorkin D (2009) Extracting key terms from noisy and multi-theme
documents. In: WWW 2009 proceedings of the 18th international conference on World Wide
Web
6. Pushpa CN, Deepak G, Kumar A, Thriveni J, Venugopal KR (2020) Improving web service
discovery by hybridization of ontology focused concept clustering and interface semantics.
IEEE. 978-1-7281-6828-9/20
7. Niwattanakul S, Singthongchai J, Naenudorn E, Wanapu S (2013) Using of Jaccard coefficient
for keywords similarity. In: 2013 Proceedings of the international multiconference of engineers
and computer scientists, IMECS 2013, vol I, 13–15 March 2013, Hong Kong
8. Deepak G, Santhanavijayan A (2020) OntoBestFit: a best-fit occurrence estimation strategy
for RDF driven faceted semantic search. Comput Commun 160:284–298
9. Pushpa CN, Deepak G, Kumar A, Thriveni J, Venugopal KR (2020) OntoDisco: improving web
service discovery by hybridization of ontology focused concept clustering and interface seman-
tics. In: 2020 IEEE international conference on electronics, computing and communication
technologies (CONECCT), pp 1–5. IEEE
10. Kumar A, Deepak G, Santhanavijayan A (2020) HeTOnto: a novel approach for concep-
tualization, modeling, visualization, and formalization of domain centric ontologies for heat
transfer. In: 2020 IEEE international conference on electronics, computing and communication
technologies (CONECCT), pp 1–6. IEEE
11. Deepak Gerard, Kasaraneni Dheera (2019) OntoCommerce: an ontology focused semantic
framework for personalised product recommendation for user targeted e-commerce. Int J
Comput Aided Eng Technol 11(4–5):449–466
12. Gulzar Z., Anny Leema A, Deepak G (2018) PCRS: personalized course recommender system
based on hybrid approach. Procedia Comput Sci 125:518–524
13. Deepak G, Teja V, Santhanavijayan A (2020) A novel firefly driven scheme for resume parsing
and matching based on entity linking paradigm. J Discrete Math Sci Crypt 23(1):157–165
14. Haribabu S, Kumar PS, Padhy S, Deepak G, Santhanavijayan A, Kumar N (2019) A novel
approach for ontology focused inter- domain personalized search based on semantic set
expansion. In: 2019 fifteenth international conference on information processing (ICINPRO),
Bengaluru, India, pp 1–5. https://doi.org/10.1109/icinpro47689.2019.9092155
15. Deepak G, Kumar N, Bharadwaj GV, Santhanavijayan A (2019) OntoQuest: an ontological
strategy for automatic question generation for e-assessment using static and dynamic knowl-
edge. In: 2019 fifteenth international conference on information processing (ICINPRO), pp
1–6. IEEE
16. Santhanavijayan AD, Naresh Kumar D, Deepak G (2021) A semantic-aware strategy for auto-
matic speech recognition incorporating deep learning models. In: Intelligent system design, pp
247–254. Springer, Singapore
17. Deepak, G, et al (2019) Design and evaluation of conceptual ontologies for electrochemistry as
a domain. In: 2019 IEEE international WIE conference on electrical and computer engineering
(WIECON-ECE). IEEE
18. Deepak G, Priyadarshini JS (2018) Personalized and enhanced hybridized semantic algorithm
for web image retrieval incorporating ontology classification, strategic query expansion, and
content-based analysis. Comput Electr Eng 72:14–25
IIMDR: Intelligence Integration Model … 717

19. Zheng C, Hu Z, Shen YP (2009) Concept similarity computation based on ontology. In: 2009
international conference on machine learning and cybernetics, vol 1, pp 165–169. IEEE
20. Strobl C, Boulesteix A, Augustin T (2007) “Unbiased split selection for classification trees
based on the Gini index” (PDF). Comput Stat Data Anal 52:483–501
21. Robertson S (2004) Understanding inverse document frequency: on theoretical arguments for
IDF. J Doc. 60(5):503–520. https://doi.org/10.1108/00220410410560582
22. Landauer TK, Foltz PW, Laham D (1998) Introduction to latent semantic analysis. Discourse
Process 25:259–284
23. Levandowsky Michael, Winter David (1971) Distance between sets. Nature 234(5):34–35.
https://doi.org/10.1038/234034a
Effective Application of Information
Technology Tools for Real-Time Project
Management

Mahmoud Ershadi , Marcus Jefferies , Peter Davis ,

and Mohammad Mojtahedi

Abstract Information technology (IT) is evolving rapidly and benefits a wide variety
of industries to break conventional communication barriers. Recent advancements
in management information systems have enabled project-based organizations to
monitor and control their projects more effectively. Although the role of IT tools
(ITT) has been explored in previous research, the importance of each type of tool
in achieving higher effectiveness in real-time management of projects has been less
studied. This study aims to explore the significance of each type of IT tools for
this purpose. Thus, a survey of principal contracting organizations in the construc-
tion industry was carried out, which indicated that reporting tools, integrated plat-
forms, analytical tools, knowledge management tools, and archiving tools are key
ITT groups that should be applied collaboratively. Although reporting tools were
perceived as the most important group from the experts’ view, the correlation analysis
also revealed that all five groups are interrelated and should be applied coherently
to achieve satisfactory outcomes. This study explains the application of ITTs for
real-time project management and directs further research on this topic.

Keywords Information technology tools · Management information systems ·

Real-time project management · The construction industry

M. Ershadi (B) · M. Jefferies · P. Davis

University of Newcastle, Newcastle, Australia
e-mail: [email protected]
M. Jefferies
e-mail: [email protected]
P. Davis
e-mail: [email protected]
M. Mojtahedi
University of New South Wales, Sydney, Australia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 719
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_65
720 M. Ershadi et al.

1 Introduction

The evolution of information technology tools has drastically enhanced the agility of
operations and the precision of management decisions. However, previous research
has identified the construction sector as a laggard in taking advantage of recent
technological advancement in this field [1]. Regarding the multi-player and risky
nature of the construction business, enterprises seek more robust tools to tackle
operational and functional complexities inherent in their activities. Discrepancies of
actual execution and baseline plans have led researchers to introduce the concept
of real-time project management in which automated monitoring technologies are
applied, project performance data is processed in real-time, and the plan is optimized
accordingly [2]. To achieve this target in construction projects, contractors need to
improve their understanding of the effective IT tools (ITT) to take full advantage of
recent IT evolutions for more effective project delivery.
Real-time action requires extensive application of advanced systems to improve
the integration of project planning and control processes and automate controls.
Updating the schedule based on actual performance is performed intermittently in
the traditional project control. While with the introduction of the Internet of Things
(IoT) and intelligent machines, it is possible to obtain real-time performance data
from construction sites to be analyzed and compared with baselines instantly. It
is useful in large projects that many tasks are being executed simultaneously and
minor daily deviations would aggregate and lead to major delays [1]. Besides, it
is not limited to time management but includes keeping the record of information,
lessons learned, and reports to be used for continuous improvement of processes.
It is not possible without the effective application of a variety of ITTs. Thus, this
study contributes to a better understanding of potential ITTs that can be used for this
purpose.
Prior studies highlighted the application of IT tools and systems in business activ-
ities. Akinsola et al. developed a process model based on a set of integrated ITTs
for enhancing the productivity of operations on construction sites [3]. In another
study, Hui Yap et al. discussed how ITT can be applied in managing knowledge
acquired from undertaking business activities to build, maintain, and advance knowl-
edge assets. They also discussed the challenges and barriers to implementing knowl-
edge management (KM) tools in a sample of construction organizations [4]. A
group of studies specifically focused on exploring the benefits of applying ITT for
construction organizations. For example, Love and Irani examined the approaches
used by such enterprises to evaluate and justify their investments in information
technology infrastructure, and the advantages that they have experienced due to IT
implementation.
In recent construction research, ITT is often being discussed in light of building
information modeling (BIM) that integrates technologies involved in the digital repre-
sentation of the features of a facility. For instance, Hartmann et al. explained the
BIM implementation strategies in practical contexts and advocated a “technology
push” implementation approach to align workflows with the functionality of this
Effective Application of Information Technology Tools … 721

system [5]. Besides, Cleveland explores emerging ITT such as portable virtual reality,
digital paper, and computer automatic virtual environment to be applied in support
of construction engineering practices [6]. Notwithstanding these studies, there is
a paucity of research on the application of ITT for real-time project management.
Such research provides practitioners with guidelines on the best approach to lever-
aging information technology for boosting the effectiveness of project monitoring
and control. The purpose of the present paper is to bridge this gap by identifying
main types of ITT, explores their significance from a project management (PM)
perspective, and discuss optimal ways to their implementation.

2 Theoretical Background

Information technology tools are those related to the analysis and transmit of stored
information. Different kinds of ITTs serve distinct areas depending on the context.
From the construction project management perspective, ITTs bring accurate real-time
data from construction sites to project leaders in support of their informed decision-
making within a timeframe. A variety of ITTs has been developed and introduced
for different applications in various construction subsectors [3]. The result of the
literature review revealed that five categories of ITTs have been introduced for project
management. These categories include (1) archiving tools, (2) analytical tools, (3)
reporting tools, (4) KM tools, and (5) integrated platforms (Table 1).
These five groups of ITTs improve PM processes and provide a suitable platform
for better integration that facilitates the establishment of real-time control systems.
Archiving tools allow systematic storing of project artifacts both during the stages
of execution and at the delivery stage. The tools in this category have the same
features as document management systems (DMS) that help organizations to locate

Table 1 A classification of ITTs for construction project management

(1) Archiving (2) Analytical (3) Reporting (4) Knowledge (5) Integrated
tools tools tools management platforms
tools
Digital Building Infrastructure A KM tool to Building
Archival of performance sustainability support Information
Project analysis software reporting tools knowledge sharing Modelling [11]
Information [7] [8] (SRTs) [9] networks [10]
Construction Earned Value Project status Project knowledge Cloud
imaging and Analysis software reporting management tool asset-enabled
archiving [12] [13] system (PSRS) [15] integrated IoT [16]
[14]
Cloud-based Construction delay Near-miss TWiki for An integrated
safety Analysis software information tool knowledge platform of
information [18] [19] management [20] collaborative PM
archive [17] [21]
722 M. Ershadi et al.

versions of important documents and keep records of structured digital documents.

The example of documents stored in project archives includes engineering design
documents, contracts, and requirement documentation. These documents are to be
classified and retrieved later for different purposes such as change management and
tracking [7].
The second category includes analytical tools that analyze collections of data in
search of useful information and insight for planning, control, and management of
projects considering constraints and limitations. In a real-time project management
context, such tools must be properly linked to data collection, data validation, and
reporting tools to decrease the time between the occurrence of events and making
timely decisions to take the control of operations via proper actions. Regarding the
application of advanced technologies and the emerging role of machine learning in
controlling onsite operations based on internet of things (IoT) concepts, the need for
instant analysis of big data is arising. In today’s business environment, undertaking
such computations requires the adoption of big data analytics software. Such ITTs
help to capture greater volumes of performance data from different agents including
both the data that is created by smart onsite meters and the data being entered by
human resources through user interfaces [22].
The third category encompasses reporting tools that act as coordinator agents
between sources of data, analytical software, and management dashboard to present
real-time performance outcomes in the form of a set of indicators. The reporting
ITTs help to capture, share, and transmit information among stakeholders as effec-
tive communication links [14]. The collection of reports overtime constitutes an
intensive knowledge base that contains issues and possible solutions/decisions in
response to them. Further, one of the key characteristics of construction projects
includes being knowledge-intensive due to the involvement of multiple technologies
and interrelated disciplines, as well as the interaction of many stakeholder parties. To
address the need for capturing lessons learned in this collaborative setting, knowl-
edge management (KM) tools (the fourth group) are applied. Such ITTs organize the
process of validating, storing, and sharing lessons learned from current and previous
projects. Nowadays, KM systems provide a web-based collaborative learning envi-
ronment by providing a broad range of users across the organization with real-time
access to the knowledge base of projects [23]. It is noteworthy that there is a close link
between knowledge management tools and archiving tools since they interoperate
to streamline project management-related processes, which necessitate the need for
implementing them consistently within an integrated platform.
The fifth group of ITTs refers to integrated platforms that consolidate other
specific tools, align their outputs, and link them with other business information
systems to ensure appropriate data integrity. This platform binds other tools together
to interoperate effectively in conjunction with each other towards real-time project
management. The importance of this category becomes more evident in real-time
control systems where incoming data from several sub-projects should be captured,
analyzed, and reported to upper management in a consolidated manner. The most
Effective Application of Information Technology Tools … 723

relevant example of an integrated platform that combines outputs of different applica-

tions in construction projects is Building Information Modelling (BIM). This model-
based system improves information sharing among civil engineers, architects, MEP
(mechanical, electrical, and plumbing) designers, and construction managers [11].
This study examines the importance and effectiveness of these five groups of ITTs
in the real-time management of construction projects.

3 Methods

This study follows a quantitative survey approach to explore experts’ viewpoints on

the effectiveness of each group of ITTs for real-time project management. The target
population includes project management professionals working in principal construc-
tion contracting enterprises in Australia. The data collection tool is an online ques-
tionnaire incorporating two sections; the first section includes demographic questions
and the second section comprises five questions related to the ITT groups identified
from the literature review. Industry experts rate the effectiveness of the following
items: Q1. providing robust analytical software, Q2. providing instant full access of
project teams to an online archive of projects that include all records and document
history, Q3. establishing an integrated platform for collecting, analyzing, and sharing
the performance project information, Q4. providing decision-makers with access to
real-time data and verified project performance information, and Q5. establishing
a knowledge management (KM) database that allows access to classified lessons
learned from previous projects.
After sending invitations to eligible experts, the dataset of the 287 completed
responses was analyzed using statistical tests. The characteristics of the experts who
participated in the survey were analyzed, which revealed that 33% were project
managers, 20% managers of the project management office department, 15% project
planning specialists, 10% were business analysts, and 8% among executives. They
have relevant job tenure in a variety of project management positions. Forty-three
percent more than 15 years of work experience, 25% between 11 to 15 years, 19%
between 6 to 10 years, and 13% between 3 to 5 years. The data analysis approach is
(1) one-sample t-test to examine the effectiveness of each group of ITTs (2) Pearson
correlation analysis to identify the dependency of ITTs using AMOS 25 software
and (3) relative importance analysis to rank ITT groups.
724 M. Ershadi et al.

Table 2. Results of the on-sample t-test in SPSS

ITTs Test value = 3
t df Sig.* Mean difference Interval**
Lower Upper
Q1. Analytical tools 13.808 286 .000 .88502 .7589 1.0112
Q2. Archive tools 12.981 286 .000 .78746 .6681 .9069
Q3. Integrated platforms 15.136 286 .000 .91986 .8002 1.0395
Q4. Reporting tools 14.654 286 .000 .95470 .8265 1.0829
Q5. KM tools 13.761 286 .000 .85366 .7316 .9758
*Significance (2-tailed), **95% confidence interval of the difference

4 Results

4.1 The Significance Level of Variables

The one-sample t-test was applied in this study to examine whether the ITTs identified
earlier were significant to be effectively applied for real-time project management.
This statistical test was carried out in SPSS 25 software (Table 2). The significance
level for all five variables is less than 0.05, which means that according to experts’
viewpoint they are essential tools for real-time management of construction projects.
This t-test revealed that the mean rating of the variables is significantly higher than
the cut-point 3 considering 95% confidence level.

4.2 The Correlation of Variables

A correlation analysis is useful to examine the relationship between ITTs. This

study conducted a Pearson correlation analysis in SPSS to identify whether there
is a significant linear correlation between ITTs, which was confirmed at the 0.01
significance level (Table 3). It can be interpreted that organizations should maintain
the coherence of these ITTs and use all of these categories in an integrated way
to achieve real-time project management. These ITTs can act as sub-systems that
are interrelated and constitute a basis for cooperative data collection, analysis, and
sharing to speed up project management processes.

4.3 Ranking of Variables

The ITTs were ranked based on the ratings given by industry experts so that their
importance level can be identified. The literature asserts that the relative importance
index (RII) can be applied rather than the mean value to provide insight into the
Effective Application of Information Technology Tools … 725

ranking of variables based on a normalized index. This index is worked out using
formula 1 in which w stands for the ratings of ITTs given by experts based on the
Likert scale, A represents the highest weight to be given to each question (5 in the
Likert scale), and N stands for the total number of experts who participated in the
survey [24].

W
RI I = (1)
A×N

The ranking of the variables based on RII revealed that reporting tools are
perceived as the most important group of ITTs, which play an important role as
a coordinator in collecting data from user interfaces, obtain results of the data anal-
ysis from analytical tools, and bring them to the attention of decision-makers. The
integrated platform followed by analytical tools was identified as the second most
important ITT. It is noteworthy that the RII score of these variables is close to each
other and higher than 0.75 which confirms their dependencies (Table 4).

Table 3 Results of Pearson correlation analysis in SPSS

ITTs Q1 Q2 Q3 Q4 Q5
Q1 Pearson correlation 1
Significance
Q2 Pearson correlation .624** 1
Significance .000
Q3 Pearson correlation .714** .602** 1
Significance .000 .000
Q4 Pearson correlation .664** .494** .692** 1
Significance .000 .000 .000
Q5 Pearson correlation .644** .528** .590** .606** 1
Significance .000 .000 .000 .000
** It means that the correlations are significant considering a 0.01 level in 2-tailed mode.

Table 4 The ranking of variables based on RII

Variables Std. deviation Std. error mean Mean RII Rank
Q1. Analytical tools 1.08584 .06409 3.8850 0.777 3
Q2. Archive tools 1.02765 .06066 3.7875 0.758 5
Q3. Integrated platforms 1.02956 .06077 3.9199 0.784 2
Q4. Reporting tools 1.10373 .06515 3.9547 0.791 1
Q5. KM tools 1.05090 .06203 3.8537 0.771 4
726 M. Ershadi et al.

5 Discussion

Adoption of the right information technology tools benefits construction organiza-

tions in fostering innovation and speeding up processes. This study argues that five
main ITTs are essential for real-time project management and facilitates the process
of collecting, analyzing, integrating, reporting, and archiving the project information
to form the organization’s knowledge base. As a result of probing experts’ opinions
and statistical analysis, it was found that the coherency of these ITTs should be
maintained to achieve the real-time flow of performance data across an organiza-
tion. The ranking of these groups using RII revealed that industry experts perceive
reporting ITTs as the most important tools, which is aligned with the studies focused
on reporting tools for managing project information [14].
The next generation of reporting tools seems to emphasize agile techniques and
visual features such as Burn Down Charts [25]. Developers need to consider emerging
demands in real-time project management such as Project Status Inquiry (PSI)
through online dashboards illustrating infographics, graphs, and indicators. Real-
time filtering, sorting, formatting, and updating the performance data help managers
focus on critical data and enable them to drill down for a more detailed review of
tasks. Proper import/export capabilities of reporting tools lead to higher flexibility
in working with a variety of data warehouses as well as analytical software.
The findings also indicated that analytical tools act as central processing units
in a project management system [18]. Some key features of this group of ITTs
include delay analysis, risk analysis, balancing and optimizing resources, analysis,
comparing actual progress with plan progress, auto-updating schedules, forecasting
performance trends, and cash flow analysis. Real-time analytics enable users of the
project management system to draw conclusions on raw performance data imme-
diately upon data availability [26, 27]. Analytical ITTs should analyze time, cost,
scope, quality, risk, and 3D models in an integrated way. Besides, the future genera-
tion of analytical ITTs is expected to focus on visualization and animation techniques
[28].
This paper introduced archiving tools as another major group of ITTs keeping
the records of key documents throughout the project lifecycle to be retrieved and
utilized in the future. This finding is aligned with previous studies that highlighted
the important role of digital archive as a central repository of useful information
[7]. We found that these tools work in conjunction with knowledge management
tools to develop the knowledge asset of a construction organization. Although prior
studies separately focused on the role of document management [17] and knowledge
management tools [10], this study asserted that they should be designed seamlessly to
optimize the process of retrieving knowledge and linking them to associated project
documents. This paper argues that KM tools should function as social media to
allow open discussion of technical teams about effective solutions to challenging
situations and complex problems. This approach enables project managers to learn
from solutions that have been already tested by other teams in the organization.
Effective Application of Information Technology Tools … 727

As another finding, the effective operation of reporting, analytical, archiving, and

KM tools requires an integrated platform to adjust their inputs and consolidate their
outputs. This integration is an indispensable part of modern project management since
the massive flow of data from multiple projects should be canalized and directed in the
correct path to maximize the value of ITTs for supporting management decisions.
This study provided an understanding of the elements essential for an integrated
real-time project management system. However, the findings are limited in terms of
identifying the specific requirements of such an integrated system. Thus, the authors
encourage future research to focus on the configuration of such an integrated system.
Notwithstanding all the research that has been conducted so far and the findings
of the current study, more in-depth studies need to be conducted to offer the best
possible solution for better transmission of information in construction projects.

6 Conclusion

The intense interactions among frequent stakeholders that are involved in construc-
tion projects have led contractors to adopt advanced IT tools for enhancing commu-
nications. This paper aims to examine the importance of ITTs and discuss their appli-
cation for a real-time project management system in this demanding environment.
Five groups of ITTs were examined in this empirical study. Project reporting tools
were perceived as the most effective categories of tools that enable higher visibility
and transparency of information between key stakeholders. Effective monitoring and
control of construction activities require collecting precise data on team performance,
risks, resource productivity, progress, needs, and issues to be addressed promptly.
Analytical tools analyze collected data to be used for planning, control, and manage-
ment of projects. The analytical ITTs that are employed for analyzing data from
multiple projects consider interdependencies among their activities, resources, and
constraints to help optimally plan their resources. This group of ITTs helps to adopt
proper strategies for allocating plant and labor resources to a portfolio of projects.
The outputs of analytical software should help managers to identify the root cause
of issues from a systematic perspective.
This study examined the effectiveness of ITTs in storing and retrieving documents
and knowledge obtained from previous and current projects. Archiving tools act
as a document management system that keeps the record of different versions of
documents for the instant access of project teams. On the other hand, knowledge
management ITTs provide a basis for better transmit of lessons learned both at an
inter-project and intra-project level. The study asserted that these two groups of
tools perform better in conjunction with each other to link knowledge and relevant
digital documents as an approach to integrated record management. Four groups
of ITTs were found to be complementary to each other and interrelated in their
inputs/outputs, which indicates that they need to be consolidated in a single integrated
platform. Integrated platforms help to centralize data and facilitate the exchange and
interoperability of project information. Such platforms nurture standardization and
728 M. Ershadi et al.

consistency in the adoption of ITTs for centralized management and real-time control
of several projects in support of their on-target delivery.

References

1. Fulford R, Standing C (2014) Construction industry productivity and the potential for
collaborative practice. Int J Proj Manag 32(2):315–326
2. Isaac S (2019) What does “real-time” project control really mean? In: European conference on
computing in construction: Chania, Crete, Greece
3. Akinsola A, Dawood N, Hobbs B (2000) Construction planning process improvement using
information technology tools. In: Proceedings of CIT, pp 40–51
4. Yap JBH, Lock A (2017) Analysing the benefits, techniques, tools and challenges of knowledge
management practices in the Malaysian construction SMEs. J Eng Des Technol 15(6):803–825
5. Hartmann T et al (2012) Aligning building information model tools and construction
management methods. Autom Constr 22:605–613
6. Cleveland A Jr (2011) Emerging tools to enable construction engineering. J Constr Eng Manag
137(10):836–842
7. Latimer D, Hendrickson C (2003) Digital archival of construction project information. Nist
Special Publication SP, pp 59–64
8. Azhar S, Brown J, Farooqui R (2009) BIM-based sustainability analysis: an evaluation of
building performance analysis software. In: Proceedings of the 45th ASC annual conference,
vol 1, no 4, pp 276–292
9. Siew RY, Balatbat MC, Carmichael DG (2013) A review of building/infrastructure sustain-
ability reporting tools (SRTs). Smart Sustain Built Environ 2(2):106–139
10. Perez-Araos A et al (2007) Designing a knowledge management tool to support knowledge
sharing networks. J Manufact Technol Manag 18(2):153–168
11. Saka AB, Chan DW (2020) Profound barriers to building information modelling (BIM) adop-
tion in construction small and medium-sized enterprises (SMEs). Constr Innov 20(2):261–284
12. Blackburn DA, et al (2009) Construction imaging and archiving method, system and program.
Google Patents
13. Subramani T, Jabasingh DS, Jayalakshmi J (2014) Analysis of cost controlling in construction
industries by earned value method using Primavera. Int J Eng Res Appl 4(6):145–153
14. Hu W, Panthi K (2020) Project status reporting system (PSRS) for pipeline relocation programs.
Built Env Proj Asset Manag 10(5):693–708
15. Wu S et al (2004) A project knowledge management tool for the construction industry. Int J IT
Archit Eng Constr 2(2):79–90
16. Xu G et al (2018) Cloud asset-enabled integrated IoT platform for lean prefabricated
construction. Autom Constr 93:123–134
17. Zou PX et al (2017) Cloud-based safety information and communication system in infrastruc-
ture construction. Saf Sci 98:50–69
18. Yang J-B, Kao C-K, Lee Y-Y (2006) System requirement analysis of a construction delay
analysis system. In: Proceeding of 23rd international symposium on automation and robotics
in construction, pp 102–106
19. Shen X, Marks E (2016) Near-miss information visualization tool in BIM for construction
safety. J Constr Eng Manag 142(4):04015100
20. Chu SKW (2008) TWiki for knowledge building and management. Online Inf Rev 32(6):745–
758
21. Hsiao DW et al (2009) An integrated platform of collaborative project management and silicon
intellectual property management for IC design industry. Inf Sci 179(15):2576–2590
22. Adhau S, Mittal M, Mittal A (2013) A multi-agent system for decentralized multi-project
scheduling with resource transfers. Int J Prod Econ 146(2):646–661
Effective Application of Information Technology Tools … 729

23. Liaw S-S, Chen G-D, Huang H-M (2008) Users’ attitudes toward Web-based collaborative
learning systems for knowledge management. Comput Educ 50(3):950–961
24. Chan DW, Kumaraswamy MM (1997) A comparative study of causes of time overruns in Hong
Kong construction projects. Int J Proj Manag 15(1):55–63
25. Schuh G, et al (2016) Next generation hardware development: Framework for a tailorable devel-
opment method. In: 2016 Portland international conference on management of engineering and
technology (PICMET). IEEE
26. Chen W-N, Zhang J (2012) Ant colony optimization for software project scheduling and staffing
with an event-based scheduler. IEEE Trans Softw Eng 39(1):1–17
27. Francis A, Miresco E (2014) Case studies for the planning and monitoring of unit-and
fixed-price contracts using project scheduling software. In: Computing in civil and building
engineering, pp 1021–1028
28. Dang T, Bargstädt H-J (2016) 4D relationships: the missing link in 4D scheduling. J Constr
Eng Manag 142(2):04015072
HCODF: Hybrid Cognitive Ontology
Driven Framework for Socially Relevant
News Validation

V. Adithya, Gerard Deepak, and A. Santhanavijayan

Abstract Nowadays everyone gets news from social media and most often it is
very much difficult to validate its credibility. Social media platforms like twitter and
WhatsApp are working on this for a long time, lots of research work has been done
by many people but still an effective technique hasn’t been found out. This paper
proposes a method to address this problem, with the use of a hybrid NLP model that
uses an Event Ontology and LSTM to validate whether a news is authentic or not. The
news dataset that has been used here are the tweets crawled and the news headlines,
where the semantic similarity between the news and tweets on specific events are
found out, which is later used to classify the tweets whether it is fake or not. This
experiment’s performance is evaluated and it is compared with baseline models and
it was found that the proposed approach is superior in terms of performance which
recorded an F-Measure and False Negative Rate of 96.14% and 0.03 respectively.

Keywords Event Ontology · Fake news detection · Jaccard similarity · LSTM ·

Social media

1 Introduction

In this modern world, anyone can publish any information, on the Web although that
is acceptable up to certain extent, but taking it too far such as posting wrong infor-
mation online in order to trigger and manipulate people, causing panic in the society,
spreading rumors about a person to tarnish their public image because of personal
vengeance can lead to everlasting and ever ending repercussions. We cannot put an
end to this problem once for all but by constantly developing and improving existing

G. Deepak (B) · A. Santhanavijayan

Department of Computer Science and Engineering, National Institute of Technology,
Tiruchirappalli, India
V. Adithya
SRM Institute of Science and Technology, Ramapuram, Chennai, India

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 731
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_66
732 V. Adithya et al.

technique’s that address this issue, impact of this problem can be reduced. Machine
Learning was the primary strategy incorporated in most Fake News Detection systems
[1].
Is all the news that we tend to come across on the social media are true? Social
media is the fastest mode of sharing and spreading information that can reach a mass
of people within minutes. In that case spreading fake or unauthentic news via social
media causes a lot of problems. Nowadays many people have come to a conclusion
that all news that comes in social media are false or very old messages that is been
circulated again. It is true in most of the cases, users are confused whether the news
is true or not.

Motivation. Social media platforms like WhatsApp and twitter are working on this
problem for a long time but still an effective technique hasn’t been found for it. The
solution to this problem can be found by using Natural Language Processing (NLP)
which is used to process text data. WhatsApp forward message that keeps circulating
in a circle, where no information of the source is included. In contrast to WhatsApp,
the twitter tweets have the information of the source and does not contain a large
number of sentences, at the same time reach a large number of audiences. Thus, by
utilizing the advantage, the tweets can be crawled easily by filtering the retweets and
can be easily processed.

Contribution. This paper proposes a methodology to identify whether a tweet is fake

or not, using an Ontology driven frame work and compare the similarity with the top
news headlines for the past five years, to its related domain and conclude whether
the news is authentic or not. Experiments show that a high accuracy can be achieved
by using an event Ontology driven framework when compared to some other NLP
or traditional machine learning techniques. This HCODF system has achieved an
overall accuracy of 96.16% where the experiments were focused on last five years
top news headlines.

Organization. The latter portion of the paper is divided into the following sections.
Section 2 addresses briefly, the relevant research done related to this topic. Section 3
explains the planned system architecture. Section 4 consists mainly of the imple-
mentation and performance evaluation. Section 5 consists of conclusion drawn from
the experiment.

2 Related Work

Jain et al. [1] have put forth an NLP and machine learning model for fake-news
detection, which uses SVM and Naïve Bias with an accuracy of 93.6%. Seo et al.
[2] have proposed a methodology that uses question and answering system, which
uses modified CNN deep learning model and they also say that their model archives
high accuracy when compared to statistical models. The model focuses on data
HCODF: Hybrid Cognitive Ontology Driven Framework … 733

augmentation batch size control and model modification. Samantaray et al. [3] have
divided the proposed model into two parts, text analysis and performance evaluation,
text similarity approach by using N-Grams, TF-IDF and cosine similarity, corpus-
based similarity and term-based similarity matchings have been used. Bahada et al.
[4] have used Bi-directional LSTM-Recurrent Neural Network model applied on two
publicly present unstructured news articles for fake news detection and say that it
produces high accuracy when compared with CNN model, vanilla-RNN model and
unidirectional-LSTM model for the same task. Ajao et al. [5] have used a hybrid
CNN and RNN models for fake news detection on twitter. The proposed architecture
issued an accuracy of 82%. Authenticity has been validated based on the contents
of the messages, semantic attributes, linguistic characteristics similar to that of the
fake news. In [6–16] several ontology models and applications have been discussed
in support of the proposed methodology.

3 Proposed System Architecture

The architecture of the proposed fake new detection HCODF system is as shown in
the Fig. 1. The proposed HCODF system consists of 3 phases, the 1st phase is slot
filling using LSTM [17] event classifier for the news database, 2nd phase is Event
Ontology creation for the twitter data and the 3rd phase is comparing both with
Jaccard Similarity.
The main purpose of this HCODF system is to find out the similarity between the
Event Ontology that is created for the tweets present in the twitter dataset and the
slot filling, which is done for the news article for the particular event by the LSTM
event classifier model which is used to classify whether it is authentic or not. The

Fig. 1 Proposed system architecture diagram of the HCODF model

734 V. Adithya et al.

first phase involves pre-processing of the raw news headlines data, by using a bag
of words model, the headlines are tokenized into a set so that the sentence structure
is lost and then these sets are combined using a union operation and lowercase
conversion, number to text conversion, punctuations removal, stop word removal,
word stemming, lemmatization is done. Then the pre-processed data is ready for the
Term Frequency-Inverse Document Frequency TF-IDF.
TF-IDF is a technique, generally used for Information retrieval, and is used to
calculate the significance of every word in the document and rank them accordingly
so that the relevancy of each word in the document is found out which can be used
as features for the dataset. The Eq. (1) represents the formulation of TF-IDF as a
product of Term Frequency and Inverse document Frequency.

TF − IDF = Term Frequency × Inverse Document Frequency (1)

Term Frequency is calculated by dividing the number of occurrences of a term t in

the document divided by the number of words in that document as shown in Eq. (2).

occurrence of term t in the document d

TF(t, d) = (2)
number of terms in the document
Document Frequency is the number of occurrences of term t in the whole set of
documents and the inverse of this is multiplied with the total number of documents.
This helps in giving low ranks to unnecessary common words (“is”, “was”, “there”).

N
IDF(t) = (3)
DF
N
IDF(t) = log( ) (4)
DF + 1
N
TF − IDF(t, d) = TF(t, d) × log (5)
DF + 1

Equation (3), (4) shows how Eq. (5) the final TF-IDF equation is derived.
After TF-IDF is encompassed, the data is used to train the LSTM such that it
generates keywords for the events that is used for slot filling. The LSTM is also
kind of RNN architecture that has feedback connections. It can process the whole
sequence of the data. Normally, a LSTM unit consists of a cell, which is used to
store values across time periods and three gates namely, input, output and forget
gates. These gates are utilized to control the flow of information into the cell and
vice versa. LSTM’s are used for classification and for prediction using time-series
data. In a similar manner, the LSTM used here is used to classify the type of the
event that the news headlines is talking about which is used later for filling the slots.
The twitter tweets are crawled for each event that has been tweeted between a
particular duration of time in English and the retweets are filtered and is saved with
its timestamp. The tweets are crawled using a tweet crawling api called as twint. The
HCODF: Hybrid Cognitive Ontology Driven Framework … 735

Fig. 2 Small visualization of the architecture’s event ontology model

twitter data is then pre-processed by performing tokenization, lower case conversion,

number to text conversion, punctuation removal, stop word removal, stemming, and
lemmatization. Ontology can be defined as a framework to represent information.
For this pre-processed dataset, Event Ontology model is created which defines the
event type and the event names are added as individuals. Figure 2 shows a small
visualization of the event ontology model used in the architecture.
The third phase of this HCODF system architecture is to compare the Ontology
and the slot filled words and say whether the news is true or not. The similarity
function that is used here is called Jaccard Similarity. Jaccard Similarity function
also called as intersection over union. This function is used to find the text similarity
between two documents.
A ∩B
Jaccard Similarity = (6)
A∪B
Equation (6) represents Jaccard similarity where A and B are two sets containing
text. It returns a float value between 0 and 1 and multiplying it by 100 will give the
percentage of similarity. Here we use 75% as the threshold, any value above 75%
will be classified as true and less than that will be considered as false.
736 V. Adithya et al.

4 Implementation and Performance Evaluation

The proposed HCODF system’s architecture was designed and successfully imple-
mented on Windows 10 operating system. The implementation has been performed
using Intel Core i5 8th Gen processor and 8 GB RAM. NLTK, Re, num2words and
sklearn are the python NLP tool libraries that has been used for pre-processing data
and performing TF-IDF. Keras, a python’s deep learning API library running on top
of TensorFlow has been utilized to implement and train the LSTM Event Classifier
model. Web Protege an open-source online Ontology editor created by the Protege
has been used to create and visualize ontologies for all the events.
The experiment was performed on the top news headlines of India dataset, for
the past 19 years, a standard dataset published in Kaggle of which the past 5 years
data has been utilized and some data set has been crawled using news crawler api
for some specific events and both are combined. The tweets were crawled for events
with a twitter crawler api called as twint by using the event’s respective hash tag
as keywords and the retweets are filtered and the data is stored with its time stamp.
The Precision, Recall, Accuracy, F-Measure and False Negative Rate are calculated,
and used as performance evaluation metrics. The Accuracy, F-Measure, and False
Negative Rate are calculated and represented as Table 1 and Table 2 using the Eqs. (7),
(8), (9), (10), (11) respectively.

True number of Positives

Precision % = (7)
True number of Positives + False number of Positives
True number of Positives
Recall % = (8)
True number of Positives + False number of Negatives

Table 1 Performance of the Proposed HCODF system for some Viral Events
Viral News Precision % Recall % Accuracy % F-Measure % FNR
Pulwama Terror Attack 87.69 90.12 88.91 88.89 0.10
Demonetization in India 92.63 96.78 94.71 94.66 0.03
Cyclone Vardha 90.14 92.63 91.39 91.37 0.07
Cyclone Winston 89.43 94.43 91.93 91.86 0.06
Amazon Forest Wildfire 91.69 96.47 94.08 94.02 0.04

Table 2 Performance of HCODF in comparison with other Baseline Approaches

Method Precision % Recall % Accuracy % F-Measure % FNR
Jain et al., [1] 92.04 94.04 93.04 93.03 0.06
Seo et al., [2] 85.43 87.43 86.43 86.42 0.13
Samantaray et al., [3] 90.41 93.04 91.72 91.71 0.07
Proposed HCODF 94.89 97.43 96.16 96.14 0.03
HCODF: Hybrid Cognitive Ontology Driven Framework … 737

Precision + Recall
Accuracy % = (9)
2
2(Precision × Recall)
F − Measure % = (10)
(Precision + Recall)

False Negative Rate = 1 − Recall (11)

The experiment was trained on several events. For this experimentation purpose, 5
events are considered. The Table 1 shows the performance evaluation for 5 individual
events namely, Pulwama Terror Attack, Demonetization in India, Cyclone Vardha,
Cyclone Winston, Amazon Forest Wildfire. It can be observed that the Precision,
Recall, Accuracy, F-Measure And False Negative Rate for Pulwama attack is 87.6,
90.12, 88.91, 88.89% and 0.10 respectively, for Demonetization in India it is 92.63,
96.78, 94.71, 94.66% and 0.03 respectively, for Cyclone Vardha it is noted to be
90.14, 92.63, 91.39, 91.37% and 0.07 respectively, for Cyclone Winston it is 89.43,
94.43, 91.93, 91.86% and 0.06 respectively and for Amazon Forest Wildfire it is
91.69, 96.47, 94.08, 94.02% and 0.04 respectively.
For this implementation, 3 existing approaches are considered as baseline as
shown in Table 2. The first approach proposed by Jain et al. [1] uses traditional NLP
with SVM and Naive Bayes and our HCODF model’s performance increase is when
compared to it is 2.85% in precision, 3.39% in recall, 3.12% in Accuracy, 3.11%
in F-measure and 0.03 in False Negative Rate. The second approach proposed by
Seo et al. [2] uses modified convolution neural networks to perform the same task
and our HCODF model’s performance-increase when compared to it is 9.46% in
Precision, 10% in Recall, 9.73% in Accuracy, 9.71% in F-measure and 0.10 in False
Negative. The Third approach proposed by Samantaray et al. [3] a hybrid of TF-IDF,
N gram vectors and cosine similarity. HCODF model’s performance-increase when
compared to it is 4.75% in Precision, 4.93% in Recall, 4.44% in Accuracy, 4.43% in
F-Measure and 0.04 in False Negative Rate.
The superiority of the HCODF system when compared to the baseline model
in terms of performance is because it uses a hybridization both LSTM and an
Event Ontology model. The HCODF model also uses Jaccard similarity instead
of cosine similarity when compared to the third approach, which is also responsible
for increase in its performance. Since the HCODF uses LSTM when compared to
the first approach that uses traditional NLP and the second approach that uses CNN,
the increase is performance is justified. The Event Ontology model adds secondary
knowledge to the system when compared to using similarity function without an
Ontology. Which is the primary reason for improvement in performance.
738 V. Adithya et al.

5 Conclusions

There is a necessity to validate the authenticity of the news that is published in social
media because the news spreads to a greater number of people when compared to any
other mode. This paper discusses an effective method for the same. This proposed
HCODF system can perform accurate predictions when compared to others because
of the Ontology driven framework, where an Event Ontology model is combined
with LSTM and used as a hybrid. Based on this, it can be concluded that in future,
ensuring techniques with hybrid approaches combined with Ontology driven frame
works can produce better results for the same purpose fulfilment. An overall F-
Measure of 96.14% with a very low FDR of 0.03 was obtained. As a future work
the proposed HCODF can be experimented with other fusion models of machine
learning, deep learning, and other similarity models.

References

1. Jain A, Shakya A, Khatter H, Gupta AK (2019) A smart system for fake news detection using
machine learning. In: 2019 2nd international conference on issues and challenges in intelligent
computing techniques (ICICT)
2. Seo Y, Seo D, Jeong C-S (2019) FaNDeR: fake news detection model using media reliability. In
Proceedings of TENCON 2018 - 2018 IEEE region 10 conference, pp 1834–1838). [8650350]
(IEEE Region 10 annual international conference, proceedings/TENCON, vol 2018-October.
Institute of Electrical and Electronics Engineers Inc. https://doi.org/10.1109/TENCON.2018.
8650350
3. Samantaray SD, Jodhani G (2019) Fake news detection using text similarity approach. Int J
Sci Res (IJSR) 8(1):1126–1132
4. Bahada P, Saxenaa P, Kamalb R (2019) Fake news detection using bi-directional LSTM-
recurrent neural network. In: International conference on recent trends in advanced computing
2019, Icrtac 2019
5. Oluwaseun A, Deepayan B, Shahrzad Z (2018) Fake news identification on Twitter with hybrid
CNN and RNN models. SMSociety, Copenhagen, Denmark, July 2018
6. Gerard D, Santhanavijayan A (2020) OntoBestFit: a best-fit occurrence estimation strategy for
RDF driven faceted semantic search. Comput Commun 160:284–298
7. Pushpa CN, Deepak G, Kumar A, Thriveni J, Venugopal KR (2020) OntoDisco: improving web
service discovery by hybridization of ontology focused concept clustering and interface seman-
tics. In: 2020 IEEE international conference on electronics, computing and communication
technologies (CONECCT). IEEE, pp 1–5
8. Kumar A, Deepak G, Santhanavijayan A (2020) HeTOnto: a novel approach for concep-
tualization, modeling, visualization, and formalization of domain centric ontologies for heat
transfer. In: 2020 IEEE international conference on electronics, computing and communication
technologies (CONECCT). IEEE, pp. 1–6
9. Deepak G, Kasaraneni D (2019) OntoCommerce: an ontology focused semantic framework
for personalised product recommendation for user targeted e-commerce. Int J Comput Aided
Eng Technol 11(4–5):449–466
10. Gulzar ZA, Leema A, Deepak G (2018) PCRS: personalized course recommender system based
on hybrid approach. Procedia Comput Sci 125:518–524
11. Deepak G, Teja V, Santhanavijayan A (2020) A novel firefly driven scheme for resume parsing
and matching based on entity linking paradigm. J Discrete Math Sci Crypto 23(1):157–165
HCODF: Hybrid Cognitive Ontology Driven Framework … 739

12. Haribabu S, Sai Kumar PS, Padhy S, Deepak G, Santhanavijayan A, Kumar DN (2019) A
novel approach for ontology focused inter-domain personalized search based on semantic set
expansion. In: 2019 Fifteenth International Conference on Information Processing (ICINPRO),
Bengaluru, India, 2019, pp 1–5. https://doi.org/10.1109/icinpro47689.2019.9092155
13. Deepak G, Naresh Kumar G, Sai Yashaswea Bharadwaj VSN, Santhanavijayan A (2019) Onto-
Quest: an ontological strategy for automatic question generation for e-assessment using static
and dynamic knowledge. In: 2019 fifteenth international conference on information processing
(ICINPRO). IEEE, pp 1–6
14. Santhanavijayan A, Naresh Kumar D, Deepak G (2021) A semantic-aware strategy for auto-
matic speech recognition incorporating deep learning models. In: Intelligent system design.
Springer, Singapore, pp 247–254
15. Gerard D et al (2019) Design and evaluation of conceptual ontologies for electrochemistry as a
domain. In: 2019 IEEE international WIE conference on electrical and computer engineering
(WIECON-ECE). IEEE (2019)
16. Deepak G, Priyadarshini JS (2018) Personalized and enhanced hybridized semantic algorithm
for web image retrieval incorporating ontology classification, strategic query expansion, and
content-based analysis. Comput Electr Eng 72:14–25
17. Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9:1735–1780.
https://doi.org/10.1162/neco.1997.9.8.1735
Understanding Customer Adoption
of E-banking

Jaride Chama, Taqi Ahmed, and El Hachimi Mohamed Amine

Abstract Financial institutions have been forced to make investments to maintain

and improve customer satisfaction through E-banking offers. This electronic channel
has demonstrated its importance by changing the traditional concept of financial
institutions. By understanding the main factors that encourage customers to use
electronic banking services, this helps to further develop this communication channel
and take advantage of its benefits by allowing banks to remain competitive and
profitable over the long term. This work is a state of the art which attempts to
answer, through previous work in the field of E-banking, the following question:
What are the factors influencing the adoption of electronic banking services among
banking customers? Identifying these factors is an essential objective to change the
characteristics of the service in order to make its adoption more attractive and increase
its level of use. The results obtained confirm the presence of several recurring factors
in the literature such as facilitating conditions, trust, perceived usefulness and social
influence. By clearly identifying the most influential adoption factors, this enabled
us to formulate a set of recommendations to the practitioners of E-banking.

Keywords E-banking · Electronic banking · Adoption factors · Technology

1 Introduction

The banking sector is one of the sectors that has taken a strong interest in digitization.
Indeed, the speed of growth of technology has more influence on banks sector than
on others [1]. In addition to that, the international economic crisis and the resulting
crisis of confidence in the markets has affected several financial institutions. It has
pushed banks to try to maintain customers relationship in different ways [2].

J. Chama (B) · T. Ahmed

Faculty of Juridical, Economic and Social Sciences, Abdelmalek Essaâdi University, Tétouan,
Morocco
e-mail: [email protected]
E. H. M. Amine
National School of Management Kenitra, Ibn Tofail University, Kenitra, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 741
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_67
742 J. Chama et al.

Today, Electronic Banking service is one of the most important E-commerce

business around the world. Indeed, E-banking activities represent an important part
in customers retail activities which is constantly increasing [3]. Therefore, banks
should give high priority to behavioral factors by considering them as key factors,
important for the successful implementation of online banking services [4].
Several researchers have studied the antecedents of technology adoption among
consumers. Models such as Theory of Reasoned Action (TRA), Theory of Planned
Behavior (TPB), Technology Acceptance Model (TAM), Diffusion of Innovations
model (DOI), Unified Theory of Acceptance and Use of technology (UTAUT)
and Information Systems Success Model (ISSM) have been used repeatedly by
researchers to study this topic. Although these different models converge in their
design, they have in common the same purpose which is to better understand and
identify the main drivers of technology adoption. This work presents first a brief
review of the existing literature before addressing the main factors of E-banking
adoption in previous work and then formulate some recommendations in favor of the
practitioners of this field of research.

2 Literature Review

E-banking is one of the services offered by banks to provide quick access to users
through electronic channels. With the development of computer technology, elec-
tronic banking services are increasingly accepted by customers who strongly support
new technologies [5]. They can be used to attract new customers because they allow
the transmission of information in real time and provide a convenient channel for
communication between the bank and its customers. Because of this, electronic
banking services encourage customers to perform financial operations more easily
and efficiently through electronic channels.
Asad, Mohajerani and Nourseresh [6] has conducted research that attempts to
explain the factors affecting customer satisfaction in electronic banking and to estab-
lish priority factors. According to these authors, several factors affect customer
satisfaction, such as usefulness, trust, security, web design, information content and
responsiveness through customer support services that must be provided.
In their turn, Ling, Fern, Boon and Huat [5] conducted a study in Malaysia
through a questionnaire survey of 200 customers who have experience using elec-
tronic banking services. Their aim was to examine the factors that influence customer
satisfaction with these services. The five factors identified by the authors are service
quality that is formed when the customer makes a comparison between the expected
service and the actual service experience, web design and content related to the
configuration of colors and graphics, security and confidentiality which is a form of
protection to ensure customers safety, convenience that corresponds to usefulness
and speed associated with network connection failures and internet banking website
browsing time.
Understanding Customer Adoption of E-banking 743

Aliyu, Rosmain and Takala [4] also developed a conceptual model based on several
previous studies related to electronic banking services, behavioral factors, banking
applications and customer service delivery. The authors conducted a hypothetico-
deductive study by administering an online questionnaire to 283 clients in Malaysia.
Their results revealed that security, cost and convenience are the major drivers of
customer service delivery in electronic banking. The study also demonstrated the
significant link between customer satisfaction and the use of electronic banking
services.
The aim of the study conducted by Al-Somali, Gholami and Clegg [7] was to
answer a set of questions, to know what factors influence the trend of customers
to accept and use online banking services as their primary banking channel? What
is the relative importance of these factors? and what is the nature and strength of
the relationship between them? The research structure was developed on the basis
of the Technology acceptance Model by integrating new control variables and then
tested with 202 clients in Saudi Arabia. The authors’ conceptual model suggests that
awareness of online banking services and their benefits is a critical component of
E-banking adoption as much as customers may not be familiar with them because
it is a new experience for many of them. Trust is a bigger issue in online banking
than in traditional branch. It is very important in the online banking industry because
by nature transactions contain sensitive information for customers. According to the
authors, social influence also plays a key role in the acceptance and behavior of
E-Banking users, it corresponds to the subjective standards defined by the opinions
of consumers entourage. The study also highlights the presence of the resistance to
change factor, that is the shift from traditional banking to online banking for some
customers who are not ready to change their current ways of operating. Of course,
demographic factors such as age, gender or level of education cannot be overlooked.
According to this study, they also have an impact on customer attitudes and behaviors
towards online banking services.
In Germany, Jünger and Mietzner [3] looked at the factors that determine the deci-
sion of households to switch from traditional banks to new financial technologies.
The authors’ results show the dominance of three main dimensions which are trust,
transparency and financial expertise. For Jünger and Mietzner, these factors signif-
icantly influence household decisions to stay or switch from traditional banking
services to new financial technologies.

3 Main Factors of E-banking Adoption

3.1 Ease of Use and Facilitating Conditions

It cannot be denied that using electronic banking services requires a minimum of

skills, such as using an electronic device, internet browsing or even handling an appli-
cation. Electronic transactions can seem complicated for many customers. Therefore,
744 J. Chama et al.

the ease felt when using a new technology will encourage its adoption. In fact, this
is the second significantly important determinant in technology adoption within the
Technology Acceptance Model. It refers to the degree to which the potential user
expects the use of the system to be effortless [8]. So, it is about the perceived ease
when using technology. Ease of use of electronic channels should therefore be sought
as a necessary determinant for the use of E-banking services. In addition, bad design
and aesthetics can prevent the user from revisiting the site [6].

3.2 Perceived Usefulness and Performance Expectancy

Perceived usefulness is the major determinant of individuals’ intention to use tech-

nology in the Technology Acceptance Model. According to Davis, Bagozzi and
Warshaw [8], this is the subjective probability of potential user that the use of a system
will increase its performance. Performance expectancy is the strongest predictor of
behavioral intention in the Unified Theory of Acceptance and Use of Technology
model of Venkatesh, Thong and Xu [9]. It is defined by these authors as the degree to
which the use of a technology will provide benefits to consumers during the execution
of a specific activity. Indeed, it is when the user finds performance gains following
the use of E-banking services that he decides to adopt them [10, 11].

3.3 Trust

Trust is a key element of success that is present in the banking sector in both traditional
and electronic channels. Previous research has shown that trust is central to build long-
term relationships between customers and companies, especially in risky electronic
banking environments [12]. Indeed, lack of trust is one of the most frequently reasons
cited by consumers who do not use electronic banking services in numerous studies
around the world [12–18].

3.4 Social Influence

Social influence is also an E-banking adoption factor which is very present in many
research studies. According to Baabdullah, Alalwan, Rana, Kizgin and Patil [19],
it is precisely the extent to which an individual perceives that important people
believe he should apply the new system. According to the authors, the preferences
and values of those around us, family members, friends, neighbors and other users
of the technology, tend to change the views and perceptions of E-banking users.
Understanding Customer Adoption of E-banking 745

3.5 Cost and Price Value

In terms of cost, several authors have stated that to encourage consumers to use new
technologies, their adoption must be reasonably priced compared to the prices of
other alternatives. The notion of price value present in the literature is defined as the
cognitive compromise of consumers between the perceived benefits of E-banking
and the monetary cost of adopting this service. With a high level of price value,
customers become more enthusiastic about continuing to use E-banking technology
[13].

3.6 Behavioral Intention

Behavioral intention is a predictor of technology adoption that has been widely and
repeatedly reported to have an important role in shaping actual use and adoption of
new systems [13]. This is a factor present in many technology adoption models such
as TAM or UTAUT and UTAUT2. Defining behavioral intention is quite simply to
answer the following question: is the client initially willing to accept and use a new
technology?

4 Recommendations

Based on the results of previous work which enabled us to identify the main adoption
factors mentioned above, we can say that today’s customers have higher requirements
towards E-banking. According to Hamidi and Safareeyeh [20], they are more knowl-
edgeable, want to get more products and more valuable information. To this end, we
can make some recommendations to practitioners in the field of E-Banking that they
can implement to support the adoption of their electronic channels.
First of all, we strongly recommend increasing the ease perceived by customers
through clear and comprehensive electronic support as well as the establishment of
customer support. Then, banking institutions must ensure the strengthening of their
electronic banking security system in order to create and maintain customer’s trust. It
is also advisable to minimize costs when customers use E-banking services in order
to increase the price value and thereby increase its use.

5 Conclusion

Due to advanced technology and innovation in the financial and banking sectors,
electronic banking services have become more familiar to customers of traditional
746 J. Chama et al.

banks [5]. Indeed, many customers are opting more and more for the novelty of
E-Banking, given the many advantages that it offers.
In order to increase the use of this electronic channel, this research has enabled us
to highlight the most common E-banking adoption factors in the literature, namely
ease of use, facilitating conditions, perceived usefulness, performance expectancy,
trust, social influence, cost, price value and behavioral intention.
However, by removing the social component of business exchange which is the
heart of relationship marketing, information and communication technologies can,
in some cases, degrade the relationship between the customer and the company [21].
In the case of E-Banking for example, customers can easily consult other websites
that offer similar services or products and this means that customer loyalty would
no longer be an easy task [19]. We can therefore ask ourselves if these technologies
really contribute to the development of the relationship with the customer and the
differentiation between banking brands in an environment where all competitors use
the same technologies.

References

1. Kirakosyan K, Dănăiaţă D (2014) Communication Management in Electronic Banking. Better

Communication for Better Relationship | Elsevier Enhanced Reader. Procedia Social and
Behavioral Sciences
2. Liébana-Cabanillas F, Nogueras R, Herrera LJ, Guillén A (2013) Analysing user trust in
electronic banking using data mining methods. Expert Syst Appl 40:5439–5447
3. Jünger M, Mietzner M (2019) Banking goes digital: The adoption of FinTech services by
German households. Finance Research Letters
4. Aliyu AA, Rosmain T, Takala J (2014) Online banking and customer service delivery in
Malaysia: data screening and preliminary findings. Procedia Soc Behav Sci 129: 562–570.
2nd International Conference on Innovation, Management and Technology Research
5. Ling GM, Fern YS, Boon LK, Huat TS (2016) Understanding customer satisfaction of internet
banking: a case study in Malacca. Procedia Econ Finance 37:80–85. The Fifth International
Conference on Marketing and Retailing (5th INCOMaR) 2015
6. Asad MM, Mohajerani NS, Nourseresh M (2016) Prioritizing factors affecting customer satis-
faction in the internet banking system based on cause and effect relationships. Procedia Econ
Finance 36:210–219. 1st International Conference on Applied Economics and Business
7. Al-Somali SA, Gholami R, Clegg B (2009) An investigation into the acceptance of online
banking in Saudi Arabia. Technovation 29:130–141
8. Davis FD, Bagozzi RP, Warshaw PR (1989) User acceptance of computer technology: a
comparison of two theoretical models. Manage Sci 35:982–1003
9. Venkatesh V, Thong JYL, Xu X (2012) Consumer acceptance and use of information
technology: extending the unified theory of acceptance and use of technology. MIS Q 36:157
10. Boonsiritomachai W, Pitchayadejanant K (2017) Determinants affecting mobile banking adop-
tion by generation Y based on the unified theory of acceptance and use of technology model
modified by the technology acceptance model concept. Kasetsart J Soc Sci (2017)
11. Oliveira T, Faria M, Thomas MA, Popovič A (2014) Extending the understanding of mobile
banking adoption: when UTAUT meets TTF and ITM. Int J Inf Manage 34:689–703
12. Rajaobelina L, Brun I, Prom Tep S, Arcand M (2018) Towards a better understanding of
mobile banking: the impact of customer experience on trust and commitment. J Financ Serv
Mark 23:141–152
Understanding Customer Adoption of E-banking 747

13. Alalwan AA, Dwivedi YK, Rana NP (2017) Factors influencing adoption of mobile banking
by Jordanian bank customers: extending UTAUT2 with trust. Int J Inf Manage 37:99–110
14. Lee KC, Chung N (2009) Understanding factors affecting trust in and satisfaction with mobile
banking in Korea: a modified DeLone and McLean’s model perspective. Interact Comput
21:385–392
15. Masrek MN, Mohamed IS, Daud NM, Omar N (2014) Technology trust and mobile banking
satisfaction: a case of Malaysian consumers. Procedia Soc Behav Sci 129:53–58. 2nd
International Conference on Innovation, Management and Technology Research
16. Mostafa R (2015) Investigating the role of trust in mobile banking acceptance, in: ideas in
marketing: finding the new and polishing the old, pp 834–842
17. Sharma SK, Sharma M (2019) Examining the role of trust and quality dimensions in the actual
usage of mobile banking services: an empirical investigation. Int J Inf Manage 44:65–75
18. Merhi M, Hone K, Tarhini A (2019) A cross-cultural study of the intention to use mobile
banking between Lebanese and British consumers: extending UTAUT2 with security, privacy
and trust. Technol Soc 59:101151
19. Baabdullah AM, Alalwan AA, Rana NP, Kizgin H, Patil P (2019) Consumer use of mobile
banking (M-Banking) in Saudi Arabia: towards an integrated model. Int J Inf Manage 44:38–52
20. Hamidi H, Safareeyeh M (2019) A model to analyze the effect of mobile banking adoption on
customer interaction and satisfaction: a case study of m-banking in Iran. Telematics Inform
38:166–181
21. Ailli S (2016) Impact des Technologies de l’Information et de la Communication sur le
marketing relationnel: une évaluation des opérateurs des banques commerciales au Maroc.
Revue Marocaine de recherche en management et marketing 19
Amalgamation of Novel Objective
Function and Multi-sink Solution
for a Reliable RPL in High Traffic
Monitoring Applications

Abdelhadi Eloudrhiri Hassani, Aicha Sahel, and Abdelmajid Badri

Abstract Recent monitoring applications are based on Low power and Lossy
networks (LLN), standardized by Internet Engineering Task Force (IETF). These
kinds of networks require a trustworthy routing protocol for nodes aggregation and
good QoS achievement. RPL gains popularity compared to other routing protocols
for its flexibility via its objective function. Default objective functions implemented
in the RPL core are based on a single metric. Consequently, the routing protocol
can’t cope with different constraints and show some issues in high traffic applica-
tions. For that, we proposed in our paper a cross approach between combined metrics
linearly and lexically for the novel objective function supported by multiple sinks
solution for network monitoring. The proposal, evaluated under the COOJA emulator
against Standard-RPL, showed a reliability improvement by increasing the PDR and
throughput, decreasing considerably the power consumption and giving stability to
the network through low number of DIO control packets.

Keywords WSN · LLN · RPL · Objective function · Cooja

1 Introduction

Data monitoring using the wireless sensor networks, considered as a great connec-
tivity potential, are tiny sensor devices [1] based networks with constraint and limited
resources [2]. Consequently, there is a need for appropriate resources management
and utilization to cope with different applications as smart grids [3], smart cities [4],
industries [5] and healthcare [6]. Thereby, in order to ensure performances efficiency
of those networks, the researchers have been interested mostly in routing protocols.
In this context, the most commonly used is RPL designed by Internet Engineering
Task Force (IETF) [7]. The Ipv6 Routing Protocol for Low power and lossy networks
(RPL) is based on IPv6 and uses the IEEE802.15.4 at the PHY and MAC layers [8].
Packets based on IPv6 cannot fit in the IEEE 802.15.4 protocol. For this reason,

A. Eloudrhiri Hassani (B) · A. Sahel · A. Badri

EEA&TI Laboratory, Faculty of Sciences and Techniques, Hassan II University,
Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 749
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_68
750 A. Eloudrhiri Hassani et al.

6LowPAN, IPv6 Low Power Personal Area Network layer [9], acts as an adaptation
between MAC and network layers. With RPL, the paths are constructed once the
network is initialized. The nodes aim to set up destination-oriented directed acyclic
graph (DODAG), a tree routing topology, using four principal ICMPv6 messages:
DODAG Information Object (DIO) holds information that enable nodes to know
the instance, configuration and preferred parent, it managed by the trickle algo-
rithm and necessary for Multiple Points to Point (MP2P) and Point-to-Point (P2P)
communications [10]. DODAG Information Solicitation (DIS) as a DIO request from
neighbors. DODAG Advertisement Object packets (DAO) used to collect downward
topology informations and DAO-ACK as a response to a DAO message. The aim of
DODAG topology is to steer data packets to one or multiple sinks. Routing paths
are created using an objective function (OF). In the core of RPL, two OF’s are
proposed namely the minimum rank with hysteresis objective function (MRHOF)
[11] based on expected transmission count as a routing metric and OF0 [12], which is
based on hop count. However, the two OF tend to minimize the cost of their metrics
which causes some paths optimization issues due to taking into account a single
constraint. Furthermore, in high traffic applications case, routing to a single sink
may show limits in terms of reliability and ensuring the nodes lifetime. Considering
these drawbacks, we have proposed an amalgamation between novel designed objec-
tive function, which uses novel strategies for metrics combination, and multiple sinks
solution for the network monitoring. The rest of this paper is organized as follow. In
Sect. 2, we present the related works. In Sect. 3, we present the proposed solution
for RPL improvement. In Sect. 4, we report the performance results and discussion,
finally a conclusion is given in Sect. 5.

2 Related Works

In this section, we present some related researches to our work that improve the
objective functions for RPL. In [13], San Martin et al. proposed an extended metric
based on the expected transmission count called Sigma-ETX which is the standard
deviation value of ETX by number of hops to sink. This technique avoids the long
hops issue that causes bottlenecks in high network densities. Also ELT metric, i.e.
Expected Lifetime, was designed by Iova et al. in [14] that consist of the node
remaining lifetime. It is injected in RPL based on multipath, which leads to enhance
the reliability despite an additional delay induced. Minimizing the delay of receiving
packets by the sink was considered by Gonizzi et al. in [15] using the AVGDELAY
as a routing metric. This technique showed a decrease in terms of end-to-end delay
but don’t address the network reliability at all. The fuzzy logic approach was consid-
ered by Bhandari et al. in [16]. Indeed, the authors proposed COOF, i.e. a novel
objective function context-oriented objective function, that considers two designed
metrics respectively remaining energy (RE) and queue fluctuation index for smart
cities applications. Using the same approach in [17], Lamaazi et al. fuzzy designed
EC-OF. It combines ETX and node energy consumption while hop count as node
Amalgamation of Novel Objective Function … 751

redirection metric. Results showed that the proposal improves the RPL performances
in terms of packet delivery ratio, latency, convergence, power consumption including
the network lifetime against MRHOF. Sankar et al. in [18] proposed Fuzzy Logic
Based Energy Aware for RPL objective function based on residual node energy,
expected transmission count ETX and traffic load. It is used for calculating the rank
increase parameter for nodes rank. The proposal shows a increase of PDR and life-
time. Congestion problem was addressed in [19] by proposing Congestion-Aware
Objective Function. Al-Kashoash et al. have used the buffer occupancy in conjunc-
tion with the expected transmission count metric. The proposed objective function
provides high packet delivery ratio and throughput while it decreases the energy
consumption. In this context, Eloudrhiri Hassani et al. In [20] designed Forwarding
Traffic Consciousness objective function, to cope with congestion due to the high
number of forwarded packets, which combines hop count, rssi and a newly designed
FTM metric. The proposed method showed a packet delivery ratio increase respec-
tively with 2 and 11% in low and high traffics, considerably reduces the power
consumption with approximately 47% as well as it achieves a good balance of traffic
managed by the relay nodes. The additive and lexical approaches were also the base
of several objective function optimizations. Indeed, Eloudrhiri Hassani et al. in [21]
proposed a new objective function based on additive combination of node and link
metrics respectively ETX and Energy consumption along the routing paths. They
proposed two way of combination, i.e. with weights and non weighted methods.
Results show that the WCMOF and NWCMOF can increase the reliability, maxi-
mize the network life time and reduce the parent changes. Smart grid application
requirements were considered by Nassar et al. in [22]. The authors designed OFQS
objective function based on power state metric, expected transmission count and
delay. The stability was considered by the hysteresis concept as in MRHOF. Results
showed improvement in terms of end-to-end delay, network lifetime and packet
delivery ratio. In [23], Al-Kashoash et al. have been interested in the congestion of
the path caused by the buffer nodes occupancy. For that purpose, Congestion-Aware
Objective Function were be proposed which consider ETX metric at a low data rate
while the buffer occupancy is considered at a high data rate.

3 Proposed Contribution

The quality of services offered by a routing protocol for IoT based on wireless sensor
networks is one of the substantial design concerns. Besides, its multiple requirements
meeting allow it to be widely used for different applications. As mentioned before, the
RPL objective function is mainly responsible for satisfying the application prerequi-
sites according to the constraint metrics used in the preferred parents selection. In a
standard manner, RPL employs an objective function with one metric to minimize.
It leads to cope with limited constraints, which certainly inflict poor performances in
term of various requirements. Furthermore, the RPL protocol hardly copes with high
traffic applications since there is one sink that monitors the network. To overcome
752 A. Eloudrhiri Hassani et al.

Fig. 1 Monitored network with multiple sinks

these challenges, we have proposed a novel objective function for routing data with
RPL supported by the multiple sinks solution as shown in Fig. 1 in order to increase
the network quality of services.

3.1 Metrics of Interest

It is important to determine how the metrics of interest should be combined to make

the best routing decisions. In this context, we have proposed in our work, linear
combination of two metrics namely rssi and hop count, amalgamated with the lexical
approach using ETX metric.

Expected Transmission Count (etx): is the number of transmissions required for

a successful packet transmission to destination. The cost of a path is the sum of the
ETX. When ETX is applied, the nodes must select the parent that has the lowest ETX
value. The metric is calculated with MAC callbacks as in Eq. 1.

ETX = 1/df × dr (1)

Where df represents the probability that a neighbor node j receive packet

successfully and dr represents the probability to receive a acknowledgment.

Number of Hops: is the only metric announced by neighboring nodes. For that
purpose, we have modified the structure of the DIO control message precisely in the
metric container part. Indeed, at reception of DIO message by a node i, it process
the hop count metric following Eq. 2.
Amalgamation of Novel Objective Function … 753

Minhcincrement cp = s
HC (i) = (2)
HC (n) + MinhcIncrement cp = s

Where n is the neighbor susceptible to be the preferred parent of node i, HC (n)

represents the number of hops to the sink advertised in the neighbor DIO whereas
MinhcIncrement is a scalar value equal to 256.

Radio Signal Strength Indicator (Rssi): Represent a link metric measured by a

node i through the CC2420 radio following Eq. 3.

rssi(i) = rssi_reg(i) − adjustment (3)

Where rssi_reg(i) is a 8 bits register that gives the signal strength between node i
and its neighbor. The antenna variation offset is represented with adjustment equal
to 45 provided from the CC2420 radio datasheet. Since the Rssi metric gives values
between 0 and −110 dBm Respectively for good and bad signal strength, this metric
must be maximized. To keep the objective function convergence, we worked with its
absolute value to follow the minimizing rule as other combined metric.

3.2 Routing and Decision Processes

During the process of parent selection, nodes are based on ranks calculated by the
objective function. Neighbors with lower rank are elected by the concerned node to
be its preferred parent. Unlike standard objective functions based on a single metric,
our proposal uses a set of metrics combined for rank processing and making the
routing decision. Indeed, at a DIO reception from a neighbor, the node calculates the
number of hops and the rssi at the MAC layer. The combination of these metrics is
done linearly with different weights following Eq. 4.

Combined(rssi, hc) = α ∗ rssi(i) + β ∗ hc(i) (4)

The weights α and β are complementary equal to 1. Subsequently, the node i

extract the rank of neighbor advertised in the DIO message and process its rank
according to Eq. 5.

Rank(i, n) = Rank(n) + Combined(rssi, hc) (5)

Where Rank (N) is the rank of the neighbor candidate to be the preferred parent
through node i could forward packets to sink. At this stage, once the node has calcu-
lated its rank based on the dio sender, if it is the first round for network establishment,
the node chooses it as the preferred parent. otherwise, the lexical approach is acti-
vated by a comparison firstly between the etx metrics of old and candidate rank, if it
754 A. Eloudrhiri Hassani et al.

Table 1 Proposed objective function algorithm

When node i receive a DIO message from neighbor
IF (neighbor! = NULL)
rssi = abs(rssi measure());
hc = neighbor.dio.hc + Minhc Increment;
etx = neighbor.dio.etx + etx measure();
Combined(rssi,hc) = α*rssi + β*hc;
Possible rank = neighbor.dio.rank + Combined(rssi,hc);
IF (Preferred parent == NULL)
Preferred parent = neighbor;
rank = Possible rank;
hc = hc;
etx=etx;
ELSE IF (etx < Preferred parent.etx)
IF (Possible rank < Preferred parent.rank)
Preferred parent = neighbor;
rank = Possible rank;
hc = hc;
etx=etx;
ELSE
/* Keep old preferred parent
exit;
ENDIF
ENDIF
/* Update the new DIO message
dio.mc.hc = hc;
dio.mc.etx = etx;
dio.rank = rank;
/* Broadcast DIO message
ENDIF
ENDIF

is higher; the candidate parent is automatically discarded. else, if the candidate parent
rank is lower too than the preferred parent rank, then the candidate parent is hold.
finally, if the condition for the parent change is fulfilled, the node updates its rank
and the hop count metric in the dio message, then broadcast a new one to advertise
neighbors that there is a change in the routing path. Table 1 show the proposed work
algorithm implemented in the RPL core.

4 Results and Discussion

In order to evaluate the performance of proposed objective functions, the simulations

are done with COOJA, a wireless network sensor simulator that provide a cross
approach emulation of the hardware platforms and programming the software level
through applications provided by the contiki operating system API. Although, in
Amalgamation of Novel Objective Function … 755

order to experiment our proposed method, we use a density of 25 nodes randomly

distributed on an area with 200 m each side. The nodes send captured data to one, two
or four sinks with a rate of 10, 20, 30 and 60 packets per minute. The proposals are
compared with the Standard-RPL in term of several parameters as packet delivery
ratio, throughput, energy consumption and number of DIO control packets.

4.1 Reliability Investigation

The first part of reliability investigation, we evaluate the Packet Delivery Ratio (PDR)
defined as the quotient of the packets sent by all the nodes of the network during the
period simulated by the number of packets received at the sink node. Figure 2 show
that our proposal with single sink provides an average PDR increase with 5.3, 4, 15
and 11% for 10, 20, 30, 60 ppm respectively against Standard-RPL. Also, according
to the multiple sinks solution added to our proposed objective function, an increase
about 30 and 37% respectively with 2 and 4 sinks is provided. These results can be
explained by the fact that, with multiple sinks, the congestion can be avoided and the
collaboration between destinations for the network monitoring can divide a single
network into several mini-networks which allows better management of the nodes
then better reliability.
The second part of reliability investigation, we evaluate the throughput, considered
as an after effect of the packet delivery ratio. Indeed, it is defined as the total received
packet by the sink or all the sinks divided by the total simulation time. Figure 3 show
that our proposal with single sink provides an average throughput increase with 6%,
3.5%, 18% and 13% for 10, 20, 30, 60 ppm respectively against Standard-RPL. Also,
according to the multiple sinks solution added to our proposed objective function,
an increase about 28% and 36% respectively with 2 and 4 sinks is provided.

Fig. 2 Packet delivery ratio for different number of sent packets per minute
756 A. Eloudrhiri Hassani et al.

4.2 Energy Investigation

Node’s energy consumption is defined as the summation of power used by each

node. There are four kinds in measurement of power. LPM indicates the radio power
consumption during sleep condition. CPU is a parameter that indicates monitors the
level of node processing. RL and RT are parameters related with node communica-
tions. The formula used to calculate the energy of nodes according to the following
Eq. (6).

Energy (mj) =3 ∗ (RT ∗ 19.5 + RL ∗ 21.5

+ CPU ∗ 1.8 + LMP ∗ 0.0545)/32768 (6)

To investigate on energy part, we measured the average of nodes power consump-

tion. Figure 4 exposes the results during the simulations. results show that our
proposed objective function with single sink can decrease the power consumption
with 17, 13, 10 and 7% for 10, 20, 30, 60 ppm respectively against Standard-RPL.

Fig. 3 Throughput for different number of sent packets per minute

Fig. 4 Power consumption of nodes for different number of sent packets per minute
Amalgamation of Novel Objective Function … 757

Fig. 5 DIO control packets for different number of sent packets per minute

Also, the multiple sinks solution can provide a decrease about 14 and 38 respectively
with 2 and 4 sinks against one sink solution.

4.3 Stability Investigation

The after effect of network stability is provided by the number of DIO control packets.
Indeed, as shown in Fig. 5, our proposal with single and multiple sinks can provide a
low number of DIO packets compared to Standard-RPL due to the optimized paths
selected using combined metrics and stability given by the mini network management
by each sink.

5 Conclusion

In this paper, an improvement of RPL protocol is proposed by a cross approach

between combined metrics and multiple sinks solution for high traffic monitoring
applications. This proposal is based on linear combination of rssi and hop count
metrics supported lexically by expected transmission count metric. Results showed
that our method has improved the reliability by I creasing the PDR and throughput,
decreasing the network power consumption and giving stability to the network
through low number of DIO control packets. As a future work, we will invest
on improving more the RPL routing protocol based on other approach of metrics
combination.
758 A. Eloudrhiri Hassani et al.

References

1. Yaqoob I et al (2017) Internet of Things architecture: recent advances, taxonomy, requirements,

and open challenges. IEEE Wirel Commun 24:10–16
2. Musaddiq A et al (2018) A survey on resource management in IoT operating systems. IEEE
Access 6:8459–8482
3. Gungor VC et al (2011) Smart grid technologies: communication technologies and standards.
IEEE Trans Industr Inf 7:529–539
4. Naeem MA et al (2018) A periodic caching strategy solution for the smart city in information-
centric Internet of Things. Sustainability (10): 2576
5. Xu LD et al (2014) Internet of Things in industries: a survey. IEEE Trans Industr Inf 10:2233–
2243
6. Qadri YA et al (2020) The future of healthcare internet of things: a survey of emerging
technologies. IEEE Commun Surv Tutor 22:1121–1167
7. Alexander R et al (2012) RPL: IPv6 Routing Protocol for Low-Power and Lossy Networks.
Internet Engineering Task Force (IETF). RFC 6550
8. Hassani AE et al (2019) Assessment of a proactive routing protocol RPL in Ipv6 based wireless
sensor networks. In Third International Conference on Intelligent Computing in Data Sciences
(ICDS), Marrakech, pp 1–7
9. Granjal J et al. (2010) Enabling network layer security on IPv6 wireless sensor networks. In
IEEE Global Telecommunications Conference GLOBECOM, pp 1–6 (2010)
10. Gaddour O et al (2012) Simulation and performance evaluation of DAG construction with RPL.
In: IEEE Third International Conference on Communications and Networking, pp 1–8
11. Gnawali O et al (2012) The Minimum Rank with Hysteresis Objective Function. Internet
Engineering Task Force (IETF). RFC 6719
12. Thubert P et al (2012) Objective Function Zero for the Routing Protocol for Low-Power and
Lossy Networks (RPL). Internet Engineering Task Force (IETF). RFC 6552
13. Sanmartin P et al (2018) Sigma routing metric for RPL protocol. Sensors (18): 1277
14. Iova O et al (2015) Exploiting multiple parents in RPL to improve both the network lifetime and
its stability. In: Proceeding of IEEE International Conference on Communications, pp 610–616
15. Gonizzi P et al (2013) Design and evaluation of a delay-efficient RPL routing metric. In: 9th
International Wireless Communications and Mobile Computing Conference
16. Bhandari KS et al (2019) A resource oriented route selection framework using contextual
information based on fuzzy logic. Electronics (8):1023
17. Lamaazi H et al (2018) OF-EC: a novel energy consumption aware objective functions for RPL
based on fuzzy logic. J Netw Comput Appl 117:42–58
18. Sankar S et al (2018) Fuzzy logic based energy aware routing protocol for internet of things.
Int J Int Syst Appl 10:11–19
19. Al-Kashoash H et al (2016) Congestion-aware RPL for 6L0WPAN networks. In Wireless
Telecommunications Symposium (WTS)
20. Hassani AE et al (2020) FTC-OF: forwarding traffic consciousness objective function for RPL
routing protocol. Int J Electr Electron Eng Telecommun 10:168–175
21. Hassani AE et al (2020) A new objective function based on additive combination of node and
link metrics as a mechanism path selection for RPL protocol. Int J Commun Netw Inf Secur
12:1–7
22. Nassar J et al (2018) Multiple instances QoS routing in RPL: application to smart grids. Sensors
18:1–6
23. Al-Kashoash H et al (2020) Congestion-Aware Routing Protocol for 6LoWPANs. Congestion
Control for 6LoWPAN Wireless Sensor Networks: Toward the Internet of Things, pp 95–107
Wireless Sensor Network Fault Detection
and Isolation for Smart Irrigation

Nassima Jihani, Mohammed Nabil Kabbaj, and Mohammed Benbrahim

Abstract Wireless Sensor Network (WSN) technology plays an important role in the
application of smart irrigation by monitoring environmental soil conditions, which
helps determine the amount of water required for an agricultural field. These sensors
are exposed to harsh environmental conditions over long period of time and became
prone to failure. These failures may lead the base station to take the wrong decision.
This Paper presents a centralized approach toward detecting and isolating sensor fault
in WSN using analytical redundancy which to achieve water saving and guarantee
a good crop quality. Simulation results are proved the effectiveness of the proposed
method.

Keywords Fault detection and isolation · Analytical redundancy · WSN · Parity

space · Smart irrigation

1 Introduction

In the last decades, the world has experienced a lot changes such as climate change
with its effect on precipitation, and the population growth that led to an increase
in food demand. These significant changes call for the adoption of a smart irriga-
tion system in the agriculture field. Smart irrigation has been the subject of multiple
research efforts recently and several research present the architecture of smart irriga-
tion system using WSN technology [1–4]. This system did not just prove its efficiency
on the yield and water saving but also it enables the automation of control and moni-
toring the crops remotely. WSN plays an important role in this application, the sensor

N. Jihani (B) · M. N. Kabbaj · M. Benbrahim

LIMAS, Sidi Mohamed Ben Abdellah University, Fez, Morocco
e-mail: [email protected]
M. N. Kabbaj
e-mail: [email protected]
M. Benbrahim
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 759
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_69
760 N. Jihani et al.

nodes are scattered over a farm in order to monitor and gather data concerning agri-
cultural conditions such as temperature, humidity, rainfall, soil PH and send them
wirelessly to the base station. Based on these monitored conditions, the base station
performs the required calculations to determine whether the agricultural field needs
water and calculate the amount of water required. To achieve the desired results, WSN
should operate correctly and accurately during the agricultural season. In smart irri-
gation system, the sensor nodes are exposed to harsh environmental conditions over
long period of time. This leads to a high probability of sensor nodes malfunction
which can cause fault measurement. The latter can cause the base station to make
faulty decisions which may damage or reduce the quality of the crops and causes
water waste. So, the WSN requires continuous monitoring to verify the quality of
measurements to ensure a reliable monitoring of the field and therefore obtaining an
adequate irrigation process, and most of the proposed smart irrigation systems have
not addressed this issue.
Fault detection and isolation (FDI) approach aims to detect and isolate faults in the
engineering systems using analytical redundancy rather than hardware redundancy
which is need the extra equipment and maintenance cost. FDI concept has widely
been used in computer science and information technology for several years [5, 6].
However, the use FDI in WSN for smart irrigation is not widely investigated.
FDI approaches in WSNs can be classified into two categories: centralized FDI
and distributed FDI. The first category, requires a central node or a base station to
collect data from the nodes, then analyzes it and takes decision if the collected data
are free faults or not. While in the second category, each node autonomously detects
and isolates its sensors faults and may also detect and isolate the sensors faults of its
neighbors.
In [7], the authors solved the FDI over WSN with centralized approach while in
[8] and [9], a distributed approach is adopted for detecting and isolating faults in
WSN.
In [8] a distributed FDI algorithm is developed in which each node detects and
isolates its sensors faults and the faults of its neighbors using parity space approach.
Each node built its model based on the measurements of its sensors and the measure-
ments of its neighbors. The same thing for [9], a model is built according to the
measurements of its neighbors using neural network approach. In some way a
distributed approach proved its efficiency on network lifetime but still has a weakness
over time in detection accuracy, nodes may make their decision based on incorrect
measurements from neighbors. Plus the limited computation capacity of the nodes
cannot support complex computation as neural network. While the main tasks of the
nodes in the most applications are to monitor environmental conditions and route
them to the base station.
Wireless Sensor Network Fault Detection … 761

However, in [7] the faults are detected and isolated with a centralized way, but
the proposed method initially requires knowing the characteristics of the faults and
their impact on the characteristics of the measured signal and based on that, for each
type of sensor fault, they created a specific algorithm to detect it.
This paper presents a centralized approach for detecting and isolating fault in WSN
using an analytical redundancy approach, the parity space approach. This proposed
method does not require a model of the monitored system or to know beforehand the
behavior of the faulty component and the characteristic of the fault as used in [7],
and the centralized approach also helps keeping detection accuracy over time and
facilitates the nodes’ tasks contrary to the distributed approach [9].
The remainder of the paper is organized as follows: Sect. 2 presents the problem
statement. Section 3, the proposed approach for FDI in a WSN with a centralized
architecture. Simulation and results discusses in Sect. 4. Finally, conclusion is given
in Sect. 5.

2 Problem Statement

The framework of the proposed smart irrigation is shown in Fig. 1. A set of sensor
nodes deployed over a farm toward measure environmental parameters such as
temperature and soil moisture, and then with the collaboration of these nodes, the
data are transformed wirelessly to the base station where the data are analyzed with
sensor fault detection and isolation (SFDI) algorithm before being used to make the
decision to irrigate the farm or not.
We consider the measurement of each sensor node writing as the following
expression:

yi (k) = Ci x(k) + f i (k) (1)

Fig. 1 Framework of smart irrigation application

762 N. Jihani et al.

Where yi (k) is the measurement vector at node i, x(k) is a vector that represents
the target quantities, Ci is the measurement matrix at node i that indicates which
quantities are measured at node i and f i (k) is the fault vector.
The gathered measurements from sensor nodes, form a model of the network in
the gateway as below:
⎡ ⎤
y1 (k)
⎢ y (k) ⎥
⎢ 2 ⎥
⎢ ⎥
⎢ . ⎥
y(k) = ⎢ ⎥ (2)
⎢ . ⎥
⎢ ⎥
⎣ . ⎦
yi (k)

3 Residual Generation

First up, we generate a global residual based on the model of the network with the
parity space approach, by multiplying the model Eq. 2 by ω:

r (k) = ωy(k) (3)

Where ω is a projection matrix which makes the residual independent of x(k) and
sensitive just to the occurrence of the faults and it is calculated as follow:

ωc = 0 (4)

c is the matrix measurement of the network.

then r (k) = ω f (k) (5)

Fault Detection. The generated residual as shown in Eq. 5, sensitive to the occur-
rence of the fault in the network. If the fault affects any sensor node, the norm of the
residual will be high than zero otherwise the norm of the residual will be equal zero.

if fi > 0 ||r || > 0 (6)

if fi = 0 ||r || = 0 (7)
Wireless Sensor Network Fault Detection … 763

Fault Isolation. Based on the generated residual and on the parity matrix (ω) which
its columns define n distinct fault signature directions, we can isolate the faulty sensor
by multiplying transpose of each column of the parity matrix by the residual.

F Ii = ωiT ∗ r i = 1, 2, . . . n. (8)

This function measure the correlation of the residual vector with fault signature
directions. So, if the fault is detected, the largest value of F I i corresponds to the
faulty sensor [10].

4 Simulation Results

In our simulation, we have a WSN consisting of 3 sensor nodes scattered in different

zones, each node contains one sensor and all of them measure the same quantity
x(k). Those sensor nodes send wirelessly their measurements to the base station.
Figure 2, illustrates the measurements recorded by the sensor nodes.
These measurements form a model of the network in the base station as follow:
⎡ ⎤
y1 (k)
y(k) = ⎣ y2 (k) ⎦ (9)
y3 (k)

Where y1 (k), y2 (k) and y3 (k) are the measurements recorded by sensor node 1,
2 and 3 respectively.
Based on this model, an implemented algorithm in the base station generates a
global residual and analyses it as explained in Sect. 3 for detecting and isolating
sensor fault.
In this paper, we examined only a single fault. So each time we inject a fault in
one sensor node, our algorithm detects and isolates it. Below, the simulation results
demonstrates the work that we did.
Based on the measurements recorded by the WSN, our algorithm generates a
residual norm as it is shown in Fig. 3. At t = 30 s a fault of type ‘abrupt’ is injected
into sensor node 1. As can be seen from Fig. 3, the generated residual norm is sensitive
to the fault and deviates from zero at t = 30 s, which means that from this instant,
there is a fault detected in the WSN.
After fault detection in WSN by the residual norm. To isolate the faulty sensor
node, our algorithm automatically calculates the correlation of the residual vector
with each column of the parity matrix (ω) as explained in Sect. 3.
764 N. Jihani et al.

Fig. 2 Measurements recorded by the wireless sensor nodes

Fig. 3 Network residual norm evolution

In our test, we have 3 fault signature directions defined by the column of the
parity matrix; the first, second and the third column of ω correspond to the fault
signature direction of the sensor 1, 2 and 3 respectively, the algorithm calculates the
correlation of the residual vector with each column of the parity matrix (ω) (Eq. 8)
and the largest value of F I i corresponds to the faulty sensor. As can be seen from
Fig. 4, F I 1 has the largest magnitude which means the faulty sensor is the senor 1.
Wireless Sensor Network Fault Detection … 765

Fig. 4 Fault isolation process

Fig. 5 Network residual norm evolution

At t = 30 s, the fault is injected into sensor node 2, the Fig. 5 shows that the global
residual norm of the network starts to react when the fault is present.
And the Fig. 6 shows that the F I 2 has the largest magnitude and that is mean, the
faulty sensor is sensor 2.
The fault is inserted into sensor 3 at t = 30 s and Fig. 7 shows that the residual
norm becomes different from zero from this instant.
766 N. Jihani et al.

Fig. 6 Fault isolation process

Fig. 7 Network residual norm evolution

The fault isolation process shows that F I 3 has the largest magnitude in Fig. 8. So
the faulty sensor is sensor 3.
Wireless Sensor Network Fault Detection … 767

Fig. 8 Fault isolation process

5 Conclusion

In this research, we developed an algorithm for detecting and isolating measurement

faults in a WSN for smart irrigation. In designing this algorithm, we took several
factors into consideration; first, we used analytical redundancy instead of hardware
redundancy which is much expensive in terms of cost. Also, we adopted a centralized
architecture rather than a distributed one to minimize the load on the sensor nodes
which have a limited capacity at the level of computation and energy source. Finally,
our approach does not require a model of the monitored system or a prior knowledge
about the behavior of the monitored system.
The simulation results obtained in this study demonstrated that the parity space
approach detects and isolates the fault in WSN in a centralized manner; when we
inject the fault into any sensor node, our algorithm detects and isolates it.
This paper examines the problem of the measurement faults caused by the
malfunctions of sensors or sensor nodes. In a WSN, a crash faults can be caused
by the failure of the communication modules of the nodes, a crash faulty sensor node
cannot participate in the network activities, and the future work may be conducted
on this subject.
768 N. Jihani et al.

References

1. Riquelme JL, Soto F, Suardíaz J, Sánchez P, Iborra A, Vera JA (2009) Wireless sensor networks
for precision horticulture in Southern Spain. Comput Electron Agric 68(1):25–35
2. Ojha T, Misra S, Raghuwanshi NS (2015) Wireless sensor networks for agriculture: the state-
of-the-art in practice and future challenges. Comput Electron Agric 118:66–84
3. Abd El-kader SM, El-Basioni BMM (2013) Precision farming solution in Egypt using the
wireless sensor network technology. Egyptian Inform J 14(3):221–233
4. Kassim MRM, Mat I, Harun AN (2014) Wireless Sensor Network in precision agriculture
application. In: 2014 international conference on computer, information and telecommunication
systems (CITS), pp 1–5. IEEE, July
5. Staroswiecki M, Hoblos G, Aitouche A (2004) Sensor network design for fault tolerant
estimation. Int J Adapt Control Signal Process 18(1):55–72
6. Isermann R (2011) Fault-diagnosis applications: model-based condition monitoring: actuators,
drives, machinery, plants, sensors, and fault-tolerant systems. Springer
7. Peng Y, Qiao W, Qu L, Wang J (2017) Sensor fault detection and isolation for a wireless
sensor network-based remote wind turbine condition monitoring system. IEEE Trans Ind Appl
54(2):1072–1079
8. Hao J, Kinnaert M (2017) Sensor fault detection and isolation over wireless sensor network
based on hardware redundancy. J Phys Conf Ser 783(1):012006. IOP Publishing
9. Smarsly K, Law KH (2014) Decentralized fault detection and isolation in wireless structural
health monitoring systems using analytical redundancy. Adv Eng Softw 73:1–10
10. Patton RJ, Chen J (1991) A review of parity space approaches to fault diagnosis. IFAC Proc
Vol 24(6):65–81
Multi-resolution Texture Analysis
for Osteoporosis Classification

Meriem Mebarkia, Abdallah Meraoumia, Lotfi Houam, Seddik Khemaissia,

and Rachid Jennane

Abstract Osteoporosis is a systemic skeletal disease characterized by low bone

mineral density (BMD) and micro-architectural deterioration of bone tissue, leading
to bone fragility and increased fracture risk. Texture analysis applied to trabecular
bone images offers an ability to exploit the information present on conventional
radiographs. In this paper, a feature extraction scheme is proposed to analyze different
bone X-ray images of two different populations in order to extract a precise and
lightweight feature. The proposed scheme uses two efficient variants of Local Binary
Patterns (LBP) and combines the information extracted from each level to improve
the classification rate. The experimental results prove the efficiency of the proposed
method in terms of classification speed and system accuracy.

Keywords Local binary pattern · Fisher score · Data fusion · Osteoporosis

1 Introduction

OSTEOPOROSIS is a skeletal disease characterized by low bone mass density

(BMD), and deterioration of bone microarchitecture and can lead to compromised
bone strength and an increased risk of fractures [1]. It is a silent disease until frac-
tures occur, resulting in significant secondary health problems and even death [2].

M. Mebarkia · S. Khemaissia
Laboratoire de Génie Electrique (LABGET), Larbi Tebessi University, Tebessa, Algeria
e-mail: [email protected]
A. Meraoumia (B) · L. Houam
LAMIS Laboratory, Larbi Tebessi University, Tebessa, Algeria
e-mail: [email protected]
L. Houam
e-mail: [email protected]
R. Jennane
I3MTO Laboratory, University of Orleans, Orléans, France
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 769
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_70
770 M. Mebarkia et al.

Several treatments have been shown to reduce the risk of osteoporotic fractures by
following the necessary steps that can be taken to diagnose osteoporosis and thus
prevent bone fracture [3, 4]. An early diagnosis of osteoporosis helps in preventing
bone fractures. The diagnosis of osteoporosis relies primarily on measuring BMD
(expressed as a T-score) using dual-energy X-ray absorptiometry (DXA). DXA is
considered as the reference-standard examination for BMD assessment. Unfortu-
nately, the high-cost, low-availability, and inconsistency of BMD measurements do
not make it a promising tool for the diagnosis of osteoporosis [5, 6]. In addition,
BMD does not cover the whole process of osteoporosis diagnosis, skeletal factors
and micro-architecture features are also of great significance in the early diagnosis
of osteoporosis.
In accordance with many studies [7–9], the particular texture analysis of trabec-
ular bone radiographs is a promising approach to assess and recognize this disease
effectively. Image feature information of trabecular bone has been known to be corre-
lated closely to bone density changes and therefore can be used for the evaluation
of osteoporosis. Many algorithms have been used for the extraction of features of
trabecular pattern, such as Gabor filter, wavelet transform and fractal models [10–12].
Song et al. have presented another approach which again used the analysis of bone
textures by the usage of Fisher encoding of local descriptors algorithm with SVM
for different classes [13], whereas Nasser et al. have proposed another approach to
assess osteoporosis based on the sparse autoencoder and SVM classifier [14]. Zheng
et al. have presented a concept based on a calculation of sparse representations, to
classify bone texture image patterns [15].
Recently, with the advancements of Machine learning, deep CNNs have been
recognized as effective tools for image classification [16]. For instance, Ran Su et al.
combined CNN features with hand-crafted features for osteoporosis classification
[17]. However, previous approaches relied mainly on handcrafted categorized feature
indexes [14, 18], and machine learning (ML) algorithms, such as Support Vector
Machine (SVM) and fuzzy classifiers [9]. In addition, most of the previous studies
require tedious process and manual operations, such as extensive preprocessing,
image normalization, and region of interest (ROI) extraction, which can drastically
affect the reproducibility of the classification method [17].
The purpose of the present work is to validate the improved methods known
as Three Patches Local Binary Patterns (TP-LBP), and Four Patches Local Binary
Patterns (FP-LBP), suggested in [19]. The approach is composed in four stages proce-
dure. First, a discrete wavelet transform (DWT) is performed to the original images
of bone radiographs. In second step, TP-LBP/-FP-LBP methods are performed on
all resulting DWT sub-bands at each decomposition level, to construct the feature
vectors as shown in Fig. 2. The third step, the Fisher score algorithm is applied to the
feature vector as supervised feature selection method. Finally, the SVM classifier is
used with the linear kernel for classification of two different populations composed
of osteoporotic (OP) and control subjects CS).
Multi-resolution Texture Analysis for Osteoporosis Classification 771

The rest of this paper is organized as follows: Sect. 2 presents and discusses
in details the proposed feature extraction process. Section 3 deals with the experi-
mental results in which unimodal and multimodal systems are discussed. Finally, the
conclusion and future work are presented in Sect. 4.

2 Feature Extraction Process

The performance of all pattern recognition systems depends mainly on the distinction
of feature vectors extracted from the data. In fact, the role of the feature extraction
process is to represent the images with a reduced and distinct representation, so
that the system can discriminate the different patterns within a reasonable time. In
our proposal, we used the texture analysis method with a multi-resolution analysis
technique (e.g. DWT).

2.1 Local Binary Pattern (LBP)

Local Binary Patterns (LBP) is an image-processing algorithm to extract texture prop-

erties for classification and segmentation purposes. LBP is computed in its simplest
form by comparing each pixel with its surrounding neighbors. The original lbp oper-
ator is defined in 3 × 3-sized rectangular neighborhoods. Each pixel in a cell is
compared to each of its 8 neighbors leading to an 8-digit binary number. Finally, an
histogram is created for all the decimal values to be used as texture description for
the image [20–22].
Due to the efficiency of this method, different variants have emerged; such as
Three-Patches LBP (TP-LBP) and Four-Patches LBP (FP-LBP) [19], as shown in
Fig. 1.

Three-Path LBP (TP-LBP). The TP-LBP is an extended method of LBP; it

compares the values of three patches with (w × w) size to produce a single bit
value for each pixel. For a parameter α, pairs of patches are taken apart along the
circle and compared to the central patch, S additional patches distributed uniformly
in a ring of radius r around it [19]. The resulting code has S bits per pixel (Fig. 1(a)).
The TP-LBP operator is given by the following equation:
S−1     
TP-LBPr,S,w,α (p) = f d Cp, Ci − d Cp , Ci+αmod(S) 2i (1)
i=0


1 if x ≥ τ
f (x) = (2)
0 if x < τ
772 M. Mebarkia et al.

Fig. 1 Extended local binary pattern. a Three-Path Local Binary Pattern (TP-LBP) and b Four-Path
Local Binary Pattern (FP-LBP) [19]

Fig. 2 Block diagram of the proposed multi-resolution bone texture analysis

Where d denotes the distance between patches, C is the patch, and f is a threshold
function and τ is the threshold of comparison.

Four-Path LBP (TP-LBP). The FP-LBP is another extended method of LBP. It uses
two rings of radius r 1 and r 2 centered in the pixel, and S patches distributed around
these two rings of (w × w) size (see Fig. 1.(b)) [19]. The FP-LBP is formed by the
comparison between two center symmetric patches in the inner circle with two center
symmetric patches in the outer circle by taking one bit in each pixel into accounts
according to the similarity of the two patches. The FP-LBP operator is given by the
following equation:

 S −1  
FP-LBPr 1,r 2,S,w,α = 2
f (d C1,i , C2,i+αmod(S) − d(C1,i+ S , C 2,i+ S +αmod(S) ))2i (3)
i=0 2 2

2.2 Proposed Feature Extraction Architecture Diagram

The structure of the proposed feature extraction method is presented in Fig. 2. The
image is first transformed using the DWT, then the features are extracted using
Multi-resolution Texture Analysis for Osteoporosis Classification 773

the LBP algorithm and finally, the extracted information is combined into the two
system’s levels (feature extraction and feature matching levels).
The input image (Ao ) of size (H × W ) is first transformed using the DWT into
several decomposition levels, as follows:

FDWT
o
(Ao ) = Ao ∈ R H ×W
i = 1, 2, . . . . (4)
FDWT (Ai−1 ) = [Ai , Hi , Vi , Di ] ∈ R 2i × 2i
H W
i

Where F DWi
T denotes the DWT transform at i level and Ai , Hi , Vi , Di are the
resulting decomposition bands at level i. Then, distinctive features of each band
are extracted separately using the TP(FP)-LBP technique.

Vψi = F TP(FP)-LBP (ψ), ψ ≡ {Ai , Hi , Vi , Di } (5)

Where FTP(FP)-LBP denotes the TP-LBP (FP-LBP) technique and one of the resulting
decomposition bands (A: Approximation, H: Horizontal, V: Vertical, D: Diagonal).
After extraction, the system can: i) either, combine (fuse) all the extracted feature
vectors into one single vector (FF: Fusion at feature extraction level) then carry on
the classification process, ii) or, continue the classification process for each vector
alone till the end and then combine the classification results for all the vectors into
a single result (FS: Fusion at the matching score level).

Fusion at Feature Extraction Level. For its simplicity, the concatenation method
is used in this study. Let V Ai , VHi , VVi , VDi be the feature vectors of the Ai , Hi , Vi , Di
bands at level i, the combined vector is obtained as follows:

V Ai = FCONCAT (V iA , VHi , VVi , VDi ) = [V Ai , VHi , VVi , VDi ] (6)

Where FCONCAT combines the vectors by placing them in front of each other.

Fusion at Matching Scores Level. Here, each band works independently, and their
results are combined to obtain a scalar score that will be used later for decision
making. In our study, the sum of weighted scores defined as follows is used:

n 1
dF = wi di , wi = n (7)
i=1 Ei 1
j=1 Ei

Where d F denotes the fused score, di is the score obtained by the subsystem i,
n the number of subsystems, Ei represents nthe system error and wi is the weight
associated with the subsystem i. Note that i=1 wi = 1 and the weights are inversely
proportional to the corresponding errors and are therefore larger for the most precise
systems.
774 M. Mebarkia et al.

3 Experimental Results

This section highlights the performance of the proposed method by being integrated
into an osteoporosis detection system. Thus, to evaluate the effectiveness of the
proposed approach, we used a dataset of 174 images acquired at the hospital of
Orleans (France) from postmenopausal women, using X-rays on the calcaneus site
[23]. The dataset comprised 87 Osteoporosis Patients OP (with fractures) and 87
Control Subjects (CS).

3.1 Tests Protocol

In this experiment, the 5-fold cross-validation procedure is adopted. The dataset is

divided into 5 equivalent folds containing 34 samples, half (17) were from healthy
subjects and the other half were osteoporotic subjects. It is worth noting that before
the feature extraction process, each image was preprocessed (normalization and
histogram equalization). The Support Vector Machine (SVM) technique is used
for the classification step. Finally, the DWT application is limited to two levels
of decomposition.
The feature vector is extracted at each level for each band separately. Fisher’s
feature selection technique is used to improve the system performance. Accordingly,
(25%, 50%, 75%, 100%) of the resulting vector components are tested for each band.

3.2 Performance Tests Results

Several experiments are conducted within two main parts. The first part of the exper-
iments determines the optimal parameters for each feature extraction methods (TP-
LBP and FP-LBP). The second part focuses on the performance evaluation of the
proposed osteoporosis detection system. This part is also divided into two sub-parts,
the first concerns the study of the unimodal system (each DWT band is treated
separately), while the second part focuses on the effect of the data fusion on the
performance of the system (multi-modal system).

Experimental Setup. Several experiments are conducted to select the appropriate

combination in order to get the best performance. The patch size (w × w) is fixed
among three predefined values (3×3, 5×5 or 7×7). The remaining parameters for the
feature extraction methods (TP-LBP and FP-LBP) are taken from the predetermined
values (r = 3, S = 6, α = 2) and (r1 = 2, r2 = 6, S = 12, α = 1), respectively. It
should be noted that all the experiments are carried out on the whole image with 25%
of the components of the vector (using Fisher feature selection). In the case of TP-
LBP, the system performance is very acceptable for almost any patch size value. For
example, when using w size 3, 5 or 7 patches; our system operates at an Area Under
Multi-resolution Texture Analysis for Osteoporosis Classification 775

the Curve (AUC) equal to 96.83% (Accuracy (ACC) = 91.76%), 97.22% (99.18%)
and 96.07% (90.59%), respectively. Compared to TP-LBP, the system performance is
very efficient for all patch sizes with FP-LBP. Thus, under the different patch sizes, the
AUC reaches 99.02% (ACC = 97.06%), 99.97% (99.41%) and 99.57% (98.24%) for
w equal to 3, 5, 7, respectively. It can be noted that excellent performance exceeding
96.00% for both TP-LBP and FP-LBP are achieved. In conclusion, it is clear that for
each feature extraction method, the value of 5 achieves the best performance.

System Performance Evaluation. The performance of the proposed osteoporosis

detection system, was evaluated for the selected parameter (w = 5 for both feature
extraction methods).

Unimodal System (Uni-Band Evaluation). In this series of tests, the performance

of the osteoporosis detection system is evaluated using the information provided by
each band extracted by the DWT (at 1 and 2 decomposition levels). Hence, Table
1 illustrates the results of all bands prior to DWT application under 1 and 2 levels
of decomposition and feature selection rates (using Fisher’s criterion) equal to 25%,
50%, 75% and 100% of vector components.
As can be seen on Table 1 for the whole i mage (with a size of 400 × 400 pixels),
the AUC is 99.22 and 99.97%, and the processing time is estimated at 0.73395 and
1.79805 s for TP-LBP and FP-LBP, respectively. Actually, the smaller the image size,
the smaller the processing time is, so that the DWT-band processing time for each
analysis level is less than the total time. For DWT-bands of level 1, the processing time
is 0.52297 and 0.74726 s, respectively for TP-LBP and FP-LBP. The time decreases
as far as the analysis levels increase (the processing time for the DWT-bands of level
2 is 0.43189 and 0.48367 s for TP-LBP and FP-LBP, respectively).
Figure 3 depicts the system performance for all decomposition bands (with 25%
feature vector’s components) for both methods (TP-LBP and FP-LBP). The compar-
ison of TP-LBP with the whole image is shown in Fig. 3(a), where the system operates

Table 1 Unimodal osteoporosis detection system performance (AUC %)

DWT TP-LBP FP-LBP
Levels
Bands 25% 50% 75% 100% 25% 50% 75% 100%
cA1 87.14 65.19 58.02 58.02 94.63 82.70 54.12 54.09
L1 cH1 88.71 62.27 47.63 47.63 90.34 73.85 45.83 45.72
cV1 84.58 64.91 54.46 54.46 87.10 71.50 42.26 42.26
cD1 87.31 58.55 47.21 47.21 93.11 81.77 59.57 59.54
cA2 60.19 49.51 49.66 49.40 75.82 61.83 56.98 56.87
L2 cH2 52.64 44.04 43.99 43.86 61.66 53.11 47.14 47.18
cV2 58.85 48.79 48.75 48.72 68.46 52.58 48.53 48.54
cD2 60.87 52.61 52.65 52.30 81.11 60.55 53.18 53.40
Whole image 97.22 91.28 68.65 68.65 99.97 99.40 91.61 54.01
776 M. Mebarkia et al.

Fig. 3 Performance of the unimodal osteoporosis detection system. a, c AUC for TP-LBP and
FP-LBP, respectively and b, d ACC when using TP-LBP and FP-LBP, respectively

with an AUC of 88.71% (ACC = 80.02%), i.e. a degradation of 8.75% in the case of
the horizontal detail band (cH). In this configuration, the best accuracy is obtained
using the detail band (cD) with an ACC equal to 82.94% (a degradation of 9.04%),
as in Fig. 3(b). In the case of the second method (FP-LBP), the system generally
remains slightly unchanged with respect to the first method (see Fig. 3 (c) and (d). In
this case the system works well (best AUC = 94.63% with ACC = 87.65%) with the
cA band, the degradation rates in AUC and ACC are 5.34% and 11.83%, respectively.
Therefore, the obtained results in all decomposition bands are suitable. In the next
subpart, the osteoporosis detection system will be examined for all the decomposition
bands when the first and second decomposition levels (L1 and L2 ) are fused. The
fusion will be done at the feature extraction level and at the matching score level,
then a comparison of the performance of the system will be addressed for the whole
image.

Multimodal System (Multi-band Evaluation). In general, a unimodal system can

be affected by certain problems such as noise, variation within the same class (low
intra-class correlation), and similarity between classes (high inter-class correlation).
These problems cause the system to operate with certain errors which results in low
Multi-resolution Texture Analysis for Osteoporosis Classification 777

classification rates. Data fusion can reduce system errors and thus improve their
efficiency. Therefore, this subpart aims to examine whether the system performance
can be improved by using a combination of information extracted from different
DWT-bands. Thus, the feature vectors extracted from these bands are combined to
make the system more efficient than a unimodal system.
Fusion at the feature extraction Level (FF) makes it possible to combine different
vectors from different processing and analysis phases to form a single feature
vector. Therefore, after decomposing the image using DWT and obtaining the four
bands, features are extracted for each band separately and then the system combines
(concatenates) the extracted vectors to get a single feature vector. Afterward, the
system continues until the final decision is made (classification). In our experiments,
all the DWT bands from each level (L1 and L2 ) are combined and the result (as AUC)
is obtained using the whole image. Results are reported in Table 2.
In fusion at matching score level (FS), the individual scores are combined to form
a single score which is then used to make the final decision. It is the most widely used
type of fusion because it can be applied to all types of systems in a simple and efficient
way. In our multimodal system, each DWT-band has its own subsystem, which means
that the overall system contains four subsystems that operate in parallel. Then the
results of the subsystems are combined for the final decision. As previously, we
also combined the DWT bands from each level of decomposition to examine the
possibility of improving the performance of the system. The results obtained are
shown in Table 2. As can be seen, it is clear that although data fusion did not improve
the classification result, the results obtained are very satisfactory in both feature
extraction methods. In the first method (TP-LBP), the system operates with an AUC
equal to 97.19% (ACC = 93.53%), while the second method (FP-LBP) remains more
efficient with an AUC = 98.69% (ACC = 94.12%). Figure 4 illustrates a comparison
between unimodal and multimodal systems.
Figure 4 shows that data fusion can give effective and promising results, especially
when considering parallel processing.

Table 2 Multimodal osteoporosis detection system performance (AUC %)

Fusion TP-LBP FP-LBP
Levels
Level 25% 50% 75% 100% 25% 50% 75% 100%
FF 98.80 73.66 57.13 57.13 98.01 86.13 50.02 50.05
L1
FS 97.19 71.45 57.25 57.25 98.69 88.35 56.06 56.00
FF 62.88 46.63 46.67 46.64 90.85 62.09 52.30 52.36
L2
FS 66.06 51.90 51.82 51.67 89.44 66.55 55.88 56.04
Whole image 97.22 91.28 68.65 68.65 99.97 99.40 91.61 54.01
778 M. Mebarkia et al.

Fig. 4 Osteoporosis detection performance under one level of DWT decomposition. a Performance
comparison using TP-LBP and b Performance comparison using FP-LBP method

4 Conclusion and Further Works

The assessment of osteoporotic disease from X-ray images is a challenging behavior

for pattern recognition and medical applications. Textured images from the bone
micro-architecture of osteoporotic and healthy subjects show a high degree of simi-
larity, greatly increasing the difficulty of classifying these textures. In this work,
we have proposed an osteoporosis detection system based on texture analysis.
In this study, two improved variants of LBP were used in conjunction with the
Discrete Wavelet Transform method to speed up the classification process. The results
obtained are very promising and can be considered as a basic cell in the development
of an effective medical diagnostic system. In our future work, we plan to fuse all first
and second level DWT-bands using other fusion rules. In addition, we will examine
the system using the principle of pyramid analysis.

References

1. NIH (2001) Osteoporosis prevention, diagnosis, and therapy. JAMA 285:785–95

2. Cauley JA (2013) Public health impact of osteoporosis. J Gerontol A Biol Sci Med Sci
68(10):1243-1251. https://doi.org/10.1093/gerona/glt093. Accessed 31 Jul 2013
3. Bliuc D, Nguyen ND, Nguyen TV, Eisman JA, Center JR (2013) Compound risk of high
mortality following osteoporotic fracture and refracture in elderly women and men. J Bone
Miner Res. 28(11):2317–2324. https://doi.org/10.1002/jbmr.1968
4. Melton LJ, Atkinson IEJ, O’Connor MK, O’Fallon WM, Riggs BL (1998) Bone density and
fracture risk in men, vol 13, pp 1915–1923
5. Sozen T, Ozisik L, Basaran NC (2017) An overview and management of osteoporosis. Eur J
Rheumatol 4:46–56
6. Tosteson AN, Melton LJ 3rd, Dawson-Hughes B, Baim S, Favus MJ, Khosla S, Lindsay
RL (2008) Cost-effective osteoporosis treatment thresholds: The United States perspective.
Osteoporos Int 19:437–447
Multi-resolution Texture Analysis for Osteoporosis Classification 779

7. Park C, Took CC, Seong J-K (2018) Machine learning in biomedical engineering. Biomed Eng
Lett 8:1–3
8. Kavitha MS, Asano A, Taguchi A, Kurita T, Sanada M (2012) Diagnosis of osteoporosis from
dental panoramic radiographs using the support vector machine method in a computer-aided
system. BMC Med Imaging 12:1
9. Muthu Subash K, Pugalendhi Ganesh K, Soon-Yong P, Kyung-Hoe H, Min-Suk H, Takio K,
Akira A, Seo-Yong A, Sung-Il C (2016) Automatic detection of osteoporosis based on hybrid
genetic swarm fuzzy classifier approaches. vol 45, p 20160076
10. Houam L, Hafiane A, Boukrouche A, Lespessailles E, Jennane R (2012) Texture character-
ization using local binary pattern and wavelets. Application to bone radiographs. In: 2012
3rd International Conference on Image Processing Theory, Tools and Applications (IPTA), pp
371–376
11. Taleb-Ahmed A, Dubois P, Duquenoy E (2003) Analysis methods of CT-scan images for the
characterization of the bone texture: First results. Pattern Recogn Lett 24:1971–1982
12. Oulhaj H, Rziza M, Amine A, Toumi H, Lespessailles E, Hassouni ME, Jennane R (2017)
Anisotropic discrete dual-tree wavelet transform for improved classification of trabecular bone.
IEEE Trans Med Imaging 36:2077–2086
13. Song Y, Cai W, Zhang F, Huang H, Zhou Y, Feng DD (2015) Bone texture characterization
with fisher encoding of local descriptors. In: 2015 IEEE 12th international symposium on
biomedical imaging (ISBI), pp 5–8
14. Nasser Y, Hassouni ME, Brahim A, Toumi H, Lespessailles E, Jennane R (2017) Diagnosis
of osteoporosis disease from bone X-ray images with stacked sparse autoencoder and SVM
classifier. In: 2017 international conference on advanced technologies for signal and image
processing (ATSIP), pp 1–5
15. Zheng K, Jennane R, Makrogiannis S (2019) Ensembles of sparse classifiers for osteoporosis
characterization in digital radiographs. Med Imag Comput Aid Diag 1095024
16. Singh A, Dutta MK, Jennane R, Lespessailles E (2017) Classification of the trabecular bone
structure of osteoporotic patients using machine vision. Comput Biol Med 91:148–158
17. Su R, Liu T, Sun C, Jin Q, Jennane R, Wei L (2020) Fusing convolutional neural network
features with hand-crafted features for osteoporosis diagnoses. Neurocomputing 385:300–309
18. Oulhaj H, Rziza M, Amine A, Toumi H, Lespessailles E, Jennane R, El Hassouni M (2017)
Trabecular bone characterization using circular parametric models. Biomed Signal Process
Control 33:411–421
19. Wolf L, Hassner T, Taigman Y (2008) Descriptor based methods in the wild. In: European
conference on computer vision workshop on faces in real-life images (ECCV), Marseille,
France
20. Ojala T, Pietikäinen M, Harwood D (1996) A comparative study of texture measures with
classification based on featured distributions. Pattern Recogn 29:51–59
21. Ojala T, Pietikainen M, Maenpaa T (2002) Multiresolution gray-scale and rotation invariant
texture classification with local binary patterns. IEEE Trans Pattern Anal Mach Intell 24:971–
987
22. Houam L, Hafiane A, Boukrouche A, Lespessailles E, Jennane R (2014) One dimensional local
binary pattern for bone texture characterization. Pattern Anal Appl 17:179–193
23. Lespessailles E, Gadois C, Kousignian I, Neveu JP, Fardellone P, Kolta S, Roux C, Do-Huu JP,
Benhamou CL (2008) Clinical interest of bone texture analysis in osteoporosis: a case control
multicenter study. Osteoporosis Int 19:1019–1028
High Performance Decoding
by Combination of the Hartmann
Rudolph Decoder and Soft Decision
Decoding by Hash Techniques

Hamza Faham, My Seddiq El Kasmi Alaoui, Saïd Nouh,

and Mohamed Azzouazi

Abstract Decoding algorithms are used to recover information after its transmission
through an unreliable radio channel. Soft decision decoding is very performant in
concatenation solutions that exploit many decoding stages. Therefore, we carry out
a combination of a symbol-by-symbol decoder, which is Hartmann Rudolph (HR)
algorithm, and the Soft Decision Decoder by Hash Techniques (SDHT) as a word-to-
word decoder. In this paper, we introduce a concatenation of HR Partially exploited
(PHR) and the SDHT algorithm to decode Bose-Chaudhuri-Hocquenghem (BCH)
codes. This scheme enables at first to employ Hartmann Rudolph algorithm using
only a part of the dual codewords, then the output of PHR is treated by the SDHT.
Our proposed serial concatenation guarantees very good performances although we
use very low percentage of the dual code space. As instance for the BCH(31, 21, 5)
code, the obtained good results are based only on 21.29% of the dual codewords.

Keywords Error correcting codes · Symbol-by-symbol decoder · Word-to-word

decoder · HR · PHR · SDHT

1 Introduction

Recently, there has been an expanding demand for perfect and reliable digital data
transmission and backup systems. This demand has grown by the emersion of exten-
sive, high throughput data networks for the exchange, treatment, and backup of binary
information in different sectors and domains. A considerable concern of standards
organizations, which develop protocols for mobile telephony, is the check of errors
so that authentic reproduction of data can be achieved. Wherefore, error correction
codes have been developed. Error correction coding is achieved by joining redun-
dancy to the transferred message using an algorithm. Several correction codes are
implemented in various equipment like smartphones, CDs, DVDs, storage media or
packets transmitted through Internet and cellular networks.

H. Faham (B) · M. S. El Kasmi Alaoui · S. Nouh · M. Azzouazi

Information Technology and Modeling Lab, Ben M’sick Sciences Faculty, Hassan II University,
Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 781
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_71
782 H. Faham et al.

There are two types of decoders employed in communication systems: soft deci-
sion and hard decision decoders. Soft decision algorithms exploit directly the received
symbols and they employ principally the Euclidian distance as a measure. While
the hard decision decoders measure is usually the Hamming distance. This type of
algorithms treat binary inputs.
The rest of the work is arranged in this way. In the second section, some decoders
are presented as cognate works. The third section presents the suggested serial
concatenation of PHR and SDHT. The fourth section gives detail about our proposed
algorithm experiments. Finally, we give a conclusion through the fifth section.

2 Related Works

Researchers have investigated soft-decision decoding by diverse methods. The

earliest solutions were based on probabilistic and algebraic approaches. For example,
Chase-2 and HR algorithms. Chase-2 [1] is a decoder that employs a list of most
probable error patterns. HR [2] is a soft decision decoder that uses a symbol by
symbol decoding.
In the last decades, metaheuristics approaches were suggested. For instance the
use of genetic algorithms, compact genetic algorithms, neural networks and algo-
rithm A*, which is an algorithm that employs a strategy priority-first search. In [3]
the authors have used this algorithm to approach the decoding problem. GADEC [4]
is a Genetic Algorithm for Decoding. This decoder presents a significant advantage
over [3], indeed, GADEC is appropriate to be implemented on several parallel archi-
tectures. The authors of [5] have introduced evolutionary decoders. These decoders
were obtained thanks to an appropriate chromosome representation and some special
mutations and crossover operations. HDGA [6] is a Hard Decoder conceived by
Genetic Algorithms. In addition, [6] authors have introduced a novel soft decoder
at the base of HDGA and Chase decoder (SDGA). DDGA [7] is a Dual Domain
decoding Genetic Algorithm. The authors of [8] have proposed two dual domain soft
decision decoders at the base of compact Genetic Algorithm (cGA). The results have
shown the benefit of using larger tournament size. CGAD is a Compact Genetic Algo-
rithm Decoder described in [9]. To examine their decoder effectiveness, the authors
have deployed it for BCH, Quadratic Residue (QR) and Reed-Solomon (RS) codes
over Additive White Gaussian Noise (AWGN) channel. RNN [10] is a Recurrent
Neural Network architecture for decoding linear block codes. This solution provides
analogous Bit Error Rate (BER) results as regards the feed-forward neural network
with notably less parameters. The same authors of [10] have presented in [11] an
architecture of recurrent neural decoding at the base of successive relaxation tech-
nique. This method can achieve important performances by employing only a single
learnable parameter. In [12] neural net decoding algorithms for nonbinary codes
were considered. The authors have suggested analytic methods for synapse weight
coefficients calculating.
High Performance Decoding by Combination of the Hartmann Rudolph Decoder … 783

SDHT [13] is a Soft decision Decoder at the base of Hash Techniques and
syndrome decoding. The originators of [13] have introduced two hard decision
decoders named HSDec [14] and HWDec [14], and Chase-HSDec decoding algo-
rithm [15], which exploits HSDec in the Chase-2 decoder with the aim of minimizing
its temporal complexity.
In [16], it was proved that off-the-shelf Deep Learning decoding algorithms can
achieve the maximum a posteriori (MAP) decoding performance for short length
codes. In [17] a performance comparison of novel Chien search and syndrome blocks
designs is given for BCH and Reed-Solomon codes.

3 The Proposed Serial Concatenation

3.1 HR Algorithm

Hartmann & Rudolph algorithm is a symbol by symbol decoding algorithm. It

employs a probabilistic approach for deciding if the bit rj of the received sequence r
is equal to 0 or 1. In fact, HR employs the totality of 2n−k dual codewords. Formula
(1) represents how HR decides if the mth bit of the decoded word c is equal to 1 or
0.
 2n−k n 1−φl c 1 ⊕δml
cm = 0 i f j=1 i=1 ( 1+φl )
jl >0

(1)
cm = 1 other wise

1 if i = j m |1)
Where δi j = and φm = Pr(r
(rm |0)
0 other wise
c⊥jl is the l th bit of the j th dual codeword.

3.2 Soft Decision Decoding by Hash Techniques (SDHT)

Let us note the linear code by C(n, k, d), k is the code dimension, n is its length
and d is the
 minimum distance between all its distinct vectors. Moreover, let us note
t = d−1 2
the error correcting capability of C. The Soft Decision Decoding by Hash
Techniques operates as following:
784 H. Faham et al.

3.3 Our Proposed Concatenation Algorithm

HR has proposed to exploit all the dual codewords. This makes the temporal
complexity of this algorithm very high especially for codes whose of parity bits
number is relatively large. With the aim of overcoming this constraint, Nouh et al.
[18] have suggested exploiting just M dual codewords, and then the HR is used only
in some symbols of the received sequence. Therefore, the algorithm will be named
Partial Hartman-Rudolph (PHR). In [19] we have applied the same idea. The formula
(1) becomes (2).
  M n 1−φl c 1 ⊕δml
cm = 0 i f j=1 i=1 ( 1+φl )
jl >0

(2)
cm = 1 other wise

The fact of reducing the used dual codewords number influences the decoding
quality. It is why we propose to reprocess the sequence returned by HR by a word-
by-word decoder, which is the SDHT algorithm.
The PHR-SDHT decoder works as follows:
High Performance Decoding by Combination of the Hartmann Rudolph Decoder … 785

4 Simulation Results and Discussions

With the intention of showing our suggested scheme efficacy, we provide in this
section the PHR-SDHT simulation results for some BCH codes. Some comparisons
of PHR-SDHT performances with other competitors are also given. Table 1 describes
the employed experiment parameters.
The performances of decoding algorithms are traced by means of BER for each
Signal to Noise Ratio (SNR = Eb /N0 ). Coding gain comparisons are achieved refer-
encing the case where the transmission is performed without encoding on the trans-
mitter side and without decoding on the receiver side. In this case, Bit Error Rate
achieves 10–5 according to an SNR = 9.6 decibels through AWGN channel. This last
is regarded like binary channel exploiting BPSK modulation.
In Fig. 1(a), (b) and (c), the PHR-SDHT decoding performances are presented
for certain BCH codes with lengths from 31 to 127. These figures show the presence
of a significant coding gain. For instance, the coding gain is superior to 4 dB for
BCH(31, 16, 7), nearly 4 dB for BCH(63, 45, 7) and nearly 3 dB for BCH(127, 113, 5).
In Fig. 2(a) and (b), results comparison between PHR-SDHT, Chase-2 [1]
and Chase-HSDec [15] decoding algorithms is presented for BCH(31, 16, 7) and
BCH(31, 21, 5). From these figures, we deduce that our proposed decoding algorithm
guarantees better performances than the two other decoders do.

Table 1 Experiment
Channel Additive White Gaussian
parameters
Noise
Modulation Binary Phase Shift Keying
(BPSK)
Minimum transferred blocks 1000
Minimum residual errors 200
786 H. Faham et al.

Fig. 1 PHR-SDHT results 10

-1

for certain BCH codes with BCH(31,26,3);s=3

BCH(31,21,5);s=5
length: a 31, b 63 and c 127
BCH(31,16,7);s=7
-2
10

BER
-3
10

-4
10

-5
10
1 2 3 4 5 6 7
SNR

(a)

-1
10
BCH(63,57,3);s=3
BCH(63,51,5);s=5
BCH(63,45,7);s=4
-2
10
BER

-3
10

-4
10

-5
10
1 2 3 4 5 6 7
SNR

(b)

-1
10 BCH(127,120,3);s=2
BCH(127,113,5);s=3

-2
10
BER

-3
10

-4
10

-5
10
1 2 3 4 5 6 7
SNR

(c)
High Performance Decoding by Combination of the Hartmann Rudolph Decoder … 787

In Fig. 3, results comparison between PHR-SDHT, SD1 [20], GADEC [4] and
PHR-ARDecGA [18] decoders is made for BCH(63, 51, 5). The comparison results
confirm that PHR-SDHT exceeds its competitors.
With the aim of showing our serial concatenation effectiveness, a comparison of
the suggested scheme PHR-SDHT and Hartmann Rudolph algorithm is presented by
means of Fig. 4 for BCH(31, 21, 5) code. This comparison indicates that the proposed
combination of Hartmann Rudolph partially exploited and SDHT guarantees great
performances using a small part of the dual codewords. The obtained good results are
based on just 218 dual codewords, which means that we have employed just 21.29%
from dual code space. Table 2 presents the reduction rate of used dual codewords.

Fig. 2 Results comparison 10

-1

Chase-2
between PHR-SDHT, Chase-HSDec
Chase-2 and Chase-HSDec PHR-SDHT;s=7
for: a BCH(31, 16, 7) and 10
-2

b BCH(31, 21, 5)
BER

-3
10

-4
10

-5
10
1 2 3 4 5 6 7
SNR

(a)
-1
10
Chase-2
Chase-HSDec
PHR-SDHT;s=5
-2
10
BER

-3
10

-4
10

-5
10
1 2 3 4 5 6 7
SNR

(b)
788 H. Faham et al.

-1
SD1
10 GADEC
PHR-ARDecGA
PHR-SDHT;s=5
-2
10

BER
-3
10

-4
10

-5
10
1 2 3 4 5 6 7
SNR

Fig. 3 Results comparison between PHR-SDHT, SD1, GADEC and PHR-ARDecGA for
BCH(63, 51, 5) code

-1
10
HR
PHR-SDHT
-2
10
BER

-3
10

-4
10

-5
10
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
SNR

Fig. 4 Results comparison between HR and PHR-SDHT for BCH(31, 21, 5)

Table 2 Reduction rate of used dual codewords

Code Number of the used dual codewords Reduction rate of used dual
Hartmann Rudolph PHR-SDHT codewords

BCH(31, 21, 5) 231–21 = 210 = 1024 218 78,71%

To make sure of the temporal efficiency of our proposed decoder, we have plotted
in the Fig. 5 the ratio of the required run time of PHR-SDHT to the one of HR
algorithm for BCH(31, 16, 7). The reduction in run time is between 80 and 84%. We
deduce that PHR-SDHT has allowed us to minimize in a very large way the run time
of the HR algorithm. This proves the efficiency and the rapidity of our concatenation
scheme.
High Performance Decoding by Combination of the Hartmann Rudolph Decoder … 789

Fig. 5 PHR-SDHT to 0,20

BCH(31,16,7)
Hartmann Rudolph required

Run Time of PHR-SDHT / Run time of HR

runtimes ratio for
0,19
BCH(31,16,7)

0,18

0,17

0,16
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
SNR

5 Conclusions

We have introduced a powerful decoder at the base of a serial concatenation between

the Hartmann Rudolph algorithm partially exploited and the SDHT algorithm in
order to decode Bose-Chaudhuri-Hocquenghem codes. Performances analysis have
shown that the suggested decoding algorithm offers very satisfying BER results with
regard to the separate use of combined decoders. In addition, the partial use of the HR
algorithm has reduced considerably its temporal complexity. For instance, we have
used just 21.29% of the dual space for the BCH(31, 21, 5) code while maintaining
good decoding quality.

References

1. Chase D (1972) A class of algorithms for decoding block codes with channel measurement
information. IEEE Trans Inf Theory 18:170–181
2. Hartmann CRP, Rudolph LD (1976) An optimum symbol-by-symbol decoding rule for linear
codes. IEEE Trans Inf Theory 22:514–517
3. Han YS, Hartmann CRP, Chen CC (1991) Efficient maximum likelihood soft-decision decoding
of linear block codes using algorithm A*, Technical Report SU-CIS-91-42, School of Computer
and Information Science, Syracuse University, Syracuse, NY 13244, December
4. Maini H, Mehrotra K, Mohan C (1994) Soft-decision decoding of linear block codes using
genetic algorithms. In: Proceedings of the IEEE international symposium on information theory,
Trondheim, Norway, p 397
5. Cardoso FA, Arantes DS (1999) Genetic decoding of linear block codes. In: Proceedings
of international conference on telecommunications congress on evolutionary computation,
Washington, DC, USA, pp 2302–2309
6. Azouaoui A, Chana I, Belkasmi M (2012) Efficient information set decoding based on genetic
algorithms. Int J Commun Netw Syst Sci 5(7):423–429
7. Azouaoui A, Belkasmi M, Farchane A (2012) Efficient dual domain decoding of linear block
codes using genetic algorithms. J Electr Comput Eng. Article ID 503834
790 H. Faham et al.

8. Berkani A, Azouaoui A, Belkasmi M, Aylaj B (2017) Improved decoding of linear block

codes using compact genetic algorithms with larger tournament size. Int J Comput Sci Issue
14(1):15–24
9. Azouaoui A, Berkani A, Belkasmi M (2012) An efficient soft decoder of block codes based on
compact genetic algorithm. Int J Comput Sci Issue 9(5), No 2:431–438
10. Nachmani E, Marciano E, Burshtein D, Béery Y (2017) RNN Decoding of linear block codes.
arXiv:1702.07560v1 [cs.IT]
11. Nachmani E, Marciano E, Lugosch L, Gross WJ, Burshtein D, Béery Y (2018) Deep learning
methods for improved decoding of linear codes. IEEE J Sel Topics Signal Process 12(1)
12. Butov VV, Dumachev VN, Fedyaeva SE (2020) Neural net decoders of nonbinary codes. J
Phys Conf Ser 1479:012089
13. El Kasmi Alaoui MS, Nouh S, Marzak A (2019) High speed soft decision decoding of linear
codes based on hash and syndrome decoding. Int J Intell Eng Syst 12(1)
14. El Kasmi Alaoui MS, Nouh S, Marzak A (2019) Two new fast and efficient hard decision
decoders based on hash techniques for real time communication systems. In: lecture notes in
real-time intelligent systems, RTIS 2017. Advances in intelligent systems and computing, vol
756. Springer, Cham
15. El Kasmi Alaoui MS, Nouh S, Marzak A (2017) A low complexity soft decision decoder for
linear block codes. In: Proceedings of the first international conference on intelligent computing
in data sciences
16. Gross WJ, Doan N, Ngomseu Mambou E, Ali Hashemi, S (2020) Deep learning techniques for
decoding polar codes. In: F-L Luo (ed.) Machine learning for future wireless communications
17. Elghayyaty M et al (2020) Performance comparison of new designs of Chien search and
syndrome blocks for BCH and Reed Solomon codes. Int J Commun Netw Inf Secur 12(2)
18. Nouh S, Aylaj B (2018) efficient serial concatenation of symbol by symbol and word by word
decoders. Int J Innov Comput Inf Control 14
19. Faham H, El Kasmi Alaoui MS, Nouh S, Azzouazi M (2018) An efficient combination between
Berlekamp-Massey and Hartmann Rudolph algorithms to decode BCH codes. Periodic Eng
Nat Sci 6(2):365–372
20. Jung B, Kim T, Lee H (2016) Low-complexity non-iterative soft-decision BCH decoder
architecture for WBAN applications. J Semicond Technol Sci 16(4)
Automated Assessment of Question
Quality on Online Community Forums

Harish Rithish, Gerard Deepak, and A. Santhanavijayan

Abstract People around the world often rely on online community forums for
answers to their queries. These forums have become hugely popular in the last decade,
leading to a spurt in the number of users and questions. For a better user experience,
quality monitoring is essential. However, manual moderation of millions of questions
is infeasible. Prior works mostly rely on handcrafted features which is ineffective
or use community feedback as part of learning which makes them unsuitable for
monitoring during question creation. In this work, we use recent deep learning tech-
niques to assess the quality of questions in online community forums at creation time.
We evaluate our model on the StackOverflow dataset that contains 60000 questions
across three qualities. Our model achieves an F1 score of 0.92 on this dataset.

Keywords Question quality prediction · Deep learning · Community Q&A ·

StackOverflow

1 Introduction

Online community forums allow for the exchange of knowledge between people
across the world. Over the past decade, an exponential rise in smartphones and
Internet penetration has made these online forums accessible to people across the
world. Question & Answer forums allow people to get answers directed specifically
to their question, rather than having to search haphazardly from a myriad of resources.
Moreover, users are able to get answers from experts in their field at free of cost.
These advantages of online forums have resulted in their substantial growth, with an
enormous number of users across the world relying on these sites for solutions [1]
(Fig. 1).
The rise in users of online forums has naturally led to a tremendous increase in the
number of questions being raised on these forums. To maintain the standard of these
questions, they are either checked manually for their quality through moderators or

H. Rithish (B) · G. Deepak · A. Santhanavijayan

Department of Computer Science and Engineering, National Institute of Technology,
Tiruchirappalli, India

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 791
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_72
792 H. Rithish et al.

Fig. 1 In this paper, we are interested in classifying user-generated questions on online forums
automatically, based on its quality

is left to the community to flag questions of poor quality. However, such manual
moderation on millions of questions is ineffective and infeasible. Questions that are
of low quality often result in users spending more time understanding them, either
by asking clarification or by editing the question. Further, questions with low-quality
have been observed to receive answers of poorer quality [2]. In contrast, high-quality
questions receive more constructive answers which also result in long-term value with
users being able to refer them in the future. Overall, low-quality questions reduce the
productivity of users on the site. Thus, we need an automated solution that can flag
potentially low-quality questions. Such a system can help analyze the quality of the
question during creation, and suggest the user to improve the quality before posting
them. They can also be used as input to ranking mechanisms during question search.
Forums such as Quora, Reddit, Stack Exchange, Yahoo Answers and Wiki
Answers serve the general audience and contain discussions on a variety of domains.
Along with such forums, there co-exists another spectrum of online forums that
are dedicated to a particular community, such as Brainly (academic), Avvo (legal),
Sharecare (health), and MadSci Network (Science). Here, we focus on questions
from the StackOverflow website, which is a dedicated forum to discuss queries in
programming. StackOverflow have tried maintaining the quality of these forums by
asking users to follow certain guidelines1 . Despite that, Correa and Sureka [3] esti-
mate around 293 K questions from the period 2009 to 2013 were deleted due to being
off-topic or containing low quality content.
In this work, we automatically assess the quality of StackOverflow questions
through recent deep learning techniques. We work with a dataset of 60000 questions
collected between 2016 and 2020 and released by the StackOverflow platform. We
develop an end-to-end pipeline that pre-processes the input question, extracts relevant
features and classifies them into one of three categories. We focus only on inputs that
an user has access to at creation time, allowing for instant flagging and suggestions.

1 http://stackoverflow.com/help/how-to-ask.
Automated Assessment of Question Quality … 793

In the remaining part of the paper, Sect. 2 discusses prior works on text classi-
fication and question quality assessment, Sect. 3 provides information on architec-
ture, Sect. 4 elaborates on dataset used and implementation details, Sect. 5 shows
experiments conducted and their corresponding results, and Sect. 6 concludes our
work.

2 Related Work

2.1 Text Classification

Question classification takes its root in text classification. This domain of text clas-
sification began when IBM stared investigating into techniques for information
retrieval in the 1950’s. One of the earliest works in this domain were by Maron
and Kuhns [4], where they relied on keywords to classify text. In the following
decades, researchers designed classification methods that only worked for a specific
field. These methods were designed with the knowledge of experts in the field by
defining field-specific rules. Naturally, these approaches did not generalize well to
other fields and started losing popularity when machine learning methods gained
prominence in the 1990’s. Machine learning was used to extract features instead of
relying on rules and then classify the text into relevant categories, based on these
features. Some of the earlier machine learning techniques used for text classification
include Naive Bayes [3], Nearest Neighbour [6], Linear Discriminant Analysis [7],
Decision Trees [8], Logistic Regression [8] and Support Vector Machines [9].
Techniques in classical machine learning for text classification evolved over the
next two decades, leading to practical deployment of these systems for a limited
number of use-cases. However, these methods did not scale well to large datasets
or content with complex semantics. With the rapid growth of deep learning in this
past decade, however, text classification systems have matured significantly and
are nowadays being deployed widely. Specifically, various improvements in vanilla
Recurrent Neural Networks have facilitated this progress in text classification [11–
13]. While all the above methods discuss several approaches for text classification,
we specifically focus on the task of assessing questions based on its quality.

2.2 Question Quality Assessment

Agichtein et al. [14] consider inputs from both the question and their answers. They
establish that there is a strong relationship between the quality of a question raised
and their feedback from users. Shah et al. [15] approach this task differently, by first
categorizing a question and then extract category-specific features for estimating
the quality of a question. Ponzanelli et al. [16] work on identifying good quality
794 H. Rithish et al.

questions that have been mistakenly flagged for review. All the above approaches,
however, rely on user answer information or community feedback for estimating
question quality.
Tausczik and Pennebaker [17] show that a user’s reputation influences the quality
of a question. They design their model using votes received by a question as a feature
to determine question quality. Ravi et al. [18] give importance to the length of the
question content and title. They extend [17] by normalizing votes received by a
question with the topic’s popularity. In both these approaches, they use information
outside of what is available at question creation time, which makes them ineffective
to provide immediate feedback while a user drafts a question. In contrast to all the
above approaches, we use recent techniques in natural language processing along
with inputs (title, body text, tags) that are available at creation time to assess the
quality of a question.

3 Automated Quality Assessment of Community Questions

See Fig. 2.

Fig. 2 We start by concatenating multiple components of our question into a single input. Our
combined input is then passed through a set of preprocessing steps. Word embeddings are obtained
for the processed text and then fed to a two-layer Bidirectional LSTM. The extracted features are
then passed to a Fully Connected layer. Finally, we assess the quality of the input question by
passing through a Softmax layer
Automated Assessment of Question Quality … 795

3.1 Task Definition

In this section, we define the quality assessment of community questions task. Given
an input title H, body text B and a set of tags T = <t1 , t2 ,… t|T| >, our goal is to classify
the question into one of C qualities.

3.2 Data Preprocessing

We start by concatenating multiple components of our question, viz., title, body text
and associated tags into a single input, making it easier for the model to process
further. We then filter out special characters and punctuation present in the input.
We follow this by tokenizing the input, based on the top K most frequent words in
the training set. We finally limit the size of our question to a maximum of L words,
by truncating if the question is longer than L words or by padding if the question is
lesser than L words.

3.3 Word Embeddings

Each word in the question is first converted to its corresponding word embedding.
Embeddings are a representation of text, with each word being encoded as a feature
vector. We employ Global Vectors for Word Representation (GloVe) [19] for this
purpose. GloVe is an unsupervised learning algorithm for obtaining vector repre-
sentations for words [19]. GloVe does not rely just on the local context, but also
incorporates global statistics to obtain word embeddings. The word embeddings are
trained on aggregated global word-word co-occurrence statistics from a corpus. In
this work, we pre-train the model on the Twitter corpus and obtain a 25-d feature
vector for each word.

3.4 Model

We extract features from a set of word embeddings by passing them through a

series of layers. We start by constructing a 2-layer deep Bi-Directional Long Short-
Term Memory (Bi-LSTM) networks. LSTMs are a special kind of Recurrent Neural
Networks that are capable of learning long-term dependencies. This allows the model
to encode very-long questions, while retaining contextual information from all parts
of the text. However, a unidirectional LSTM only has past contextual information
since the only inputs it has seen are only from the past. To overcome this, we use
Bi-Directional LSTMs which runs the questions in two ways, one from the beginning
796 H. Rithish et al.

to the end and another from the end to the beginning. Through this, we obtain two
hidden states, one with past context and another with future context, which we then
combine through concatenation. We use a hidden state of 128 units and thus, after
concatenation, we obtain a 256-d feature vector. Additionally, we incorporate dropout
and recurrent dropout to minimize overfitting. The sequence of L outputs from the
earlier computation, are passed through another layer of Bi-LSTM to increase the
complexity of the model function. We retain the same structure for the second layer,
and only tweak hyper-parameter values. We pass the output of the 2-layer Bi-LSTM
through a Fully Connected layer, whose output forms the final feature vector for
our input question. The Fully Connected layer transforms the 256-d feature vector
into a |C|-d vector, with each unit representing the unnormalized log value of a its
classification label. Finally, we obtain the classification labels for the question by
passing the unnormalized log values through a Softmax layer.

3.5 Loss Function

Since our task is classification, we employ the standard Cross Entropy loss.

4 Datasets and Implementation Details

4.1 Datasets

Stack Overflow is a widely used question and answering site for programming ques-
tions. The site is a community-based platform, where users can raise new questions
or provide answers to existing questions of other users. Recently, Stack Overflow has
collected a dataset of 60000 questions between 2016 and 2020 and released the data
publicly. A question consists of a title, body and one or more tags. The task involves
classifying a question into one of the three categories:
• HQ: High-quality posts with 30+ score and without a single edit.
• LQ (Close): Low-quality posts that were closed by the community without a single
edit.
• LQ (Open): Low-quality posts with a negative score and with multiple community
edits. However, they still remain open after the edits.
The dataset is balanced, with 60000 questions being evenly distributed between the
three categories. The training and validation set contain 45000 and 15000 questions
respectively, with the questions from each category being sampled in a stratified
manner (Fig. 3).
Automated Assessment of Question Quality … 797

(a) HQ (b) LQ (Close) (c) LQ (Open)

Fig. 3 Word clouds for different question categories

(a) Loss vs Epoch graph (b) Accuracy vs Epoch graph

Fig. 4 Training plots on the Stack Overflow dataset

4.2 Implementation Details

For tokenization, we set the number of words K = 10000. We limit the number of
words in each question, L to 300. For the Bi-LSTM module, we set dropout and
recurrent dropout to 0.4 for the first layer and decrease the value to 0.2 for the second
layer. We use Adam for optimization, with an initial learning rate of 10−2 and reduce
the learning rate by a factor of 0.4 whenever the loss plateaus. We train the model
on an NVIDIA-1080 Ti GPU for 15 epochs, with a batch-size of 256. We show the
training plots of our model in Fig. 4.
798 H. Rithish et al.

5 Experiments and Results

5.1 Quantitative Results

We present the quantitative results of the model on the StackOverflow dataset in

Table 1. However, on low-quality questions that are closed as well as high quality
questions, the model performance drops to an F1-score of 0.87 and 0.84, respectively.
We observe that the model performs extremely well on low-quality questions that
are still open with an F1-score of 0.99. We now present the confusion matrix for the
predictions in Fig. 5. Validating the results seen in the table, we observe that the model
performs well in identifying low-quality questions that are still open. Interestingly,
the model confuses high-quality questions with low-quality questions that are closed,
and vice-versa.

Table 1 Quantitative results

Class Precision Recall F1-score
on the Stack Overflow
validation set HQ 0.88 0.80 0.84
LQ (Close) 0.88 0.87 0.87
LQ (Open) 0.99 0.99 0.99
Overall 0.92 0.92 0.92

Fig. 5 Confusion matrix for

predictions on the
StackOverflow validation set
Automated Assessment of Question Quality … 799

Fig. 6 Qualitative examples of HQ questions that are correctly predicted by our model

Fig. 7 Qualitative examples of LQ (Close) questions that are correctly predicted by our model

Fig. 8 Qualitative examples of LQ (Open) questions that are correctly predicted by our model

5.2 Qualitative Results

We show qualitative examples of correct model predictions on the StackOverflow

validation set, with high-quality questions in Fig. 6, low-quality questions that are
closed in Fig. 7 and low-quality questions that are still open in Fig. 8.

6 Conclusion

There has been a rapid growth in users and their queries in online community forums,
necessitating an automated quality control mechanism during question creation time.
In this paper, we design a model that takes title, body text and tags associated with a
question as input and predicts the quality of a question. Our model preprocesses the
inputs, embeds them into word vectors using GloVe, extracts features using a 2-layer
deep bi-directional LSTM and finally uses a Softmax layer for classification. We
evaluate our model on the StackOverflow dataset that consists of 60000 questions
800 H. Rithish et al.

and 3 categories of quality. Our model achieves an F1-score of 0.92, indicating that
our system can be deployed for practical use.

References

1. Agichtein E, Castillo C, Donato D, Gionis A, Mishne G (2008) Finding high-quality content

in social media. In: Proceedings of the 2008 international conference on web search and data
mining, pp 183–194
2. Movshovitz-Attias D, Movshovitz-Attias Y, Steenkiste P, Faloutsos C (2013) Analysis of the
reputation system and user contributions on a question answering website: Stackoverflow. In:
International conference on advances in social networks analysis and mining, pp 886–893
3. Correa D, Sureka A (2014) Characterization and modeling of deleted questions on stack
overflow
4. Maron ME, Kuhns JL (1960) On relevance, probabilistic indexing and information retrieval. J
ACM (JACM) 7(3):216–244
5. Li YH, Jain AK (1998) Classification of text documents. Comput J 41(8):537–546
6. Weiss S, Kasif S, Brill E (1996) Text classification in use net newsgroups: a progress report.
In: Proceedings of the AAAI spring symposium on machine learning in information access,
pp 125–127
7. Hull D, Pedersen J, Schutze H (1996) Document routing as statistical classification. In: AAAI
spring symposium on machine learning in information access, vol 12, pp 49–54
8. Lewis DD, Ringuette M (1994) A comparison of two learning algorithms for text categorization.
In: Third annual symposium on document analysis and information retrieval, vol 33, pp 81–93
9. Schutze H, Hull DA, Pedersen JO (1995) A comparison of classifiers and document represen-
tations for the routing problem. In: Proceedings of the 18th annual international ACM SIGIR
conference on Research and development in information retrieval, pp 229–237
10. Zhang D, Lee WS (2003) Question classification using support vector machines. In: Proceed-
ings of the 26th annual international ACM SIGIR conference on Research and development in
information retrieval, pp 26–32
11. Chen T, Xu R, He Y, Wang X (2017) Improving sentiment analysis via sentence type
classification using BiLSTM-CRF and Cnn. Expert Syst Appl 72:221–230
12. Ranjan MNM, Ghorpade Y, Kanthale G, Ghorpade A, Dubey A (2017) Document classification
using LSTM neural network. J Data Mining Manage 2(2):1–9
13. Zhou C, Sun C, Liu Z, Lau F (2015) A C-LSTM neural network for text classification. arXiv
preprint arXiv:1511.08630
14. Srba I, Bielikova M (2016) A comprehensive survey and classification of approaches for
community question answering. ACM Trans Web (TWEB) 10(3):1–63
15. Shah C, Kitzie V, Choi E (2014) Questioning the question–addressing the answer-ability of
questions in community question-answering. In: 2014 47th Hawaii international conference
on system sciences, pp 1386–1395. IEEE
16. Ponzanelli L, Mocci A, Bacchelli A, Lanza M, Fullerton D (2014) Improving low quality stack
overflow post detection. In: 2014 IEEE international conference on software maintenance and
evolution, pp 541–544. IEEE
17. Tausczik YR, Pennebaker JW (2011) Predicting the perceived quality of online mathematics
contributions from users’ reputations. In: Proceedings of the SIGCHI conference on human
factors in computing systems, pp 1885–1888
18. Ravi S, Pang B, Rastogi V, Kumar R (2014) Great question! Question quality in community
Q&A
19. Pennington J, Socher R, Manning CD (2014) Glove: global vectors for word representation.
In: Proceedings of the 2014 conference on empirical methods in natural language processing
(EMNLP), pp 1532–1543
QFRDBF: Query Facet Recommendation
Using Knowledge Centric DBSCAN
and Firefly Optimization

Deepak Surya, Gerard Deepak, and A. Santhanavijayan

Abstract The Internet contains approximately 40 trillion gigabytes of data. A vast

amount of information is presented to the user for their query. With such enormous
search results, it becomes impossible for the user to refine the results to find the
contents that match their interest. Almost 92% of search traffic clicks of Google
are confined only to the first page. Also, Google identifies that about 16% to 20%
of searches are new each year. Hence, a method that enables the users to filter the
search results based on some attributes to obtain the web content that matches their
intention is vital. One such method is query facets. Query facets enable the user to
filter the web results, making it easy for them to seek their desired results. This paper
presents a novel approach to extract query facets. Query words are obtained from the
queries upon query preprocessing. TF-IDF is applied to the preprocessed dataset and
the query words to reorder the terms based on the frequency. The reordered terms
are sorted based on concept similarity, and a knowledge centric DBSCAN algorithm
is employed on the sorted items to generate facets.

Keywords Domain knowledge · DBSCAN · Facet recommendation · Knowledge

centric approach

1 Introduction

Query facets categorize web searches based on some attributes. Faceting technology
facilitates the user to narrow down the web searches based on their interest. It is
widely used in e-commerce websites, online libraries, and social networks. Facets
are generated based on the web content and they describe the properties of the
information. Most recent systems [1] have used query facets in order to achieve
feasible and promising recommendation results in the era of Semantic Web. The
web information for the search undergoes text analysis to identify the latent patterns.

D. Surya · G. Deepak (B) · A. Santhanavijayan

Department of Computer Science and Engineering, National Institute of Technology,
Tiruchirappalli, India

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 801
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_73
802 D. Surya et al.

For example, in an e-commerce website, facets are generated based on customer

reviews, ratings, and product descriptions. The faceted search enables the user to
classify the search content based on brands, size, average customer rating, packaging
option, condition, seller, discount, availability, price range, etc. An item can belong
to multiple faceted categories, and the user can set multiple facet classifications to
refine the search results. There is no predefined order to set the facets; instead, facet
classifications are set according to the user’s convenience to navigate the search
results.
Motivation: Queries that capture the user intention are collected. Upon preprocessing
the compiled queries, query words are obtained. Each query word’s significance in
the query collection is determined by employing the TF-IDF measure to the query
collection. Upon evaluation, the query terms are reordered based on their frequency.
The reordered items are classified based on concept similarity. Finally, a knowledge
centric DBSCAN clustering algorithm is employed to the reordered items, and Firefly
with concept similarity is applied to the domain knowledge to generate query facets.
Contribution: This paper proposes a novel approach for the automatic generation of
facets from query words. The approach encompasses tags, keywords, and data labels
along with domain knowledge for query facet generation. A combination of concept
similarity and DBSCAN clustering algorithm is employed. Finally, the query facets
enrichment is performed based on the firefly optimization computation with concept
similarity. An overall F-Measure of 91.1% is achieved for the User Q dataset.
Organization: The remainder of the paper is organized in following manner. Section 2
of the paper presents the Related Works. Section 3 consists of the Proposed System
Architecture. The Implementation of the work is highlighted in Sect. 4 discusses
and the Results and Performance Evaluation are described in Sect. 5. Finally, Sect. 6
concludes the paper.

2 Related Work

Jiang et al. [1] used knowledge bases to mine query facets. Their approach covers a
wide range of information to generate query facets rather than only the top results
of the web search. Ramya et al. [2] put forth a method to extract facets for user
queries automatically. The lists are classified to generate the user query facets based
on HTML tags, free-text patterns, and repeat regions. Cosine Similarity is employed
to evaluate the similarity between the lists. The High-Quality Clustering algorithm
classifies the items, and the best items in each cluster are recommended to the user
QFRDBF: Query Facet Recommendation Using Knowledge … 803

as the facets. Vandic et al. [3] put forth a framework dynamic facet ordering in
e-commerce web applications. An automated algorithm ranks facets based on speci-
ficity and dispersion measures. Some specific issues to e-commerce, such as copious
numerical facets, classification of facets by their mutual attributes, and the chance of
multiple clicks, are also addressed by their work. Hervey et al. [4] study the extrac-
tion and ranking of facets in geospatial data portals. Their study proposes methods
of alternative interfaces and query processing pipelines to enhance the search for
geospatial data. Babu et al. [5] present a method to improve the user search experi-
ence by incorporating sub-queries and facets. Galán et al. [6] presented evaluation
approaches to compare region query for the DBSCAN algorithm and a novel region
query strategy. The overall number of region queries and the nearest neighbor search’s
complexity for each region query is minimized. The query strategies in each region
are comparatively assessed based on clustering effectiveness and efficiency. Cherrat
et al. [7] proposed an technique to advance the quality of segmenting the finger-
print images by employing K-means and DBSCAN algorithm. The quality of the
fingerprint images are enhanced by the application of Soble and TopHat filtering algo-
rithms. K-means clustering algorithm is used to obtain the precise segmentation of the
foreground and background region for each block in the image. The computing time
is reduced by using the local variance thresholding technique. Upon implementing K-
means clustering, DBSCAN clustering is applied to overcome the flaws of the former
one. Xue et al. [8] proposed a technique to match biomedical ontologies using the
firefly algorithm. Their work introduced the compact firefly algorithm (CFA), which
outperforms the existing state of art swarm intelligent algorithms. Kaushal et al. [9]
proposed a segmentation technique based on Firefly optimization to segment medical
images of breast cancer irrespective of the image type or modality. Their approach
is compared with the results obtained from the current advanced techniques, and the
results obtained certifies the performance of the proposed method. Xie et al. [10]
addresses the two main problems presented by the K-means clustering model: the
initialization sensitivity problem and local optima traps. These problems are fixed
by the methods proposed by the author in this paper Iwendi et al. [11] proposed
a TF-IDF algorithm with the temporal Louvain method to analyze the text from
various intelligent sensing systems. The proposed work enables analysts to make
reliable decisions by classing documents into hierarchical structures and presenting
the relationships among the variables in the document. The results obtained vali-
date the proposed work’s accuracy and shorter execution time over other existing
approaches. In [12–26] several semantic and ontological approaches in support of
the existing literature are depicted.
804 D. Surya et al.

3 Proposed System Architecture

Queries that represent the user intention are obtained. The collected queries undergo
query preprocessing to generate query words. The pre-processed dataset and query
words are assessed using TF-IDF and reordered based on the frequency. The TF-IDF
is a statistical measure used to determine how much each word in a document is related
to it. It is implemented in information retrieval, keyword extraction, and automated
text analysis. The term frequency and inverse document frequency convey a particular
word’s frequency in one document and a body of documents, respectively. Most of
the machine learning algorithms do not work with text but works with numbers.
The TF-IDF is used to vectorize the text. Text Vectorization process converts a text
document to a vector or array of numbers. It can represent a document as a vector of
numbers to be fed to a machine-learning algorithm. The term frequency and inverse
term frequency are computed as shown in the Figs. 1 and 2, respectively. The value
of TF-IDF is computed by estimating the product of the TF (term frequency) and
IDF (inverse document frequency).

(1)

(2)

(3)

The rearranged items are grouped based on concept similarity, and a knowledge
centric DBSCAN algorithm is employed to generate facets. DBSCAN is an agglom-
erative clustering algorithm which can create automatic categories and determine the
outliers. Two parameters are required to run the algorithm : the radius (maximum) of
a local neighborhood, the number of points (minimum) needed to grow the cluster.
Minimum domain knowledge is adequate to set the hyperparameters. The number
of clusters formed is not determined before training the dataset. It is very efficient
on large data sets as it works excellent with arbitrary clustering, cut noise, and label
outliers. The samples are mutually exclusive and non-exhaustive. Also, the domain
knowledge is fed to the firefly algorithm and concept similarity algorithm to create
facets.
QFRDBF: Query Facet Recommendation Using Knowledge … 805

Fig. 1 Proposed Architecture for QFRDBF Framework

4 Implementation

The dataset used in this work is the User Q dataset. The User Q dataset contains
90 queries in XML format, and each of those queries contains 100 documents.
The id, title, URL, description, rank, and text of the document, repeat region,
converted HTML list queries are found in each document. The implementation is
performed in Google’s Collaboratory environment in a computer with i7 processor
and 16 GB RAM. Experiments were conducted for News Domains, and Ontologies
were modelled for the Domains of various News. OntoCollab is used for modeling
and visualization of Ontology. The OntoCollab dynamically models ontologies and
Customized Web Crawlers are used to extract metadata for Ontology Modeling.
Hermit Reasoner is used for reasoning of Ontologies. NLTK framework is used to
perform the Natural Language Processing tasks. The concept similarity (CS) has
been employed to identify the semantically similar concepts in different Ontologies
and measure the similarity between concepts in an Ontology. The neighborhood set
806 D. Surya et al.

Fig. 2 F1 Score vs No. of Recommendations graph

and the feature set of a concept are used to determine its similarity with another
concept. The most related concepts to a user query can be obtained using the CS
measure. Concept maps of the ontologies can be exploited to extract the concepts.
Mostly, experts design the concept map with the help of the ontology developed by
the domain experts. The concept maps can also be automatically derived from the
ontologies that are used for comparison. An XML file is created for some specific
classes that comprise the representative and characteristic set of some classes in
Ontology. The characteristics and representatives obtained from the previous step
are utilized to calculate the concept similarity measure value. The obtained concept
similarity score is exploited to determine the most related concept to the user query.
QFRDBF: Query Facet Recommendation Using Knowledge … 807

Algorithm 1: The proposed algorithm for Query Facet Recommendation

Input: Multiword user query, dataset, and domain knowledge
Output: Query centric user relevant query facets.
Step 1: The queries obtained are synonymized using WordNet
Step 2: Real world knowledge is obtained from DBpedia.
Step 3: Jaccard between the queries, synonymized terms and the real-world knowledge
is calculated.
Step 4: Only the datapoints that are semantically similar are considered, i.e., A
datapoint in the data set is considered if Jaccard value is greater than 0.5.
Step 5: The database is modified to store only the semantically similar terms.
Step 6: DBSCAN clustering is employed on the modified database.
DBSCAN (MDB, jaccardFunc, epsilon, simPts)
clusterCount := 0
for each item I in the modified database DB
if label(I) ≠ undefined then continue
Neighbor Points NP := QueryRange(DB, jaccardFunc, I, epsilon)
if |NP| < simPts then
label(I) := Noise
continue
clusterCount := clusterCount + 1
label(I) := clusterCount
SeedSet S := NP \ {I}
for each point J in S
if label(J) = Noise then label(J) := clusterCount
if label(J) ≠ undefined then continue
label(J) := clusterCount
Neighbors NP := QueryRange(MDB, jaccardFunc, J, epsilon)
if |NP| ≥ simPts then
S := S ∪ NP
Step 7: The result obtained from the previous step is optimized using the Firefly
algorithm.
while (t < Max_Generation)
for u = 1 : N (Total no of fireflies N)
for v = 1 : u (N fireflies)
if (Uv > Uu) :
Vary attractiveness with distance Ω via log(sim(r,Ω));
Vary attractiveness with Pearson’s coefficient C via
exp(α – 0.25*C);
Push firefly u towards v;
Evaluate the latest solutions obtained, and accordingly,
update the intensity of light ;
end if
end for v
end for u
Rank the fireflies to determine the best;
end while
Results post-processing;
Results visualization;
end

The queries collected are synonymized using WordNet 3.0. WordNet 3.0 is used
to obtain semantically similar terms to the query, such as synonyms, hyponyms, and
meronyms. DBpedia is exploited to obtain real-world knowledge. Jaccard similarity
808 D. Surya et al.

between the queries, synonymized terms, and real-world knowledge is calculated.

Only the semantically similar data points are considered, i.e., A data point in the data
set is considered if the Jaccard value is greater than 0.5. The database is modified
to collect the semantically similar data points that are identified in the last step. The
DBSCAN clustering algorithm is employed on the modified database. The DBSCAN
clustered result is optimized by applying the firefly algorithm, where the objective
function is WebPMI.

5 Performance Evaluation and Results

The clusters obtained are evaluated using PRF, wPRF, Purity, R1, F1, F5, and NMI
to determine its quality and efficacy. Normalized Mutual Information (NMI) is an
external measure used to determine the quality of the clusters. Two step clustering
with different numbers of clusters can be compared using NMI values as it is a
normalized score. The NMI score for a cluster CL with respect to a class label CA is
calculated as depicted in Eq. 4. The similarity between two clusters is measured by
Random Index (RI). RI is a statistical measure used widely in clustering to determine
the accuracy. RI value can be calculated, as shown in Eq. 5. TRP, FLP, FLN, TRN
represents the number of True Positives, False Positives, False Negatives, and True
Negatives, respectively. The F-measure estimates the clusters’ accuracy describing
the class labels from the test’s harmonic mean of the precision and recall. The F-
measure of the cluster CL with respect to a class CA is computed as depicted in
Eq. 6.

(4)

(5)

The weighted total of maximum F-measure score for a cluster C gives the overall
F-Measure value. The overall F-measure value can be computed as shown in Eq. 7.
Of all the generated facets, some are useful than the other. Normalized Discounted
Cumulative Gain (nDCG) is employed to identify and extract the most useful facets
from a collection of facets. The nDCG method can be applied to the facet collection
to rank the facets and identify the useful ones. It is applied as depicted in Eq. 8.
Here, DCGk represents the cumulative gain obtained by correct ordering, and iDCGk
depicts the discounted gain of the ith facet. The nDCG measure can be segregated into
two kinds oF-Measure s: first, purity nDCG (fp−nDCG) and second, recall purity
nDCG (rp − nDCG). The former is based on each cluster’s original representation,
while the latter is based on overall facets. The product of DGi with the correctly
assigned item percentage yields the purity of the ith facet. The weights associated with
each item in fp-nDCG and rp-nDCG are illustrated in the Eq. 9 and 10, respectively.
QFRDBF: Query Facet Recommendation Using Knowledge … 809

(6)

(7)

DC G k
n DC G k = (8)
n DC G i
|C L| ∩ f i
wi = (9)
fi
|C L ∩ f i | |C L ∩ f i |
wi = (10)
fi CL

Table 1 compares the results obtained among the proposed approach QFRDBF,
QDMiner, and AEFTM (Automatic Extraction of Facets for user Query in Text
Mining) frameworks on the user Q dataset. The nDCG score of the QFRDBF is
higher than the nDCG score of AEFTM and QDMiner. The DCG score of QFRDBF,
AEFTM, QDMiner are 0.691, 0.738, and 0.938. The PRF and wPRF score of
QFRDBF is much greater than the other baseline models. The PRF score of the
proposed QFRDBF approach is 0.78 while that of QDMiner and AEFTM are 0.392
and 0.436, respectively. The wPRF scores of QDMiner, AEFTM, QFRDBF are
0.392, 0.436, and 0.78. The purity value of QFRDBF is greater than AEFTM and
QDMiner, as the similarity of the items found in clusters obtained from the proposed
approach is higher than the items belonging to the clusters generated from AEFTM
and QDMiner. The purity values of QDMiner, AEFTM, QFRDBF are 0.912, 0.932,
and 0.972. The proposed approach QFRDBF has a better R1 measure value than
AEFTM and QDMiner. The R1 measure value of QFRDBF is 0.967, while the R1
scores of AEFTM and QDMiner are 0.937 and 0.923. The F1 and F5 measure of
FRDBH are 0.941 and 0.882. FRDBH has higher scores than QDMiner and AEFTM
as 0.735, 0.689 are the scores of the former and 0.792, 0.59 are the scores of the
latter. Also, FRDBH has the highest NMI scores among the other baseline methods,
such as AEFTM and QDMiner. The NMI scores of FRDBH, AEFTM, and QDMiner
are 0.984, 0.819, and 0.878.

Table 1 Measures Indicating the Quality of Query Facets Recommended

nDCG PRF wPRF Purity R1 F1 F5 NMI
QDMiner 0.691 0.384 0.392 0.912 0.923 0.735 0.689 0.819
AEFTM 0.738 0.470 0.436 0.932 0.937 0.792 0.590 0.878
Proposed 0.938 0.720 0.780 0.972 0.967 0.941 0.882 0.984
810 D. Surya et al.

6 Conclusions

The queries that capture the user’s interest are collected, and query words are obtained
by preprocessing them. The query words are synonymized using WordNet 3.0, and
the real-world knowledge is obtained from DBpedia. The Jaccard similarity among
the query words, synonymized terms, and real-world knowledge are computed. Based
on the results, the semantically similar terms are identified and collected in a modi-
fied database. The DBSCAN clustering algorithm is employed on the modified
database, and the firefly optimization technique is applied to the result obtained
to generate query facets. The results achieved validates the efficacy and the perfor-
mance of the proposed framework, i.e., Facet Recommendation Using Knowledge-
Centric DBSCAN and Firefly (QFRDBF). It can be inferred from the results that the
proposed framework functions better than the other baseline frameworks. QFRDBF
has the highest PRF, wPRF, Purity, R1, F1, F5, and NMI scores among QDMiner
and AEFTM. A F1 and F5 measure of 91.1% and 88.2% are achieved.

References

1. Jiang Z, Dou Z, Wen JR (2016) Generating query facets using knowledge bases. IEEE Trans
Knowl Data Eng 29(2):315–329
2. Ramya RS, Raju N, Sejal N, Venugopal KR, Iyengar SS, Patnaik LM (2019) Automatic
extraction of facets for user queries [AEFUQ]. In: 2019 Fifteenth International Conference
on Information Processing (ICINPRO), pp 1–6. IEEE
3. Vandic D, Aanen S, Frasincar F, Kaymak U (2017) Dynamic facet ordering for faceted product
search engines. IEEE Trans Knowl Data Eng 29(5):1004–1016
4. Hervey T, Lafia S, Kuhn W (2020) Search facets and ranking in geospatial dataset search.
In: 11th International Conference on Geographic Information Science (GIScience 2021)-Part
I. Schloss Dagstuhl-Leibniz-Zentrum für Informatik
5. Babu D: Web search using automatically generated facets (2019). In: 2019 2nd International
Conference on Intelligent Computing, Instrumentation and Control Technologies (ICICICT),
vol 1, pp 37–40
6. Galán SF (2019) Comparative evaluation of region query strategies for DBSCAN clustering.
Inf Sci 502:76–90
7. Cherrat EM, Alaoui R, Bouzahir H (2019) Improving of fingerprint segmentation images based
on K-means and DBSCAN clustering. Int J Electric Comput Eng 2088–8708:9
8. Xue X (2020) A compact firefly algorithm for matching biomedical ontologies. Knowledge
and Information Systems, pp 1–17
9. Kaushal C, Kaushal K, Singla A (2020) Firefly optimization-based segmentation technique to
analyse medical images of breast cancer. Int. J. Comput. Math. 1–16
10. Xie H, Zhang L, Lim CP, Yu Y, Liu C, Liu H, Walters J (2019) Improving K-means clustering
with enhanced Firefly Algorithms. Appl Soft Comput 84:
11. Iwendi C, Ponnan S, Munirathinam R, Srinivasan K, Chang CY (2019) An efficient and
unique TF/IDF algorithmic model-based data analysis for handling applications with big data
streaming. Electronics 8(11):1331
12. Deepak G, Santhanavijayan A (2020) OntoBestFit: A Best-Fit Occurrence Estimation strategy
for RDF driven faceted semantic search. Computer Communications, vol 160, 284–298
QFRDBF: Query Facet Recommendation Using Knowledge … 811

13. Pushpa CN, Deepak G, Kumar A, Thriveni J, Venugopal KR (2020) OntoDisco: improving web
service discovery by hybridization of ontology focused concept clustering and interface seman-
tics. In: 2020 IEEE International Conference on Electronics, Computing and Communication
Technologies (CONECCT), pp 1–5. IEEE
14. Kumar A, Deepak G, Santhanavijayan A (2020) HeTOnto: a novel approach for conceptualiza-
tion, modeling, visualization, and formalization of domain centric ontologies for heat transfer.
In: 2020 IEEE International Conference on Electronics, Computing and Communication
Technologies (CONECCT), pp 1–6. IEEE
15. Deepak G, Kasaraneni D (2019) OntoCommerce: an ontology focused semantic framework
for personalised product recommendation for user targeted e-commerce. Int J Comput Aided
Eng Technol 11(4–5):449–466
16. Gulzar ZA, Anny L, Gerard D: Pcrs: personalized course recommender system based on hybrid
approach. Procedia Comput Sci 125, 518–524 (2018)
17. Deepak G, Teja V, Santhanavijayan A (2020) A novel firefly driven scheme for resume parsing
and matching based on entity linking paradigm. J Discrete Math Sci Cryptography 23(1):157–
165
18. Haribabu S, Sai Kumar PS, Padhy S, Deepak G, Santhanavijayan A, Kumar N (2019) A
novel approach for ontology focused inter- domain personalized search based on semantic set
expansion. In: 2019 Fifteenth International Conference on Information Processing (ICINPRO),
Bengaluru, India, 2019, pp 1–5. https://doi.org/10.1109/icinpro47689.2019.9092155
19. Deepak G, Naresh Kumar G, Sai Yashaswea Bharadwaj VSN, Santhanavijayan A (2019)
OntoQuest: an ontological strategy for automatic question generation for e-assessment using
static and dynamic knowledge. In: 2019 Fifteenth International Conference on Information
Processing (ICINPRO), pp 1–6. IEEE
20. Santhanavijayan A, Naresh Kumar D, Deepak G (2010) A semantic-aware strategy for auto-
matic speech recognition incorporating deep learning models. In: Intelligent System Design,
pp 247–254. Springer, Singapore
21. Deepak G, et al. (2019) Design and evaluation of conceptual ontologies for electrochemistry as a
domain. In: 2019 IEEE International WIE Conference on Electrical and Computer Engineering
(WIECON-ECE), IEEE
22. Deepak G, Priyadarshini JS (2018) Personalized and enhanced hybridized semantic algorithm
for web image retrieval incorporating ontology classification, strategic query expansion, and
content-based analysis. Comput Electric Eng 72:14–25
23. Kaushik IS, Deepak G, Santhanavijayan A (2020) QuantQueryEXP: a novel strategic approach
for query expansion based on quantum computing principles. J Discrete Math Sci Cryptography
23(2):573–584
24. Varghese L, Deepak G, Santhanavijayan A: an IoT analytics approach for weather forecasting
using Raspberry Pi 3 Model B+. In: 2019 Fifteenth International Conference on Information
Processing (ICINPRO), pp 1–5. IEEE (2019)
25. Shreyas K, Deepak G, Santhanavijayan A (2020) GenMOnto: A strategic domain ontology
modelling approach for conceptualisation and evaluation of collective knowledge for mapping
genomes. J Stat Manage Syst 23(2):445–452
26. Deepak P, Zakir M (2016) Enhanced neighborhood normalized pointwise mutual information
algorithm for constraint aware data clustering. ICTACT J Soft Comput 6(4)
SentiFusion (SF): Sentiment Analysis
of Twitter Using Fusion Techniques

Sampurn Anand, Chaya Vijaykumar, Gerard Deepak,

and A Santhana Vijayan

Abstract The Internet has become a domain for learning, idea exchange, feedback,
sharing thoughts or opinions, and many more. In the current world, Twitter has
been the most liable platform where people express their views and ideas about a
various range of products. In this paper, different feature extraction methods are
used to get more accurate sentiment analysis results. Firstly, tweet pre-processing
is performed using Euclidean Vector Length Similarity and other dictionaries. The
feature extraction is performed using (Jaccard Similarity and Trigrams) and (K-
means and KL-divergence). Then using tf-idf and Shannon Entropy, feature fusion
and weighting are done. At last, before final Sentiment Computation, Bagging is
accomplished using SVM classifiers and Decision Trees.

Keywords Tf-idf · K.L. Divergence · Sentiment analysis · SentiFusion · Trigrams

1 Introduction

In a layman’s language, Sentiment analysis is the analysis of emotions of different

texts. Technically it is the process of recognizing and classifying different opin-
ions using different computational algorithms expressed on a topic to determine
the author’s feelings towards it. Social media is a platform that generates a variety
of data-rich sentiments, which include tweets, blogs, comments, feedback, product
reviews, and many more. It also provides an opportunity for businesses by connecting
with their customers for advertising and improvising their products. One of the fore-
most and versatile social media platforms is Twitter. Amongst all the social media
platforms, Twitter is one of the many famous microblogging sites. Twitter Social
Feature Extraction [1] is quite important and has played a primary role for Feature

S. Anand · G. Deepak (B) · A. Santhana Vijayan

Department of Computer Science and Engineering, National Institute of Technology
Tiruchirappalli, Tiruchirappalli, Tamil Nadu, India
C. Vijaykumar
Department of Computer Science and Engineering, Bangalore Institute of Technology,
VV Puram, Bengaluru, India

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 813
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_74
814 S. Anand et al.

Extraction. It has more than 150 users with 2,400 tweets per day. Analyzing Twitter
is the best option because it contains various opinions, feedback, and reviews.
Extracting fascinating information and non-trivial patterns from unstructured data
is the primary concern of Opinion Mining and Sentiment analysis which are the
branches of Texting mining. Opinion mining (analyzing people’s opinions) is the
automated process of identifying and extracting subjective information from a given
document. The main aim of doing opinion mining is to classify the tweets as per
their sentiment values. The sentiment values will help in making the decision-making
process for different businesses and other applications.
In SentiFusion, we propose a combination of several methods, namely the Jaccard
Similarity, Trigrams, K-means Clustering with Silhouette coefficient, and Kullback
Leibler Divergence, which help in extracting the right features and makes the data
well-structured for classification. Following the feature extraction, the tf-idf has been
encompassed to act as a feature weighting scheme. Bagging is done using the two
most vital algorithms, which are Support-vector classifiers and Decision Trees. The
experiments have been conducted on Sentiment140 dataset models and have been
baselined with other models to show the proposed SentiFusion approach outperforms
different approaches. SentiFusion has achieved better performance measures in terms
of Precision, Recall, Accuracy, and F-Score.
In Sect. 2, Related Work has been described. Section 3 presents the proposed
system Sect. 4 consists of a description of Sentiment Computation. Implementation
with experimental results and Performance evaluation is discussed in Sect. 5. The
paper is concluded in Sect. 6.

2 Related Work

In [1], (Savitha et al. 2018) have proposed measures to extract sentiment variation’s
reasons on Twitter. The author used two approaches, Cosine Similarity and n-gram
similarity with Latent Semantic Analysis, to notice the similarities between tweets
and then concluded the reasons for sentiment variation by getting the highest simi-
larity value. This approach leads to an overall increase in the time complexity of the
system, which can be reduced. In [2], (Amit and Deshmukh 2013) used Pointwise
Mutual Information (PMI) and Chi-Square (X2) for feature selection. Further, for
sentiment classification, Naive Bayes (NB), Support Vector Machine (SVM), and
Maximum Entropy (ME) were used. The approach was mainly Lexicon-based. This
system lagged in the overall accuracy of judgment. POS was introduced as specific
polarity features in [3] by (Apoorv et al. 2011). They used tree kernels for feature
extraction. This model had outperformed the general state-of-the-art baseline but had
lower F-Scores. (Öztürk and Ayvaz 2018), in [4] performed comparative sentiment
analysis of the tweets retrieved and analyzed tweets in Turkish Language. For this
purpose, they used the Turkish sentiment analysis lexicon then simply classified data
using term frequencies. It had lower F-scores even if only the analysis of Tweets is
concerned. The effects of pre-processing were ignored in [5]. In it, the main focus was
SentiFusion (SF): Sentiment Analysis of Twitter … 815

on classification. (Jianqiang and Xiaolin 2017) tried to depict the six pre-processing
methods’ performances while using two feature models and four classifiers on five
different Twitter datasets. In this paper, the accuracy and F1-measure proposed
methods using classification classifiers were improved by expanding acronyms and
replacing negation. It doesn’t change the URLs, numbers, or stop words, which was
why it increased the time complexity of the system. The authors showed that Logistic
Regression and SVM classifiers are less sensitive than Naive Bayes and Random
Forest classifiers when various pre-processing methods were applied. Textual infor-
mation of Twitter messages and sentiment diffusion patterns were fused to perform
better [6]. But apart from the classifications, other sections of the system decreased
the overall F-scores. (Wang et al. 2020) used sentiment reversal phenomenon. They
proposed an iterative algorithm, which was named SentiDiff, to predict polarities of
the sentiment of Tweets. The PR-AUC improvement (Precision-Recall Area Under
Curve) between 5.09 and 8.38% was obtained using this model on Twitter senti-
ment classification tasks. Common sense knowledge (Domain Specific Ontology),
a very innovative sentiment analysis method, was used in [7] by (Ramanathan and
Meyyappan 2019). They created their Ontology based on ConceptNet. The combined
sentiment lexicon approach finally analyzed sentiments. In the final step, concerning
the domains, semantic orientations of domain-specific features were incorporated.
The only problem with this system was it led to an increase in the time complexity
because of the new Ontology. In [10–21] several ontological supporting strategies
have been discussed.

3 Proposed System Architecture

The architecture of the proposed system is depicted in Fig. 1. It can be interpreted as a

probabilistic model wherein the concepts of association rule mining, trigram models,
and statistical analysis have been integrated. It is an aspect level text classification.
Our proposed system will obtain a decimal value as a sentiment value of a particular
search option. This decimal value can describe the sentiments related to the search
better as it can be put on a linear scale of sentiments and then interpreted. This way,
the accuracy of the system can be increased.
The first and foremost step is gathering the datasets. Tokenization and pre-
processing Twitter data (Normalization using Euclidean Vector Length Similarity)
are performed next. Feature extraction is implemented using two methods explained
further. By combining the scores of all the tokens and keywords, feature-fusion is
accomplished. For feature weighting scheme Shannon Entropy is carried out. This
feature weighting method is used to compute the entropy of a probability distribution
of the appearance of different clusters in different sentiments. Lastly, Bagging using
SVM classifiers and Decision trees is conducted. All these steps rely on each other.
The output of one step corresponds to the input of the very next step. Implementation
of the methods mentioned here is done using the chosen data-set and different Python
816 S. Anand et al.

Fig. 1 Proposed system

architecture

libraries for text mining. For reference purposes, the proposed system architecture
will be called SentiFusion or SF in further mentions.

3.1 Data Preparation and Pre-processing

The gathered data set is fed to the next stage of pre-processing the data. In Tokeniza-
tion, tweets are broken down into words that are termed as ‘tokens’. Tokens are
normalized to standard form (all lower cases) using Euclidean Vector Length Simi-
larity. Acronyms are converted to their full forms using the Normalization process.
Using the stop word dictionary, each word is tagged as ST (Stop-words) or NST
(Non-Stop-words) for their Parts-of-Speech. Using the emoticon dictionary, emoti-
cons are removed and taken in as input to their sentiment values for each tweet. The
URLs are removed from the tweets. The inflected forms of words are grouped to
find out the base term using the procedure of Lemmatization. Each token is now
classified as (word, POS tag, English-word, Stop-word, Sentiment values). Directly,
these tokens with their classifications are fed as inputs to the next step of the process.
Dataset Description: In SentiFusion, the Sentiment140 data set is being used. This
data set has already been extracted using Twitter API after selecting the default
language as “English”. It contains 1,600,000 tweets. The six fields used from the
dataset here in the proposed model are the target, ids, date, flag, user, and text. The
SentiFusion (SF): Sentiment Analysis of Twitter … 817

link for the data set is https://www.kaggle.com/kazanova/sentiment140. The Target

field contains the tweet’s polarity, which won’t be utilized in this proposed system
architecture as the classification will be done in the fifth step. The other fields are
also not used in our approach. We are using the dataset for the tweets mainly, which
is pre-processed and used further for classification.

3.2 Feature Extraction

In this step for Extracting Features, two methods are adopted. In the First Method,
Jaccard Similarity and Trigrams are being used. In the Second Method, K-means
Clustering with the Silhouette coefficient and Kullback Leibler divergence is used.
Tf-Idf is common to both the methods for Classification. Data from the previous
step is fed to Method 1, where the initial level of feature extraction is done. The
data is provided to Method 2, which refines the feature. This step aims to extract and
summarize as many features from the tweets as possible, which will be fused and
classified in the next steps. Using two methods of feature extraction ensures better
accuracy.

Jaccard Similarity
Jaccard Similarity is the method of grouping the same data from two different data
sets. This algorithm is used to increase the accuracy of the proposed system and
shorten the given data sets. The pre-processed data from the previous step is compared
with different tweets and assigned different weights. The whole file is read for various
tweets to make comparisons between them for similar words. Every tweet will be
assigned a Jaccard Similarity ratio given by dividing the size of the intersection of
the two sets by the size of its union. Since we have already normalized our tweet
in the first step, repeating a word in a single tweet is avoided without changing the
results, a distinct feature of SentiFusion. So, by now, for a detailed analysis, all the
tokenized words of tweets will be grouped in arrays with an ID number representing
the number of appearances of the word present in the whole document. Each array
is termed a cluster [8].

Trigrams
An n-gram is collected from a text corpus, a contiguous set of n items from a given
sample of text. Trigrams are a type of n-grams with n = 3. N-grams are used in
text analysis for grouping up different letters of a word. For example, for the string
“abcdefg” trigrams (n = 3) will be efg, def, cde, and so on. These language models
are used in Natural Language Processing Models to reduce the algorithm’s overall
time complexity. In SentiFusion, trigrams are used to group different elements of the
arrays obtained from the previous step. Each array element will be converted into
chunks of 3 and assigned a weight as per the array’s ID number. Then all the chunks
with their weights will be inserted into a stack. This step is done so that there will be
the least number of comparisons in the next step.
818 S. Anand et al.

K-means Clustering with Silhouette Analysis Technique

The stacks obtained from the previous step are converted into clusters using K-means
Clustering. With the assistance of SentiWordNet, a score is assigned to each cluster
by its aggregated scored tweets within the next steps. A score of every term within
the tweet will be reclaimed with its tag (i.e., noun, adjective, verb, and adverb). The
results of K-means clustering are validated and interpreted using Silhouette analysis.
This analysis gives a mean score of the clusters. If the typical score is 1, then we will
say that the info point is well assigned to the primary cluster, whereas a score on the
brink of −1 means data is aligned to the second cluster.

Kullback–Leibler Divergence
Kullback-Leibler divergence (relative entropy), in short, called KL divergence,
measures the difference between any two probability distributions. Given two discrete
probability distributions, p(x) and q(x), D(p||q), the relative entropy between p and
q is defined in Eq. (1).

D( p||q) = p(x) ∗ log( p(x)/q(x)) (1)

From Eq. (1) it can be conclude that, D(p||q) is always non-negative and if p = q
then D(p||q) = 0. It shows the variation of p(x) with respect to q(x). If KL divergence
is small, there is a high similarity between the distributions of two variables, and
the converse of it is also True. This probability distribution is used to assign more
precise values to the classified clusters obtained from the previous step. Every token
of the clusters is now assigned a decimal value (KL—scores) of their related entropy,
which gives a more precise analysis of Sentiments.

TF-IDF
We implement Term frequency-inverse document frequency, a numerical statistic
for keyword extraction. TF-IDF is related to a bag of words (Bagging). The main
idea behind using TF-IDF as a numeric feature generator is that—more importance
should be given to the word that appears more in one text document and less in
another, which eventually becomes useful for classification. In SentiFusion, this
model is used for emphasizing the tokens, which are frequently repeated. The tokens
in SentiFusion are terms, and similarly, each cluster is considered as a document
here. For every token and cluster, a tf value and idf value are calculated, respectively.
When these values are multiplied together, we get a score that is highest for tokens
that frequently appear in the clusters and low for tokens that occur less often in every
cluster, allowing us to find tokens necessary in the cluster. Hence, if the token is
very common, then the value will approach 0; else, it nears 1. Now, the “tf-idf score”
is assigned to all the tokens, and a distinction is made between normal tokens and
keywords for all the clusters. This tf-idf score is multiplied with the KL-scores from
the previous step. This clustered data is now fed to machine learning algorithms for
further feature fusion and classification.
SentiFusion (SF): Sentiment Analysis of Twitter … 819

3.3 Feature Fusion

This step is for feature fusion by combining the scores of all the tokens and keywords
obtained in the previous step of Feature Extraction. The sentiment scores are multi-
plied with the scores obtained from the earlier steps. A final decimal sentiment score
is obtained by the end of this step. The keywords will naturally have higher scores
as they already have a higher combined tf-idf and KL scores. Let these scores of all
the tokens be called “Senti-scores” for future references. All the clusters are fed to
the next step of feature weighting and bagging to summarize the results of sentiment
analysis of the queried topic.

3.4 Feature Weighting Using Shannon Entropy

Shannon and Weaver proposed this feature weighting method in 1963. It acts as a
criterion for measuring the degree of uncertainty represented by a discrete probability
distribution. Entropy here measures existent contrasts between different sentiment
datasets. Let p(x) or pij be the probability mass function of the random variable x or
xij .
Steps for feature weighting through Shannon Entropy are:
1. Normalizing the arrays of performance indices to obtain the outcome pij based
on Eq. (2)
xi j
p(x) = pi j = m (2)
i=1 x i j

2. For Computing entropy measure, Eq. (3) is used

m
E j = −k ∗ pi j ln( pi j ) (3)
i=1

Where k = ln(m)
1

3. Eq. (4) defines the weight based on Shannon Entropy

(1 − E j )
w j = n (4)
j=1 (1 − E j )

This feature weighting method is used to compute the entropy of a probability

distribution of the appearance of different clusters in different sentiments. Applying
the concepts of Shannon Entropy, we will be using the following formula:
N
entr opy(g) = H ( p(S|g)) = − p(Si |g)logp(Si |g) (5)
i=1
820 S. Anand et al.

where N is the number of sentiments (in SentiFusion, N = 3 as trigrams were

used in feature extraction).
A higher value of a cluster’s entropy in different sentiment datasets indicates that
a distribution of the appearance of that cluster is close to uniform. Therefore, such
clusters would not contribute much to the classification. On the other hand, a low
value of the entropy will indicate that the cluster appears in some of the datasets more
frequently than in other datasets. So it highlights a proper sentiment (or objectivity).
Hence, to increase the overall accuracy the clusters with low entropy values are only
passed on to the next step. This is achieved by setting a threshold value and then
filtering out all clusters with entropy. As the number of features used is reduced, this
step lowers the recall. Nevertheless still the primary attention is on high accuracy, as
the dataset used is enormous.

3.5 Bagging

Support Vector Machines or SVMs are supervised learning models. For the clas-
sification of data and then performing Regression analysis as per the need, SVMs
analyze the data and recognize the patterns. In this step, using the SVM classifiers,
we are creating a supervised learning model of three classes—strong sentiments,
neutral sentiments, and Weak sentiments. This classification is done for each token
based on Senti-sores (product of tf idf and KL divergence scores). This model is now
trained to accept and classify any new tokens based on Senti-scores. For each of the
three classes, the tokens are placed and interlinked in space in the form of Decision
trees.

4 Sentiment Computation Using SentiWordNet

SentiWordNet [9], derived from the WordNet database, is a lexical resource used
for opinion mining. It assigns to each synset (a word or a term which are inter-
changeable synonyms) three sentiment scores: positivity (+1), negativity (−1), and
objectivity(0). In this final step after feature selection, a ranking of the selected tokens
is done using the SentiWordNet dictionary. The selected tokens include both positive
and negative rankings. These score values of the products are inputted to SVC and
Decision Tree classifiers in the three classes. The polarity of the word signifies the
characteristic of a given tweet.
SentiFusion (SF): Sentiment Analysis of Twitter … 821

5 Implementation and Performance Evaluation

The whole program was run on a system with an i7 Processor/16 GB RAM. It was
connected with the DCX1 server of supercomputing to supercomputers. Stanford’s
Python Natural Language Toolkit has been used for carrying out NLP as well as Pre-
processing Tasks. Pre-processing of blank space and special character tokenizers
were customized. Similarly, a customized algorithm was used for stemming or
lemmatization and stop-word removals.
The performance of the proposed SentiFusion has been computed using Precision,
Recall, Accuracy, and F-Score as the desired matrix. In Table 1, a comparison between
CSNM paper and SentiFusion paper is carried out. The standard formulae of the
respective terms are used to calculate the performance measure.
Table1 depicts a comparison between the existing method-CNSM and the
proposed method—SentiFusion. In total, 2500 data are compared, of which the 500
datasets of CNSM yield 86.87% of Precision, 88.02% Recall, 87.445% Accuracy,
and an F-score of nearly 87.44%. In contrast, we can see a Precision of 94.14%,
97.18% of Recall, 95.66%, and 95.64% of Accuracy and F-Score, respectively in
SentiFusion. We can observe a considerable difference rate in Accuracy and F-Score
of both the models. With the more reliable F-Score, we can conclude that the proposed
system has higher performance measures.
The SentiFusion method performed better than the CSNM method. This happened
mainly because of the Bi-way verification system in SentiFusion. In the CNSM
method, the author employs n-gram similarity matching and Cosine similarity using
Latent Semantic Analysis methods. This made the model a bit slower in terms of
Time complexity. Topic modeling is used with ranking for feature selection. There
is no robust classifier included in the existing system of CNSM. The proposed
SentiFusion has a concrete Bi-Classifier and Ensemble Classifier with SVM that
harvests boundary features, granular and sub-granular features, unlike LSA. Whereas
in feature extraction of the SentiFusion model, two different methods are used, which

Table 1 Comparison of CNSM and SentiFusion

No. of Performance measures
Data
Points
Precision % Recall % Accuracy % F-Score in %
CNSM SF CNSM SF CNSM SF CNSM SF
500 86.87 94.14 88.02 97.18 87.44 95.66 87.44 95.63
1000 85.54 93.89 87.45 96.85 86.49 95.37 86.48 95.35
1500 84.44 93.07 87.01 95.18 85.72 94.12 85.7 94.11
2000 83.17 92.74 85.45 94.85 84.31 93.79 84.29 93.78
2500 82.21 91.45 84.39 93.17 83.3 92.31 83.28 92.3
822 S. Anand et al.

give a higher accuracy rate. The time complexity of the SentiFusion model is also
less as the comparisons are less because of Shannon Entropy.

6 Conclusions

A SentiFusion model was trained which can now be used to analyze sentiments
for any topic properly. The model was trained using two different methods. The
first method consisted of Jaccard Similarity and Trigrams while the second method
had K-means Clustering and KL divergence implemented in it. Then feature fusion
and weighting were facilitated by tf-idf and Shannon Entropy. Bagging in the last
step was done using SVM Classifiers. This model was found to give an accuracy
of 92% in extensive datasets whereas 95% accuracy was easily obtained for smaller
datasets. The average F-score of our model is 94.23 which is higher than the others.
The proposed model has 95% F-score for smaller datasets and 92% for extensive
datasets. This model also has very high precision value. Hence, it is concluded that a
two-way Feature Extraction combining two different methods provides better results.

References

1. Shirbhate AG, Deshmuk SN (2013) Feature extraction for sentiment classification on twitter
data. Int J Sci Res (IJSR), pp 2184–2185
2. Mathapati S, Anil D, Tanuja R, Manjula SH, Venugopal KR (2018) CNSM: cosine and n-gram
similarity measure to extract reasons for sentiment variation on Twitter. Int J Comput Eng
Technol (IJCET) 9:150–161
3. Agarwal A, Xie B, Vovsha I, Rambow O, Passonneau R (2011) Sentiment analysis of Twitter
data. In: Proceedings of the workshop on languages in social media, LSm 2011, pp 30–38
4. Öztürk N, Ayvaz S (2018) Sentiment analysis on Twitter: a text mining approach to the Syrian
refugee crisis. Telemat. Inform. 35(1), 136–147
5. Jianqiang Z, Xiaolin G (2017) Comparison research on text pre-processing methods on Twitter
sentiment analysis. IEEE Access 5:2870–2879
6. Wang L, Niu J, Yu S (2020) SentiDiff: combining textual information and sentiment diffusion
patterns for Twitter sentiment analysis. IEEE Trans Knowl Data Eng 32(10):2026–2039
7. Ramanathan V, Meyyappan T (2019) Twitter text mining for sentiment analysis on people’s
feedback about Oman tourism. In: 2019 4th MEC international conference on big data and
smart city (ICBDSC), Muscat, Oman, pp 1–5
8. https://python.gotrained.com/nltk-edit-distance-jaccard-distance/#Example_1_Character_
Level
9. https://github.com/aesuli/SentiWordNet
10. Deepak G, Santhanavijayan A (2020) OntoBestFit: a best-fit occurrence estimation strategy
for RDF driven faceted semantic search. Comput Commun 160:284–298 (2020)
11. Pushpa CN, Deepak G, Kumar A, Thriveni J, Venugopal KR (2020) OntoDisco: improving web
service discovery by hybridization of ontology focused concept clustering and interface seman-
tics. In 2020 IEEE international conference on electronics, computing and communication
technologies, pp 1–5, July 2020
SentiFusion (SF): Sentiment Analysis of Twitter … 823

12. Deepak G, Kasaraneni D (2019) OntoCommerce: an ontology focused semantic framework

for personalised product recommendation for user targeted e-commerce. Int J Comput Aided
Eng Technol 11(4–5):449–466
13. Gulzar Z, Anny Leema A, Deepak G (2018) Pcrs: personalized course recommender system
based on hybrid approach. Procedia Comput Sci 125:518–524
14. Deepak G, Teja V, Santhanavijayan A (2020) A novel firefly driven scheme for resume parsing
and matching based on entity linking paradigm. J Discrete Math Sci Cryptography 23(1):157–
165
15. Haribabu S, Sai Kumar PS, Padhy S, Deepak G, Santhanavijayan A, Kumar DN (2019) A
novel approach for ontology focused inter- domain personalized search based on semantic set
expansion. In: 2019 fifteenth international conference on information processing
16. Deepak G, Kumar N, VSN Sai Yashaswea Bharadwaj G, Santhanavijayan A (2019) OntoQuest:
an ontological strategy for automatic question generation for e-assessment using static and
dynamic knowledge. In: 2019 fifteenth international conference on information processing
(ICINPRO), pp 1–6 (2019)
17. Santhanavijayan A, Kumar DN, Deepak G (2020) A semantic-aware strategy for automatic
speech recognition incorporating deep learning models. In: Intelligent system design. Springer,
pp 247–254
18. Deepak G, et al (2019) Design and evaluation of conceptual ontologies for electrochemistry as
a domain. In: 2019 IEEE international WIE conference on electrical and computer engineering
19. Deepak G, Priyadarshini JS (2018) Personalized and Enhanced Hybridized Semantic Algorithm
for web image retrieval incorporating ontology classification, strategic query expansion, and
content-based analysis. Comput Electr Eng 72:14–25
20. Kaushik IS, Deepak G, Santhanavijayan A (2020) QuantQueryEXP: a novel strategic approach
for query expansion based on quantum computing principles. J Discr Math Sci Cryptography
23(2):573–584
21. Deepak G, Kumar DN, Santhanavijayan A (2020) A semantic approach for entity linking
by diverse knowledge integration incorporating role-based chunking. Procedia Comput Sci
167:737–746
Using Internet of Things to Increase
Efficient Collaboration in PLM

Narjiss Tilioua, Fatima Bennouna, and Zakaria Chalh

Abstract Nowadays, the automotive sector needs more and more digitalized solu-
tions, and this is especially with regard to PLM software and its effectiveness in
communicating information related to new products under development, even if this
is the main role of PLM but to increase its effectiveness and efficiency of collaboration
between automotive companies, their subcontractors and suppliers. Because effective
collaboration leads directly to the rapid success of product lifecycle management.
In this paper we will show the need to better digitalize the communication between
the different actors of the PLM wheel and we will also mention three principles that
will help us to better collaborate in the context of product lifecycle management.

Keywords PLM · Supplier · References · Industry 4.0 · Collaboration · IOT

Things · Automotive

1 Introduction

In a world where the only constant thing is change, the platform Product Lifecycle
Management PLM has been created and designed over the past two decades to
manage the entire product lifecycle efficiently and cost-effectively from ideation,
design, manufacture and then service to disposal. In fact, the huge need of an inno-
vative solution as PLM systems comes from the rapid evolution of the production
worldwide. One reason that pushes organization to adapt their strategies to the context
and the new challenges such as the international concurrence, the segmented market,
the big flow of information and data, the intensive resources difficult to manage and
control, etc.

N. Tilioua (B) · F. Bennouna · Z. Chalh

LISA Laboratory, Sidi Mohamed Ben Abdellah University, Fes, Morocco
e-mail: [email protected]
Z. Chalh
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 825
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_75
826 N. Tilioua et al.

As a solution to all of this, PLM systems consist of integrating products, processes

and people into a single system, enabling the co-engineering and the collaboration
between internals, suppliers and customers in order to enable enterprises to drive
innovation, increase productivity, reduce costs, improve quality, accelerate growth,
ensure compliance and most importantly shorten the time to market.
However, putting all the actors into the single system is not sufficient to get to
the optimal results. As a matter of example, we know that PLM systems contain
all the product references and their presentation of products in 3D format. So when
engineers or technicians in the supply chain management desire to order products,
they search for the latest updates of the products with their respective references
on the PLM platform. But, the information flow of product datais often interrupted
during the various stages of product testing at OEMs. As a consequence, these stages
are not yet processed in some companies by PLM systems at the time of search and
order.
The rest of this document is structured as follows:
Section 1: assesses the status and challenges of PLM systems and strategy imple-
mentations according to three main principles. In Sect. 2, we present the discussion
on the artificial innovation to be added to increase the performance of automotive
companies while improving the use of PLM.

2 Literature Review

This paper is inspired by several scientific publications, a literature review of several

sectors that use PLM, and our objective is to compare the solutions already commu-
nicated by other authors and to propose the best solution that will be based on digital
technology to improve lifecycle management efficiency in the automotive sector.
Knowing that the life cycle of automotive products is composed of several interre-
lated stages, roughly speaking we can mention three stages: Beginning of Life (BOL),
Middle of Life (MOL) and End of Life (EOL). And to link the data from these three
phases, the software and PLM solution has been created as a single information
space that allows all actors in the value chain of the product under development to be
involved in the research and development phase integrated in the first stage (BOL)
discussed in this document. Our document therefore goes in the direction of opti-
mizing the procurement time of manufactured products that are in the testing phase
with the new references launched by suppliers.
At the research and development stage, digital solutions can be implemented in the
form of communication and information exchange on technical changes in products
that is more interactive and fast, an example: connecting the phones or laptops of the
players in this value chain to PLM software in order to alert them about new products
that interest them in real time.
Using Internet of Things … 827

It is thanks to the rapid and advanced evolution of information technology that the
automotive sector has become more competitive in recent years. Internet technology
has contributed to better linking and virtually connecting processes even with distant
actors and to the development of new strategies for the development of products
that meet customer requirements and aim to better satisfy them. And as Mr. Sharma
lied in his article, the real challenge of technology lies in its ability to link people,
processes and information [1].
The current state of PLM software is based mainly on the BOL and MOL phases,
but according to existing research, the future vision of PLM will also give importance
to the EOL phase.
The literature sets up several challenges and opportunities of PLM, including
data and information management issues [2] and requirements management [3].
The definition of PLM presented by Corallo [4] groups together the information
and collaboration management of products throughout their life cycle as well as the
inherent aspects of managing. However, it should be noted that providing and sharing
information will enable effective PLM processes, both within an organization and
with partners and stakeholders across the enterprise.
The evolution of PLM also addresses the increasing complexity of companies,
which is mainly due to the following factors increasing organizational growth
(collaboration, decentralization) increasingly complex products [5].
Product lifecycle management systems play an important role in the integration
and process-based organization of information about a series of development and
evolution phases of products, processes and resources.
New information technologies, such as Digital Twin, are encouraging industrial
digital transformation and effective lifecycle management that must now take into
account all phases of the project, where The use of the 3D model in some of them is
no longer the best solution [6].
Digital Twin (DT) has a great capacity to realize Industry 4.0. to simulate working
conditions in real time and make intelligent decisions, when a cost-effective solution
can be easily provided. Knowing that the term PLM allows an application of the
Digital Twin as companies are moving towards green manufacturing with the Digital
Twin which enablesmaximizing production quality throughout a complete loop cycle
[7]. Thanks to the simulations of the Digital Twin tool and the opportunity it gives to
the actors of the value chain to make decisions relevant to the topics concerning the
life cycle of products, based on this opportunity, they canmake intelligent products
with self-awareness [8].
828 N. Tilioua et al.

3 Internet of Things in PLM Tool Towards Efficiency

in Management Data Flows

The Automotive products lifecycle begins with the Research and Development stage.
At the first stage, it is necessary for the vehicle company’s engineers and technicians
to search for the composition of the products,and we know that suppliers develop
new references of these products after any EOM negative return or any innovation of
that product, so the idea is to link and facilitate the references searching to OEM engi-
neers and technicians commands directly with suppliers by using PLM tool connected
with their personal computers, as well as the future manufacturing technology. At
this stage, applications are filed within patent offices to protect the intellectual prop-
erty created during the drug development. It will also require the organization of
pilot production and procurement since at the development stages, sample batches
of automotive products are always required in those tests. It is also necessary to
conduct a search and optimal supplier name selection for the product being developed
with the subsequent necessary measures implementation for registering a command.
Procurement stage is followed by banc tests of parts required.
Once the idea research stage of a new part project is completed, a feedback is made
to OEMs and suppliers. At this stage of the product life cycle, a future study protocol
is created, and after its approval, research and development studies are conducted
in accordance with the requirements of good practice. Component data of product
samples from the research and development stage.
This stage is followed by the registration of the product in the PLM, during which
the actors are responsible for their modifications. After the evaluation, the engineers
issue a marketing authorization for the automotive product. The process of collecting
the documents in the file is complicated and requires specialists with a high technical
level. At this stage, it is necessary to ensure that all the data generated by the previous
steps are integrated.
The production phase normally takes place after registration, but its prepara-
tion takes a considerable amount of time. The organization of the production is
usually done in parallel with the previous steps. The production of automotive parts
is controlled by testing, and the main component of this system is quality assurance
at all stages of production, from the purchase of raw materials and materials and their
control through to the final product.
To remain competitive in today’s digital age, knowledge sharing and knowledge
reuse between different companies in internal or external collaboration needs to be
strengthened, and the knowledge needs of the different phases of the product lifecycle
need to be identified and detailed in a clear way.
We will compare three articles in this sense, the first article chose to deal with the
two BOL phases of the R&D companies because this phase is well known as the most
important stage on which the entire product life cycle is based, and the MOL phase
with logistics companies and it was noticed that the staff in both phases insist on the
need for more advanced integration of the suppliers this is shown in the article by the
following sentence: The R&D staff focused on new materials and innovations from
Using Internet of Things … 829

suppliers to apply them more quickly to their own product development (M2), while
the logistics staff focused on the transport capacity of suppliers because sometimes
their own fleet cannot meet the customers’ requirements.
And for the same article the author focused his vision on expertise as a type of
knowledge to be shared to better streamline and help manage product lifecycles,
saying that R&D expertise was mainly related to design, development, technology
and manufacturing process. For the logistics people, it was mainly related to import
and export, and export, insurance, and political and legal aspects. This by insisting
in the side of logistics companies that the expertise acquired from competitors is
more useful because it is more relevant to them unlike R&D companies which have
confidentiality in their expertise compared to their competitors. And the latter can
only obtain knowledge through other channels, especially through person-to-person
communication. However, logistics personnel expect the government to organize
special meetings to obtain faster updates and among the things that show the need
for supplier integration in logistics are hygiene, safety and temperature in transport,
limiting the weight and length of aircraft.
And an example was given in this article that knowledge sharing is done in a hier-
archical way from the internal base of the company to the external side by specifying
reference persons and insisting that Knowledge must be shared within the depart-
ment, within the company, with the branch, with the supplier, with the customer and
even with the competitor.
In the other article called: A Green PLM approach, which focuses on product
lifecycle management from an environmental perspective, his contribution suggests
a framework for sustainable product development that takes into account the entire
product lifecycle and will enable companies to be more resource efficient.
For him the Factory of the Future is a set of competent regional small and medium
sized companies that could compete but, at the same time, be available to collaborate
when needed by sharing knowledge and resources. with customers, stakeholders and
suppliers and thus enable sustainability through green products and processes (Green
Factory certified).
The author in this article emphasizes a special attention to green manufacturing
which must be considered critical as it is the most polluting stage of the product
life cycle due to greenhouse gas emissions and liquid and solid waste that are not
sufficiently controlled.
In the 3rd article: Convergence of IoT and product lifecycle management in
medical health care, the author highlights the technological and artificial side to
better develop the role of PLM while saying that with the modern cloud and IOT
Intelligent connected devices improve collaboration and flexibility throughout the
value chain, Greg Cline [9] presents that things already generate more data than
people or applications.
The main contribution of this paper is to raise the importance of IoMT integra-
tion with Product Lifecycle Management (PLM), information and data sharing, and
collaboration from medical experts to patients and between devices.
830 N. Tilioua et al.

The author thus insists on the link between the Internet of Things (IoT) and PLM
in the medical market because with Information Communication Technologies (ICT)
and sensors that are driving a dynamic change in industry and academia. As research
on PLM and IoT has been conducted for a long time. But these feature-based devices
consume more energy and a majority of small nodes are battery-powered, which is
a critical and interesting challenge.
The e-Cloud with IoT-enabled devices is being developed; in which devices can
communicate with each other while sharing and exchanging very important informa-
tion and the power and life cycle of these devices will be managed from the beginning
to the end of the medical treatment process thus keeping in mind the ability to gather
knowledge and information.
Finally, I think it is important to combine these three ideas of IOT and lean
manufacturing in green manufacturing by ensuring supplier integration and effective
collaboration between all the collaborators in the value chain with the control of
environmental experts and industrial managers through the use of intelligent sensors
and connecting them directly to the devices of all these people each on their own
side of the collaboration.
Giving the importance to collaboration in the purpose of having the effective
management of products throughout their life cycle. The literature on the efficiency
of integration and communication between suppliers and stakeholders and OEMs its
objective in this article is facilitating the selection of the latest references that can be
adapted to the context of product life cycle management. In Table 1, we summarize
their relevance and importance for decision support in the context of LCM.
And for the problem of efficient supplier integration and communication between
test engineers on new vehicle projects and suppliers of the latest 3D part modification
or creation, I propose to integrate one more step in the PLM tool process which is the
effective communication to order parts directly without the need to switch to another
system to order them, and these steps are outlined as follows (Fig. 1):
As a result of this proposition there will be an efficiency in rapid exchange of
references flows between suppliers and OEMs, and also in logistics. to approve that
we.
The hyper-connected world of products, human actors, and operating environ-
ments, enabled by IoT technologies [13], creates a potential for explosive growth
in the generation of product—related data streams. It is not sufficient anymore to
seek to integrate such information centrally. Instead, part of the information inte-
gration is best performed at the point of data consumption. For such integration to
produce meaningful results, the complexity of relevant product information needs
to be managed. Directly relevant to establishing efficient architectures, indexing,
and big data management capabilities for IoT—generated data is the key concept of
context, adopted in context aware computing [13].
Table 1 Collaboration support issues and principles
Issues Example Principles
Current PLM systems must enable collaboration, C.Vila, An approach to Sustainable Product See Principle 1 below Wherever possible,relevant
Using Internet of Things …

mandatory consultations, and capture inputs, Lifecycle Management (Green PLM), Procedia communications should be captured at an
contributions, and changes in all system Engineering 132 (2015) 585–592 [10] appropriate level of detail;
interactions it can be used to enable better knowledge retrieval.
Collaboration 4.0 concepts can provide a context
for knowledge reuse throughout the product
lifecycle, and time and money optimization
A number of participants can contribute to PLM YanXin, Dealing with Knowledge Management See Principle 2 below
decisions—groups, deparments, and otherrelevant Practices in Different Product Lifecycle Phases In the stages of an automotive part Development
stakeholders within the organizational structures within Product-service Systems, Volume 83, 2019, process, involve the relevant participants in these
and hierarchies of different Pages 111–117 [11] stages and give each participant his or her role and
the opportunity to send messages and order emails
directly to the suppliers of these parts. Assign a
participant as a key decision-maker
Problems using PLM manifest themselves in ALIHASSANSODHRO, Convergence of IoT and See Principle 3 below To improve the integration of
many forms different product lifecycle management in medical actors and minimize training expenses, and also to
healthcare, 2018 [12] ensure the integration of suppliers who are a
different organization by adding a training page in
the PLM tool according to the user’s role
831
832 N. Tilioua et al.

Fig. 1 Proposition of collaboration support solution

4 Conclusion

This work contributes to provide the collaboration in PLM. Time optimization and all
actors’ integration in automotive company’s value chain are essential conditions to
have an efficient product management. The paper highlights how supporting concepts
in-group collaboration and integration facilities can help to meet the requirements of
choice and reuse of PLM decisions and decision-making processes, thus contributing
to organizational knowledge accumulation and organizational learning. The princi-
ples presented to support decision-making in a PLM context provide a foundation for
effective PLM decision support with significant potential for integration with existing
PLM systems and for exploiting emerging and future PLM ontologies, which may
have broad implications for organizations’ PLM processes and protocols in thefuture.

References

1. Sharma (2005) Collaborative product innovation: integrating elements of CPI via PLM
framework 37(13), 1425–1434
2. Marra M (2012) Supply chain knowledge management: a literature review 39(5): 6103–6110
3. Silventoinen A (2014) Challenges of information reuse in customer-oriented engineering
networks 34(6):720–73
Using Internet of Things … 833

4. Corallo A (2013) Defining product lifecycle management: a journey across features. Def
Concepts, Volume 2013 |Article ID 170812
5. Meier U (2017) Twenty Years of PLM–the good, the bad and the ugly. Product Lifecycle
Manage Ind. Fut, pp 69–77
6. Tchana Y (2019) Designing a unique Digital Twin for linear infrastructures lifecycle
management. Procedia CIRP84, pp 545–549
7. Papinniemi J, Hannola L, Maletz M (2014) Challenges in integrating requirements management
with PLM. Int J Prod Res 52(15):4412–4423
8. Mabkhot M (2018) Requirements of the smart factory system: a survey and perspective
9. Cline G (2017) Industry 4.0 and industrial IoT in manufacturing: a sneak peek
10. Vila C (2015) An approach to sustainable product lifecycle management (Green PLM). Procedia
Eng 132:585–592
11. Xin Y (2019) Dealing with knowledge management practices in different product lifecycle
phases within product-service systems 83:111–117
12. Alihassansodhro. Convergence of IoT and product lifecycle management in medical healthcare
(2018)
13. Koshechkin K. Implementation of digital technologies in pharmaceutical products lifecycle
14. Leal AG, Santiago A (2014) Integrated environment for testing IoT and RFID technologies
applied on intelligent transportation system in Brazilian scenarios
ECC Image Encryption Using Matlab
Simulink Blockset

Sara Chillali and Lahcen Oughdir

Abstract To implement hardware operations on various Xilinx FPGAs, we provide

a method of software simulation, providing models (set of Simulink blocks) and
presenting the concept of software simulation using Matleb Simulink for processing
and encryption of images.
This work presents an efficient architecture of various image algorithms, en-
cryption, decryption and Keys generate using an elliptic curve, for all kinds of color
and grayscale images with a minimum number of generator blocks possible. Example
and illustration of these architecture implemented were presented.

Keywords Matlab Simulink · Image processing · ECC · Encryption · Decryption

1 Introduction

Cryptography is a science that ensures the security of sharing and storage of data by
individuals in the presence of adversaries.
This is done through techniques that ensure the confidentiality, authenticity and
integrity of the data.
As a science of secrecy, cryptology is divided into two; cryptography which aims
to build and prove Cryptosystems and cryptanalysis whose goal is to "break" the
systems built by cryptographers.
For a long time and in all previous eras, secret codes were the basis of cryp-
tography, until 1970, the development of an encryption and signature system was
an essential objective of cryptography [3, 9, 10], therefore cryptanalysts wage
their wars in the shadows by controlling communication networks and information
transformation.

S. Chillali (B)
Sidi Mohamed Ben Abdellah University, FP, LSI, Taza, Morocco
e-mail: [email protected]
L. Oughdir
Sidi Mohamed Ben Abdellah University, ENSAF, Fez, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 835
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_76
836 S. Chillali and L. Oughdir

Fig. 1 Encryption and decryption protocol

The emergence of digital information sent through insecure channels like the
Internet has facilitated communication between different parties, in fact, there are
security holes in networks, which facilitates access to information.
Mathematics with difficult inverse problems and a support of computer science
and applied physics gave a power to cryptography which became a science attached
to mathematics and computer science; it is a science of encrypting messages, other
than cryptanalysis, which is used to decrypt them. [7, 8].
Our contribution in this work is the encrypted of different types of images of
different sizes, using a mathematical object called elliptic curve, or we can give
several definitions, studied in algebraic geometry, [4].
For more than forty years Diffie-Hellman, [1] gave the appearance of the key
exchange protocol which was developed at the end of the 1980s: ECC, Elliptic Curve
Cryptography. These cryptographic protocols use group structures whose security is
ensured by the difficulty of the discrete logarithm problem on these groups, (DLP),
which shows the success of these elliptic curves in public key cryptographic systems,
hence the interest of the arithmetic study of these geometric curves.
Image processing is certainly the most innovative that man has known, [5, 6] moti-
vated by his needs in various fields, including imaging and the following, generates
sensitive problems solved by the various image processing techniques.
The digital image is represented by a matrix of points called pixels (Picture
Element), each having as characteristic a gray level or a color coming from the
corresponding location in the real image or calculated from an internal description
of the scene to represent [2].
We can represent an image by a matrix M(mi,j ); mi,j ∈ {0, 1, 2, . . . , 255}.
ECC Image Encryption Using Matlab Simulink Blockset 837

In [11], the authors gave a new image encryption algorithm based on a secure
variant of Hill Cipher (HC) and three enhanced one dimensional chaotic maps (1D).
Our approach is to propose an encryption scheme based on a safer concept, namely
discrete logarithm problem on elliptic curves. Thus, the simulation results carried
out on a database of images in colors and in grayscale give very effective security
results, this is the subject of an article already submitted.
In the following, thanks to the discrete logarithm problem of this elliptic curve,
a new method of image encryption will be constructed, using the Diffie-Hellman
protocol on such a curve, the security of this type of encryption is proven because
the resolution of this problem is difficult on these curves (Fig. 1).

2 Elliptic Curve on F P

The definition of an elliptical curve is formal and uses mathematics, supported by

algebra and geometry to define the group law on such a curve. In this part we will
define an elliptic curve on the finite field FP , p ≥ 5 is a prime number [4].
In the following, we can define an elliptic curve on the finite field FP by the
equation of the form:

y 2 = x 3 + ax + b, (1)

where (a, b) ∈ F2p and  = −16(4a3 + 27b2 ) is invertible in FP .

Group law:
We denote this elliptic curve by: Ea,b (p), we can write:
 
E a,b ( p) = (x, y) ∈ F2p /y 2 = x 3 + ax + b ∪ {[0 : 1 : 0]} (2)

We define the law + on Ea,b (p) by:

Let P(x1 , y1 ) and Q(x2 , y2 ) be two points on the curve, then P + Q = R(x3 , y3 )
such that:
• R = [0 : 1 : 0], for x1 = x2 ,y2 = −y1 ; (P is the opposite of Q, even for Q)
• R(x3 , y3 ); x3 = t 2 − x1 − x2 and y3 = −tx3 − s, with:
 y −y  x y −x y
2 1
x2 −x1
, i f P = Q 2 1 1 2
x2 −x1
, i f P = Q
t = 3x12 +a and s = ax1 −x13
2y1
,if P = Q 2y1
,if P = Q
A series of k consecutive additions of point P is called multiple point of order k
is defined as follow:

kP = P
 +P+
· · · + P (3)
k times
838 S. Chillali and L. Oughdir

Fig. 2 Affine points over y 2 = x 3 − x on F89

Ea,b (p), + is an abelian group, [0:1:0] is the point at infinity it is the neutral
element for the + law.
Let Q also be an element of a cyclic subgroup G =< P >. An integer k that
solves the equation kP = Q is called a discrete logarithm of Q in the base P. One
writes k = logP Q.
Let P and Q, find x satisfying xP = Q, is called discret logaritm problem (DLP)
(Fig. 2).

3 Construction of a Secret Key

Let p be a large prime number, Ea,b (p) is an elliptic curve on the field FP defined by
equation (1), whose discrete logarithm problem is difficult in this curve.

3.1 Diffie-Hellman Key Exchange

Aicha and Omar publish a point P which generates a subgroup of known order r of
an elliptic curve Ea,b (p).
• Aicha chooses a private number t ∈ {1, 2, . . . , r − 1} and calculates tP.
• Omar chooses a private number s ∈ {1, 2, . . . , r − 1} and calculates sP.
• Aicha sends tP to Omar.
ECC Image Encryption Using Matlab Simulink Blockset 839

• Omar sends sP to Aicha.

At the end, Aicha and Omar build their shared secret key: K = tsP = stP.

3.2 Secret Encryption Matrix

Each for him, the two entities build the secret encryption matrix T by the following
two steps:
• Step1:
Build a matrix A = (ai,j )i=1,...,f;j=1,2,3 .
For 1 ≤ i ≤ f , if i K = (xi , yi ), then ai,1 = i, ai,2 = xi and ai,3 = yi , i.e.:
⎛ ⎞
1 x1 y1
⎜2 y2 ⎟
⎜ x2 ⎟
⎜ ⎟
⎜: : : ⎟
A=⎜ ⎟
⎜i xi yi ⎟
⎜ ⎟
⎝: : : ⎠
f xf yf

• Step2:
Build a Key matrix T that is the same size as the image.
One of the algorithms at this work is to find the key matrix T from matrix A.
Let [n, m, p] the size of the image to be encrypted, then T (Fig. 3).
• Flowchart Algorithm:

Fig. 3 Flowchart algorithm

840 S. Chillali and L. Oughdir

4 Encryption Algorithm

We represent our software simulation approach to show the multiple image

encryption, see Figs. 4, 5, 6 and 7.
Figure 8, we use the available Simulink generator block set.
We explore the important aspects related to the hardware implementation with a
MATLAB Simulink combination, we obtain an architecture that offers an alternative
through a graphical user interface.
Aicha wants to send a confidential image “image” to Omar, she will construct
Embedded functions and follows the structure of the algorithm described by the
following steps:
• Setp1:
Aicha transforms the image, ‘image’ to the matrix M(mi,j ); mi,j ∈
{0, 1, 2, . . . , 255}.
• Step2:
Aicha transforms M to matrix Mt and T to matrix Tm using two Embedded
functions:

This function is composed of an entry which represents the original image to be

encrypted and three exits which represent the three make up of the matrix Mt.

• This function is composed of an entry which represents the original image to be

encrypted, another entry which represents the ECC key A and three exits which
represent the three make up the matrix Tm.
• The sizes of the dies are:size(Mt) = size (Tm) = [n ∗ m ∗ 8, 1, 3].
• Setp3:
Aicha calculates the column vector, R = mod(Tm + Mt, 2).
• Setp4:

Aicha transforms R into a matrix Rt and then encrypt the image “encryptimage”
by the following encryption system:

Fig. 4 Flowchart Embedded

function1
ECC Image Encryption Using Matlab Simulink Blockset 841

Fig. 5 Flowchart Embedded

function2

Fig. 6 Flowchart encryption system

Encrypt System is composed of 7 inputs which represent the image, which we

want to encrypt, the components of the matrix Mt, the components of the matrix Tm
and two outputs which represent the image encrypt and its transform into matrix.

5 Decryption Algorithm

Omar receive, "encryptimage" sent by Aicha, he has the private key which allows him
to calculate the decryption function and to find the matrix M, which can transform
it into the image, “image”, see Fig. 9.
We can summarize the encryption and decryption method by a flowchart see
Fig. 10.
842 S. Chillali and L. Oughdir

Fig. 7 Component of encryption system

Fig. 8 Block diagram approach for dual image encryption

6 Example and Illustration

In this section, we assume that; p = 2543, a = 1758 and b = 254.

E 1758,254 (2543) : y 2 = x 3 + 1758x + 254

So;
 
E 1758,254 (2543) = (x, y) ∈ F22543 /y 2 = x 3 + 1758x + 254 ∪ {[0 : 1 : 0]}
ECC Image Encryption Using Matlab Simulink Blockset 843

Fig. 9 Block Diagram Approach for dual image decryption

Fig. 10 Flowchart encrypt and decrypt system

Let K = (146, 2377) ∈ E1758,254 (2543), is Diffie-Hellman’s secret key between

Aicha and Omar, these two entities count iK = (xi , yi ), i ∈ {1, 2, . . . , 256} and
determine their key matrix:
844 S. Chillali and L. Oughdir
⎛ ⎞
1 146 2377
⎜ 2 144 ⎟
⎜ 460 ⎟
⎜ ⎟
⎜ : : : ⎟
A=⎜ ⎟
⎜ i xi yi ⎟
⎜ ⎟
⎝ : : : ⎠
256 867 401

Figure 11 shows input images respective encrypted and decrypted images.

The histograms of the encrypted image are uniform distributed in all gray levels,
therefore they are random, one can observe in Fig. 12, the histograms of the original
and encrypted images which are flat and are derived from the encrypted images.
To assess the effectiveness of image encryption algorithms, a study should be
made on the following points:
• Differential Analysis,
• Statistical Analysis,
• Mean Square Error (MSE) and Peak Signal to Noise Ratio (PSNR)

Fig. 11 Results obtained from the images

ECC Image Encryption Using Matlab Simulink Blockset 845

Fig. 12 Histograms of input and encrypted images

• Image Entropy.
This study is a very important task.
Such an evaluation of these encryption algorithms proposed in this work was
carried in an article entitled: “Image encryption algorithm based on elliptic curves”
and submitted for publication in a journal Soft Computing.

7 Conclusion

Using Matlab Simulink System Generator, we performed an ECC image encryption

approach, this study presents a software simulation for encryption and decryption,
we noticed that the histograms of the encrypted image are uniformly distributed in
all grayscale, the efficiency of this type of encryption was studied in another article
submitted in a journal Soft Computing.

Acknowledgments The authors thank the reviewers and everyone who contributed to this work.

References

1. Diffie W, Hellman M (1976) New directions in cryptography. IEEE Trans Inf Theory
2. Chillali S, Oughdir L (2019) A diagram of confidentiality of information during a traffic offence.
AIP Conf Proc 2074:020028
3. Zeriouh M, Chillali A, Boua A (2019) Cryptography based on the matrices. Bol Soc Paran Mat
37(3):75–83
846 S. Chillali and L. Oughdir

4. Boulbot A, Chillali A, Mouhib A (2020) Elliptic curves over the ring R. Bol Soc Paran Mat
38(3):193–201
5. Hua ZY, Zhou YC, Pun CM, Chen CLP (2015) 2D sine logistic modulation map for image
encryption. Inf Sci 297:80–94
6. Zhang Y (2018) The unified image encryption algorithm based on chaos and cubic. S-Box Inf
Sci 450:361–377
7. Zhang L, Zhang F (2009) A new certificate less aggregate signature scheme. Comput Commun
32(6):1079–1085
8. Xiong H, Guan Z, Chen Z, Li F (2014) An efficient certificate less aggregate signature with
constant pairing computation. Inform Sci 219:225–235
9. He D, Chen J, Zhang R (2012) An efficient and provably-secure certificate less signature scheme
without bilinear pairings. Int J Commun Syst 25:1432–1442
10. Karati A, Islam SKH, Biswas GP (2018) A pairing-free and provably secure certificate less
signature scheme. Inf Sci 450:378–391 (2018)
11. Essaid M, Akharraz I, Saaidi A, Mouhib A (2019) Image encryption scheme based on a new
secure variant of Hill cipher and 1D chaotic maps. J Inf Secur Appl 47:173–187
Development of Large Chaotic S-boxes
for Image Encryption

Younes Qobbi, Abdeltif Jarjar, Mohamed Essaid, and Abdelhamid Benazzi

Abstract In this work, we suggest a novel algorithm for construction of large chaotic
confusion and diffusion S-boxes for gray scale and color image scrambling. The
modification of the value of a pixel by these S-boxes, depend not only on its gray
level intensity but also on its position in the original image. In order to raise the
impact of avalanche effect and to make our method robust against differential attack,
we have developed a strong link between the ciphered pixel and the next original.
The simulations carried out by our system on a large number of randomly selected
images show the effectiveness of this crypto system against any known attack.

Keywords S-box · Substitution · Diffusion · Image encryption · Image decryption

1 Introduction

With the rapid development of information and communication technology, securing

data transmission in particular digital images trough communication networks
becomes a very important hot topic. Digital images are characterized by a large
amount of information and also by a strong correlation between adjacent pixels.
Image encryption based on the use of chaotic dynamic systems is an effective ways to
ensure the security of this transmission [5]. Chaotic systems obtained by deterministic
equations have crucial advantages in the generation of pseudorandom numbers; these

Y. Qobbi (B) · A. Benazzi

Mohamed First University, HSTO, AMSPCS Laboratory, Oujda, Morocco
e-mail: [email protected]
A. Benazzi
e-mail: [email protected]
A. Jarjar
High School Moulay Rachid, Taza, Morocco
M. Essaid
Sidi Mohamed Ben Abdellah University, LSI, Taza, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 847
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_77
848 Y. Qobbi et al.

systems are characterized by extreme sensitivity to initial conditions and to control

parameters. These systems are widely used in image encryption systems [1]. Several
chaos-based crypto systems are proposed. Fridrich [2] propose an image encryption
system based on complex permutations created by using two-dimensional chaotic
maps, and a simple diffusion mechanism. In ref [3], authors proposed an encryption
method using the bit-level and on the use of a new hyper chaotic map. In ref [4]
authors presented a new image encryption algorithm based on the improvement of
several chaotic maps (logistics map, sine map and Chebyshev map) and also on the
use of a new version of Hill cipher, which is deemed more secure. In ref [5], authors
proposed an enhancement of the classical Hill algorithm, this improvement is due to
the use of an invertible matrix (3 × 3) and of a vector of dynamic translation chaotic.
Through to the important cryptographic characteristics of the substitution tables,
S-boxes (Substitution boxes) have attracted the attention of designers of crypto
systems. Several approaches for construction S-boxes based on chaos are proposed.
In ref [6], authors have proposed a new method for construction an S-box with high
performances. In ref [7], authors propose an image encryption system based on the
use of dynamic S-boxes, so that S-boxes are generated according to the size of the
original image. This algorithm is based on a method for scrambling the pixels of the
plain image and another method for substitution of the pixels by using the previous
generated S-box. In ref [8], authors are used a new chaotic system Sine-Tent noted
(STS) for construction of double S-boxes. In ref [1], the author presents a new image
crypto system using the same encryption and decryption mechanism. This crypto
system named Unified image encryption system. This system is based on the use of
the cubic S-box. In ref [9], author presents an encryption algorithm based on the use
of multi-parameter chaotic systems for the construction an S-box. In ref [10], authors
propose an algorithm of image encryption using permutation and substitution, in this
algorithm a new chaotic system is used to generate an S-box of high performances. In
ref [11], authors propose an efficient technique to construct S-boxes based on the use
of a cubic polynomial map. In ref [12], authors propose an encryption scheme based
on permutation-substitution. In the first phase the plain image is permuted row- and
column-wise. In the second phase an encryption of pixels is established by using the
dynamics S-boxes generated by a novel technique based on the logistic map.
In this work, a novel technique is developed to create three S-boxes of dynamic
size. Our approach based on Substitution-Diffusion. Pixel’s substitution established
by using previously generated S-boxes and a new substitution formula. The three
S-boxes changed their places to establish the Diffusion process.
The rest of this article is organized as follows. In the next Sect, the technique of
construction of three S-boxes is described in detail. In the Sect. 3, the encryption
and decryption processes are presented. The performance analysis and simulation
are described in Sect. 4. At the end, conclusion of the present contribution is drawn
in Sect. 5.
Development of Large Chaotic S-boxes for Image Encryption 849

Fig. 1 Preparation of plain

image

2 Proposed Method

Our technique is based on the use of three large chaotic S-boxes generated from
the chaotic maps to establish substitution and diffusion. The substitution of a pixel
depends not only on its value but also on its position. This technique is based on the
following steps:
• Preparation of Plain Image.
• Generation of Chaotic Sequences.
• Encryption Settings.
• Encryption Process.
• Decryption Process.

2.1 Preparation of Plain Image

After loading of the plain image of size h × w: where h and w present the height and
width of plain image respectively. This image is subdivided into three vectors VR,
VG and VB of size 1 × t : where t = h × w. which are represents the red, green and
blue channels respectively. This transformation is illustrated in the Fig. 1.

2.2 Chaotic Sequences Generation

For the generation of encryption keys, our method is based on the use of chaotic
dynamic systems. Indeed we use two chaotic maps.

Logistic Map. Is a chaotic dynamic system most used for the generation of pseudo-
random number sequences. It is given by the following expression:

{xn+1 = µ1 xn (1 − xn ) : µ1 ∈ [3.57, 4] and x0 ∈ [0.5, 1[ (1)

850 Y. Qobbi et al.

Tent Map. It is the simple chaotic dynamic system used in image encryption. It is
described by the recursion relation:

µ2 yn i f yn < 0.5
yn+1 = : µ2 ∈ [0, 2] and y0 ∈ ]0, 1]. (2)
µ2 (1 − yn )Other wise

2.3 Encryption Settings

Our approach is based on the use of three S-boxes (SR, SG and SB) of size t × 256.
Where t = h × w. Which are used to encrypt the channels (Red, Green and Blue).
Generating technique of three s-boxes is similar. In the next section, we present the
method of generating the one of the three S-boxes dented SX. Where SX represents
SR, SG or SB.

Method of Generating S-box (SX). The proposed method to generate the SX is

described by steps below:

Step 1: the initial conditions and the control parameters are set for the logistic
and tent maps.
Step2: the logistic and tent maps are iterated for 3 × 256 times to generate two
random sequences X and Y of size (3 × 256).
Step 3: The two sequences X and Y are combined to generate three chaotic vectors
P, Q and R of size (1, 256) by using the following algorithm:
P = (int)((max(X(i),Y(i)) × 1012 ) i = 0 to 255
Q = (int)((max(X(i),Y(i)) × 1012 ) i = 256 to 2 × 256–1
R = (int)((max(X(i),Y(i)) × 1012 ) i = 512 to 3 × 256–1
Step 4: By sorting in ascending order the elements of the vectors P, Q and R, three
permutation vectors U, V and W of size (1, 256) are obtained.
Step 5: the elements of the vectors U, V and W, which contain the values
in [[0, 255]] are used to initialize the three first rows of SX. The other rows are
generated by the following algorithm:
SX(i,j) = SX(i-2 , SX(i-1,j)) i = 3 to t-1 and j = 0
to 255
Steps 1 to 5 are repeated three times to generate SR, SG and SB.
Development of Large Chaotic S-boxes for Image Encryption 851

Initialization vector. In order to increase the performance of our algorithm, a diffu-

sion phase is proposed. Indeed, an initialization vector Iv of size (1, 3) is generated
from three vectors VR, VG and VB by using the following algorithm:
Iv(0) = Iv(0) ⊕ VR(i) i = 1 to t-1
Iv(1) = Iv(1) ⊕ VG(i) i = 1 to t-1
Iv(2) = Iv(2) ⊕ VB(i) i = 1 to t-1

2.4 Encryption Process

After decomposing the Plain image into three vectors VR, VG and VB and generation
the encryption parameters, the proposed encryption process is applied according to
the following steps:
Step 1: Modification of the first pixel of each vector VR, VG, and VB by using
the initialization vector (Iv) to start the diffusion process.
Step 2: The three S-boxes (SR), (SG) and (SB) are used to encrypt the pixels of
three color vectors VR, VG and VB. This encryption of a pixel depends not only on
its value but also on its position in the original image, using the following algorithm:
VR(i) = SR (i,VR(i)) where i = 0 to t-1
VG(i) = SG (i,VG(i)) where i = 0 to t-1
VB(i) = SB (i,VB(i)) where i = 0 to t-1
Step 3: In order to establish a diffusion process, the encrypted pixel is linked to
the next original pixel, by using the strong expression as follows:
VR(i + 1) = SB(i, VR(i)) ⊕ VR(i + 1) where i = 0 to t-2
VG(i + 1) = SR(i, VG(i)) ⊕ VG(i + 1) where i = 0 to t-2
VB(i + 1) = SG(i, VB(i)) ⊕ VB(i +1) where i =0 to t-2
Where ⊕ is the Xor operation.
The diagram below (Fig. 2) represents the encryption mechanism of the suggested
scheme.

2.5 Decryption Process

Our approach is based on diffusion process, which requires starting the decryption
process from the last pixel to the first. This process goes through the following steps:
Step 1: split the cipher image of size (h, w) into three vectors VR, VG and VB
of size (1, t) where t = h × w.
852 Y. Qobbi et al.

Fig. 2 Encryption process

Step 2: Construction the inverse S-boxes (RS, G S and B S) by using the

following algorithm:

Step 3: The three large invers S-boxes (RS, G S and B S) are used to find the
original pixel by using the following algorithm:
Development of Large Chaotic S-boxes for Image Encryption 853

3 Experimental Results

To prove the effectiveness of our scheme against any known attack. We present
several simulation results obtained by application of our crypto system on Lena (256
× 256), Peppers (512 × 512), and Baboon (512 × 512), The results are given in the
following diagram (Fig. 3).

4 Security Analysis

The main objective of any encryption system is to secure the information from any
known attack. Some security analysis has been applied on the proposed algorithm
which includes correlation, histogram, key space, sensitivity and entropy analysis.
The result analysis of our system is described as follows.

4.1 Correlation Analysis

The adjacent pixels in data image are generally highly correlated. So an efficient
scheme is the one that allows removing this correlation to resist against statistical
attacks. The correlation between adjacent pixels is given by the following equation:

E((x − E(x))(y − E(y)))

Corr x y = √ (3)
D(x) × D(y)
854 Y. Qobbi et al.

Fig. 3 Plain, Ciphered and decrypted images

Table 1 Correlation analysis

Horizontal Vertical Diagonal
Plain Lena (256 × 256) 0.9354 0.9576 0.8964
Cipher Lena (256 × 256) −0.0107 −0.0059 −0.0063
Plain Peppers (512 × 512) 0.8397 0.7195 0.6985
Cipher Peppers (512 × 512) 0.0123 −0.0052 −0.0023

1 N 1 N
E(x) = xi and E(y) = yi (4)
N 1 N 1

1 N 1 N
D(x) = (xi − E(x))2 and D(y) = (yi − E(y))2 (5)
N 1 N 1

Where E is the expected value operator and D is the variance of variable.

The results prove that the values of these coefficients are very close to zero. So
our algorithm is robust against statistical attacks. These results are shown in Table 1.
Development of Large Chaotic S-boxes for Image Encryption 855

Fig. 4 Histogram analysis

4.2 Histogram Analysis

An image histogram is a graphical representation of the distribution of pixels

according to their grayscale values in that image. The histograms of all the images
encrypted by our crypto system are uniform as shown in the figure (Fig. 4). Which
prove that our approach is robust against attacks by histogram.

4.3 Entropy Analysis

Entropy information indicates the level of randomness in a source of information. It

is given by the following expression:
255
H(m) = − Pr(mi ) × logPr(mi ) (6)
i=1

Where Pr(mi ) is the probability of random variable m of ith index.The ideal entropy
for 256 symbols representing the different gray levels in an 8-bit image is H(m) = 8.
856 Y. Qobbi et al.

Table 2 Entropy analysis

Cipher Image Proposed algorithm Ref [1] Ref [12]
Lena 7.9990 – 7.9046
(256 × 256)
Baboon 7.9998 7.999285 –
(512 × 512)
Peppers 7.9998 7.999314 –
(512 × 512)

The entropy of the different ciphered images obtained by our scheme is near to 8.
That prove our scheme can resist against the attacks by entropy. The information
entropy of our algorithm and some references are listed in the Table 2 below.

4.4 Brute Force Attack

Key Space
A good encryption system should have a key space large enough to remain strong
against brute force attacks. Indeed the key space must be greater than 2100 [13].The
key space of our method is composed of two initial conditions x0 and y0 and two
control parameters µ1 and µ2 . These parameters are the float (32 bit). Then the key
space of our system is 2128 . This makes our system safe to any brute force attack.

4.5 Differential Attacks

A strong encryption system is one that is sensitive to a small change in the original
image such as a single pixel change in the original image produces entirely different
ciphered image. This sensitivity is evaluated through a Number of Pixels Change
Rate (NPCR) and Unified Average Changing Intensity (UACI). Their definitions are
presented as follows.
w  h
1 Di j
N PC R = 1
× 100% (7)
w∗h
  

1 i j I C1i j − I C2i j
U AC I = × 100% (8)
w×h 255

IC1 = [IC1i j ] and IC2 = [IC2i j ] express two ciphered images corresponding to
two plain images with a tiny difference, and Di j is calculated by

1 i f I C1i j = I C2i j
Di j = (9)
0 else
Development of Large Chaotic S-boxes for Image Encryption 857

Table 3 NPCR and UACI analysis

Image Proposed algorithm Ref [1] Ref [12]
NPCR UACI NPCR UACI NPCR UACI
Baboon (512 × 512) 100 33.6019 99.6100 33.4643 --------- --------
Peppers (512 × 512) 100 33.5819 99.6089 33.4689 -------- -------
Lena (256 × 256) 100 33.5862 ---------- --------- 98.9182 32.7865

The following Table 3 below shows the measurement of NPCR and UACI between
two cipher images of the Babon (512 × 512) Lena (256 × 256) and Peppers (512 ×
512). When making a slight modification in the original image.

5 Conclusion

In this article, a novel technique is developed to create three chaotic S-boxes. Which
are used to perform two main processes of any image encryption system: Substitution
and diffusion. The main contributions of our approach are: First, the modification of a
value of a pixel depends on two parameters which are: its value and its position in the
plain image. Second, our algorithm uses three large chaotic S-boxes of dynamic size.
The simulation results and security analysis presented by a correlation coefficient
close to zero, a flat histogram of the ciphered image and an entropy value near to 8,
prove that our algorithm is strong against any known attack.

References

1. Zhang Y (2018) The unified image encryption algorithm based on chaos and cubic S-Box. Inf
Sci 450:361–377
2. Fridrich J (1998) Symmetric ciphers based on two-dimensional chaotic maps. Int J Bifurcat
Chaos 8(06):1259–1284
3. Cao C, Sun K, Liu W (2018) A novel bit-level image encryption algorithm based on 2D-LICM
hyperchaotic map. Signal Process 143:122–133
4. Essaid M, Akharraz I, Saaidi A (2019) Image encryption scheme based on a new secure variant
of Hill cipher and 1D chaotic maps. J Inf Secur Appl 47:173–187
5. Hraoui S, Gmira F, Abbou MF, Oulidi AJ, Jarjar A (2019) A new cryptosystem of color image
using a dynamic-chaos hill cipher algorithm. Procedia Comput Sci 148:399–408
6. Belazi A, Abd El-Latif AA (2017) A simple yet efficient S-box method based on chaotic sine
map. Optik 130:1438–1444
7. Wang X, Wang Q (2014) A novel image encryption algorithm based on dynamic S-boxes
constructed by chaos. Nonlinear Dyn 75(3):567–576
8. Zhu S, Wang G, Zhu C (2019) A secure and fast image encryption scheme based on double
chaotic s-boxes. Entropy 21(8):790
9. Khan M (2015) A novel image encryption scheme based on multiple chaotic S-boxes. Nonlinear
Dyn 82(1–2):527–533
858 Y. Qobbi et al.

10. Lu Q, Zhu C, Deng X (2020) An efficient image encryption scheme based on the LSS chaotic
map and single S-box. IEEE Access 8:25664–25678
11. Zahid AH, Arshad MJ (2019) An innovative design of substitution-boxes using cubic
polynomial mapping. Symmetry 11(3):437
12. Rehman AU, Khan JS, Ahmad J, Hwang SO (2016) A new image encryption scheme based on
dynamic s-boxes and chaotic maps. 3D Res 7(1), 7.
13. Wu X, Zhu B, Hu Y, Ran Y (2017) A novel color image encryption scheme using rectangular
transform-enhanced chaotic tent maps. IEEE Access 5:6429–6436
Evaluation of Feature Extraction
Methods Combined with Support Vector
Machines for Powerline Component
Recognition in Aerial Images

Jamila Garfaf, Lamyae Fahmani, and Hicham Medromi

Abstract As part of its efforts to maintain the stability and sustainability of power
generation for its consumers, the Moroccan National Office of Electricity and
Drinking Water (ONEE) conducts regular inspections of its transmission and distri-
bution networks. Compared to other inspection methods, UAV inspection enables
regular inspection with minimized effort and lower cost and provides more detailed
images of network components safely and securely. However, human analysis of the
amount of data produced by the UAV may not be efficient. Therefore, the image
analysis process’s automation should be implemented by developing advanced algo-
rithms to detect components and their defects. In this work, we focused on object
recognition, which we consider a key step in the development of more complex tasks.
We evaluated HOG, LBP, and SIFT extractors’ recognition performance in the iden-
tification of power line components, by combining each of these extractors with the
SVM classifier. The results showed that combining the HOG and SVM can give the
best results compared to the LBP-SVM and SIFT-SVM methods with an accuracy
rate of 97.98% and a shorter acquisition time for image processing. These results
indicate that this combination can meet electric companies’ technical requirements
related to accuracy and speed.

Keywords Object recognition · SVM · Feature extraction · HOG · LBP · SIFT ·

Unmanned Aerial Vehicles (UAV) · Powerline equipment

1 Introduction

In order to maintain the reliability, stability, and sustainability of electricity produc-

tion for its consumers, the National Office of Electricity and Drinking Water (ONEE),

J. Garfaf (B) · L. Fahmani · H. Medromi

Research Foundation for Development and Innovation in Science and Engineering (FRDISI), 16
469 Casablanca, Morocco
Engineering Research Laboratory (LRI), System Architecture Team (EAS), National and High
School of Electricity and Mechanic (ENSEM) Hassan II University, 8118 Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 859
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_78
860 J. Garfaf et al.

the pillar of the energy strategy and the armed arm of the state in the water and sanita-
tion sector in Morocco [1], conducts regular diagnoses and visual inspection of their
transmission and distribution networks. These inspection missions within the ONEE
were until recently carried out by patrols on the ground and by helicopters. However,
these traditional methods have several drawbacks. In addition to the fact that these
methods are expensive and too slow, they jeopardize the safety of individuals given
that most components are located in hard-to-reach places and that these inspection
operations are often carried out under more difficult meteorological conditions [2, 3].
Accordingly, the ONEE devotes significant investment and effort to carry out these
types of operations.
To ensure inspection and maintenance operations with the lowest possible cost
and effort as well as ensure the safety of agents and operators, the ONEE has resorted
to the implementation of UAV technology at its transmission and production units.
Compared to other inspection methods based on foot patrols and helicopter-assisted
inspections, operating costs are lower, and the UAVs can fly safely as close as possible
to the power line element, allowing more detailed photos of the components to be
taken [3–6]. However, given the enormous amount of data produced by the UAV,
manual human analysis can be ineffective. This explains the use of artificial intelli-
gence as an engineering solution to replace humans and optimize results, as is the
case for many other engineering problems [7]. Indeed, it is necessary to develop effi-
cient and advanced algorithms to automate the process of identifying the presence
of anomalies or damage on inspected power lines by exploiting the aerial vector of
photos and videos taken and produced remotely by the UAV. Numerous studies are
being conducted on the monitoring and inspection of power transmission systems.
In this work, we are focusing on the detection and recognition of power network
components as an essential part of the development of complex algorithms.
In [8], a supervised learning approach has been proposed to detect and classify
pylons. This approach uses Histogram of Oriented Gradients (HOGs) to train two
MLP (Multilayer Perceptron) neural networks. The first classification separates the
foreground from the background, and the second multilayer MLP distinguishes the
four different types of electrical pylons described in the paper. The same feature
extractor has been used in [9] for the detection of molten insulators by the local
binary model (LBP). The extracted features were transferred to the SVM (Support
Vector Machine) classifier. An evaluation of the proposed method was performed on
500 images, which gave a detection rate of 89.1%.
Using the infrared images introduced by the UAV, ZHOU et al. [10] designed a
system for fault diagnosis and remote isolator detection. Each isolator is located using
the Scale-invariant Feature Transform (SIFT) method. The obtained results showed
that the choice of SIFT allows the identification and localization of isolators with a
high accuracy. This boost is thanks to the invariance of scale, and to the invariance
to variations in rotation angle and image brightness. Because of its performance, the
features of SIFT are also used by [11]. They are mainly used to detect the tilt angle
of the electric pylon during video surveillance in the substation.
Among the most prominent traditional object extractors, are: Scale-Invariant
Features Transform (SIFT) [12], Histogram of Oriented Gradients (HOG) [13], Local
Evaluation of Feature Extraction Methods … 861

Binary Patterns (LBP) [14] which have been used in many works, we can cite among
others [15–17]. In order to evaluate their recognition performance in the identifica-
tion of power line components, each of these extractors will be combined with the
famous SVM classifier which has proven in many works its performance in computer
vision applications. The experiments were conducted on a set of images of power
line components captured from videos taken by the UAV.
We organized this paper as follows: the second section describes in detail the
methodology adopted and the algorithms studied and used. The results are presented
and discussed in the third section. The last section concludes the paper while
highlighting future research directions.

2 Methodology

In the computer vision domain, object recognition is a crucial and challenging task.
It consists of determining the predefined category to which an object in a given
image belongs [18]. The recognition process consists of two main steps: the feature
extraction step, followed by the classification step [19]. Indeed, before the received
image is fed into the recognition process, a pre-processing phase is necessary to
improve the quality of the input images so that the relevant features can be better
extracted from the image. This preprocessing phase consists firstly of converting the
images produced by the UAV, which are generally in RGB colors, into grayscale
[20, 21] and secondly of resizing them to a smaller fixed size of 256 * 256 so that
no important information is missing. This allows to prevent the overloading and
saturation of the network and to reduce the computing time. Then, a feature extractor
is placed whose objective is to represent the recovered image as a descriptive vector
based on the extracted features. This vector is afterwards passed to the classifier to
recognize the class to which the image belongs (see Fig 1).
For an image recognition system to be effective, it must have a well-defined
and highly targeted feature extraction step. To achieve this, the features must be as
distinctive and informative as possible in order to obtain reliable identification. In fact,
if the features are not well selected, this could significantly impair the performance
and efficiency of the recognition system.
Among the most popular traditional object extractors are SIFT, HOG, LBP, which
have been used in many works. These extractors will be the main subject of this paper.
An evaluation of their recognition performance will be performed in our paper by

Fig. 1 Block diagram of the recognition process

862 J. Garfaf et al.

combining them with the famous SVM classifier, which is becoming more and more
popular [22, 23] due to its performance in the classification of noisy and high-
dimensional data [24]. The experiments are conducted on a data set constructed
from aerial images of power line components.
A detailed description of the HOG, LBP and SIFT extractors and the SVM
classifier is given in the following sections.

2.1 Histogram of Oriented Gradients Extractor

The gradient histogram was proposed initially by Dalal and Triggs [13] for human
detection. The gradient-oriented histogram is used to extract features using informa-
tion about the orientation of the gradient from the appearance and shape of the local
object.
The calculation of HOG descriptors is based on four key operations: calculation
of the amplitude and orientation of the image gradient, generation of cell histograms,
histogram normalization and descriptor generation. First, a detection window of 64
* 128 pixels is divided by cells of 8 * 8 pixels, forming 8 * 16 = 128 cells. For
each pixel (x, y), the gradient magnitude m(x, y) and the orientation θ(x, y) can be
calculated using Eqs. (1) and (2), respectively:

m(x, y) = G x (x, y)2 + G y (x, y)2 (1)
 
G y (x, y)
(x, y) = arctan (2)
G x (x, y)

where G x (x, y)et G y (x, y) represent the gradient of the image along the two
directions x and y successively as given below:

G x (x, y) = I (x + 1, y) − I (x − 1, y) (3)

G y (x, y) = I (x, y + 1) − I (x, y − 1) (4)

With I (x, y) represent the intensity of the image in the pixel (x,y). Then, a
histogram is generated for each cell containing nine boxes. The amplitude of each
cell in the histogram is calculated from the accumulation of the amplitudes of the
pixels with directions belonging to the interval of the cell. A 9 × 1 vector is then
obtained for each cell and presented as a histogram. The blocks are normalized to
give a 36 * 1 vector in order to have a better invariance to changes in illumination and
contrast. The process is repeated by passing through all the pixels of the image. In
the end, all 36 × 1 normalized vector is grouped into a single vector corresponding
to the HOG descriptor.
Evaluation of Feature Extraction Methods … 863

Fig. 2 LBP Operator calculation example

2.2 Local Binary Pattern Extractor

Local Binary Pattern (LBP), originally proposed by Ojala in 1996 [14], allows char-
acterizing the spatial structure of the image texture using a 3 * 3 neighborhood around
a central pixel. Based on a 3 * 3 neighborhood, the LBP compares the gray values
of the 8 neighboring pixels to the central pixel, which is taken into account as a
threshold.
Let gc , g0 ,…, g7 denote respectively the grey levels of the central pixel and its
eight neighboring pixels. The LBP code of the central pixel characterized by its
coordinates (xc , yc ) is calculated as follows:


7
L B P(xc , yc ) = S(gc − gi ) × 2 p (5)
p=0

With S(z) is the sign function defined as follows:


1x ≥0
S(z) = (6)
0x <0

The Fig 2 shows an example for the calculation of the LBP code
After generating the LBP codes of all pixels of the input image, these codes are
collected in a histogram in the form of a texture descriptor.

2.3 Scale Invariant Feature Transform Descriptor

SIFT (Scale Invariant Feature Transform) is an algorithm introduced by David Lowe

in 2004 [25]. The principle of the SIFT algorithm consists in converting an image into
a set of feature descriptive vectors that are invariant to scale, rotation and partially
invariant to illumination. The four steps describing the SIFT method are described
below.

Scale-Space Extrema Detection: :This step aims to scroll through the image at
different scales and in different octaves in order to extract the points of interest that
864 J. Garfaf et al.

are invariant to the scale. For this, the input image is convoluted with a centered
Gaussian with a progressive variation of σ. A succession of images with different
levels of blur is obtained, thus constructing a Gaussian pyramid. The points of interest
are considered as the local maxima and minima in DoG images that appear at different
scales through the application of the Gaussian difference between two successive
images of the same scale.
Keypoints Localization: The first step of the SIFT algorithm produces a vast
number of points of interest, some of which are unstable and cannot be retained as
characteristics. An adjustment is, therefore, necessary to obtain more useful points.
Low contrast points and points located on low curvature contour edges are therefore
eliminated in this step.
Orientation Assignment: In this step, each selected point is assigned one or more
orientations based on the local features of the image. Indeed, for each point of interest,
the algorithm takes a neighborhood, then calculates the amplitude and direction of
the gradient at each pixel to finally build an orientation histogram with 36 graduations
covering 360°. The peaks of the histogram characterize the dominant orientation in
the neighborhood of a point of interest. In order to have more information available,
Lowe suggests using the peaks of the histogram that represent at least 80% of the
dominant peak to create new points of interest that differ in their main orientation.
A set of key points that are invariant in scale and rotation is then created.
Keypoints Descriptor: the descriptors are calculated according to the orientations
of the key points. Indeed, the algorithm defines a neighborhood window of 16 * 16
pixels around the point of interest. This window is divided into 16 blocks of 4 *
4 pixels where the orientation and norm of the gradients are calculated for each of
these blocks. These orientations are then placed in a histogram with 8 graduations
and are weighted by the norm of the gradient at the point of interest to reduce the
illumination change effects. The result is a SIFT feature vector of 16 ∗ 8 = 128
values for each point of interest.

2.4 Support Vector Machines Classifier

SVM was originally developed by Vapnik [26] for two-class classification problems
but was later extended to include multi-class classification problems. SVM aims to
define decision boundaries by providing hyperplanes to classify data samples in data
distinctly. The data of the input space is projected into a larger space using the kernel
function [24].
For a training set labeled (xi , yi ), i = 1, 2… where xi  Rn and yi  {1, −1}, the
SVM proceeds to solve the optimization problem that may be presented as follows:

WTW 
Min w,b,ξ +C ξi (7)
2 i
Evaluation of Feature Extraction Methods … 865

Subject to

yi (w T φ(xi ) + b) ≥ 1 − ξi , ξi ≥ 0 (8)

where w is the weight vector, and b is the bias. The training vector xi is mapped to a
higher dimensional space by the function φ, and the coefficient C is set to adjust the
error tolerance.
Several kernel functions are applied according to the SVM method. The RBF
kernel is used in our paper, whose function is as follows:

  −γ xi −x 2j
K xi , x j = e ,γ > 0 (9)

3 Experiments and Results

Following a detailed study and presentation of feature extraction algorithms in the

classical approach, namely SIFT, HOG and LBP and the famous SVM classifier, we
compare in this section and evaluate their performance and recognition capabilities
by combining each of the three extractors with the SVM classifier.
The three models that have been designed, namely the HOG-SVM model, the
LBP-SVM model and the SIFT-SVM model have been trained and tested using
aerial images of power line components.
Initially, we built the database from videos and images produced by the UAV
during inspection missions carried out on two transmission lines at ONEE stations.
We have also used the data set designed by [27] and [28], which is available for
research purposes. By using data augmentation methods such as flipping (horizon-
tally and vertically) and cropping, we collected 5850 aerial images, divided into three
categories: towers, insulators and conductors. These images were grouped into two
sets: a training set and a test set. Details of the data set are presented in Tables 1 and
2. Some samples of these images are shown in Fig 3.
In order to evaluate and compare the recognition and classification capabilities of
the three models studied in our paper, we used the following statistical parameters:

Table 1 Details of the electrical components DataSet

Type Tower Insulator Conductor Total
Numbers of images 1899 1845 2106 5850

Table 2 Details of the

DataSet Training DataSet Testing DataSet
training and the testing
DataSet Training DataSet 3919 1931
866 J. Garfaf et al.

Fig. 3 Samples of electrical power line equipment aerial images

Fig. 4 Accuracy for the 120 97.98

three models
80
45.78
32.37 Accuracy(%)
40

0
HOG-SVM SIFT-SVM LBP-SVM

precision, recall, F1-score, and accuracy which are defined as follows [29]:

TP
Pr ecision = (10)
T P + FP
TP
Recall = (11)
T P + FN
2T P
F1 − scor e = (12)
2T P + T N + F P + F N
TP +TN
Accuracy = (13)
T P + T N + FP + FN

Where TP, FP, TN and FN are the numbers of true positives, false positives, true
negatives, and false negatives in the detection results, respectively.
As shown in Fig 4, it can be seen that the HOG-SVM model achieves the best
performance in power line recognition with 97.98% accuracy, compared to SIFT-
SVM and LBP-SVM, which achieved an accuracy of 45.78% and 32.37%, respec-
tively. In addition, Table 3 shows the other evaluation parameters related to the
Evaluation of Feature Extraction Methods … 867

Table 3 Performance results

Models Precision Recall F1-score
of the three models
HOG-SVM 97.99% 97.96% 97.97%
SIFT-SVM 46.26% 44.34% 35.32%
LBP-SVM 10.79% 33.33% 16.30%

Table 4 Image computation

HOG-SVM SIFT-SVM LBP-SVM
time
Computation time 0.349 s 2.885 s 0.392 s

accuracy, recall and f1 scores. It shows that the combination of HOG and SVM
presents a good compromise between accuracy and recall. Recall reaches 97.96%,
indicating that most components are correctly detected.
Besides, given that the solution will be implemented in real-time applications [30],
the image processing time for each model must also be calculated and compared. As
shown in Table 4, the image processing time for the HOG-SVM model is 0.349 s,
which is less than the other two models. Both the SIFT-SVM model and the LBP-
SVM are very time-consuming, resulting from the large amount of data and the high
resolution of the collected images.
Experimental results, based on the comparative study and evaluation of models
designed using different evaluation measures, show that the method in which HOG
is the extractor combined with SVM gives the best results in terms of accuracy
and computing time. These results indicate that this combination allows utilities to
meet their technical requirements for accuracy and speed when performing off-line
analysis for transmission line inspection.

4 Conclusion

In this paper, we have focused on component recognition, which we see as an essential

step towards the overall goal of developing an intelligent system for the identification
of power line component dysfunctions. We have evaluated HOG, LBP, and SIFT
extractors’ recognition performance in the identification of power line components,
combining each of these extractors with the famous SVM classifier.
We aim to study and evaluate the deep approach with convolutional neural network
architectures on larger power line equipment datasets in our future work. This is in
order to study in more detail the efficient model that could be implemented in future
algorithms for detection and identification of malfunctions and anomalies and to
collect and design large power line equipment datasets that we want to make available
for research purposes.

Acknowledgement This work was supported by the National Center for Scientific and Technical
Research (CNRST, Morocco) and the National Office for Electricity and Drinking Water (ONEE,
868 J. Garfaf et al.

Morocco). We want to express our gratitude to the ONEE teams for their collaboration and support
for successfully conducting our research project.

References

1. The official Website of the ONEE - Electricity sector (2020). http://www.one.org.ma/. Accessed
26 Oct 2020
2. Nguyen VN, Jenssen R, Roverso D (2018) Automatic autonomous vision-based power line
inspection: a review of current status and the potential role of deep learning. Int J Electric
Power Energy Syst 99: 107–120
3. Fahmani L, Garfaf J, Boukhdir K, Benhadou S, Medromi H (2020) Modelling of very high
voltage transmission lines inspection’s quadrotor. SN Appl Sci 2: 1425
4. Liu X, Miao X, Jiang H, Chen J (2020) Review of data analysis in vision inspection of power
lines with an in-depth discussion of deep learning technology. arXiv:2003.09802[cs,eess]
5. Jalil B, Leone G, Martinelli M, Moroni D, Pascali M, Berton A (2019) Fault detection in power
equipment via an unmanned aerial system using multi modal data. Sensors 19: 3014
6. Fahmani L, Garfaf J, Boukhdir K, Benhadou S, Medromi H (2020) Unmanned aerial vehicles
inspection for overhead high voltage transmission lines. In: 2020 1st International Conference
on Innovative Research in Applied Science, Engineering and Technology (IRASET). pp 1–7.
IEEE, Meknes, Morocco
7. Chergui M, Chakir A, Medromi H (2019) Smart IT governance, risk and compliance semantic
model: business driven architecture. In: 2019 Third World Conference on Smart Trends in
Systems Security and Sustainablity (WorldS4), pp 297–301. IEEE, London, United Kingdom
8. Sampedro C, Martinez C, Chauhan A, Campoy P (2014) A supervised approach to electric tower
detection and classification for power line inspection. In: 2014 International Joint Conference
on Neural Networks (IJCNN), pp 1970–1977. IEEE, Beijing, China
9. Tiantian Y, Guodong Y, Junzhi Y (2017) Feature fusion based insulator detection for aerial
inspection. In: 2017 36th Chinese Control Conference (CCC), pp 10972–10977. IEEE, Dalian,
China
10. Shen-pei Z, Xi L, Bing-chen Q, Hui H (2017) Research on insulator fault diagnosis and remote
monitoring system based on infrared images. Procedia Comput Sci 109: 1194–1199
11. Yu P, Dong BG, Xue YJ (2012) Electric power tower inclination angle detection method based
on SIFT feature matching. AMM. 236–237: 759–764
12. Lowe DG (1999) Object recognition from local scale-invariant features. In: Proceedings of
the Seventh IEEE International Conference on Computer Vision, pp 1150–1157, vol 2. IEEE,
Kerkyra, Greece. https://doi.org/10.1109/ICCV.1999.790410
13. Dalal N, Triggs B (2005) Histograms of oriented gradients for human detection. In: 2005 IEEE
Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2005), pp
886–893. IEEE, San Diego, CA, USA
14. Ojala T, Pietikäinen M, Harwood D (1996) A comparative study of texture measures with
classification based on featured distributions. Pattern Recogn 29: 51–59
15. Jones MJ, Viola P (2010) Robust Real-time Object Detection, 30
16. Taşkiran M, Çam ZG (2017) Offline signature identification via HOG features and artifi-
cial neural networks. In: 2017 IEEE 15th International Symposium on Applied Machine
Intelligence and Informatics (SAMI), pp 000083–000086
17. Zhang H, Qu Z, Yuan L, Li G (2017) A face recognition method based on LBP feature for CNN.
In: 2017 IEEE 2nd Advanced Information Technology, Electronic and Automation Control
Conference (IAEAC), pp 544–547
18. Liu L, Ouyang W, Wang X, Fieguth P, Chen J, Liu X, Pietikäinen M (2019) Deep Learning for
Generic Object Detection: A Survey. arXiv:1809.02165[cs]
Evaluation of Feature Extraction Methods … 869

19. Yao G, Lei T, Zhong J (2019) A review of Convolutional-Neural-Network-based action

recognition. Pattern Recogn Lett 118: 14–22
20. RGB to grayscale — skimage v0.19.0.dev0 docs. https://scikit-image.org/docs/dev/auto_exam
ples/color_exposure/plot_rgb_to_gray.html
21. Poynton C (2015) Frequently asked questions about color. 24
22. Malathi V, Marimuthu NS (2010) Wavelet transform and support vector machine approach for
fault location in power transmission line. Int J Electric Electron Eng 6: 512–513
23. Ray P, Mishra DP (2016) Support vector machine based fault classification and location of a
long transmission line. Eng Sci Technol Int J 19: 1368–1380
24. Niu X-X, Suen CY (2012) A novel hybrid CNN–SVM classifier for recognizing handwritten
digits. Pattern Recogn 45: 1318–1325
25. Lowe DG (2004) Distinctive image features from scale-invariant keypoints. Int J Comput Vis
60: 91–110
26. Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20: 273–297
27. Lee SJ, Yun JP, Choi H, Kwon W, Koo G, Kim SW (2017) Weakly supervised learning with
convolutional neural networks for power line localization. In: 2017 IEEE Symposium Series
on Computational Intelligence (SSCI), pp 1–8. IEEE, Honolulu, HI
28. Tao X, Zhang D, Wang Z, Liu X, Zhang H, Xu D (2020) Detection of power line insulator
defects using aerial images analyzed with convolutional neural networks. IEEE Trans Syst Man
Cybern, Syst 50: 1486–1498
29. Powers DM (2011) Evaluation: from Precision, Recall and F-measure to ROC, Informedness,
Markedness and Correlation
30. Chergui M, Chakir A, Medromi H, Radoui M (2017) A new approach for modeling strategic
IT governance workflow. In: El-Azouzi R, Menasche DS, Sabir E, De Pellegrini F, Benjillali M
(eds.) Advances in Ubiquitous Networking, vol 2, pp 285–298. Springer Singapore, Singapore
A Novel Brain Tumor Detection
Approach Based on Fuzzy C-means
and Marker Watershed Algorithm

Hanae Moussaoui, Mohamed Benslimane, and Nabil El Akkad

Abstract Segmentation has a very essential role in medical imaging. Image segmen-
tation can be defined as the process that gives the ability to separate a random image
into parts or objects that make up the image, which means separating objects from
the background, so we can later analyze each of these components separately. In
this work we’ll propose a method based on the combination between fuzzy c-means
clustering and marker watershed algorithm. The first thing to do is to apply the fuzzy
c-means clustering which is an unsupervised clustering method that is usually used
in medical imaging. Then the first output image will be going through the second
algorithm which is the marker watershed that uses concepts from edge detection
and mathematical morphology. The proposed method was applied and tested on
several medical images to detect anomalies such as tumors, and according to the
results, we can say that it gives a satisfactory and good results of detection. To eval-
uate the performance of our proposed method, we used a set of metrics such as the
“Dice coefficient”, “Sensitivity” and “Specificity”; which all have shown satisfactory
results.

Keywords Image segmentation · Medical imaging · Fuzzy c-means · Watershed

algorithm · Thresholding

1 Introduction

When we talk about Medical imaging we’ll be definitely looking for a technique or
a process of making the interior of the body visually represented. Medical imaging
is used for clinical analysis and medical intervention, such as revealing any structure
hidden by bones or skin. It’s also very helpful when doctors try to diagnose and treat

H. Moussaoui (B) · N. El Akkad

LISA, Engineering, Systems and Applications Laboratory, ENSA of Fez, Fez, Morocco
e-mail: [email protected]
M. Benslimane
LTI Laboratories, EST, Sidi Mohamed Ben Abdellah University, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 871
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_79
872 H. Moussaoui et al.

a disease. Image processing is used in several areas of computer vision, in particular,

image segmentation [1–4], 3D reconstruction [5, 6], camera Self-calibration [7–13]
and cryptography [14–16]. Image segmentation has many approaches to segment
an image such as thresholding, clustering based methods, edge detection and region
growing methods. In this paper we’re going to present a method based on watershed
algorithm which takes into account the gradient of the desired image. The second
algorithm is the fuzzy c-means clustering method which is an unsupervised segmen-
tation technique which allows pixels to be in more than one cluster. In this paper
we will talk in details about the proposed method. Starting by explaining the fuzzy
c-means algorithm, followed by marker controlled watershed and thresholding; after
that we’ll present our proposed method and results, in the end the paper will end up
with a conclusion.

2 Related Works

In the past 10 years many brain tumor detection methods have been proposed, we
can start by fuzzy c-means algorithm which plays a very important role in medical
image segmentation, nevertheless it has many drawbacks. By using fuzzy c-means
algorithm we need first to set the number of clusters which is considered as a disad-
vantage, and also it’s known with it sensitiveness about noise and other artifacts. M.
Sreerangappa, M. Suresh and D. Jayadevappa [17] have proposed a modified fuzzy
c-means algorithm known as a spatial FCM clustering to avoid all the drawbacks
listed above. For this methodology they started by pre-processing the MRI image by
applying the Median Filter followed by Wavelets Transform and Spatial FCM. Nilesh
Bhaskarrao Bahadure, Arun Kumar Ray and Har Pal Thethi [18] have proposed a
combination between Berkeley Wavelet Transformation (BWT) and SVM that helps
to classify MRI images to healthy or infected tumor. P. Kumar and B. Vijayakumar
[19] have proposed a brain tumor segmentation and classification methodology, based
on Principal Component Analysis (PCA) and Radial Basis Function (RBF).

3 Fuzzy C-means Algorithm

Fuzzy c-means is an extension of k-means clustering is considered as a soft computing

technique, which means that each pixel in the image can belong to multiple clusters.
Fuzzy c-means is widely used in medical imaging [20–22], especially for brain tumor
detection. Fuzzy c-means clustering works in five steps:
Step (1): Initialize randomly the membership matrix using the following equation:


c
uj (xi ) = 1 i = 1, 2, . . . , k (1)
j=1
A Novel Brain Tumor Detection Approach Based on Fuzzy C-means … 873

Step (2): Calculate the centroid using the equation:

  m
i u j (x i ) x i
Cj =   m (2)
i u j (x i )

Step (3): Calculate dissimilarity between the data points and centroid using the
Euclidean distance:

Di = (x2 − x1 )2 (y2 − y1 )2 (3)

Step (4): Upgrade the new membership matrix. For this we use the equation
bellow:

 1/
1 m−1
d ji
u j (xi ) = (4)
c  1/
1 m−1
k=1 dki

Here m is a fuzzification parameter. The range m is always [1.25, 2].

Step (5): Repeat the process from step 2, unless the centroids keep the same value
(Fig. 1).

Fig. 1 Fuzzy c-means

diagram Start

Initialize randomly the membership matrix

Calculate the centroid

Calculate”Euclidean distance”

‘”Update the new membership matrix”

Yes Centroids changement

No
Output
874 H. Moussaoui et al.

4 Marker Controlled Watershed Algorithm

Marker watershed is a robust morphological gradient-based segmentation technique

that is used especially in the case when we have to segment objects that are so
close and touching one another. Marker controlled Watershed algorithm is highly
recommended for medical image segmentation [23–25]. The watershed algorithm
works by finding catchment basins and also watershed ridge lines in the input image
by treating it as a topographic surface, where the light pixels are high and the low
pixels are low. Watershed algorithm works in five steps:
Step (1): Transform the input image to Grayscale.
Step (2): Apply the Gradient Magnitude.
Step (3): Mark the main objects of the foreground.
Step (4): Calculate the background markers.
Step (5): Determine the watershed transform.
The main disadvantage of watershed algorithm is that it generally results over
segmentation. For this, we usually de-noise the image before applying watershed
algorithm [26].

5 Thresholding

Image thresholding is a simple form of image segmentation; it’s a subset of image

segmentation [20]. Threshold segmentation is a way to create a binary image based
on the histogram or the threshold value on the pixel intensity of the original image
[28]. Image binarization is the process of separating pixels’ values into two groups,
black for background and white for objects or foreground. Image thresholding can
be grouped in local and global thresholding (Fig. 2).
In this paper will be using Otsu’s thresholding for our image segmentation. Otsu’s
thresholding is based on finding that one threshold that minimizes the weighted within
class variance as well as maximizing the between class variance.
Frequency Formula:


T
ω= P(i) when P(i) = n i /N (5)
i=0

N is the total pixel number.

n i is the number of pixels in level i.
Mean Formula:


T
μ= i P(i)/ω (6)
i  =0
A Novel Brain Tumor Detection Approach Based on Fuzzy C-means … 875

Threshold segmentation

Local Global
”Bernsen”
thresholding thresholding
”Chow and Kanelo”
”Eikvil”
”Mardia and
Hainsworth”
Traditional Multistage Iterative
”Niblack”
(Otsu) (Quadratic ratio (iterative triclass
”Yanowitz and
technique) thresholding)
Bruckstein”
”Tr Singh”
”Sauvola”

Fig. 2 Thresholding diagram

Total Variance Formula:


T
σt2 = (i − μ)2 P(i) (7)
i=0

6 The Proposed Method

In this section we’ll present our proposed method which is made of two stages.
First of all we’ll start by applying fuzzy c-means clustering on our input image
which is supposed to be a MRI image, and then the first output image will be going
through the second algorithm which is the marker controlled watershed. The reason
behind choosing fuzzy c-means clustering wasn’t arbitrary; but, because of it robust
characteristics for ambiguity. This segmentation method is highly recommended
when we’re dealing with medical imaging, also, it’s a soft clustering method that
assigns a membership of every pixel to each cluster. The second stage is to apply
the watershed algorithm to the segmented image. The proposed method gives good
detection results. As we can see that the final result shows clearly the tumor area well
segmented and surrounded by a marker which defines the borders. The flow chart of
the proposed method is as follow:
876 H. Moussaoui et al.

Input image

First Stage
Applying fuzzy c-means segmentation

The first segmented image

De-noise the image with Median Otsu’s thresholding

Second Stage

Watershed algorithm

Generate Markers

Final segmentation

Fig. 3 Flowchart of the proposed method

7 Experimental Results and Discussion

In this experiment part we’ll be using brain web images that have a tumor. As we
can see in the flow chart in the top (Fig. 3), the proposed method is divided into two
stages. In the first stage we’ll apply fuzzy c-means algorithm to the input image,
here, we tried many values of k and the conclusion we can draw is that by choosing
a small value of k (k = 3) we will have good results at the end. After applying
fuzzy c-means we de-noise the first output image using Median filter. In the second
stage of the proposed algorithm, we’ll apply both thresholding and marker watershed
algorithm to identify and segment the tumor using markers in the brain image. As
we can see bellow in (Fig. 4), the proposed algorithm gives a good segmentation and
identification of tumors in different locations.
To evaluate the performance of the proposed method, we used a set of metrics
such as “Dice coefficient”, “Sensitivity” and “Specificity”.
Dice Similarity:

2 ∗ GT ∗ CS /|GT| + |CS| (8)

CS: represents the tumor segmentation and GT is the ground truth.

Sensitivity: is an evaluating metric that indicates the true positivity, which means
that sensitivity gives us the probability that a segmented pixel belongs to the brain
tumor.

Sensitivity = TP/(TP + TN) (9)

A Novel Brain Tumor Detection Approach Based on Fuzzy C-means … 877

Original image Fuzzy c-means watershed After using Final

Otsu threshold segmentation

Fig. 4 A table showing the obtained results using the proposed method

Dice Sensitivity Specificity

The proposed 0.7906 0.9813 0.8664

method
FCM 0.7501 0.9630 0.3317

watershed 0.7626 0.9979 0.6623

Fig. 5 A table showing the obtained results for the first image

Specificity: an evaluating metric that indicates the true negativity, which means that
sensitivity gives us the probability that a segmented pixel doesn’t belong to the brain
tumor area, but belongs to the background.

Specificity = TN/(FP + TN) (10)

TP: indicates the true positivity, which refers to the number of pixels that belong
to the tumor area.
TN: indicates the true negativity, which refers to the number of pixels that don’t
belong to the tumor area.
FP: The false positive, which indicates the number of pixels that were wrongly
detected as brain tumor pixels (Figs. 5 and 6).
878 H. Moussaoui et al.

Fig. 6 The graphic 1.2

representation for the first 1
table 0.8
0.6 Dice Coefficient
0.4 Sensitivity
0.2
Specificity
0
The Fuzzy c- Watershed
proposed means
method

8 Conclusion

In this paper we’ve been talking about our proposed method which is based on fuzzy
c-means clustering and marker controlled watershed algorithm. The results obtained
show satisfactory abnormalities detection. We’ve used several metrics to evaluate
the performance for the proposed method like “Dice coefficient”, “Sensitivity” and
“Specificity”.

References

1. Khrissi L, Elakkad N, Satori H, Satori K (2019) Image segmentation based on kmeans and
genetic algorithms. In: Proceedings of ESAI 2019, embedded systems and artificial intelligence,
pp 489–497 (2019)
2. Zheng X, Lei Q, Yao R, Gong Y, Yin Q (2018) Image segmentation based on adaptive K means
algorithm. EURASIP J Image Video Process 2018:1–10. https://doi.org/10.1186/s13640-018-
0309-3 Article number: 68
3. Zaitoun NM, Aqel MJ (2015) Survey on image segmentation techniques. In: International
conference on communication, management and information technology (ICCMIT 2015).
ScienceDirect
4. Ghosh S, Das N, Das I, Maulik U (2019) Understanding deep learning techniques for image
segmentation. ACM Comput Surv 52(4):1–35. https://doi.org/10.1145/3329784
5. El Akkad N, El Hazzat S, Saaidi A, Satori K (2016) Reconstruction of 3D scenes by camera
self-calibration and using genetic algorithms. 3D Res 6(7):1–17
6. Merras M, Saaidi A, El Akkad N, Satori K (2016) Multi-view 3D reconstruction and modeling
of the unknown 3D scenes using genetic algorithms. Soft Comput 22(19):6271–6289
7. El Akkad N, Merras M, Baataoui A, Saaidi A, Satori K (2017) Camera Self-calibration having
the varying parameters and based on homography of the plane at infinity. Multimed Tools Appl
77(11):14055–14075
8. El Akkad N, Merras M, Saaidi A, Satori K (2014) Camera self-calibration with varying intrinsic
parameters by an unknown three-dimensional scene. Vis Comput 30(5):519–530
9. El Akkad N, Merras M, Baataoui A, Saaidi A, Satori K (2018) Camera self-calibration having
the varying parameters and based on homography of the plane at infinity. Multimed Tools Appl
77(11):14055–14075
10. Akkad NE, Merras M, Saaidi A, Satori K (2013) Robust method for self-calibration of cameras
having the varying intrinsic parameters. J Theor Appl Inf Technol 50(1):57–67
11. Akkad NE, Merras M, Saaidi A, Satori K (2013) Camera self-calibration with varying
parameters from two views. WSEAS Trans Inf Sci Appl 10(11):356–367
A Novel Brain Tumor Detection Approach Based on Fuzzy C-means … 879

12. El Akkad N, Saaidi A, Satori K (2012) Self-calibration based on a circle of the cameras
having the varying intrinsic parameters. In: Proceedings of 2012 international conference on
multimedia computing and systems, ICMCS. pp 161–166
13. Merras M, El Akkad N, Saaidi A, Nazih AG, Satori K (2014) Camera calibration with varying
parameters based on improved genetic algorithm. WSEAS Trans Comput 13:129–137
14. Es-sabry M, El Akkad N, Merras M, Saaidi A, Satori K (2019) A novel color image encryption
approach based on random numbers generation of two matrices and bit-shift operators. Soft
Comput 24(5):3829–3848
15. Es-Sabry M, El Akkad N, Merras M, Saaidi A, Satori K (2018) A novel text encryption
algorithm based on the two-square cipher and caesar cipher. Commun Comput Inf Sci 872:78–
88
16. Es-Sabry M, El Akkad N, Merras M, Saaidi A (2018) Grayscale image encryption using shift
bits operations. In: 2018 international conference on intelligent systems and computer vision,
ISCV, pp 1–7
17. Sreerangappa M, Suresh M, Jayadevappa D (2019) Segmentation of brain tumor and perfor-
mance evaluation using spatial FCM and level set evolution. Open Biomed Eng J 13(1):M6.
https://doi.org/10.2174/1874120701913010134
18. Bahadure NB, Ray AK, Thethi HP (2017) Image analysis for MRI based brain tumor detection
and feature extraction using biologically inspired BWT and SVM. Int J Biomed Imaging
2017:12. Article ID 9749108. https://doi.org/10.1155/2017/9749108
19. Kumar P, Vijayakumar B (2015) Brain tumour Mr image segmentation and classification using
by PCA and RBF kernel based support vector machine. Middle-East J Sci Res 23(9):2106–2116
20. Maolood I, Al-Salhi Y, Lu S (2018) Thresholding for medical image segmentation for cancer
using fuzzy entropy with level set algorithm. Open Med J 13(1):374–383. https://doi.org/10.
1515/med-2018-0056
21. Jobin Christ MC, Parvathi RMS (2011) Fuzzy c-means algorithm for medical image
segmentation. https://doi.org/10.1109/ICECTECH.2011.5941851
22. Zhou H, Schaefer G, Shi C (2009) Fuzzy c-means techniques for medical image segmentation.
In: Jin Y, Wang L (eds) Fuzzy systems in bioinformatics and computational biology. Studies
in fuzziness and soft computing, vol 242. Springer, Heidelberg. https://doi.org/10.1007/978-3-
540-89968-6_13
23. Grau V, Mewes AUJ, Alcaniz M, Kikinis R, Warfield SK (2004) Improved watershed transform
for medical image segmentation using prior information. IEEE Trans Med Imaging 23(4):447–
458. https://doi.org/10.1109/TMI.2004.824224
24. Zhanpeng H, Qi Z, Shizhong J, Guohua C (2016) Medical image segmentation based on the
watersheds and regions merging. In: 2016 3rd international conference on information science
and control engineering (ICISCE), Beijing, pp 1011–1014. https://doi.org/10.1109/ICISCE.
2016.218
25. Lu Y, Jiang Z, Zhou T, Fu S (2019) An improved watershed segmentation algorithm of medical
tumor image. In: IOP conference series: materials science and engineering, vol 677, no 4
26. Ravi S, Khan AM (2014) Bio-medical image segmentation using marker controlled watershed
algorithm: a case study. Int J Res Eng Technol 3:26–30. eISSN: 2319-1163, pISSN: 2321-7308
27. Sandhya G, Kande GB, Savithri S (2017) A novel approach for the detection of tumor in
MR images of the brain and its classification via independent component analysis and kernel
support vector machine. Imaging Med 9(3):33–44 ISSN: 1755-5191
28. Senthilkumaran N, Vaithegi S (2016) Image segmentation by using Thresholding techniques
for medical images. Comput Sci Eng: Int J (CSEIJ) 6(1):1–13
Image Retrieval Based on MPEG-7
Feature Selection Using Meta-heuristic
Algorithms

Naoufal Machhour and M’barek Nasri

Abstract The continuous growth of the internet and its diversified domains and
tools; and the rapid development of new technologies have enormously increased
the amount of digital images and consequently the dimension of the image data
sets. Therefore, finding the relevant images to a query image in such database is a
challenging task. Content-based image retrieval is the powerful method so far which
use the visual characteristics of the image. The choice of these visual descriptors is
a decisive phase in this area. This work presents a method which extracts the three
image features which are: color, texture and shape based on the MPEG-7 standards.
Then, the genetic algorithm is implemented for feature selection to provide the best
extracted features to avoid unnecessary ones and reduce calculations and retrieval
time. Meanwhile, the k-nearest neighbors algorithm is used to search for the relevant
images. Finally, we will apply our method on two different image data-sets which
contain images belonging to different domains to show the efficiency of selecting
the correct features using meta-heuristic algorithms.

Keywords Content-based image retrieval · Feature selection · Genetic algorithm ·

K-NN algorithm · MPEG-7 standards

1 Introduction

The image bases are becoming very wide and diversified. In order to well exploit
these large databases and extract the relevant images, content-based image retrieval
(CBIR) is the powerful method so far. This method which used the visual descriptors
of the image is able to find the closest images to a query image (QI) in wide image
data-set. All CBIR systems follow two steps as represented in Fig. 1. In the offline
phase, the CBIR system extract the visual features of the database images, then store

MATSI Laboratory, ESTO, University of Mohammed Premier, Oujda, Morocco

N. Machhour (B) · M. Nasri

Laboratory MATSI, ESTO, University of Mohammed Premier, Oujda, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 881
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_80
882 N. Machhour and M. Nasri

Fig. 1 CBIR system diagram

them in a new features database. In the next online phase, the CBIR system calculates
the similarity measures between the extracted features of all data-set images and
the QI ones to determine the similar images to the inputted QI. Therefore, color
descriptor which is used widely in CBIR systems provide good results. Yue et al. [1]
demonstrated that color features are introduced in pretty nearly all CBIR systems
due to their advantages, such as color histogram [2] and color moments such as the
mean, variance and third moments. Also, texture is an essential descriptor in CBIR
systems. The gray level co-occurrence matrix (GLCM) which give more details
about the arrangements of colors [3] can provide some important informations such
as: energy, entropy, contrast and homogeneity. Moreover, shape features are used to
describe particular regions in image. Thus, many techniques are developed such as:
edge detection, filters, Fourier transform or Hu moment.
There are many works developing CBIR systems which use a single image
descriptor such as color, as developed in the work of Sharma et al. [4] which
used the conventional color histogram (CCH) and Lei et al. [5] which applied a
quantization method on the HSV color space into 36 non-uniform bins, etc. These
studies give encouraging results but ignore other descriptors which can enhance the
retrieval precision. On another hand, several works merged a lot of features simulta-
neously, these methods can increase the search time and provide inefficient results by
using unnecessary descriptors. The work in hand used an efficient method based on
the feature selection using meta-heuristic algorithms applied to MPEG-7 standards
which increasingly become a general framework for content description. The aim
of this study is to enhance the precision rate and decrease the query time by main-
taining intelligently the relevant features from the MPEG-7 standards to avoid the
unnecessary descriptors which can influence the accuracy of the retrieved relevant
images and increase the response time by performing unnecessary calculations. For
this, the proposed CBIR system extracts the image features based on the MPEG-7
standards. Then applied the genetic algorithms (GAs) for feature selection to retain
the relevant descriptors. Finally, the k-nearest neighbors algorithm (k-NN) is applied
for image retrieval.
Image Retrieval Based on MPEG-7 Feature Selection … 883

2 Related Works

Multiple and various methods have been implemented in this domain. These studies
combined different visual features and used several search techniques. Alsmadi [6]
proposed a CBIR system using a technique based on the neutrosophic clustering algo-
rithm. They extract color descriptor based on the YCbCr using Canny edge histogram
(CEH) and discrete wavelet transform (DWT). The CEH is utilized to describe shape
characteristics and the GLCM to extract texture features. Machhour et al. [7], estab-
lished a method using the color string coding (CSC) technique combined to genetic
algorithm in the first work. Then, in an extended work, the authors in [8] improved
the color string coding technique by grouping each color series separately and devel-
oping a segmentation process based on the HSV color space. The GLCM is used
for extracting the texture characteristics. Meanwhile, they performed a comparison
between the simulated annealing (SA) and GAs in term of precision and response
time for image retrieval. In Ashraf et al. [9], the representation and image retrieval are
based on the artificial neural network (ANN) and the bandelet transform. ElAlami
[10] extracted the texture and color signature using the Gabor filter and the 3-D color
histogram. Then, they used the GAs to provide the optimal boundaries of the numer-
ical domains ranges with discrete values represented by nominal features. Finally,
they extracted the relevant features using 2 consecutive operations: the preliminary
and deeply reduction. In the aim of maximizing the precision and simplifying the
computation of the image search, the study in Lin et al. [11] is based on the feature
selection using the sequential forward selection technique which is developed on 3
characteristics: color co-occurrence matrix, pixels variance of scan pattern and color
histogram for k-mean.

3 Materials and Methods

The efficiency of a CBIR system is linked to several factors such as the features
extracted, retrieval method, response time, etc. As we know, many of these features
that are extracted from the image are irrelevant, which influence the quality of the
retrieved images and increase the calculation and response time. In the proposed
CBIR system as shown in Fig. 2, we implement a technique which is based on the
MPEG-7 standards to extract 6 descriptors representing the color, texture and shape
features of the image. Therefore, our technique is based on two phases. In the first
one which is done offline, we extract the image features which are: the dominant
color descriptor (DCD), color structure descriptor (CSD) and color layout descriptor
(CLD) for color features; texture edge histogram descriptor (EHD) and 2D-discrete
wavelet transform (2D-DWT) for texture features; and the curvature scale space
(CSS) for contour-shape descriptor. Then, we implement the genetic algorithms for
feature selection to avoid the unnecessary descriptors and retain the relevant ones.
884 N. Machhour and M. Nasri

Fig. 2 Proposed CBIR system diagram

In the second phase which is done online, we extract the query image features, and
apply the k-nearest neighbors algorithm (k-NN) to find the relevant images.

3.1 Features Extraction

Color Features Extraction

Dominant Color Descriptor (DCD). This descriptor allows to limit the number of
representative values of color characteristics by regrouping them into a small number
of colors as depicted by Manjunath et al. [12]:

F = {(ci , pi , vi ), s}, (wher e i = 1, . . . , N ) (1)

N , represents the number of dominant color ci ; pi , vi and s are the percentages,

color variances and the optional spatial coherency of ci respectively. Manjunath et al.
[12] proved that eight dominant colors are sufficient to describe an image. The value
of the percentage pi is between 0 and 1. The pi value is the fraction of image elements
which correspond to color ci where:
N
pi = 1 (2)
i=1

The color variance vi depicts the variation of the values of image elements assigned
to the class of the dominant color ci ; while s which represents the spatial coherency
shows the homogeneity of this dominant color ci .
Color Structure Descriptor (CSD). This descriptor is used to represent an image by
the color distribution [13, 14]. According to Messing et al. [14], this feature can be
defined as a generalization of the color histogram which can be represented by the
color structure histogram described by the formula below:

h(m), (wher e m = 1, . . . , M) (3)

Image Retrieval Based on MPEG-7 Feature Selection … 885

Where, M is the quantized color cm (M ∈ {256, 128, 64, 32}), and h(m) is the
bin value representing the number of the structuring elements having pixels with
color cm .
Hence, the CSD is based on the information of the spatial structure of the colors
and their frequency occurrence. Messing et al. [14] proved that the performance of
this descriptor can be enhanced based on the non-uniformly quantization histogram
amplitudes using the nonlinear HMMD color space.
Color Layout Descriptor (CLD). The discrete cosine transform (DCT) is developed
to extract this descriptor to describe the spatial distribution of colors. For this, Manju-
nath et al. [12] specified four steps which are: image partitioning technique, repre-
sentative color detection method, discrete cosine transform, and then, the nonlinear
quantization of the zigzag-scanned coefficients.
Texture Features Extraction
Edge Histogram Descriptor (EHD). This descriptor describes a range of 5 types of
edges in each part of image which is divided into n × n blocks. The 5 types of these
edges developed by Won et al. [15] and shown in Fig. 3 are: vertical (a), horizontal
(b), 45-degree diagonal (c), 135-degree diagonal (d), and non-directional edges (e).
Therefore, every local histogram contains 5 bins which correspond to one of 5
edge types. Won et al. [15] demonstrated that dividing the image into 16 blocks
provide a total of 80 histogram bins. Therefore, to extract these features, the image
is converted to gray level, then dividing this image into sub-blocks. Finally, these
descriptors are provided using the directional and non-directional edge features.
2D-Discrete Wavelet Transform (2D-DWT). This descriptor is very efficient in many
domains including image processing. This technique can decompose an image in
various frequency sub-bands. Also, 2D-DWT depicts the frequency and spatial range
and can be used efficiently for non-stationary signals. The technique is based on
dividing the image into rows and columns by applying the average intensity of the
image (which is illustrated by low pass) and high temporal resolution describing the
edges of image (high pass). Figure 4 illustrate an example of two-level decomposition
of the DWT. This method utilizes filters such as Haar, Daubechies, etc. In this study,
we used the low-pass (L) and the high-pass (H ) filters based on the Haar Wavelet
Descriptor as shown below:

L = [1, 1] (4)

Fig. 3 Directional and non-directional types of edges

886 N. Machhour and M. Nasri

Fig. 4 Two-level decomposition of DWT

H = [−1, 1] (5)

Shape Features Extraction. In this study, the shape features extraction is based on
the MPEG-7 standards. Thus, we utilize the curvature scale space (CSS) for contour-
shape descriptor. This latter is characterized by its robustness against noise, scale and
change in translation and orientation. Firstly, we select N -samples equidistant point
on the contour, next we choose randomly a point and follow the contour clockwise.
Then, we create two series (x) and (y) to group and store separately for each selected
point two coordinates (x and y coordinates) respectively. Consequently, we get an
arc-length parameterization depicted by the formula below where (u) is an arbitrary
parameter:

r (u) = (x(u), y(y)) (6)

Finally, this curve parametric representation is undergoing the convolution process

with the Gaussian function to compute the CSS representation. In order to extract the
curvature zero-crossing points of these curves, Mokhtarian et al. [16] demonstrated
that the standard deviation of the Gaussian varies from small to large value. Below,
the formula for calculating the curvature function as developed by Mokhtarian et al.
[16]:

ẋ(u) ÿ(u) − ẋ(u) ẏ(x)

k(u) = (7)
 3
ẋ 2 (u) + ẏ 2 (u) /2

3.2 Feature Selection

This important machine learning technique which is feature selection is a powerful

and efficient tool for selecting a subset of relevant characteristics and features. The
choice of using this method is based on several reasons: elimination of the unnec-
essary features, simplification and reducing the complexity of the retrieval method,
decreasing the search time by diminishing the calculations, avoiding the problem of
dimensionality and improving the accuracy of relevant images by choosing the right
subset of features. In this area, various techniques and methods were implemented
Image Retrieval Based on MPEG-7 Feature Selection … 887

and classified in 3 categories of algorithms: wrappers, filters and embedded methods.

In the work in hand, we implement the wrappers technique which is based on the
search method using the genetic algorithms.

Genetic Algorithms (GAs). This popular and efficient meta-heuristic algorithm

which is based on a random search for the optimal solution in a large population
was introduced by Holland in 1962 [17]. Therefore, the GA is based on the evolu-
tionary processes using natural operations which are: evaluation, selection, crossover,
mutation and the best survival individual [17]. The steps followed by the GAs are:
– Generate randomly the initial population
– Repeat until termination criteria:
• Evaluate population
• Generate the new population
• Apply the genetic operations (crossover and mutation)
• Get the population offspring.
– Get the best chromosome.

Population and Chromosome Representation. The population which is generated

randomly is represented by the database features which contains vectors where each
one contains 6 MPEG-7 features of each image int the data-set. This population is
composed of a set of chromosomes, each one consists of a number of genes containing
the image features. The GA begins the optimization process by calculating the fitness
value for each chromosome to select the best ones to generate the next population.
Therefore, the fitness function used to evaluate each individual in the population is
the standard deviation (σ ) as illustrated below:

1 N
σ = (xi − μ)2 (8)
N i=1

Where, μ represents the chromosome mean, xi the gene value and N the size of
the chromosome.
Natural Operations of the GAs. The roulette wheel technique is used in the selection
phase to generate randomly an array (Ri) containing values between 0 and 1. Then,
we calculate the selection probability (Pi) for each chromosome to provide the
cumulative probabilities (Qi). The comparison between (Ri) and (Qi) select the
chromosomes to yield the new generation.
In the next phase, the population is exposed to the crossover operation. The GA
chooses 2 chromosomes to exchange their genes between them from a crossing point
chosen randomly. The probability of the crossover operation (Pc) is 0.75, that means
that on average 75% of the population will undergo this operation.
The mutation operation is an essential process, because it makes random changes
in the population. Thus, all individuals in the population have the same chance to
be chosen and the mutation probability is 0.01. Then, the gene in a position chosen
888 N. Machhour and M. Nasri

randomly of the offspring chromosome is replaced by another gene chosen randomly

from all the population.
The application of these genetic operations provides a better fitness value which
is higher than the previous one of the old generations. Therefore, we can observe
that the new best chromosome has a better evaluation than the best one in the old
population. Thus, we store for each population generated the fittest chromosome.
This latter contains the relevant features for image retrieval.

3.3 Proposed Image Retrieval Algorithm

In CBIR systems, many retrieval methods have been developed such as tradi-
tional techniques or new approaches using artificial intelligence techniques or meta-
heuristic algorithms. Machine learning methods have shown their effectiveness in
this type of problems. In this study, we develop the k-nearest neighbors algorithm
for image retrieval. This non-parametric and supervised machine learning algorithm
was introduced in 1967 by Thomas Cover et al. [18] and developed for regression
and classification problems. The k-NN algorithm is among the simplest methods in
machine learning; and it is used massively in several domains because of its ease of
implementation and use and low execution and calculation time. The steps followed
by the k-NN algorithm are:
– Load the database of selected features.
– Initialize the value of k.
– For each image in the data-set.
• Compute the distance value of the current image and the QI.
• Store this distance in the distance data-set.
– Sort the distance data-set.
– Select the first k images from the distance data-set.

4 Results and Discussion

4.1 Image Data-Set and Evaluation

In this work, we apply our method on two different image bases which are: Corel
and Caltech image data-sets. These large image bases contain hundreds of images
belonging to different domains; and allow us to evaluate and test the performance
of the feature selection method and compare the results obtained in this study with
other methods. For this, we calculate two measures: precision (P) and recall (R).
Image Retrieval Based on MPEG-7 Feature Selection … 889

Table 1 The average recall and precision values using Corel image data-set
Classes Africa Food Horses Elephant Flower Beach Building Nature Bus Dinosaur
(P) 0.76 0.76 0.84 0.84 0.88 0.76 0.76 0.76 0.88 0.96
(R) 0.19 0.19 0.21 0.21 0.22 0.19 0.19 0.19 0.22 0.24

Table 2 The average recall and precision sample using Caltech image data-set
Classes Minaret Tick Flamingo Bike Okapi Chandelier Ball Rhino Rooster Panda
(P) 0.92 0.84 0.80 0.80 0.76 0.72 0.96 0.72 0.80 0.80
(R) 0.30 0.42 0.29 0.32 0.48 0.16 0.37 0.30 0.40 0.52

Table 3 The recall and precision rates compared to other works

Methods Yue et al. [1] Lin et al. ElAlami Machhour Ashraf et al. Proposed
[11] [10] et al. [19] [9] method
Precision 0.650 0.722 0.735 0.78 0.82 0.82
Recall 0.129 0.144 0.146 0.187 0.164 0.205

N umber o f Per tinent I mages Found

P= (9)
T otal I mages Retrieved
N umber o f Per tinent I mages Found
R= (10)
N umber o f Per tinent I mages in Data − Set

The average precision and recall computed in this paper are: 0.82 and 0.205 for
Corel image data-set and 0.812 and 0.356 for Caltech image data-set respectively.
Table 1 and Table 2 represent the precision and recall rates computed for Corel and
Caltech data-sets respectively. Furthermore, Table 3 depicts a comparison of the
results found in this study and other CBIR systems using the Corel image data-set.
As evidenced by the results, the proposed method provides good recall and preci-
sion rates which shows the efficiency of feature selection technique based on GAs.
Furthermore, applying several times the GAs for feature selection, we notice that the
EHD descriptor is efficient in images including distinct objects and CSD descriptor
provide pertinent results in images containing principal colors. However, combining
the 3 best selected features in this study which are: CSD for color features, EHD
for texture descriptor and CSS for contour-shape descriptor, provide relevant results
for almost any type of images as shown in results above. Moreover, the use of the
feature selection reduces the number of descriptors which decreases enormously the
calculations, computations of image retrieval and execution time. Furthermore, the
k-NN algorithm is a relevant choice as image retrieval method because it provides
pertinent and satisfactory results in low execution and response time.
890 N. Machhour and M. Nasri

5 Conclusion

In the work in hand, we have developed an efficient CBIR system based on feature
selection technique. Once the QI is entered, the proposed system extracts 6 MPEG-
7 features for color, texture and shape signature. Meanwhile, these descriptors are
subject to natural operations of the genetic algorithm to retain the relevant features
and remove the unnecessary ones. Finally, the k-NN algorithm is applied for image
retrieval. The experiments are done using the Corel and Caltech image data-sets. As
illustrated in the previous section, the results obtained have demonstrated that the
GAs are powerful and intelligent method for selecting the correct and relevant image
features. Furthermore, removing the needless image characteristics has improved
enormously the retrieval time by avoiding additional calculations and increased the
accuracy of retrieved images. Also, the results found by the k-NN algorithm show
the efficiency of this machine learning technique. As depicted by the retrieval results,
especially the average recall and precision rates which are 0.205 and 0.82 respec-
tively, combining the MPEG-7 standards for image features extraction, GAs for
feature selection and the k-NN method for image search provide satisfactory results
which demonstrates and proves the efficiency of the proposed method which also
exceeds other proposed CBIR systems. In the prospective works, other techniques
and optimization methods will be utilized for feature selection; and combined to other
image descriptors while using different retrieval methods to increase the accuracy of
relevant images in content-based image retrieval system.

References

1. Yue J, Li Z, Liu L, Fu Z (2011) Content-based image retrieval using color and texture fused
features. Math Comput Model 54(3–4):1121–1127
2. Swain MJ, Ballard DH (1991) Color indexing. Int J Comput Vis 7(1):11–32
3. Shapiro LG, Stockman GC (2001) Computer vision. Prentice-Hall, Upper Saddle River
4. Sharma NS, Rawat PS, Singh JS (2011) Efficient CBIR using color histogram processing.
Signal Image Process 2(1):94
5. Lei Z, Fuzong L, Bo Z (1999) A CBIR method based on color-spatial feature. In: Proceedings
of IEEE. IEEE Region 10 conference. TENCON 99. Multimedia technology for Asia-Pacific
information infrastructure (Cat No 99CH37030), vol 1. IEEE, pp 166–169
6. Alsmadi MK (2020) Content-based image retrieval using color, shape and texture descriptors
and features. Arab J Sci Eng 45:3317–3330
7. Machhour N, M’Barek N (2020) Content based image retrieval based on color string coding
and genetic algorithm. In: 2020 1st international conference on innovative research in applied
science, engineering and technology (IRASET). IEEE, pp 1–5
8. Machhour N, Nasri M (2020) A novel content-based image retrieval based on a new approach
of color string coding and meta-heuristic algorithms. Int J Adv Trends Comput Sci Eng
9(1.5):128–137
9. Ashraf R, Bashir K, Irtaza A, Mahmood MT (2015) Content based image retrieval using
embedded neural networks with bandletized regions. Entropy. 17(6):3552–3580
10. ElAlami ME (2011) A novel image retrieval model based on the most relevant features. Knowl-
Based Syst 24(1):23–32
Image Retrieval Based on MPEG-7 Feature Selection … 891

11. Lin CH, Chen RT, Chan YK (2009) A smart content-based image retrieval system based on
color and texture feature. Image Vis Comput 27(6):658–665
12. Manjunath BS, Salembier P, Sikora T (2002) Introduction to MPEG-7: multimedia content
description interface. Wiley
13. Manjunath BS, Ohm J-R, Vasudevan VV, Yamada A (2001) Color and texture descriptors.
IEEE Trans circuits Syst Video Technol 11(6):703–715
14. Messing DS, Van Beek P, Errico JH (2001) The mpeg-7 colour structure descriptor: Image
description using colour and local spatial information. In: Proceedings 2001 international
conference on image processing (Cat No 01CH37205), vol 1. IEEE, pp 670–673
15. Won CS, Park DK, Park S (2002) Efficient use of MPEG-7 edge histogram descriptor. ETRI J
24(1):23–30
16. Mokhtarian F, Abbasi S, Kittler J (1996) Robust and efficient shape indexing through curvature
scale space. In: In British machine vision conference. Citeseer
17. Holland JH (1975) Adaptation in natural and artificial systems. University of Michigan Press,
Ann Arbor
18. Cover T, Hart P (1967) Nearest neighbor pattern classification. IEEE Trans Inf Theory 13(1):21–
27
19. Machhour N, Nasri M (2020) New color and texture features coding method combined to the
simulated annealing algorithm for content based image retrieval. In: 2020 fourth international
conference on intelligent computing in data sciences (ICDS). IEEE, pp 1–8
A Powerful and Efficient Method
of Image Segmentation Based
on Random Forest Algorithm

Zahra Faska, Lahbib Khrissi, Khalid Haddouch, and Nabil EL Akkad

Abstract Segmentation is the heart of an automatic image analysis system. There

are several segmentation techniques. The contour segmentation approach aims to
separate regions of different gray levels and relatively homogeneous. The region
segmentation approach consists in grouping the adjacent pixels of the image into
distinct regions. The cooperative approach is used in order to improve the result of
segmentation and the segmentation based on classification methods; in this case,
classes are defined by the maximum sets of related pixels belonging to the same
class. In this work by studying the dual problem, we develop a simple but efficient
cooperative approach between a Random forest classification method and a set of
contour detection methods as Canny, Prewitt and Sobel. Firstly, original image is
initially segmented by hybridization of Canny, Prewitt, and Sobel’s algorithms for
edge detection. Then, we will use the output image obtained by another supervised
classification segmentation process, which is random forest. To compare between
the results obtained by the different methods, we used a several evaluation metrics
such us: entropy, MI, IoU and DSC. The advantage and robustness of the proposed
method is demonstrated by in-depth experimentation on a set of images.

Keywords Segmentation · Random forest · Edge detection · Canny · Prewitt ·

Sobel · Cooperation

1 Introduction

In an image processing system, the most important step is image segmentation. A

good segmentation leads to good image analysis because it is from the segmented

Z. Faska (B) · K. Haddouch · N. EL Akkad

Laboratory of Engineering, Systems and Applications, ENSA, Sidi Mohamed Ben Abdellah
University, Fez, Morocco
e-mail: [email protected]
L. Khrissi
LIIAN, Department of Computer Science, Faculty of Science, Dhar El Mahraz, Sidi Mohamed
Ben Abdellah University, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 893
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_81
894 Z. Faska et al.

image that the measurements are carried out for the extraction of the discrimi-
nating parameters with a view to the classification or interpretation. Therefore, image
segmentation aims to find homogeneous regions and contours. A region is a set of
adjacent pixels, which are grouped into distinct and homogenous regions according
to pre-defined criteria (texture, color, etc.).We can summarize the segmentation task
as follows given an image; the objective of segmentation is to establish a compact
description that is representative of its information content, which is more useful
than all of its points. This involves extracting relevant visual indices (primitives),
sufficiently correlated with the entities that make up the scene from which the image
is taken. Moreover, image segmentation is a vast subject of study. In the last years,
several works dealing with this subject have been published in different fields as:
camera self-calibration [1–5], 3D reconstruction [6, 7] and Cryptography [8, 9],…
etc. We can distinguish the following four different segmentation techniques that
exist in the literature (Fig. 1):
In the contour segmentation approach, we consider that the primitives to be
extracted are the lines of contrasts separating regions of different and relatively
homogeneous gray levels, or else regions of different texture. In practice, it is a
question of recognizing the transition zones and of locating the border between
the regions as well as possible. In the contour-based segmentation approach, there
are three categories of methods are derivative, analytical, and deformable methods
[10, 11].
In the region approach a grouping, the process is repeated until all pixels in the
image are included in regions. This approach, therefore, aims to segment the image
based on the intrinsic properties of the regions. We can distinguish the following
methods: region growing, merge, the division into regions, and split/merge [12, 13].
The segmentation by region-contour cooperation arouses great interest. It
combines the two segmentation techniques (region-based, contour-based), takes the

Segmentation techniques

Contour Region Cooperation Classification

segmentation segmentation region-contour

• Derivative • Region • Sequential • SVM

methods growing cooperation.
• KNN
• Deformable • Merge region • Mutual
• Random
methods cooperation.
• Division into Forest
• Analytical regions. • Cooperation
• … Etc.
methods results
• Split and merge

Fig. 1 Classification of segmentation approaches

A Powerful and Efficient Method of Image Segmentation ... 895

benefits of the complementary nature of region and contour information. Thus, a

segmentation by region-contour cooperation is a mutual aid between these two tech-
niques to obtain a better final segmentation result. This approach can be a solu-
tion to overcome the disadvantages of both techniques. We can cite three forms of
region-contour cooperation: sequential, mutual, results cooperation [14–16].
Segmentation based on classification methods, in this section we consider pixels of
the image as objects to be classified. The techniques of segmentation by classification
aims to provide a partitioning of the image, where each class contains pixels that have
similar gray levels. We can cite supervised methods (KNN [17], SVM [18], Random
Forest [19]), and unsupervised methods (Kmeans [20–22] [23], Fuzzy C -means
(FCM) [24]).
There is no one-classification method that can be applied to any type of image
and that can provide optimal and most natural partitioning. This explains the great
diversity of classification methods that exist in the literature. Previous work often used
simple local features for node testing; such as pixel differences or differences between
the sums of accumulated filter responses on local rectangular regions [25–27].
In contrast, [28, 29] present a dataset containing over 20K images (each pixel has
a semantic tag) with a large and unlimited open vocabulary. Figure 2 shows examples
of this dataset.
This paper focuses on a cooperative segmentation between a segmentation by
Canny, Prewitt, Sobel contour detection models, and a segmentation that treat the
problem of segmentation based on classification methods, which the discriminant
classifier used is Random forest.
In the second section, we present the segmentation procedure by the proposed
method as well as how each of these algorithms works. The third section will focus
on the experimental and interpretations of the results. In the last section, we will
present the conclusion.

(a)

(b)

(c)

Fig. 2 Examples of images in the ADE20K dataset. The a row shows the sample images, b row
shows the annotation of objects, and the c row show the annotation of object parts
896 Z. Faska et al.

2 Proposed Approach

In this section, we will treat our proposed method that is used to combine two
complementary processes. The first process to make the segmentation of the contours
by combining the canny algorithm, Prewitt and Sobel’s filters. The second process
is a supervised classification segmentation by using the results of the first process,
which is Random forest. For this, we worked with the ADE20K Dataset [28, 29].
The flowchart of the method proposed is illustrated as follows (Fig. 3):

2.1 Canny Method

The Canny filter (or Canny detector) [10] is used in image processing for edge
detection. The algorithm was designed by John Canny in 1986 to be optimal according
to three clearly explained criteria:
• Good detection: the operator must detect as precisely as possible all the contours
of an image.
• Good localization: optimization of the precision with which the contour is
detected.
• Uniqueness: the contour should elicit a unique operator response.
These three criteria are expressed in the joint optimization of three functions for
defining the optimal linear filter for the detection of a stair-step under the assumption
of an independent additive noise signal.

Color image

Edge detection with Canny’s algorithm

Combination of methods
Initial contour
segmentation

Edge detection with Prewitt and Sobel’s filters

Contour Extraction

Segmentation by classification using the model

RandomForestClassifier

Image segmented

Fig. 3 Procedure for implementing image segmentation by the proposed method

A Powerful and Efficient Method of Image Segmentation ... 897

If we consider that the filter has an impulse response ψ(x), these functionalities
are written:
∞
 ψ(x)d x
Good detection : =  0 (1)
∞
−∞ ψ 2 (x)d x

  
ψ (0)
Good localization :  =  (2)
∞ 
−∞ ψ 2(x)d x
   0
ψ (0) ψ(x)d x
Uniqueness :  = k −∞ (3)
∞ ∞
−∞ ψ 2 (x)d x −∞ ψ 2 (x)d x

The solution that verifies these three criteria, proposed by Canny is as follows:
x x −x −x
ψ(x) = a1 e σ sin ωx + a2 e σ cos ωx + a3 e σ sin ωx + a4 e σ cos ωx (4)

Where the coefficients ai and ω are determined from the size of the filter. The
parameter is a parameter of great importance that we find in all other filters derived
from the approach of Canny. This scale parameter indicates how far short of two
parallel contours will be merged into one. Canny shows that the derivative of a
Gaussian is a good approximation of its filter.
Starting from other initial conditions than those of Canny, Deriche’s detector
meets exactly the same quality criteria as Canny’s, but which has an infinite impulse.
It could therefore be synthesized in a particularly efficient recursive manner [11].
The Deriche filter has a general expression of the form:
 2
−α|X | 1 − e−α
ψ(x) = −C X e With c = (5)
e−α

2.2 Sobel and Prewitt Operators

Noise, interference between adjacent edges, and precision errors due to the discreet
character of the image definition grid are among the factors that can corrupt the
performance of edge detection in real images. To detect intensity transitions by
numerical differentiation, derivative methods are the most used. These local methods
scan the image with a mask defining the area of interest. At each position, an operator
is applied to detect significant transitions at the level of the chosen discontinuity
attribute.
898 Z. Faska et al.

Numerous contour extraction techniques exist in the literature. They can be clas-
sified into the following categories: Algorithms based on the gradient (or first-
order operators), Algorithms based on the Laplacian (or second-order operators),
Multi-scale algorithms, and Algorithms using sophisticated filters.
In the case of gradient-based algorithms (or first-order operators), the gradient of
an image gives the rates of change in gray level per unit distance in the directions of
the coordinate axes. It is defined as a vector characterized by its amplitude (related to
the amount of local variation in gray levels), and direction (orthogonal to the border,
which passes at the point considered.
We can therefore obtain a perfect knowledge of the image, which is calculated as
follows:
∂ I (x, y) ∂ I (x, y)
∇x = Et ∇ y = (6)
∂x ∂y

Thus, at each point of the image (x, y), we define two partial derivatives, along x
and y. The direction of the gradient vector maximizes the directional derivative, and
its norm is the value of this derivative.
The various existing variants, including the Roberts operator [30], provide
improvements to the discrete estimation of the gradient (derivative) and to the consid-
eration of the orientation of the contour. The Prewitt operator [31] and the very
popular Sobel operator [32] allow local estimation of the norm of the two-dimensional
spatial gradient of a gray-level image. They amplify regions of strong local variations
in intensity corresponding to the contours. These operators consist of a pair of 3 ×
3 convolutional masks. Applying each of the masks separately gives an estimate of
the horizontal and vertical components of the gradient by simple linear filtering to
moderate the effects of noise.

2.3 Random Forest Algorithm

The Random Forest classifier is a powerful machine-learning tool for solving clas-
sification and regression problems. Uses adaptive (boosting) or random (bagging)
strategies. It consists of using a combination or an aggregation of a large number of
models while avoiding overfitting.
Bagging is an ensemble method introduced by Breiman in 1996 [33]. The word
Bagging is the contraction of the words Bootstrap and Aggregating. Bootstrap is
a principle of statistical resampling [34] traditionally used for the estimation of
quantities or statistical properties. The idea of bootstrap is to use multiple sets of
data sampled from all observed data using a random draw with replacement. Thus,
each elementary classifier of the set will be trained on one of the bootstrap samples
so that they are all trained on a different training set. Aggregating these classifiers
allows for a more efficient predictor.
A Powerful and Efficient Method of Image Segmentation ... 899

A Random Forest is a predictor made up of a set of elementary classifiers of the

decision tree type. In the specific cases of CART models (binary trees) [33, 35], for
building trained Random Forest model, two steps of randomness are used:
• Step1: Create randomly from the dataset b samples with replacement Z i , i = 1…,
B. (each sample having n points).
• Step2: Create randomly q attribute among the existing p and we build the CART
tree G i (x) on these attributes.
• Aggregation by average in regression.

1 B
G(x) = G i (x) (7)
B i=1

• Aggregation by vote in classification.

G(x) = Ma jorit y V ote(G 1 (x), G 2 (x), . . . , G B (x) (8)

3 Results and Interpretations

The method we have proposed is a combination of the segmentation of contours by

the Canny, Sobel, and Prewitt’s algorithm and the segmentation by classification. We
demonstrated our approach by experimenting on a set of the existing labeled images
using the Random Forest algorithm. The advantage and robustness of the proposed
method is demonstrated by in-depth experimentation on a set of images. From the
images in Fig. 4, we can see that the combination of the Random Forest algorithm
with other filters leads to a further increase in performance.
The quality and performance of the segmented image are measured with four
metrics as IoU, DSC, Entropy, and MI presented in the expressions (9), (10), (11)
and (12).
• Intersection over Union (IoU): to compare the similarity between two arbitrary
shapes (volumes) A, B ⊆ S e R n

|A ∩ B|
I oU = (9)
|A ∪ B|

The Intersection over Union (IoU) also called Jaccard index, is an evaluation
metric used to measure the accuracy of an object detector on a particular dataset. It
measures the degree of similarity between the target mask and our prediction output.
• The Dice similarity coefficient (DSC), also called the Sorensen - Dice index or
simply the Dice coefficient, is a statistical tool that measures the similarity between
two sets of data. This index has undoubtedly become the most widely used tool in
the validation of image segmentation algorithms, A, B ⊆ S e R n it is expressed
by:
900 Z. Faska et al.

Edge Detection Segmentation by

Original Image using Canny, Sobel, classification using The proposed method
and Prewitt’s Random Forest
algorithms algorithm

Fig. 4 Image segmentation results obtained by the different methods

2|A ∩ B|
DSC = (10)
|A| + |B|

Here, | A | is the number of elements of A. The index can vary from 0 (when A
and B are disjoint) to 1 (when A and B are equal).
• Entropy: was introduced by Shannon (1948), is a mesure of image information
content, which is interpreted as the average uncertainty of information source,
were the higher value of entropy means more detailed information [35].
 L−1
Entr opy = − p f (i)log2 p f (i) (11)
i=0

With p f the ratio of the number of pixels to the value of the gray level over the
total number of pixels.
• Mutual information (MI): is a quantity that measures a relationship between the
original image and image segmented, were the higher value of MI means the
relationship between the images exists.
 PAF (a, f )
MI = PAF (a, f )log (12)
af PA (a)PF ( f )

With PAF (a, f ) the joint histogram of the merged image F and the source image.
In order to show the quality of our approach on a set of images, which takes the
same size of the original image in all methods:
A Powerful and Efficient Method of Image Segmentation ... 901

Table 1 Comparison of three segmentation methods on the living room image

Metrics IoU DSC MI Entropy
Canny, Sobel and Prewitt 0.6016 0.7512 0.2158 5.3829
Random forest 0.7123 0.8320 0.5818 5.9333
Our approach 0.7440 0.8532 0.3617 6.2322

Table 2 Comparison of three segmentation methods on seaside image

Metrics IoU DSC MI Entropy
Canny, Sobel and Prewitt 0.4834 0.6518 0.2253 4.4726
Random forest 0.8188 0.9004 0.5482 6.6097
Our approach 0.8526 0.9204 0.4591 6.6271

Table 3 Comparison of three segmentation methods on the museum image

Metrics IoU DSC MI Entropy
Canny, Sobel and Prewitt 0.6612 0.7960 0.2663 5.8720
Random forest 0.6990 0.8228 0.4592 5.3751
Our approach 0.8970 0.9457 0.6026 5.9905

– The living room image is 256 * 256 in TrueColor, the Seaside Image is 225 * 300
in TrueColor, and the Museum image is 256 * 256 in TrueColor.
The best result of segmentation is where the values of IoU and dice are close to
the value 1, while the values of MI and Entropy are high. The metric values of some
images presented in Fig. 4 are given in the following Tables 1, 2, and 3:
Tables 1, 2, and 3 presented the results of the different methods obtained from
the four measurements IoU, Dice similarity coefficient, MI, and Entropy, where the
value in Italic in each of the rows shows the best segmentation result obtained by the
three methods.
We have found that our proposed segmentation cooperative approach gives the
best segmentation result compared to other methods. From these tables, we notice
that the values of IoU and DSC obtained by our approach are better than the other
techniques, as well as the value of entropy. Therefore, from the comparison between
different methods, the proposed cooperative approach provides an improvement in
the image segmentation results compared to other methods.
902 Z. Faska et al.

4 Conclusion

In this work, we have proposed a cooperative approach between several methods

widely used in the field of image segmentation. To date, many researchers have
developed different segmentation algorithms, and to respond to the limits and inade-
quacies of these proposed methods, the tendency is to make a combination between
two or more methods. In this point, we proposed an approach combining the contours
segmentation by the Canny’s algorithm, Prewitt, and Sobel’s derivatives methods,
and the segmentation by classification with the Random Forest algorithm. We tested
our method on an existing labeled image dataset. From analysis and a comparison of
the results obtained by the different methods, we can say that our proposed approach
gives better segmentation results and leads to a greater increase in performance.

References

1. El Akkad N, Merras M, Baataoui A, Saaidi A, Satori K (2018) Camera self-calibration having

the varying parameters and based on homography of the plane at infinity. Multimed Tools Appl
77(11):14055–14075
2. Akkad NE, Merras M, Saaidi A, Satori K (2013) Robust method for self-calibration of cameras
having the varying intrinsic parameters. J Theor Appl Inf Technol 50(1):57–67
3. Akkad NE, Merras M, Saaidi A, Satori K (2013) Camera self-Calibration with Varying
Parameters from Two views. WSEAS Trans Inf Sci Appl 10(11):356–367
4. El Akkad N, Saaidi A, Satori K (2012) Self-calibration based on a circle of the cameras
having the varying intrinsic parameters. In: Proceedings of 2012 international conference on
multimedia computing and systems, ICMCS, pp 161–166
5. Merras M, El Akkad N, Saaidi A, Nazih AG, Satori K (2014) Camera calibration with varying
parameters based on improved genetic algorithm. WSEAS Trans Comput 13:129–137
6. El Akkad N, El Hazzat S, Saaidi A, Satori K (2016) Reconstruction of 3D scenes by camera
self-calibration and using genetic algorithms. 3D Res 7(1):6
7. Merras M, Saaidi A, El Akkad N, Satori K (2018) Multi-view 3D reconstruction and modeling
of the unknown 3D scenes using genetic algorithms. Soft Comput 22(19):6271–6289
8. Es-Sabry M, El Akkad N, Merras M, Saaidi A, Satori K (2018) Grayscale image encryp-
tion using shift bits operations. In: 2018 International conference on intelligent systems and
computer vision, ISCV, pp 1–7
9. Es-Sabry M, Akkad NE, Merras M, Saaidi A, Satori K (2018) A novel text encryption algorithm
based on the two-square cipher and caesar ciphe. Commun Comput Inf Sci 872:78–88
10. Canny J (1986) A computational approach to edge detection. IEEE Trans Pattern Anal Mach
Intell 8(6):679–698
11. Deriche R (1987) Using Canny’s criteria to derive a recursively implemented optimal edge
detector. Int J Comput Vis 1:167–187
12. Muñoz X, Freixenet J, Cufi X, Marti J (2003) Strategies for image segmentation combiningre-
gion and boundary information. Pattern Recogn Lett 24:375–392
13. Fan J, Yau DKY, Elmagarmid AK, Aref WG (2001) Automatic image segmentation by
integrating color-edge extraction and seeded region growing. IEEE Trans Image Process
10:1454–1466
14. Sebari I (2007) Les approches de segmentation d’image par coopération régioncontour
15. Muñoz X, Cufi X (2000) A new approach to segmentation based on fusing circum-
scribed contours, region growing and clustering. In: IEEE International conference on image
processing, pp 800–803
A Powerful and Efficient Method of Image Segmentation ... 903

16. Saber EA (1998) Integration of color, edge, shape, and texture features for 94 automatic region-
based image annotation and retrieval. J Electron Imaging 7(3):684–700
17. Wazarkar S, Keshavamurthy BN, Hussain A (2018) Region-based segmentation of social
images using soft KNN algorithm
18. Yang HY, Wang XY, Wang QY, Zhang XJ (2012) LS-SVMbased image segmentation usingcolor
and textureinformation. J Vis Commun Image Represent 23:1095–1112
19. Breiman L (2001) Random forests. Mach Learn 45:5–32
20. Dhanachandra N, Manglem K, ChanuImage YJ (2015) Segmentation using K-means clustering
algorithm and subtractive clustering algorithm
21. Khan SS, Ahmad A (2004) Cluster Centre Initialization Algorithm for K-means Cluster. Pattern
Recognit Lett 25:1293–1302
22. Khrissi L, El Akkad N, Satori H, Satori K (2020) Image Segmentation based on k-means and
genetic algorithms. In: Embedded Systems and Artificial Intelligence. Springer, Singapore, pp
489–497
23. Khrissi L, Akkad NE, Satori H, Satori K (2019) Color image segmentation based on hybridiza-
tion between Canny and k-means. In: 2019 7th Mediterranean congress of telecommunications
(CMT). IEEE, pp 1–4
24. Khrissi L, El Akkad N, Satori H, Satori K (2020) Simple and efficient clustering approach based
on cuckoo search algorithm. In: 2020 fourth international conference on intelligent computing
in data sciences (ICDS). IEEE, pp 1–6
25. Shotton J, Winn J, Rother C, Criminisi A (2006) TextonBoost: joint appearance, shape and
context modeling for multi-class object recognition and segmentation. In: Proceedings of
ECCV, pp 1–15
26. Viola P, Jones M (2001) Rapid object detection using a boosted cascade of simple features. In:
Proceedings of CVPR, pp 511–518
27. Winn J, Criminisi A (2006) Object class recognition at a glance. In: Proceedings of CVPR
28. Zhou B, Zhao H, Puig X, Fidler S, Barriuso A, Torralba A (2017) Scene parsing through
ADE20K dataset. In: Computer vision and pattern recognition (CVPR)
29. Roberts LG (1965) Machine Perception of Three Dimensional Solids. In: Tippet JT (ed) Optical
and electrooptical information processing. MIT Press, pp 159–197
30. Prewitt JMS (1970) Object enhancement and extraction. In: Lipkin BS, Rosenfeld A (eds)
Picture processing and psychopictorics. Academic Press
31. Sobel (1990) An isotropic image gradient operator. In: Freeman H (ed) Machine vision for
three-dimensional scenes, pp 376–379. Academic Press
32. Breiman L (1996) Bagging predictors. Mach Learn 24(2):123–140
33. Efron B, Tibshirani RJ (1993) An introduction to the bootstrap. Chapman & Hall, New York
34. Breiman L, Friedman JH, Olshen RA, Stone CJ (1984) Classification and regression trees.
Chapman and Hall, New York
35. Yang L, Guo BL, Ni W (2008) Multimodality medical image fusion based on multiscale
geometric analysis of contourlet transform. Neurocomputing 72:203–211
Digital Technologies in Battle Against
COVID-19
Towards Overlapped Objects
Recognition Using Capsule Network

Merieme Mansouri, Samia Benabdellah Chaouni, and Said Jai Andaloussi

Abstract Many approaches have proposed successful systems to recognize objects

from an image. However, Overlapped objects recognition is still a main challenge for
all kinds of objects. It is difficult for a machine to recognize two or more objects that
partly hide each other. On the other hand, the new artificial neural network capsule
network (CapsNet) has shown promising results in recognizing overlapped digits
from multi-MNIST dataset. However, CapsNet has achieved low performance when
it recognizes an object from an image with complex background. In this paper, we
propose a contribution containing GrabCut to extract foreground from background,
and then use CapsNet to recognize the overlapped foreground. As a result, we will
address the overlapped objects recognition challenge and the complex background
limit of CapsNet.

Keywords Object recognition · Overlapped objects · Capsule network

1 Introduction

Artificial intelligence is making anything artificial not natural like machines have the
intelligence of humans, the ability to understand, think and learn. Machine learning
is a subset of artificial intelligence. Number of definitions of machine learning have
been presented based on everyone’s vision. The most powerful one is Arthur Samuel’s
definition in 1959: “machine learning: field of study that gives computers the ability
to learn without being explicitly programmed”. [1] Learning is a specific human
property, having machines with human properties will make our life much easier.
Machine learning has different types of algorithms. Two most important ones are:
supervised learning and unsupervised learning. Supervised learning is teaching the
computer how to do a task while unsupervised learning learns to do it but itself [2].
Machine learning serves many fields: Data Mining [3], Natural Language Processing

M. Mansouri (B) · S. B. Chaouni · S. Jai Andaloussi

Faculty of Science Ain Chock, Hassan II University, Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 907
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_82
908 M. Mansouri et al.

(NLP) [4], and so on. One of the most powerful fields of AI is Computer Vision; it is
training machines to interpret the visual world like humans using images and videos.
We are interested in object recognition from images, an essential domain in
computer vision that recognizes objects from an image based on several tasks such as:
detection, segmentation and classification. Two types of approaches were developed
to recognize objects: conventional approaches and deep learning based approaches.
Conventional or traditional approaches are any machine learning methods or
approach before the appearance of Deep learning-based approaches. Deep learning
is a subset of machine learning inspired by the construction of the human brain. The
difference between these two concepts is that classical machine learning takes the
input data and parses it to determine its features, then trains the machine learning
with that data, eventually the machine will apply what it has learned to make a deci-
sion. While deep learning has come to ensure end-to-end learning, it means from the
input image to the output, it handles the both feature extraction and decision phases.
Object recognition has a major role in different life domains: security [5], space
exploration [6], medical diagnosis [7] and so on. Some works recognize food to help
people who suffer from obesity and diabetes to control their nutrition, by calculating
for example number of calories of the recognized food to control the taken quantity
by day [8–11] while others try to recognize animals and give all information you
need about them [12, 13]. There are researches in the traffic monitoring environment
to recognize traffic signs [14, 15]. And others to detect and recognize general objects
[16–18]. The common challenge in object recognition is the overlapped objects. All
those mentioned works have shown impressive results in recognizing simple objects
or multiple non-overlapped objects in an image, but it is very challenging to define
two or more objects that partially hide each other. The missing parts of objects will
make it unfamiliar for machines. There are methods that arrive to detect and classify
overlapped objects, those methods may ensure state of the art accuracy but they work
either on one type of object shape [19] or require specific equipment [20].
Among overlapped objects recognition methods, the new artificial neural network
Capsule network (CapsNet) has shown great results on overlapped digits dataset
multi-MNIST [21], a special version of Modified National Institute of Standards
and Technology (MNIST) dataset. CapsNet has improved high accuracy compared
to convolutional neural networks (CNN). It has come to solve some CNN limits,
but nothing is perfect. CapsNet requires more computational resources and it does
not show high performance when it recognizes objects from an image with complex
background. In this paper, we propose the use of capsule networks to extract features
and classify six categories of overlapped fruits. In order to overcome CapsNet limits
that consist in complex background, we propose extracting the foreground (object)
from the background using GrabCut [22] segmentation algorithm. Our contribution
will overcome the challenge of overlapped objects recognition and CapsNet limit.
This paper can be summarized as follows: Sect. 2 contains related works, Sect. 3
describes our proposed system, and conclusion and future works are presented in
Sect. 4.
Towards Overlapped Objects Recognition … 909

2 Related Works

In this section, several existing works have been chosen based on specific comparison
criteria, there are works on general object recognition and others on specific objects
such as food, traffic signs, animals, and so on. The existing works of our problematic:
“overlapped object recognition” are presented while some approaches of capsule
network have been studied as well. The summary of those works are collected and
presented in a synthesis table giving a brief explanation of their limitations according
to our problematic and contribution.

2.1 Different Objects Recognition

Yu et al. [8] propose a classifier to learn incrementally the user’s data and each
user has 300 new types of food. The system used convolutional neural network
(CNN) to extract features and the combination of nearest class mean (NCM) clas-
sifier and nearest neighbor (NN) classifier for classification, trained and tested with
FoodLog dataset achieving with that the state of the art accuracy. Zhou et al. [16]
discuss the detection of the main object in an image contains different objects,
the authors propose region convolutional neural network (RCNN) to recognize the
multi-objects in an image, based on four steps, search selective for detection, CNN
for feature extraction, support vector machine (SVM) for classification and Non-
maximal suppression(NMS) to remove redundant candidate region and find the best
position of object detection. To obtain the main object from all detected objects, they
put a system based on the ratio of object volume and the image, and the rarity of
all the objects to mark the main object. The experiment was trained and tested with
ImageNet2012 and Flickr8k datasets and achieves 90% accuracy. Hayat et al. [17]
constructed a deep learning framework based on CNN to extract feature and recognize
nine general objects, they worked with Caltech-101 dataset for training and testing,
their work contains also a comparison of the proposed system and variant approaches
of Bag of Words (BOW) based on SVM, the results show that the proposed method is
much better with an accuracy of 90.12%. Liu et al. [9] propose an improved CNN to
extract features and recognize food using two datasets UEC-256 and Food-101 and
achieving 76.3 and 77.4% consecutively. Huang et al. [12] present a deep learning
system for recognizing 27 kinds of birds from images. The authors used the Internet
of Birds (IoB) dataset and preprocessed the images by flipping, rotating, translating,
and noise removing, and then feature extraction and recognition have done with
CNN achieving 95.3% accuracy. Vennelakanti et al. [14] worked on traffic signs
detection and recognition. First, they preprocess their images by converting RGB
image to HSV image, then detect the signs using color and shape based detection,
and finally extract features and recognize signs using CNN, for training and testing,
they used Belgium and German data sets achieving 98% accuracy. Sadeq et al. [10]
present the problem of size estimation in food recognition. Authors propose a system
910 M. Mansouri et al.

that employs capturing solid and liquid food items with fixed posture and estimate
the volume with simple geometric information, the system detects the food item
with ellipse detector, and segment it with color based K-means clustering, Otsu
thresholding, and Canny edge Detector, then uses CNN for features extraction and
classification. The used dataset was FooDD reaching 96% for recognition and 4.46%
error for size estimation. Taheri et al. [13] propose a system that classifies 19 classes
of animals using score-level fusion: fusion of two classifiers, one based on features
extracted from CNN and second based on appearance and shape-based features, to
make a decision and give final results, authors used nearest neighbor (NN) classi-
fier, and preprocessed images with resizing, converting from RGB to grayscale and
histogram equalization. For training and testing they employed LHI-Animal-Faces
dataset and achieved 95.31% accuracy. Ege et al. [11] present a review of five size
estimation methods of food; we are going to focus just on the two last new methods.
The first method used the two backside cameras in the iPhone to obtain depth infor-
mation. Authors used the convolutional neural network U-NET for extraction and
segmentation, and UECFood-100 for training, the second method used grains of
boiled rice as a reference object to estimate the real size based on CNN and its own
dataset trained. The approach achieves 10% estimation error.

2.2 Overlapped Objects Recognition

Yukitoh et al. [20] present the difficulty of overlapping objects, the upper object
changes its angle and the lower object is partially hidden which leads to difficulty
in object recognition. To solve that, authors used two RGB-D sensors to recog-
nize different kinds of overlapped bread from own dataset with simple background,
the opposite viewpoints of the sensors were concatenated by iterative closest point
(ICP) algorithm, to recognize the upper object, authors used posture adjustment
and the two sensors and achieve 96.9% accuracy, for lower object, they used the two
sensors according to the degree of hidden, they reach 84.4% accuracy for 50% hidden
object. Changhui et al. [19] focused on overlapped citrus recognition for automatic
harvesting; they used threshold segmentation in HSV space, gift-wrapping convex
hull algorithm and transform method for citrus fruit localization and reconstruction.
They used their own dataset, as a result, the positioning and reconstruction error
rate were 6.51 and 8.84% respectively. Zhou et al. [23] present a novel segmenta-
tion method for overlapped fish images, authors captured images in real time using
camera attached to a computer, then preprocessed image by extracting background
and filtered the image from noise, then segmented overlapped fish applying several
methods: determine overlapped area using a shape factor, detect corners using the
curvature scale space algorithm then extract skeleton Zhang-Suen algorithm to obtain
as a conclusion the intersection points and the segmented overlapped fish. Authors
obtained 90% in accuracy as results. Song et al. [24] propose an improved version of
non-maximum suppression algorithm (NMS) in order to detect overlapped objects.
Authors grouped detected boxes into clusters and selected the center cluster with the
Towards Overlapped Objects Recognition … 911

maximum score. A predefined threshold is used to specify the degree of overlap-

ping, any box exceeding the threshold will be incorporated into the corresponding
cluster finally they applied soft-NMS to remove unclear boxes. Authors combined
this method with two object detection networks Faster R-CNNs and R-FCNs to
demonstrate its effectiveness using MS COCO dataset.

2.3 Object Recognition Using Capsule Networks

Kumar et al. [15] propose a system that detects and recognizes traffic signs with
Capsule Network; they preprocessed their input images by enhancing brightness and
contrast and augmented the dataset by rotation and flipping. They used Germany
traffic signs dataset and achieved 97.6% accuracy. Ahmed et al. [18] present a 3D
capsule network for 3D object recognition, they proposed four 3D capsule network
architecture, containing different numbers of convolution layers, a dynamic routing
algorithm which is used between a primary capsule layer and class layer, squash
function was added and a decoder for object reconstruction. They used ModelNet-
10 and ModelNet-40 dataset which contain general objects 3D models, achieving
91.37 and 89.66% accuracy respectively. In order to reduce computational time and
improve classification, Hoogi et al. [25] propose a novel Capsule Network architec-
ture that contains self-attention mechanism which select the salient regions, and
CapsNet analyze only those regions, the proposed system was applied on three
medical datasets, MNIST and SVHN and achieve 90 94 92% accuracy on liver,
brain, lung datasets respectively and 99.5% on MNIST and improve the baseline
CapsNet accuracy in SVM with 2.4%.

2.4 Comparison Criteria

In our search of related works to our thesis, we chose several approaches, mentioned
in the previous section, based on different comparison criteria; as a result, we built
a table of synthesis as described in Sect. 4.5. The comparison criteria proposed are:
Object Type (OT). The type of the recognized object in the approach.
Detection Method (DM). Methods used for detection of the object from an image.
Segmentation Method (SM). Methods used for segmentation of the object from
an image.
Feature Extraction Method (FEM). The method used to extract features.
Classification Method (CM). The method used for classification.
Overlapped Objects (OvO). The existing or not of detecting and classifying
overlapped objects:
Method Used (MU). The method used to recognize the overlapped objects.
Background (B). Define if the background is complex or simple.
Datasets (D). The used datasets.
912 M. Mansouri et al.

Accuracy/Error (A/E). The average accuracy or test error results of the

experiment in each approach.

2.5 Table of Synthesis

As shown in Table 1, we synthesized and presented the related approaches in the

table of synthesis based on the comparison criteria mentioned in the previous section.
16 selected approaches of different types of object recognition have been analyzed
based on the previous mentioned comparison criteria. As we can see, some used
conventional methods, and most of them decided to work with deep learning to
reduce time and effort from designing features separately [9, 10, 14], others used
the new deep learning network CapsNet [15, 18, 25] except that these contributions
cannot be applied in the context of overlapped objects. However, only four approaches
have worked on the overlapped objects. Yukitoh et al. [20] presented the difficulty
of overlapped objects, but the system requires specific equipment which conducts to
a practical limitation in real-world conditions. In addition, Changhui et al. [19] and
Zhou et al. [23] recognize overlapped objects based on one specific shape or type.
While Song et al. [24] consider each overlapped two objects or more should be from
the same class. In the following section, we will present the proposed approach that
overcomes these limitations.

3 Proposed Approach

To recognize overlapped objects, we propose to segment the object and separate

it from the background based on GrabCut, then extract features and classify them
using capsule networks. In this way, we address the overlapped objects recognition
challenge and CapsNet limit in recognizing objects with complex background. In
Fig. 1, we present the architecture of the proposed system giving an example of the
supposed input and output of the system.

3.1 Used Methods

This paragraph defines the main used methods in our system.

GrabCut. Is a semi-automatic algorithm to extract foreground from background,

User has to draw a bounding box around the desired object, the pixels outside the
box are defined as background and the pixels inside are defined as unknown pixels.
A Gaussian Mixture Model (GMM) is used to estimate the background and the fore-
ground based on color distribution, a graph is generated containing nodes (pixels)
Table 1 Table of synthesis
Ref OT DM SM FEM CM OvO A/E D
MU B
[8] Food --- --- CNN NCM --- --- --- FoodLog
NN
[16] General objects Selective search --- CNN SVM --- --- 90% ImageNet2012
Flickr8k
[17] General objects --- --- CNN CNN --- --- 90.12% Caltch-101
[9] Food Bounding box --- CNN CNN --- --- 76.3% UEC-256
77.4% Food-101
[12] Birds --- --- CNN CNN --- --- 95.37% Internet of
Birds(IoB)
[14] Traffic Sign Color based detection --- CNN CNN --- --- 98% Belgium German
Shape based detection
[10] Food Ellipse detector K-means clustering CNN CNN --- --- 96% for recognition FooDD
Otsu thresholding 4,46% error for size
Towards Overlapped Objects Recognition …

Canny edge estimation

Detector
[13] Animal --- --- KFA NN --- --- 95.31% LHI-Animal-Faces
CNN
[11] Food Edge detection K-means clustering CNN CNN ---- --- 10% error for calorie UECFood-100
Selective search GrabCut estimation
U-Net
[20] Breads --- --- --- --- Two RGB-D sensors Simple *Upper object: 96.9% Own dataset
*Lower object:
84.4%(for half hidden)
[15] Traffic sign --- --- CapsNet CapsNet --- --- 97.6% German traffic sign
[18] General objects --- --- CapsNet CapsNet --- --- 91.37% ModelNet-10
89.66% ModelNet-40
(continued)
913
Table 1 (continued)
914

Ref OT DM SM FEM CM OvO A/E D

MU B
[25] Medical objects --- --- CapsNet CapsNet --- --- 90% Liver, Brain,
Digits Self-attention 94% Lung datasets
mechanism 92% MNIST
99.5% SVHM
[19] Citrus fruit --- HSV space threshold --- --- Gift-wrapping Convex Simple 6.51% Own dataset
hull 8.84%
Distance transform
method
[23] Fish --- --- --- --- Shape factor Simple 90% ---
Curvature scale space
Zhang-Suen
[24] General objects --- --- --- FastR-CNNs NMS Complex --- MS COCO
R-FCNs Soft-NMS
M. Mansouri et al.
Towards Overlapped Objects Recognition … 915

Fig. 1 Proposed system architecture. The images are examples of the supposed input and output
of the system

and edges, the weight of edges is defined based on the probable definition of each
pixel, a graph-cut is run to find a new classification of the pixels, after several itera-
tions, the unknown pixels are classified either background or foreground. Users can
correct results by marking the missing parts for optimization [22].

Capsule Network (CapsNet). Convolutional neural network(CNN) has shown

great results in different types of object recognition [12, 14, 17], however, it still
have some limits, it doesn’t give importance to spatial hierarchical between features
due to Max Pooling layer of CNN which loses those information, the model can be
able to recognize object in different orientation but it needs a lot of training examples
[26]. In order to overcome those limits, Jeffery Hinton et al. [21] proposed Capsule
Networks, It is a new artificial neural network works with capsules instead of scalar
features, the length of vectors represents the probability of existing, and their orien-
tation represent a specific parameter of the object (position, size..), CapsNet uses
dynamic routing between capsules instead of max pooling, it takes low features and
routing it to high features, if there is an agreement, the weight will be augmented and
the high capsule will be activated. Another property for CapsNet is the squashing
function, it is an activation function similar to ReLu function but it is suitable for
vectors. Although Capsule Network has some limits; even it needs much less samples
for training than CNN and has simpler architecture, CapsNet requires more compu-
tational resources and it does not show high performance when it recognizes objects
from images that contain complex background like CIFAR dataset [26].

3.2 Used Dataset

Our main dataset is fruit-360 dataset [27], it is a vegetables and fruits dataset
containing 120 categories, 61488 images for training and 20622 for testing of
simple objects with simple background. We chose just six categories as a start:
apple, avocado, banana, orange, lemon and strawberry. In order to avoid overfit-
ting in our network where the model learns the training set too well so it fails in
predicting other data of the same object, we expanded the principal dataset with a
916 M. Mansouri et al.

Fig. 2 Extraction foreground from complex background using GrabCut

maid dataset composed of collected images from the internet. Moreover, the second
dataset contains overlapped objects to enrich the training phase. We applied GrabCut
on the second dataset to extract only the foreground and put it on a simple back-
ground as shown in Fig. 2. In order to train CapsNet we combined images from the
two datasets, each training class contains 100 images, and each testing class contains
25 images. All images have been resized to 100 × 100.

3.3 Segmentation and Recognition

In the training phase, we preprocess our dataset by extracting foreground from back-
ground using GrabCut, and then feed our CapsNet for training. The CapsNet model
contains three layers, and a decoder, first layer is a convolution layer with ReLu
activation function, second layer is a primary layer contain a squash and reshape
function, the feature maps from convolution layer output will be reshaped to vectors,
and then normalized their length between 0 and 1 with squash function. Third layer
is capsule layer, here we use routing between capsules algorithm, the capsules in
primary layer will predict the output of the capsule layer, if there is an agreement,
the capsule predicted will be activated, and the capsule weight will be augmented,
this stage will be repeated after a number of iterations. The decoder is based on three
fully connected layers with two ReLu activation functions and one sigmoid function
to reconstruct the input segmented object. Figure 3 represents the used architecture

Fig. 3 Capsule network architecture

Towards Overlapped Objects Recognition … 917

of Capsule Network. To test a new image, we extract foreground from background

using GrabCut, then we use CapsNet to predict the objects classes.

4 Conclusion and Future Works

In this paper, we present a contribution in order to overcome the overlapped object

recognition challenge and handle the complex background limit of capsule networks.
Our proposed approach uses GrabCut to separate objects from the background, and
then feed it to the capsule network in purpose to extract features and give the predicted
class for each object in an image. Our next work consists of continuing the realiza-
tion phase and recording the results obtained using the machine learning platform
Tensorflow-GPU and running on top of it the neural network interface Keras written
in python, GrabCut has applied with the help of OpenCV. In addition we will perform
a comparison with the most common object recognition networks. We intend to apply
this contribution on overlapped food items on plates and extract nutrition informa-
tion such as number of calories in order to help people who suffer from obesity and
diabetes to control their nutrition.

References

1. Samuel A (1959) Some studies in machine learning using the game of checkers. IBM J Res
Dev 3(3):210–229
2. COURSERA Homepage https://www.coursera.org/learn/machine-learning/home/welcome
3. Zhao L, Kong W, Wang C (2020) Electricity corpus construction based on data mining and
machine learning algorithm. In: 2020 IEEE 5th Information Technology and Mechatronics
Engineering Conference (ITOEC), pp 1478–1481. IEEE
4. Lansley M, Kapetanakis S, Polatidis N (2020) SEADer++ v2: detecting social engineering
attacks using natural language processing and machine learning. In: 2020 International
Conference on Innovations in Intelligent Systems and Applications (INISTA), pp 1–6. IEEE
5. Liu B, Wu M, Tao M, Wang Q, He L, Shen G, Yan J (2020) Video content analysis for
compliance audit in finance and security industry. IEEE Access 8:117888–117899
6. Yang X, Nan X, Song B (2020) D2N4: a discriminative deep nearest neighbor neural network
for few-shot space target recognition. IEEE Trans Geosci Remote Sens 58(5):3667–3676
7. Szymkowski M, Saeed E, Omieljanowicz M, Omieljanowicz A, Saeed K, Mariak Z (2020) A
novelty approach to retina diagnosing using biometric techniques with SVM and clustering
algorithms. IEEE Access 8:125849–125862
8. Yu Q, Anzawa M, Amano S, Ogawa M, Aizawa K (2018) Food image recognition by person-
alized classifier. In: 25th International Conference on Image Processing (ICIP), pp 171–175.
IEEE, Greece (2018)
9. Liu C, Cao Y, Luo Y, Chen G, Vokkarane V, Ma Y (2016) Deepfood: deep learning-based food
image recognition for computer-aided dietary assessment. In: 14th International Conference
on Smart Homes and Health Telematics, pp 37–48. Springer, Cham (2016)
10. Sadeq N, Rahat FR, Rahman A, Ahamed SI, Hasan MK (2018) Smartphone-based calorie
estimation from food image using distance information. In 5th International Conference on
Networking, Systems and Security (NSysS), pp 1–8. IEEE, Bangladesh
918 M. Mansouri et al.

11. Ege T, Ando Y, Tanno R, Shimoda W, Yanai K (2019) Image-based estimation of real food size
for accurate food calorie estimation. In: Conference on Multimedia Information Processing
and Retrieval (MIPR), pp 274–279. IEEE, USA
12. Huang YP, Basanta H (2019) Bird image retrieval and recognition using a deep learning
platform. IEEE Access
13. Taheri S, Toygar Ö (2018) Animal classification using facial images with score-level fusion.
IET Comput Vision 12(5):679–685
14. Vennelakanti A, Shreya S, Rajendran R, Sarkar D, Muddegowda D, Hanagal P (2019) Traffic
sign detection and recognition using a CNN ensemble. In: International Conference on
Consumer Electronics (ICCE), pp 1–4. IEEE, USA
15. Kumar AD (2018) Novel deep learning model for traffic sign detection using capsule networks.
arXiv preprint arXiv:1805.04424
16. Yu L, Chen X, Zhou S (2018) Research of image main objects detection algorithm based on
deep learning. In: 3rd International Conference on Image, Vision and Computing (ICIVC) pp
70–75. IEEE, China
17. Hayat S, Kun S, Tengtao Z, Yu Y, Tu T, Du Y (2018) A deep learning framework using convo-
lutional neural network for multi-class object recognition. In: 3rd International Conference on
Image, Vision and Computing (ICIVC), pp 194–198. IEEE, China
18. Ahmad A, Kakillioglu B, Velipasalar S (2018) 3D capsule networks for object classification
from 3D model data. In: 5fnd Asilomar Conference on Signals, Systems, and Computers, pp
2225–2229. IEEE, USA
19. Changhui Y, Youcheng H, Lin H, Sa L, Yanping L (2017) Overlapped fruit recognition for
citrus harvesting robot in natural scenes. In: 2nd International Conference on Robotics and
Automation Engineering (ICRAE), pp 398–402. IEEE, China
20. Yukitoh M, Oka T, Morimoto M (2017) Recognition of overlapped objects using RGB-D sensor.
In: 6th International Conference on Informatics, Electronics and Vision & 2017 7th Interna-
tional Symposium in Computational Medical and Health Technology (ICIEV-ISCMHT), pp
1–4. IEEE, Japan
21. Sabour S, Frosst N, Hinton GE (2017) Dynamic routing between capsules. In: Advances in
Neural Information Processing Systems, pp 3856–3866
22. Rother C, Kolmogorov V, Blake A (2004) Grabcut: interactive foreground extraction using
iterated graph cuts. ACM Trans Graph (TOG) 23(3):309–314
23. Zhou C, Lin K, Xu D, Liu J, Zhang S, Sun C, Yang X (2019) Method for segmentation of
overlapping fish images in aquaculture. Int J Agric Biol Eng 12(6):135–142
24. Song Y, Li X, Gao L (2020) Improved non-maximum suppression for detecting overlapping
objects. In: Twelfth International Conference on Machine Vision (ICMV 2019), v ol. 11433,
p. 114330F). International Society for Optics and Photonics (2020).
25. Hoogi A, Wilcox B, Gupta Y, Rubin DL (2019) Self-Attention Capsule Networks for Image
Classification. arXiv preprint arXiv:1904.12483
26. Mukhometzianov R, Carrillo J (2018) CapsNet comparative performance evaluation for image
classification. arXiv preprint arXiv:1805.11195
27. Mureşan H, Oltean M (2018) Fruit recognition from images using deep learning. Acta
Universitatis Sapientiae, Informatica 10(1):26–42
The Spread of the Corona Virus Disease
(Covid-19) and the Launch of 5G
Technology in China: What Relationship

Abdelhakim Moutaouakil, Younes Jabrane, Abdelati Reha,

and Abdelaziz Koumina

Abstract In this paper, we highlight the technology of the fifth generation of mobile
communications (5G), its impact on human health and its association with the emer-
gence of the Corona virus in China. We have found through this study, that this
technology uses frequencies below 6 GHz used also in other previous generations of
mobile networks and they do not lead to the emergence of dangerous diseases like
cancer or the corona virus, like what some people claim, and it turned out that this
virus comes from the pangolin, which transmitted the infection to humans in Wuhan,
China, in late 2019.

Keywords 5G · Millimeter waves · Antenna · COVID-19

1 Introduction

The last 20 years have witnessed important changes in mobile phone networks with
a profound transformation from the second generation (2G) (which enables us to
communicate with voice), to the third generation (3G) (which was distinguished in
addition to voice communication data transmission), and then to the fourth generation
(4G) (Which has been known to provide and transfer large-scale communications)
[1]. In fact, from a phone initially designed to conduct a voice conversation between
two users without providing any service other than sending/receiving SMS, today’s
modern smartphone has become a true mobile data center that provides access to
several services and applications (browser, camera, internet and games). This gener-
alization of smartphone internet access and video call access are driving a higher

A. Moutaouakil (B) · Y. Jabrane

Modelisation of Complex Systems Laboratory, Cadi Ayyad University, Marrakech, Morocco
A. Reha
Laboratory of Innovation in Management and Engineering for the Enterprise, ISGA,
Marrakech, Morocco
A. Koumina
Nanostructures Physics Laboratory, Cadi Ayyad University Marrakech, Marrakech, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 919
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_83
920 A. Moutaouakil et al.

data rate. That is why the mobile phone industry has chosen to develop the fifth-
generation network (5G) to provide a technical solution to the problem facing 4G
today [2, 3].
But as soon as China announced the launch of this technique, it received a series
of accusations, according to which it would have caused a group of diseases, such
as cancer, heart diseases, the most recent of which is the Corona virus, in reason
for using this technique for strong millimeter wave. For some, it has developed into
burning 5G towers and assaulting telecom worker [4].
Today, we are witnessing a wide spread of the Corona virus (Covid-19), which has
caused the death of one million of people around the world, young and old, and has
made countries confused and Sacrificing their economy by closing the borders in the
absence of a vaccine so far. The disease first appeared in November 2019 in Wuhan,
central China and It is the city where 5G was launched in the same month. The World
Health Organization (WHO) declared this epidemic a public health emergency of
international concern on January 30, 2020 [5]. The thing that made everyone wonder
about the source of this epidemic, how it spreads, how to prevent it and how to treat
it, and whether it has anything to do with the launch of 5G technology.
In this paper, we will study the fifth generation (5G) and how it works and what
damage it does, especially for human health, and then compare it with the symptoms
caused by the Corona virus and determine if there is a connection between them or
not.

2 5G Technology

In telecommunications, 5G is the fifth generation of mobile communications, since

it will allow subscribers to take advantage of ultra-high speed, while limiting the
energy consumption of smart phones. Its maximum speed should be 1 Gbit/s for
downloads, and 500 Mbit/s for uploads. Today, the 5G network is still under test,
even if it is used during military operations [6].
The 5G will provide wireless connectivity for a wide range of new applications
as illustrated in Fig. 1 such as: Smart Home Systems, road safety and autonomous

Fig. 1 Targeted applications

for the fifth-generation
network
The Spread of the Corona Virus Disease (Covid-19) … 921

Fig. 2 5G Spectrum

car, E health, Smart Cities, industrial process control, entertainment, and of course
the Internet of Things (IoT).
In order to provide and support the ever-increasing traffic and ever-increasing
demand for bandwidth, 5G will need to expand its frequency range beyond the bands
normally used in mobile communications. This includes both the frequency bands
below 6 GHz and the millimeter frequencies.
As shown in the following figure, the lower part of this frequency range, below
30 GHz, is preferred from the point of view of propagation properties. In addition,
it is important to understand that high frequencies, especially above 10 GHz, can
serve as a complement to the lower frequency bands, and will be primarily used to
provide additional system capacity and very wide band transmission for higher data
rates. Millimeter frequencies can be used for indoor applications to provide very
high-speed connections [7].
The two basic types are referred to as “sub-6 GHz” (which just uses frequencies
below 6 GHz) and “millimeter waves”, which are 24 GHz and higher (this name
comes from the fact that the lengths of individual radio waves can be measured in
millimeters). According to [8], the only types of 5G radio signals currently used in
5G networks in China are the less than 6 GHz variants].
To ensure, at millimeter frequencies or the frequencies sub-6 GHz, a high-speed
connection and the ubiquity of the urban network, 5G will use "massive MIMO"
(Multiple Input Multiple Output) technology using multi-antenna networks [9].
The design of a new antenna for 5G wireless devices is more demanding at MMW
frequencies where the flexibility of the system architecture, efficiency, reliability
and compatibility with MIMO systems are the main considerations for achieving
efficiency. High spectral frequency, as well as reduced multipath fading. The most
important requirements for 5G MIMO antenna are greater bandwidth to allow simul-
taneous operation of many system services, high gain to cope with high frequency
atmospheric attenuation and absorption.
The following figure shows examples of antennas used for 5G.
The spread of this kind of antennas elicited a cry from many scientists. Over 200
scientists and physicians demanded a 5G moratorium. These scientists highlight in
particular the risks of cancer, genetic damage and neurological disturbances due to
the use of this type of antenna which sends and receives millimeter waves, but the
922 A. Moutaouakil et al.

Fig. 3 Examples of 5G
antennas

Fig. 4 Corona virus

(Covid-19) [11]

truth is that the energy associated with these waves is too weak to cause cell damage
or break the weak energy bonds that maintain molecular clusters among living things.
As well that we have been using wireless signals less than 6 GHz around the
world for years, which China is currently using for fifth generation as we have said
in a large number of applications. For example, all existing 4G cellular networks use
signals in this range, as do WI-Fi (operating at 2.4 GHz and 5 GHz) and a home
microwave oven [10].
So, if these signals could cause a kind of metabolic change in animals or people,
as some say, then we all have had their effect long ago. It did not happen because due
to the signal strength used by all of these systems, there was no measurable negative
connection between these signals and our health, and there was no way that could
lead to something that would have created the corona virus.

3 Covid-19

In late December of 2019, some similar symptoms appeared in many people in the
Chinese city of Wuhan, which made doctors believe that it was a virus that infected
the city and research began to show that later we are really facing a global epidemic
that is spreading rapidly as the World Health Organization said.
Covid-19 is an infectious disease caused by a strain of the corona virus, called
SARS-CoV-22, contagious by human-to-human transmission via respiratory droplets
or by touching contaminated surfaces and then touching one’s face. The most
common symptoms are fever, cough and difficulty breathing, but they can cause
acute respiratory distress which can lead to death [11].
The Spread of the Corona Virus Disease (Covid-19) … 923

Covid-19 can be identified by this ring of small pointed proteins - called spicules
- which surrounds it. Once the viral particles enter the mouth, nose or eyes, they will
move to the back of the nasal passages, but also to the mucous membranes located
at the back of the throat. Thanks to their small spikes, the particles will cling to the
cell membranes of the respiratory tract, which produce a protein called ACE2. Once
hooked, they will release genetic material similar to DNA, called RNA, which will
enter the human cell. This genetic material will then "divert" cells from their usual
functions to give them a new role: to create and release millions of copies of the
virus.
The investigation into the animal origins of SARS-CoV 2, the etiological agent
of Covid-19, is still current. Bats, notably the species Rhinolophus affinis, and
the Pangolin are the only two animal species harboring corona viruses very close
to SARS-CoV-2. While the bat is the most likely source of the virus, although no
formal evidence exists, the role of Pangolin in the chain of transmission remains to
be determined [12].
A study, the publication of which has been accelerated in Nature, analyzes the
composition of the firm in several samples of pangolins seized during anti-poaching
operations. Despite its protected species status, the Pangolin is the mammal that
suffers the most from illegal trade. Its flesh is consumed in Asia and its scales are
used in Chinese medicine. Genetic analysis of the Pangolin samples identified six
strains of corona viruses, all of which belong to the same phylo-genetic group as
SARS-CoV-2, beta-corona viruses. At the genomic level, the genes of the six strains
of penguins are organized in the same way as those of SARS-CoV-2 [13, 14].
Scientists are therefore now certain of the appearance of Covid-19, a process
which has absolutely nothing to do with the installation of antennas in 5G phase
because the epidemic comes from an animal.

4 Conclusions

The world now knows a host of challenges, the most important of which is how to
tackle this epidemic, which is spreading wildly and killing people. As health concerns
become so frightening, it is easy to fall into the trap of thinking that everything is a
potential threat. Fortunately, the scientific development that the world knows answers
us about these fears, thanks to which we have learned that the source of this epidemic
is some animals, and there is no connection to launching 5G in its spread, and we
have also confirmed that this technology does not cause any diseases currently and
in particular they only use waves which frequency is below 6 GHz used by 4G and
other recent applications.
924 A. Moutaouakil et al.

References

1. Rappaport TS, Sun S, Mayzus R et al (2013) Millimeter wave mobile communications for 5G
cellular: It will work. IEEE Access 1:335–349
2. Khan F, Pi Z (2011) An introduction to millimeter wave mobile broadband systems. IEEE
Commun Mag 49:101–107
3. Haraz OM, Ali MMM, Alshebeili S, Sebak AR (2015) Design of a 28/38 GHz dual band
printed slot antenna for the future 5G Mobile communication networks. Paper presented at:
IEEE International Symposium on Antenna and Propagation & USNC/URSI National Radio
Science Meeting 10:1532–1533
4. Wang C-X, Haider F, Gao X et al (2015) Cellular architecture and key technologies for 5G
wireless communication networks. IEEE Commun Mag 52: 122–130
5. WHO: Statement on the second meeting of the International Health Regulations 2005
Emergency Committee regarding the outbreak of novel coronavirus (2019-nCoV) (2020)
6. Wen-Shyang C, Chun-Kun W, Kin-Lu W (1999) Square-ring microstrip antenna with a cross
strap for compact circular polarization operation. IEEE Trans Antennas Propag, vol 47, no 10,
pp 1566–1568
7. Cheng X, Tang C, Zhang Z (2019) Accurate channel estimation for millimeter-Wave MIMO
systems. IEEE Trans Veh Technol 68:5159–5163
8. Seker C, Güneser MT, Ozturk T (2018) A review of millimeter wave communication for 5G. In:
Proceedings of the 2nd International Symposium on Multidisciplinary Studies and Innovative
Technologies (ISMSIT), Ankara, Turkey, 19–21 October
9. Li B, Fei Z, Zhang Y (2019) UAV Communications for 5G and beyond: Recent advances and
future trends. IEEE Internet Things J 6:2241–2263
10. Huang J, Wang C-X, Feng R, Sun J, Zhang W, Yang Y (2017) Multi-frequency mmWave
massive MIMO channel measurements and characterization for 5G wireless communication
systems. IEEE J Sel Areas Commun 35:1591–1605
11. Li Q, Guan X, Wu P et al (2020) Early Transmission dynamics in Wuhan, China, of novel
corona virus infected pneumonia N Engl J Med
12. Zhu N, Zhang D, Wang W et al (2020) A novel coronavirus from patients with pneumonia in
China, 2019 N Engl J Med, 382 2020, pp 727–733
13. Lai C-C, Shih T-P, Ko W-C, Tang HJ, Hsueh PR (2020) Severe acute respiratory syndrome
coronavirus2 (SARS-CoV-2) and corona virus disease-2019 (COVID-19): the epidemic and
the challenges International Journal of Antimicrobial Agents, 105924
14. Parry J (2020) China coronavirus: cases surge as official admits human to human transmission.
British Medical Journal Publishing Group
Analysis of Covid-19 Impact on Gender
(Male and Female) in South Asian
Country Pakistan

Saba Malik , Mubbra Habib, Mehmood Ahmed Husnain Hashmi,

Muhammad Tariq Saeed, Anwaar ul Huda, Syeda Anam Zahra,
Muhammad Nisar , Muhammad Sajid, Shahbaz Ahmad Shahzad,
Saddam Hussain , and Muhammad Sakandar Majid

Abstract The impact of Coronavirus (COVID-19) disease has been reported for
different countries. Males are at high risk of death due to COVID-19 as compared
to females. There is no data available for Pakistan, the South Asian region. We are
therefore undertaking this analysis to determine the gender effect on the outcomes of
COVID-19 in the South Asian nation of Pakistan. An online survey was performed
across Pakistan (Punjab, Sindh, Khyber Pakhtunkhwa, Balochistan) and national
figures from the Pakistani government. In total, N = 113 individuals were included
(92 males and 21 females). The descriptive statistics, association tests, and Pie chart
indicate that males are more impacted by COVID-19 as compared to females. The
doctor’s response indicates that 81.4% of males and 18.6% females are COVID-19
effective in Pakistan. Information from Pakistani national government statistics indi-
cates that the number of male cases are 78%. Therefore, from the entire studies, we
can infer that COVID-19 attacks more on males rather than females in the South Asian
nation Pakistan. However, more quantitative studies are recommended to measure

S. Malik (B) · M. Habib · M. A. H. Hashmi · M. S. Majid (B)

Department of Zoology, Wildlife, and Fisheries, University of Agriculture
Faisalabad, Faisalabad, Pakistan
M. T. Saeed · S. A. Zahra
Department of Diet and Nutritional Sciences, University of Lahore Islamabad Campus,
Islamabad, Pakistan
A. ul Huda
College of Pharmacy, University of Sargodha, Sargodha, Pakistan
M. Nisar
Pakistan Studies Islamia College, Peshawar, Pakistan
M. Sajid
Department of Psychology, Institute of Southern Punjab Multan, Multan, Pakistan
S. A. Shahzad
Department of International Relation, Iqra University, Islamabad, Pakistan
S. Hussain
Department of Irrigation and Drainage, University of Agriculture Faisalabad, Faisalabad, Pakistan

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 925
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_84
926 S. Malik et al.

the extent to which sex would result in COVID-19 outcomes among Pakistan’s South
Asian countries.

Keywords COVID-19 cases · Global gender deference · Sex-disaggregated data ·

Gender

1 Introduction

In December 2019, the source of severe pandemic of respiratory disease in Wuhan

City, China, was confirmed to be a modern bacterial β-coronavirus, now recognized as
Severe Acute Respiratory Syndrome Coronavirus (SARS-CoV-2) [1]. SARS-COV-
2 is responsible for the most significant lethal complication and is also known as
the 2019 coronavirus (COVID-19). Since its emergence, SARS-CoV-2 has been
extended in 196 countries and yet was declared a pandemic in the World Health
Organization (WHO) by 11 March 2020 [2, 3], respectively. This contributed to
approximately 2 million reported illnesses and approximately 130,000 fatalities in
global cases, including 2/3 in Europe (as of 15 April 2020).
There are currently no recorded SARS-CoV-2 antiviral drugs, although several
therapeutic substances such as redeliver, lopinavir-ritonavir, an IL-6 receptor hybrid
protease inhibitor, colchicine, and tocilizumab are known to be investigated [3, 4].
The fatality of COVID-19 is more than the global level of seasonal flu by 3.4%
worldwide. The state of acute respiratory distress (ARDS), coagulopathy, septic
shock, and metabolic acidosis is a consequence of the occurrence of COVID-19. In
7–17% of hospital patients, COVID-19 cardiac complications included arrhythmia,
acute heart illness, and shockwaves [5].
The estimated death rate in Italy was 7.2% [6], in South Korea [3] it was 0.9%
and in China it was 2.3%. Case pregnancy is higher in individuals > 80 years of
age (14.8% in China, 20.2% in Italy) and cancer. Moreover, all comorbidities, the
most prevalent heart condition in the aged was this high-risk population, which has
a fatality incidence of 10.5%, correlated to negatives [8].
The first findings from China revealed a gender gap concerning observed cases and
the COVID-19 patient fatality rate [7, 8]. Up till now, however, only a small number
of publications discussed gender differences in COVID-19 occurrence and course of
the disease and a detailed study of the root factors is incomplete now [9, 10]. There
are a total of 260,000 confirmed positive cases in the country with 184,000 rescued
and 5,475 deaths up till Friday, July 17, 2020, according to the Ministry of Health,
Government of Pakistan. The largest cases were reported in the province of Sindh
(109,000), followed by Punjab (88,539), Khyber Pakhtunkhwa (31,217), Balochistan
(11,322), Gilgit Baltistan (1775), Federal (82), and Azad Jammu & Kashmir (1,808)
confirmed cases [11]. Outcomes are seen in the Fig. 1.
To date, the largest number of deaths in Punjab has been at 2,059, followed
by Sindh (1,922) KPK (1,124), Islamabad (157) Balochistan (128), AJK (46), and
Gilgit Baltistan (39). In the province of Sindh, a total of (74,076) infected persons
Analysis of Covid-19 Impact on Gender … 927

Fig. 1 COVID-19 incident number in all provinces of Pakistan [11]

were recovered, followed by Punjab (64,815), KPK (22,613), Islamabad (11,599),

Balochistan (8,161), and AJK (1,077) recovered cases to date as summarized in Table
1 [11].
Statistics from several countries throughout the globe propose males and females
will receive equal COVID-19 while males would be at high risk of severe illness &
death [12]. Male had worse results than females in previous coronavirus results or
serious respiratory syndromes (SARS) [13] (Fig. 2).
COVID-19 Case Fatality Rate (CFR) for males and females across 38 regions
with sex-disaggregated data on COVID-19 deaths are shown in Fig. 3. The CFR was
calculated by the following formula:
[33]
The male CFR is higher than the female CFR in 37 of the 38 nations, with the
average male CFR 1.7 times higher than the average female CFR (P < 0.0001,
Wilcoxon signed the ranking test) [21] Denmark and Greece.

Table 1 COVID-19 case all over the South Asian Country Pakistan as per July 17, 2020 [11]
Sr. No Province Confirm cases Active cases Deaths Recoveries
1 Sindh 1099,000 34,070 1,922 74,076
2 Punjab 88,539 22,149 2,059 64,815
3 KPK 31,217 7,749 1,124 22,613
4 Balochistan 11,322 3,096 128 8,161
5 Gilgit Baltistan 1,775 340 391 396
6 AJK 1,808 685 46 1,077
928 S. Malik et al.

Fig. 2 Sex-disaggregated data of confirmed COVID-19 across the whole world [14]

Fig. 3 Comparative analyses of COVID-19 impact on gender whole world [21]

In older people, the proportion of Male/Females mortality proportion continues

to raise it is generally over two in people 60 years of age and beyond, and more than
three in Italy for males and females aged 70–79 [14]. A lot of work has been done on
COVID-19 gender influence; they show that males have more cases of COVID-19
than females. In China, Italy, Germany Switzerland, and Spain, France, the recorded
case fatality rates are relatively homogeneous and vary from 1.7 to 1.8. It supports
Analysis of Covid-19 Impact on Gender … 929

the belief that there is a strong biological pattern, responsible for the higher rate of
deaths in humans, irrespective of country-specific demographics and approaches to
science (Fig. 3) [12, 16–18]. The COVID-19 ratio of males and females in the table
below indicates the gender equality trend.
In Fig. 4 hospitalizations, Intensive Care Unit (ICU) deaths, and CFR are listed
in multiple nations. However exact numbers are given. Throughout all indicators,
sexually disaggregated data was not accessible [18] (Fig. 5).

Fig. 4 Gender difference of COVID-19 in Europe (Italy, Germany Switzerland, Spain, France) and
East Asia (China)

Fig. 5 The male vs female data varies significantly by East Asian countries (China & North Korea),
European countries (Italy, Spain, Germany, Sweden), Oceania countries (Australia), Middle East
country (Iran), and North American country (Canada) [14]
930 S. Malik et al.

Fig. 6 The sex-disaggregated data at global level

During past coronavirus epidemics, the outcome of the outbreak was worse
for men than for women: men had worse outcomes with Serious Acute Respira-
tory Syndrome (SARS) and a higher risk of dying from Middle East Respiratory
Syndrome (MERS) [15]. Pakistan is unique in this collection of results, there are
concerns around the accuracy of COVID-19 coverage of females. Overall, though
more studies, the finding that COVID-19 is just like widespread for males and females
all over the world would need to be confirmed. People diagnosed with COVID-19
are more likely to die in all states that record rates of death by gender.
The data in Fig. 6 showed that for every 10 females worldwide there are more cases,
death, hospitalization, and ICU admissions of males. Male had worse results than
females in coronavirus or serious respiratory syndromes (SARS) attacks. Statistics
from several countries throughout the globe propose that males would be at high risk
of death due to COVID-19. East Asian countries (China & North Korea), European
countries (Italy, Spain, Germany, Sweden, Switzerland, and France), Oceania coun-
tries (Australia), Middle East country (Iran), and North American country (Canada)
has a higher number of male cases than female attacked by COVID-19 [18, 19]. But
there is no data available for the South Asian country Pakistan. Hence, we conduct
this study to evaluate the sex gender impact on COVID-19 results in the South Asian
country Pakistan.

2 Methodology

We have collected data from different sources for this analysis. Firstly, we obtained
data on 23rd march from doctor’s doing jobs in hospitals (n = 113) all over Pakistan
(Punjab, Sindh, Khyber Pakhtunkhwa, Balochistan). All the doctors associated with
private and government hospitals are included in the study. We excluded nurses,
healthcare workers, and the doctors, which are doing house jobs. We obtained data
Analysis of Covid-19 Impact on Gender … 931

from the doctors with the help of a Google form-based questionnaire. In this data
we apply descriptive statistics first and then association tests to achieve the goal of
gender and sex influence on COVID-19 outcomes in Pakistan. Secondly, the data
on sex-specific percentages for Pakistan was collected directly from the national
government statistics of 25 April 2020 and African population of health center 25
January, 2021, which are added to the aggregated data issued by the Government of
Pakistan on 27 May 2020.

3 Results

In Table 2 expresses the results, which are obtained from our first source from the
doctors. The Table 2 describes the distribution of frequencies for which COVID-19
is more influenced by gender. Here, 81.4% of men respondents were impacted and
COVID-19 impacted 18.6% of women’s respondents in the South Asian country
Pakistan. The variable’s mean and standard deviation is (1.1858) and (.39071),
respectively.
The below pie chart diagram explains our results. As shown in Fig. 6 shows that
more males are affected by COVID-19 than females in the South Asian country
Pakistan (Fig. 7).
Secondly, the data from the national government statistics of 25 April 2020 and
data issued by the Government of Pakistan on 27 May 2020 was analyzed. Our
analyses indicate that approximately 70% of all COVID-19 cases in the South Asian
country of Pakistan are male, with very close statistics for sex-specific hospital admis-
sion, mortality, and post-infection recovery as shown in Table 2. Although the number
of male cases in Pakistan are 78%. From across South Asian nation Pakistan, with
sex-specific statistics on the number of all COVID-19 related deaths, the rate of
male deaths ranged from 65–76%. In the South Asian nation of Pakistan, the fatality
pattern indicates that the male-to-female event fatality ratio is 0.9 (1.9% for males
vs. 2.1% for females). In the below table Case Fatality represents the percentage of
positive cases that die afterward (Figs. 8, 9, 10 and 11).

Table 2 Descriptive analysis of gender affected to COVID-19

Gender Frequency Percent Mean Standard deviation
Male 92 81.4 1.1858 .39071
Female 21 18.6
Total 113 100.0
932 S. Malik et al.

Fig. 7 Pie chart of gender affected by COVID-19 in South Asian country Pakistan

Fig. 8 COVID-19 cases and deaths rates per 100,000 in men/women in Pakistan

4 Discussion

Researcher [31] mentioned in his research that worldwide, medical statistics warn
us that males suffer more from diseases of the respiratory system than females
(25%), particularly those caused by acute viral infections. Females are less vulner-
able to infectious diseases than Men, relying on distinct innate immunity, gonad
development of steroid hormones, and sex chromosome related factors.
Researcher [32] discusses in his study that there was increased predisposition
and susceptibility to heart disease in clinical trials where testosterone levels were
Analysis of Covid-19 Impact on Gender … 933

Fig. 9 COVID-19 Cases by age and sex (rates per 100,000) in Pakistan

Fig. 10 COVID-19 Total cases (% male/female) in Pakistan

high, which is more common in men who are more prone to COVID-19. Those
devoid of COVID-19 cardiac disease have a better prognosis associated with patients
with heart failure and hypertension [32]. Discussion in his paper that the sex and
gender inequalities found in COVID-19 prevalence highlight the need to consider
the effect of gender and sex on the occurrence of the disease and death, and to adapt
gender and sex care to fit. Experiences from previous diseases and pandemics have
strongly illustrated the value of integrating a gender and sex study into health policy
preparedness and response efforts.
Nevertheless, the gendered effects of illness rise in, epidemics or diseases have
not yet been discussed by legislation and public health initiatives. Such countries do
not disaggregate data by ethnicity and age as other countries do. Also, policymakers
will disaggregate and evaluate data for gender and age gaps in all countries. In fact,
934 S. Malik et al.

Fig. 11 COVID-19 deaths (% male/female) in Pakistan

as prophylactic and clinical trials proceed. Many studies identify the autoimmunity
effect of sex hormones.
In addition, data from many studies also shows that male COVID-19 patients and
other acute viral infections have greater vulnerability to mortality than females. In
men compared with women, it is considered to be more prevalent. Researchers [21]
explain in his study that there are many reasons behind this variance, like behavioral,
genetic, hormonal, as well as immunological variables. The mortality rates for men
are two to three times higher than for women, but tend to fluctuate in feature of popu-
lation area and demography. Social variables linked to ethnicity, immune differences,
hormonal variations and lifestyle behaviors including smoking and drunkenness will
play a role. Researcher [22] illustrates in his research that male populations can in any
way be vulnerable to the COVID-19 pandemic as a result of genetic polymorphisms
identified by gender.
Researcher [23] stated in his paper that beyond the biological danger the disease’s
epidemiological evidence of transmission indicates that the area is not similar and
that its effect on multiple populations is markedly interdependent. Researcher [24]
justify the occurrence of COVID-19 more on males, researcher stated based on its
socio-economic status the risk of males contracting and dying from SARS-CoV-2 is
significantly increased. In England and Wales, males throughout the deprived areas
have a mortality rate of 77.6 per 100,000 (44.3 per 100,00 for females) rising to 172
per 100,000 in its most deprived areas (97 per 100,000 for females).
Data shared by a globally chosen company Global Health 50/50 encouraging
balance between men and women in health, found a higher share of deaths in men
than in women in virtually every region. The details in the Italy Integrated Monitoring
Newsletter (April 23, 2020 update reveals that men’s deaths are about twice as many
as women’s deaths (17.1 vs. 9.3%). Researcher [25] found similar results in Greece,
the Netherlands, Denmark, Belgium, Spain, China and the Philippines. A sample
of 4,880 respiratory patients or near contact with Covid-19 patients in a hospital in
Analysis of Covid-19 Impact on Gender … 935

Wuhan showed that positive rates of SARS-CoV-2 in men and older populations were
substantially higher (>70 years) but only age is defined as a risk factor (>70 years).
Likewise, a recent retrospective observational analysis of researcher [26] showed
that 67% of critically ill patients with SARS-CoV-2 were men. Furthermore, a study
of researcher [27] results for 1.099 Covid-19 patients found 58.1% to be males of 173
extreme incidents, 57.8% have affected this population. Furthermore, reported infor-
mation from survival research [28] found that male deaths are considerably higher and
symptoms greater than female deaths. Finally, the researcher [29] recently recorded
a 1.7-fold higher case fatality rate for males than females (P < 00001). However,
researchers [30] recommend a walk-through gate for mitigating the COVID-19 effect.

5 Conclusion

We conduct this study to evaluate the sex gender impact on COVID-19 results in the
South Asian country Pakistan. Mixed methodology was used in this study, primary
data from doctors doing jobs all over Pakistan (Punjab, Sindh, Khyber Pakhtunkhwa,
Balochistan) and secondary data from national Pakistani government statistics. The
doctor’s response shows that 81.4% of males and 18.6 females are affected by
COVID-19 in Pakistan. Descriptive analysis between gender and COVID-19 results
in a mean of 1.1858 and a standard deviation of .39071. The data from national
Pakistani government statistics show that the number of male cases in Pakistan
is 78%, the rate of COVID-19 related male deaths ranged from 65–76%. From
the whole analysis, we can conclude that COVID-19 attacks more on males rather
than females in the South Asian country Pakistan. East Asian countries (China &
North Korea), European countries (Italy, Spain, Germany, Sweden, Switzerland, and
France), Oceania countries (Australia), Middle East country (Iran), South Asian
country (Pakistan) and North American country (Canada) has a higher number of
male cases than female attacked by COVID-19. However, we recommend further
comprehensive studies to measure the degree to which sex will lead to COVID-19
results among South Asian country Pakistan.
Limitations of the Study
The key limitation of this study applies to the comparatively low number of
samples carried out in the field, indicating that the overall number of tests and figures
can be significantly higher than what we report. However, we do not expect the central
point of this study to change considerably.

Acknowledgment The authors were very thankful to Dr. Imran Hussain, Dr. Wajid Yaqoob Bajwa,
Dr. Sheerbaz, Dr. Ali Hassan Shah, and Dr. Saqib Ishaq for providing data for our research.

Conflict of Interest Statements Author Saba Malik (Corresponding author), Mubbra Habib,
Mehmood Ahmed Husnain Hashmi, Muhammad Tariq saeed, Anwaar ul huda, Syeda Anam zahra,
Muhammad Nisar, Muhammad Sajid, Shahbaz Ahmad Shahzad, Saddam Hussain, and Muhammad
Sakandar Majid declares that they have no conflict of interest.
936 S. Malik et al.

References

1. Guan WJ, Ni ZY, Hu Y, Liang WH, Ou CQ, He JX, et al (2020) Clinical characteristics of
coronavirus disease 2019 in China. N Engl J Med. https://doi.org/10.1056/NEJMoa2002032
2. Who director-generals opening remarks at the media briefing on COVID-19 - 11 March 2020.
World Health Organisation (2020). https://www.who.int/dg/speeches/detail/who-director-
generals-opening-remarks-at-the-media-briefing-on-COVID-19%2D%2D-11-march-2020.
Accessed 22 Mar 2020
3. Coronarvirus Resource Center, Johns Hopkins University Medicine. 2020. Johns Hopkins
University Medicine. Medicine JHU. https://coronavirus.jhu.edu/map.html. Accessed 22 Mar
2020
4. Wu C, Chen X, Cai Y, Xia J, Zhou X, Xu S, et al (2020) Risk factors associated with acute
respiratory distress syndrome and death in patients with coronavirus disease 2019 pneumonia
in Wuhan, China. JAMA Intern Med. https://doi.org/10.1001/jamainternmed.2020.0994
5. Wang D, Hu B, Hu C, Zhu F, Liu X, Zhang J, et al (2020) Clinical characteristics of 138
hospitalized patients with 2019 novel coronavirus-infected pneumonia in Wuhan, China. Jama.
https://doi.org/10.1001/jama.2020.1585
6. Grasselli G, Pesenti A, Cecconi M (2020) Critical Care Utilization for the COVID-19 Outbreak
in Lombardy, Italy: Early Experience and Forecast During an Emergency Response. Jama.
https://doi.org/10.1001/jama.2020.4031
7. Zhao S, Cao P, Chong MK, Gao D, Lou Y, Ran J, et al (2020) The time-varying serial interval of
the coronavirus disease (COVID-19) and its gender-specific difference: a data-driven analysis
using public surveillance data in Hong Kong and Shenzhen, China from January 10 to February
15, 2020. Infect Control Hosp Epidemiol.https://doi.org/10.1017/ice.2020.64:1-8
8. Mo P, Xing Y, Xiao Y, Deng L, Zhao Q, Wang H et al (2020) Clinical characteristics of refractory
COVID-19 pneumonia in Wuhan. China Clin Infect Dis. https://doi.org/10.1093/cid/ciaa270
9. Conti P, Younes A (2020) Coronavirus Cov-19/Sars-Cov-2 affects women less than men:
clinical response to viral infection. J Biol Regul Homeost Agents 34(2)
10. Cai H (2020) Sex difference and smoking predisposition in patients with COVID-19. Lancet
Respir Med. https://doi.org/10.1016/s2213-2600(20)30117-x
11. World Health Organization. WHO Coronavirus Disease (COVID-19) Dashboard, 17 July 2020
https://covid19.who.int/. Accessed 17 July 2020
12. The epidemiological characteristics of an outbreak of 2019 novel coronavirus diseases (COVID-
19)—China, 2020. 2020. The Novel Coronavirus Pneumonia Emergency Response Epidemi-
ology Team, China CDC Weekly. https://www.ne.jp/asahi/kishimoto/clinic/cash/COVID19.
pdf. Accessed 03 Apr 2020
13. Globalhealth. (2021). The COVID-19 Sex-Disaggregated Data Tracker | Global Health 50/50.
https://globalhealth5050.org/the-sex-gender-and-covid-19-project/the-data-tracker/. Accessed
18 April 2021
14. Karlberg J, Chong DS, Lai WY (2004) Do men have a higher case fatality rate of severe acute
respiratory syndrome than women do? Am J Epidemiol 159(3):229–231. https://doi.org/10.
1093/aje/kwh056
15. Chen X, Chughtai AA, Dyda A, MacIntyre CR (2017) Comparative epidemiology of Middle
East respiratory syndrome coronavirus (MERS-CoV) in Saudi Arabia and South Korea. Emerg
Microbes Infect 6(6):e51. https://doi.org/10.1038/emi.2017.40
16. Epidemia COVID-19 (2020) Aggiornamento Nazionale. Dipartimento Malattie Infettive e
Servizio di Informatica, Istituto Superiore di Sanità. https://www.epicentro.iss.it/coronavirus/
bollettino/Bollettino-sorveglianza-integrata-COVID19_30-marzo-2020.pdf. Accessed 30 Mar
2020
17. Coronavirus disease 2019 (COVID-19) daily situation report of the Robert Koch Institute 2020.
Robert Koch Institute. Institute RK. https://www.rki.de/DE/Content/InfAZ/N/Neuartiges_C
oronavirus/Situationsberichte/2020-04-01-en.pdf?__blob=publicationFile. Accessed 03 Apr
2020
Analysis of Covid-19 Impact on Gender … 937

18. Gebhard C, Regitz-Zagrosek V, Neuhauser HK et al (2020) Impact of sex and gender on

COVID-19 outcomes in Europe. Biol Sex Differ 11:29 https://doi.org/10.1186/s13293-020-
00304-9.
19. Pennell LM, Galligan CL, Fish EN (2012) Sex affects immunity. J Autoimmun 38(2–3):J282–
J291. https://doi.org/10.1016/j.jaut.2011.11.013
20. Haque A, Akçeşme FB, Pant AB (2018) A review of Zika virus: hurdles toward vaccine
development and the way forward. Antivir Ther 23(4):285–293
21. Scully EP, Haverfield J, Ursin RL et al (2020) Considering how biological sex impacts immune
responses and COVID-19 outcomes. Nat Rev Immunol 20:442–447. https://doi.org/10.1038/
s41577-020-0348-8
22. Agrawal H, Das N, Nathani S, Saha S, Saini S, Kakar SS, Roy P (2020) An assessment on
impact of COVID-19 infection in a gender specific manner. Stem Cell Reviews Reports 1–19.
https://doi.org/10.1007/s12015-020-10048-z
23. Wenham C, Smith J, Davies S, et al (2020) Women are most affected by pandemics—lessons
from past outbreaks. Nature 583(7815):194–198
24. Bowleg L (2020) We’re not all in this together: on COVID-19, intersectionality, and structural
inequality. Am J Public Health 110(7):917
25. ONS (2020) Deaths involving COVID-19 by local area and socioeconomic deprivation: deaths
occurring between March and May 2020. Accessed 24 Jul 2020
26. Liu R, Han H, Liu F, Lv Z, Wu K, Liu Y, et al (2020) Positive rate of RT-PCR detection of
SARS-CoV-2 infection in 4880 cases from one hospital in Wuhan, China, from Jan to Feb
2020. Clin Chim Acta 505:172–175. https://doi.org/10.1016/j.cca.2020.03.009
27. Xie J, Tong Z, Guan X, Du B, Qiu H (2020) Clinical characteristics of patients who died of
coronavirus Disease 2019 in China. JAMA New Open 3:e205619. https://doi.org/10.1001/jam
anetworkopen.2020.5619
28. Jin JM, Bai P, He W, Wu F, Liu XF, Han DM et al (2020) Gender differences in patients with
COVID-19: focus on severity and mortality. Front Public Health 8:152. https://doi.org/10.3389/
fpubh.2020.00152
29. Scully EP, Haverfield J, Ursin RL, Tannenbaum C, Klein SL (2020) Considering how biological
sex impacts immune responses and COVID-19 outcomes. Nat Rev Immunol 20:442–447.
https://doi.org/10.1038/s41577-020-0348-8
30. Hussain S, Cheema MJM, Motahhir S, Iqbal MM, Arshad A, Waqas MS, Malik S (2020)
Proposed Design of Walk-Through Gate (WTG): Mitigating the Effect of COVID-19. Appl
Syst Innov 3(3):41. https://doi.org/10.3390/asi3030041
31. Fung TS, Liu DX (2014) Coronavirus infection, ER stress, apoptosis and innate immunity.
Front Microbiol 5:296. https://doi.org/10.3389/fmicb.2014.00296
32. Gebhard C, Regitz-Zagrosek V, Neuhauser HK et al (2020) Impact of sex and gender on
COVID-19 outcomes in Europe. Biol Sex Differ 11:29. https://doi.org/10.1186/s13293-020-
00304-9
Deep Residual Convolutional Neural
Network Based Detection of Covid-19
from Chest-X-Ray Images

Valaparla Rohini, M. Sobhana, Ch. Smitha Chowdary, Mukesh Chinta,

and Deepa Venna

Abstract The Coronavirus or 2019-nCoV (COVID-19) is a contagious disease.

This new strain outbreak pressuring the health community in the world because
this virus not identified early and also spreading to a large number of countries and
territories, here some of them are on the edge of failing to control the spread of this
virus. This virus mostly affects the respiratory system. So, it is possible to diagnose
the virus, infected in the respiratory system using CT scan and chest-x-ray imaging
approaches. Chest-X-ray images are amply available and swift imaging time than CT
scan. The coronavirus pneumonia infected chest-x-rays has taken as the input data.
Convolutional Neural Network (CNN) pre-trained ResNet-50 model is the proposed
method, and it uses cross-validation for analysis. However, this model is employed
to get high performance with more accurate result.

Keywords Corona virus · Chest-X-ray · Diagnosis · ResNet-50

1 Introduction

COVID-19 has been an epidemic disease caused by the foremost recently discov-
ered Severe Acute Respiratory Syndrome CoronaVirus2 (SARS-CoV2) [1]. There
are different diseases find by coronaviruses. Including a Severe Acute Respiratory
Syndrome (SARS), discovered from musk cats to humans in 2003, the Middle East
Respiratory Syndrome (MERS) discovered from dromedary to humans in 2012, and
COVID-19 was caused by bats [2]. Coronavirus first unveiled in Wuhan, December
2019. COVID-19 is presently an epidemic affecting many nations globally. With the
information of the World Health Organization (WHO) approximately in ten months,

V. Rohini (B) · M. Sobhana · M. Chinta · D. Venna

Department of CSE, V R Siddhartha Engineering College, Vijayawada 520007, Andhra Pradesh,
India
Ch. Smitha Chowdary
Department of CSE, Koneru Lakshmaiah Education Foundation, Guntur 522502, Andhra Pradesh,
India

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 939
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_85
940 V. Rohini et al.

there are more than 35.4 million people infected across 214 nations and territo-
ries. Coronavirus caused more than 1.04 million deaths, 24.6 million individuals
have recovered. The number of individuals infected by this Coronavirus disease is
increasing, and it turns to be increasingly hard to end the diagnostic process within
the constrained time. Currently, many efforts and more attention paid to identifying
the COVID-19 affected patients. The COVID-19 most common signs and symptoms
are dry coughs, fever, tiredness and less common signs and symptoms are aches and
pains, a headache, sore throat, diarrhea, conjunctivitis, lack of taste or smell, a rash
on the skin or discoloration of palms or toes [3]. Signs and symptoms are difficulty
respiration, chest pain, and lack of speech or movement. This virus is unfolding
among people at some point of contact, commonly through tiny droplets through
coughing, sneezing and talking. This contamination can also additionally attain to
pneumonia. Lungs was effected by the pneumonia infection and also in the human
body lungs was mostly effected by the coronavirus. So that CT scan and Chest-x-ray
images are used to diagnose the pneumonia.
Therefore, it requires expertise and necessitates for different algorithms to analyze.
Consistent with this context, Machine Learning algorithms have obtained better
performance in detection of pneumonia, and provides accurate detection rate. By
using machine learning algorithms CT scan method is used to diagnose the chest
pneumonia [4]. But it is the time taken to process and cost-effective. And CT scanner
may not available in many regions as well. For that purpose, the contribution of our
paper used chest-x-ray images because chest-x-ray images are widely available and
take less time than the CT scan [5, 7]. So that, our paper follows penalized tuned
versions of Convolution Neural Network-based models InceptionV3, Resnet50.

1.1 Deep Convolutional Neural Network

A deep convolutional neural network used to examine the visual images and also for
classifying images. Convolutional Neural Networks (CNNs) is the used weights for
each pixel are fully connected. The layers in the convolutional neural network are
fully connected, pooling and Convolutional layers. These layers are used to perform
a different tasks on the input data. When deep transfer learning [6] techniques applied
to medical data, meaningful results are come out and these techniques [7] are used
for classification, segmentation and infection detection of medical data.

2 Related Work

Here some related research work for COVID-19 disease. S. Sai Thejeshwar et al. [8]
presented a KE Sieve Neural Network using the transfer learning for features which
extracted and weights on the VGG-19 pre-trained CNN model. This algorithm is not
capable of some specified dataset.
Deep Residual Convolutional Neural Network … 941

Al-Bawi, Ali, et al. [9] the classical VGG neural network performed with the
convolutional COVID block, and it obtained 95.51% accuracy for covid-19, normal,
pneumonia chest x-ray images. It was time taking process, and training deep neural
network was difficult.
Fang Y. et al. [10] proposed an RT-PCR technique used to identify the viral
RNA and serology testing will identify the visible representation of antibodies. In
this study when observing the patient samples, there is a difference in the detection
rate for initial CT (50/51 [98%]) patients greater than first RT-PCR (36/51 [71%])
patients. Collected annotated data is very expensive or difficult validation CT datasets
collected in one system developed to quantify infections only.
Zhang, Jianpeng, et al. [11] Rapid and accurate detection of the virus used
pervasive prevention and large-scale screening. Without fine-tuning, anomaly three
modules achieved an accuracy of 83.61%. It is cost-effective.
Shan F, Gao Y, Wang J, et al. [12] Deep learning-based method used for segmen-
tation and quantification of the respiratory system using chest CT scans [4]. This
system used a “VB-Net” neural network obtained dice similarity coefficients 91.6%
± 10.0% and with accuracy 86.7% on CT images. It is less sensitive to detect virus
from the throat.
Mohamed Loey et al. [13] GAN (Generate Adversarial Network) is for prepro-
cessing, deep transfer model for training, testing and validation. GAN has Image
Generation network and discriminator network. The Image Generation network
generates false data. Discriminator network describes the difference between real
and fake data. The algorithm takes some set of transfer models. The considered
dataset was 30 times larger than the original dataset, so it creates the fake dataset
images.

3 Proposed Methodology

In this method, the deep convolutional neural network (DCNN) based pre-trained
models have used for the categorization of chest-x-ray images. By using the Deep
Transfer learning method, it controls inadequate data and training time. By using
the pre-trained models of the CNN i.e., ResNet50 [14] and InceptionV3 [15] the
coronavirus was detected from chest-x-ray images.
ResNet50 is the best technique [16] of CNN because it uses shortcuts and also
saves the time of training. It intercepts the distortion of a system which is deeper
and complex. In the ResNet-50 model, bottleneck blocks used for faster training
[17]. ResNet50 contain 50-layer deep neural network and it has trained on the Image
Net database [18]. The database has contained 14 million images created for image
recognition competitions which are belonging to more than 20000 categories [19].
InceptionV3 is CNN pre-trained model. It consists of several convolutional steps
and maximum pooling steps. A Fully connected convolution network has connected
at the final stage. The InceptionV3 pre-trained model used to extract the features of
images [20] (Fig. 1).
942 V. Rohini et al.

Fig. 1 Flow diagram for

prediction of Covid-19
Clinical
Symptom

Chest-x-ray images

Load pre-trained models

Cross-Validation

Training dataset

Validation dataset

Estimated
Probability

3.1 Data Set

The input dataset has taken from open source Kaggle dataset repository [21]. The
repository contains covid-19 and regular chest-x-ray images. It is an unbalanced
dataset. Here the output is a probability of accuracy (Figs. 2 and 3).

Fig. 2 Covid-19 positive

chest-x-ray images
Deep Residual Convolutional Neural Network … 943

Fig. 3 Normal Patient’s

chest-x-ray images

Table 1 5-Fold Cross-validation Training and Testing Data.

Iteration1 Iteration2 Iteration3 Iteration4 Iteration5
Test fold1 Train fold1 Train fold1 Train fold1 Train fold1
Train fold2 Test fold2 Train fold2 Train fold2 Train fold2
Train fold3 Train fold3 Test fold3 Train fold3 Train fold3
Train fold4 Train fold4 Train fold4 Test fold4 Train fold4
Train fold5 Train fold5 Train fold5 Train fold5 Test fold5

3.2 Performance Evaluation

The introduced method using python3 programming language with Ubuntu 16.04
operating system. The considered dataset contains chest x-ray images. The 30 epochs
carried out to over fitting pre-trained models. The dataset differentiated into 2 parts
80% training and 20% testing for performance evaluation. The 5-fold cross-validation
the method is chosen k value from 1 to 5 (Table 1).

1
5
performance = performancek
5 k=1

Specificity = True − Negative/(True − Negative + False − Positive) (1)

Precision = True − Positive, /(True − Positive + False − Positive) (2)

Recall = True − Positive/(True − Positive + False − Negative) (3)

F1 − Score = 2x((Precision x Recall)/(Precision + Recall)) (4)

Accuracy =(True − Negative + True − Positive)/(True − Negative + True − Positive

+ False − Negative + False − Positive) (5)
944 V. Rohini et al.

Here True-Positive correctly labelled as Covid-19, False-Positive represent

normal but mislabeled positive, True-Negative correctly labelled normal, but False-
Negative mislabeled negative by the proposed model. The above equations defined
for obtaining the performance.

4 Results and Discussion

In this proposed work, data is trained and tested on the chest-x-ray images. Training
accuracy and loss values of pre-trained ResNet50, InceptionV3 for fold-3 shown
in Figs. 4 and 5. To avoid the over-fitting 30epochs are taken for the training.
When compared to InceptionV3, Resnet50 obtained the highest training accuracy.
The ResNet50 gives a fast training process than InceptionV3. The loss values are
decreased in ResNet50 and InceptionV3 at the training stage. ResNet50 pre-trained
model reduces the training loss values quickly than the InceptionV3.

Fig. 4 Fold-3 performance

of training accuracy for
pre-trained models

Fig. 5 Fold-3 performance

of training loss values for
pre-trained models
Deep Residual Convolutional Neural Network … 945

Table 2. Confusion Matrix Performance results

Models/Fold Confusion matrix and performance results (%)
TP TN FP FN Acc Rec Spe pre F1
Inception V3 Fold 1 7 10 0 3 85 70 100 100 82
Fold 2 10 10 0 0 100 100 100 100 100
Fold 3 10 10 0 0 100 100 100 100 100
Fold 4 10 10 0 0 100 100 100 100 100
Fold 5 10 10 0 0 100 100 100 100 100
Performance 97 94 100 100 96.4
ResNet50 Fold 1 8 10 0 2 90 80 100 100 89
Fold 2 10 10 0 0 100 100 100 100 100
Fold 3 10 10 0 0 100 100 100 100 100
Fold 4 10 10 0 0 100 100 100 100 100
Fold 5 10 10 0 0 100 100 100 100 100
Performance 98.2 96 100 100 97.8

The other detailed way of given test data shown in Table 2. The highest accuracy of
98.2%. In training and testing, ResNet50 provides superiority over the InceptionV3.
Here when observe the given table confusion matrix results are described from the
pre-trained models fold-1 to 5.

5 Conclusion

For the detection of coronavirus, the proposed method deep convolutional neural
network based approach using chest-x-ray images has obtained a good result on the
considered dataset automatically. ResNet-50 is the one of best algorithm in deep
learning having a good result. The performance of ResNet50 pre-trained model
attains with the highest accuracy 98.2%. It is an accurate, rapid and effective clinical
support system for covid-19 detection. The training dataset is limited for proposed
work. Further, will try to work on different CNN models can be tested by using the
more number of Covid-19 and chest-x-ray images in the dataset.

References

1. Yan L, Zhang HT, Xiao Y, Wang M, Guo Y, Sun C, Tang X, Jing L, Li S, Zhang M, Xiao Y, Cao
H, Chen Y, Ren T, Jin J, Wang F, Xiao Y, Huang S, Tan X, Huang N, Jiao B, Zhang Y, Luo A,
Cao Z, Xu H, Yuan Y (2020) Prediction of criticality in patients with severe covid-19 infection
using three clinical features: a machine learning-based prognostic model with clinical data in
Wuhan. MedRxiv 2020.02.27.20028027
2. Costa S, Gomes, V, Moreli ML, Saivish, MV (2020) The emergence of SARS, MERS and
novel SARS-2 coronaviruses in the 21st century
946 V. Rohini et al.

3. Bouaziz JD, Duong T, Jachiet M, Velter C, Lestang P, Cassius C, Arsouze A et al (2020)

Vascular skin symptoms in COVID-19: a French observational study. J Eur Acad Dermatol
Venereol
4. Shuai W, Bo K, Jinlu M, Xianjun Z, Mingming X, Jia G, Mengjiao C, Jingyi Y, Yaodong L,
Xiangfei M, Bo X (2020) A deep learning algorithm using CT images to screen for corona
virus disease (covid-19), medRxiv
5. Zhang J, Xie Y, Li Y. Shen C, Xia Y (2020) COVID-19 screening on chest X-ray images using
deep learning based anomaly detection, arXiv
6. Chuanqi T, Fuchun S, Tao K, Wenchang Z, Chao Y, Chunfang L (2018) A survey on deep
transfer learning In: The 27th international conference on artificial neural networks (ICANN).
Lecture Notes in Computer Science. Springer
7. Oh Y, Park S, Ye JC (2020) Deep learning covid-19 features on cxr using limited training data
sets. IEEE Trans Med Imaging
8. Sai Thejeshwar S, Chokkareddy, C, Eswaran K (2020) Precise prediction of covid-19 in chest
x-ray images using ke sieve algorithm. ResearchGate
9. Al-Bawi A, Al-Kaabi KA, Jeryo, M, Al-Fatlawi A (2020) CCBlock: an effective use of
deep learning for automatic diagnosis of covid-19 using x-ray images. arXiv preprint arXiv:
2009.10141
10. Fang Y, Zhang H, Xie J, Lin M, Ying L, Pang P, Ji W (2020) Sensitivity of chest CT for
COVID-19:comparison to RT-PCR. Radiology
11. Zhang J, Xie, Y, Liao Z, Pang G, Verjans J, Li, W, Sun Z et al (2020) Viral pneumonia screening
on chest x-ray images using confidence-aware anomaly detection. arXiv
12. Fei S, Gao Y, Wang J, Shi W, Shi N, Han M, Xue Z, Shi Y (2020) Lung infection quantification
of covid-19 in ct images with deep learning. arXiv preprint arXiv: 2003.04655
13. Loey M, Smarandache F, Khalifa NEM (2020) Within the lack of chest covid-19 x-ray dataset:
a novel detection model based on gan and deep transfer learning. Symmetry 12(4):651
14. Wen L, Li X, Gao L (2019) A transfer convolutional neural network for fault diagnosis based
on ResNet-50. Neural Comput Appl. Springer
15. Wang C, Chen D, Hao L, Liu, X, Zeng Y, Chen J, Zhang G (2019) Pulmonary image
classification based on inception-v3 transfer learning model. IEEE Access 7
16. Wu Z, Shen C, Van Den Hengel A (2019) Wider or deeper: Revisiting the resnet model for
visual recognition. Elsevier
17. Yamazaki M, Kasagi A, Tabuchi, A, Honda T, Miwa M, Fukumoto N, Tabaru T, Ike A,
Nakashima K (2019) Yet another accelerated SGD: Resnet-50 training on imagenet in 74.7
seconds. arXiv: 1903.12650
18. Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, Huang Z, Karpathy A, Khosla A,
Bernstein M, Berg AC, Fei-Fei L (2015) Imagenet large scale visual recognition challenge. Int
J Comput Vis 115:211–252
19. Krizhevsky A, Sutskever I, Hinton GE (2017) ImageNet classification with deep convolutional
neural networks. Commun ACM 60(6):84–90
20. Dong N, Zhao L, Wu, CH, Chang JF (2020) Inception v3 based cervical cell classification
combined with artificially extracted features. Elsevier (2020)
21. Kaggle home page (2018) https://www.kaggle.com/paultimothymooney/chest-xray-pne
umonia
A New Optimized Approach to Resolve
a Combinatorial Problem: CoronaVirus
Optimization Algorithm
and Self-organizing Maps

Omayma EL Majdoubi, Farah Abdoun, and Otman Abdoun

Abstract The optimization provides resolutions to complex combinatorial prob-

lems that generally deal within large data size and expensive operating processes.
Metaheuristics target the promoting of new applied algorithms to resolve NP-Hard
problems in order to improve resolutions requiring enhanced search strategies. Trav-
elling Salesman Problem (TSP) is a common combinatorial problem, applied on
several benchmark networks, namely the transportation networks and the routing
vehicle problem in order to establish new intelligent computing methods as well as,
to prove studies on their performances and efficiencies. Collective intelligence has
proven satisfactory resolutions wherewith metaheuristics, although their algorithms
complexity. The aim of this work is to solve the Euclidean TSP, classified as a NP-
hard problem by means of Self-Organizing Maps (SOM) which is a Kohonen-type
network. The resolution is also computed corresponding to a new bio-inspired evolu-
tionary strategy so-called coronavirus optimization algorithm combined with SOM
algorithm. The present approach is combining an unsupervised learning strategy
within the new coronavirus optimization algorithm to replicate iteratively new
infected individuals and to generate diversification on the search space. This new
hybrid method presents a good approximative resolutions, proved by applying tests
for TSPLIB instances wherein the exact optimum is defined corresponding to each
TSP data. However, the present resolutions are complex specifically for large scale
by means of increasing the size of input data or size parameters.

Keywords Travelling salesman problem · CoronaVirus optimization algorithm ·

Kohonen network · Neural network · Self-organizing maps · Parallel computing ·
Combinatorial optimization problem · Adaptation · TSPLIB

O. EL Majdoubi (B) · O. Abdoun

Computer Sciences Department, FPL, Abdelmalek Essaâdi University, Larache, Morocco
F. Abdoun
Research Center STIS, M2CS, ENSAM, Mohammed V University, Rabat, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 947
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_86
948 O. El Majdoubi et al.

1 Introduction

The optimization problem is defined as a maximization or minimization problem

admitting a set of input design variables, the set of admitted solutions and their
definitions via an associated computed measurement [1]. Many fields of operations
research and artificial intelligence emphasize on presenting resolution of combina-
torial optimization problems, in order to define a solution vector depending on the
given objective function and within the respect of predefined constraints [2]. Various
techniques are used to solve exactly the combinatorial optimization problems and to
determine the optimum. These methods perform an implicit research among the set of
space of solutions. Yet, the optimum solutions are proved only for small or medium
scale of problem instances and then many interesting problems are not computed
by exact methods. Therefore, the aim of the optimization is based on developing
heuristics or metaheuristics in order to find good eligible approximations that are not
necessary the optimal solution [3]. Collective intelligence, perceived on biological
cycle of social insects, is inspiring researchers to develop metaheuristics algorithms
in the context area of swarm intelligence (SI) and conducting efficient resolutions
of various optimization problems [4]. The paper is structured as following: The first
section concerns a literature review on combinatorial optimization problems and
TSP resolutions. The second section is a review related to Kohonen-type network.
The third section is a review related to coronavirus optimization algorithm. The next
section is related to discuss and analyze the numerical results. Finally, a conclusion
and future work are presented.

2 Combinatorial Optimization Problems

The theory of computing complexity reveals that several problems are complex and
proven to be not efficiently processed by sequential or parallel deterministic algo-
rithm [5]. The interest of the researchers surrounds the proposal of methods to solve
complex problems in order to reach optimal or sub-optimal solutions within a reason-
able run time. The exact methods are dedicated for small problems. Thus, the approx-
imate methods including heuristics and metaheuristics achieve solutions within a
reasonable execution time [6]. Commonly, there is no efficient method to attain opti-
mality in a quick polynomial run time for resolving NP-Hard problems, including
graphic coloring, Knapsack problem, integer programming, TSP and its variants. It is
imperative to speed up the execution time of proposed resolving methods, in partic-
ular for large scale problems [7]. The embedding of parallel processing techniques
is recommended and possibly depending on parallel architecture of the machine,
for resolving expensive algorithms and providing feasible solutions [8]. TSP is a
classical optimization problem, applied to many area contexts of computing, logis-
tics and industrial engineering. For its various applications, the researchers solved
this problem through branch and bound method, linear programming and dynamic
A New Optimized Approach to Resolve a Combinatorial Problem … 949

programming, as well as, the genetic algorithms, ant colony optimization and artifi-
cial bee colony optimization wherein the input scale is increasing [9–14]. Further-
more, the aforementioned problem is a complex combinatorial optimization problem
among the NP-Hard problems, that requires exploring new resolution methods,
specifically the bio-inspired metaheuristics. The Cat Swarm Optimization inspired
from real cat behaviors, reveals to be efficient for the TSP resolution [6]. Spider
Monkey Optimization is a new method among SI algorithms, is proved to solve the
TSP with good performances [4]. Indeed, the bio-inspired methods used to solve TSP,
for instance genetic algorithms, ant colony optimization, particle swarm optimization
and bat algorithm optimization, are not necessary providing the best solutions [15].
Differential evolution is an evolutionary algorithm, usually used to solve continuous
optimization problems. A discrete version of the differential evolving algorithm is
proposed to improve resolving of complex discrete TSP [16]. The TSP problem is
described as a salesman who is required to choose an itinerary of minimum distance
to carry out the distribution of its merchandises. The salesman must travel from the
depot, visit each city once and return to the starting point [17]. The enumeration of all
tours for a given problem is not a possibly appropriate method. Corresponding to an
n-city TSP, the possible permutations is equal to (n − 1)! Routes [18]. In addition, the
exact methods are requested to constitute a search space of all possible permutations
of cities, then to find the corresponding optimal solutions. The resulting run time is
a polynomial factor of O(n!), as n is the length of the route. The exact resolutions of
large size problems are inoperable [7]. As the genetic algorithm is well performing
in achieving good solutions for TSP, this genetic approach is also inconvenient in
case of premature convergence, expensive evolving routines and non-performance
of local search techniques. Thus, an immune based genetic algorithm is attested to
exceed these inconvenient and improving TSP solutions [15]. Admittedly, the TSP
is formulated mathematically as following [19]:
n n
Minimize : z = ci j xi j
i=1 j=1
n
Subject to : xi j = 1, j = 1, 2, . . . .n,
i=1
n
xi j = 1, i = 1, 2, . . . .n,
j=1
xi j ∈ {0, 1}, i, j = 1, 2, .......n.

x f or ms a H amiltonien cycle (1)

Where cij and xij are the costs and the decision variables respectively linking
position i to position j. Vector x̃ has the whole sequence of the resulting route which
is a solution for the TSP [20]. In what follows, a literature review is achieved for
studying a Kohonen-type neural network in order to investigate its performances in
resolving TSP.
950 O. El Majdoubi et al.

3 Self-organizing Maps: SOM

The Artificial Neural Network (ANN) is described as a processing information

system, bio-inspired based on the properties of biological nervous systems [21].
Also, the neural network is designed as the model of brain information processing,
consisted of samples of communicating computing units within a connected network
[22]. In fact, ANN simulate biological networks system of neurons, in order to permit
computers to learn similarly as human-reinforcement learning. The neural network
is basically compound of perceptron that is consisted of inputs, processor and single
resulting output [23]. The SOM algorithm introduced by Kohonen, is described
as an unsupervised learning method for algorithms improvement associated to a
specific input data [24]. SOM algorithm is a class of competitive neural network
aiming to explore network topologies and distribution strategies of input patterns
[25]. Particularly, the SOM are derived from neural network fundaments and are
structured in two-layers of competitive learning network [26]. Generally, the SOM
are distinguished by its abilities of recognition via an unsupervised method, are a
full monolayered connected network [27], where the output nodes are in contest
explained by the update of the winner node and its neighbors [28]. Variants of SOM
are commonly able to evaluate input patterns of the network via various strategies of
auto-organization based on the learning approaches. Generally, the self-organizing
learning is based on frequent update of synaptic weights of the considered network
and in accordance with the input patterns by means of learning strategies [29]. The
Fig. 1 is a representation of SOM network.
The SOM algorithms are used to solve TSP especially by integrating efficient
initialization methods and adequate adaptation parameters to better explore solutions
also, to enhance fast convergence [29]. Applying this unsupervised learning method
to TSP, leads to define an input vectors including the coordinates for the given cities
and the corresponding weights of network nodes. The winner neuron is conquering
the output nodes for a given city. Then, the neighboring routine is applied to find the
closer node to the winner node by updating neurons weights. This SOM procedure is
applied till up attaining a given max error or a predefined maximum of evolvement
iterations [30]. SOM algorithms are efficient in solving planning path problems
argued by its performances of adaptation procedures and on exploring iteratively
winner nodes and their corresponding neighboring [30]. Figure 2 illustrates SOM
algorithm applied to the TSP [30].

Fig. 1. Self-organizing map

network on n input pattern
A New Optimized Approach to Resolve a Combinatorial Problem … 951

Fig. 2. SOM algorithm applied to TSP [30]

The geometric tour is modeled as a stretching ring of nodes while the network
is evolving. The input data are presented randomly in the network. The neuron
network competition is based on considering Euclidean distance between nodes of
the network. The winner node J is obtained with the minimum distance to the present
city as formulated below [29]:
  
J = Argminj xi − yj 2 (2)

Where (xi, yi ) is the coordinate of city i and ||.|| is the Euclidean distance [29]. The
SOM adaptation procedure is applying firstly via an inhibition mechanism in order
to explore a winner node corresponding to each city up to represent all the cities
of the network once in one vector. The winner node is selected when the Euclidean
distance connecting the given city and the candidate node is minimum. Furthermore,
the adaptation procedures result in moving the winner node and its neighbors towards
the presenting city as following [30]:
 
vj∗ = vj + μf(., .) c − vj (3)

Where μ is the corresponding learning rate. The neighboring function is defined

as:
 
ex p −d 2 /G 2 , d < 0.2 m
f (G, d) = (4)
0, else

Where G is the gain parameter and d is the cardinal distance measured along the
ring consisted of m nodes. The gain parameter is initialized by G0 = 0.06 + 12.41 m
and updated for each iteration. The predefined values of learning and decreasing rates
are μ = 0.6 and α = 0.1 [30]. The competitive training algorithm is used to allow the
winning neuron to learn as well as the neurons located within a predefined radius from
the winning neuron and to update the related adaptation measures simultaneously
[29]. The cardinal distance measured along the ring between nodes j and J, where m
952 O. El Majdoubi et al.

is the number of the neurons is obtained as following [29]:

d = min{||j − J||, m − ||j − J||} (5)

Importantly, the neighborhood function has an impact on planning for the shortest
path by exploring closer nodes to each other’s [29]. It is important to develop heuris-
tics for parameters adaptation and adjustment in order to reach a faster convergence.
The main adaptation parameters are updated as described on Eqs. 6 and 7. The
learning rate is related to the number of iterations. Furthermore, the neighborhood
function variance is defined depending to an appropriate initialization equal to 10
[29].

1
αk = √
4
(6)
k

σk = σk−1 (1 − 0.01 × k), σ0 = 10 (7)

The resolution of Euclidean TSP is also made via hybrid methods, for instance,
by combining SOM procedure within an evolutionary dynamic algorithm, by the
integration of a mapping operator, a fitness evaluation and a selection procedure.
This method is interesting justified by the efficiency and the performances relative to
computed solutions and run time [31]. The hybridization of evolutionary algorithms
is a powerful method for the resolution improvement. The recombination of algo-
rithms is possibly made by implementing unrelated methods to previous information
of parents, evolutionary process, or designing interaction models within the set of
solutions [32].

4 Coronavirus Optimization Algorithm: CVOA

Recently, the world faced a new respiratory virus so-called coronavirus (COVID-
19). At the event of the outbreak, a novel bio-inspired metaheuristic is developed,
modeling the spread of coronavirus. Since an initial infected individual called Patient-
Zero (PZ), the coronavirus infects new individuals according to a specific rate and
create a search space of infected individuals. Each individual belongs to infected
population, is possibly constrained to die, infect or recover. The main properties of
CVOA are as follows: The probabilities and parameters are defined also updated by
scientific community, the exploration of search space is handled as long as the infected
population is not null and the high rate of expansion ensures better use of search space
leading to the intensification of the resolution. Also, the concept of parallel strains
reformulated as applying the processing algorithms in different processors in order to
generate diversification of the resolutions. The Coronavirus optimization algorithm
begins by generating a random initial population since the individual PZ. The next
step concerns the exploration of dead individual depending on die probability(P_DIE)
A New Optimized Approach to Resolve a Combinatorial Problem … 953

and make their removal from the infected population. This step so-called intensifica-
tion is also handled for the detection of recovered individual who possibly can infect
others, depending on probabilities of spreading (P_SUPERSPREADER). Also, for
ensuring diversifications of individuals, it is appropriate to define an immigration
probability (P_TRAVEL) for detecting individuals enabling the spread of coron-
avirus in other solutions. The update of populations is ensured for each iteration.
Each dead individual is appended in dead population to avoid reusing it. After each
iteration, the infected population are sent to recovered population. The recovered
individual can be reinfected if the probability of reinfection (P_REINFECTION) is
acquired. Thus, each individual is subjected to distancing or isolation measures if the
probability of isolation is achieved. Finally, each infected individual is sent to new
infected population. The stop criterion is expected if there is no evolution of the infec-
tion, explained by the decreasing size of infected population and new infected popu-
lation, or if the individuals in the new infected population cannot transmit anymore
the virus. Generally, the stop criterion is achieved throughout a specific duration
called PANDEMIC_DURATION [33]. The main CVOA algorithm [33] is presented
as following (Fig. 3):

Fig. 3. Main CVOA procedure [33]

954 O. El Majdoubi et al.

5 Results and Discussions

This work consists on applying a new method hybridizing Coronavirus Optimization

Algorithm and SOM algorithm to solve the Euclidean TSP. The conceived approach
is based on considering subsequently a random path of n city as an input pattern.
Corresponding to each generated tour, the adaptation parameters are calculated to
select the winner node of which the distance is minimum to the located point of the
tour. The adaptation procedure is then applied to detect the neighbors in order to
create a new candidate solution. The implementation of the proposed method is an
evolutionary algorithm based on probabilistic procedures, that requires a definition
of the replicate function, which is suggested to insert new individuals to be subjected
to CVOA evolution. The replicate function is thus conceived to generate candidate
solution since SOM algorithm. The results are obtained by programming in Matlab
using the TSPLIB instances. The Table 1 presents the numerical results of SOM
algorithm by applying adaptation functions.
The following section is presenting the obtained results of TSP resolution using
the proposed approach combining the CVOA and SOM algorithms. Corresponding
to primary executions, the implementation requires an expensive run time, especially
when the size of the tour is increasing or the number of iterations since SOM imple-
mentation(it_Max) as well the number of iterations since the present hybrid algorithm
(PANDEMIC_DURATION). Hence, the multi-core parallelism via SPMD is applied.
The obtained results also are obtained by considering: it_Max = 10, number of cores
= 4, PANDEMIC_DURATION = 250, and by considering a predefined learning
parameter: μ = 0.6 and α = 0.1.
The results as shown in Table 2 are better improved compared to SOM imple-
mentation. The following section is presenting the aforementioned implementation
by considering the adaptation functions of learning parameters.
The presented solutions are also better improved by considering the adaptation
of learning parameters. The evolution of the convergence related to the considered
approach is described in Table 3 and Fig. 4.
The convergence evolution corresponding to small and medium data size
(Burma14, Ulysses16, Ulysses22) is similar as the average deviation are 0.09, 0.08

Table 1. The Numerical

TSPLIB Present Exact Numbers of
results of SOM algorithm
instance solution optimum iterations
applied to TSP
Burma14 33.6037 30.8785 3000
Berlin52 20985 7542 2500
Ulysses16 74.2578 74.1087 3000
Ulysses22 99.7789 75.5975 3000
Att48 108250 10628 2500
Rat99 6738. 3 1211 2500
Pr76 467350 108159 2500
A New Optimized Approach to Resolve a Combinatorial Problem … 955

and 0.06 related to the Euclidean TSPLIB instance Burma14 and corresponding to
results of Tables 1, 2 and 3 respectively. Also, the average deviation related to the
TSPLIB instance Ulysses16 are 0.002, 0.06 and 0.017 corresponding to SOM resolu-
tion and to hybrid proposed algorithm considering a predefined and adapted learning
parameter respectively. The calculated average deviation related to the TSPLIB
instance Ulysses22 are 0.3, 0.4 and 0.3 resulting since the three considered imple-
mentation tests respectively. The considered approaches presented in this work are
efficient to resolve the TSP corresponding to small and medium data scale. Indeed,
corresponding to the TSPLIB instances namely Berlin52, Att48, Rat99 and Pr76,
the SOM resolution is not good performing. Thus, the results of the aforementioned
instances are better improved by the implementation of the proposed hybrid method
especially, whereas the learning parameters are adapted. The obtained average devi-
ation is 0.3 and 0.5 corresponding to Pr76 and Rat99 respectively, resolved by
the present combined algorithm and where the learning parameters are adapted.
Therefore, it is important to improve the present proposed resolution by tuning of
the considered parameters or experiment the successful neighborhood procedures.
Importantly, it is necessary to apply the techniques of parallel computing in order to
achieve fast convergence and to improve resolutions.

Table 2. The Numerical

TSPLIB instance Present solution Exact optimum
results of hybrid CVOA-SOM
algorithm applied to TSP Burma14 33.5524 30.8785
Berlin52 20624 7542
Ulysses16 79.1248 74.1087
Ulysses22 107.8736 75.5975
Att48 108920 10628
Rat99 1915 1211
Pr76 147060 108159

Table 3. The Numerical

TSPLIB instance Present solution Exact optimum
results of adapted hybrid
CVOA-SOM algorithm Burma14 32.8596 30.8785
applied to TSP Berlin52 20985 7542
Ulysses16 75.3930 74.1087
Ulysses22 100.4639 75.5975
Att48 24246 10628
Rat99 1915 1211
Pr76 147060 108159
956 O. El Majdoubi et al.

Fig. 4. The convergence evolution of the proposed algorithm to solve TSP

6 Conclusion and Future Work

The present work is emphasizing on the importance of hybridizing bio-inspired

methods within learning strategies. The proposed approach is requiring parallel
processing especially when the complexity of the algorithm is increasing. Also,
the improvement of this hybrid approach depends on the improvement of SOM algo-
rithm, on the selection and search strategies as well as on tuning parameters. Hence,
the improvement of the proposed hybrid metaheuristic is the prospect for future
research.

References

1. Bovet DP, Clementi A, Crescenzi P, Silvestri R (1996) Parallel approximation of optimization

problems. In: Ferreira A, Pardalos PM (eds) Solving combinatorial optimization problems in
parallel. LNCS, vol 1054, pp 248–274. Springer, Berlin
2. Lüling R, Monien B, Reinefeld A, Tschöke S (1996) Mapping tree-structured combinatorial
optimization problems onto parallel computers. In: Ferreira A, Pardalos PM (eds) Solving
combinatorial optimization problems in parallel. LNCS, vol. 1054. Springer, Berlin, Heidelberg
3. Laursen PS (1996) Parallel heuristic search – introductions and a new approach. In: Ferreira A,
Pardalos PM (eds) Solving combinatorial optimization problems in parallel, LNCS, vol 1054.
Springer, Berlin, Heidelberg, pp 248–274
4. Akhand MAH, Ayon SI, Shahriyar SA, Siddique N, Adeli H (2020) Discrete spider monkey
optimization for TSP. Appl Soft Comput vol 86, p 105887
5. Clementi A, Rolim JDP, Urland E (1996) Randomized parallel algorithms. In: Ferreira A,
Pardalos PM (eds) Solving combinatorial optimization problems in parallel. LNCS, vol 1054.
Springer, Heidelberg
6. Bouzidi A, Riffi ME (2019) Improved CSO to solve the TSP. In: Ezziyyani M (eds)
AI2SD Conference 2018. Advances in Intelligent Systems and Computing, vol 714. Springer,
Singapore
7. Othman A, Haimoudi E, Mouhssine R, Ezziyyani M (2019) An effective parallel approach to
solve multiple TSP. In: Ezziyyani M (eds) AI2SD Conference 2018. Advances in Intelligent
Systems and Computing, vol 714. Springer
8. Ferreira A, Pardalos PM (1996) SCOOP: Solving combinatorial optimization problems in
parallel. In: Ferreira A, Pardalos PM (eds) Solving combinatorial optimization problems in
parallel. LNCS, vol 1054. Springer, Berlin
A New Optimized Approach to Resolve a Combinatorial Problem … 957

9. Durbin R, Willshaw D (1987) An analogue approach to the travelling salesman problem using
an elastic net method. Nature 326:689–691
10. Wang J, Ersoy OK, He M, Wang F (2016) Multi-offspring genetic algorithm and its application
to TSP. Appl Soft Comput 30:484–490
11. Ezugwu AES, Adewumi A, Frîncu M (2017) Simulated annealing based symbiotic organisms
search optimization algorithm for traveling salesman problem. Expert Syst Appl 77:189–210
12. Choong SS, Wong LP, Lim CP (2019) An artificial bee colony algorithm with a modified choice
function for the traveling salesman problem. Swarm Evol Comput 44:22–635
13. Chen WL, Dai SG (2006) Survey of algorithms for traveling salesman problem. J Chuzhou
Univ 8(3):1–6
14. Guo JY (2006) An overview of traveling salesman problem. Pop Sci Technol 8:229–230
15. Lahjouji A, Tajani C, Krkri I, Fakhouri H (2019) Immune based genetic algorithm to solve a
combinatorial optimization problem: Application to traveling salesman problem. In: Ezziyyani
M (eds) AI2SD Conference 2018, Advances in intelligent systems and computing, vol 915.
Springer
16. Ismail MA, Daryl E, Kathryn K (2020) A novel design of differential evolution for solving
discrete TSP. Swarm Evol Comput 52:100607
17. Hoffman KL, Padberg M, Rinaldi G (2013) Traveling salesman problem. In: Gass S I, Fu MC
(eds) Encyclopedia of operations research and management science. Springer, Boston
18. Gilmore PC, Gomory RE (1964) A solvable case of the traveling salesman problem. Proc Natl
Acad Sci 51:178–181
19. Ahuja RK, Magnanti TL, Orlin JB (1993) Network flows. Prentice-Hall
20. Bazaraa MS, Jarvis JJ, Sherali HD (1990) Linear programming and network flows. Wiley, New
York
21. Koprinkova HP, Mladenov V, Kasabov NK (2015) Artificial neural networks. Springer Series
in Bio-/Neuroinformatics
22. Trick M (2008) Oper Res Lett 36(2):276–277
23. Khourdifi Y, Bahaj M (2020) Analyzing social media opinions using hybrid machine learning
model based on artificial neural network optimized by particle swarm optimization. In:
Ezziyyani M (eds) AI2SD Conference 2019. Advances in Intelligent Systems and Computing,
vol 714. Springer, Singapore
24. Kohonen T (1990) The self-organizing map. Proc IEEE 78(9):1464–1480
25. Kohonen T (2001) Self-organizing maps. Springer , Heidelberg
26. Somhom S, Modares A, Enkawa T (1997) A self-organising model for TSP problem. J Oper
Res Soc 48(9):919–928
27. Kohonen T, Kaski S, Lagus K, Salojarvi J, Honkela J, Paatero V, Saarela A (2000) Self
organization of a massive document collection. IEEE Trans Neural Netw 11(3):574–585
28. Moukhafi M, Yassini KE, Bri S, Oufaska K (2019) Artificial neural network optimized by
genetic algorithm for intrusion detection system. In: Ezziyyani M (eds) AI2SD Conference
2018. Advances in Intelligent Systems and Computing, vol 714. Springer, Singapore
29. Yanping B, Wendong Z, Zhen J (2006) A new self-organizing maps strategy for solving TSP.
Chaos Solitons Fractals 28(4):1082–1089
30. Jan F, Miroslav K, Vojtěch V, Libor P (2011) An application of the self-organizing map in the
non-euclidean TSP. Neurocomputing 74(5):671–679
31. Créput JC, Koukam A (2009) A memetic neural network for the euclidean TSP. Neurocom-
puting 72(4–6):1250–1264
32. Tinós R (2020) Artificial neural network based crossover for evolutionary algorithms. Appl
Soft Comput 95:106512
33. Martínez-Álvarez F, Asencio-Cortés G, Torres JF, Gutiérrez-Avilés D, Melgar-García L, Pérez-
Chacón R, Rubio-Escudero C, Riquelme JC, Troncoso A (2020) Coronavirus optimization
algorithm: a bioinspired metaheuristic based on the COVID-19 propagation model. Big data
8(4):308–322
Study of a Dual-Function Intelligent
Electronic Pin to Help Compliance
with Security Measures Related
to Covid-19
Laince Pierre Moulebe, Abdelwahed Touati, Eric Obar Akpoviroro,
and Nabila Rabbah

Abstract COVID-19 is a virus that developed in December 2019, the disease has
spread around the world and there are many cases per day. The latest information
reports that there are more than sixty-nine million cases of infection worldwide
according to an official diagnosis. The new coronavirus pandemic has killed at least
one million six peoples worldwide since its emergence, according to a report by AFP
(Agence France Presse), but the actual number estimate higher, knowing that some
countries are exceeded by the number of cases in hospitals and that others do not
have the resources to test the majority of their population. Several treatment methods
to slow down and eliminate Covid-19 exist to date. Indeed, following the appearance
of the new Corona Virus, the world scientific community has carried out several
research and development for eliminated the virus, there are smart masks, respirators,
gels, mobile applications, vaccines that are gradually developing and much more
research. However, according to studies, transmission is mainly by direct contact,
including contact of uninfected hands with the face, non-compliance with distancing
measures. And, according to psychologists, one of the most effective methods to
fight against the Covid-19 is to prevent these frequent forget fulnesses. On the other
hand, research shows that hand-to-face contact is a frequent human behavior. Hence,
the main objective of this article is to propose a prevention technique through the
study of an electronic device that allows users to keep their safety distance, while
reminding them of the approach of a hand towards their face. It is an electronic pin
featuring PIR sensors, with lenses designed to detect at distances 1.5 m. Containing
a battery for power, a button, which turn on or off the pin, depending on whether you
are in a bus for example or in an airport. Besides, this simple electronic component
designed, pin will have suitable cost for all social categories.

Keywords Covid-19 · Prevention · Pin electronic · Distancing measure · Hand

detection · PIR sensor

L. P. Moulebe (B) · A. Touati · E. O. Akpoviroro · N. Rabbah

Laboratory of Structural Engineering, Intelligent, Systems & Electrical Energy ENSAM,
Hassan II University, Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 959
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_87
960 L. P. Moulebe et al.

1 Introduction

Covid-19 was declared by the World Health Organization on March 11, 2020 as
a pandemic following the number of deaths which has multiplied 13 times more
than that recorded in China [1]. Covid-19 continues to increase today, with more
than sixty-nine million cases positive infection that have been officially diagnosed in
208 countries and territories since the start of the epidemic. According to the AFP,
the actual number of contaminations is much higher than this value, knowing that a
large number of countries now only testing cases requiring hospital treatment. Among
these cases, at least forty-four million are now considered cured. Covid-19 is usually
spread through close contact (shaking hands, by the airway, coughing and sneezing).
Several researches are established in order to fight this pandemic [2], through some
suggestions for health workers because they are less exposed and their availability
is essential [3, 4], moreover according to [5] it is important to take an interest in
the behavior of people working in hospitals, on their knowledge about the epidemic.
Despite the containment including about 4.5 billion of the population, awareness of
behaviors and habits, respect for hygiene, we note that the number of cases continues
to increase. Practical innovative systems such as automatic disinfectants suitable for
entering large spaces [6], smart bracelets allowing hands to be disinfected at any time
[7], it is clear that innovations are continually emerging. All these systems aim to
fight against the pandemic which affects more than 300 people every minute [8]. To
date several countries have developed vaccines which are being distributed [9–11],
however populations remain hesitant about these vaccines [12, 13], because some
of the first vaccines were not effective from the start. This distrust of the population
shows that it remains important to offer solutions in addition to vaccines which are
the subject of speculation, this is why in this article we propose another innovation in
the prevention of Covid-19. According to [14], touching the face is a typically human
behavior that expresses self-awareness; however, it is one of the best ways to prevent
coronavirus spread. Several studies on self-contact have been established over the
time [15–19], showing that, it is a behavior that is repeated frequently, but further
research shows that in self-contact the contact of hand to face is the most frequent
[20]. In [21] a study shows the behavior concerting the contact of the hands with
the face to perform on 60 peoples. We see that it is indeed one of the most frequent
human behaviors, but the best way to fight covid-19 is to limit it, even avoid contact
of non-disinfected hands with the face because each forget (Fig. 1) may be likely to
cause multiple casualties. Hence, to limit the errors and forgetting when faced with
Covid-19 concerning touching the face, in this article we propose the study of an
electronic device that would allow social distancing (Fig. 2) while allowing users
to always clean their hand before touching their face and keep the distancing. The
electronic system studied mainly consists of a PIR sensor, a lens, an on/off button,
a power supply battery, a microcontroller and a sound. The studied device will be
worn like a pin, its size is around 50 mm.
Study of a Dual-Function Intelligent Electronic Pin … 961

Fig. 1 Example of behavior

This work is divided into three parts, the first paragraph presenting the description
of the system, then in the second the proposed technical solution followed by a
conclusion.

2 Description Studied

Devices used to remind humans to perform tasks are found a lot in our environment,
in the health sector we use connected objects such as Medi’Rappel, Medisafe, to
remind to take these drugs or to respect the dosage, warning lights used in cars,
control systems in industry. The Covid-19 is mainly transmitted between people by
respiratory droplets and contact [21–25],In an analysis of 75,465 cases of Covid-19
in China, no transmission by air was reported [26]. The device is designed to detect
peoples over a distance of 0 to 1.5 m, the device mainly consists of a PIR (Fig. 3)
detector which is ideal motion sensors for humans and animals [27], a battery, a
microcontroller allowing the sounder to be put on and standby for 5 after detection
of the approach of a hand towards the face or of a person within a radius of 1.5 m
according to the simple algorithm (Fig. 4) and a microphone (Fig. 5 b). Note that
these are miniaturized devices with the smallest sizes. The detection distance is an
important parameter who is calibrate with the lens of the PIR sensor. Through the
work established by [27, 28], we can evaluate the characteristics (Fig. 5 a) of the
battery for a detection distance between 0–1,5 m which is around 5 V. The electronic
pin is designed to allow users to keep their safe distance. Therefore, the user wearing
the pin is automatically alerted by a song in the event of presence of a person within
1.5 m, but also when a hand approaches the face, knowing that the hand passes
through the field of the lens, this sensor can be shut down if we are in a bus or public
transport where compliance with the measures is difficult.
In this flowchart we show the system that puts the sensor on stand-by, (1) the
sensor is active and detects a hand, then the alert is issued and (2) the user cleans
his hands. Subsequently, through a microcontroller, the sensor is put on stand-by for
5 min (3). And finally the controller activates the sensor after 5 min (4).
962 L. P. Moulebe et al.

Fig. 2 Safety distance

Fig. 3 PIR sensor

3 Technical Solution

In this part we present an electronic diagram of the proposed solution to perform on

easyEDA Fig. 6.
The proposed circuit is similar to a basic motion detection circuit with some
modification, in fact the proposed circuit mainly consists of a PIR sensor (IRA), a
push button (KEY1) to turn on/off the pin, a power supply battery (BAT), a sounder
to emit the alert song (BUZZER) and the components necessary for the operation of
the usual motion sensor circuit. One of the particularity of this system is the function
of being a pin placed on a garment, and the lens which must limit the detection to
110°/1.5 m.
Study of a Dual-Function Intelligent Electronic Pin … 963

1 sensor on hand/human detected

Alert

2 Hand clean/distance taken

Microcontroller desactiveted sensor

3 sensor off Delay 5 min

Microcontroller activeted sensor

4 sensor on

Fig. 4 Algorithm of microcontroller

Fig. 5 a Output of a PIR sensor in case of passages at different distances, b sounder

As mentioned in the introduction, the device is of the order of a few centimeters,

in order to have a shape similar to the common pins, the 3D solution appears as
shown in Fig. 7.
964 L. P. Moulebe et al.

Fig. 6 Electronic circuit of pin

Fig. 7 3D pins model

4 Conclusion

In this work we were able to highlight the fact that COVID-19 is mainly spread by
contact, and that despite several restriction such as containment, awareness to clean
hands and safety distance, the number of cases continues to increase, then it has been
shown that touching the face is a typically human behavior. This allowed us to show
the need for a simple system allowing users to have a simple and easily portable
system that reminds them to clean their hands before touching their face but also to
respect social distancing measures.
Thus, omissions of safety measures in supermarkets, in front of banks or any other
space will be limited as well for contact of the hands with the face. In definitive this
solution will help to reduce the propagation of the Covid-19.
Study of a Dual-Function Intelligent Electronic Pin … 965

References

1. Ho LTF, Chan KKH, Chung VCH, Leung TH (2020) Eur J Integr Med 101116. https://doi.org/
10.1016/j.eujim.2020.101116
2. Yen M, Schwartz J, Chen S, King C, Yang G, Hsueh P (2020) ScienceDirect Interrupting
COVID-19 transmission by implementing enhanced traffic control bundling: implications for
global prevention and control efforts. J Microbiol Immunol Infect. 3. https://doi.org/10.1016/
j.jmii.2020.03.011
3. Xu K, Lai X, Zheng L (2020) Suggestions on the prevention of COVID-19 for health care
workers in department of otorhinolaryngology head and neck surgery. World J Otorhinolaryngol
Neck Surg. https://doi.org/10.1016/j.wjorl.2020.03.002
4. Tebala GD, Bond-Smith G ( 2020) Guidelines and recommendations during the COVID-19
pandemic: a word of caution. Am J Surg 220(6):1526–1527. https://doi.org/10.1016/j.amjsurg.
2020.06.005
5. Shi Y et al (2020) Brain, Behavior, & Immunity - Health Knowledge and attitudes of medical
staff in Chinese psychiatric hospitals regarding COVID-19. Brain Behav Immun Heal 4:100064.
https://doi.org/10.1016/j.bbih.2020.100064
6. Hussain S et al (2020) Proposed design of walk-through gate (WTG): Mitigating the effect of
COVID-19. Appl Syst Innov 3(3): 1–7. https://doi.org/10.3390/asi3030041
7. El Majid B, Motahhir S, El Hammoumi A, Lebbadi A, El Ghzizal A Preliminary design of a
smart wristband disinfectant to help in Covid-19 Fight. https://doi.org/10.3390/inventions50
30032
8. CovInfo - Suivez l’évolution du Coronavirus dans le monde avec notre bulletin quotidien.
https://www.covinfo.fr/monde. Accessed 06 Dec 2020
9. Reddy KA, Srinivas K, Ayyappan GS (2015) Simulation of power management circuit in a
standalone solar photovoltaic-fuel cell-battery hybrid power system. In 2015 IEEE IAS Joint
Industrial and Commercial Power/Petchem Product Industries ICPSPCIC, 168–173. https://
doi.org/10.1109/CICPS.2015.7974070
10. “Covid-19 vaccine: who are countries prioritising for first doses? | World news | The
Guardian.” https://www.theguardian.com/world/2020/nov/18/covid-19-vaccine-who-are-cou
ntries-prioritising-for-first-doses. Accessed 11 Dec 2020
11. Vaccins contre le Covid-19 : quels sont les pays qui ont déjà passé commande ?
- L’Express.” https://www.lexpress.fr/actualite/monde/vaccins-contre-le-covid-19-quels-sont-
les-pays-qui-ont-deja-passe-commande_2138952.html. Accessed 11 Dec 2020
12. On Kwok K et al (2020) Journal Pre-proof Influenza vaccine uptake, COVID-19 vaccination
intention and vaccine hesitancy among nurses: a survey TITLE Influenza vaccine uptake,
COVID-19 vaccination intention and vaccine hesitancy among nurses: a survey. Int J Nurs
Stud (2020). https://doi.org/10.1016/j.ijnurstu.2020.103854
13. Vaccine hesitancy in the University of Malta Faculties of Health Sciences, Dentistry and
Medicine vis-à-vis influenza and novel COVID-19 vaccination | Elsevier Enhanced Reader.
https://reader.elsevier.com/reader/sd/pii/S0378378220307623?token=22F51A1FB4FEFC9
15C562BAC3928A46C9FB5C26CBB714C789EB335A44A2423C702966C5AD8E77760B
4D65FC3F6264393. Accessed 07 Dec 2020
14. Why you can’t stop touching your face, even though it’s one of the best ways to prevent coron-
avirus spread. https://www.businessinsider.fr/us/psychological-reason-why-you-cant-stop-tou
ching-face-coronavirus-2020-3?utm_source=feedburner&utm_medium=referral. Accessed
09 Apr 2020
15. Ekman P, Friesen WV (1974) Nonverbal behavior and psychopathology. In: Kats MM,
Friedman RJ (eds) The psychology of depression. Contemporary theory and research, New
York, pp 203–224
16. Barroso F, Freedman N, Grand S (1980) Self-touching, performance, and attentional processes.
Percept Mot Skills 50(3):1083–1089
17. Hampson E, Kimura D (1984) Hand movement asymmetries during verbal and nonverbal tasks.
Canadian J Psychol 38(1):102–125 PMID: 6713294
966 L. P. Moulebe et al.

18. Harrigan JA (1982) (1985) Self-touching as an indicator of underlying affect and language
processes. Soc Sci Med 20(11):1161–1168
19. Leavens DA, Aureli F, Hopkins WD, Hyatt CW (2001) Effects of cognitive challenge on self-
directed behaviors by chimpanzees (Pan troglodytes). Am J Primatol 55(1):1–14. https://doi.
org/10.1002/ajp.1034. PMID: 11536312
20. Goldberg S, Rosenthal R (1986) Self-touching behavior in the job interview: antecedents and
conse- quences. J Nonverbal Behav 10(1):65–80
21. Mueller SM, Martin S, Grunwald M (2019) Self-touch: contact durations and point of touch of
spontaneous facial self-touches differ depending on cognitive and emotional load. PLoS One
14:3. https://doi.org/10.1371/journal.pone.0213677
22. Liu J, et al. (2020) Community Transmission of Severe Acute Respiratory Syndrome Coro-
navirus 2, Shenzhen, China, 2020. Emerg Infect Dis 26(6). https://doi.org/10.3201/eid2606.
200239
23. Modes of transmission of virus causing COVID-19: implications for IPC precaution recom-
mendations https://www.who.int/news-room/commentaries/detail/modes-of-transmission-of-
virus-causing-covid-19-implications-for-ipc-precaution-recommendations. Accessed 11 Apr
2020
24. Li Q et al (2020) Early transmission dynamics in Wuhan, China, of novel coronavirus-infected
pneumonia. N. Engl J Med 382(13):1199–1207. https://doi.org/10.1056/NEJMoa2001316
25. Chan JFW et al ( 2020) A familial cluster of pneumonia associated with the 2019 novel
coronavirus indicating person-to-person transmission: a study of a family cluster. Lancet
395(10223):514–523. https://doi.org/10.1016/S0140-6736(20)30154-9
26. Frieden TR, Lee CT (2020) Identifying and Interrupting Superspreading Events—Implications
for Control of Severe Acute Respiratory Syndrome Coronavirus 2. Emerg Infect Dis. 26(6)
https://doi.org/10.3201/eid2606.200495
27. Gami H (2017) Movement direction and distance classification using a single PIR sensor. IEEE
Sensors Lett. 2(1):1–4. https://doi.org/10.1109/lsens.2017.2782179
28. Zappi P, Farella E, Benini L (2010) Tracking motion direction and distance with pyroelectric
IR sensors. IEEE Sens J 10(9):1486–1494. https://doi.org/10.1109/JSEN.2009.2039792
DenTcov: Deep Transfer Learning-Based
Automatic Detection of Coronavirus
Disease (COVID-19) Using Chest X-ray
Images

Youssra El Idrissi El-Bouzaidi and Otman Abdoun

Abstract On 31 December 2019, COVID-19, a novel coronavirus, appeared for the

first time in the Chinese city of Wuhan, to act as a preliminary warning and affected
a wider human being in the world. This virus, declared a pandemic by the auspices
of the World Health Organization (WHO), given its high rate of transmissibility.
The protocol most often used to detect the virus is PCR. It is a time-consuming and
less sensitive procedure with high false-negative results. These problems are solved
through radiographic imaging techniques to detect radioactive symptoms related
to COVID-19. Furthermore, significant time is required to complete the analytical
task, and mistakes can occur, meaning that automation is necessary. The use of
advanced Artificial intelligence tools can significantly accelerate both the time and
quality of the analysis. We suggest DenTcov, a computer-aided approach to detect
COVID-19 infection via chest X-ray images. Our model is a two-phase process:
Phase (1) Pre-Processing and data augmentation, Phase (2) COVID-19 detection
based DensNet121, a pre-trained model, then trained with the dataset prepared by
us. During the experimental phase of DenTcov, we measure the performances of
the architecture by calculating a set of common metrics, both 2-class and 3-class
classification. The experimental assessment confirms the DenTcov model offers a
96.52 and 99% higher classification accuracy for three and two classes, respectively,
compared to other proposed methodologies.

Keywords Artificial intelligence · Deep learning · CNNs · Transfer learning ·

DenseNet121 · COVID-19

1 Introduction

Recently, coronavirus (COVID-19) caused a veritable worldwide panic on all the

continents, disrupting lives and causing high mortality. After initial screening, the
virus has spread to almost every country, killing more than 1,104,061 citizens out of

Y. El Idrissi El-Bouzaidi (B) · O. Abdoun

Laboratory of Advanced Sciences and Technologies, Polydisciplinary Faculty, Abdelmalek
Essaadi University, Larache, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 967
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_88
968 Y. El Idrissi El-Bouzaidi and O. Abdoun

an estimated 39,164,496 confirmed cases as of October 16, 2020, according to the

statistics published on the World Health Organization’s website [1]. A rapid expan-
sion of the virus was expected. Therefore, it is necessary for prevention, quarantine,
and treatment procedures to identify early and effectively any infected person.
Many developed countries have seen their health infrastructures deteriorate, the
lack of equipment and kits (PCR) has become a problem. To diagnose this virus, one
of the following tests is considered justified: radiographic imaging, or X-rays, and
computed tomography or CT images. Chest radiography is preferable to computed
tomography because CT imaging generally takes much longer than x-ray images,
making it expensive, the availability of needs-based CT scanners is restricted to
industrialized countries. On the other hand, radiographic images remain largely the
most common and accessible method of diagnostic imaging for the most practical
and clinical needs [2].
One problem with the X-ray images is that sometimes the characteristics that
describe the existence of COVID-19 are mixed with other respiratory diseases such
as pneumonia, and therefore, radiologists find it difficult to diagnose this disease.
Deep learning techniques help to solve these problems, and their predictive accuracy
is the same, and sometimes better than that of a radiologist [3]. Diagnostic using
Artificial Intelligence (AI) solutions are getting more and more successful in the
modern medical field [4]. In particular, image classification by convolutional neural
networks is the most interesting tool of deep learning [5]. Image-based diagnostics
using deep learning and computer vision techniques have proven their great value
in providing a rapid and precise disease diagnosis consistent with a radiologist’s
reliable accuracy [3]. Advanced learning methods are not currently being replaced
by qualified medical diagnostic clinicians and are intended to complement strategic
decision-making. DenTcov Model based on deep learning, more precisely convolu-
tional neural network-based an advanced deep-transfer learning algorithm to auto-
matically extract features from the radiographic images that describe the appearance
of the disease and indicate whether it is a COVID-19 infection, which is very useful
in helping healthcare clinicians screen, identity, and track positive cases.
The paper was organized into the following sections: The proposed model is
outlined and detailed below starting Sect. 3. The parameters relating to the experi-
mental model with performance measures appear in Sect. 4. For more discussion of
the proposed DenTcov model and related results, see Sect. 4.2 and 4.3, respectively.
See Sect. 5 to conclude.

2 Related Work

Since computer-assisted image and deep learning technologies have proven their
superiority for medical image analysis, with the reliability to detect COVID-19 into
chest X-ray image [6, 7], researchers use advances of these techniques as a tool to
diagnose COVID-19. Applying artificial intelligence offered significant advantages
in detecting the disease and measuring infection rates, with promising results.
DenTcov: Deep Transfer Learning-Based Automatic Detection .. 969

Farooq and Hafeez [8] proposed by successive stages based on the architecture of
ResNet-50 a model, called COVIDResNet, using a collection of radiographic image
datasets that classifies patients as normal, bacterial, viral Pneumonia, and COVID-19
positive infections. An accuracy of 96.23% was achieved.
In another Paper [9], transfer learning techniques applied with the concatena-
tion of pre-trained models, ResNet50V2 and Xception, the proposed approach was
compared to Xception Res-Net50V2 models as a means of classifying chest X-ray
images according to the normal, COVID-19, and non-COVID pneumonia classes.
They used five-fold cross-validation as part of their classification testing process.
This technique resulted in approximately an accuracy of 91.4%.
Narin et al. [10] showcased several deep learning approaches, including three
distinct deep learning architectures, ResNet-50, Inception-ResNetV2, and Incep-
tionV3, The binary classification is used to discriminate COVID-19 distinct to the
healthy image. The evaluation indicates that ResNet50 had the highest classification
efficiency with an accuracy of 98.0%, recall of 96%, and accuracy of 100%, against
the accuracy of 97.0 and 87% for InceptionV3 and Inception-ResNetV2, respectively.
Chowdhury et al. [7] demonstrate some techniques with a transfer technique
used to detect an infection caused by COVID-19 on X-ray images of the lungs. They
demonstrate using pre-trained architecture such as AlexNet, DenseNet-201, ResNet-
18, and SqueezeNet, employed as classifiers. A classification accuracy of 98.3% was
achieved using the SqueezeNet network, with a specificity of 99%.
Asif and Wenhui [11] propose a system to automatically diagnose COVID-19,
the Transfer Learning Based Inception V3 was used to diagnose the patient’s chest
infection. This model was tested against normal radiographic imaging, pneumonia,
and COVID-19. Classification accuracy of 96% was achieved.
Khan et al. [12] have come with a classifier of chest X-ray images called CoroNet,
classifying images as Normal, Bacterial, Viral Pneumonia, and COVID-19 based on
the Xception architecture. As a result, they obtained an overall performance accuracy
of up to 89.6%.
Sethy and Behera [13], a number of various CNN models for extraction features
were trained using radiographic imagery of different architectures together in
conjunction with the Support Vector Machine for detecting COVID-19 positive cases,
thus achieving the accuracy of 95,38%.

3 Proposed COVID-19 Disease Detection Model

In this section, the DenTcov model is illustrated and described in detail. The archi-
tecture of the different parts is illustrated (see Fig. 1). The model was defined in two
phases. Phase I consists of (1) pre-processing and (2) data augmentation. In phase
II, (1) a deep convolutional network is specified as the first step, (2) the best deep
transfer learning model has been defined in the second step. a background discussion
on the process for each part of the model is given.
970 Y. El Idrissi El-Bouzaidi and O. Abdoun

3.1 Model Architecture Overview

An advantage offered by an automated diagnostic solution based on the radiographic

image using deep learning approaches, helping to significantly reduce examination
processing time and cost-efficiency, such solutions allow network weights to be
trained from large datasets. Only a limited amount of positive radiographic image
data relating to COVID-19 are available at present. An effective way to achieve
significant results in classification problems with a lack of data in a given domain is
to exploit existing knowledge from a similar domain, a technique known as transfer
learning [14], which is processed by employing a pre-trained CNNs and fine-tuning
weights on small datasets. Given a model founded on transfer learning, with a limited
training dataset, overfitting was a major issue. Therefore, the technique of augmenting
the data on a limited set of training data both solved the overfitting challenge and
improved the generalizability of the model.

Train Set
Validation Set
Test Set Automatic Learning rate

Train Loss
Update

Pre-Processing Augmentation
Input
X-ray
images

(COVID-19 Probabilities)

Best Model Obtained

Dense block AdaptiveMaxPool Covid-19

Conv + ReLu AdaptiveAvgPool Normal
MaxPool BatchNorm2d
Pneumonia
Transition Softmax

Decision-Making of the proposed DenTcov

Phase I Phase II

Fig. 1 DenTcov architecture

DenTcov: Deep Transfer Learning-Based Automatic Detection .. 971

Fig. 2 A Representative chest x-ray images after performing the Augmentation techniques

Table 1 Augmentation and

Data augmentation Pre-processing
Pre-processing parameters
used in the proposed model. Random flip Resizing (to 224 × 224 × 3)
Random rotation (maximum Normalization (mean and
angle was 10) standard deviation)
Random zoom (5%)
Change in brightness and
contrast

3.2 Data Augmentation

Data augmentation has occurred in the creation of new data using modifications made
to the dataset. Basically, with reasonable modifications and by applying different
transformations randomly to our training set, we create new augmented data (see
Fig. 2).
The following Table 1 summarizes the transformations we used in this work.

3.3 Network Architecture

For the purpose of this paper, for feature extraction, the knowledge from an existing
set of Convolutional Neural Network architectures pre-trained was used, with excel-
lent results achieved by using a variety of classification tasks as alternatives to
proposing our own architecture from scratch. For this reason, The DensNet121 archi-
tecture [17] has been used as an effective way for feature extraction by using its
weights that are pre-trained in the ImageNet dataset [15] using reduced parameters.
Another reason to use this network is to take advantage of the network’s condensed
structure, offering easy training, especially due to the reuse of features by other
layers, which increases the variety of its content between layers and improves the
network’s performance and generalization. we also adopt advanced high-speed tech-
niques introduce in Fastai [16], allowing us to improve the accuracy of our model
through every few epochs. The subsequent section provides a short summary of the
DenseNet121 architecture.

DensNet121. Transmits the characteristics to all the next layers in each dense block.
In its architecture, DensNet121 is defined as follows: there are four dense blocks,
972 Y. El Idrissi El-Bouzaidi and O. Abdoun

each dense block has a different number of layers (6, 12, 24, 16 consecutive layers).
Between each dense layer, there are transition layers, which allow reducing the size of
the features that are generated by the concatenation. The classification layers contain
a Global Average-Pooling plus a Fully Connected layer. For details please see [17].

Pre-trained Model. Our DenTcov has two components: the first consists essen-
tially of exploiting the capabilities offered by the DensNet121 for the transformation
of visual elements into feature vectors, while the second component consists of a
fully connected layer whose primary objective is to predict classification. As part
of this work, we have added a sequence of layers to the early portion of a typical
DensNet121, namely an adaptive Average/Max pooling, Batch normalization, and
drop out according to [16].
The second part of the DensNet121 is replaced with a fully connected network
so that the probabilities are calculated according to our classes, Normal, Pneumonia,
and COVID-19. As a predictable result, between each of these classes, is the one
having the highest probability. During the training phase, the modified parts of the
model were fine-tuned with pre-trained weights, leaving the original part untrained.
We obtain the best convolutional neural network architecture for transfer learning
and its hyperparameter by training the model and selecting an optimal learning rate
using the cyclic learning rate technique proposed in [18]. We configure the training
process of our model with the parameters given in the following Table 2.

Table 2 Hyperparameters
Parameters Value
details used in DenTcov
model. Batch size 64
Optimizer Adam, SGD
Momentum Betas = (0.9, 0.99)
Classes 3 Classes (Normal, Pneumonia,
Covid-19)
Loss function Crossentropy. Useful in
multi-classification problems
Training epochs 15
Dropout 0.3
Classifier Softmax (layer outputs the
probability distribution)
Activation function ReLu
DenTcov: Deep Transfer Learning-Based Automatic Detection .. 973

4 Experimentation and Discussion

4.1 Experiments

Dataset. The dataset was obtained from GitHub, the public source repository that
Dr. Joseph et al. has shared [19] which contains chest X-rays of patients infected by
COVID-19, also X-rays belonging to other COVID 19 families, such as ARDS,
SARS, and MERS. In our dataset, we only included radiographic images with
COVID-19, representing a total of 190 images, 264 normal x-ray images, and 275
bacterial and viral pneumonia infections, obtained from the Kaggle “Chest XRay
Images (Pneumonia)” [20].

Pre-processing. During the pre-processing phase, the images are distributed

randomly, 80% for training, and 20% of the data to validate. At the input of the
DenseNet121, the images are reduced to 224 × 224 × 3 pixels, and the model is
tuned.

Training. During the training, only the newly added layers of the network and last
layers group; Fully connected layers, are trained for 5 epochs while preserving the
weights of ImageNet for the rest of the network. We use the cyclic learning rate
technique introduced in [18], to help select the optimal learning rate. Using these
discriminatory learning rates, which keep a low learning rate for the initial layers
because they need less tuning and progressively increase the learning rate on the later
layers that need more tuning, especially those that are fully connected. By using the
learning rate (Max_lr) in the range 1e−4, 1e−3, all the network is fine-tuned, as first
layers with 1e−4 and the last layers with 1e−3. Each intermediate layer has been
trained with learning rates between these values for 10 epochs.

Evaluation Metrics. We have provided the four evaluation matrices to judge the
achievement of the results obtained from the DenTcov model. The true positives,
true negatives, false positives, and false negatives correspond to T.P, T.N, F.P, and
F.N, respectively, which are as follows in Table 3.

Table 3 Hyperparameters details for the DenseNet121 pre-trained model

Metric Value Metric Value
Recall (3)
Accuracy (1)
Precision (2) F1 (4)
974 Y. El Idrissi El-Bouzaidi and O. Abdoun

4.2 Results

The DenTcov was evaluated for two scenarios. The first is the classification of the
radiographic images into three classes: pneumonia, COVID-19, and Normal. The
second scenario is the classification of both COVID-19 and Pneumonia.
Figures 3, 4 illustrates the curves representing the loss convergence relating to
the training and the validation phases for 3-class (see Fig. 3) and 2-class (see Fig. 4),
respectively, the loss function going into 70 epochs and the highest classification
achieved an Accuracy of 96.52% (see Fig. 5).
Figure 6 gives an overview of the confusion matrix of the COVID-19 class
compared to the other classes.
We show how the Pneumonia class is not classified well (see Fig. 6), This problem
may be due to the fact that the class group two types of pneumonia: Viral and
Bacterial. Moreover, people who develop Pneumonia can have different levels of
progression. It is also notable that the classes COVID-19 achieves a perfect sensitivity,
and Class Normal already have a good level of correct classification. An excellent
compromise between the true and predicted values by DenTcov; which made it
possible to verify the effectiveness of the model, these results provide an encouraging
Loss

Batches processed

Fig. 3 Convergence graph of training and validation losses for class 3

Loss

Batches processed

Fig. 4 Convergence graph of training and validation losses for 2-class

DenTcov: Deep Transfer Learning-Based Automatic Detection .. 975

Error_rate
Accuracy

Batches processed Batches processed

Fig. 5 Accuracy and Error rate convergence graph for 3-class

Covid-19 Covid-19

Actual
Actual

Normal
Pneumonia

Pneumonia

Covid-19

Pneumonia
Covid-19

Pneumonia
Normal

Predicted Predicted

Fig. 6 The confusion matrix of DenTcov for 3-class & 2-class classification

Table 4 Precision, Recall,

Metrics COVID-19 (%) Normal (%) Pneumonia (%)
and F1-score corresponding
to different classes. Recall 100.0 92.0 89.09
Precision 97.29 93.24 89.09
F1-Score 98.62 92.61 89.09

as a tool of reducing stress on the health care system. Table 4 details the precision,
the recall, and the F1 score, respectively, for each class.

4.3 Discussion

Our current work focuses on automated detection from X-ray images of the charac-
teristics that could be significantly related to COVID-19 disease compared to other
respiratory diseases, in order, to classify it effectively, the use of the DensNet121
architecture was effective in this task, due to the high accuracy that has been achieved,
96.52 and 99% for three and two classes, respectively. Another key advantage, getting
a higher recall value signifies fewer False Negative, which is a significant result in
976 Y. El Idrissi El-Bouzaidi and O. Abdoun

Table 5 A comparison of State-of-the-art and DenTcov according to different criteria

State-of-the-art approaches Accuracy 3-Class (%) Accuracy 2-Class (%) Param (Million)
VGG-19 [6] 93.48 98.75 143
Xception + ResNet50V2 [9] 91.40 NA NA
Xception [12] 89.60 99.00 33
DarkNet [21] 87.02 98.08 1.1
DenTcov 96.52 99.00 8

medical applications. As a result, our model is a guiding decision whose performance

is comparable (sometimes even superior) to human experts, allowing for a timely
diagnosis. The model DenTcov proposed is higher compared with other references
(see Table 5).

5 Conclusion

COVID-19 grew worldwide in a short period of time resulting in a significant number

of patients who eventually died. The WHO reports that this can be prevented through
prompt diagnosis. DenTcov automates the detection of COVID-19 in chest X-rays
using Artificial intelligence techniques, involving complex structures with fewer
parameters that require less computing power with greater results. A combination of
transfer learning and data augmentation techniques are used to address the overfitting
issue. In the experiments, an accuracy of 96.52 and 99% for 3-class and 2-class,
respectively, were achieved with a high F1 of 98.62%, confirming the effectiveness
of the proposed DenTcov model. Although several methodologies were developed
to operate, the proposed model achieved the best results.

References

1. WHO Coronavirus Disease (COVID-19) Dashboard. https://covid19.who.int. Accessed 16 Oct

2020
2. Cherian T et al (2005) Standardized interpretation of paediatric chest radiographs for the
diagnosis of pneumonia in epidemiological studies. Bull World Health Organ 83:353–359
3. Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts HJWL (2018) Artificial intelligence
in radiology. Nat Rev Cancer 18:500–510
4. Kallianos K et al (2019) How far have we come? Artificial intelligence for chest radiograph
interpretation. Clin Radiol 74:338–345
5. Shin H-C et al (2016) Deep convolutional neural networks for computer-aided detection: CNN
architectures, dataset characteristics and transfer learning. IEEE Trans Med Imaging 35:1285–
1298
6. Apostolopoulos ID, Aznaouridis S, Tzani M (2020) Extracting possibly representative COVID-
19 Biomarkers from X-Ray images with deep learning approach and image data related to
pulmonary diseases. J Med Biol Eng 40:462–469
DenTcov: Deep Transfer Learning-Based Automatic Detection .. 977

7. Chowdhury MEH et al (2020) Can AI help in screening Viral and COVID-19 pneumonia?
IEEE Access 8:132665–132676
8. Farooq M, Hafeez A (2020) COVID-ResNet: A Deep Learning Framework for Screening of
COVID19 from Radiographs. arXiv:2003.14395 [cs, eess]
9. Rahimzadeh M, Attar A (2020) A modified deep convolutional neural network for detecting
COVID-19 and pneumonia from chest X-ray images based on the concatenation of Xception
and ResNet50V2. Inform Med Unlocked 19:100360
10. Narin A, Kaya C, Pamuk Z (2020) Automatic Detection of Coronavirus Disease (COVID-19)
Using X-ray Images and Deep Convolutional Neural Networks. arXiv:2003.10849 [cs, eess]
11. Asif S, Wenhui Y, Jin H, Tao Y, Jinhai S (2020) Classification of COVID-19 from Chest X-ray
images using Deep Convolutional Neural Networks. medRxiv 2020.05.01.20088211. https://
doi.org/10.1101/2020.05.01.20088211
12. Khan AI, Shah JL, Bhat M (2020) CoroNet: A deep neural network for detection and diagnosis
of COVID-19 from chest x-ray images. Comput Methods Programs Biomed 196:105581
13. Sethy PK, Behera SK, Ratha PK, Biswas P (2020) Detection of Coronavirus Disease (COVID-
19) Based on Deep Features and Support Vector Machine
14. Du SS, Koushik J, Singh A, Poczos B (2017) Hypothesis transfer learning via transformation
functions. In Guyon I, et al (eds) Advances in Neural Information Processing Systems, vol 30,
pp 574–584. Curran Associates, Inc.
15. Deng J, et al (2009) ImageNet: A large-scale hierarchical image database. In: 2009 IEEE
Conference on Computer Vision and Pattern Recognition 248–255 (2009). https://doi.org/10.
1109/CVPR.2009.5206848
16. Howard J, Gugger S (2020) Fastai: a layered API for deep learning. Information 11:108
17. Huang G, Liu Z, Van Der Maaten L, Weinberger KQ (2017) Densely Connected Convolutional
Networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
pp 2261–2269. https://doi.org/10.1109/CVPR.2017.243
18. Smith LN (2017) Cyclical Learning Rates for Training Neural Networks. arXiv:1506.01186
[cs]
19. Cohen JP (2020) ieee8023/covid-chestxray-dataset (2020). Accessed 02 Oct 2020
20. Chest X-Ray Images (Pneumonia) | Kaggle. https://www.kaggle.com/paultimothymooney/
chest-xray-pneumonia. Accessed 02 Oct 2020
21. Ozturk T et al (2020) Automated detection of COVID-19 cases using deep neural networks
with X-ray images. Comput Biol Med 121:103792
Network Technology
Explainable Deep Learning Model
for COVID-19 Screening in Chest
CT Images

Mostafa El Habib Daho , Amin Khouani, Mohammed El Amine Lazouni,

and Sidi Ahmed Mahmoudi

Abstract In this work, we proposed an Explainable model based on Deep Learning

for fast COVID-19 screening in chest CT images. We first collected a database of 360
COVID and Non-COVID images at the Tlemcen hospital in Algeria. This database
was merged with two other public datasets (the first one has been collected from
several articles published on medRxiv, bioRxiv, NEJM, JAMA, Lancet, etc. The
second one was obtained from infected cases in hospitals in Sao Paulo, Brazil). We
also conducted a comparative study between Deep Learning classification models
that are widely used in the state of the art such as VGG16, VGG19, Inception v3,
ResNet50, and DenseNet121. We also proposed an interpretable architecture based
on the ResNet50 model and the GradCam explanation algorithm. Experimentations
showed promising results and prove that the introduced model can be very useful for
the diagnosis and follow-up of patients with COVID-19.

Keywords COVID-19 · GradCam · Deep learning · Classification · Chest CT-scan

1 Introduction

COVID-19 is an infectious and contagious virus and, given its global magnitude, the
World Health Organization (WHO) has announced it as a pandemic. The declaration
of a pandemic also highlighted the deep concern about the dangerous speed of spread
and severity of COVID-19. To date (mid-November), the number of patients has
exceeded 54 million worldwide.
Worldwide, the most effective screening technique is Reverse Transcription Poly-
merase Chain Reaction (RT-PCR). Nevertheless, the poor sensitivity of RT-PCR
and the lack of test kits in pandemic regions increases the workload of screening,

M. El Habib Daho (B) · A. Khouani · M. E. A. Lazouni

Biomedical Engineering Laboratory, University of Tlemcen, Tlemcen, Algeria
e-mail: [email protected]
S. A. Mahmoudi
University of Mons, 20 Parc Sq., B7000 Mons, Belgium

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 981
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_89
982 M. El Habib Daho et al.

so numerous sick persons are not immediately confined [14] and this accelerates
the spread of COVID-19 as many infected subjects cannot obtain immediate medi-
cation. In this circumstance, it is important to use other diagnostic techniques to
better control the pandemic. Healthcare departments recommend the use of medical
imaging (computed tomography (CT) and chest X-Ray) as a complement to RT-
PCR. Recent studies have shown that CT has a high diagnostic and prognostic value
for COVID-19. Computed Tomography is an easy exam and quick to obtain at no
additional cost and has shown a much higher sensitivity than RT-PCR for the inves-
tigation of COVID-19 [9]. The CT-scan image can distinguish potential high-risk
subjects that are predisposed to become critical and require instant medical support.
With advances in artificial intelligence, machine learning (ML), and computer
vision, the build of smart prediction and diagnostic models has grown at an acceler-
ated velocity [11]. Improving these intelligent models for disease prediction and diag-
nosis requires the use of ML techniques [9, 14]. However, to get better ML models,
efficient and effective feature extraction techniques are needed. Deep learning (DL),
as an ML method, has given encouraging performances in facilitating the diagnosis
of lung diseases using CT-scan images of the chest [2, 13]. Deep learning, unlike the
classic ML, can extract relevant characteristics automatically without expert feature
engineering. In this work, we introduce a diagnostic support system for COVID-19
by classifying CT-scan images.
Since the appearance of the COVID-19 pandemic, several works on the classifi-
cation and analysis of scanner images have been carried out. Among these works,
researchers such as [16] and [1] have opted to use classic machine learning tech-
niques. In Barstugan et al. [1], the authors proposed several feature extraction tech-
niques based on the Grey Level (GL Co-occurrence Matrix, GL Run Length Matrix,
GL Size Zone Matrix), Discrete Wavelet Transform, and Local Directional Pattern.
The classification task was performed using SVM.
Other researchers have used deep learning techniques for this task. Among these
works, we find the work of Jin et al. [7], Xu et al. [22] and Wang et al. [21]. They
have used CNN models (Convolutional Neural Networks) as DL techniques for the
screening of COVID-19 using CT-scan images. Pathak et al. [12], Ying et al. [19]
and Li et al. [10] have performed a transfer learning on the ResNet-50 architecture.
Zheng et al. [24] and Jin et al. [8] have used U-Net and CNN as deep learning models
to achieve good performances.
Works cited above focus on classifying COVID-19 images using either machine
or deep learning techniques trained on public datasets. This work aims to propose an
explainable deep learning model for the classification of COVID-19 CT-scan images.
For that, we have collected a new CT-scan dataset for the CHU of Tlemcen in Algeria.
We have also proposed an explainable deep learning classifier based on the Resnet50
architecture and the GradCam algorithm.
Explainable Deep Learning Model for COVID-19 Screening in Chest CT Images 983

2 Methods

Deep learning methods have proven their effectiveness in several fields such as
Computer Vision (CV) and Natural Language Processing (NLP). Deep Convolu-
tional Neural Networks, or CNN for short, are the most popular networks for various
computer vision tasks. These models are reputed to be very good for visual recog-
nition since once a segment in a particular area of an image is learned, the CNN can
recognize that segment anywhere else in the image. Among the commonly used CNN
architectures for the classification of images, we can find VGG16, Resnet50, Incep-
tionV3, and Densenet. We have chosen to use these architectures in this work since
they have been very successful in the challenges of ImageNet and it was proven that
using these pre-trained models in other classification problems (transfer learning)
gives good results.

2.1 VGG

Simonyan and Zisserman [17] proposed the VGG architecture. This model is simple;
it uses only 3 × 3 stacked convolutional layers. The reduction of the convolution
matrix dimension is achieved by a pooling operation (Max Pooling). The fully
connected layers are the two last layers, containing 4,096 neurons each, followed
by a softmax function. VGG16 and VGG19 are the two variants of this architecture.
The “16” and “19” represent the number of layers in the model.

2.2 ResNet

As image recognition becomes more complex, Deep Learning training becomes very
difficult, since extra deep layers are needed to calculate and enhance the accuracy of
the model. Residual training is a concept created to solve this issue and the resulting
model is known as ResNet [5]. ResNet consists of several residual modules where
each module represents a layer. The depth of a ResNet can vary considerably, the
most used are Resnet50 [5] and the one developed by Microsoft which contained 152
layers [5]. Such depth was infeasible before the appearance of ResNet because of the
problem of vanishing gradients. Therefore, ResNet solves this problem by providing
skip connections, applied to each layer before the ReLu activation function, allowing
to preserve the gradient [4].
984 M. El Habib Daho et al.

2.3 Inception

The first version of this model was named GoogLeNet [20], but later variants were
simply denominated Inception. Google published successively new versions of this
architecture. This model uses some convolution filters (1 × 1, 3 × 3, and 5 × 5) in
the same inception module to extract multi-level features.

2.4 DenseNet

DenseNet [6] is composed of blocks called “dense”. In these blocks, each layer is
connected with all the previous layers. To summarize, the DenseNet architecture
makes maximum use of the residual mechanism by making each layer (in the same
dense block) connect to its following layers.

2.5 Proposed Method

The proposed model used the ResNet50 architecture as shown in the Fig. 1. We have
performed a transfer learning of the ResNet50 model using the ImageNet weights.
Then we added a GlobalAveragePooling layer and two Dense layers of size 1024
and 512. The last classification layer contains two neurons and uses the sigmoid
activation function. For better interpretability of the results, we used the GradCam
algorithm to explain the decision of the model.

Fig. 1 Our proposed method

Explainable Deep Learning Model for COVID-19 Screening in Chest CT Images 985

3 Results and Discussion

3.1 Dataset

In this study, we have used three different datasets; two of them are available for free
access on the internet and the third one we collected from the Hospital of Tlemcen
(CHU Tlemcen).
Yang et al. [23] published the first dataset; it contains 349 COVID and 397 non-
COVID images. This database has been collected from several articles published on
medRxiv, bioRxiv, NEJM, JAMA, Lancet, etc.
The second one is also public and has been published by Soares et al. [18], it
contains 1252 COVID and 1230 non-COVID images. This dataset is gathered from
cases inside hospitals in Sao Paulo, Brazil.
We have collected the third dataset from the Hospital of Tlemcen in Algeria; it
contains 180 COVID and 180 non-COVID images. For training and testing, we used
the three datasets with data augmentation. During the test, we validated the different
models on 355 COVID-19 and 363 Non-COVID images.
Figure 2 shows examples of images with and without the presence of COVID-
19 in the three datasets. On the figure, we can see that the three datasets present
variations in color and texture, which will allow the model to be more robust to the
difference in quality and type of the image presented during the test phase.

(a) Dataset 1_COVID (b) Dataset 2_COVID (c) Dataset 3_COVID

(d) Dataset 1_ NON-COVID (e) Dataset 2_NON-COVID (f) Dataset 3_ NON-COVID

Fig. 2 Images from different datasets

986 M. El Habib Daho et al.

Fig. 3 Block diagram of this work

3.2 Experiments

The experiments were performed on a local machine with a 4820k i7 processor,

56 GB RAM, an 8 GB GTX1070 GPU and Windows 10 operating system.
The block diagram of this work is presented in Fig. 3.
First, the database is pre-processed, it contains images of different sizes, and
we have readjusted it so that all images will be 224 × 224 in size. For a better
generalization, we made a data augmentation by applying random transformations
on the input images, several transformations have been performed such as rotation
and zoom with a probability of 0 to 1 to ensure the diversity of the data. During the
acquisition process, the input images can suffer important variations on the color
due to the sensor quality (Fig. 2). To solve this problem we have performed color
normalization, this process can eliminate variations in luminosity and contrast. After
the pre-processing step, a training process is carried out on each model, and then an
evaluation is made on a test set. We have trained all models using batches of size
32 for 50 epochs using the Stochastic Gradient Descent method to reduce the loss
function. In Fig. 6, the curves show that all models have learned well from the dataset
and suffer neither from over- nor under-fitting. This can be explained by the fact that
SGD, as explained in [3], contains the error, stays stable, and prevents overfitting
even without any regularization term.
From the confusion matrix in Fig. 4, we can evaluate the quality of each model
by measuring the accuracy, the sensitivity, the specificity and the F-score on the
validation set using the following equations:
Explainable Deep Learning Model for COVID-19 Screening in Chest CT Images 987

With:
True positive (TP): Prediction is COVID and the image is COVID.
True negative (TN): Prediction is Non COVID and the patient is healthy.
False positive (FP): Prediction is COVID and the patient is healthy.
False negative (FN): Prediction is Non COVID and the image is COVID.
Table 1 shows the obtained performances of different models. We can notice that
our approach gives better results compared to the other architectures on accuracy,
specificity, and F-score. For sensitivity, our model is ranked second after Inception
v3, which is good since it is first on the other metrics. In medical diagnostic aid in
general and decision support for COVID-19 in particular this metric is very impor-
tant. Indeed, high sensitivity is equivalent to a very low false-negative rate. False
negatives are patients who have been diagnosed as Non-COVID by the system but
are positive. Such an error can cause the death of the patient and does not have the

(a) VGG16 (b) VGG19 (c) DenseNet121

(d) Inception V3 (e) ResNet50 (f) Our Method

Fig. 4 Confusion matrix of different models

Table 1 Results
Accuracy % Sensitivity % Specificity % F-score %
VGG 16 96.10 96.90 95.31 96.08
VGG 19 95.96 94.92 96.96 95.87
DenseNet 121 97.35 97.18 97.52 97.32
Inception V3 97.49 97.74 97.24 97.47
ResNet 50 96.23 96.33 96.14 96.20
Our work 97.63 97.46 97.79 97.60
988 M. El Habib Daho et al.

Fig. 5 Explanation with heatmaps on COVID-19 images using Grad-CAM

(a) VGG16 (b) VGG19

(c) DenseNet121 (d) Inception V3

(e) ResNet50 (f) Our Method

Fig. 6 Train and validation accuracy of different models

same impact as a false positive error that can be quickly corrected with additional
tests. We can therefore deduce that our model and Inception V3 are the best in terms
of performance.
Explainable Deep Learning Model for COVID-19 Screening in Chest CT Images 989

To enhance the explainability of our architecture and to highlight the relevant

areas driving the decision of our model, we have adopted the Grad-CAM [15].
Without extra manual annotation, this heatmap is entirely produced by the DL model.
Network predictions are interpreted by generating heatmaps to visualize the most
representative areas of the image using gradient-weighted class activation maps.
Figure 5 presents the heatmap of the suspect regions for the COVID examples.
These heatmaps illustrate the fact that our algorithm focuses on infected areas while
neglecting regular regions as demonstrated in the three images in Fig. 5.

4 Conclusion

In this study, we introduced a benchmarking analysis and study on different models

for the classification of CT-scan images of COVID-19. Such a tool is necessary for
the management of the COVID-19 crisis given the worldwide shortage of RT-PCR
screening kits.
In this work, we have exploited the power of feature extraction in deep learning-
based methods. The lack of all deep learning models is the absence of explainability.
To overcome this problem, we used a heatmap, generated by the GradCam algorithm,
to visualize relevant regions in the images driving the final decision of the model.
The obtained results by our explainable model outperform those achieved by the
other techniques used in this work and prove that a machine learning model using
the CNN architecture allows us to properly classify COVID-19 images.

References

1. Barstugan M, Ozkaya U, Ozturk S (2020) Coronavirus (covid-19) classification using CT

images by machine learning methods
2. Hagerty JR, Stanley RJ, Almubarak HA, Lama N, Kasmi R, Guo P, Drugge RJ, Rabinovitz
HS, Oliviero M, Stoecker WV (2019) Deep learning and handcrafted method fusion: higher
diagnostic accuracy for melanoma dermoscopy images. IEEE J Biomed. Health Inf 23(4):1385–
1391
3. Hardt M, Recht B, Singer Y (2016) Train faster, generalize better: Stability of stochas- tic
gradient descent. In: International conference on machine learning, pp 1225–1234
4. He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition. In: 2016
IEEE conference on computer vision and pattern recognition (CVPR), pp 770–778
5. He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition. In:
Proceedings of the IEEE conference on computer vision and pattern recognition, pp 770–778
6. Huang G, Liu Z, Van Der Maaten L, Weinberger KQ (2017) Densely connected convolutional
networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition,
pp 4700–4708
7. Jin C, Chen W, Cao Y, Xu Z, Tan Z, Zhang X, Deng L, Zheng C, Zhou J, Shi H, Feng J (2020)
Development and evaluation of an AI system for covid-19 diagnosis. medRxiv. https://doi.org/
10.1101/2020.03.20.20039834
990 M. El Habib Daho et al.

8. Jin S, Wang B, Xu H, Luo C, Wei L, Zhao W, Hou X, Ma W, Xu Z, Zheng Z, Sun W, Lan L,

Zhang W, Mu X, Shi C, Wang Z, Lee J, Jin Z, Lin M, Jin H, Zhang L, Guo J, Zhao B, Ren
Z, Wang S, You Z, Dong J, Wang X, Wang J, Xu W (2020) Ai-assisted CT imaging analysis
for covid-19 screening: Building and deploying a medical AI system in four weeks. medRxiv.
https://doi.org/10.1101/2020.03.19.20039354.
9. Kaur M, Gianey HK, Singh D, Sabharwal M (2019) Multi-objective differential evolution based
random forest for e-health applications. Mod Phys Lett B 33(05):1950022
10. Li L, Qin L, Xu Z, Yin Y, Wang X, Kong B, Bai J, Lu Y, Fang Z, Song Q, Cao K, Liu D, Wang
G, Xu Q, Fang X, Zhang S, Xia J, Xia J (2020) Artificial intelligence distinguishes covid-19
from community acquired pneumonia on chest CT. Radiology 296(2) (2020). https://doi.org/
10.1148/radiol.2020200905
11. Mei X, et al (2020) Artificial intelligence–enabled rapid diagnosis of patients with covid-19.
Nat Med. https://doi.org/10.1038/s41591-020-0931-3
12. Pathak Y, Shukla P, Tiwari A, Stalin S, Singh S, Shukla P (2020) Deep transfer learning based
classification model for covid-19 disease. Ingenierie Rec. Biomed. https://doi.org/10.1016/j.
irbm.2020.05.003, h
13. Pezeshk A, Hamidian S, Petrick N, Sahiner B (2019) 3-D convolutional neural networks for
automatic detection of pulmonary nodules in chest CT. IEEE J Biomed Health Inf 23(5):2080–
2090
14. Kumar Shukla P, Kumar Shukla P, Sharma P, Rawat P, Samar J, Moriwal R, Kaur M (2020) Effi-
cient prediction of drug–drug interaction using deep learning models. IET Syst Biol 14(4):211–
216. https://doi.org/10.1049/iet-syb.2019.0116. IET Digital Library, https://digital-library.the
iet.org/content/journals/10.1049/iet-syb.2019.0116
15. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D (2017) Grad-CAM: visual
explanations from deep networks via gradient-based localization. In: 2017 IEEE international
conference on computer vision (ICCV), pp 618–626
16. Shi F, Xia L, Shan F, Wu D, Wei Y, Yuan H, Jiang H, Gao Y, Sui H, Shen D (2020) Large-
scale screening of covid-19 from community acquired pneumonia using infection size-aware
classification
17. Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556
18. Soares E, Angelov P, Biaso S, Higa Froes M, Kanda Abe D (2020) SARS-CoV-2 ct- scan
dataset: a large dataset of real patients CT scans for SARS-CoV-2 identification. https://doi.
org/10.1101/2020.04.24.20078584
19. Song Y, Zheng S, Li L, Zhang X, Zhang X, Huang Z, Chen J, Zhao H, Jie Y, Wang R, Chong Y,
Shen J, Zha Y, Yang Y (2020) Deep learning enables accurate diagnosis of novel coronavirus
(covid-19) with CT images. medRxiv. https://doi.org/10.1101/2020.02.23.20026930
20. Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, Erhan D, Vanhoucke V, Rabinovich
A (2015) Going deeper with convolutions. In: Proceedings of the IEEE conference on computer
vision and pattern recognition, pp 1–9
21. Wang S, Kang B, Ma J, Zeng X, Xiao M, Guo J, Cai M, Yang J, Li Y, Meng X, Xu B (2020)
A deep learning algorithm using CT images to screen for corona virus disease (covid-19).
medRxiv. https://doi.org/10.1101/2020.02.14.20023028
22. Xu X, Jiang X, Ma C, Du P, Li X, Lv S, Yu L, Chen Y, Su J, Lang G, Li Y, Zhao H, Xu K,
Ruan L, Wu W (2020) Deep learning system to screen coronavirus disease 2019 pneumonia
23. Yang X, He X, Zhao J, Zhang Y, Zhang S, Xie P (2020) Covid-ct-dataset: a CT scan dataset
about covid-19
24. Zheng C, Deng X, Fu Q, Zhou Q, Feng J, Ma H, Liu W, Wang X (2020) Deep learning-based
detection for covid-19 from chest Ct using weak label. medRxiv. https://doi.org/10.1101/2020.
03.12.20027185
A Novel Tool for Automating the Trace
File Analysis in Vehicular Ad Hoc
Networks with Multi-agents Method

Sanaa Achchab, Souad El Houssaini, Amal Tmiri, Souad Ajjaj,

and Mohammed-Alamine El Houssaini

Abstract Vehicular Ad hoc Networks (VANETs) are a nascent area of research

that has gained considerable attention in recent years due to their role in the design
of Intelligent Transportation Systems (ITS). NS-3 (Network Simulator 3) is one of
the most widely used discrete event simulators for the VANET Simulation. Trace
files, which capture events occurring in these networks, can be used to study the
performance. This article introduces a new tool for automating the analysis of trace
files in VANETs using the concept of a multi-agents system. This tool allows to
calculate the following performance metrics: throughput, packet delivery rate, loss
of packets, packets sent and packets received in a very short time for all different
file sizes. In addition, it facilitates and speeds up the entire task of analyzing a high
number of network simulations along with offering the possibility of storing all the
trace files and all the measurements generated.

Keywords VANET · Trace file · Multi-agents system · NS-3

1 Introduction

Vehicular Ad hoc Networks are a subclass of MANETS, and they are considered as
one of the components of Intelligent Transport Systems (ITS). These systems aim to
incorporate new information and communication technologies in the (transportation
field) to improve the safety and the comfort of the road users. A VANET is an ad hoc
type network. It can opportunistically use communications with infrastructures, thus

S. Achchab (B) · S. El Houssaini · A. Tmiri

Department of Computer Science, Faculty of Sciences, Chouaib Doukkali University, El Jadida,
Morocco
S. Ajjaj
ENSAM, Hassan II University, Casablanca, Morocco
M.-A. El Houssaini
ESEF, Chouaib Doukkali University, El Jadida, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 991
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_90
992 S. Achchab et al.

allowing access to other networks and also to the Internet. The various infrastructures
used within the framework of the VANETs are called “Roadside Units” (RSU). These
units can be traffic lights, parking meters, or any other terminal placed at the edge of
the road.
Network simulation [1, 2] in telecommunication and computer networks is a
technique in which the behavior of a real network is modeled using a simulator.
Network simulation tools [1, 2] are used to model real networks with great ease.
There are several network simulators (open source or commercial), which have
different functions [3], such as NS-3 [4], NS-2 [5], OPNET, QualNet, OMNeT,
J-Sim [6] and NetSim.
The NS-3 is one of the most used and popular simulators. The main objective of
this NS-3 project is to give network researchers an open simulation environment to
test their network protocols. After simulating a network using a network simulator
such as NS-3, a trace file is created which contains the events that occurred during the
simulation and based on these events, users study the network performance. There
are different types of files: trace file.tr and file. Pcap.
In this article, we have developed a new tool for automating the analysis of the
NS-3 trace file.tr in VANETs using multi-agent systems that automatically calculates
VANET performance parameters such as throughput, packet delivery rate, packet
loss, packets sent and packets received, it can also draw a graph. This tool calls Ns-3
Analyzer using Multi-Agents System (NsAMAS).
Our article is structured as follows Sect. 2 presents previous work. Section 3
presents the concept of agent and multi-agents system. Section 4 introduces
NsAMAS, presents its concepts and explains its architecture. In Sect. 5 we high-
light the most important and unique characteristics of NsAMAS. Section 6 explains
the uses and the obtained results. Finally, a conclusion and the future work are given
in Sect. 7.

2 Related Work

There are several analysis tools that have been developed to generate statistics on
network simulation performance metrics or parameters [7, 8]. For example, NS-2
Trace Analyzer JTrana, and Trace Graph. They use NS-2 trace files.
JTrana [9] is a NS-2 trace analyzer built from scratch in Java used to study wireless
network simulation using a graphical user interface. Its input is an NS-2 wireless
trace, and its output consists of many graphs or data sets of the network.
Trace Graph [10, 11] is another NS-2 trace file analyzer. This software offers
different analysis options. It is developed in MATLAB 6.0 [12]. This software also
allows the user to retrieve useful statistics from a specific NS-2 trace file.
NS-2 Trace Analyzer [13, 14] is a tool written in the C/C++ language. Similar
to the precedent tools. This tool does not generate graphs related to the recovered
statistical results.
A Novel Tool for Automating the Trace File Analysis ... 993

All the software mentioned above are designed for the analysis of the NS-2 trace
file. In addition, it does not allow the user to store the results of each trace file
analyzed in a database locally, so that they can reuse it without having to repeat the
analysis. To extract statistics from another previously analyzed trace file, the user
must download the trace file again. This task may be acceptable for a small trace file,
but this process can take a long time for simulations that generate large trace files.
In this article, we will try to resolve these issues.

3 Multi-Agents System (MAS)

The topic of multi-agents system (MAS), although not new, is currently a very active
research area. It is a specialty which is interested in collective behavior evoked by
the interaction between several autonomous and flexible entities called agents, these
interactions are manifested on cooperation, competition or coexistence between these
agents. We are going to present the key concepts needed by agents and multi-agents
system (MAS) to design a practical architecture for analyzing the NS-3 trace file.

3.1 Agent

The concept of agent is a whole concept born decades ago. In the literature, many
works have focused on agents and multi-agents system. Definitions differ from author
to author and from field of application to another. We can cite the most consensual
definitions that we have found [15] as follows:
– An agent is an autonomous entity or simply an independent entity, abstract or
real, able to act for himself and his environment and can communicate with other
agents in a multi-agents universe.
– An agent is an entity which presents the following characteristics: independence,
reactivity and anticipation and has a specific social capacity.

3.2 Multi-Agents System

Multi-Agents System (MAS) [16] is a system made up of agents of the same type
or not, which cooperate and interact in a coordinated manner in order to achieve a
common goal or objective in an environment of which they have the same perception
and point of view.
A multi-agents system (or MAS) is a system that composed of the following
components [16]:
– An environment E, is a space which generally has a criterion.
994 S. Achchab et al.

– A set of objects O. These objects are localized. For each one it is possible to assign
a position in E, at a precise moment. These objects can be perceived, destroyed
and modified by agents.
– A set A of agents, is specific objects which contain the active entities of the system.
– A set of relations R that bring things together.
– A set of Op operations allow O objects to be interpreted, produced, consumed,
transformed and manipulated by agents.

4 NsAMAS Overview

NsAMAS is a new tool for automating the analysis of NS-3 trace file.tr in VANETS
using multi-agents system. The objective of NsAMAS is to perform a practical and
automatic analysis of the .tr trace file produced by a simulation using graphical
interfaces.

4.1 Architecture

This software is built on three layers: a presentation layer, a business layer and a
database access layer, as shown in Fig. 1.
• Presentation Layer
This layer presents a graphical interface which allows the user to interact with the
software. This layer contains the master agent to read the file, extract the fields
and transfer the data to another agent. User requests move from the presentation
layer to the actual processing layer. It also provides data in table and graph form
to study the results.
• Treatment Layer
This layer processes the data received from the source and stores the results in
the database. It contains two agents, one for calculating metrics and the other for
drawing graphs. The primary processor is responsible for analyzing the trace files
and storing the results, which is one of the basic software processes. It is also
responsible for studying general simulation information’s and each simulation
node. The Charts agent is responsible for plotting all types of graphics.
• Database Access Layer
This layer is used to store analysis results in the local database or to retrieve
requested data. It also interacts with the agent which calculates the performance
parameters to store the results, and with the master agent to retrieve the data
and store the information for each trace file.tr. The database used in NsAMAS is
MySQL and contains a table for each trace file entered by the user.
A Novel Tool for Automating the Trace File Analysis ... 995

Fig. 1 NsAMAS architecture

4.2 Issues with the Design

One of the main objectives of NsAMAS is to facilitate the analysis of trace files.
Currently, NsAMAS focuses on NS-3.tr trace files.
The trace file in ASCII code contains 12 fields [2]. each file line begins with an
event (+, -, d, r) descriptor followed by the simulation time (in seconds) of that event,
and the from and to node. Then we have the type, the size of the packet (in bytes)
and Flags (ignore). The next field is the packet class, which can be used to identify
a particular TCP connection. The next two fields are the source address and the
destination address. Finally, the sequence number of the protocol and the identifier.
When a trace file.tr is sent as input, the agent reads the file, extracts the 12 fields
mentioned above for use in the processing phase, and finally stores the information
about this file in the database.

5 Trace File Analysis

In this part, we identify the most important characteristics of NsAMAS and we also
present the architecture and the use of this software.
996 S. Achchab et al.

Fig. 2 Simulation execution steps

5.1 Simulation Execution

The first step of the analysis is to download the NS-3 simulation.tr trace file. Then the
file will be read and the information processing will start using an agent. The data on
the produced documents will be recorded in the database at the end of the processing
and analysis step. The general idea of the system, for analyzing, processing and
storing the file information, is shown in Fig. 2.

5.2 Identification and Storage

In this step our software starts to extract 12 fields from the trace file which were
provided as input. Once the trace file has been selected, an agent first tries to extract
the fields, then another agent uses these fields to process and analyze the data. Once
the processing phase of the trace file is completed, the results must be stored in the
database. Each trace file has its own table.
A Novel Tool for Automating the Trace File Analysis ... 997

5.3 Statistics Calculation

In the trace file processing step, the NsAMAS can automatically calculate the general
simulation metrics for the.tr trace file using an agent and without any user intervention
and it calculates other information regarding the metrics for each node.

General Simulation Information

The general simulation data contains the file path and the general simulation time,
information on packets transferred or dropped during simulation, the recorded infor-
mation is the number of packets sent, the number of packets received and the number
of packets dropped.

Node Specific Information

After processing of the.tr trace file, the user of the simulation data can obtain informa-
tion corresponding to a particular node. This information contains statistics such as
TX Packets, RX Packets, Total dropped packet, Data received, Data sent, Throughput,
Goodput and Packets Delivery Ratio (PDR).

6 NsAMAS Usages and Results

In this section, we explain in detail the different functions of NsAMAS and also
describe the information that NsAMAS can retrieve such as general simulation infor-
mation, general information about each node and graphics. The Table 1 provides the
simulation parameters that we used in our experiment.

6.1 Simulation Performance Metrics Extraction

To measure the network performance metrics, using our NsAMAS trace file analyzer.
An agent calculates various performance parameters: the general simulation param-
eters and the specific parameters to each node (Fig. 3). As an example of these

Table 1 Simulation
Parameters Values
parameters
Version of simulator NS-3.27
Routing protocol DSR
Number of vehicles 50
Simulation time 120 s
Physical and MAC layer 802.11p
Packet size 1000 bytes
Mobility model Manhattan Grid
998 S. Achchab et al.

metrics: throughput, packet drop rate, good throughput, packet delivery rate (PDR),
packet loss, packet loss rate (PLR), normalized routing load (NRL) and concentra-
tion activity. These metrics are the most common requested by researchers and the
most useful for evaluating the performance of VANETs. User can also retrieve the
results of previously scanned trace files from a local database as shown in Fig. 4.

Fig. 3 Specific node metrics

Fig. 4 Selection of a file in the previously analyzed database

A Novel Tool for Automating the Trace File Analysis ... 999

6.2 Chart Plotting

At this point, the user can draw various graphs in the form of a curve or a histogram
as shown in Figs. 5 and 6. Once all parameters are set, graphics can be drawn for
general simulation or for a particular node. The graphics provided by this tool are

Fig. 5 Chart plotting

Fig. 6 Bar chart

1000 S. Achchab et al.

Fig. 7 Processing time

result for the trace file.tr

Fig. 8 File processing time

comparison

the packet delivery rate and the bit rate (bits/s) versus time. Finally, it is also possible
to export graphics as an image.

6.3 NsAMAS Results

In this part, we provide synchronization results related to the basic steps described in
Sect. 5, which contains the reading of the trace file, analyzing and finally storing the
data in the local database. We have proved that the software not only handles trace
file.tr, but it does so extremely quickly. We present the results in Fig. 7 regarding the
above processes.
Based on the above results, it is clear that our NsAMAS software operates even
with large trace files to reduce the processing time. And we notice that the analysis
time increases with the size of the file.
We also provide a comparison of NsAMAS with other popular simulation analysis
tools. These tools are illustrated in Sect. 2 and are some of the best-known tools
dealing with simulated trace files. One of their main drawbacks is the time spent in
analyzing the trace file. The comparison shows that our NsAMAS software gives
good results in processing time Fig 8.
A Novel Tool for Automating the Trace File Analysis ... 1001

7 Conclusion and Future Work

In this work, we presented a new tool labeled NsAMAS to automate the analysis of
VANETs simulation trace files, performed with NS-3 using the multi-agents system
concept. The main objectives of this tool are: offering a simple interface to facilitate
the processing of.tr trace files in a very short time for all file sizes. Furthermore,
displaying the results automatically in the form of tables and graphs to help students
and researchers studying the performance of VANETs under NS-3. Results can be
stored in the local database to reuse them without the re-analysis phase. Our NsAMAS
software provides good results in terms of processing time and analysis compared
to other popular simulation analysis tools.
Building on the success of this software, we intend to upgrade it to calculate other
metrics, such as transmission capacity and power consumption. We also intend to
use other big data tools to analyze the file and draw other types of diagrams.

References

1. Weingärtner E, vom Lehn H, Wehrle K (2009) A performance comparison of recent network

simulators. In: Proceedings of the IEEE international conference on communications 2009
(ICC 2009). IEEE, Dresden
2. Bouras C, Charalambides S, Drakoulelis M, Kioumourtzis G (2013) Simulation design and
execution the case of TRAFIL. In: International conference on simulation and modeling
methodologies, technologies and application
3. Siraj S, Gupta A, Badgujar R (2012) Network simulation tools survey. Int J Adv Res Comput
Commun Eng 1:199–206
4. https://www.nsnam.org/. Accessed 11 July 2020
5. Simulator Network 2, Ns-2. https://www.isi.edu/nsnam/ns. Accessed 4 Oct 2020
6. Dorathy I, Chandrasekaran M (2018) Simulation tools for mobile ad hoc networks: a survey. J
Appl Res Technol 16(5):437–445
7. https://www.nsnam.org/docs/manual/html/tracing.html. Accessed 10 July 2020
8. https://www.nsnam.org/docs/release/3.9/tutorial/tutorial_23.html. Accessed 10 July 2020
9. Qian H, Fang W (2008) Jtrana: a java-based NS2 wireless trace analyzer
10. Malek J, Nowak K (2003) Trace graph-data presentation system for network simulator NS.
In: Proceedings of the information systems – concepts, tools and applications (ISAT 2003),
Poland
11. Cicconetti C, Mingozzi E, Vallati C (2009) A 2 k· r factorial analysis tool for ns2measure.
In: Proceedings of the fourth international ICST conference on performance evaluation
methodologies and tools. ICST (Institute for Computer Sciences, Social-Informatics, p 4
12. The Mathworks, Inc. (2012) 3 Apple Hill Drive, Natick, MA 01760-2098, USA
13. Salleh AU, Ishak Z, Din NM, Jamaludin MZ (2006) Trace analyzer for NS-2. In: Proceedings
of the 4th student conference on research and development (SCOReD 2006), Shah Alam,
Selangor, Malaysia, 27–28 June 2008
14. Pujeri UR, Palaniswamy V (2015) Trace analyzer for NS3
15. Xie J, Liu C-C (2017) Multi-agent systems and their applications, 14 July 2017
16. Dorri A, Kanhere SS, Jurdak R (2018) Multi-agent systems: a survey
A Fuzzy On/Off Switching Strategy
for Green Cellular Networks

Soufiane Dahmani , Mohammed Gabli, Abdelhafid Serghini,

and El Bekkaye Mermri

Abstract In the current generation of cellular networks, practitioners and

researchers have shown a keen interest in green wireless communication due to
its capability to create eco-efficient networks. To adapt to the increase in traffic and
services for all mobile subscribers, Base stations (BSs ) and relay stations (RSs ) must
be deployed more and more in order to meet the growth in this demand. However,
increasing the number of BSs or RSs can increase energy consumption and reduce
efficiency as it is responsible for large carbon dioxide (CO2 ) emissions. In this paper,
we introduce a new approach to the management of the base stations and the relay
stations using a fuzzy dynamic distribution of the BSs and RSs in the center according
to the distribution of users. We use, secondly, a sleep mode activation of these stations
which is seen as the key to reducing grid power consumption. The performance and
the effectiveness of the proposed approach are clarified by a simulation example that
reveals the capacity of our strategy in reducing energy consumption.

Keywords Green optimization · Fuzzy logic · Base station and relay station
deployment · Energy consumption · On-off switching strategy

S. Dahmani (B) · A. Serghini

ANAA Research Team, ESTO, FSO, LANO Laboratory, University Mohammed Premier, Oujda,
Morocco
e-mail: [email protected]
A. Serghini
e-mail: [email protected]
M. Gabli
FSO, Department of Computer Science, University Mohammed Premier, Oujda, Morocco
e-mail: [email protected]
E. B. Mermri
FSO, Department of Mathematics, University Mohammed Premier, Oujda, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1003
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_91
1004 S. Dahmani et al.

1 Introduction

In recent years, mobile data traffic is growing rapidly and yields very high produc-
tivity. Wireless networks have become a challenge for operators in current and future
years. That’s why researchers are trying to develop the network every time and
everywhere.
A basic requirement for green wireless communications is low energy consump-
tion and low air pollution. As the number of mobile stations has swelled and seen the
needs of users, mobile operators have to increase the number of Base Stations (BSs )
or Relay Stations (RSs ), which implies an increase of the energy consumer in the
wireless cellular network. According to [1], the annual number of BSs has enhanced
by two million for the period from 2007 to 2012, which conducts to uncontrolled
production of carbon dioxide.
The BSs used to cover remote areas are mainly operated by diesel generators. The
annual operating cost of these power grids is often very high.
As pointed by [2], in 2002, the production of CO2 related to information and
communication technologies (ICT) reached 151 metric tons of carbon dioxide
(MtCO2 ), nearly half of this value is due to the wireless connection. CO2 emis-
sions are also expected to attain 349 metric tons in 2020. The authors in [3] detect
that information and communication technology (ICT) consumes approximately 600
Terawatt-hour (TWh) of electrical energy or three percent of the total worldwide elec-
trical energy, this number is expected to grow to 1700 TWh in 2030. These statis-
tics convince researchers around the world to develop new ways to reduce energy
consumption and reduce its negative impacts on the environment.
In the telecommunication system, the RSs are distributed with BSs to increase
the connectivity rate of these networks as shown in Ref. [4, 5]. Therefore the mobile
stations (MSs) can send their data directly to the base station (BS) or by relay stations
(RSs) to relay their data to the BS as shown in Fig. 1. This technique is called a
multi-hop relay network.
The CO2 emissions and energy consumption by RS are lower than those by BS.
Moreover, the installation is easier and faster and the maintenance costs are lower.
Unlike BS, RS does not require a direct connection to the core network [6]. The
use of RS in a cellular network in parallel with the use of BS clearly has its own
unique characteristics. However, incorrect deployment in unsuitable locations can
lead to lower energy efficiency, throughput, transmission delay, transmission loss,
power consumption, and increased cost of deployment expenses. In order to properly
utilize the advantages of RS and BS, an effective strategy is required to choose the
method and how to deploy BS and RS to green cellular networks, to achieve the
desired target and throughput, with a reasonable deployment budget, less energy
wastage, and most of all, minimal environmental pollution.
Regarding the deployment of base stations in an environment, there is a set of
deployment strategies such as randomly, or at the edges of cells as indicated in
[7, 8].
A Fuzzy On/Off Switching Strategy … 1005

Fig. 1 Example of the multi-hop relay networks model

Fig. 2 BS or RS deployment in area divided into grids

To install BS, the central position is chosen because it reduces the number of
stations to be installed and at the same time it covers a large number of mobile
stations with less disturbance [9]. Figure 2 shows an example of a base stations and
relay stations deployment site in area divided into grids.
1006 S. Dahmani et al.

Our first challenge, in this paper, is to propose an efficient placement scheme for
the deployment of BSs and RSs on a selected site taking into account the uncertainty
and imprecision of the coverage rate and traffic demand.
The second challenge is to minimize the power consumption of these stations by
using the on-off switching strategy while maintaining the network coverage as used
in our paper [10].
In Sect. 2 we describe the system model and problem formulation. Section 3
presents our proposed approach based on fuzzy logic and on-off switching strategy.
Section 4 shows the simulation Results and application. The conclusion is drawn in
the final Section.

2 System Description and Problem Formulation

We consider the network of base stations or relay stations in a space or position of

any size, with the random distribution of the users and variable traffic in a day. To
find the publishing site, the area is divided into networks of similar dimensions.

Step 1:
We divide the area into divers’ grids, the publishing site is already chosen as a focal
point. The decision to install a BS or a RS in each site depends on the number of
users deployed in each network, as shown in Fig. 3.

Fig. 3 BS or RS deployment in an area divided into grids depending on the number of mobile
stations (MSs ) in each grid.
A Fuzzy On/Off Switching Strategy … 1007

Fig. 4 Binary logic to decide to install a base station or a relay station.

In the literature, this deployment mainly depends on the threshold value of the
number of users which is calculated as follows:
T NU
Tr = (1)
T NG
where Tr is the threshold value, TNU is the total number of users and TNG is the
total number of grids.
The authors in [9] relied directly on this threshold to define the type of station
that will be installed. If the number of users (TNU) on each grid is higher than the
threshold value, BS is published. On the other hand, if this number is lower than the
threshold, the RS is displayed, as shown in Fig. 4.
In the existing model to define the number of users, the decision maker monitored
the traffic for a defined period (e.g. one month) and calculated the average value.
In our view, the average does not accurately reflect the user traffic. Therefore, we
consider that this value is an imprecise and uncertain value and not a constant one.

Step 2:
In the classic case, after the final publication of base stations and relay stations,
an installed station remains active all the time, even for periods when the traffic is
low, which leads to a loss of energy. The next step (challenge) of our work is the
implementation of the on-off switching approach for these stations to reduce energy
consumption (see [10]).
1008 S. Dahmani et al.

3 Proposed Approach Based on Fuzzy Logic and On-Off

Switching Strategy

3.1 Fuzzy Threshold

As we mentioned earlier, traffic is an ambiguous and uncertain value. To address this

situation, we provide a fuzzy logic model that considers each threshold value as an
imprecise one. In this case, the threshold value is given by:


r = T N U
T (2)
T NG


where, T N U is a fuzzy number with the following membership function:
  T NU −t

max0, 1 − α i f t ≤ T NU
μT (t)
NU max 0, 1 − t−T NU
β
if t > T NU

where the two values α and β are positive fixed values signifying respectively the
left and right deviations of the fuzzy numbers. The membership function μT NU (t)
is represented in Fig. 5.


Fig. 5 Illustration that represents the membership function of the fuzzy number T NU.
A Fuzzy On/Off Switching Strategy … 1009

3.2 Base Station and Relay Station On/Off Switching Strategy

Each BS or RS station has two main powers, the basic and transmitting energy. The
basic energy is small energy related to these stations when they are in an idle state,
while the transmission power is related to them when sending information to the
users and thus it is in an active state, this means that the power consumption can be
reduced by controlling the BS or RS mode. When BS or RS is not fully used or not
connected to the maximum number of users, it can be kept in standby or idle mode
to minimize power consumption.
We will rely on the same strategy used in our previous article [8], which is based
on the concept of traffic states, the reactive approach and the proactive approach.

Traffic States: Suppose that in a given area and for one day, the traffic is distributed
as follows (see Fig. 6).

Between 00 am and 10 am, the traffic does not exceed 450 GBytes;
Between 10 am and 2 pm, the traffic is between 450 and 900 GBytes;
Between 2 pm and 8 pm, the traffic exceeds 900 GBytes;
Between 8 pm and 00 am, the traffic is between 450 and 900 GBytes;
The maximum traffic during the day is 1200 GBytes.

From this data, traffic variations can be divided into three different states s1 , s2
and s3 where s1 = [0; 450], s2 = [450; 900], and s3 = [900; 1200].

Reactive Approach: We consider the highest traffic state. Next, we determine the
optimal set of BS for that state. As the traffic conditions decrease, we need to
deactivate a number of BS while still meeting the coverage requirements.

4 Simulation Results and Application

4.1 Data Description and Computational Results

In order to carry out our simulation study, we have considered a large network area
to be covered. The size of this area is 10 × 10 km2 divided into 40 grids. The number
of mobile stations (MS) to be covered is 1200. Fixed power consumption (P) at the
BS and RS is set, respectively, to 68.8 W and 20 W (see [11] for instance). For
traffic reports, we randomly generated this traffic over a day spanning one month
and we calculated the average traffic per each hour. We obtained three traffic states
as follows:

s1 = [0, 450], s2 = [450, 900], and s3 = [900, 1200].

1010 S. Dahmani et al.

Fig. 6 Representation of traffic variations over a day.

In the classical model, Tr = 120040

= 30. Therefore, if the number of users in a
grid is greater than 30, then a BS must be installed. Otherwise, we must install a RS.
Using our fuzzy approach with α = 1 and β = 1, we found that 32 BSs and 8 RSs
needed to be installed.
In the reactive approach, the spread of BSs or RSs is initially implemented based
on the highest traffic state s3 . Therefore, the 32 BSs and 8 RSs which define the
largest possible group of stations can be active at any time. Then, when the traffic
demand decreases, some base stations and relay stations of the primary stations are
deactivated. The output results for traffic states s2 are 20 BSs and 6 RSs and the
output results for s1 are 11 BSs and 5 RSs .

4.2 Analysis of CO2 Emissions

Base stations and relay stations consume the majority of energy in the cellular
network. Therefore, activating and deactivating these components will lead to signif-
icant energy savings and reduced CO2 emissions. Electrical energy is derived from
fuel oil, where each one kilowatt hour represents 620 g of carbon dioxide (see [12]).
The expression of carbon dioxide emissions for each traffic state is defined by:

Nsi × PB S/R S × Tsi × 620

C O 2 emissions[kg/day](si ) = (4)
106
A Fuzzy On/Off Switching Strategy … 1011

where,
CO2 (si ): The amount of CO2 produced by the traffic state si ;
Nsi : The number of BSs and RSs started from the traffic state si ;
PBS (W): The energy consumption of the BS;
PRS (W): The energy consumption of the RS;
Tsi (H): The duration of the traffic state.
In the classic approach, the decision makers decide to keep all the 40 BSs active
throughout the day, and then the total CO2 emission, in KgCO2 /Day, is

40 × 68.8 × 24 × 620
= 40.94976
106
When we use the switching on-off strategy considering only the BSs, the CO2
emissions for each traffic state is (Fig. 6).

40 × 68.8 × 6 × 620
C O 2 (s3 ) = = 10.2374
106
26 × 68.8 × 8 × 620
C O 2 (s2 ) = = 8.87244
106
16 × 68.8 × 10 × 620
C O 2 (s1 ) = = 6.82496
106

CO2 (Day) = 10.23744 + 8.872448 + 6.82496 = 25.934848

In our approach, after the fuzzy installation of relay stations in parallel with base
stations and the application of the reactive approach during the day, the total CO2
emission is:

32 × 68.8 × 6 × 620 8 × 20 × 6 × 620

C O 2 (s3 ) = + = 8.189952 + 0.5952 = 8.785152
106 106
20 × 68.8 × 8 × 620 6 × 20 × 8 × 620
C O 2 (s2 ) = + = 6.82496 + 0.5952 = 7.42016
106 106
11 × 68.8 × 10 × 620 5 × 68.8 × 10 × 620
C O 2 (s1 ) = + = 4.69216 + 0.62 = 5.31216
106 106

CO2 (Day) = 8.785152 + 7.42016 + 5.31216 = 21.517472

By comparing these results, we see that the CO2 emission is reduced by 17%,
which shows the efficiency of our approach.
1012 S. Dahmani et al.

5 Conclusion

In view of the notable environmental pollution in recent years, we have looked in

this article, at the issue of environmental protection which is a basic problem that
arises in cellular network systems. The main challenge here concerns improving
energy efficiency in order to reduce the emission of CO2 . Indeed, we have developed
a system that takes into account both BSs and RSs deployment using the possibility
theory concerns as well as the on-off switching strategy.
The effectiveness of our approach has been assessed taking into account energy
consumption and CO2 emissions. The results show that the fuzzy deployment of BSs
and RSs and the on-off switching strategy reduced the total power consumption in
the cellular network and reduced the CO2 emissions.
Future research in this area may focus on (i) including other objectives like
reducing the total cost of installing base stations, and (ii) to apply our approach
to realistic test data generated by a real case study.

References

1. Mclaughlin S, Grant PM, Thompson JS, Haas H, Laurenson DI, Khirallah C, Hou Y, Wang
R (2011) Techniques for improving cellular radio base station energy efficiency. IEEE Wirel
Commun 18:10–17
2. Alsharif MH, Nordin R, Ismail M (2013) Survey of green radio communications networks:
techniques and recent advances. J Comput Netw Commun
3. Humar I, Ge X, Xiang L, Jo M, Chen M, Zhang J (2011) Rethinking energy efficiency models
of cellular networks with embodied energy. IEEE Netw 25:40–49
4. Yang Y, Hu H, Xu J, Mao G (2009) Relay technologies for WiMAX and LTE-advanced mobile
systems. IEEE Commun Mag 47:100–105
5. Chen L, Huang Y, Xie F, Gao Y, Chu L, He H, Li Y, Liang F, Yuan Y (2013) Mobile relay in
LTE-advanced systems. IEEE Commun Mag 51:144–151
6. Arthi M, Arulmozhivarman P (2017) Power-aware fuzzy based joint base station and relay
station deployment scheme for green radio communication. Sustain Comput Informatics Syst
13:1–14
7. Son K, Oh E, Krishnamachari B (2011) Energy-aware hierarchical cell configuration: from
deployment to operation. In: 2011 IEEE conference on computer communications workshops
(infocom wkshps), pp 289–294 (2011)
8. Mohammed G, Soufiane D, El Bekkaye M, Abdelhafid S (2019) Optimization of multi-
objective and green LTE RNP problem. In: 2019 International Conference on Wireless Tech-
nologies, Embedded and Intelligent Systems, WITS 2019 (2019) https://doi.org/10.1109/
WITS.2019.8723811.
9. Ameer R, Rolly RM, Jacob J (2018) Dynamic base station or relay station deployment and
small cell On/Off strategy for green ultra dense networks. Procedia Comput Sci 143:827–834
10. Dahmani S, Gabli M, Mermri EB, Serghini A (2020) Optimization of green RNP problem for
LTE networks using possibility theory. Neural Comput Appl 32. https://doi.org/10.1007/s00
521-018-3943-x
11. Ratheesh R, Vetrivelan P (2019) Energy efficiency based on relay station deployment and sleep
mode activation of eNBs for 4G LTE-A network. Automatika 60:322–331
12. Koutitas G (2010) Low carbon network planning. In: 2010 European Wireless Conference
(EW), pp 411–417
Multi-band Planar-Inverted-F-Antenna
Design for WIFI WIMAX and WLAN
Applications

Asma Khabba, Layla Wakrim, Saida Ibnyaich, and Moha M’Rabet Hassani

Abstract In this paper we propose a new triple band Planar-Inverted-F-Antenna

(PIFA) to be employed for WIFI/WIMAX/WLAN applications. The multi-band oper-
ation is obtained by etching three L-shaped slots from the radiating patch. The first
band is centered at 2.47 GHz, the second band is centered at 3.51 GHz and the third
band is centered at 5.23 GHz. The radiating patch is mounted on the top of low cost
FR4-epoxy substrate acting as ground plane. The patch is connected to the ground
plane using a short circuit and powered with a feeding plate. The overall design has
an area of 41.66 × 50 mm2 . The proposed antenna is studied in terms of reflection
coefficient, impedance matching, radiation pattern and distribution current. While
the evolution steps has been studied to find the optimized version of the proposed
design.

Keywords PIFA · L-shaped slots · Multiband

1 Introduction

With the rapid development of modern wireless communication technology, multi-

band antennas have become one of the most important circuit elements and have
attracted great interest [1–3]. Meanwhile, the need of different services in mobile
communication terminals, has turned the antennas design towards multi-band struc-
tures as well as miniature structures [4]. Due to its high performance, simple structure
and low manufacturing costs, PIFA antenna is frequently used for mobile terminals.

A. Khabba (B) · L. Wakrim · S. Ibnyaich · M. M. Hassani

Instrumentation, Signals and Physical Systems (I2SP) Team, Faculty of Sciences Semlalia,
Cadi Ayyad University, Marrakesh, Morocco
S. Ibnyaich
e-mail: [email protected]
M. M. Hassani
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1013
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_92
1014 A. Khabba et al.

The multiband PIFA antenna has attracted great interest for the employment in
wireless devices, in addition its miniature dimensions are compatible with mobile
terminals and warrants versatile functions of a compact mobile handset [5]. Several
techniques are used to achieve multi-band performance such as slot insertion [6–8],
the addition of a parasitic element [9], the addition of a capacitive load [10].
In this manuscript, a novel triple-band PIFA antenna using L-shaped slots is
designed and studied using CST Microwave Studio software. The Proposed antenna
is designed in such manner to accommodate the three applications WIFI, WIMAX
and WLAN in single antenna. The antenna is marked with high gain at three resonant
frequencies. While the antenna evolution process is carried out to investigate the
effect of the etched slots on the resonant frequencies. In addition, the simulated
results concerning current distribution and radiation patterns are discussed.
The rest of this paper is organized as follows: Sect. 2 presents the antenna structure
and geometry. The antenna operating mechanism is discussed in Sect. 3. While Sect. 4
describes the simulating results including reflection coefficient, input impedance and
distribution current. The gain and radiation patterns are presented in Sect. 5. Finally,
Sect. 6 concludes the study.

2 Antenna Configuration and Geometry

The antenna geometry is presented in Fig. 1. The suggested design consists of a radi-
ating plate which is connected to the ground plane by a short circuit of width 2.5 mm
and powered with a plate of width 4.363 mm. The ground plane is made up using the
low cost FR4 epoxy substrate with dielectric constant of 4.3, loss tangent of tanδ =
0.018 and thickness of 1.2 mm. The dimension of the radiating plate, ground plane,
shorting plate and the powered plate is displayed in Fig. 1(a). While the dimensions
of the L-shaped slots is showed in Table 1.

3 Evolution Process of the Proposed Antenna

Figure 2 presents the stepwise geometries through which the proposed design has
been obtained. While the reflection coefficient of the antennas at the four stages is
presented in Fig. 3. As can be seen the design is started with a conventional PIFA
antenna which exhibits a single band operation where the resonant frequency is
obtained at 4.75 GHz with 850 MHz of bandwidth. The second step is came out
with the target to create a novel operating band. The second antenna in Fig. 2(b) is
obtained by inserting an L-shaped slot of size 19.365 × 5 mm2 in the radiating plate.
As can be seen in Fig. 3 the loaded slot leads to create a novel resonant frequency at
2.8 GHz, however the new resonant mode does not have an appropriate impedance
matching (S11 = −8 dB). The third step is performed by inserting a second L-shaped
slot of size 22.8 × 5 mm2 in the radiating plate (Fig. 2(c)). As remarked from Fig. 3,
Multi-band Planar-Inverted-F-Antenna Design … 1015

Fig. 1 Configuration of the proposed PIFA antenna

Table 1 Slots dimensions

Parameters Values (mm) Parameters Values (mm)
X1 1.671 X4 5
Y1 9.457 Y4 1
X2 8 X5 1
Y2 1.671 Y5 19.365
X3 1 X6 5
Y3 22.8 Y6 1

the effect of the second slot is presented in the creation of the third resonance at 2.3
GHz while the impedance matching does not reach an appropriate value. Finally, the
last step is done with the cutting of another L-shaped slot of size 9.457 × 8 mm2 .
It is clearly remarked that this cutting slot has a crucial role to adjust and improve
the impedance matching of the three resonant modes. Which makes it ready to be
employed for WIFI, WIMAX and WLAN standards. The details about the three
resonant modes is described in the next subsection.
1016 A. Khabba et al.

Fig. 2 Configuration of the proposed PIFA antenna a conventional antenna, b antenna 2 c Antenna
3, d proposed antenna

Fig. 3 Reflection coefficients of the stepwise geometries

4 Simulation Results

The reflection coefficient of the proposed PIFA antenna is presented in Fig. 4. As

depicted, the antenna exhibits three resonant frequencies. The first resonant frequency
Multi-band Planar-Inverted-F-Antenna Design … 1017

is obtained at 2.47 GHz (S11 = −16.08 dB) which corresponds to WIFI standard,
with 120 MHz of bandwidth from 2.41 GHz to 2.53 GHz. While, the second resonant
frequency is obtained at 3.51 GHz (S11 = −21.77 dB) which corresponds to WIMAX
standard with 40 MHz of bandwidth from 3.49 GHz to 3.53 GHz and the third
resonant frequency is obtained at 5.23 GHz (S11 = −32.11 dB) which corresponds
to WLAN standard with 390 MHz of bandwidth from 5.04 GHz to 5.43 GHz.
The input impedance of an antenna is expressed as Z in = Rin + j X in , where
Z in is the input impedance, Rin is the resistance (real part) and X in is the reactance
(imaginary part). In order to obtain a good impedance matching at the resonant
frequency the input impedance should be close to 50  which means the input
resistance Rin should be close to 50  and the input reactance X in should be close
to 0. The input impedance of the proposed PIFA antenna is presented in Fig. 5.

Fig. 4 Simulated reflection coefficient of the Suggested design

Fig. 5 Simulated input impedance Zin of the Suggested design

1018 A. Khabba et al.

Fig. 6 Current distribution at the resonant frequencies, a 2.47 GHz, b 3.51 GHz, c 5.23 GHz

As can be seen, the input resistance at the three resonant frequency is Rin is 39.15
, 54.74  and 48.01  respectively and the input reactance X in is −8.96 , 7.12
 and −1.4  respectively. It should be noted that the best impedance matching is
reached at the third resonance frequency where Rin is close to 50  and X in is close
to 0  which generates a low reflection coefficient of −32.11 dB.
To better understand the antenna behavior, the distribution of current at the three
resonant frequencies is depicted in Fig. 6. As depicted in Fig. 6(a) at the first resonant
frequency the maximum of current is concentrated at the horizontal part of the second
slot and a slight current intensity is concentrated in the horizontal part of the third
slots. While at the second and third resonant frequencies (Fig. 6b, c), it is remarked
that the maximum of current is intensified around the vertical and horizontal parts
of the second and third slots). Therefore, it can be concluded that the second and the
third slots are responsible to generate the resonant frequencies while the first slot
is responsible to enhance the impedance matching as well as tuning the resonance
frequencies values.
The 3D radiation patterns at 2.47 GHz, 3.51 GHz and 5.23GHz are presented
respectively in Fig. 7(a), (b) and (c), while the 2D radiation patterns at the two cutting
plan phi = 0° and phi = 90° are presented in Fig. 7(d), (e) and (f). As can be seen
the proposed antenna provides high radiation performance in term of gain where the
max gain value obtained at 2.47 GHz, 3.51 GHz and 5.23 GHz are respectively 4.09
dB, 5.45 dB and 8.5 dB while the maximum radiation is oriented toward broadside
direction (phi = 0°, theta = 0°) for the first and third resonant frequencies and toward
the positive part of x axis for the second resonance.
Multi-band Planar-Inverted-F-Antenna Design … 1019

5 Radiation Patterns

6 Conclusion

A triple band PIFA antenna covering WIFI/WIMAX/WLAN standards has been

presented for mobile communications. The antenna is characterized with an appro-
priate area which can be easily fitted inside wireless devices. The three slots position
and dimensions are chosen carefully to operate at the desired frequencies. In addi-
tion the antenna proves high radiation performance in terms of 3D far-field radiation
pattern at the three resonances while the 2D radiation pattern is computed to show
the directional radiation behavior.

Fig. 7 Current distribution at the resonant frequencies, a 2.47 GHz, b 3.51 GHz, c 5.23 GHz
1020 A. Khabba et al.

References

1. EL bakouchi RJ, Ghammaz A (2015) A quad-band compact PIFA operating in the

GSM1800/GSM1900/UMTS/LTE2300/LTE2500/2.4-GHz WLAN bands for mobile termi-
nals. In: The proceeding of the 2015 Third World Conference on Complex Systems (WCCS),
pp 1–4. IEEE, Piscataway
2. Mondal T, Samanta S, Ghatak R, Bhadra Chaudhuri SR (2015) A novel tri-band hexag-
onal microstrip patch antenna using modified sierpinski fractal for vehicular communication.
Progress in Electromagnetics Research C, vol 57, 25–34
3. Nashaat D, Elsadek, H, Ghali, H (2003) Dual-band reduced size PIFA antenna with U-slot for
Bluetooth and WLAN applications. In : IEEE Antennas and Propagation Society International
Symposium. Digest. Held in conjunction with: USNC/CNC/URSI North American Radio Sci
Meeting (Cat No 03CH37450), pp 962–965. IEEE
4. Manouare AZ, Ibnyaich S, Idrissi AE et al (2016) Miniaturized triple wideband CPW-fed
patch antenna with a defected ground structure for WLAN/WiMAX applications. J Microwaves
Optoelectron Electromagn Appl 15(3):157–169
5. Ben Ahmed M, Bouhorma M, Elouaai F, Mamouni A (2011) Low SAR planar antenna for
multi standard cellular phones. Eur Phys J Appl Phys 53:33604
6. Wakrim L, Ibnyaich S, Hassani Moha M (2017) Multiband operation and performance enhance-
ment of the PIFA antenna by using particle swarm optimization and overlapping method. Appl
Comput Intell Soft Comput
7. Shagar AC, Wahidabanu SD (2011) Novel wideband slot antenna having notch-band function
for 2.4 GHz WLAN and UWB applications. Int J Microwave Wirel Technol 3:451–458
8. Fregoli L, Peixeiro C (2004) Small Multi-band Planar Inverted-F Antenna for mobile
communication systems and WLAN/WPAN applications. URSI EMTS
9. Meshram MK, Animeh RK, Pimpale AT, Nikolova NK (2012) A novel quad-band diversity
antenna for LTE and W-Fi applications with high isolation. IEEE Trans Antennas Propag
60:4360–4371
10. Ibnyaich S, Ghammaz A, Hassani MM (2012) Planar inverted-F antenna with J-shaped slot and
parasitic element for ultra-wide band application. Int. J Microwave Wireless Technol 4:613–621
A New Approach for Evaluating
the Performance of AODV Routing
Protocol in VANETs Using
the Plackett-Burman Method
Souad Ajjaj, Souad El Houssaini, Mustapha Hain,
and Mohammed-Alamine El Houssaini

Abstract In order to evaluate the performance of routing protocols in Vehicular Ad

hoc Networks (VANETs), the common method is to simulate different scenarios by
varying one factor at a time. Hence, the number of simulations increases as the number
of factors rises. This method attains useful results, but it is demanding and time
consuming. In this paper, we present a new approach to address performance analysis
tasks in VANET routing protocols, by applying the experimental plan methodology
particularly the Plackett-Burman Method (PBM). The main objective of our work
is to build mathematical models that allow studying the effect of multiple factors
simultaneously with a minimum number of experiments. To implement our work,
we select NS-3 for the Network simulation and SUMO for generating the mobility
model. Additionally, we opt for AODV as a VANET routing protocol. We involve four
factors, which are the Black Hole Attack, the Worm Hole Attack, the Node Density
and the Number of Connections. The outcomes of our study show the relevancy of
our method in the performance analysis of VANET routing protocols.

Keywords Vehicular Ad hoc Network (VANET) · Experimental plan ·

Performance analysis · AODV · Plackett-Burman Method (PBM)

1 Introduction

Vehicular Ad hoc Networks (VANETs) are a special case of Mobile Ad hoc Networks
(MANETs), where the mobile nodes are intelligent vehicles. VANETs have particular

S. Ajjaj (B) · M. Hain

ENSAM, Hassan II University, Casablanca, Morocco
e-mail: [email protected]
S. El Houssaini
Department of Computer Science, Faculty of Sciences, Chouaib Doukkali University,
El Jadida, Morocco
M.-A. El Houssaini
ESEF, Chouaib Doukkali University, El Jadida, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1021
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_93
1022 S. Ajjaj et al.

characteristics such as the high mobility, the dynamic topology, the delay constraints
and the intermittent connectivity [1]. Thus, the routing in VANETs must be as effi-
cient and reliable as possible. In this context, the present work introduces a new
approach for studying the effect of factors on the performance of the routing process
in VANETs. Our proposed method will offer the possibility to simultaneously assess
the effect of numerous factors with fewer experiences, by applying the Plackett-
Burman method [2]. The aim of our new approach is to model mathematically the
effect of factors on variables called responses. To perform our model, we conduct
eight experiments integrating four independent factors namely the Black Hole Attack,
the Worm Hole Attack, the Node Density and the Number of Connections. Accord-
ingly, we select Throughput, Packet Loss Ratio, Average End to End delay and
Routing Overhead as response variables. To implement our work, experiences are
built using Network Simulator NS-3, SUMO and AODV routing protocol [3].
The remainder of this paper is structured as follows. In Sect. 2, we present AODV
routing protocol. In Sect. 3, we will discuss the related work. Section 4 then introduces
the Plackett-Burman method. Section 5 will deal with the details of our proposed
work. Analysis of the acquired results is detailed in Sect. 6. Finally, we conclude
with synthesis of the work achieved and future research directions.

2 Overview of AODV

Ad hoc On Demand Distance Vector (AODV) is a reactive protocol in which routes

are established only when required. The [3] details the working of AODV.
AODV uses the concept of sequence numbers in order to maintain newer and
fresher routes. When a source node needs to send data to a destination, two mech-
anisms are invoked: the route discovery and the route maintenance. In the route
discovery process, the source node seeks in its routing table if a valid route exists for
this destination, then it sends a RREP (Route Reply) message; otherwise, it broadcasts
a RREQ (Route Request) message to all its neighboring nodes. Each intermediate
node rebroadcasts the RREQ message until the message reaches the destination or an
intermediate node that has a valid route to the destination. Hence a RREP message
is sent in the reverse path. If the source node does not receive a response within a
specified timeout, it retransmits another RREQ message. This procedure is repeated
RREQ-RETRIES times before delivering an error message (RERR).In Route main-
tenance, AODV maintains routes as long as they are active, a route is active if data
packets periodically transit from source to the destination. AODV uses the HELLO
messages to check the activity of the routes.
A New Approach for Evaluating the Performance of AODV Routing Protocol … 1023

3 Related Work

Multiple researches have dealt with the performance analysis of AODV in VANETs.
For instance, the authors in [4] evaluated the performance of three VANET routing
protocols, AODV, DSR and OLSR under two scenarios: urban and highway, in terms
of Packet Delivery Ratio and End to End delay in different node density. They
conclude that AODV has better performances regarding to packet delivery ratio in
both scenarios, while DSR has better delay, whereas OLSR showed average perfor-
mances in both scenarios. In [5], the authors studied the effect of node density on
Throughput, Packet Delivery Ratio (PDR), End to End delay and Packet Loss Ratio
using DSDV, OLSR and AODV. The comparison shows that AODV gives better
results. [6], compared throughput and delay of AODV and DSR. The results show
that AODV has better Throughput than DSR but more delay. In [7], the researchers
simulated a VANET using NS2 and VanetMobiSim. They compared three routing
protocols DSDV, AOMDV and AODV in terms of Packet Delivery Ratio and End to
End Delay. They stated that reactive routing protocols provide good results in low
and middle node densities. In [8] a modified protocol of AODV for multi-path was
proposed to overcome the problem of AODV high routing overhead. The authors
of paper [9] studied the effect of Black Hole attack on the performance of AODV
and ZRP, in terms of average Throughput, Packet Delivery Ratio, End-to-End Delay,
Normalized Routing Load and Average Path Length. The findings show that the
performance of AODV is decreased when Black Hole attack is launched. [10], illus-
trates the vulnerability of AODV to Black Hole Attack in VANETs and propose a
new algorithm to secure it. In this paper [11], authors evaluated the impact of Worm
Hole Attack on AODV, in terms of Average Number of Hopes per Route, Average
Delay, Average Route Discovery time, Average Throughput, Average Retransmis-
sion Rate, Average Data Dropped, and Average Traffic received. They emphasize the
degradation of AODV under Worm Hole Attack.

4 The Plackett-Burman Method (PBM)

Experimental plan methodology is used to plan and structure the experiments in

order to extract maximum information with a minimum number of experiences. It
is widely employed in research and industries. The goal is to study the effect of
variables called factors on other variables called responses. In this paper, we adopt
the Plackett-Burman Method (PBM). PBM [2] allows studying up to N-1 factor in
a number N of experiments, for example, an 8-experiment plan can handle up to 7
Factors. The number N of experiments is a multiple of 4 but not a power of 2, i.e., 8,
12, 16, 20, etc.…
To implement our work using PBM, the steps listed below are required:
1. Identify the Factors and the Responses of the experimental Plan, for each Factor
two levels are defined: the low level coded (–1) and the high level coded (+1).
1024 S. Ajjaj et al.

2. Build the PB matrix of experiments: the table where columns represent the
Factors and rows indicate the number of experiments. The matrix contains only
+1 and −1.
3. Perform the matrix of effects denoted X. It is based on the matrices of
HADAMARD [12]. We build it by adding a column containing only +1 to
the left of the matrix of experiments. This matrix is then used to calculate the
coefficients of the mathematical model. X is such that:

XT X = nIn (1)

with In : identity matrix of order n, and X T is the transposed matrix of X.

4. Calculate the coefficients of the model mathematically expressed in the form:

ŷ = b0 + b1 x1 + b2 x2 + b3 x3 + · · · + bk xk (2)

Where ŷ is the response, b0 is the constant of the model, b1 , b2 , .. bk are the

coefficients of the variables x1 , x2 … xk This equation is written in a matrix form
as follows:

Y = BX (3)

Y is the matrix of experimental responses, B is the vector of coefficients, and X

is the matrix of effects. The coefficients are then determined by the method of
least squares [13]:
−1
B = (X T X ) (X T Y) (4)

X T : is the transposed matrix of X.

4. Calculate the main effect of each Factor, that is the difference between the
average of the response values where factor A is set at the high level (+1) and
the average of the response values where factor A is set at the low level (−1):


N 
N
EA = 2 × [ (yi) − (yi)]/N (5)
i =1 i =1
A = +1 A = −1

Where N is the number of experiments and y represent the values of the response.
A New Approach for Evaluating the Performance of AODV Routing Protocol … 1025

5 Proposed Work

One of the main constrictions in VANETs is security wherein the vulnerability mainly
originates from the wireless communication and the lack of infrastructure [1]. There-
fore, various types of attacks can be launched in the network, e.g., the Black hole
Attack and the Wormhole Attack. In the Black hole Attack, the malicious vehicle
falsely pretends to have a short and fresh route to the destination and sends a fake
route reply with a high sequence number to the source node. The source node begins
the data transmission, believing that the optimal path is through this malicious node.
The malicious node then captures all routing packets and drops them. Worm Hole
Attack is carried out in a cooperative way, where two colluding nodes create a virtual
tunnel between them (at remote location) and let other nodes think that the routing
is faster via them. Therefore, fake routes are created resulting in a virtual traffic in
the network. The goal is to change the network topology and confuse the routing
process.
Other constraints influence the routing in VANETs. The findings in [7] and [14]
show that the routing performances vary considerably in terms of density and the
number of connections.
In our study we choose to study the effect of four factors: the Black Hole Attack,
the Worm Hole Attack, the Node Density and the Number of Connections.
For each factor, we define the levels low (−1) and high (+1) (Table 1).
For Black Hole and Worm Hole attacks, we use logical values (0 for level −1 and
1 for level +1). Node Density varies from 50 to 150 vehicles, whereas Number of
Connections varies from 4% of the total number of vehicles in level −1, and 20% in
level +1.
In our study, we select Throughput (y1), Packet Loss Ratio (y2), Average End to
End Delay (y3) and Routing Overhead (y4) as response variables.
1. Throughput (Th) is the total number of bits successfully received during
simulation time, it is measured in kilobits per second (kbps).
2. Packet Loss Ratio (PLR) represents the ratio of the number of lost packets to
the total number of sent packets.
3. End to End Delay (E2E delay) is the average time needed to transmit data from
source to destination, it is measured in (ms).
4. Routing overhead (RO) is the total number of control packets generated by
AODV routing protocol during route discovery and route maintenance.

Table 1 The values for levels (+) and (–) of all factors
N Factor High level (+1) Low Level (−1) Factor code
1 Black hole attack 1 0 X1
2 Wormhole attack 1 0 X2
3 Node density 150 50 X3
4 Number of connections 20% 4% X4
1026 S. Ajjaj et al.

Table 2 Simulation parameters in NS3

N Parameter Value
1 Network simulator NS3.29
2 Mobility simulator SUMO-0.32.0
3 Propagation model TwoRayGroundPropagationLossModel
4 Wifi channel YansWifi
5 Mac and physic layer IEEE 802.11p
6 Simulation time 200 s
7 Packet size 1024
8 Routing protocol AODV

Table 3 The matrix of the effects with experiment responses

N° X0 X1 X2 X3 X4 Y1(TH) Y2(PLR) Y3(delay) Y4(Overhead)
1 +1 +1 −1 – 1 +1 22,459 55,115 72,678 26227,000
2 +1 +1 +1 −1 −1 18,970 62,090 5,090 13046,000
3 +1 +1 +1 +1 −1 10,993 78,033 39,931 1082724,000
4 +1 −1 +1 +1 +1 17,282 65,465 126,736 1428075,000
5 +1 +1 −1 +1 +1 15,964 68,093 76,912 541073,000
6 +1 −1 +1 −1 +1 19,666 60,697 121,866 81667,000
7 +1 −1 −1 +1 −1 25,474 49,085 58,955 259989,000
8 +1 −1 −1 −1 −1 30,886 38,279 41,685 11356,000

In this study, A simulation map of El Jadida city Morocco is imported from Open
Street Map (OSM). Simulation parameters are described in (Table 2).

6 Experimental Results

6.1 The Matrix of the Effects

6.2 Coefficients of the Mathematical Model

The coefficients of the mathematical model are calculated by applying the method
of least squares. The calculations are done using a spreadsheet. The resulting model
for each Response ŷ is presented below:

ŷ1 (throughput) = 34,4312 − 6,2302x1 − 6,9679x2 − 0,0557x3 − 17,1105x4 (6)

A New Approach for Evaluating the Performance of AODV Routing Protocol … 1027

ŷ2 (PLR) = 31,1904 + 12,4515x1 + 13,9283x2 + 0,1112x3 + 34,1914x4 (7)

ŷ3 (EED) = 19,2327 − 38,6577x1 + 10,8485x2 + 0,1530x3 + 394,5790x4 (8)

ŷ4 (RO) = −703589,188 − 29504,25x1 + 441716,75x2 + 7948,9125x3 + 1109260,94x4

(9)

where: x1, x2, x3 and x4 are respectively the factors of Black Hole Attack (BKH),
Worm Hole Attack (WMH), Node Density (ND), and Number of Connections (NC).

6.3 Graphical Representation of the Results

Plots of Effects
Figure 1 illustrates the effect of all Factors on the AODV throughput. We notice that
the throughput decreases in + l level of all Factors. For Worm Hole and Black Hole
Attacks, this decrease is due to the presence of malicious nodes that try either to drop
packets in case of Black Hole Attack or capture and tunnel the packets to a distant
location in case of Worm Hole Attack. As a result, the number of successfully received
packets reduces causing the degradation of the throughput. AODV Throughput also
cuts down with the increase in Node Density and Number of Connections.
According to Placket Burman Method, we can rank these factors based on the
average effects. The effect of Worm Hole Attack comes first, then the Black Hole
Attack, the Node Density and finally the Number of Connections. The results

Fig. 1 Plot of main effects

on throughput

Fig. 2 Plot of main effects

on PLR
1028 S. Ajjaj et al.

achieved using PBM agree with several works in literature. For example, the authors
in [5] reported that the throughput decreases under the effect of Node Density. The
same conclusion is achieved in [14] regarding to the Number of Connections. Further-
more, the study carried out in [10] confirmed that AODV throughput is better in
scenarios free of routing security attacks.
Examining Fig. 2 related to the effects on AODV Packet Loss Ratio. It is clear
that the four Factors have a positive effect in their +1 level. Meaning that, the Packet
Loss Ratio increases in the presence of Black Hole and Worm Hole Attacks as well
as when the network is dense or the Number of Connections is high. According
to PBM, the most influencing factors are respectively Worm Hole and Black Hole
Attacks. Node Density comes in third position and Number of Connections in the
last rank.
The outcomes revealed by Plackett-Burman Method are in line with other
researcher’s study. For instance, the results in paper [5] emphasize that the number
of lost packets grows proportionally with the number of nodes. Additionally, [15]
stated that the Worm Hole Attack is responsible for rising AODV packet loss. In [9]
the conclusion is that AODV performance significantly degrades under Black Hole.
Figure 3 indicates the effect on End To End delay. We notice that the effect of
Black Hole Attack is negative unlike the three other factors with positive effect. In
other words, the End To End delay decreases in the level +1 of Black Hole Attack,
while it increases in the level +1 of the other factors. This rise is very important for
Number of Connections, while it slightly grows for Worm Hole Attack and Node
Density.
The reason is that AODV uses the mechanism of route discovery before sending
data packets. As the number of connections increases, the route search time will also

Fig. 3 Plot of Main Effects

on E2E delay

Fig. 4 Plot of Main Effects

on Overhead
A New Approach for Evaluating the Performance of AODV Routing Protocol … 1029

increase leading to high values of delay. Additionally, in Worm Hole Attack, packets
are tunneled through the malicious nodes that are located at remote positions; the
End To End Delay then becomes higher.
The results of PBM show that AODV performance decrease in the presence of
Worm Hole Attack, also in high dense networks and increased Number of Connec-
tions. The exception found is that the End To End Delay for Black Hole Attack is
reduced when the attack is launched. This is explained by the behavior of Black Hole
node that sends fake and fast Route Reply (RREP) to source node, this significantly
decreases both the route discovery process and the delay. This conclusion is compat-
ible with [9] where it is concluded that delay under Black Hole is better than in a
normal scenario. [4] Confirms the increase of AODV End To End delay with the rise
in the number of nodes and number of connections.
The authors in [16] studied the Worm Hole delay and proposed an improved
AODV to preserve the network from this attack.
The results of the effect on Routing Overhead are shown in Fig. 4. We observe
that the Routing Overhead increases in +1 level of all the Factors except for Black
Hole Attack that shows insignificant effect. This result copes with the study in [17],
which proved that the Routing Overhead decreases under Black Hole Attack because
malicious node sends a fake RREP as fast as possible. Consequently the Overhead
decreases because the route discovery packets are reduced.
We also note that Node Density has the most significant effect, followed by
Worm Hole Attack and finally the Number of Connections. This rise in Overhead is
caused by the large number of Route Request messages generated by AODV, mainly
in VANET environment qualified by high mobility and frequent connections and
disconnections of nodes.
The outcomes of our study match with several previous studies. [8] stated that
the Routing Overhead of On-Demand routing protocols rises significantly when
the number of nodes increases. In [18] the researchers emphasized the importance
of implementing new mechanisms to reduce AODV control messages overhead.
The authors developed two improved versions of AODV protocol (AODV-LAR and
AODVLine) to overcome the issue of high routing over-head. The authors in [16]
confirm the negative impact of Worm Hole Attack on AODV overhead. They also
presented a new approach to reduce AODV Overhead, and secure it against Worm
Hole Attack.

Plots for Model Validation

The plots below are used to check whether the experiment values of the responses fit
with the estimated responses (from the mathematical model). X-axis represents the
experimental values and Y-axis shows the estimated values of the responses (Figs. 5,
6, 7 and 8).
From the above figures, we can say that the values of the responses estimated by
our model fit perfectly with those experimentally obtained for: Throughput, Packet
Loss Ratio and End To End delay. Otherwise, the Routing Overhead shows acceptable
correlation between experimental and estimated results.
1030 S. Ajjaj et al.

Fig. 5 Experiment Vs
Estimated throughput

Fig. 6 Experiment Vs
Estimated PLR

Fig. 7 Experiment Vs
Estimated E2E delay

Fig. 8 Experiment Vs
Estimated overhead
A New Approach for Evaluating the Performance of AODV Routing Protocol … 1031

7 Conclusion

In this paper, we presented a new approach for the performance analysis of routing
protocols in Vehicular Ad hoc networks. As per our humble knowledge, our proposed
work unprecedentedly applied the Plackett-Burman Method, in which the findings
eventually cope with previous studies in the same area. This method allows to simul-
taneously evaluate the effect of variables called factors on other variables called
responses through a few number of experiences. Moreover we performed a mathemat-
ical model that describes the relationship between each response and the coefficients
of the factors.
In this study, we investigated the effect of four factors namely the Black Hole
Attack, the Worm Hole Attack, the Node Density and the Number of Connections on
the following performance metrics of AODV routing protocol: Throughput, Packet
Loss Ratio, End To End Delay and Routing Overhead. The achieved results demon-
strate the degradation of AODV performances when the attacks are launched in the
networks, as well as when the network is dense with high number of communica-
tions. The present work can be extended to appraise the effect of more factors on the
performance of other VANET routing protocols.

References

1. Arif M, Wang G, Zakirul Alam Bhuiyan M, Wang T, Chen J (2019) A survey on security attacks
in VANETs: Communication, applications and challenges. Veh Commun 19:100179
2. Plackett RL, Burman JP (1946) The design of optimum multifactorial experiments. Biometrika
33(4):305–325. https://doi.org/10.1093/biomet/33.4.305lastaccessed2020/10/26
3. Perkins C, Belding-Royer E, Das S (2003) Ad hoc On-Demand Distance Vector (AODV)
Routing. RFC 3561, RFC Editor https://www.rfc-editor.org/rfc/rfc3561.txt. Accessed 26 Oct
2020
4. Pranav K, Kapang L, Tuithung T (2011) Simulation based analysis of Adhoc routing protocol
in urban and highway scenario of VANET. Int J Comput Appl 12(10):42–49
5. Gupta P, Chaba Y (2014) Performance analysis of routing protocols in vehicular ad hoc
networks for CBR applications over UDP connections. Int J Eng Comput Sci 3(6)
6. Ghori MR, Sadiq AS, Ghani A (2018) VANET routing protocols: review, implementation and
analysis. J Phys Conf Ser 1049:012064
7. Vidhale B, Dorle SS (2011) Performance analysis of routing protocols in realistic environment
for vehicular ad hoc networks. In: 2011 21st International Conference on Systems Engineering,
pp 267–272
8. Hu S, Jia Y, She C (2017) Performance analysis of VANET routing protocols and implemen-
tation of a VANET terminal. In: 2017 International Conference on Computer Technology,
Electronics and Communication (ICCTEC), pp 1248–1252
9. Purohit KC, Dimri SC, Jasola S (2017) Mitigation and performance analysis of routing proto-
cols under black-hole attack in vehicular ad-hoc network (VANET). Wireless PersCommun
97:5099–5114
10. Tyagi P, Dembla D (2017) Performance analysis and implementation of proposed mechanism
for detection and prevention of security attacks in routing protocols of vehicular ad-hoc network
(VANET). Egyptian Inf J 18:133–139
1032 S. Ajjaj et al.

11. Ojha M, Kushwah RS (2014) Impact and performance analysis of WORMHOLE attack on
AODV in MANET using NS2. Int J Sci Res (IJSR) 3(6):1823–1826
12. Kharaghani H, Tayfeh-rezaie B (2005) A Hadamard matrix of order 428. J Comb Des 13
13. Dodge Y, Rousson V (1993) Alternative Methods of Regression. A Wiley-Interscience
Publication
14. Spaho E, Ikeda M, Barolli L, Xhafa F (2013) Performance comparison of OLSR and AODV
protocols in a VANET crossroad scenario. In Park JJ, Jong H, Barolli L, Xhafa F, Jeong HY
(eds) Information Technology Convergence, Lecture Notes in Electrical Engineering. Springer
Netherlands, Dordrecht, pp 37–45 (2013)
15. Narayanan SS, Murugaboopathi G (2018) Modified secure AODV protocol to prevent
wormhole attack in MANET. Concurr Comput Pract Exp 32
16. Safi SM, Movaghar A, Mohammadizadeh M (2009) A novel approach for avoiding wormhole
attacks in VANET. In: 2009 First Asian Himalayas International Conference on Internet, pp
1–6
17. Chavan AA, Kurule DS, Dere PU (2016) Performance Analysis of AODV and DSDV Routing
Protocol in MANET and Modifications in AODV against Black Hole Attack. Procedia Comput
Sci 79:835–844. Proceedings of International Conference on Communication, Computing and
Virtualization (ICCCV)
18. Ayash M, Mikki M, Yim K (2012) Improved AODV routing protocol to cope with high overhead
in high mobility MANETs. In 2012 Sixth International Conference on Innovative Mobile and
Internet Services in Ubiquitous Computing, pp 244–251
Design and Simulation of a New Dual
Band Fractal Antenna for TM
Applications

Abdelhakim Moutaouakil, Younes Jabrane, Abdelati Reha,

Abdelaziz Koumina, and Nadia El Makoussi

Abstract In this paper, we present the three first iterations design of a new dual-band
antenna based on U fractal geometry feeding by microstrip inset feed. The proposed
antenna is printed on FR4 substrate with a dielectric constant of 4.4. At the third
iteration, the studied antenna has a dual-band behavior with tow resonant frequencies:
2.25 and 3.29 GHz with a good impedance matching. The simulated results performed
by CADFEKO a Method of Moments (MoM) based Solver. The antenna offers low
profile, low coast and a very light weight which satisfy the requirement of MicroSat
Telemetry (TM) applications.

Keywords Fractal antenna · Dual-band · TM

1 Introduction

Interest in multi-band antennas is growing, in particular with the aim of reducing the
number of onboard and ground antennas and by combining several applications on the
same antenna. In particular, the increasing diffusion of radio navigation systems such
as GPS (Global Positioning System) has made it possible to flood the world market
with a large number of new receivers combining good performance, small size and
reduced consumption [1–3]. In addition, the deployment of the new radio navigation
system European Galileo (planned for 2013) and its combination with GPS will
allow the implementation of more reliable, high-precision applications. Thus, the
development of new antenna architectures combining two or more radio navigation

A. Moutaouakil (B) · Y. Jabrane · N. El Makoussi

Modelisation of Complex Systems Laboratory, Cadi Ayyad University, Marrakech, Morocco
A. Reha
Laboratory of Innovation in Management and Engineering for the Enterprise, ISGA, Marrakesh,
Morocco
A. Koumina
Nanostructures Physics Laboratory, ENS, Cadi Ayyad University
Marrakech, Marrakech, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1033
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_94
1034 A. Moutaouakil et al.

systems is in full swing. Indeed, most satellite communications systems, whether

geostationary, in low or elliptical orbits, require circularly polarized antennas. As
part of the new Galileo satellite navigation system, the CNES Antennas service
offered to analyze the possibility of combining the GPS, Galileo and Telemetry
(TM) applications of the Microsoft family of satellites [4].
In this paper, an original topology of compact multi-band antennas is presented
the form of a U-shaped fractal. Fractal antennas elements are currently beneficial to
antenna designers & researchers. The concept of fractal geometry was introduced for
the first time by Mandelbrot in [5]. Generally, using fractal geometries in antennas
tends to miniaturize their physical sizes and produce multiband response.

2 Antenna Design

Initially, we designed a rectangular patch antenna fed by a microstrip line by adding

at the level of the ground plane a rectangular slot of width Wp+10 and length Lp+8,
while respecting the different parameters for each antenna elements.
The patch antenna is designed using FEKO Solver. Here, we use FR4 epoxy as
the substrate (εr = 4.4), The parameters characterizing our patch antenna (Fig. 1) are
presented in Table 1 [6, 7]:
The previous rectangular patch antenna Fig. 1 will be modified based on the addi-
tion of a fractal U-shape. Applying fractal theory for antenna design is an interesting
method of improving antenna performance thanks to the property of self-similarity
which characterizes fractal geometries [8–10]. As shown in Fig. 2, the first iteration
replaces the horizontal segment by 5 segments reduced to two thirds of the initial
length, which means that the fractal dimension is: [11].

ln(N ) ln(5)
d=   = = 1.464
ln 1 R ln(3)

Fig. 1 Geometry of the patch antenna. a Top view b Bottom view

Design and Simulation of a New Dual Band … 1035

Table 1 Design parameters

Design parameters Value (mm)
& corresponding values
Length of substrate Ls 56.02
Width of substrate Ws 72.44
Length of patch Lp 28.01
Width of patch Wp 36.22
Thickness of the substrate h 1.6
Width of supply line Wf 2.8
Length of inset point 1

Fig. 2 Iteration steps to get Carpet geometry

3 Simulation Results

The following figures represent a comparison of the reflection coefficient versus

frequencies for different iteration numbers in the band [1–4 GHz]. We note that
when the number of iterations increases the first band becomes larger with good
adaptation (Fig. 3).
It is worth noting from the simulation that the resonant frequency increases with
the number of iterations increases. We also notice a widening of the bandwidth in
the third iteration equal to 499 MHz. The adaptation to the second band is improved
at the second iteration, but decrease at the third iteration.
The performances of a linearly polarized antenna are described in terms of its
main E and H plans. The E-plan contains the electric field vector while the H-plan
contains the magnetic field vector in the direction of the maximum radiation [9].
Figure 4 shows the simulated radiation pattern of the third iteration antenna
its resonance frequencies. From the simulation, we see that our antenna has a
bi-directional radiation pattern for the two resonant frequencies.
The Table 2 below summarizes some results on the fractal antenna performance.
1036 A. Moutaouakil et al.

Fig. 3 Simulated reflected coefficient (S11) versus frequencies for three first iterations

Fig. 4 Simulated radiation pattern in E-plan and H-plan at: a 2.25 GHz b 3.29 GHz

Table 2 Summary of results

Iterations Resonance Gain (dB) Bandwidth (MHz)
frequency’s (GHZ)
0 2.01 3.7 [1830–2201 MHz]
3.29 1.20 [3125–3485 MHz]
1 2.02 3.78 [1814–2206 MHz]
3.27 1.72 [3135–3409 MHz]
2 2.03 4.69 [1834–2228 MHz]
3.14 2.54 [2992–3406 MHz]
3 2.25 7.08 [1887–2386 MHz]
3.29 3.8 [3013–3486 MHz]
Design and Simulation of a New Dual Band … 1037

4 Conclusions

In this work, the setup of U-shape fractal on the patch antenna allows a behavior
multiband. The simulated results performed by CADFEKO a MoM based Solver are
encouraging. We see that with the development of this antenna in terms of form, it
can be used for the MicroSat Telemetry (TM) application, which is characterized
by a resonance frequency equal to 2.245 GHz, and a bandwidth wide [2200–2290]
MHz.

References

1. Kumar D, Singh RB, Kaur R (2019) Global positioning system. In: Spatial Information
Technology for Sustainable Development Goals. Springer, Cham
2. Hadj Abderrahmane L, Brahimi A A new design of dual band fractal antenna for LEO applica-
tions. The Eighth International Conference on Systems and Networks Communications. ISBN
978-1-61208-305-6
3. Wibisono G, Wildan M, Wahyudi J, et al (2020) Co-design structure of dual-band LNA and
dual-band BPF for radio navigation aid application. Wireless Pers Commun
4. Hebib S, Aubert H, Pascal O, Fonseca N, Marc Lopez J (2007) Antennes multi-bandes pour
application GPS/Galileo/TéléMesure MicroSat. 15èmes Journées Natio ales Microondes, May
2007, Toulouse, France, pp 1–4
5. Mandelbrot BB (1983) The Fractal Geometry of Nature, San Francisco, pp 152–180
6. Constantine Balanis A (1997) Antenna Theory Analysis and Design. Wiley, Hoboken
7. Moutaouakil A, Jabrane Y, Reha A, Koumina MA (2019) Effect study of a ring on the
performances of a patch antenna. In 2019 International Conference of Computer Science and
Renewable Energies (ICCSRE), Agadir, Morocco, pp 1–3
8. Chen WL, Wang GM (2008) Small size edge-fed Sierpinski carpet microstrip patch antenna.
PIERS C, vol. 3
9. Gupta S, Kshirsagar P, Mukherjee B (2018) Sierpinski fractal inspired inverted pyramidal DRA
for wide band applications. Electromagn 38(2):103–112
10. Moutaouakil A, Jabrane Y, Reha A, Koumina MA (2020) Design and simulation of a fractal
multiband antenna printed on a circular patch. Int J Adv Sci Technol 29(06), 5290–5296
11. Boutejdar A, Halim BI (2019) Design and manufacturing of a dual patch ring antenna using
parasitic ring resonators and partial ground plane for multiband applications. In: 2019 IEEE
International Symposium on Phased Array Systems and Technology
Design and Simulation of Three Bands
Planar Inverted F Antenna Array for 5G
Communication Systems

Sara Arsalane, Nabil Arsalane, and Mounir Rifi

Abstract The fifth generation of wireless mobile communication is one of the most
recent technologies in the field of wireless communications; this advance should
provide a very great service in terms of speed and diversity.
The old phones had an exterior antennas as an antenna dipole, but now most of
them had an integrated antennas, these antennas are PIFA (Patch Inverted F antenna),
a patch short—circuited in a corner (to decrease its size) and a power supply beside
it.
This paper is focused on this type of antenna that is frequently used in the field
of mobile communication, and it aims to present a design and simulation of a three-
band planar inverted F antenna array, which will be used as equipment under test for
electromagnetic compatibility tests in the fifth generation context.

Keywords Wireless network · PIFA antenna · PIFA array antenna · 5G · CEM

1 Introduction

A wireless network is a network within—at least—two devices can communi-

cate without a cables, through networks wireless, a user has the capability to stay
connected while moving around a geographic perimeter that is why occasionally we
hear about “mobility” [1].
Wireless networks are based on a link make use of radio waves in place of the
usual cables. There are some technologies depending on various factors like the
transmission frequency used as a well as the throughput and range of transmissions.
Wireless networks simplified the connection between equipment for ten meters and

S. Arsalane (B)
Labo RITM, University Hassan II, Casablanca, Morocco
N. Arsalane
Labo LREA, Institut Superieur de genie applique, Casablanca, Morocco
M. Rifi
Labo RITM, University Hassan II, Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1039
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_95
1040 S. Arsalane et al.

a few kilometers. either, the installation of such networks does not need a heavy
upgrading of existing infrastructure as is the case of wired networks (digging trenches
to route cables, trucking and connectors…), which guide to a rapid development
of this type of technologies. On the other hand, there is a problem of regulation
relating to radio transmissions. Indeed, radio electric transmissions save for a large
applications (military, amateurs…), however they are sensitive to interference, that’s
why regulation is necessary in each country in order to define the frequency ranges
and the powers at which it is possible to transmit for each category of use [2].
Mobile telephony or cellular telephony is a way of telecommunication more accu-
rately of radio communication. The transmission of voice and data using radio waves
(frequencies in the UHF bands ranging from 700 to 2600 MHz) between a base
stations that can cover an area of several tens of kilometers in radius and the user’s
mobile phone [3].
Antennas are a cornerstone of wireless telecommunications systems, these
elements allows accessing radio, television telephony wherever they are, and they
are present at both ends of a link. Wireless transmission cannot be performed without
the use of transmitting and receiving antennas. On transmission, it is the only inter-
face capable of ensuring signal transfer between a guided propagation medium and
free space. At the reception, it takes care of the interfacing between the free space
and the guided environment [4]. Among the most used antennas, it is PIFA antenna
(inverted F antenna) which is increasingly used in the mobile phone market, and it
is resonant at a quarter wavelength (thus reducing the required space needed on the
phone) [5]. This paper aims to study the PIFA antenna characteristics and simulation
of a three band PIFA antenna array for 5G communications systems that will be used
as equipment under test for electromagnetic compatibility tests in incoming paper.
This article is organized in three parts:
• Presentation of PIFA antenna and its characteristics.
• Implementation of a model of the PIFA antenna and PIFA array antenna.
• Results and discussion.

2 Presentation of Technical Details of a PIFA Antenna

In Recent years, the development of mobile telephony has growing rapidly, new
bands have been added, and the market is demanding small mobile phones with
more services, which requires the use of miniature antennas, multiband and which
minimize the risk of parasitic radiation to the human body [6].
PIFA antennas (Planar Inverted F antennas) are today the most used in mobile
telephony, due to their simple and flexible structure: reduced volume, low manu-
facturing cost, facility of integration; however, their major disadvantage is having a
narrow bandwidth [7].
Design and Simulation of Three Bands Planar Inverted F Antenna Array … 1041

Fig. 1 Typical inverted F antenna (left) and typical PIFA antenna (right)

2.1 Antenna Description

The PIFA antenna is the consequence of the modification of the IFA antenna (inverted
F antenna) from a horizontal wire element to a planar structure in order to rectify its
losses mismatch and improve its radiation characteristics [8] (Fig. 1).
This structure consists of a metallic radiating element parallel to the plane of
mass. One of the edges of the patch is connected to the ground by a short circuit
plane which constitutes a particularities of this antenna. The upper element is usually
excited by a probe coaxial that penetrates the dielectric substrate [9].
The variation of its dimensions, the power supply position, the plane of short
circuit, the substrate height alter the performance of PIFA.
These antennas have a several advantages like:
• Their compactness, low manufacturing cost.
• Their performance can be improved by adding loads judiciously.
• In research carried out on mobile telephony: PIFA represent relatively large losses
in the hand, but week in the head.

2.2 Implementation of a Model of the PIFA Antenna

This paragraph aims to choose the model of the PIFA antenna and its parameters.
The structure of antenna consists to choose the metal of the plate which induce
a short circuit between the radiating patch and the ground plane, as in the Fig. 2,
also the metal of (substrate, patch, the ground, the feed rod, the coaxial cable) [10]
(Table 1).
And to improve the performance of the PIFA antenna as well as to solve the
problem of having a narrow bandwidth there are several techniques:
• Increase the height of the substrate to increase the volume of the antenna.
• Reduce the size of the ground plane.
• Inserts slots in the round plane.
• Use of material of low relative permittivity as substrate.
• Adding a parallel plane between the patch and the ground plane (load capacitive).
• Inserting slots into the patch.
1042 S. Arsalane et al.

Fig. 2 The inverted F antenna

Table 1 Different material of

Component Material Mue Epsilon
such layer of the antenna
Feed rod Copper (pure) 1 –
Coaxial cable Copper (pure) 1 –
Ground Copper (pure) 1 –
Patch Copper (pure) 1 –
Substrate FR4 1 4.3
Short Copper (pure) 1 –

For the antenna presented in this paper we chose to insert slots in the patch to
have a multiband antenna.
And as the dimensions of the antenna had an impact on the directivity of the
antenna, so it is important to choose it carefully (Table 2).
After determination of all the factors that will be used to simulate the antenna and
also the dimensions of PIFA, so this paragraph will be concentrated on the conception
and simulation of the antenna in the chosen frequency range (2 GHz–6 GHz).
The calculations are made in far field (Fraunnhofer area) (Fig. 3).
Concerning the adaptation of the antenna, the following figure shows the
magnitude of the reflection coefficient in the frequency range (2 GHz–6 GHz) (Fig. 4).
The measurements are made in the frequency range (2 GHz−6 GHz) which
includes the second frequency range of the 5G technology.
The magnitude of the reflections coefficient is in decibel with a frequency up to
6 GHz. As we can notice that the magnitude of reflection coefficient S (1, 1) is less
than (−10 dB) for three frequencies (2.4 GHz, 3.4 GHz and 5.4 GHz) which means
that our antenna is a tri band antenna, and it is adapted for this three frequencies
and as a consequence we have the transmission of the maximum of electromagnetic
radiation.
For the radiation pattern that it obtained by the calculation of the electric field
in two plane E and H. that it represent the energy radiated by an antenna. It’s a
Design and Simulation of Three Bands Planar Inverted F Antenna Array … 1043

Table 2 Dimension of the

Name Description Value Expression
PIFA antenna
Lg Ground length 90
Wg Ground width 60
R Radius of outer cylinder 5
ri Inner cylinder radius 2
Hs Substrate height 2 Icc+Hs+Hg
Hg Ground height 2
Hc Cylinder height 10
Hpole Monopole height 9
Icc Short circuit length 5
Wcc Short circuit width 10
Lpatch Patch length 50
Wpatch Patch width 50
Ws Substrate width 55
Ls Substrate length 65
Wslot1 Slot1width 40
Wslot2 Slot2 width 30
Lslot1 Slot1lengh 39
Lslot2 Slot 2 length 35

Fig. 3 PIFA antenna

diagrammatical representation of the transportation of radiated energy into space,

according to the energy radiation [11] (Fig. 5).
This shape represent a three dimensional radiation pattern, for our PIFA antenna
the large section of the radiated field, that covers a huge area, is the main lobe or
major lobe. This is the part where the maximum radiated energy exists. The direction
of this lobe stipulates the directivity of the antenna.
The angle of aperture of an antenna is described as the area, aligned perpendicular
to the direction of an incoming electromagnetic waves [12] (Fig. 6).
1044 S. Arsalane et al.

Fig. 4 Reflection coefficient magnitude of PIFA antenna

Fig. 5 The 3D PIFA antenna radiation pattern

Fig. 6 The 2D PIFA

antenna radiation pattern

For the opening angle it fluctuate between 60 to 120 theta but it is intensive from
50 to 100, so the opening angle is 50 theta in the far field.
After the simulation of our antenna and obtaining the different results, we will
approach the periodization of this antenna in order to obtain an antenna array, which
will be functioning in the 5G frequency band and used subsequently in tests as an
equipment under tests.
Design and Simulation of Three Bands Planar Inverted F Antenna Array … 1045

2.3 Implementation of a Model of the PIFA Array

PIFA or formed beam antennas are used for several purposes, an antenna array is
a collection of separate antennas and fed synchronously, the electromagnetic field
produced by an antenna array is the vector sum of the fields produced by each of the
elements.
In addition these antennas are generally not very cumbersome in comparison with
horn type aperture antennas or reflector antenna which would fulfill the same mission
[13].

Principle
Antenna array are the result of the judicious association of several antennas to
synthesize a radiant opening [14].
Antenna array are made up of N radiating sources (dipoles, patches, PIFA…)
distributed in space. The electrical quantities (signals) injected or collected on the
access of each of these source are weighted in amplitude and (or) in phase before to
be summoned among themselves. A beam splitter is used to perform coherent and
weighted summations of the signals collected by a receiving systems, and the splitter
will be injected into each source of antenna. Due to this weighting, the antenna arrays
can produce diagrams of radiations having the desired shape [15]. In Particular, it
will be possible to create several lobes simultaneously or a lobe in the direction of
the incident signal and a zero in the direction interference.
Array antennas can have several geometric configurations that can be grouped as
follows:
• Linear Arrays: alignment of sources on a straight line.
• Planar Arrays: the sources are arranged on a plane.
• Circular Arrays: the sources are arranged on a circle.
• Volume Arrays: the sources are distributed in a volume.
Several factors contribute to the formation of the antenna radiation pattern:
• The geometric configuration (linear, circular, planar, volume).
• The distance between the sources.
• The amplitude of excitation for each element.
• The polarization phase for each element.
For our case we choose to implement a linear array, it is the most frequently used
geometry in the design of antenna arrays because it is easy to implement, it is made
up of N sources equidistant with a lattice patched. These sources are supplied with
the same amplitude and with a progressive phase gradient for a point M located in
the far radiation zone, the total radiated field is given by the sum of each elementary
field [16].
Therefor this paragraph present the case of PIFA array in a linear geometric. The
linear array composed of four elements (PIFA antenna) distributed on an axis and
separated by a distance d (pitch of the array) (Fig. 7).
1046 S. Arsalane et al.

Fig. 7 PIFA array

• The mutual coupling between the antenna elements:

The principle of an antenna is to change electromagnetic energy into an induced
voltage or current to be measured and vice versa. Nevertheless, the voltage calculated
in each antenna piece will depend on the incident field, and also on the tensions of
other elements. However, the voltage received on each element will induce a current
that affects the neighboring element that in turn radiates [17].
• The distance between the antenna elements:
Concerning the design of the antenna array, there is a major factor to consider
in the calculation. This factor is the spacing between the elements of the array that
acts on the radiation pattern and gain. Noting that when we elevate the number of
radiating elements, the gain is increased by 3 dB but this without taking into account
losses [18].The maximum gain is obtained when the spacing is between 0.4 and
0.9 λ, if the elements are too close to each other, a coupling phenomenon reduces
the value of the gain and when they are too far apart, secondary lobes appear and
therefore reduce the directivity.

Implementation of a Model of PIFA Array

This part aims to present the results of calculation of the PIFA array. The
measurements are made in far field in the frequencies band (2 GHz–6 GHz) (Fig. 8).
This figure present the reflection coefficient S(1, 1) given in decibel with a
frequency up to 6 GHz, and as we can observe that the amplitude is less than 3 dB
for three frequencies (2.4 GHz, 3.4 GHz, 5.3 GHz), so our PIFA array is adapted for
this three frequencies. And as a result we managed to correct the major disadvantage
of is type of antenna which is the narrow bandwidth.
For the radiation pattern we chose to present it for the frequency 3.4 GHz (Fig. 9).
By definition the angle of aperture of an antenna is the angular size for which the
radiated power is half of the radiated power in the most favorable direction and for
our case it varies from 65 to 110, but concentrated from 70 to 100 (Fig. 10).
Design and Simulation of Three Bands Planar Inverted F Antenna Array … 1047

Fig. 8 Reflection coefficient magnitude of PIFA array

Fig. 9 The 2D PIFA array radiation pattern

Fig. 10 The 3D PIFA array radiation pattern

And for the 3D presentation it shows the concentration of the energy in a specific
direction, that represent to the width of the major lobe which is determinate by the
angular width of each boundary of the lobe whose intensity decreases by half, a
decrease of 3 decibel.
1048 S. Arsalane et al.

Table 3 The energy balance

Frequencies in GHz Energy in dB
of each frequency
2.4 54,13 e−6
3.4 65,22 e−6
5.3 84,26 e−6

For the energy balance of the antenna which means the radiated power carried by
the electromagnetic wave for a specified frequency. The energy of each frequency it
presented in this table (Table 3):
The radiated electric energy depends on the frequency, and that what we observe
in the table, the energy balance increases with the frequency.

3 Conclusion

Innovation in communication systems, required an advanced studies in the field of

antennas. In this context, we propose a design of PIFA antenna array (4 * 1) operate
in three frequency bands (2.4 GHz, 3.4 GHz, 5.3 GHz), therefore we remedy to a
main drawback of the PIFA antenna that is the narrow bandwidth by inserting slots
into the patch. And as a perspective of this work we have:
– The implementation of this antenna array.
– Use of this antenna as an equipment under test for the electromagnetic compati-
bility tests.

References

1. Imrich chlamtac (2019) Wireless networks. J Mobile Commun Comput Inf

2. Muscatello J, Martin J (2005) Wireless networks security, 20 April
3. Vasseur, JP, Dunkels A (2010) what are smart objects
4. Dai HN, Wu MY, Mingluli (2013) An overview of using directional antennas in wireless
networks, April
5. Firoozy N, shirazi M (2011) Planar inverted F antenna (PIFA) design dissection for cellular
communication application, January
6. Preethi V, Annapurna Devi S (2018) PIFA antenna for wireless communications, June
7. Kumar N, Thakur A, der shama J (2017) Study of planar inverted F (PIFA) for mobile devices,
June
8. IEEE (2017) Design of multi resonant PIFA antenna for mobile telecommunication networks,
March
9. Agarwal M, Singh R, Meshram MK (2013) Dual band linearly polarized planar inverted F
antenna (PIFA) for GPS/WIMAX applications
10. Mohamed G (2016) Conception d’une antenne PIFA multibandes pour applications sans fils,
27 June
11. Vallozi L, Rogier H (2016) Latest developments in the field of textile antenna
Design and Simulation of Three Bands Planar Inverted F Antenna Array … 1049

12. Luo Y, Zhang Z, Chen H, Huangfu J, Ran L (2009) High directivity antenna with small aperture,
November
13. Sayidmarie KH, Maewan NJ, Aabd R, See C (2019) Antennas for emerging 5G systems
14. Reckeweg M, Rohner C (2015) Antenna Basics
15. Kumar ST, Susila M (2017) Design of multi resonant PIFA for mobile telecommunication
networks
16. Rhee E (2020) Miniaturized PIFA for 5g communication networks, 15 February
17. Singh H, Sneha HL, Jha RM (2013) Mutual coupling in phased arrays, April
18. Abimbola RA (2017) Effects of spacing and number of elements on the fields radiated by linear
arrays of equally spaced elements, October
RFID System for Hospital Monitoring
and Medication Tracking Using Digital
Signature

Safae El Abkari, Soufiane Kaissari, Jamal El Mhamdi, Abdelilah Jilbab,

and El Hassan El Abkari

Abstract The enhancement of patient safety and the improvement of medical care
services remain one the main concerns of the World Health Organization. Thus,
given the healthcare critical impact on patients, there is a growing need for high-risk
and expensive medicine management. For this purpose, we propose an innovative
system prototype for the reduction of medication error possibility from prescription,
validation to its preparation. We present a Radio Frequency IDentification monitoring
system for hospital and pharmacy using digital signature. Our designed system is
constituted of RFID readers, ESP8266 modules, and tags to ensure an automated
medication management and a good inventory control.

Keywords Tracking · Location · Medication · Monitoring · Management ·

RFID · Wi-Fi · Hospital

1 Introduction

Nowadays, healthcare sector can be described as one of the fast-growing domains.

Therefore, there is a consistent need of innovative solutions and applications [1–3]
to improve the medical care and the clinical process quality [4].
In healthcare intuitions, medication errors’ percentage is 56% for ordering, 34%
for administration, 6% for transcribing, and 4% for dispensing [5]. While statistical
data of hospitals show that computerized data is utilized in most healthcare institu-
tions, the Institute of Medicine (IOM) records that there is an occurrence of many
medication errors during the clinical process [6]. Hospital monitoring and manage-
ment systems can be a solution to efficiently speed up data flow in hospitals, reduce
errors and im-prove the inventory control of high risk or expensive medications [7].
For this purpose, the following actions are necessary to develop to ensure patient and
medication safety:

S. El Abkari (B) · S. Kaissari · J. El Mhamdi · A. Jilbab · E. H. El Abkari

E2SN, ENSAM, ST2I, Mohammed V University in Rabat, Rabat, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1051
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_96
1052 S. El Abkari et al.

• Prevent and identify adverse events to reduce medication errors.

• In case of errors, fix and evaluate the medication process.
Traditionally, most medical institutions manage medications in warehouses [8].
Thus, Radio Frequency IDentification (RFID) comes as an innovative solution for
medications, equipment or medical staff identification and tracking. In fact, it is a
technology for an automatic and flexible monitoring system in healthcare due to the
following advantages:
• Direct access to data in the system in a short time which reduce human error risks.
• Automatic non-touch object and people identification within various sensing
ranges.
• Information data can be coded and encrypted which permits high secured data
system.
• High storage capacity (up to 1 Mbits).
Our main contributions are summarized as follows:
• A pharmacy and a hospital medication monitoring system using digital signature
and RFID technology.
• Organization and fast implementation of the medication process.
• Reduction of manpower with a digitalized and efficient monitoring-management
system.
• Material and information flow speed up.
• Medication safety improvement with an automated and an intelligent medication
monitoring protocol.
The rest of the paper is organized as follows. Section 2 presents related works.
We introduce our proposed RFID system in Sect. 3. Finally, Sect. 4 concludes our
paper.

2 Wireless Monitoring Systems

Many studies on medication monitoring systems have evolved and achieved high
performances by utilizing different wireless technologies. Table 1 below present
different wireless identification technology related management systems.

3 Methods and Results

3.1 RFID Technology (Radio Frequency Identification)

RFID or Radio Frequency IDentification (electronic label) is a wireless technology

for automatic identification of people and object using radio waves. RFID system is
RFID System for Hospital Monitoring and Medication Tracking … 1053

Table 1 Wireless Identification Technology related management systems

Ref. Technology Proposition Description
[9] Barcode Medication identification Reads barcodes (Identifiers) using
smartphone cameras
[6] Wi-Fi Medication Management Presents frameworks for personalized
medication treatment management
It could also be applied for existing
personalized care system
[10] NFC Medication tracking Tracks medications using NFC tags
and a microcontroller
Transfers data to smartphone
[11] RFID Medication tracking Evaluates an RFID application in
tracing medication and patient in an
intensive care unit
[12] RFID/NFC Medication tracking Uses RFID tags (Identifiers) and NFC
smartphones as readers
[13] RFID Medication stocks’ monitoring Proposes an automated system for
system medication stocks’ monitoring
Uses two sensors’ types which ensure
a wide range of medical product
management (medical equipment and
medicines)
[14] RFID Healthcare moni-toring system Creates monitoring cycles and
monitors health status of patients
using mi-crocontrollers, sensors and
Internet of Thing (IoT)

Fig. 1 The process of RFID

identification

constituted of a tag, a reader and a software application [15]. RFID tags include chips
for data storage and antennas while RFID readers (responders) read and can write
information of multiple tags simultaneously without line-of-sight communication
[16, 17]. RFID technology offers great advantages such as its reusability, ease of
data transmission, and security [18]. Figure below illustrates the RFID identification
process which consists of reading an RFID tag attached to a person or an object
without any physical contact (Fig. 1).
1054 S. El Abkari et al.

We classify RFID tags depending on their features into different categories:

• Frequency bands

– Microwave: 2.4–5.8 GHz.

– Ultra-High Frequency: 865–954 MHz.
– High Frequency: 13.6 MHz.
– Low Frequency: 125–134 kHz.
• Power Source

– Passive: wherein tag and reader possess same energy source.

– Active: wherein tag and reader do not possess same energy source.
– Semi-active: wherein tag energy source is used for chip power supply and reader
energy source is used for data transmission.
• Storage Capacity: Kbits, Megabits …
RFID technology has a significant potential for reducing medical errors and
improve medical service quality in the healthcare field [19]. Thus, RFID can
be highly effective for rapid tracking of medical staff, medications or equipment.

3.2 System Components

In order to design an efficient medicine monitoring system prototype, it is neces-

sary to select the right components that will properly operate during the medica-
tion process (ordering, administration, transcribing, and dispensing). We used the
following com-ponents (Table 2):

3.3 Methodology

Our RFID management system consists of four main parts (Fig. 3):
• Pharmacy system (Fig. 2. b) placed in the pharmacy,
• Nurse/doctor modules (Fig. 2. a) carried by the medical staff,
• Tracking modules placed at each room door,
• Monitoring system which collects and processes data collected using RFID and
Wi-Fi devices.
For the pharmacy system, we establish communication between RFID module and
Arduino Uno board with Serial Peripheral Interface (SPI) using SS, clock, MISO,
and MOSI. ESP8266 module and Arduino Uno board [20] communicate with TX
and RX pins. Thus, ESP8266 modules are utilized to send RFID IDs to pharmacy
and other system components to track medications.
RFID System for Hospital Monitoring and Medication Tracking … 1055

Table 2 Functions of used hardware

Item Function
Arduino Uno card Temporary medication data processing
8-bit microprocessor
32k Flash Memory
2K RAM -16 MHz Clock Speed
3.3–5 V
14 Digital + 6 Analog I/O Pins
13.56 MHz RFID Reader Medication management
(NFC –PN532 Module)
Passive RFID tags
13.56 Mhz
ESP8266 module Nurses/Doctors locating
(Wi-Fi module) Wireless network communication
Computer Medication data processing
(Toshiba satellite skull-candy l50- b- 11g)
Computer Locating data processing
(HP elite 8200 – Windows 10) Medication data storage

Fig. 2 a RFID nurse/doctor module, b Pharmacy monitoring system.

Medications (Fig. 4), patients, nurses, and doctors have their respective tags and
IDs (Fig. 5). Medical staff and medications are tracked and monitored by PN532-
NFC and ESP8266 modules. Each patient, doctor, and nurse have a unique digital
signature (unique identification code (ID)).
Traditionally, prescriptions are handwritten, nurses verbally inform the pharma-
cist about a patient medication. In addition, records of token medication are done
manually by the pharmacist (Table 3). It means that it is mainly based on human
resources which leads to many errors during the medication process. Our system
attempts to locate, manage, and monitor medications by improving the traditional
practices through an automated monitoring system. Tables 3 and 4 show traditional
and proposed medication monitoring processes.
The medication process is constituted of two main phases: Prescription phase and
medicine dispensing-tracking phase.
• Prescription phase
1056 S. El Abkari et al.

Fig. 3 RFID monitoring system.

Fig. 4 Medication RFID IDs

Fig. 5 Nurse, doctor and patient RFID IDs

RFID System for Hospital Monitoring and Medication Tracking … 1057

Table 3 Traditional medication monitoring system

Step Description
After diagnosing a patient, the doctor prescribes medicines and
gives prescription to nurse in charge.

The pharmacist prepares medications.

Nurse receives medications.

The patient receives his medications.

Table 4 RFID medication monitoring system

Step Description
After diagnosing a patient, the doctor prescribes medicines
and gives prescription to nurse in charge.

If needed, doctor changes or updates prescriptions.

The pharmacist prepares medications.

When medications are ready for pickup, the nurse receives

a notification via her/his mobile module.

The nurse checks for any errors or changes in the prescrip-

tion, confirms medication reception.

The patient receives his medications.

In this phase, the doctor prescribes medications after diagnosing a patient. The
pharma-cist then receives prescriptions signed with doctors’ IDs, verify IDs,
prepares medications, and sends a notification to the nurse in charge (Fig. 6.).
If there is an error in the prescription, the doctor receives a query to update it.
1058 S. El Abkari et al.

Fig. 6 Prescription received at the pharmacy system level

Fig. 7 Prescription
information received by the
nurse

• Phase of medication dispensing-tracking

The nurse receives a notification from the pharmacist (Fig. 7.), verifies if the
prescription has no errors (patient, medications, dosage and time are correct) and
alerts the pharmacist by his/her module. When the nurse picks up medications,
she/he confirms the reception.

Once the nurse receives medications prepared by the pharmacist, the monitoring
system starts tracking those medications. Tracking modules send then notifications
of medication locations with correspondent nurse ID to the monitoring system.
ESP8266 modules are also placed in different places in the hospital to permit
communication and data transfer within the network.
Our proposed system shows an improvement in patients’ safety in hospitals and
the medication process. Table 5 below illustrates features of our proposed system
compared to existing medication monitoring systems.
RFID System for Hospital Monitoring and Medication Tracking … 1059

Table 5 Comparison of our system to existing monitoring systems

Ref. [22] [23] [24] [9] [25] Our system
Medication prescription
(list, dose, instruction)
Medication identification
Medication locating
Handling of medication
errors
Medication validation
Medication updates

4 Conclusion

In this paper, RFID technology is implemented in a monitoring system for the

real-time medication tracking, reducing medical errors during medical or clinical
processes, and enhance the pharmacy inventory control.
The proposed system constituted of ESP8266 and RFID modules allows the infor-
mation flow speed up between the medical staff, thus improves the quality service
of patient care and work performances. In addition, we designed the proposed RFID
system according to the recent healthcare trend requirements and we experimentally
demonstrated its effectiveness for the medical process improvement.
For future works, we aim to integrate Internet of Things (IoT) features to our
doctor/ nurse modules which are an early stage. We will also focus on developing
the tracking of people within the hospital (medical staff and patients).

References

1. Husar J, Iakovets A (2019) Proposal for monitoring the movement of biological samples in
hospitals. In: Automation and Management in Theory and Practice. In: 13th Anniversary Con-
ference of Universities, High Schools and Practice, Stará Lesná, SR
2. Balog M, Husár J, Iakovets A (2020) RFID System for Effective Managing of Medical Facilities.
New Approaches in Management of Smart Manufacturing Systems, pp 1–23. Springer, Cham
3. Mhamdi J, El Abkari S (2015) Contriving an RFID system for Alzheimer patients tracking.
In: 3rd International Workshop on RFID and Adaptive Wireless Sensor Networks, pp 23–28.
IEEE, Agadir
4. Bates DW, Cullen DJ, Laird N, Petersen LA, Small SD, Servi D, Vander Vliet M (1995)
Incidence of adverse drug events and potential adverse drug events: implications for prevention.
JAMA 274(1):29–34
5. Omotosho A, Ayegba P (2019) Medication adherence: a review and lessons for developing
countries. Int J Online Biomed Eng (iJOE) 15(11):104–123
6. Koutkias VG, Chouvarda I, Triantafyllidis A, Malousi A, Giaglis GD, Maglaveras N (2009)
A personalized framework for medication treatment management in chronic care. IEEE Trans
Inf Technol Biomed 14(2):464–472
1060 S. El Abkari et al.

7. Tejesh BSS, Neeraja SJAEJ (2018) Warehouse inventory management system using IoT and
open-source framework. Alexandria Eng J 57(4):3817–3823
8. Balog M, Matiskova D, Semanco P (2017) System of Workpiece Automation with RFID
Identifier Published Patent Application SK 50063–2016 Darina Matisková, A., Michal Ba-log,
Pavol Semančo Banská Bystrica: IPO SR 7
9. Sarzynski E, Decker B, Thul A, Weismantel D, Melaragni R, Cholakis E, Tewari M, Beckholt
K, Zaroukian M, Kennedy AC (2017) Beta testing a novel smartphone applica-tion to improve
medication adherence. Telemed e-Health 23:339–348
10. Morak J, Schwarz M, Hayn D, Schreier G (2012) Feasibility of mHealth and Near Field
Communication technology-based medication adherence monitoring. In: Annual International
Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pp 272–275.
IEEE
11. Yang J, Feng Z, Ma X, Zhang X (2018) Indoor positioning method based on wireless signal.
Int J Online Biomed Eng (iJOE) 14(10):53–67
12. Schreier G, Schwarz M, Modre-Osprian R, Kastner P, Scherr D, Fruhwald F (2013) Design and
evaluation of a multimodal mHealth based medication management system for patient self-
administration. In: 35th Annual International Conference of the IEEE Engineering in Medicine
and Biology Society (EMBC), pp 7270–7273
13. Mirea A, Albu A (2018) Acquisition of physical data in an automated system for monitoring
medication stocks. In: 12th International Symposium on Applied Computational Intelligence
and Informatics (SACI), pp 000179–000182
14. Khan SF (2017) Health care monitoring system in Internet of Things (IoT) by using RFID. In:
International Conference on Industrial Technology and Management (ICITM), pp 198–204
15. McCarthy JF, Nguyen DH, Rashid AM, Soroczak S (2002) Proactive displays and the
experience UbiComp project. ACM SIGGROUP Bull 23(3):38–41
16. RFID Journal, 2002–2007, “referred 6.8.2007” https://www.rfidjournal.com. Accessed 25 July
2020
17. El Abkari S, Jilbab A, El Mhamdi J (2018) Real time positioning over WSN and RFID network
integration. In: 4th International Conference on Advanced Technologies for Signal and Image
Processing (ATSIP), pp 1–5. IEEE, Tunisia
18. Cheng-Ju L, Li L, Shi-Zong C, Chi Chen W, Chun-Huang H, Xin-Mei C (2014) Mobile
healthcare service system using rfid. In: International Conference on Networking, Sensing and
Control, vol 2, pp 1014–1019
19. Castro L, Lefebvre E, Lefebvre LA (2013) Adding intelligence to mobile asset management
in hospitals: the true value of RFID. J Med Syst 37:1–17
20. El Abkari S, Jilbab A, El Mhamdi J (2020) Wireless Indoor localization using fingerprinting
and Trilateration. J Adv Res Dyn Control Syst 12(7):2597–2602
21. Devos P, Jou AM, De Waele G, Petrovic M (2015) Design for personalized mobile health
applications for enhanced older people participation. Eur Geriatr Med 6:593–597
22. Abbey B, Alipour A, Gilmour L, Camp C, Hofer C, Lederer R, Rasmussen G, Liu L, Nikolaidis
I, Stroulia E (2012) A remotely programmable smart pillbox for enhancing med-ication adher-
ence. In: Computer-Based Medical Systems (CBMS). In: 25th International Sym-posium, pp
1–4
23. Neubeck L, Coorey G, Peiris D, Mulley J, Heeley E, Hersch F, Redfern J (2016) Devel-opment
of an integrated e-health tool for people with, or at high risk of, cardiovascular dis-ease: the
consumer navigation of electronic cardiovascular tools (CONNECT) web application. Int J
Med Informatics 96:4–37
24. Ferreira F, Almeida N, Rosa AF, Oliveira A, Teixeira A, Pereira JC (2013) Multimodal and
adaptable medication assistant for the elderly: a prototype for interaction and usability in
smartphones. In: Information Systems and Technologies (CISTI), pp 1–6
Building a Domain Ontology
for the Construction Industry: Towards
Knowledge Sharing

Hayat El Asri, Fatine Jebbor, and Laila Benhlima

Abstract As the construction industry has crucial interplays with many related
fields, all in a constant state of flux, the use of a domain ontology to capture various
relevant entries from related subdomains has become of paramount importance. The
purpose of this article is to conceptualize the utility of ontologies in the construction
industry as a solution for inter-sectorial transactions. Thereafter, upon presenting the
process of ontology development, a real-life company was taken as a case study to
implement the developed ontology. The purpose behind this study is to fully automate
the generation of a domain ontology from Excel to OWL files. The methodology and
process followed are thoroughly explained in the paper. The process is meant to
be generic in such a way it could be reused in other projects for other industries,
provided that the data is hierarchical and presented under Excel.

Keywords Domain ontology · Construction · Knowledge sharing

1 Introduction

The construction industry is constantly evolving, and new technologies are being
implemented and used on a regular basis. One of those trendy technologies is ontolo-
gies. In fact, because the construction industry generally implicates stakeholders
from different backgrounds and disciplines, having a “common representation of
concepts” is crucial and has proven to be effective and time-saving.
The use of ontologies in the construction industry is gaining popularity. Indeed,
their ability to describe relationships and their high interconnectedness make them

H. El Asri (B) · F. Jebbor · L. Benhlima

Mohammadia School of Engineering, Mohammed V University, Rabat, Morocco
e-mail: [email protected]
F. Jebbor
e-mail: [email protected]
L. Benhlima
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1061
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_97
1062 H. El Asri et al.

the bases for modeling high-quality, linked and coherent data [1]. Research in this
area has been going on since the 2000s; however, it was not as advanced as it is today.
In fact, before 2000, research was focused on Artificial Intelligence (AI) and software
requirements. After the 2000s, AI was still in vogue, but several technologies emerged
like ontologies that became more popular with the rise of knowledge management.
Further, with the emergence of BIM (Building Information Modeling) technologies
[2], the use of ontologies has become even more important. However, the discrepancy
in file and item labelling is a seeming problem that the solution proposed in this article
tries to solve.
This paper is structured as follows. Section 2 discusses key concepts in ontolo-
gies. Section 3 discusses some relevant previous works in the field of construc-
tion. Section 4 presents the case study for which a general ontology was elaborated.
Section 5 discusses the domain ontology implementation with a focus on the mapping
rules and the implementation process. Last but not least, Sect. 6 presents the future
works.

2 Ontologies: Key Concepts

2.1 Definition

The definition of ontologies evolved tremendously over the past decades. In 1995,
Gruber defined ontology as an “explicit specification of a conceptualization”. A
few years later, in 1998, [1] added that ontologies can be used in many applica-
tions to “capture relationships, boost knowledge management, and make knowl-
edge assets intelligently accessible to people in organizations”. In 2001, the defi-
nition evolved and started encompassing conceptual models and implementations.
[3] defined ontologies as the implementation of a conceptual model that satisfies
the engineering trade-offs of a running application. In the past few years, ontolo-
gies have started being linked to knowledge management. [4] said that “Ontologies
play an important role in knowledge management and representation. In practice,
ontology means a hierarchy of concepts with a set of properties and relations that
represent a domain. Properties of each class describe the features and attributes of the
class and their restrictions.” This is, however, not always true as it is not a hierarchy.
What is interesting about ontologies is the inference process.
In a nutshell, ontologies can be seen as a “unifying structure” that enables
researchers to represent information in such a way that it provides common under-
standing and representation [5]. Furthermore, ontologies can represent any informa-
tion. They are often associated with taxonomic hierarchies of classes, although they
are not limited to this only.
Building a Domain Ontology for the Construction Industry ... 1063

2.2 Types and Uses

Different types of ontologies exist and each is used to serve a specific purpose.
The following are the well-known types. General ontologies are known to contain
and provide general knowledge rather than domain-specific knowledge. Domain
ontologies that are focused and applicable to a specific application domain, which
helps in defining how a group of users conceptualize and visualize some specific
phenomenon. Core reference ontologies that are usually linked to a specific domain;
yet, integrates different viewpoints related to groups of users. It is considered to be
the result of numerous domain ontologies. Information ontologies focus on concepts,
instances, and their relationships with the aim of proposing a general overview.
Last but not least, linguistic or terminological ontologies that can be glossaries,
dictionaries, controlled vocabularies, taxonomies, folksonomies, thesaurus, or lexical
databases. According to [6], “the roles of linguistic ontologies are twofold: The first
one is to present and define the vocabulary used. Second, linguistic ontology is the
result of a terminology agreement between a users’ community. This agreement
defines which term is used to represent a concept in order to avoid ambiguity.”
It is worth noting that an ontology and a thesaurus are completely different in
several aspects. On the one hand, ontologies are multidimensional and flexible.
They provide “controlled vocabulary” to describe a specific domain in a hierar-
chical manner. Furthermore, ontologies do not only provide definitions, but also
relationships, axioms, and restrictions. On the other hand, a thesaurus is rigid and
single-dimensional. It does not provide definitions, but rather relationships. It should
be noted that there exist three types of relationships only: BT (Broader Term), RT
(Related Term), and NT (Narrower Term).

2.3 Tools

Several languages, mechanisms, and tools are available for developing, constructing,
and implementing ontologies. The main languages used are: XML, RDF, RDFS, and
OWL. Many tools based on the aforementioned languages have been developed and
widely used like Protégé, WebOnto, WebODE, Swoop, and OWLGrEd, to mention
but a few.
1064 H. El Asri et al.

Table 1 below presents a broad comparison of the commonly-used tools:

2.4 General Ontology Development

According to [8], there are 5 main phases for building and implementing an
ontology: methodology, pre-development, development, implementation, and post-
development. The first phase, the methodology, is all about the extraction of the
domain-specific terminology from the different documents available and from the
web. In the pre-development stage, a bottom-up conceptualization is used to construct
the core ontology, which will define the semantics of the referential ontology. In this
step, a conceptual graph is also formalized. In the development phase, all requirement
specifications are analyzed and the taxonomic structure is elaborated. Representing
the ontology using a specific language, OWL for example, is carried out and the
conceptual model is officially elaborated. Then, the implementation process starts
and is focused on two main aspects: the operations and the exploitation. In the oper-
ationalization process, making the core ontology context-specific is key. Then, in the
exploitation process, the focus shifts towards integrating the referential ontology into
a knowledge based-system that includes a fact base and a rule base. Maintenance
happens in the post-development phase. Last but not least, determining the need for
change after implementation happens in the integration process.
For this project, we are interested in developing and implementing a semi-
automatic process that would enable the generation of domain-specific ontologies
for the construction industry.

3 Previous Works

In the past two decades, researchers got interested in ontologies and have started to
investigate their uses and how they can help the construction industry move forward.

Table 1 Comparison of ontology tools (adapted from [7])

Base Language Storage Availability Library
Protégé OKBC+ CLOS DBMS and Yes
Files
WebOnto OCML Files only Yes
WebODE HTML + Java DBMS only No
Swoop OWL HTML models No
OWLGrEd OWL Files only Yes
Building a Domain Ontology for the Construction Industry ... 1065

For instance, [9] proposed a systematic domain ontology for requirements change
management process in the global software development environment. A hybrid
method combining methontology, which is a well-known approach, and the 101
method was used for its development, the Web Ontology Language (OWL) for its
representation, and Protégé for its implementation. According to the researchers,
there are several benefits for building such ontology; for example, it can be used
to mitigate miscommunication and misunderstanding issues that may arise within
organizations.
Besides, [10] presented a method for developing an ontology of web-pages
mapped to a domain knowledge. The researchers focused on solutions of semi-
automatic ontology construction using web usage data; an experiment of Microsoft
Web data was implemented and evaluated. The goal behind this study, that was
successfully achieved, was to show the efficiency of page recommendations based
on a domain ontology. Similarly, [11] have also conducted a study on ontologies as
a crucial element in the semantic web. An ontology for the health tourism industry
in Thailand was implemented. The final results have shown that the ontology was
efficient and further works are getting implemented.
From another perspective, [12] presented a comprehensive analysis of the align-
ments between one domain ontology from the OAEI Conference track and three well
known top-level ontologies, namely: DOLCE, GFO and SUMO. The researchers
mentioned that the reuse of the knowledge available in top-level ontologies along
with that of domain ontologies reduces the modelling time and complexity, hence the
heterogeneity issue of knowledge representation. The problem of matching domain
and top-level ontologies has been addressed and a viable solution was proposed.
In addition, [13] investigated the conversion from EXPRESS, a schema that
defines the standard IFC (Industry Foundation Classes), into OWL for the construc-
tion industry. Although this has previously been investigated by researchers, no
results turned into a referential yet. [13] analyzed the available research and came
up with recommendations in order to get a usable ontology. The result of this study
is a list of criteria and requirements to be followed in order to generate a referential
ontology.
Furthermore, [14] investigated an approach for building domain ontologies for
guiding requirements’ elicitation and extracting semantic graphs from textual tech-
nical standards to generate compatible baseline domain ontologies. To achieve that,
three main steps were followed: term extraction, term clustering, and taxonomy
building. The researchers identified three core properties of domain ontologies
that they called “explicit relational expression, qualified relation identification, and
explicit temporal and spatial expressions.” The results of the paper suggest that the
methodology/approach presented in the paper could help reduce the hassle that comes
with building a domain ontology from scratch.
The number of research papers available that discuss ontologies in the construction
industry are not that numerous. Those that investigated automation processes are
limited. What we are proposing in this article is to provide an automated process that
could be reused in other domains where data is hierarchical in Excel files.
1066 H. El Asri et al.

4 Case Study

A Rabat-based construction company was taken as a case study. This company uses
many documents (mainly in excel, word, and pdf formats) that are shared among
several collaborators. As an example, Fig. 1 presents a snapshot of one of the
numerous documents used by the company, which is in Excel format: the “bor-
dereau”. It is one of the main documents used in all the works that the company
does.
The column “designation des ouvrages” (2nd column) represents the different
classes and subclasses while the column “N° des prix” (1st column) provides the
identification of each class/subclass. For instance, the class “Gros oeuvre – Charp-
ente Métallique” is a main class that contains 2 direct sub-classes, namely: “terrasse-
ments et remblais” (identifier 100.1) and “bétons et acier en infrastructure” (identifier
100.2). Each of these sub-classes has several sub-classes of its own.
This document is the basis of an important process in the company and the problem
that can occur is that collaborators sometimes use different names for the same
concept. Therefore, a domain ontology is of great help. Our objective was then to
develop a domain ontology based on several documents provided.

5 Building-Domain Ontology Development

The main objective behind developing this domain ontology for the construction
industry is to provide a common and unified terminology to enable the experts who
work on the same project(s) to use the same terminology.

UNITE DES
N° DES PRIX DESIGNATION DES OUVRAGES EN TOUTES LETTRES
MESURES
100 GROS-ŒUVRE - CHARPENTE MÉTALLIQUE
100.1 TERRASSEMENTS ET REMBLAIS
100.1.1 TERRASSEMENTS
100.1.1.1 FOUILLES EN PLEIN MASSE TOUT TERRAINS m³
100.1.1.2 FOUILLES EN PUITS ET EN RIGOLE TOUT TERRAINS m³
100.1.2 REMBLAIS
100.1.2.1 REMBLAI D'APPORT EN TOUT VENANT COMPACTE m³
100.1.2.2 REMBLAIS EN MATERIAUX PROVENANT DES FOUILLES m³
100.1.2.3 ÉVACUATION DES DÉBLAIS m³
SOUS-TOTAL HT DE TERRASSEMENTS ET REMBLAIS
100.2 BÉTONS ET ACIERS EN INFRASTRUCTURE
100.2.1 BÉTON DE PROPRETÉ m³
100.2.2 GROS BÉTON m³
100.2.3 BÉTON ARMÉ EN INFRASTRUCTURE
100.2.3.1 BÉTON POUR SEMELLES ISOLÉES, FILANTES ET RADIERS m³

Fig. 1 Snapshot of an excerpt of the “bordereau” of the structural works

Building a Domain Ontology for the Construction Industry ... 1067

5.1 Building-Domain Ontology Development Methodology

The ontology development was divided into two main phases: the specification phase,
in which the domain knowledge is acquired, and the conceptualization phase, in
which the knowledge is structured and organized. Figure 2 presents the general
methodology followed to build this ontology.
The process started with the gathering of the user requirements from the company.
A thorough analysis was conducted and a pre-building phase started. To make sure
the pre-building phase was heading in the right direction, several interviews and
document checks were carried out as part of the user requirements gathering. After
making sure that all the specifications were taken into account, the ontology develop-
ment and implementation was carried out. Last but not least, an ontology evaluation
was conducted.
Because the majority of the company’s data is available in Excel files, our goal
was to automate the domain ontology implementation process to make it easier, time-
saving, and reliable to have access to the needed data/information. The first step was
to find the mapping rules.

5.2 Mapping Rules

The following are the mapping rules used.

1. concepts -> classes: each designation is a concept. This latter is gotten by reading
each element of the column “designation des ouvrages”.
2. relationships between concepts - > object properties: there are two types of
relation-ships, namely ISA and ISAD:

“C1 ISA C2” if C1 is a sub-concept of C2.

“C1 ISAD C2” if C1 is a direct sub-concept of C2.

Fig. 2 General process of the proposed approach

1068 H. El Asri et al.

We are interested in ISAD relationships because if a concept C1 is a direct sub-

concept of C2, then C1 is automatically a sub-concept of C2. The algorithm in Fig. 4
was used to extract ISAD relationships from the “bordereau”.

5.3 Domain Ontology Implementation Process

In this section, we are especially interested in the building-domain ontology imple-

mentation, based on the mapping rules to automatically extract concepts and
relationships between these latter from the Excel files, and to generate an owl file.
In order to map the “bordereau” into a conceptual model, three main steps were
followed for each line in the Excel file. The general process of the proposed approach
is explained in Fig. 3.
The following are the steps followed, in a loop of iterations, in the proposed
approach.
Step 1: Concept extraction from the Excel file & integration in the OWL file.
The function getDesignation() returns the designation of a specific price number.
Step 2: ISAD relation extraction & concepts linking.
After checking if the extraction concept is linked via an ISAD relation with other
concepts, the identified relations are integrated in the OWL file.
Step 3: ontology evaluation.
The file obtained (DOnto4CI.owl) is checked and evaluated by a human expert.
The result of the implementation process is presented in Fig. 5 and 6.

Fig. 3 General process of

the proposed approach
Building a Domain Ontology for the Construction Industry ... 1069

Algorithm 1: extraction of ISAD relations

1: Read num=N°_prix
2: Do
3: If num is an integer Then
4: num<- next_num
5: While (num contains a dot) {
6: getDesignation(num) ISAD getDesignation(num-suffix(num))
7: num<- next_num
8: }
9 : EndIf
10: While (!end of table)

Fig. 4 Algorithm for ISAD relationships extraction

Fig. 5 Snapshot of an excerpt of the class hierarchy

Figure 5 presents a snapshot of the class hierarchy in Protégé. A number of classes

are shown, and each of the classes has several subclasses of its own. For instance,
the class “étanchéité” contains 10 subclasses such as “isolation thermique”, “écran
pare-vapeur”, and “cuvelage”.
Figure 6 presents a snapshot of a very small portion of the elaborated ontology.
The developed ontology contains 9 classes, each with up to 34 subclass. Therefore,
it was impossible to include it all in the paper.
1070 H. El Asri et al.

Fig. 6 Snapshot of an excerpt of the elaborated ontology

6 Conclusion and Future Works

The use of ontologies in the construction and building industry is growing. Because
of the multi-participants from different areas of expertise and disciplines that work in
the construction industry, and the documents that are usually written and checked by
different ex-perts coming from various backgrounds, having a uni-fied representation
of the general terms used can be time-saving and of great help. This paper presented
a domain ontology developed for a specific construction company, following their
requirements and using their data based on a semi-automatic process. At the moment,
we are working on the implementation of the entire process using Java, Apache POI
for reading the Excel files, and Jena to generate the OWL file. Furthermore, and
since its apparition, Building Information Modelling (BIM) has been considered as
a shifting point in sharing and representing knowledge in the construction industry.
Although not widely used yet, companies and researchers have realized the impor-
tance of all the benefits that come with using it. Integrating ontologies within the BIM
environment can only be beneficial and will add to the advantages that BIM already
offers. It will allow the creation of a “common jargon”, which will help in standard-
izing the meaning concepts and terminologies, encourage knowledge sharing, and
promote the reuse of existing knowledge. For this project, researchers are currently
working on the integration of the domain ontology into BIM. Last but not least,
Building a Domain Ontology for the Construction Industry ... 1071

ontology development and implementation is a continuous and long process, and the
next step will be to develop a full-fetched domain ontology, to be integrated in BIM.

References

1. Benjamins V, Fensel D, Perez A (1998) Knowledge management through ontologies.

In: Proceedings of the 2nd International Conference on Practical Aspects of Knowledge
Management (PAKM 98), vol 5, pp 1–12
2. Kubba S (2017) Building Information Modeling (BIM). Handbook of Green Building Design
and Construction, pp 227–256
3. Guarino N, Welty C (2001) A formal ontology of properties. In: Proceedings of 12th Interna-
tional Conference on Knowledge Engineering and Knowledge Management, Lecture Notes on
Computer Science. Springer
4. Hafidh R, Sharif M, Alsallal M (2019) Smart holistic model for children and youth with special
educational needs and disabilities. In: International Conference on Computing, Electronics
Communications Engineering (iCCECE), pp 130–135. IEEE
5. Anumba CJ, Issa RRA, Pan J, Mutis I (2008) Ontology-based information and knowledge
management in construction. Constr Innov 8(3):218–239
6. Roussey C, Pinet F, Kang M-A, Corcho O (2011) Ontologies in Urban Development Projects.
Springer, London
7. Slimani T (2015) Ontology development: a comparing study on tools, languages and
formalisms. Indian J Sci Technol 8(24):1–12
8. Bozzato L, Ferrari M, Trombetta A (2008) Building a domain ontology from glossaries: a
general methodology. In: Proceedings of the 5th Workshop on Semantic Web Applications and
Perspectives (SWAP2008), Italy, pp 1–11
9. Alsanad AA, Chikh A, Mirza A (2019) A domain ontology for software requirements change
management in global software development environment. IEEE Access 7:49352–49361
10. Nguyen TTS, Lu H (2016) Domain Ontology Construction Using Web Usage Data. In: Kang B,
Bai Q (eds) AI 2016: Advances in Artificial Intelligence, vol 9992. Lecture Notes in Computer
Science. Springer, Heidelberg, pp 338–344
11. Chantrapornchai C, Choksuchat, C (2016) Ontology construction and application in practice
case study of health tourism in Thailand. SpringerPlus 5(1): pp 1–31. Springer
12. Schmidt D, Trojahn C, Vieira R (2016) Validating top-level and domain ontology alignments
using WordNet. In: Proceedings of the Brazilian Seminar on Ontologies
13. Pauwels P, Terkaj W (2016) EXPRESS to OWL for construction industry: towards a
recommendable and usable ifcOWL ontology. Autom Constr 63:100–133
14. Omoronyia I, Sindre G, Stålhane T, Biffl S, Moser T, Sunindyo W (2010) A domain ontology
building process for guiding requirements elicitation. Requir Eng Found Softw Qual 188–202
Optimization of the Ad Hoc Network
by Using Hybridization of Genetic
Algorithm with a Two-Optimization
Algorithm

Hala Khankhour, Otman Abdoun, and Jâafar Abouchabaka

Abstract This article presents a hybrid genetic algorithm (GA) with two-
optimization algorithm, to find a solution to the problem routing in Ad Hoc network
big map, the objective is to find the shortest possible distance between a trans-
mitter head (cluster, node, intermediate or advanced) and the base station. From
studies and remarks, it has been proved that in the topologies of the large Ad Hoc
network, the node responsible for transmitting to the base station sometimes used
up its total energy and became inactive. So, there are many techniques to solve the
problem routing in Ad Hoc networks, like GA (Genetic Algorithm), SA (Simulated
Annealing), and BA (Bat Algorithm). In this article, we are focus on the (GA) that
are present in the basic approaches of our work. We propose an analysis of these
metaheuristics by identifying the fundamental ideas and propose ways to guide the
choice of a metaheuristic in practice, particularly in Ad Hoc big map network, to
find the best path between two nodes in communication. Thereafter, we try to more
optimizer our algorithm by using a 2-optimizer algorithm, and we compare between
two method GA and GA-2opt, the result show much better quality of our solution to
adjust the problem of routing in big map Ad Hoc network.

Keywords Artificial intelligence · Genetic algorithm · Two-optimization

algorithm · Ad Hoc network

1 Introduction

Nowadays, the use of wireless technology has occupied the telecommunication

network in the world, routing is one of the most important problems [1] because
wireless communication occurs only between two neighboring nodes, therefore a

H. Khankhour (B) · J. Abouchabaka

Computer Science Department, University IBN Tofail, Science Faculty, Kenitra, Morocco
e-mail: [email protected]
O. Abdoun
Computer Science Department, Faculty Polydisciplinary, University Abdelmalek Essaadi,
Larache, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1073
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_98
1074 H. Khankhour et al.

perfect routing algorithm must find an optimum path in communication between

a source node and destination node. [2] and to kip up the node always active so
as not to lose its energy. On the other side, for the large network which contains
several nodes (n), the time traveled between two nodes in communication, is an
exponential function of n, if the network is large, more the time is increased too,
according to the work studied. The routing problems in the Ad Hoc network were
resolved at first using Dijkstra algorithm, then the Bellman–Ford–Moore algorithm
employed to solve more complex problems, they can be effective in wireless or
wired networks. Unfortunately, the exact methods can not solve this type of complex
problems mostly for real-time communications or for topologies rapidly evolving [3].
and support other constraints such as voice, video, and teleconferencing. The appear-
ance of neural networks (NNs), proved their effectiveness to solve these complex
problems. In addition, the quality returned by the artificial NNs doesn’t influence
by the number of nodes in the network, but they have limited resources. The (GA)
are considered a suitable solution to overcome these issues and they are adapted
according to the size of the population.
For this reason, we propose in this article, an analysis of these metaheuristics
by identifying the fundamental ideas and new design about the (GA). The first GA
appeared around the 1970s. The (GA) idea has come from Darwin’s theory, Charles
Darwin observed that the (GA) is based on genetics and natural selection. It is a
combination of fitness and other components randomly but structured which is called
by population. the characteristics are then used in gene sequences which will be
combined with other genes to form chromosomes and subsequently individuals,
each solution is associated with an individual, and that individual is evaluated and
ranked based on their similarity to the best solution. In the 1960s John Holland used
the computer program with crosses and mutation [5].
Researchers have applied GAs to various problems, like the multicast Ad Hoc
problem, bandwidth, and other dynamic ones [6], rapid topology change with variable
length chromosomes (Munemoto’s algorithm [7]).
The paper is organized as follows, Sect. 2 we explain the operation of (GA), in
Sect. 3 present the GA for Ad Hoc network, after that in Sect. 4 speak about two-
optimization algorithm, in Sect. 5 present the formalization the two-opt Algorithm,
the new design algorithm is proposed in Sect. 6, and Sect. 7 gives the implementation
and discussion, the article is concluded in Sect. 8.

2 Operation of the Genetic Algorithm (GA)

Firstly we created the population, the population is made up of several individuals;

by using the roulette method, we arbitrarily crossed two individuals to give a new
generation having the characters of both parents. In addition to this crossing, we have
added the mutation operator for two individuals by chance from the last generation.
Optimization of the Ad Hoc Network by Using Hybridization … 1075

Then, this new population undergoes a selection according to the fintess of its indi-
vidual, that means that only the individuals best adapted to the environment survive
[1, 2].
In Ad Hoc network we have considered chromosomes as routes (Strings) which
are variable, and their genes as nodes (parameters) such that the distance between
nodes is different to zero, first we randomly generalize a population, which contains
several chromosomes (path), we used the single point crossing by chance (0.8), to
better improve the quality of our solution we added the mutation operator (0.05)
from the last population and we evaluated each chromosome (path) by the function
of fitness (1), the best solution must be chosen from a source node and a destination
node in Ad Hoc Network.
1
f i = li−1 (3)
j=1 C gi( j),gi( j+1)

3 Genetic Algorithms (GA) for Ad Hoc Network

According to researchers, the GA is powerful for this kind of complex problems in

particular NP-hard problems [8]. In this paper, we applied the GA on the big map Ad
Hoc Networks to minimize the total duration of a tour, from the source node to the
destination node. This challenge is also to find the shortest possible duration among
all the possible visits in the given map [9].
Description of parameters used between (GA) and Ad Hoc networks
• Individual → A possible route
• Population → A set of journeys
• Mutation → Random modification in an individual’s journey
• Adaptation → The shorter the journey, the more adapted it is to the problem
• Natural selection → Elimination of the least adapted individuals
The principle of operation of (GA) is as follows [4]
1. We randomly generate a first population of individuals
2. We select half of the individuals (natural selection)
3. We arrange the individuals remaining in a couple and we create for each couple a
couple of children who represent journeys having characteristics of the journey
of both parents
4. Children can have a one in q chance of having a mutation (user-defined q
parameter)
5. We go back to step 2.
The procedure used is depicted in Fig. 1.
To further evaluate our work, we have added to our (GA) the valuable features of
the Tow-optimization algorithm.
1076 H. Khankhour et al.

Fig. 1 Flow chart of the

genetic algorithm (GA)

Fig. 2 Example of
permutation in algorithm
with two-optimization

4 Two Optimization Algorithm (2opt)

In 1958, the two optimizations (2-opt) algorithm has been proposed by Cross to solve
the traveling salesman problem. The ultimate idea is to take a road that is crossing
and rearrange it [10] (Fig. 2).
2-optimization is an iterative algorithm: at each step, we remove two edges from
the current solution and reconnect the two turns thus formed. This method makes
it possible to improve the solutions by avoiding the path when tour inequality is
respected between a source node and a destination node (see figure above). On the
Optimization of the Ad Hoc Network by Using Hybridization … 1077

diagram on the right, the route (a ->b->e->d->c->f->g) is modified to the route (a-
>b->c->d->e->f->g) by reversing the order of visit to tour e and c. More generally,
reversing the running order of two tours need to reverse the running order of all the
tours between them [11].

5 Formalization of the 2-Opt Algorithm

Let a graph G = (V, E) and H be a Hamiltonian cycle in G provided with a cost

function returning the cost of the weights of the paths composing the tour.
We define a 2-permutation in the Hamiltonian cycle H as the replacement of two
edges a1, a2 ∈ H by two edges a3, a4 ∈ E such that the resulting turn is always
Hamiltonian in G.
In our case of the Ad Hoc network problem, we consired the edges like a node
and their diagonals like a path.

6 New Design Algorithm (GA-2opt)

In this paper, our new GA-2opt approach aims to solve the routing problem in Ad
Hoc network, between the source node and the destination node in a short time too
[12], we have proposed a new algorithm design as follows:

Begin with the constraint limits is set for the sp route

GenerationSize = 1000; PopulationSize=40; Number of iterations = 100;
CrossoverRate =0.8; MutationRate = 0.05; NodeSource = 3;
NodeDistination = 35;
Repeat GetBestCycleHamiltonian(GenerationSize, populationSize, NodeSource,
NodeDestination, CrossoverRate, MutationRate);
6.1 function
SubsectionSample
2-opt (G: graph, H: GetBestCycleHamiltonian)
improvement: boolean: = true
Please note that the first paragraph of a section or subsection is not indented. The first
as long as improvement = real doing
paragraphsimprovement:
that follows a =table,
falsefigure, equation etc. does not have an indent, either.
Subsequent
for any vertex xi of H do are indented.
paragraphs, however,
for any vertex xj of H, with j different from i-1 from i and from i + 1 do
if disatance (xi, x (i + 1)) + distance (xj, x (j + 1))> distance (xi, xj) +
distance (x (i + 1), x (j + 1)) then
replace the edges (xi, x (i + 1)) and (xj, x (j + 1)) by (xi, xj) and
(x (i + 1), x (j + 1)) in H
improvement: = true
end.
1078 H. Khankhour et al.

Table 1 The computational results between GA et GA-2opt

TSPLIP Optimal GA GA-2opt
Berlin52 7 542 7 885 7 600
Pr76 108 159 119 988 109 500
Rat99 1 211 1 400 1 220
Bier127 118 282 139 821 120 233
Eil51 426 430 426
Lin318 42 029 47 510 43 215

Fig. 3 Comparison result between GA and GA-2opt

7 Implementation and Discussion

To see the effect of the 2-optimizer algorithm on the Ad Hoc network, we have built
a program that uses the uniform (GA) as a heuristic, and in the crossing step we
check both children by the “2- optimizer”, if there is an optimization, replace the
child (route) with the new solution.
Table 1 shows the great effect of this optimize.
From the results in Fig. 3, we notice that the distance of the genetic algorithm 2-
opt is much less than the (GA) mostly for the big city like Bier127. it means that
Optimization of the Ad Hoc Network by Using Hybridization … 1079

Fig. 4 Topology with 20 nodes

in a big map network using 2-opt optimization we can find the shortest path in the
shortest time too, from the source node to the destination node.
After testing our algorithm with the BENCHMARK instances, we applied our
algorithm to the ad hoc network using the topology Fig. 4, used in this article [9]
after several execution of our program, from the above results we can say that the
delay for our new design algorithm AG2OPT is good, we find the best path, the cost
is 142, in short time too 0,020 s better than result in this article [9].

8 Conclusion

This article proposes a new conception of the hybridization between the metaheuristic
(GA) algorithm and two optimization algorithms, and we choose to further optimize
our algorithm using the advantages of the 2 opt optimizer, for this purpose the goal is
to find the shortest path from a source to a destination in quick time too, mainly for
a large network of cardswhose goal that the nodes do not lose its energy, choosing
the best path, to avoid shocks, collision and of course not to saturate the bandwidth
in Ad Hoc network in a big map.
1080 H. Khankhour et al.

References

1. Alander JT (2001) An indexed bibliography of genetic algorithms in economics, Technical

Report Report
2. Lerman I (1995) Algorithmes génétiques séquentiels et parallèles pour une représentation,
Rapport de Recherche de l’INRIA Rennes
3. Abdoun O, Abouchabaka J, Tajani C (2012) Analyzing the performance of mutation operators
to solve the travelling salesman problem. IJES, vol. 2(1)
4. Liu K, Xu H, Gao H, Peng X (2009) Research on confirmation of tension leveller basic techno-
logical parameters based on neural network and genetic algorithm, Le Yao in Computational
Mechanics, pp 422
5. Dorigo MG (1997) Ant colony system: acopperative learning approach to the traveling salesman
problem. IEEE Trans Evol Comput
6. Fausto A, Araújo R, Garrozi C (2010) MulRoGA: a multicast routing genetic algorithm
approach considering multiple objectives. Appl Intell 32, 330–345
7. Ge X, Xue G (2019) An improved genetic algorithm for pro-active dynamic vehicle routing
problem. In: Proceedings of the Twelfth International Conference on Management Science and
Engineering Management
8. Khankhour H, Abouchabaka J, Abdoun O (2020) Genetic Algorithm for Shortest Path in Ad
Hoc Networks, Lecture Notes in Networks and Systems, vol. 92, pp. 145–154
9. Chang WA, Ramakrishna RS (2002) A genetic algorithm for shortest path routing problem and
the sizing of populations. IEEE Trans Evol Comput 6(6), 12
10. Okdem S, Karaboga D (2009) Routing in wireless sensor networks using an Ant Colony
Optimization (ACO) router chip. Sensors, 9(2), 909–921. https://doi.org/10.3390/s90200909
11. Mavroidis I, Papaefstathiou I, Pnevmatikatos DN (2007) Hardware implementation of 2-
opt local search algorithm for the traveling salesman problem. In: 18th IEEE International
Workshop on Rapid System Prototyping (RSP 2007), 28–30 May 2007, Porto Alegre, RS,
Brazil
12. Kumar S, Mehfuz S (2016) Intelligent probabilistic broadcasting in mobile ad hoc network: a
PSO approach. J. Reliable Intell. Environ. 2, 107–115
Design of a Microstrip Antenna Two-Slot
for Fifth Generation Applications
Operating at 27.5 GHz

Salah-Eddine Didi, Imane Halkhams, Mohammed Fattah, Younes Balboul,

Said Mazer, and Moulhime El Bekkali

Abstract Currently, industry and academia have been engaged in significant

research and development activity in anticipation of the next generation of wireless
micro and picocellure networks (5th generation). Microstrip antennas are widely
employed for non-wired communications because of their affordability, reduced
profile, and simplicity of realization. However, weak characteristics such as low
bandwidth, a bad gain, etc. limit their possible uses. Since the future generation (5G)
needs a high gain antenna, this study’s principal reason is to conceive a sensitive
gain microstrip antenna. This article presents a patch antenna design with two slots
for the 5G application and provides a bandwidth of around 27.5 GHz. The role of
these slots improved the antenna’s performance. This antenna has been conceived
on a polyimide substrate of relative permittivity of εr = 4.3 and a height of the
substrate of (h = 0.15 mm). The metrics of this offered antenna are 27.91 GHz,
−25.4052 dB, 0.96 GHz, 6.6 dB, and 6.9 dB are working frequency, reflection
coefficient, bandwidth, gain and directivity respectively.

Keywords 27.5 GHz · Millimeter wave · Microstrip patch antenna · 5G

1 Introduction

The role of an antenna is to convert the electrical energy of a signal into electromag-
netic energy, or conversely to convert electromagnetic energy into electrical energy.
A transmitting antenna is a device that ensures the transmission of energy between
a transmitter and the free space where this energy will propagate. Conversely, a

S.-E. Didi (B) · Y. Balboul · S. Mazer · M. E. Bekkali

IASSE Laboratory, Sidi Mohamed Ben Abdellah University, Fez, Morocco
e-mail: [email protected]
I. Halkhams
SED Laboratory, Private University of Fez, Fez, Morocco
M. Fattah
IMAGE Laboratory, Moulay Ismail University, Meknes, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1081
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_99
1082 S.-E. Didi et al.

receiving antenna is a device that transmits the energy of a wave propagating in space
to a receiving device. The patch antenna is a flat antenna whose radiating element is
generally a square conductive surface, separated from a conductive reflector plane by
a dielectric substrate. Its realization resembles a double-sided printed circuit board,
substrate, and is therefore suitable for industrial production. The use of patch antennas
is almost generalized in all mobile communication systems. These antennas are light,
compact, and inexpensive [1].
The 5th generation (5G) wireless technology network is scheduled to be deployed
by 2020. It is the most ambitious wireless communication network project ever
carried out and will have to meet the expectations of a mass market. The latest spec-
ification trends focus on network densification, increased data throughput (10 Gb/s,
10 times higher than 4G), increased network coverage, latency between 1 ms and
5 ms and reduced power consumption [2].
Frequency bands: It is planned to free up frequencies for 5G that are located prac-
tically all along the radio spectrum. The National Frequency Agency (ANFR) indi-
cates that they are divided into two major classes. The low frequencies offer extensive
coverage and excellent transmission in the indoor environment. For high frequen-
cies, they are strong, but the transmission in residential environments is restricted.
Therefore, the chosen technology strategy is to combine the new radio spectrum as
well as frequencies currently available for 2G, 3G, and 4G mobile data transmission.
Currently, Two new funding tranches will apply to 5G: the 3,5 GHz frequency band is
between 3,4 GHz and 3,8 GHz and the 26 GHz frequency band is in the [24,25 GHz;
27,5 GHz] frequency range [3, 4].
In this article, we present a patch antenna with two slots for 5G. This proposed
antenna must resonate at 27.5 GHz and has a size of 4.678 mm × 4.55 mm ×
0.15 mm structure. In the first phase, we introduce the required introduction, in the
second phase; we explain the design methodology of this antenna offer, in the third
phase, we carry out the simulation of this created antenna through an HFSS (High-
Frequency Structure Simulator) software tool. The details of these simulations will be
discussed, and we thus compare the different results acquired from this new antenna
with those already available. In the fourth phase, we finish this work by concluding
with a conclusion.

2 Antenna Design and Performance

In this paragraph, we present the most commonly used method for designing
microstrip patch antennas, including the procedure for determining their parame-
ters (width and length), and also the dimensions of the ground plane, and those of the
feed lines, and the type of power supply for this design, as well as how to glue the
patch antenna to the substrate of your choice. For simplified analyses and antenna
performance predictions, we propose a rectangular patch antenna with two slots oper-
ating at 27.5 GHz as part of the fifth generation 5G application, as illustrated in Fig. 1.
Design of a Microstrip Antenna Two-Slot for Fifth Generation Applications … 1083

Fig. 1 The suggested shape of the Microstrip patch antenna

The proposed antenna has been printed on a substrate of permittivity εr = 4.3. Poly-
imides are commonly used for coating metal substrates, and use a microstrip line for
their feed. Quarter Wave Converter is a microstrip cable employed to ensure proper
impedance matching. To determine the width and length of this quarter-wavelength
(WT, LT) and those of the micro-strip feed cable (Wf, Lf), and those of this offered
antenna (Wp, Lp), we apply the formulas in [5, 6].
• by applying Eq. (1), it is possible to find the value of the width Wp:

c 2 1
W p = 0.5 ( )2 (1)
f r εr + 1

• by applying Eq. (2), it is possible to find the value of the width Lp:

c
L p = L e f f − 2L p = √ − 2L p (2)
2 f r εe f f

(εe f f + 0.3)( WH sp + 0.264)

With : L p = 0.412 (3)
(εe f f − 0.258)( WH sp + 0.8)
εr + 1 εr + 1 H s −1
And : εe f f = + (1 + 12 )2 (4)
2 2 Wp

• Ground plan dimensions:

1084 S.-E. Didi et al.

Table 1 Suggested antenna

Parameters Values (mm)
size in this work
LE 4.678
WE 4.55
Hs 0.15
Wp 2.578
Lp 3.35
LT 1.1
WT 0.5
Lf 0.9
Wf 0.2
S1 1.7 mm × 0.2 mm
S2 0.6 mm × 0.3 mm

WE = Wp + 6Hs et LE = Lp + 6Hs (5)

where: the symbols Wp and Lp represent the width and length of the rectangular
patch antenna respectively, fr represents the working frequency, c shows that the
celerity of light is approximately equal to: 300000000 m/s, Lp is shown the stretch
path, Hs is the height of the material substrate, εeff indicates the set value of the
dielectric constant of the substrate material, WE is the latitude of the earth plane,
LE is the longitude of the earth plane (Table 1).

3 The Results of the Simulation and the Debates

In this part, we show how we performed simulations on the design of a two-slot

patch antenna with S11 return loss, VSWR, gain and directivity. For this design, we
have chosen HFSS (High-Frequency Structure Simulator) as the tool of choice. The
latter it is a tool for simulating three-dimensional shapes that allows the creation
of circuits with finished ground planes, any shape, various types of materials and
insulators of different thicknesses. It applies the Finite Element Technique (FEM).
This Finite Element Technique (FEM) enables to digitally perform the following
partial differential equations.

3.1 Return Loss of This Offered Antenna

The return loss S11 of an antenna makes it possible to quantify the quantity of
the reflected signal in relation to the incident signal. This return loss is equal to −
10 of the fund value, indicating that about 10% of the received power is returned,
Design of a Microstrip Antenna Two-Slot for Fifth Generation Applications … 1085

Fig. 2 Reflection coefficient S11 of this offered antenna in this research

that is to say that 90% as much power is absorbed through the antenna, something
that is believed to be very good for wireless mobile communication. the reflection
coefficient at the antenna level, if there is a perfect match there is no reflection, that
is to say that in this case the reflection coefficient will be equal to zero, if there is no
match the reflection coefficient is different from zero, we can define the qualification
at the match level of an antenna either by attributing its impedance characteristic,
generally equal to 50 ohms, or by attributing to its level of reflection coefficient. This
antenna functions at 26.91 GHz with a return loss of −25.4052 dB, a bandwidth of
0.96 GHz (27.35 GHz–26.39 GHz) as shown in Fig. 2 below.

3.2 VSWR (Stand-Wave Ratio)

VSWR occurs due to the reflection of a large amount of the generated power. The
value of this VSWR on the efficiency bandwidth must be between 1 and 2. In order
for the transmission of the antenna power to be maximized, this ratio must tend
towards 1. Figure 3 illustrates VSWR as a function of frequency. The VSWR value
is 1.095 at the operating frequency of this antenna.

3.3 Gain

The gain is equal to the ratio between the power density radiated by the antenna in
a specific direction and the power density emitted by a reference antenna. The gain
of this study for this new antenna is 6.6 decibel, as shown in Fig. 4 below. The gain
of this antenna obtained is quite suitable and its use is the most appropriate for 5G.
1086 S.-E. Didi et al.

Fig. 3 Offered Antenna Standing Wave Ratio (VSWR)

Table 2 Comparative analysis of the results of this antenna in relation to those of the existent
results

Fig. 4 Three-dimensional radiation curve of this proposed antenna

Design of a Microstrip Antenna Two-Slot for Fifth Generation Applications … 1087

3.4 Directivity

The directivity of an antenna in a specific sense is the density of power radiated

by the antenna in this sense relative its average value in all space. Directivity is a
concentration of energy in an antenna or the ability to concentrate it and each antenna
has a specific directivity in a preferred direction. The value of the directivity of this
antenna is equal to 6.9 dB as shown in Fig. 5 below:
Table 2 contains the results of the few preceding studies concerning gain, direc-
tivity, bandwidth, S11, overall antenna size and operating frequency. The perfor-
mances of the proposed antennas, such as gain, bandwidth and reflection coefficient
S11, are preferable to those of antennas [6, 8, 9] and [11]. Thus, a second advantage
is that the proposed antenna is smaller in size than antennas [8, 6, 9] and [11]. For the
antenna [7] has a higher reflection coefficient compared to the proposed antenna, but
the size of the latter is smaller than that of the antenna [7]. The gain of the antenna
[3] is higher than that of the suggested antenna of shorter size. The antenna [10]
has a better return loss than the proposed antenna, but the latter also has a higher
gain than the antenna [10], so the proposed antenna has a very important feature,
namely its smaller size. Consequently, the aim of this work is to compensate for the
decrease in performance and the reduction in size of the microstrip patch antenna
for 5G technology.

Fig. 5 The
three-dimensional directivity
pattern of the offered
microstrip antenna
1088 S.-E. Didi et al.

4 Conclusion

In this article, I have shown the definition and characteristics of the antenna as well
as some of the advantages of the fifth generation, these advantages being the date of
its publication, the increase of the data transmission rate, the increase of the coverage
area, the decrease of the power consumption, also the steps of the transition from the
first to the fifth generation, He also mentioned the presentation of the fifth generation
frequency bands. I specified the steps in the design of a patch antenna as well as the
study of the simulations on the HFSS software and I compared the results obtained
with those of some other antennas designed during work carried out in previous
years.
In this article, rectangular microstrip antenna is presented with two slots S1 and
S2 that improve antenna performance. The improved bandwidth allows for increased
user occupancy, indispensable for 5G portable communications. The global dimen-
sions of the offered antenna are 4.55 mm × 4.678 mm × 0.115 mm. The antenna
described in this document operates over 26,91 GHz and a reflection coefficient of −
25.4052 dB, a bandwidth of 0.96 GHz, a gain of 6.6 dB, and a directivity of 6.9 dB.
Through this study, we have shown that the two-slot antenna obtains the best
results in the relevant frequency band (24,25 GHz–27,5 GHz) on which it can work
well with the fifth generation portable communication system. The proposed antenna
would therefore be a good candidate for wireless applications at 27.5 GHz.
For future work, we will use this antenna in the design of patch antenna arrays to
improve the performance of patch antennas that meet the requirements of the fifth
generation 5G.

References

1. Odile P (2009) Antennas Theory, design and application. French center for the exploitation of
coupier law, Paris. https://www.dunod.com/
2. Faisal MMA, Nabil M, Kamruzzaman Md (2018) Design and simulation of a single element
high gain microstrip patch antenna for 5G wireless communication. In: Faisal MM (ed) Inter-
national Conference on Innovations in Science. Engineering and Technology (ICISET). IEEE,
Chittagong, pp 290–293
3. Iftikhar A, Sun H, Qasim A, Abdul S (2020) Design of umbrella shape single element patch
antenna with high gain and high efficiency for 5G wireless communication in 28 GHz. In:
17th International Bhurban Conference on Applied Sciences and Technology (IBCAST), pp
710–713. IEEE, Islamabad, Pakistan
4. Muhammad Z, Sultan S, Yasar A, Yuriy V, Peter E (2020) Ultra-wideband circular antenna
for future 5G. In: Conference of Russian Young Researchers in Electrical and Electronic
Engineering (EIConRus), pp 2280–2283. IEEE, St. Petersburg and Moscow, Russia
5. Ghanendra K, Chakresh K (2020) Design and analysis of u-slot mircostrip patch antenna for
mobile communication at 60 GHZ. Bull Eng
6. Dennis NA-C, Jorge LA-C (2018) High Gain 4 × 4 rectangular patch antenna array at
28 GHz for future 5G applications. In: XXV International Conference on Electronics, Electrical
Engineering and Computing (INTERCON), pp 1–4. IEEE. Lima, Peru
Design of a Microstrip Antenna Two-Slot for Fifth Generation Applications … 1089

7. Merlin Teresa P, Umamaheswari, G (2020) Compact slotted microstrip antenna for 5G

applications operating at 28 GHz. IETE J Res 1–9
8. Omar D, Dominic B, Onyango K, Franklin M (2019) A 28 GHz rectangular microstrip patch
antenna for 5G applications. Int J Eng Res Technol ISSN 0974–3154, vol 12, no 6, pp 854–857
9. Khraisat YSH (2018) Increasing microstrip patch antenna bandwidth by inserting ground slots.
J Electromagn Ana Appl. ISSN: 1942-0730
10. Jyoti S, Agarwal SK (2017) T and L slotted patch antenna for future mobile and wire-
less communication. In : 8th International Conference on Computing, Communication and
Networking Technologies (ICCCNT), pp 1–5. IEEE, Delhi, India
11. Safpbri J, Muhammad AJ, Shaifol II, Mohd NM, Norhafiza H (2018) 28 GHz microstrip patch
antennas for future 5G. J Eng Sci Res 2(4): 01–06
Performance Improvement
of Corner-Truncated Sierpinski Carpet
Fractal Antenna Using U-tree Slots
for Vehicular Communications
and Other Bands

Fatima Ez-Zaki , Abdelilah Ghammaz, and Hassan Belahrach

Abstract This paper describes the design of a Corner-truncated Sierpinski Carpet

Fractal Antenna which resonates at multiple frequencies (1.84, 4.09, 5.9, 7.18, 10.88,
13.17, 15.15, and 15.60 GHz). Different performance parameters like radiation
pattern, Voltage Standing Wave Ratio (VSWR), return losses (S11) are observed
at all the operating frequencies. The FR4 glass epoxy with relative permittivity of
4.4 and a thickness of 1.6 mm is used as substrate material. The proposed antenna has
a simple structure. Antenna investigation is done between 0.5 and 16 GHz frequen-
cies. The antenna is fed by 50 Ohms microstrip feed line and simulated using CST
Microwave Studio. The established antenna can be used for multiple applications
and bands such LTE, WIFI, and DSRC for vehicular communication, in addition to
satellite, radar, and space communications.

Keywords Multiband fractal antenna · Sierpinski carpet fractal · Defected ground

plane · Vehicular communication

1 Introduction

Nowadays, wireless technology requires antenna with wider bandwidth and smaller
dimensions. Therefore characteristics such as light weight, low profile, low cost,
easy to be manufactured and to be integrated with RF devices makes the Microstrip
antenna very attractive research area [1, 2]. Besides self-similarity and space fillings,
frequency independent, compact size make fractal antenna an excellent candidate to
minimize the size of the antenna or to improve its characteristics [2]. Furthermore
Fractal geometry is also combined with electromagnetic theory in order of investi-
gating a new class of radiation, propagation, and scattering problems. Recent studies

F. Ez-Zaki (B) · A. Ghammaz · H. Belahrach

Electrical Systems, Energetic Efficiency and Telecommunication Laboratory, Faculty of Science
and Technology, UCAM, Marrakech, Morocco
H. Belahrach
Royal School of Aeronautics, 40150 Marrakesh, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1091
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_100
1092 F. Ez-Zaki et al.

on the microstrip fractal antenna have been recently reported in [3–11]. Different
fractal geometries, such as Hilbert, or Sierpinski curves, have been used to bring
the benefits of miniaturization [3], wide bandwidth [4–6], and multiband operation
[5–10]. So far the bandwidth of the microstrip fractal antenna is reported to be
narrow. Therefore multiple techniques are used to resolve this task such as using
slots, notches, defected ground structure (DGS) [12–14].
In this paper, we propose a novel truncated antenna based on the Sierpinski carpet
shape of the second order, and a slotted ground. The proposed antenna was simulated
using CST Microwave Studio. Impact of fractal iterations and defected ground plane
was analysed in details.

2 Antenna Design Procedure

Analysis of the proposed structure is obtained by using CST Microwave Studio. A

patch antenna with square shape is designed by using the following equations [15],
the procedure assumes that the specified information includes the dielectric constant
(εr ) of the substrate, fr which represents the resonant frequency and the thickness
(h) of the substrate. The Sierpinski carpet is constructed using square geometry as
figured in Fig. 1. Figures 2, 3, and 4 describe the process of construction for corner-
truncated Sierpinski carpet fractal antenna. The antenna finale structure is depicted
in Fig. 5 including the defected ground plane by inserting U-tree slots.
Antenna 1: The design of desired antenna starts with eventually rectangular patch
which operates at resonant frequency of 5.8 GHz. By using the following equations
for the values of εr = 4.4, f r = 5.8 GHz, h = 1.6 mm and c = 3 × 108 m/s the
rectangular patch dimensions are calculated as following.
c
W =  (1)
2 f r ε2r

Effective dielectric constant:

 
εr + 1 εr − 1 h 1/2
εreff = + 1+ (2)
2 2 W

Patch effective length:

c
Lef f = √ (3)
2 f r εr e f f

Patch length extension because of the fringing effect:

W 
+ 0.264 (εreff + 0.3)
L = 0.412h  Wh  (4)
h
+ 0.8 (εreff − 0.258)
Performance Improvement of Corner-Truncated Sierpinski Carpet Fractal Antenna … 1093

Patch length:

L = L e f f + 2L (5)

For the used frequency of 5.8 GHz the patch width and length are:
W = 15 mm and L = 12 mm, the substrate is of 20 × 20 mm2 , the dimensions of
the full ground plane are 20 × 20 mm2 . The thickness of substrate (h) is of 1.6 mm.
Antenna 2: four semi-circle corners are cut to provide as a perturbation as shown
in Fig. 2.
Antenna 3: To get 1st iteration of the antenna, the base patch geometry is divided
into 9 congruent sub-rectangles in a 3-by-3 grid, and remove the central rectangle as
shown in Fig. 3.
Antenna 4: the same procedure as step 3 each rectangle patch is divided of 1st
iteration geometry into 9 congruent sub-rectangles, and remove the central one as
illustrated in Fig. 4.
Antenna 5: To get the finale geometry a U-tree slots in the ground plane as shown
in Fig. 5.

Fig. 1 Antenna 1 geometry

Fig. 2 Antenna 2 geometry

Fig. 3 Antenna 3 geometry

1094 F. Ez-Zaki et al.

Fig. 4 Antenna 4 geometry

Fig. 5 Antenna 5 geometry

3 S-Parameter and VSWR Characteristics of Various

Antenna Design Stages

3.1 Various Antennas S11 and VSWR Response

Simulated S11 and VSWR versus frequency plots of all the antenna construction
stages are depicted in Figs. 6, and 7 respectively. The S11 for the eventual square
patch are less than −10 dB at 0.82, 5.543, 9.10, and 15.08 GHz frequencies whereas
the corner truncated patch antenna resonates at 0.84, 5.8, 9.39 GHz. The Sierpinski
corner truncated patch first iteration works at 0.87, 5.36, 12.35 GHz similarly to the
second iteration as it can be seen in Fig. 6 for antenna 3 and 4. So far the U-tree
slots Sierpinski corner truncated antenna shows more improved performances and
presents 7 resonant frequencies at 2.15, 4.37, 7.61, 11.3, 13.36, 15.743 GHz (Table 1).

3.2 Effect of Feed Line and DGS in the Investigated Antenna

Response

So further the effect of the feed line dimensions is studied. As can be noted clearly
from the Fig. 8 and Fig. 9 the dimensions of the feed line have relatively smaller
effects on the antenna’s response.
As for the selection of the U-tree dimensions it is greatly dependent on the required
operating bands and they mainly affect the directivity of the antenna’s radiation
patterns. However, the U-tree dimensions are choose in order to get more improved
Performance Improvement of Corner-Truncated Sierpinski Carpet Fractal Antenna … 1095

Fig. 6 S11 results of various Antennas

Fig. 7 VSWR results of various Antennas

1096 F. Ez-Zaki et al.

Table 1 Various antennas results comparison

Antenna 1 Antenna 2 Antenna 3 Antenna 4 Antenna 5
Resonant Frequencies (GHz) 0.82 0.84 0.87 0.87 2.15
5.543 5.8 5.36 5.35 4.37
9.10 9.39 12.35 12.39 7.61
15.08 11.3
13.36
15.743
S11(dB) −14.34 −13.25 −13.77 −13.63 −18.79
−16.017 −13.32 −10.79 −10.19 −11.87
−32.93 −12.39 −21.30 −15.68 −18.54
−14.34 −39.65
−26.83
−22.06
VSWR 1.53 1.55 1.5 1.52 1.26
1.4 1.54 1.8 1.9 1.68
1.18 1.62 1.39 1.39 1.27
1.04 1.02
1.1
1.17

Fig. 8 VSWR results of various Antennas

Performance Improvement of Corner-Truncated Sierpinski Carpet Fractal Antenna … 1097

Fig. 9 VSWR results of various Antennas

antenna features and to attain the desired frequency band of vehicular communica-
tions at 5.9 GHz. In sum simulated results for S11 (dB) and VSWR of proposed
antenna are analysed in next section.

4 Proposed Antenna Results and Discussion

In this section, the simulation results of proposed antenna are analyzed in details.

4.1 S11 and VSWR of the Proposed Antenna

Looking at Fig. 10, it is apparent that as we approach to antenna 5, the antenna reveals
more multiband and wideband behavior due to the increase in paths and lengths of
the antenna geometry by using the DGS technique. As it can be easily noticed from
the graph of the VSWR results the proposed antenna presents a good impedance
matching described by a VSWR closer to 1 within the operating bands.
1098 F. Ez-Zaki et al.

Fig. 10 Investigated antenna S11

Fig. 11 Investigated antenna VSWR

4.2 Directivity of Proposed Antenna

The simulated radiation pattern of the proposed antenna at resonant frequencies are
illustrated in Fig. 12. The designed antenna shows a nearly directional characteristics
in Phi = 90 with a very good directivity of a peak value of 6.87 dBi at almost operating
frequencies. As can clearly observed that the radiation patterns have multiple lobes
at higher frequencies than lower frequencies which leads that the radiation pattern is
distorted at high frequencies (Fig. 11).
Performance Improvement of Corner-Truncated Sierpinski Carpet Fractal Antenna … 1099

Fig. 12 Investigated antenna Directivity

5 Conclusion

The current paper discussed the construction of corner-truncated Sierpinski carpet

antenna steps, besides the effect of inserting U-tree slots on the ground plane
1100 F. Ez-Zaki et al.

has been analyzed. The key features of this antenna is compact size, simple in
construction using Sierpinski fractal geometry. The investigated antenna works at
multiple frequencies (1.84, 4.09, 5.9, 7.18, 10.88, 13.17, 15.15, and 15.60 GHz)
and get improved performance for multi band (C, S, X and Ku) applications. Inves-
tigated antenna can be used for multiple applications and bands such LTE, WIFI,
and DSRC for vehicular communication, in addition to satellite, radar, and space
communications.

References

1. Werner DH, Ganguly S (2003) An overview of fractal antenna engineering research. IEEE
Antennas Propag Mag 45(1):38–57
2. Anguera J, Andújar A, Jayasinghe J et al (2020) Fractal antennas: an historical perspective.
Fractal Fractional 4(1):3
3. Singh K, Grewal V, Saxena R (2009) Fractal antennas: a novel miniaturization technique for
wireless communications. Int J Recent Trends Eng 2(5):172
4. Bhatt S, Mankodi P, Desai A et al (2017) Analysis of ultra wideband fractal antenna designs
and their applications for wireless communication: a survey. In: 2017 International conference
on inventive systems and control (ICISC). IEEE, pp 1–6
5. Soni BK, Singhai R (2018) Design and analysis of minkowskized hybrid fractal like antenna
for multiband operation. Prog Electromagnet Res 80:117–126
6. Ali JK, Abdulkareem S, Hammoodi A et al (2016) Cantor fractal-based printed slot antenna
for dual-band wireless applications. Int J Microwave Wirel Technol 8(2):263–270
7. Bhutani P, Sagar S, Kumar A (2019) Performance analysis of sierpinski carpet fractal antenna
for wireless communication. In: Applications of computing, automation and wireless systems
in electrical engineering. Springer, Singapore, pp 749–758
8. Dhankhar A, Kaur J (2015) Design of sierpinski carpet fractal antenna for multiband
applications
9. Radonic VK, Palmer GS, Crnojevic-Bengin V (2012) Flexible Sierpinski carpet fractal antenna
on a Hilbert slot patterned ground. Int J Antennas Propag 11:7 p, article ID 980916
10. Sivia JS, Kaur G, Sarao AK (2017) A modified sierpinski carpet fractal antenna for multiband
applications. Wireless Pers Commun 95(4):4269–4279
11. Wong S, Ooi BL (2001) Analysis and bandwidth enhancement of Sierpinski carpet antenna.
Microwave Opt Technol Lett 31(1):13–18
12. Gupta A, Joshi HD, Khanna R (2017) An X-shaped fractal antenna with DGS for multiband
applications. Int J Microwave Wirel Technol 9(5):1075
13. Arya AK, Kartikeyan MV, Patnaik A (2010) Defected ground structure in the perspective of
microstrip antennas: a review. Frequenz 64(5–6):79–84
14. Garg C, Kaur M (2014) A review of defected ground structure (DGS) in microwave design.
Int J Innovative Res Electr Electron Instrum Control Eng 2(3):1285–1290
15. Balanis CA (2005) Antenna Theory: Analysis and Design, 3rd edn. Wiley-Interscience,
Hoboken
A Novel System Based V2V
Communications to Prevent Road
Accidents in Morocco

Zakaria Sabir and Aouatif Amine

Abstract Due to the deaths and injuries caused by road accidents every year, road
safety became an important field of research for many institutes. The development
of smart solutions can remarkably enhance road safety. Vehicular networks are one
of those solutions. It can bring advanced services, namely collision detection, traffic
management, communication between vehicles, and so on. This paper introduces an
implementation of a complete vehicle communication system. The prototypes are
developed using Raspberry Pi boards and other sensors. The results demonstrate the
efficiency of the presented safety scenarios in dealing with the dangers of the road.

Keywords Connected vehicles · V2V · ITS · Prevent accidents · Road safety ·

Vehicular communications

1 Introduction

Road safety has become a very important subject during the last years. Every year,
millions of people die and thousands are injured due to road accidents. The World
Health Organization (WHO) reported that 1.25 million people died in 2013 and
more than one million accidents were recorded in 2014 in Europe, resulting in 1.4
million injuries and 25 900 deaths. In Morocco, 3593 persons were killed in 2016
against 3499 in 2017 [1], which represents a limited reduction of 2.6% year to
the other. Even if the statistics show a minor decrease, a lot of worrying facts are
hidden [2]. These statistics are so flagrant that the Ministry of Equipment, Transport,
Logistics, and Water (METLE) dedicated a call for projects especially in the field
of road safety, in cooperation with the National Center for Scientific and Technical
Research (CNRST) and Moroccan universities. The aim of those projects is first

Z. Sabir (B) · A. Amine

ILM, ENSA, Ibn Tofail University, Kenitra, Morocco
e-mail: [email protected]
A. Amine
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1101
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_101
1102 Z. Sabir and A. Amine

of all decreasing the number of accidents that occur every year in Morocco. Our
project entitled “SafeRoad: Multiplatform for Road Safety (MRS)” is one of those.
His main goal is to reduce the collision rate using the anticipation and the detection
of accidents. And since the collection of statistics of the corporal accidents of road
traffic carried out by the METLE in 2016 [3] reported that there were 80680 accidents
caused by obstacles, we considered also working on the detection of obstacles using
vehicular communications.
The automation/semi-automation of vehicles can help humans avoiding errors
that cause most road accidents. Constructors are developing smart vehicles that will
be supplied by a variety of technologies and connected. They will assist drivers
by alerts indicating the dangers on the road. Principally, they will be connected to
a variety of sensors within the vehicle: warning lights, stability, diagnostic tools,
airbags, and others. Thus, a vehicle is able to automatically disseminate information
to others, like the presence of a slippery roadway, if it is down or in an accident. If the
technologies of networking, electronics, and computing are well integrated, the roads
will be safer and the impact on the environment will be minimized [4]. Recently,
diverse definitions have been given to a “connected” vehicle. A description that
involves all other connections of a smart vehicle is provided by the US Department
of Transportation (USDOT): “Connected vehicle applications provide connectivity
between vehicles to prevent accidents, between vehicles and infrastructure, in order to
generate safety, mobility and environmental benefits; between vehicles, infrastructure
and wireless devices to provide continuous real-time connectivity to all users of the
system” [5]. The system of connected vehicles [6] will operate using Vehicle to
Infrastructure (V2I) [7] and Vehicle to Vehicle (V2V) [8] communications.
In this paper, we present our solution to the problem of road accidents, especially
those caused by obstacles. We introduce an implementation of a vehicular commu-
nications testbed. We focus on V2V technology, and we carry out our experiments
using mini cars based on Raspberry Pi boards. We add different sensors and features
to develop our complete prototypes and we propose two road safety scenarios that
can help to avoid accidents. The next sections of the paper will be as follow: Sect. 2
discusses the related work, Sect. 3 recalls some vehicular network characteristics.
The proposed system is presented in Sect. 4, Sect. 5 Concludes the paper.

2 Related Work

In the early 1980s, research that treaded inter-vehicular communications was

conducted in Japan by the Association of Electronic Technology for Automobile
Traffic and Driving [9]. In the 1990s, other projects, like the California PATH program
[10] and the European “Chauffeur” [11], have deployed and considered automated
platooning systems via the dissemination of data over vehicles. Since 2000, many
research institute, private companies, and automobile manufacturers have supported
worldwide projects concerning vehicular safety using wireless communications:
A Novel System Based V2V Communications … 1103

• FleetNet: This German project was initiated by six companies and several research
institutes [12]. The project aimed to develop and demonstrate a wireless ad-
hoc network for vehicular communications through a platform of cooperative
driver-assistance systems, in order to ensure comfort and security for drivers and
passengers.
• The Car2Car communication consortium was inaugurated by European automo-
bile manufacturers. The goal was to improve road safety and to manage effectively
the traffic thanks to the use of inter-vehicle communications [13]. This project had
several missions and intends to develop a V2V system prototype for road safety
applications.
• CarTalk2000 [14] is a European project which aimed to establish, test, and evaluate
an organized and pure ad hoc wireless network and also to develop cooperative
driver assistance systems, in order to enhance global traffic safety and to increase
driving comfort and transport efficiency.
• NoW (Network on Wheels) [15] is a German research project supported by the
government and implemented by research institutes and universities, providers,
and vehicle constructors. The goal is to develop a vehicular communication system
dedicated to road safety, traffic efficiency and infotainment applications, radio,
security, and privacy.
• The Integrated Project PReVENT [16] is a European automotive industry activity,
started in February 2004 with the goal of establishing and validating preventive
safety applications and technologies. Co-funded by the European Commission,
PReVENT stands for Preventive safety and act in accordance with the objective
of the safety initiative to pull down accidents and traffic death in Europe.
• MobiVIP is a research project of Predit 3 (Integration of the Communication
and Information systems Group) [17]. Its objective is evaluating and showing the
influence of the information and communication technologies on a new service
formulated by small urban vehicles for mobility in town centers.
In our work, different from those researches, we focus on the Moroccan case and
we propose a system that we will be adapted to fit the Moroccan situation since
a foreign system cannot be applied directly in Morocco. Many factors are to take
into consideration such as the culture, the environment, the driver behavior, and the
architecture of the cities. We succeeded in the first step using mini cars and we will
move to real tests with real vehicles in the next step.

3 Vehicular Networks Characteristics

Vehicular ad hoc networks represent a special case of mobile ad hoc networks.

However, research works studied and accomplished in this area cannot be employed
straight in the field of vehicular networks. There are several characteristics of vehic-
ular networks that must be taken into consideration before applying the classical
architectures and protocols of ad hoc networks. We present in the following some
1104 Z. Sabir and A. Amine

features and restrictions related to vehicular networks which differentiate them from
ad hoc networks:
• Neighborhood recognition: drivers can receive various information that will
help them ameliorate their visibility and behave accordingly to any adjustments
in the surroundings of the vehicle. This information could be the state of the road,
weather conditions… etc.
• Elevated processing capacity: vehicles can provide powerful communication,
sensing, and computing capacity. Nowadays, they are capable of analyzing and
interpreting the assembled data in order to help drivers to make good decisions.
• Extensive scope: as opposed to other ad hoc networks that adopt generally a
restricted network scope, vehicular networks are able to reach the whole route
system and thus cover a lot of users.
• Unlimited transmission power and adequate storage: vehicles have big storage
and can provide stable power to computing and communication equipment. They
can have also different communication interfaces such as Bluetooth, Wi-Fi, and
others.
• Changes of the topology: vehicular networks are distinguished by fast move-
ments due to the vehicles’ speed that can reach 200 km/h. Therefore, nodes connect
and disconnect quickly and the topology changes frequently. This points to poor
connectivity and minimal road life.
• Anticipated movement: usually, vehicles’ mobility is restricted by the roadways,
they are often connected to Road Side Units (RSUs), roadside infrastructures
(like panels and traffic lights), or highways. This makes the anticipation of their
movement easier, in contrast to classic mobile ad hoc networks, where nodes
move arbitrarily.
• Diverse communication status: the two habitual communication status in vehic-
ular networks are outside and inside the town. In the first one, the area is almost
easy and the roads are usually straight (especially on highways). However, in the
second one, the circumstances are quite difficult. Inside the town, obstacles such
as buildings may isolate the paths. Consequently, communication is not always
continuous.

4 Proposed System

Today’s traveling manner can be transformed and improved through the use of
connected vehicles. They will help to create a wireless communication network, safe
and interoperable, which includes various components (trains, cars, buses, traffic
light …). Connected vehicle technologies address some huge challenges in several
fields such as environment, safety, transportation, and mobility.
The aim of our work is part of the road safety field, particularly in the Moroccan
case in order to resolve the problems mentioned above. We propose a new system
which can be used to notify Moroccan drivers in various circumstances. It will
contribute to reducing the number of accidents and avoiding the crash that frequently
A Novel System Based V2V Communications … 1105

Fig. 1 Example of V2V

communication. The first
vehicle stops and sends a
warning to other vehicles
after detecting an obstacle on
the road

occurs between vehicles or between vehicles and bikes, animals, pedestrians, or other
objects. Our system uses V2V communications and it is based on reel time informa-
tion. Vehicles interchange important information that is valuable to predict a risky
situation. Thanks to this system, the danger of unstructured roads in Morocco can
be avoided by delivering the aforementioned information to drivers.
Considering the statistics cited in [3], the first road safety scenario in our system
can alert a driver of an obstacle in the road. An example is depicted in Fig. 1, where
three vehicles are traveling in one lane and the first vehicle stopped because of a cow
crossing the road. The second vehicle is blocking the vision; hence the third vehicle
can’t see the first one. Thanks to V2V communication, the last vehicle is capable of
receiving a warning about the hard braking and avoiding a possible accident.
The second road safety scenario we propose aims to alert a driver about a specific
traffic condition on the road. The driver will then reduce the speed and change
his behavior regarding the situation (accident, road work, …). He may consider
changing the route and taking a safer one. Due to V2V communication, the warning
is transmitted by a vehicle ahead that has already discovered the problem. Figure 2
shows an explicative scheme of our system.

4.1 Experimental Setup

In order to validate our proposition, we started by testing our system in prototypes

before moving to real case testing. To develop our complete prototype, we used two
mini cars from Sunfounder [18] (Fig. 3). Each one is based on a Raspberry Pi 3 B+
[19] and has the indispensable configuration to authorize the control of movements,
communication, and sensors. The specifications of the used boards are as follow:
they are based on an ARM Cortex-A53 quad-core processor which is operating at
1.4 GHz and have 4 USB port 2.0, an Ethernet port, a memory of 1 GB (RAM), a
micro-SD slot, a camera interface, and an audio jack. A micro USB port is used for
power (5V/2.5 A).
Each car is equipped with a DC motor driver which power its movement, and
since they have integrated Wi-Fi, they have the ability to communicate within the
1106 Z. Sabir and A. Amine

Fig. 2 Explicative scheme of V2V Communication

Fig. 3 Mini car from

Sunfounder [18]

Fig. 4 Our car prototype

test zone using a local ad hoc network. They are also equipped with a step-down
DCDC converter module, a USB camera, and a PCA9685-based servo controller.
In order to build our complete prototype, we had to add some other sensors since
the original kit didn’t come with all the necessary equipment. In Fig. 4 we can see the
added sensors which are the infrared obstacle detector and the line follower sensors.
This figure depicts also the location of each sensor. The infrared sensor will be used
A Novel System Based V2V Communications … 1107

to detect any obstacles on the road while the line follower sensors will help to keep
the car on the road. The details of the sensors are provided in Table 1.

Table 1 The specifications

Sensor name Line follower sensor Infrared obstacle
of the used sensors
detector
Reference TCRT5000 FC-51
Detection distance 1 mm–25 mm 2 cm–30 cm,
Angle: 35
Operating voltage 5V 3.3–5 V
Quantity 5 1

Fig. 5 V2V communication

in our testbed. V1 detects the
obstacle on the path and
sends the warning to V2

Fig. 6 V2V communication

in our testbed. V1 detects the
obstacle on the path and
sends the warning to V2
which changes the path taken
by V1 to get to the
destination
1108 Z. Sabir and A. Amine

4.2 Results and Discussion

Using V2V communications, our road safety scenarios aim to disseminate an alert
informing the drivers of the existence of an obstacle on the roadway. As a first try,
we attempt to run the cars on a map, so they follow pre-programmed distances.
The result was that the cars couldn’t move following a straight line, due to the
little difference between the motors of each car. Therefore, we preferred to use line
follower sensors to help cars following the road. The first road safety scenario is
about the hard-braking warning. We put an obstacle on the road and ran the two
cars at once. Since the first car (V1) is equipped with an infrared sensor (see Fig. 4),
it detects the obstacle and sends the alert to the second car (V2) using the Wi-Fi
interface. The moment it received the alert, V2 stops as well (see Fig. 5). Receiving
the warning message depends on the Wi-Fi range (the maximum distance between
V1 and V2).
The second road safety scenario aims to send a warning to the drivers, so they
can choose the safer road. Similarly, we put an obstacle on the path before the final
destination and we ran V1 and V2 at the same time.
Once the infrared sensor senses the obstacle, V1 stops instantly and sends the
alert to V2 which changes the current path (taken by V1) and chooses the safer one.
Figure 6 illustrates the second road safety scenario.

Table 2 Comparison of our work with other studies

Study Communication Implementation Tools Scenario Purpose Special
mode case
[20] V2V Real time RPi; Blind spot Improving Not
Ultrasonic detection traffic specified
sensors; congestion
Audible
Alarm;
GPS
Sensor; RF
Module
[21] V2V Real time OsmAnd; Displaying Warning Not
Android location of about specified
Devices; relevant emergency
GRCBox vehicles vehicles
[22] V2V Real time RPi; Autonomous Smooth Not
Controller; driving traffic flow specified
IR Sensor (level 3)
Our V2V Real time RPi mini Obstacle Preventing Moroccan
work cars; IR detection accidents Case
sensor;
Line
Followers
A Novel System Based V2V Communications … 1109

We present a comparison of our proposed system with other studies in the literature
in Table 2.

5 Conclusion

We reached our target and succeeded in achieving the solution. The primary consider-
ation of the paper was to introduce an implementation of a complete vehicle commu-
nication system using V2V technology in the context of the SAFEROAD project.
Thanks to the partnership with the METEL, we obtain real information about road
accidents in Morocco. Our solution is developed especially for the Moroccan case
and it is the first of its kind in our country. The aim was to send an alert to the drivers
in time, in order to help them avoiding road dangers and so decreasing the number
of accidents. We conducted our experiments using mini cars based on Raspberry Pi
as well as infrared obstacle detector and line follower sensors. The results affirm the
usefulness of our system in the field of road safety. In future work, we will confirm
the reality on the ground by applying this system to real cars and test it in different
cities, especially those where a grand number of accidents occur.

Acknowledgment This research work is supported by the “SafeRoad: Multiplatform for Road
Safety (MRS)” Project under contract No: 24/2017, financed by the Ministry of Equipment, Trans-
port, Logistics and Water (METLE), and the National Center for Scientific and Technical Research
(CNRST).

References

1. World Health Organization (2018) Global status report on road safety 2018 (WHO). World
Health Organization, Geneva
2. Leconomiste: Accidents routiers. https://www.leconomiste.com/flash-infos/accidents-routiers-
au-maroc-tue-presque-autant-qu-en-france-avec-7-fois-moins-de. Accessed 08 July 2019
3. Ministry of Equipment, Transport, Logistics, and Water, Direction of Roads (2017)
Compendium of Statistics of Road Traffic in Personal Accidents
4. Guerrero-Ibáñez JA, Flores-Cortés C, Zeadally S (2013) Vehicular ad-hoc networks (VANETs):
architecture, protocols, and applications. In: Chilamkurti N, Zeadally S, Chaouchi H (eds)
Next-generation wireless technologies. Springer, London, pp 49–70
5. Brookes R, Pagani P (2014) What becomes a car. In: Proposed paper for: BIT 2014 conference
workshop-technology enabled business models: platforms, analytics and performance, March
2014
6. Sabir Z, Amine A (2020) NDN vs TCP/IP: which one is the best suitable for connected vehi-
cles? In: Proceedings of the first international conference on technology, engineering, and
mathematics, Kenitra, Morocco, 26–27 March 2018. Springer, Cham, pp 151–159
7. Sabir Z, Dafrallah S, Amine A (2019) A novel solution to prevent accidents using V2I in
Moroccan smart cities. In: 2019 international conference on computational intelligence and
knowledge economy (ICCIKE), Dubai, UAE, pp 621–625
1110 Z. Sabir and A. Amine

8. Sabir Z, Amine A (2018) Connected vehicles using NDN for intelligent transportation systems.
In: International conference on industrial engineering and operations management, Paris,
France, pp 2433–2441
9. Tsugawa S (2005) Issues and recent trends in vehicle safety communication systems. IATSS
Res 29:7–15
10. Hedrick JK, Tomizuka M, Varaiya P (1994) Control issues in automated highway systems.
IEEE Control Syst Mag 14:21–32
11. Gehring O, Fritz H (1997) Practical results of a longitudinal control concept for truck
platooning with vehicle to vehicle communication. In: Proceedings of conference on intelligent
transportation systems, Boston, MA, USA. IEEE, pp 117–122
12. Enkelmann W (2003) FleetNet - applications for inter-vehicle communication. In: IEEE IV2003
intelligent vehicles symposium. Proceedings (Cat. No. 03TH8683), Columbus, OH, USA.
IEEE, pp 162–167
13. C2C-CC: CAR 2 CAR Communication Consortium. https://www.car-2-car.org/. Accessed 24
June 2019
14. Reichardt D, Miglietta M, Moretti L, Morsink P, Schulz W (2002) CarTALK 2000: safe and
comfortable driving based upon inter-vehicle-communication. In: IEEE intelligent vehicle
symposium, 2002, vol 2, pp 545–550
15. Festag A, Noecker G, Strassberger M, Lübke A, Bochow B, Torrent-moreno M, Schnaufer
S, Eigner R, Catrinescu C, Kunisch J (2008) ‘NoW – network on wheels’: project objectives,
technology and achievements, Hamburg, Germany
16. Schulze M, Nöcker G, Böhm K (2005) PReVENT: a European program to improve active
safety. In: Proceedings of 5th international conference on intelligent transportation systems
telecommunications, France
17. Despeyroux T, Masseglia F, Rossi F, Senach B (1992) User-centered design, improvement and
analysis of information systems. Bull Sociol Methodol/Bulletin de Méthodologie Sociologique
37:55–57
18. Sunfounder: Smart Video Car for Raspberry Pi. https://www.sunfounder.com/learn/category/
Smart-Video-Car-for-Raspberry-Pi.html. Accessed 15 July 2019
19. Buy a Raspberry Pi 3 Model B+ – Raspberry Pi. https://www.raspberrypi.org/products/raspbe
rry-pi-3-model-b-plus/. Accessed 01 Aug 2019
20. Ghatwai NG, Harpale VK, Kale M (2016) Vehicle to vehicle communication for crash avoid-
ance system. In: 2016 international conference on computing communication control and
automation (ICCUBEA), pp 1–3
21. Hadiwardoyo SA, Patra S, Calafate CT, Cano J, Manzoni P (2017) An android ITS driving safety
application based on vehicle-to-vehicle (V2V) communications. In: 2017 26th international
conference on computer communication and networks (ICCCN), pp 1–6
22. Tayeb S, Pirouz M, Latifi S (2017) A raspberry-pi prototype of smart transportation. In: 2017
25th international conference on systems engineering (ICSEng), pp 176–182
Detection of Underground Cavities Using
Electromagnetic GPR Method in Bhalil
City (Morocco)

Oussama Jabrane, Driss El Azzab, Mahjoub Himi, Mohammed Charroud,

and Mohammed El Gettafi

Abstract The presence of underground cavities of natural or anthropogenic origin in

urban areas presents a major risk, which causes severe problems for civil engineering
and environmental management, damage to properties and infrastructures, and can
trigger major building collapse.
A geophysical survey was carried out, based on the sequential application of
magnetic, low-frequency ground penetrating radar (GPR) to locate and detect under-
ground cavities and voids below the surface which have a high relative dielectric
permittivity (RDP). Ten profiles were carried out in a karstic location in Bhalil
Medina listed as an environmentally sensitive area, giving 10 radargrams as results.
The latter are processed to extract the maximum information and to be easily corre-
lated with the geological study (geological section and the lithostratigraphic column).
The observed RDP anomalies can be related to the presence of air-filled cavities
within the shallow subsurface, small geological structures, or a type of formation of
high reflection of the signal, and revealed the presence of a high degree of karstifi-
cation in the inspected site. The presence of such features prevents the extension of
urbanizations, and endangers the existing habitats, especially in the old town like the
Medina of Bhalil.

O. Jabrane (B) · D. El Azzab · M. Charroud

Faculty of Science and Technology, SIGER Laboratory, Sidi Mohamed Ben Abdellah University,
BP2202 Fez, Morocco
e-mail: [email protected]
D. El Azzab
e-mail: [email protected]
M. Charroud
e-mail: [email protected]
M. Himi
Economic and Environmental Geology and Hydrology Group, University of Barcelona,
Barcelona, Spain
e-mail: [email protected]
M. El Gettafi
FPN, G-ES Laboratory, Mohammed Premier University, 62999 Nador, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1111
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_102
1112 O. Jabrane et al.

Keywords Ground penetrating radar · Cavities · Bhalil · Karst · Voids · Survey

1 Introduction

Karst processes often result in the formation of underground cavities due to the effect
of chemical erosion by the circulation of acidic water (dissolution) on carbonate
rocks.
The urban area of Bhalil continues to develop locally on more or less unstable
grounds, in particular its old town, subjected to multiple constraints. These are linked
to rapid population growth coupled with the fragile geological nature of the under-
lying formations. Another critical aspect of these cavities lies in the fact that they
can provide a safe and consistent habitat for many people in this area.
Underground cavities detection is one of the primary objectives of geophysical
investigation techniques because of the danger that threatens the inhabitants causing
soil instability, which can provoke dramatic events such as subsidence and even total
collapse to the ground surface [6, 15]. The diagnosis of these buildings condition is
carried out regularly by the authorities sometimes forces them to evacuate certain
houses deemed to be in danger. For the subsurface identification and mapping of
such karst features, the non-invasive and high-resolution geophysical techniques
have appeared as an appropriate choice [12].
The present case study, applied Ground-penetrating radar (GPR) for the site char-
acterization, which uses radar pulses to image the subsurface. It is a non-intrusive
prospecting technique based on the propagation of electromagnetic waves in the
underground to explore the presence of a medium RDP [13]. It has become one of
the most important geophysical methods for cavities and voids detection, to delimit
them laterally and in-depth according to its efficiency, high resolution, rapid execu-
tion and simultaneous result by imaging the subsurface. Mapping dangerous areas
in the Old Medina are an asset in terms of future urban development in the city,
protection of people’s properties, and especially their lives.
Our investigation consists of two complementary parts: i/ Geological study to
identify the origin of the observed voids and their consequence at the surface; ii/
prospecting and locating the underground cavities using indirect and non-destructive
electromagnetic GPR method.

2 Geology of the Study Area

2.1 Site Description

The study area is located in a small parcel around the old town of Bhalil, which is
28 km south of the city of Fez. The town of Bhalil is 950 m above sea level, it is a
transition zone between the tabular Middle Atlas, to the south, and the Sais basin, to
the north.
Detection of Underground Cavities Using Electromagnetic GPR Method … 1113

Karst features presence greatly increases the risk of underground cavities devel-
opment by a cluster of inter-related processes, including bedrock dissolution, soil
sapping, rock collapse, and soil collapse. The presence of underground cavities in
the basement of an urban area can cause severe effects and threatens the lives of the
inhabitants.

2.2 Geological Setting

The geological deposits layers, tectonic and geomorphological settings, combined

with the anthropic action, makes the old town of Bhalil and its surroundings a region
at high risk of potentially destructive collapse because of its karstic nature.
The Bhalil basin is limited to the east by an accident in the direction of N170
parallel to the P.20 road, to the west by the Bhalil-Sefrou fault, to the northwest by
the Bhalil inlier, and limited to the south by Aggay River. The region knows a diverse
lithostratigraphy (Fig. 1B) ranging from the Paleozoic deposits outcrops in 3 inliers
(Kandar, Beni Mellala and Bhalil), then the Jurassic layers which appears in the
South and the North, overcoming geological formations of Triassic age, composed
of red gypsum-salt clays with intercalation of often altered doleritic basalts. These
Paleozoic and Mesozoic terrains are drowned northward under recent formations of
the Upper Miocene and Plio-Quaternary age.
The geological formations of the Upper Miocene are mainly marine and composed
at the base of a mollasse, made up of conglomerates and sandstone, which is first
surmounted by pinkish to beige fossiliferous limestones then ocher silts and blue
sandy marls [3, 4, 9, 11].
The old town of bhalil is built on recent geological formation (Fig. 1A), mainly
formed by, Miocene age limestones with conglomeratic units, early quaternary

Fig. 1 A-Geological section showing the layout of the outcrops in the study area [7], B-
Lithostratigraphic column showing the main formations described in the study area
1114 O. Jabrane et al.

(villafranchien) tufa and travertine deposits. The latter, contains many voids and
cavities, due to the natural karst phenomenon, or to the human action consisting in
digging into the rocks for underground habitats. A final recent layer consists of late
quaternary (colluvial deposits that conceal the cavities).

3 Methodology

3.1 Basic Theory

This study was conducted to detect subsurface cavities using GPR (Ground Pene-
trating Radar) designed by the Swedish company Mala geosciences. The advantage
of this technique is its high-resolution results, faster operation, equipment portability
and simultaneous picturing of sub-soil in terms of RDP.
This technique is based on the use of two antennas, one for transmitting electro-
magnetic (EM) waves, and the other for receiving these waves after reflection on
the RDP targets. The two antennas are connected to a console, itself connected to a
laptop that allows data visualization during acquisition, or to configure the console. A
simple automotive battery supplies the power, the transmitter sends electromagnetic
waves, ranging between two frequencies from 10 MHz to 2 GHz. The signal is not a
single sinusoid at the designated frequency, but a sum of sinusoids whose spectrum
is centered on that frequency [2].
The investigation depth range using GPR method depends on the DPR which is
related to the electrical conductivity of the underground layers, transmitted center
frequency of antennas and the radiated power [8, 10], while the EM velocity is
conditioned by the dielectric permittivity of the soils [1, 5, 14].
The dielectric permittivity can be expressed as:

ε = ε = εr ε0 (1)

Where εr is the relative permittivity, and ε0 is the dielectric permittivity in space

(ε0 = 8.854187817 × 10−12 F.m−1 ).
The GPR method suffers from a major handicap as it is attenuated by various factor
and especially the electrical conductivity. While the attenuation constant depends on
the physical properties of the media. In general, the attenuation constant can be
expressed as:

  1/ 2
με  σ 2 1/ 2
∝= ω 1+ −1 (2)
2 ωε

Where ε, μ and σ are absolute permittivity, magnetic permeability and electrical

conductivity of the subsurface material respectively.
Detection of Underground Cavities Using Electromagnetic GPR Method … 1115

Radiowaves propagate through different materials at different speeds. The velocity

of the radiowaves depends on the physical properties of the medium. In general, the
velocity of radiowaves through a homogeneous material is given by:

⎡ ⎤−1/2
  σ 2 1/2
2 ⎣
V = 1+ − 1⎦ (3)
με ωε

When σ is small as compared with ωε, the EM velocity can be expressed as:

C
V =√ (4)
εr

Where V is relative velocity of propagation of electromagnetic wave pulse in a

solid medium (m/s), C is velocity of the electromagnetic wave in the air, equal to
light 0,3 m/ns, and εr is relative dielectric constant of the material.

3.2 Field Investigations

The measurements were carried out in a parcel of the Rquiba district (Fig. 2), using
two antennas RAMAC GPR. For all ground-penetrating radar profiles, it was neces-
sary to choose some characteristics of implementation in the field. After several tests
to avoid sever attenuation and to get the best resolution, we opted for a recording

Fig. 2 A Satellite image showing the survey area, B map showing profiles arrangement in Rquiba
district
1116 O. Jabrane et al.

time of 200 ns. It corresponds, in the context of the study area, to an investigation
depth of about 10 m and bistatic type antennas with a central frequency of 200 MHz.
The profiles have different lengths and different orientations, 6 profiles which are
oriented NNE-SSW perpendicular to the length of the recognized cavities in the
studied area, separated by a distance of 1.2 m, and 4 profiles perpendicular to the
previous profiles separated by a distance of 3 m.

3.3 Processing

The geophysical ground penetrating radar work carried out as part of this study
enabled us to record a large amount of data. The radargrams collected systematically
underwent the processing sequence.
To improve the image provided by the radiograms, it is necessary to apply certain
processing to the raw data from the ground penetrating radar, in order to increase
the signal/noise ratio of the radar sections, and to make reading and interpretation
easier. They help to attenuate the physical phenomena inherent in the propagation of
electromagnetic waves in the ground so as to obtain an image that reflects the object
to be identified. The processing applied in our work include in particular:
• Time-zero correction for shifting the traces along the time axis.
• Removing the low-frequency components (“wow” operation) of the GPR record;
it reduces the data to a mean zero level.
• Band-pass filtering (1D filter band pass frequency): Eliminate the high-frequency
(noise).
• Signal gain «energy decay»: Applying a gain to compensate the signal attenuation.
The final sections obtained were then carefully analyzed in order to extract as
much information as possible on the structure of the subsoil investigated and on the
possible existence of underground cavities within the plot.

4 Results and Discussion

The ground-penetrating radar profile shows areas of strong linear reflections and
areas of very low amplitude. We associate these with the voids contained in the
travertines layer, numbered A1, A2, and A3 (Fig. 3). We can follow spectacularly
the first decimeters of travertines formation before encountering disturbances linked
in our case to the presence of 3 anthropic cavities.
Alongside the anomalies related to voids, discontinuities have been highlighted
that we associate to small fractures, which are sloping towards the North East.
Strong reflections continue up to 20 ns. We measured the thickness of the roof
of the cavities related to a small subsurface travertine formation, which was about
0.5 m on average.
Detection of Underground Cavities Using Electromagnetic GPR Method … 1117

Fig. 3 GPR Profile 1 with depth scale, showing the arrangement of the 3 cavities

Fig. 4 GPR Profiles (a, b, c, d. and e.) showing all A1 anomalies, (a, and b. profiles) showing the
second cavity anomaly A2
1118 O. Jabrane et al.

Fig. 5 Perpendicular GPR Profiles (f, j, h, and i.), radar-gram (f, and h.) showing the same anomaly
A2

We can follow these voids in profiles 3, and 4 (Fig. 4), in which we note the
absence of anomaly A3. As a result, the cavity linked to this anomaly is very limited
in space. The cavities corresponding to anomalies A1 and A2 continue deeper to
appear in profile 4, implying a length of about 6 m. From the fifth profile, we note
that all kind of anomalies disappear, so that indicates a more or less homogeneous
level of the same travertine formation.
The profiles from 7 to 10 (Fig. 5), are perpendicular to the first ones, so they follow
the cavities but present little disturbance. The profile 7 and 9 which are perpendicular
to profile 1 and passes through the middle, show an anomaly appears close to the
surface, attributing in our case to anomaly A2 of profile 1 which indicates the presence
of an underground cavity in the same travertine formation.

5 Conclusion

In this work we demonstrated the efficiency of the GPR application as a very powerful
tool for mapping distribution of different anomalies related to cavities in urban area.
Detection of Underground Cavities Using Electromagnetic GPR Method … 1119

The results are very encouraging, as most of the cavities reported by the inves-
tigation have been confirmed. We were able to follow these voids limit in the three
directions of space at the scale of the plot studied. It was therefore possible to select
the largest cavities from the small ones thanks to the large number of profiles and
their arrangements. The very fine spacing of the measurements (20 cm) also made
it possible to highlight small fractures that affect the travertines, which are very
limited in space (a few meters) and sloping NE direction, which is consistent with
the relatively large structures mentioned in previous geological work in the area.
In this work the description of danger zones (areas with potential risk evidenced
by cracks and voids in-depth), requires additional GPR surveys at the western central
limit of the urban area in the old Bhalil, to define the necessary preventive measures.
A better geotechnical knowledge of the underlying layers would also be beneficial
seeking to minimize the risk of the possible collapse and landslides. From a geotech-
nical point of view, all the new infrastructure and facilities must be located in a stable
area to avoid construction disorder which represents a high financial impact for many
owners.

Funding This research was funded by AECI project: A/012172/07: Geophysical investigating for
the delimitation of the unstable zones of the city of Fez and vicinity.

References

1. Annan AP (2003) Ground-penetrating radar principles, procedures & applications. Sensors and
Software Inc, Mississauga
2. Bano M (1996) Constant dielectric losses of ground-penetrating radar waves. Geophys J Int
124(1):279–288
3. Charriere A (1990) Héritage hercynien et évolution géodynamique alpine d’une chaîne intra-
continentale: le Moyen Atlas au SE de Fès (Maroc). Thèse, Doctorat d’Etat, Toulouse, p
589
4. Charroud M, Cherai B, Benabdelhadi M, Charroud A, El Moutaouakkil N, Falguères C, Lahrach
A (2006) Sedimentary evolution of a fore-chain Sais basin during plio-quatrenary and modali-
ties of tectonic inversion (Sais basin, Morocco). In: Geophysical research abstracts, European
geosciences union, vol 8, p 10039
5. Daniels DJ (2004) Ground penetrating radar. The Institution of Electrical Engineers, 2nd edn.
London
6. De Bruyn IA, Bell FG (2001) The occurrence of sinkholes and subsidence depressions in the far
West Rand and Gauteng province, South Africa, and their engineering implications. J Environ
Eng Geosci 7(3):281–295
7. Fassi D (1999) Les formations superficielles du Saïs de Fès et de Meknès. Du temps géologique
à l’utilisation actuelle des sols. Notes et Mémoires du Service Géologique, Maroc, pp 389–527
8. Franke JC, Yelf RJ (2003) Applications of GPR to surface mining. In: Proceedings of the 2nd
international workshop on advanced ground penetrating radar, University of Technology, Delft,
The Netherlands, pp 115–119
9. Gourari L, Boushaba A, Ramon J, Akdim B (2000) Hydrochimie, processus, facteurs de
précipitation des encrôutements travertineux et les causes géo-environnementales du ralen-
tissement de leur formation actuelle dans la bassin karstique de l’Oued Aggai (Causse de
Sefrou, Moyen-Atlas, Maroc). Comm Inst 115–133
1120 O. Jabrane et al.

10. Gregoire C, Halleux L (2002) Characterisation of fractures by GPR in a mining environment.

First Break 20(7):467–471
11. Hinaje S (2004) Tectonique cassante et paléochamps de contraintes dans le Moyen et le haut
Atlas (Midelt-Errachidia) depuis le Trias à l’Actuel. Thèse d’Etat Es-Sciences, Rabat, p 425
12. Pazzi V, Di Filippo M, Di Nezza M, Carlà T, Bardi F, Marini F, Fontanelli K, Intrieri E,
Fanti R (2018) Integrated geophysical survey in a sinkhole-prone area: Microgravity, electrical
resistivity tomographies, and seismic noise measurements to delimit its extension. Eng Geol
243:282–293
13. Pueyo-Anchuela O, Casas-Sainz AM, Soriano MA, Pocovi-Juan A (2010) A geophysical
survey routine for the detection of doline areas in the surroundings of Zaragoza (NE Spain).
Eng Geol 114(3–4):382–396
14. Vasudeo AD, Katpatal YB, Ingle RN (2009) Uses of dielectric constant reflection coefficients
for determination of ground-penetrating radar. World Appl Sci J 6(10):1321–1325
15. Waltham T, Bell F, Culshaw M (2005) Sinkholes and Subsidence. Karst and Cavernous Rocks
in Engineering and Construction. Springer, Berlin, p 382
An Adapted Routing Protocol for Mobile
Edge Computing

Ali Ouacha

Abstract In the current time, technological progress produces devices that have
become smaller in size. Performing tasks by these devices is in most cases a deli-
cate task due to their limited performances (such as processing capacity, storage and
connectivity), especially when dealing with large data size and when having time
constraints. However, Mobile Edge Computing (MEC) comes to overcome these
weaknesses by delegating the execution of heavy tasks to other peripheral nodes
with more powerful in computing capacity (Edge Server). In addition, these devices
form a mobile network in which communication between them is based on routing
protocols. Thus, in this paper, we take benefits from the control information
exchanged between different nodes of the network to dissimulate information
concerning the processing capacities of nodes, and to take advantage of physical
links characteristics between nodes. All these information is combined into a metric
allowing the selection of the most privileged neighbor node to perform various tasks
of a particular node. Intensive simulations are done by the NS-3 simulator showing
the effectiveness of this idea especially in runtime.

Keywords Mobile Cloud Computing (MEC) · Computation offloading · Routing

protocol · Mobility · Optimized Link State Routing Protocol (OLSR)

1 Introduction

Technological progress have led to the birth of smaller equipment that generates large
amount of data. This has created the need to work together and collaborate to achieve
a number of goals. Indeed, in an environment made up of several devices, some of
them have very limited computational performance. To be able to process a set of
tasks within acceptable deadlines, it must rely them on another powerful equipment.
It is the key concepts of Mobile Edge Computing (MEC) [1, 2]. In addition, in order

A. Ouacha (B)
Computer Science Department,
Faculty of Sciences, Mohammed V University, Rabat, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1121
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_103
1122 A. Ouacha

to communicate, all these devices use a routing protocol. One of these protocols and
most commonly used is the Optimized Link State Routing Proto-col (OLSR) [3, 4]
It is classified under the proactive protocols category whose principle is based on the
periodic exchange of information on the state of the links and on the neighborhood
through several types of control messages. Thus, so that each node can learn about
the performance of the other nodes of the network, several types of information are
added to the exchanged control messages. So, the main objective of this work is to use
these information that present the computation characteristics of each neighboring
node and combine them with other information extracted from the characteristics
of the physical link that connects the concerned node to other neighboring nodes
in order to be able to choose the best node that is capable of perforing a task list.
Moreover, since this is a mobile environment, the time required to process the tasks
by a neighbor node must be taken into consideration. This is what the predicted time
to quit the neighborhood is used for. It allows the node to avoid selecting a neighbor
node which risks leaving the neighborhood without even finishing execution of tasks
assigned to it.
The whole article has several sections in addition to the introduction and
the conclusion. The first section illustrates an overview of the OLSR protocol.
The second section provides a state of art that summarizes the various works on
the concept of treatment sharing. The third section presents the formulation of the
problem as well as the proposed solution whereas the fourth section presents the
various tests and simulations carried out as well as the obtained results.

2 OLSR Overview

OLSR [3] is a proactive routing protocol, designed to operate in a Mobile Adhoc

NETworks (MANETs) [5] without any central controlling entity. It represents an
adaptation and optimization of the principle of links state routing for ad hoc networks.
The optimization is that in this protocol, the nodes will only declare a sub-set of
their neighborhood by using MPR (Multipoint Relay). The OLSR protocol defines
an unique general format of the packet for all control messages circulating in the
network. It defines four types of control messages. In addition to HELLO and
Topology Control (TC) messages sent respectively every two and five seconds, the
OLSR protocol offers two other messages: Multiple Interface Declaration (MID) and
Host and Network Association (HNA).
The multipoint relays concept of OLSR aims to reduce the number of unnecessary
control messages during flooding in the network. The principle is based on the fact
that each node chooses a minimal sub-set of its symmetrical neighbors with one hop,
so as to be able to reach the whole neighborhood with two hops. As illustrated in
Fig. 1, to reach all two hop neighbor nodes (nodes with blue color), the node N just
needs to send its message to the subset of nodes selected as MPR (nodes with green
color) among nodes of the first neighborhood. The MPR allow an optimized broadcast
and minimize the used bandwidth by avoiding periodic sending of messages control.
An Adapted Routing Protocol for Mobile Edge Computing 1123

Fig. 1 Categories of nodes in OLSR protocol

3 Related Works

The authors of the research made in [6] present an algorithm based on a task
scheduling approach that is founded on dynamic voltage and frequency scaling
(DVFS). It aims to achieve energy reduction by allocating tasks to the appropriate
cores of mobile devices or offloaded it to the mobile cloud computing (MCC) envi-
ronment. The authors of the research made in [7] propose an heuristic based approach
named Edge Cloud Heuristic Offloading (ECHO) to find a tradeoff solution between
application runtime, battery lifetime, and user cost according to the user’s prefer-
ences. In the research made in [8], the authors propose a solution based on game
theory for offloading tasks among a swarm of Unmanned Aerial Vehicles (UAVs).
They opt to reduce their energy consumption and guarantee QoS for users by consid-
ering the UAVs as aerial capacitated cloudlets pooling their computational resources
to execute tasks from the ground internet of things (IoT) devices in their coverage.
The authors of [9] propose an online energy efficient task assignment and computing
offloading strategy in the aim to minimize the time average energy consumption of
the multimedia services. The proposed strategy can adaptively decide the task assign-
ment, coordinate and optimize the wireless and computation resource allocation by
taking the dynamic wireless condition and service delay constraints into considera-
tion. In [10] authors propose a framework to provide guaranteed services at minimum
system cost. To address the challenges of configuring the edge optimization capa-
bility and determining the type and amount of clouds in which resource provision
within the Cloud Assisted Mobile Edge (CAME) framework, the authors provide
an Optimal Resource Provisioning (ORP) algorithm that is used for optimization of
the capacity of adjacent hosts and provides dynamic cloud renting strategies. The
authors of the research made in [11] propose a model that uses the Markov approx-
imation method for full implementation of tasks and computations in MEC. This
model provides a multi-server system in which only one mobile is brought close to
1124 A. Ouacha

the neighboring servers for complete computing. The use of the Markov approxi-
mation technique provides a common optimization of the allocation of computing
tasks and the frequency of the processor’s scalability to minimize the runtime and
mobile power consumption. The research made in [12] proposes a new MEC archi-
tecture based on intelligent computational offloading. Depending on the size of the
data from the mobile user computing task and the performance characteristics of
the periphery computing nodes, a task-predicting and task migration algorithm are
proposed. In the work done in [13], the authors propose an Energy-aware Collabora-
tive Routing (ECoR) scheme for optimally handling task offloading between source
and destination UAVs in a grid-locked UAV swarm. The proposed ECoR acts on two
objectives: routing path discovery and routing path selection. In the first one (routing
path discovery), the scheme selects the most optimal path based on the maximization
of residual energy. In the second one (routing path selection), the scheme ensures
balanced energy utilization between members of the UAV swarm and enhances the
overall path lifetime without incurring additional delays in doing so.
Like these researches, our goal is to find for a mobile node (MN), whose compu-
tational performance is limited, the best neighbor node (NN) that presents the ideal
environments to perform its tasks. But unlike all these works, we seek taking advan-
tage of the normal operating of the OLSR protocol and insert in the control messages
exchanged between nodes the characteristics relating to the processing capacity.

4 Problem and Resolution

4.1 System Model

As shown in Fig. 2, the Mobile Node (MN) contains a list of N independent tasks,
where each one is denoted τi with i = {1, 2, . . . , N }. Those tasks that are supposed
to be computationally greedy and sensitive to delays should all be achieved. The
execution time of the task list cannot exceed a maximum required time t max and each
task τi can be processed remotely by an elected Neighbor Node (NN).
Each data task is atomic (i.e. it cannot be divided into subtasks) and it is charac-
terized by the following two parameters τi  di , λi . The first one denoted di [bits]
represents the size of the task (quantity of input parameters and program codes) to
be transferred from the MN in question to the NN. The second parameter, denoted
λi [cycles], specifies the workload referring to the amount of computation necessary
to accomplish the processing of this task. In addition, the execution time of a task τi
also depends on the frequency f j [H er t z] of the N N j . Considering the constraint of
a mobile environment, the NN should be chosen from nodes in the first neighboring
so that it stays in the neighboring of the MN as long as possible.
and for all the tasks to be executed on the N N j , we have:
An Adapted Routing Protocol for Mobile Edge Computing 1125

Fig. 2 Node with list of tasks offloaded to a Neighbor Node

N λi
t exec
j = [Second] (1)
i=1 fj

Since the task τi is offloaded to the N N j , the time needed to transmit it is given
by:

di
ti,comm
j = [second] (2)
rj

In this equation r j corresponds to the uplink data rate [14]. It is calculated by:

PT × gj
r j = w × log(1 + ) (3)
σj

With: w is the bandwidth of the link between the MN and N N j , note that the
MN and all nodes of the neighborhood use the same channel. P T represents the
transmission power of the MN, σ j the noise of the channel to which the MN and the
N N j are connected and g j the gain between the MN and the N N j .
and for all the tasks to be offloaded on the neighbor node N N j , we have:

 N di
t comm
j = [Second] (4)
i=1 r j
1126 A. Ouacha

Thus, the completion time of the process of offloading and executing a single task
τi , which will be called Time Metric (TimeMetric) is given by:

ti, j = ticomm + tiexec (5)

Consequently, for a single task τi we have:

di λi
ti, j = + [Second] (6)
rj fj

and for all offloaded tasks, we have:

 N  di λi

tj = + [Second] (7)
i=1 r j fj

4.2 Problem Formulation

In this section, we present our formulation of the optimization problem that aims at
selecting the convenient neighbor node N N j whitch minimizes the time t j it takes
for an offloaded task set to be executed. In other words, this is to select the node
 corresponds to the lowest value t j from the possible set of available values
that
T t1 , t2 , . . . , tj , . . . , tM and that minimizes our TimeMetric.
The obtained problem is formulated as follows:
⎧
⎪
⎪ ArgMin {T }
⎪
⎪ { f j ,r j ,λi ,di }
⎨
P1 (C1 ) f j > F min (8)
⎪
⎪
⎪
⎪ (C2 ) t j > T min
⎩
(C3 ) e M N > E min

In this work, each of the tasks is offloaded to a N N j so that it can be executed.

Thus, each feasible offloading must meet the below constraints:
• The constraint (C1) indicates that the frequency of the N N j f j must be greater
than a minimum authorized value F min in order to avoid the Neighbor Node whose
execution capacity is very limited.
• The constraint (C2) shows that the execution time of all the offloaded tasks must
be less than the given latency requirement T min . This constraint is necessary and
especially when it comes to a mobile environment where the probability of the
N N j to leave the neighbourhood of the MN is very greater.
• The last constraint (C3) is important especially if the N N j battery is critical. It
imposes that the total remaining energy e M N of the MN must be greater than a
An Adapted Routing Protocol for Mobile Edge Computing 1127

threshold E min in order to avoid the nodes that risk shutting down before achieving
the execution of the tasks. The threshold E min for a single task is calculated by
the Eq. 10 below.

P T λi
E imin = [J ] (9)
fj

and for all offloaded tasks, we have:

N P T λi
E min = [J ] (10)
i=1 fj

4.3 Problem Resolution

In the formulation part of the problem, Constraint (C2) specifies that the execution
time of all the offloaded tasks must be less than threshold T min in order to avoid
choosing the NN which risks leaving the neighbourhood of the MN because of
mobility constraint. The time T min in question will be represented by the predicted
Rest Time to Quit (RTTQ) calculated as illustrated in Fig. 3 [14]. It is estimated based
on the cartesian coordinates exchanged through the HELLO messages received from
NN.
The TimeMetric is calculated when the MN has a list of tasks to transmit to be
executed on the corresponding neighbor node N N j using algorithm 1. Note that the
frequency f j , λi as well as the node coordinates x j and y j of each neighbor node
are retrieved from the HELLO message sent and they are stored by the MN.

Fig. 3 Predict remaining

time estimation
1128 A. Ouacha

5 Simulations and Results

Wherever Times is specified, Times Roman of Times New Roman may be used. If
neither is available on your word processor, please use the font closest in appearance
to Times. Avoid using bit-mapped fonts if possible. True-Type 1 fonts are preferred.

5.1 Simulations

To test the solution presented in this work, we used as a simulator NS-3. It has the
advantage of being more suitable for network environments. It contains an imple-
mentation of a very large number of protocols, most often used in various types of
networks. In addition, it is a free and open source. The network architecture used in
NS-3 is based on existing software and hardware in the real world. This facilitates
the use of the simulator and allows easy integration of new models.
The simulation environment as well as the simulation parameters, as shown in
Table 1, consists of a network of 50 mobile nodes according to the Random Waypoint
Mobility model [15]. All nodes have a single interface that uses OLSR as the routing
protocol. A variable number of nodes between 5 and 25 with a step of 5 are chosen
to send a set of tasks represented by a data packet of size 1000 Bytes. Each node
sends 10 packets with a frequency of 1 packet for one (1) second.
An Adapted Routing Protocol for Mobile Edge Computing 1129

Table 1 Simulation
Parameter Value
parameters
Network simulator NS −3.29
Simulation time 100 s
Simulation area 1500 × 1500 m × m
Number of nodes 50
Transmission range 500 m
Mobility model Random WayPoint (RWP)
Max speed 20 m/s
Max pause time 0s
Packet size 1000 bytes
WiFi 802.11b

To simulate the calculations performance, two numbers are assigned to each node.
The first one represents the frequency f j of the N N j processor. It is initialized by
a randomly generated value between 0.5 and 3.0 GHz. The second represents the
number of cycles λi necessary to process a single task. It is initialized by an integer
value generated randomly between 1 and 8.
The simulations consist in choosing a number of source nodes; each one generates
10 data packets constituting the tasks to be executed. Before sending each set of
tasks, each node researches through the nodes of its first neighborhood for the one
that has the most minimal possible TimeMetric. Given the mobile environment and
to avoid the choice of a neighboring node which risks leaving the neighborhood even
before giving the result of the processing, only the one whose RTTQ (algorithm 1)
is twice higher than the TimeMetric is chosen. In order to be able to calculate the
communication time which is part of the TimeMetric, the information that will be
needed is extracted from the physical layer of MN as well as from the frame received
from the NN containing the hello message.

5.2 Results

At the end of the simulations under NS-3, a set of results are obtained. Thus, the
graph of Fig. 4 clearly shows that the average time necessary for the execution of the
tasks by the selected NN always remains lower than the average time calculated at
the level of all the other neighboring nodes necessary for the execution of the same
set of tasks regardless of the number of source nodes.
In addition, Fig. 5a shows that the TimeMetric has experienced a slight increase
when the number of source nodes increases. Indeed, once the number of source nodes
increases, the number of generated packets also increases which causes pressure on
the elected NN. Thus several packets will be directed to this node which generates
waiting times whether it is at the level of the reception of the packets by the lower
1130 A. Ouacha

Average Time Metric Neighbor Node

All Neighbor Nodes
Average Time 3

0
5 10 15 20 25
Number of Source Nodes

Fig. 4 Average time required for the execution of tasks by a neighbor node compared to the average
time calculated at the level of other neighboring nodes

0.66 2.18
0.64 2.16
0.62 2.14
0.6 2.12
0.58 2.1
5 10 15 20 25 5 10 15 20 25
Number of Source Nodes Number of Source Nodes
a. Selected Sink node b. All Neighbor Nodes

Fig. 5 Average of the time required for the execution of tasks

layer or at the level of the order of execution of the tasks on the microprocessor.
Figure 5b also shows a slight decrease in the average of the calculated time metric
at all neighboring nodes as the number of source nodes increases.
Despite the gain recorded in terms of the TimeMetric, this system has the disad-
vantage of an additional load that will be added to the control traffic generated by the
OLSR protocol itself. Indeed, Fig. 6 shows a slight increase in the average throughput
exhibited by the Modified OLSR version compared to the standard OLSR version.
This is due to the fact that the HELLO control messages sent every two seconds are
enriched by other information used for the calculation of the time metric and the
RTTQ (current position of the sending node of the HELLO message, its calculation
frequency, the number of cycles of its processor, …).
An Adapted Routing Protocol for Mobile Edge Computing 1131

Average Throughput OLSR Standart Vertion

OSLR Modified Version
Average Throughput
1

0.5

0
5 10 15 20 25
Number of Source Nodes

Fig. 6 Average flow rate comparison between the standard version of OLSR Protocol and the
modified version

6 Conclusion

This article illustrates the case of a node whose computational performance is very
limited (the case of a smart watch or a smart garment …) and that has a set of tasks to
perform. To do this, it seeks among the nodes from the first neighbourhood the one
which presents the smallest metric of time. Thus, a first solution consists in taking
advantage of the control messages of the OLSR protocol to exchange the calculation
performances of each node of the network. So each node has everything it needs to
decide which node will run this task list.
The results of the simulations carried out by NS3 show a gain in terms of the time
metric against a price paid by a slight increase in the average flow. These results also
show that (Disadvantages) in a cell, the neighbour node which has the smallest time
metric will be elected by all the other neighbouring nodes. This presents a resulting
pressure by sending multiple sets of spots to that same node.
As Perspectives, a forthcoming study aims to integrate a mechanism to balance
the loads so as to share the processing of tasks between several neighbouring nodes.
A first proposition consists in finding the ideal node or nodes for the execution of the
tasks in the set of nodes of the first and second neighbourhood. Another perspective
this time is to also take advantage of other metrics such as energy performance and
those of the main memory in the search for the execution node.

References

1. Cao J, Zhang Q, Shi W (2018) Edge Computing: A Primer, 1st edn. Springer, Cham
2. Nunna S, Ganesan K (2017) Mobile edge computing. In: Thuemmler C, Bai C (eds) Health 4.0:
how virtualization and big data are revolutionizing healthcare. Springer, Cham, pp 187–203
3. Clausen T, Jacquet P (2003) RFC3626: Optimized Link State Routing Protocol (OLSR), vol
RFC3626: RFC Editor
1132 A. Ouacha

4. Clausen T, Dearlove C, Jacquet P, Herberg U (2014) The Optimized Link State Routing Protocol
Version 2. Internet Engineering Task Force (IETF), April 2014
5. Corson S, Macker J (1999) RFC2501: Mobile Ad hoc Networking (MANET): Routing Protocol
Performance Issues and Evaluation Considerations: RFC Editor
6. Lin X, Wang Y, Xie Q, Pedram M (2015) Task scheduling with dynamic voltage and frequency
scaling for energy minimization in the mobile cloud computing environment. IEEE Trans Serv
Comput 8:175–186
7. De Maio V, Brandic I (2018) First hop mobile offloading of DAG computations. Presented
at the 2018 18th IEEE/ACM international symposium on cluster, cloud and grid computing
(CCGRID), Los Alamitos, CA, USA
8. Ma W, Liu X, Mashayekhy L (2019) A strategic game for task offloading among capacitated
UAV-mounted cloudlets. In: 2019 IEEE international congress on internet of things (ICIOT),
pp 61–68
9. Sun Y, Wei T, Li H, Zhang Y, Wu W (2020) Energy-efficient multimedia task assignment and
computing offloading for mobile edge computing networks. IEEE Access 8:36702–36713
10. Ma X, Wang S, Zhang S, Yang P, Lin C, Shen XS (2019) Cost-efficient resource provisioning
for dynamic requests in cloud assisted mobile edge computing. IEEE Trans Cloud Comput.
https://doi.org/10.1109/TCC.2019.2903240
11. Zhou W, Fang W, Li Y, Yuan B, Li Y, Wang T (2019) Markov approximation for task offloading
and computation scaling in mobile edge computing. Mob Inf Syst 2019:8172698
12. Miao Y, Wu G, Li M, Ghoneim A, Al-Rakhami M, Hossain MS (2020) Intelligent task predic-
tion and computation offloading based on mobile-edge cloud computing. Future Gener Comput
Syst 102:925–931
13. Mukherjee A, Misra S, Chandra VSP, Raghuwanshi NS (2020) ECoR: energy-aware collab-
orative routing for task offload in sustainable UAV swarms. IEEE Trans Sustainable Comput
5(4):514–525
14. Ouacha A, Lakki N, Abbadi JE, Habbani A, Bouamoud B, Elkoutbi M (2013) Reliable MPR
selection based on link lifetime-prediction method. In: 2013 10th IEEE international conference
on networking, sensing and control (ICNSC), pp 11–16
15. Roy RR (2011) Handbook of mobile ad hoc networks for mobility models, 1st edn. Springer,
Heidelberg
Advanced Technologies in Energy
and Electrical Engineering
Improved Geographic Routing Protocol
for Wireless Sensor Networks

Kenza Redjimi, Mohammed Redjimi , and Mehdi Boulaiche

Abstract Nowadays, several researches focus on Wireless Sensor Networks

(WSNs) protocols. These small sensors are disseminated in a large geographical
area, sense local data, and send them over multi-hop routes to the base station (BS).
Furthermore, the BS processes data and sends results to remote centers. The most
problem in the WSNs is the energy optimization of the sensor nodes for collect,
process, and information broadcast. Each node has limited energy autonomy and
most of the existing protocols focus on the optimal management of energy consump-
tion aiming to maximize the network lifetime. We present, here, a geographic routing
protocol dedicated to the WSNs. This approach exploits the neighboring node’s posi-
tions, which are in the direction of the BS to send the data packets with minimum
energy dissipation. In order to forward packets efficiently around voids, new mech-
anisms are introduced. The shortest path between the sender node and the desti-
nation node is calculated by taking in account only the local knowledge of sensor
nodes (neighborhood). Simulation results show that this protocol is very efficient
and outperforms other geographical protocols in terms of energy management.

Keywords Wireless sensor networks · Geographic routing protocols · GSM

Localization · Energy consumption · Energy efficiency

1 Introduction

Wireless Sensor Networks (WSNs) [1] are systems made up of several indepen-
dently functioning elements, which are disseminated over large areas. These elements
include limited computing and storage resources (processing units, memories, etc.)
and are fitted with sensors specific to the environments in which they operate and

K. Redjimi · M. Redjimi · M. Boulaiche (B)

Department of Computer Science, LICUS Laboratory, University 20 August 1955,
21000 Skikda, Algeria
e-mail: [email protected]
K. Redjimi
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1135
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_104
1136 K. Redjimi et al.

to data, which they must measure. They are also equipped with wireless means for
sending and receiving data. The information collected by the various nodes of the
network undergoes a first elementary processing and then is transmitted to the base
station. These sensors can measure various ranges of quantities such as humidity,
temperature, level, pressure, wind speed, objects detection, etc. Consequently, these
sensor networks are useful in several fields and their applications are multiple, for
example in monitoring, control and decision-making. Thus, they are used in medicine
[2], military field, forest or flood fire monitoring, smart homes and cities [3, 4] and
industrial automation among other fields [5].
Batteries supply these elements with energy and thus allow them certain autonomy
of operation. Optimal management of the energetic charge of these batteries remains
one of the most important problems to be resolved with regard to WSNs [6].
A sensor node expends a maximum of energy when it is in the phase of sending
data. The energy expended during the reception is less. However, to optimize the
energy consumption in a sensor node and, consequently, to increase the network
lifetime, it is essential to consider all the energy expenses at all levels of the architec-
ture of the WSN. Indeed, these sensors are sometimes inaccessible and recharging
their batteries can be very restrictive especially when they are located in hostile areas.
As a result, many researchers have carried out the implementation of proto-
cols making it possible to better managing energy consumption in wireless sensor
networks. Note, however, that several solutions provided are in fact improvements
of other pre-existing solutions.
Most of the developed routing protocols for WSNs [7] require node localization
coordinates since events are directly related to the node location. The positions of
two nodes make it possible to calculate the distance between them, which makes it
possible to estimate the energy expended during the communication between these
two nodes. This work exploits this idea to improve the energy consumption in WSN’s.
Our protocol increases packets delivery ratio to the base station.
The following sections are organized as follows: Sect. 2 briefly presents general
notions concerning routing protocols in WSNs. Section 3 is devoted to present the
proposed approach. In Sect. 4; the results and their discussion are presented and
Sect. 5 concludes this article.

2 Backgrounds

Generally, three classes of routing protocols can be distinguished: flat, hierarchical,

and geographic routing protocols (also called location-based routing protocols).
There are two categories in flat routing protocols; reactive (On-Demand) protocols
such as Dynamic Source Routing (DSR) [8] and Ad hoc On-demand Distance Vector
(AODV) [9] or proactive (Table-Driven) protocols such as Destination Sequenced
Distance Vector (DSDV) [10]. In the hierarchical routing protocols, the nodes are
regrouped into clusters and in each cluster, a cluster head (CH) is periodically elected.
The cluster head plays specific roles such as the aggregation of data transmitted by
Improved Geographic Routing Protocol for Wireless Sensor Networks 1137

the cluster nodes and their routing to the base station. Low-energy adaptive clustering
hierarchy (LEACH) [11] and its variants are typical examples of this class. In the
case of location-based routing protocols, the sensor nodes are accessible thanks to
their positions via the Global Positioning System (GPS), thanks to tables containing
their coordinates or other specific tools. These positions allow the implementation of
information routing. For example, Greedy Perimeter Stateless Routing (GPSR) [12]
is a geographic routing protocol. In Geographic Routing, each node is concerned
only with its one-hop neighbors. Therefore, in order to make forwarding decisions,
geographic routing protocols, only require the position of the packet’s destination and
the position of its one-hop neighbors. This property has made Geographic Routing
very assuring. Geographic Routing protocols operate according to two functioning
modes: i) the packets are routed from one node to its neighbor, which can be the
closest node to destination or the node that has the highest energy, etc. ii) if there are
no nodes with these prescriptions, the algorithm goes into recovery mode.
Geographic Routing protocols give interesting and attractive solutions [13–18]
for the WSNs because to forward information, a node needs only to know its own
position, and its neighbors and the BS locations. The localization information can be
obtained thanks to the Global Positioning System (GPS). This reduces the complexity
of control management at the level of each node very significantly.
The greedy protocols [18] use the shortest paths to distribute the packets between a
source and a destination node. In the Nearest with Forward Progress (NFP) protocol
[18], the algorithm chooses the nearest neighbor among the nodes that are situ-
ated in the forward direction of the destination to send the packets. In the Most
Forward within the transmission Range (MFR) routing protocol [1], a node selects
the neighbor that is nearest from the destination as the following routing node. In the
case of the normalized advance (NADV) for geographic routing protocol, the neigh-
boring nodes of the source node are selected taking into account their proximity as
well as the cost of the link.
The protocol proposed here is based on EEGR [19] and brings improvements to
it:
– The next neighbor is calculated each round rather than calculating it for each
forwarded packet,
– In order to forward packets around a void through the optimal path, a new
mechanism is used,
– In the proposed approach, the algorithm tries to find a global optimum path without
any assumptions.
1138 K. Redjimi et al.

2.1 The Network Model

A WSN is represented as a graph G = (E, V), where E is the set of edges that
corresponds to the links between nodes and V the nodes. The proposed protocol
adopts the model of Ivan Stojmenovic et al. [20, 21] that generalized the energy
model of Rodoplu and Meng’s [22] by adjusting a linear factor for the transmission
and the reception of a radio signal between two nodes separated by a distance d:

u(d) = ad a + c (1)

α is an attenuation factor (2 ≤ α ≤ 6).

c is a constant (also related with other energies consumed for data processing at
each node), and a is a linear factor.

3 The Proposed Approach

In the proposed approach, the WSN is considered as a connected-graph. Therefore,

to transmit a packet from a source node to the BS, the algorithm tries to find a global
optimum path among the entire locally intermediate nodes. The aim of this operation
is to minimize energy consumption. In a greedy mode, each concerned node uses its
neighbor localization that is closest to the BS and builds iteratively the optimal path.

3.1 Approach Overview

As explained above, in geographic protocols, a node considers a closest neighbor to

the BS for a next hop. Although, this technique does not make it possible to guarantee
energy optimization on the one hand and comes up against the so-called void problem
[23], when there is no neighbor closer than the current node. The proposed approach
tries to avoid these problems.
In the proposed approach, at first a Minimal Spanning Tree (MST) calculation
process is initiated. Each node builds a local sub-graph G = (E , V ), including the
neighbors, which are closest to the BS than the source node and that are situated
in the BS direction. (See Subsect. 3.2). This area is referred to as a routing area.
Therefore, by using Prim’s algorithm [24], a minimal spanning tree is calculated
using the local sub-graph G’ (Fig. 1). As a result, we obtain the paths that start from
the current node and which traverse a set of nodes whose sum of weights is minimal
(see Subsect. 3.2) is the hop neighbor.
Improved Geographic Routing Protocol for Wireless Sensor Networks 1139

Fig. 1 The routing and recovery areas

A node that encounters a void must expand its subgraph by δ degrees to include
other neighbors; this zone is called the recovery area. The same algorithm is applied
by using the source routing to route packets. According to the shortest path that is
included in the packet’s header, the next hop transmits the packet.

3.2 Sub-graph Formation Algorithm

In greedy routing techniques, each node participates for packet forwarding towards
the BS. At the network setup time, the nodes select their next forwarding neighbor.
This neighbor forwards all received packets. In order to balance the load of neigh-
boring nodes, the selection is initialized at each round. The duration of a round
depends on the intensity of the network traffic. Note that a node plays the rotation role
if its next forwarding neighbor energy decreases. In each node, a table that contains
the neighbor’s identifier, position and remaining energy is periodically updated. In
addition, each node periodically broadcasts a message displaying its position and
remaining energy.
In the local sub-graph G = (E ,V ), V corresponds to the set of neighbors in
ARouting_Area/ARecovery_Area, E to the set of edges between these neighbors.
The distance between two nodes u and v is identified as dist(u, v), and r is the sensor
node’s communication range.
1140 K. Redjimi et al.

3.3 Edge Weights

To balance and to regularly distribute the consumption energy on all the neighbors,
we add a new metric. The energy required for the transmission and the reception of
data between two nodes is represented in Eq. (2):

(ad a + c)
Edgeweight = (2)
R Eu

c represents the cost of transmission/reception between two nodes u, v?V  ,

REu is the remaining energy for the node u.
Note that G is constructed only at the network setup since sensor nodes are static
but the edge weights are recalculated at each round in order to balance the role of
next forwarding neighbor among multiple neighbors.

3.4 The Routing Strategy

In the WSN, data packets are forwarded from source nodes to the BS via multi-hop
ways. In the proposed approach, we distinguish two modes for packets forwarding:
In the normal mode, the packets are forwarded through the Routing Area whereas
in the recovery mode, the packets are routed through the Recovery Area. A packet
is forwarded in normal mode but when it encounters a void at a node p, it will be
balanced to recovery mode to forward it around the void. When the packet gets to a
node q having at least a neighbor closer to the BS than node p, the packet is returned
back to greedy mode forwarding again. In that case, the packet is forwarded to that
neighbor. To do that, a node p that balances a packet into recovery mode includes in
the packet’s header its position, i.e. the position where the packet was balanced to
recovery mode. The node p includes also in the header the minimum path through
which the packet will be forwarded in recovery mode. Afterwards, each node q that
receives the packet checks first whether it has at least a neighbor closer to the BS than
node p using the position included in the packet’s header. If so, it balances the packet
to normal mode and forwards the packet to that neighbor. Otherwise, it forwards
the packet according to the minimum path included in the header. The algorithm 1
presented bellow summarizes the routing strategy of this approach.
Improved Geographic Routing Protocol for Wireless Sensor Networks 1141

Algorithm 1 : Route(Packet “P”);

Si : the node that is handling packet P;

N(Si ): neighbors that are closest to the BS;
R(Si ): neighbors that are in the recovery area;
H(Si ): the one hop neighbors;

If N(Si Then // No neighbor is closer to the BS;

If Si is the source of the packet Then Next_Neighbor := Nearest Neighbor in H(Si );
Else If (P.mode is in normal mode) Then
Apply sub-graph G’ Formation Algorithm from R(Si );
Apply Prim’s Algorithm;
Find the minimum path in theMST found;
Next_Neighbor := First nod in the minimum path in the MST found;
Include the minimum path inthe packet’s header;
Else // P.mode is in recovery mode
Check the address contained in the packet’s header;
If there is at least a neighbor closer to BS than P.Rec_Addrr Then
Balance the packet to normal mode;
Next_Neighbor := That_Neighbor;
Else
Next_Neighbor := first node in the path included in packet’s header;
EndIf
EndIf
EndIf
Else // there is neighbors closer to the BS;
Apply sub-graph G’ Formation Algorithm from N(Si ) ;
Apply Prim’s Algorithm;
Find the minimum path in the MST found;
Next_Neighbor := First node in minimum path in the MST found;
EndIf
Transmit P to Next_Neighbor;

Table 1 Simulation
Parameters Values
parameters
Network nodes number 20 to 160
Generation mode Randomized
Trigger nodes 05 trigger nodes
Size of the network 100 m × 100 m
Transmission range of nodes 30 m
Radio Transmission power 0.028 w
Radio Reception power 0.036 w
MAC sub-layer MAC IEEE 802.11
Angle δ 2π/3
1142 K. Redjimi et al.

4 Results and Discussion

Aiming to evaluate the proposed protocol performance, a set of network’s simula-

tions was conducted by using the J-Sim simulation software [25]. We considered
homogenous sensor nodes as, which have the same initial amount of energy, the
same hardware resources, the same transmission range and the energy of the base
station is considered as unlimited. These networks are considered deployed in an area
of 100 m × 100 m creating topologies containing 20 to 160 nodes. The generation
of the positions of nodes is generated randomly. The other factors used to calculate
edge weights ‘α’, linear factor ‘a’, and the constant ‘c’ are set to 2, 1 and 1000
respectively. The Table 1 presents these parameters.
The simulation results of our proposed approach called Improved Greedy Routing
protocol (IGR) are compared with four geographic protocols: EGGR, NFP, GPSR
and MFR as reported in [15]. The Figs. 2 and 3 represent the energy efficiency
allowing evaluating these protocols.
The Fig. 2 shows the average of energy consumption when we change the number
of nodes in the network and the Fig. 3 depicts this average per time. From these
figures, we can note that the proposed protocol (so-called IGR) gives best results
when compared with the other protocols (EGGR, GPSR, NFP, and MFR). It improves
average energy consumption in the network. The justification of this improvement
lies at the fact that in IGR protocol, the packets are forwarded through the optimal
paths, whereas in EGGR protocol, the calculation of the optimal path takes into
account the assumption that the cost of the retransmission between each neighbor

Fig. 2 Average of the energy consumed compared to the number of nodes in the network
Improved Geographic Routing Protocol for Wireless Sensor Networks 1143

Fig. 3 Average energy consumed with time

and the BS is optimal, which is not the fact in the reality. In the other protocols
(GPSR, NFP, and MFR), a neighbor closest to the BS is selected without regarding
the energy consumption criteria. Therefore, the forwarded packets can be expensive
in terms of energy consumption, especially in the case of voids (GPSR protocol
applies recovery mechanism to get out of voids).

5 Conclusion

The proposed approach (called IGR: Improved Greedy Routing geographic protocol)
allows an efficient energy management for the geographic routing protocol. As a solu-
tion to resolve the local minima problem by creating an additive area called recovery
area to forward packets. Mechanisms are introduced to elaborate the optimal way
during each round. After realizing several simulations and analyzing the simula-
tion results, we observe that the proposed protocol contributes to the reduction of
the energy consumption at the sensor nodes and consequently, increase the network
lifetime.
In future works, we project to realize a series of simulations aiming to demonstrate
the effectiveness and the efficiency of this protocol in increasing the packet’s delivery
rate especially in the case of complex networks.
1144 K. Redjimi et al.

References

1. Obaidat MS, Misra S (2014) Principles of wireless sensor networks. Cambridge University
Press, Cambridge
2. Dhobley A, Ghodichor NA, Golait SS (2015) An overview of wireless sensor networks for
health monitoring in hospitals via mobile. Int J Adv Res Comput Commun Eng 4(1):169–171
3. Belghith A, Obaidat MS (2016) Wireless sensor networks applications to smart homes and
cities. In: Smart cities and homes, pp 17–40. https://doi.org/10.1016/b978-0-12-803454-5.000
02-x
4. Lasek A, Cercone N, Saunders J (2016) Smart restaurants: survey on customer demand and
sales forecasting. In: Smart cities and homes, pp 361–386. https://doi.org/10.1016/b978-0-12-
803454-5.00017-1
5. Kandris D, Nakas C, Vomvas D, Koulouras G (2020) Applications of wireless sensor networks:
an up-to-date survey. Appl Syst Innov 3(1):14
6. Khan JA, Qureshi HK, Iqbal A (2015) Energy management in wireless sensor networks: a
survey. Comput Electr Eng 41:159–176. https://doi.org/10.1016/j.compeleceng.2014.06.009
7. Bhushan B, Sahoo G (2018) Routing protocols in wireless sensor networks. Studies Compu-
tational Intelligence, pp 215–248. https://doi.org/10.1007/978-3-662-57277-1_10
8. Boukerche A, Turgut B, Aydin N, Ahmad MZ, Boloni L, Turgut D (2011) Routing protocols
in ad hoc networks: a survey. Comput Netw 55:3032–3080. https://doi.org/10.1016/j.comnet.
2011.05.010
9. Mulert J, Welch I, Seah WKG (2012) Security threats and solutions in manets: a case study
using aodv and saodv. J Netw Comput Appl 35(4):1249–1259. https://doi.org/10.1016/j.jnca.
2012.01.019
10. Ade SA, Tijare PA (2010) Performance comparison of AODV, DSDV, OLSR and DSR routing
protocols in mobile ad hoc networks. Int J Inf Technol Knowl Manag 2(2):545–548
11. Tyagi S, Kumar N (2013) A systematic review on clustering and routing techniques based upon
LEACH protocol for wireless sensor networks. J Netw Comput Appl 36(2):623–645. https://
doi.org/10.1016/j.jnca.2012.12.001
12. Karp B, Kung H (2000) GPSR: greedy perimeter stateless routing for wireless networks. In:
Proceedings of the sixth annual international conference on mobile computing and networking.
ACM Press, pp 243–254. https://doi.org/10.1145/345910.345953
13. Milocco RH, Costantini H, Boumerdassi S (2014) Improved geographic routing in sensor
networks subjected to localization errors. Ad Hoc Netw 13:476–486. https://doi.org/10.1016/
j.adhoc.2013.10.001
14. Lee S, Bhattacharjee B, Banerjee S, Han B (2010) A general framework for efficient geographic
routing in wireless networks. Comput Netw 54(5):844–861. https://doi.org/10.1016/j.comnet.
2009.09.013
15. Boulaiche M, Bouallouche-Medjkoune L (2015) EGGR: energy-aware and delivery guarantee
geographic routing protocol. Wireless Netw 21(6):1765–1774. https://doi.org/10.1007/s11276-
014-0880-1
16. Tao S, Ananda AL, Chan MC (2010) Greedy face routing with face identification support
wireless networks. Comput Netw 54(18):3431–3448. https://doi.org/10.1016/j.comnet.2010.
07.004
17. Kleerekoper A, Filer NP (2015) Perfect link routing for energy efficient forwarding in
geographic routing. Ad Hoc Netw 30:46–62. https://doi.org/10.1016/j.adhoc.2015.03.001
18. Al-Shugran M, Ghazali O, Hassan S, Nisar K, Arif ASM (2013) A qualitative comparison
evaluation of the greedy forwarding strategies in mobile ad hoc network. J Netw Comput Appl
36(2):887–897. https://doi.org/10.1016/j.jnca.2012.10.008
19. Zhang H, Shen H (2007) Eegr: energy-efficient geographic routing in wireless sensor networks.
In: 2007 international conference on parallel processing (ICPP 2007). IEEE, p 67. https://doi.
org/10.1109/ICPP.2007.37
20. Bose P, Morin P, Stojmenović I, Urrutia J (2001) Routing with guaranteed delivery in ad hoc
wireless networks. Wireless Netw 7(6):609–616. https://doi.org/10.1023/A:1012319418150
Improved Geographic Routing Protocol for Wireless Sensor Networks 1145

21. Stojmenovic I, Lin X (2001) Power-aware localized routing in wireless networks. IEEE Trans
Parallel Distrib Syst 12(11):1122–1133. https://doi.org/10.1109/71.969123
22. Rodoplu V, Meng TH (1999) Minimum energy mobile wireless networks. IEEE J Sel Areas
Commun 17(8):1333–1344. https://doi.org/10.1109/49.779917
23. Chen D, Varshney PK (2007) A survey of void handling techniques for geographic routing
in wireless networks. IEEE Commun Surv Tutor 9(1):50–67. https://doi.org/10.1109/COMST.
2007.358971
24. Prim RC (1957) Shortest connection networks and some generalizations. Bell Syst Tech J
36(6):1389–1401
25. https://www.physiome.org/jsim/. Accessed 8 Dec 2020
Supervision and Monitoring
of Photovoltaic Systems Using Siemens
PLC and HMI

Ahmed Bouraiou, Ammar Neçaibia, Saad Motahhir,

Mohammed Salah Bouakkaz, Issam Attoui, Nadir Boutasseta,
Rachid Dabou, and Abderrezzaq Ziane

Abstract Automation is the top priority in modern industries, the debate on this
phenomenon is always controversial as the machine tends to replace man in several
tasks, and he unfortunately cannot perform the tasks that the machine does in several
fields. The supervision and monitoring of the industrial systems are two very impor-
tant fields in industrial automation. For ensuring the reliability and stability of PV
system operation, it is indispensable to install supervision and monitoring systems.
The work in this paper presents a simple platform for supervision and control in real-
time of pumping system parameters (current, voltage, irradiance, temperature and
flow) based on programmable logic controller (PLC) and human-machine interface
(HMI) of Siemens tools.

Keywords PV system · Supervision · Monitoring · PLC · HMI

1 Introduction

The production technology and the complexity of the operations to be performed, lead
to the implementation of devices and systems for the automation of manufacturing
or production workshops. Automation is the scientific and technological field that

A. Bouraiou (B) · A. Neçaibia · R. Dabou · A. Ziane

Unité de Recherche en Energie Renouvelables en Milieu Saharien, URERMS, Centre de
Développement Des Energies Renouvelables, CDER, 01000 Adrar, Algeria
S. Motahhir
Engineering, Systems and Applications Laboratory, ENSA, SMBA University, Fez, Morocco
e-mail: [email protected]
M. S. Bouakkaz
Department of Electrical Engineering, Laboratoire D’électrotechnique de Skikda «LES»,
Université du 20 Août 1955, 21000 Skikda, Algeria
I. Attoui · N. Boutasseta
Research Center in Industrial Technologies CRTI, P.O. Box 64, Cheraga, Algeria

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1147
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_105
1148 A. Bouraiou et al.

performs and controls technical tasks, by machines operating without human inter-
vention, or with the aid of reduced intervention. The supervision and monitoring of
the industrial systems are two very important fields in industrial automation. They
are based on the application of different analysis of many signals and parameters
like temperature, current, vibrations, voltage, etc. Many research works have been
done on this area like [1–10]. On the other hand, the transition to renewable energy
resources utilization has become an indispensable task due to the negative effect of
conventional sources (fossil) in terms of environmental pollution and global warming
[11, 12]. Many scientific papers have been published on this topic like [13–25]. Solar
energy is one of the most popular renewable energies especially in countries that
characterized by huge solar energy potential [26–28]. The supervision and control
of the photovoltaic system is essential key for ensuring the reliability and stability of
PV system operation [5]. The paper presents a simple platform for supervision and
control of pumping system parameters (Current, voltage, irradiance, temperature and
flow) based on a programmable logic controller (PLC) and human-machine interface
(HMI) of Siemens tools. After this introductory section, the paper is organized as:
The first section is ‘The used software’ which divided into 3 parts: Step 7 software,
WinCC Flexible software and Project Creation and Hardware (HW) configuration.
The second section concerns ‘The developed platform’. At the end of this paper the
conclusion is drawn.

2 The Used Software

The used software in this work is given below:

2.1 Step 7 Software

STEP7 is the basic software package for configuring and programming of S7 300 and
S7 400 programmable logic controller. It is part of the SIMATIC software industry.

2.2 WinCC Flexible Software

WinCC Flexible is a software compatible with the STEP7 environment, it offers the
configuration of various operator panels and SCADA system, moreover creating the
graphical interface (HMI) and the variables means being able to read the process
values via the PLC, display them so that the operator can supervise, control and
adjust them.
Supervision and Monitoring of Photovoltaic Systems … 1149

Fig. 1 SIMATIC STEP7 project

2.3 Project Creation and Hardware (HW) Configuration

The implementation of an automation solution with STEP7 requires the realization

of following basic tasks:
• Creation of the SIMATIC STEP7 project
• Hardware (HW) configuration
Figure 1 presents the picture of SIMATIC STEP7 project (PV system).
The picture of the hardware part is illustrated in Fig. 2.

3 The Developed Platform

Before giving the results of this work, it is more suitable to provide the following
detail: The CPU 314C-2 PN/DP PLC (Siemens) is used for programming using
Ladder language the process step, the main component of this PLC is given Table 1
below:
1150 A. Bouraiou et al.

Fig. 2 Hardware (HW) configuration

Table 1 PLC main

Main component Number Specification
component
Digital input (DI) 24 24 DC
Digital output (DO) 16 24 DC
Analog input (AI) 5 10 V or 0–20 mA
Analog output (AO) 2 10 V or 0–20 mA

In this work, the five analog input is used for measurement of the parameters like
current, voltage, irradiance, temperature and flow, the detail of these parameters via
symbolic configuration in Step 7 software is presented in Fig. 3.
Supervision and Monitoring of Photovoltaic Systems … 1151

Fig. 3 Symbol of parameters

The simulation of the PLC behavior in Step 7 is realized by S7-PLCSIM, it can

run the STEP 7 user program and test it in a simulated programmable controller,
S7-PLCSIM also offers a simple interface to the STEP 7 user program used to view
and modify various objects such as input and output variables.
The simulation of the different analog input in S7-PLCSIM for our application is
given Fig. 4.
The scaling of these input parameters was done in step 7 using the FC 105 (see
Fig. 5). The realization of the human-machine interface (HMI) is performed using
WinCC Flexible. The principal interface (Home) contains the general information
about the system, (see Fig. 6).
1152 A. Bouraiou et al.

Fig. 4 S7-PLCSIM
interface

The second interface (system) allows the supervision of the system parameters’
values in real-time like current, voltage, irradiance, temperature and flow, as presented
in Fig. 7. An example of real-time plotting curves is illustrated in Fig. 8.
Supervision and Monitoring of Photovoltaic Systems … 1153

Fig. 5 Input parameters scaling

Fig. 6 Main interface

1154 A. Bouraiou et al.

Fig. 7 Real-time system supervising

Fig. 8 Real-time system supervising

Supervision and Monitoring of Photovoltaic Systems … 1155

4 Conclusion

The supervision and monitoring of the industrial systems are two very important
fields in industrial automation. In this paper, a simple platform based on Siemens
PLC and HMI was realized in order to supervise and monitor the solar pumping
system parameters. The real realization of this platform requires the sensors signal
converting to universal signals such as 4–20 mA or 0–10 V. This study shows that
can use the PLC and HMI for resolving the supervision and monitoring issues in
photovoltaic systems.

References

1. Atoui I, Meradi H, Boulkroune R, Saidi R, Grid A (2013) Fault detection and diagnosis in
rotating machinery by vibration monitoring using FFT and Wavelet techniques. In: 2013 8th
international workshop on systems, signal processing and their applications (WoSSPA). IEEE,
pp 401–406. https://doi.org/10.1109/WoSSPA.2013.6602399
2. da Silva AM, Povinelli RJ, Demerdash NAO (2008) Induction machine broken bar and stator
short-circuit fault diagnostics based on three-phase stator current envelopes. IEEE Trans Ind
Electron 55:1310–1318. https://doi.org/10.1109/TIE.2007.909060
3. Attoui I, Boutasseta N, Fergani N, Oudjani B, Deliou A (2015) Vibration-based bearing fault
diagnosis by an integrated DWT-FFT approach and an adaptive neuro-fuzzy inference system.
In: 2015 3rd international conference on control, engineering & information technology (CEIT).
IEEE, pp 1–6. https://doi.org/10.1109/CEIT.2015.7233098
4. Attoui I, Boudiaf A, Fergani N, Oudjani B, Boutasseta N, Deliou A (2015) Vibration-based
gearbox fault diagnosis by DWPT and PCA approaches and an adaptive neuro-fuzzy inference
system. In: 2015 16th international conference on sciences and techniques of automatic control
and computer engineering (STA). IEEE, pp 234–239. https://doi.org/10.1109/STA.2015.750
5177
5. Madeti SR, Singh SN (2017) Monitoring system for photovoltaic plants: a review. https://doi.
org/10.1016/j.rser.2016.09.088
6. Attoui I, Oudjani B, Fergani N, Bouakkaz S, Bouraiou A (2020) Novel predictive features
using a wrapper model for rolling bearing fault diagnosis based on vibration signal analysis.
Int J Adv Manuf Technol 106:3409–3435. https://doi.org/10.1007/s00170-019-04729-4
7. Attoui I, Fergani N, Boutasseta N, Oudjani B, Bouakkaz MS, Bouraiou A (2021) Multiclass
support vector machine based bearing fault detection using vibration signal analysis. https://
doi.org/10.1007/978-981-15-6403-1_61
8. Attoui I, Boutasseta N, Fergani N (2021) Novel machinery monitoring strategy based on time-
frequency domain similarity measurement with limited labeled data. IEEE Trans Instrum Meas
70:1–8. https://doi.org/10.1109/TIM.2020.3011874
9. Attoui I (2020) Novel fast and automatic condition monitoring strategy based on small amount
of labeled data. IEEE Trans Syst Man Cybern Syst 1–10 (2020). https://doi.org/10.1109/TSMC.
2020.3018102
10. Boutasseta N, Ramdani M, Mekhilef S (2018) Fault-tolerant power extraction strategy for
photovoltaic energy systems. Sol Energy 169:594–606. https://doi.org/10.1016/J.SOLENER.
2018.05.031
11. Bouraiou A, Necaibia A, Boutasseta N, Mekhilef S, Dabou R, Ziane A, Sahouane N, Attoui I,
Mostefaoui M, Touaba O (2020) Status of renewable energy potential and utilization in Algeria.
J Clean Prod 246:119011. https://doi.org/10.1016/J.JCLEPRO.2019.119011
1156 A. Bouraiou et al.

12. Touaba O, Ait Cheikh MS, Slimani ME-A, Bouraiou A, Ziane A, Necaibia A, Harmim A
(2020) Experimental investigation of solar water heater equipped with a solar collector using
waste oil as absorber and working fluid. Sol Energy 199:630–644. https://doi.org/10.1016/j.sol
ener.2020.02.064
13. Attoui I, Omeiri A (2014) Modeling, control and fault diagnosis of an isolated wind energy
conversion system with a self-excited induction generator subject to electrical faults. Energy
Convers Manag 82:11–26. https://doi.org/10.1016/j.enconman.2014.02.068
14. Attoui I, Omeiri A (2014) Contribution to the fault diagnosis of a doubly fed induction gener-
ator for a closed-loop controlled wind turbine system associated with a two-level energy
storage system. Electr Power Components Syst 42:1727–1742. https://doi.org/10.1080/153
25008.2014.950361
15. Attoui I, Omeiri A (2015) Fault diagnosis of an induction generator in a wind energy conversion
system using signal processing techniques. Electr Power Components Syst 43:2262–2275.
https://doi.org/10.1080/15325008.2015.1082161
16. Bouakkaz MS, Boukadoum A, Boudebbouz O, Bouraiou A, Boutasseta N, Attoui I (2020)
ANN based MPPT algorithm design using real operating climatic condition. In: Proceedings
of the 2020 international conference on mathematics and information technology, ICMIT 2020,
pp 159–163. https://doi.org/10.1109/ICMIT47780.2020.9046972
17. Bouakkaz MS (2020) Fuzzy logic based adaptive step hill climbing MPPT algorithm for PV
energy generation systems. In: 2020 International conference on computing and information
technology (ICCIT-1441), pp 248–252
18. Fergani N, Charef A, Attoui I (2015) Desired closed-loop based self-tuning fractional PID
controller for wind turbine speed control. In: 2015 4th international conference on systems and
control, ICSC 2015. https://doi.org/10.1109/ICoSC.2015.7152771
19. Boutasseta N, Bouakkaz MS, Bouraiou A, Attoui I, Fergani N (2020) Practical implementation
of computational algorithms for efficient power conversion in photovoltaic energy generation
systems
20. Abderrezzaq Z, Mohammed M, Ammar N, Nordine S, Rachid D, Ahmed B (2017) Impact of
dust accumulation on PV panel performance in the Saharan region. In: 2017 18th international
conference on sciences and techniques of automatic control and computer engineering (STA),
pp 471–475. https://doi.org/10.1109/STA.2017.8314896
21. Mostefaoui M, Ziane A, Bouraiou A, Khelifi S (2018) Effect of sand dust accumulation on
photovoltaic performance in the Saharan environment: southern Algeria (Adrar). https://doi.
org/10.1007/s11356-018-3496-7
22. Mostefaoui M, Necaibia A, Ziane A, Dabou R, Rouabhia A, Khelifi S, Bouraiou A, Sahouane
N (2018) Importance cleaning of PV modules for grid-connected PV systems in a desert
environment. In: 2018 4th International Conference on Optimization and Applications, pp
1–6. https://doi.org/10.1109/ICOA.2018.8370518
23. Dabou R, Sahouane N, Necaibia A, Mostefaoui M, Bouchafaa F, Rouabhia A, Ziane A,
Bouraiou A (2017) Impact of partial shading and PV array power on the performance of
grid connected PV station. In: 2017 18th International Conference on Sciences and Techniques
of Automatic Control and Computer Engineering, pp 476–481. https://doi.org/10.1109/STA.
2017.8314901
24. Motahhir S, Chouder A, Hammoumi AEl, Benyoucef AS, Ghzizal AEl, Kichou S, Kara K,
Sanjeevikumar P, Silvestre S (2020) Optimal energy harvesting from a multistrings PV gener-
ator based on artificial bee colony algorithm. IEEE Syst J 1–8 (2020). https://doi.org/10.1109/
jsyst.2020.2997744
25. Motahhir S, El Hammoumi A, El Ghzizal A (2020) The most used MPPT algorithms: review
and the suitable low-cost embedded board for each algorithm. https://doi.org/10.1016/j.jclepro.
2019.118983
26. Blal M, Khelifi S, Dabou R, Sahouane N, Slimani A, Rouabhia A, Ziane A, Neçaibia A,
Bouraiou A, Tidjar B (2020) A prediction models for estimating global solar radiation and
evaluation meteorological effect on solar radiation potential under several weather conditions
at the surface of Adrar environment. Measurement 152:107348. https://doi.org/10.1016/J.MEA
SUREMENT.2019.107348
Supervision and Monitoring of Photovoltaic Systems … 1157

27. Sahouane N, Dabou R, Ziane A, Neçaibia A, Bouraiou A, Rouabhia A, Mohammed B (2019)

Energy and economic efficiency performance assessment of a 28 kWp photovoltaic grid-
connected system under desertic weather conditions in Algerian Sahara. Renew Energy. https://
doi.org/10.1016/j.renene.2019.05.086
28. Bouakkaz MS, Boukadoum A, Boudebbouz O, Fergani N, Boutasseta N, Attoui I, Bouraiou A,
Necaibia A (2020) Dynamic performance evaluation and improvement of PV energy generation
systems using Moth Flame Optimization with combined fractional order PID and sliding mode
controller. Sol Energy 199:411–424. https://doi.org/10.1016/j.solener.2020.02.055
Maximum Power Point Tracking Using
SEPIC Converter and Double Diode
Solar Cell Model

Khalid Chennoufi and Mohammed Ferfra

Abstract In this paper a hybrid controller for tracking the maximum power point of
photovoltaic modules is proposed. For an accurate modelling, the double diode model
was adopted, and so as to decrease the current undulation the single-ended primary-
inductor-converter was used. The proposed controller is developed by integrating
sliding mode and Artificial Neural Network, the latter has been designed to delivers
an optimal voltage, which corresponds to the voltage of the maximum power point,
whereas the sliding mode was developed to track the signal of the reference voltage
by computing the duty cycle of the converter. The simulation has been carried out
using Simulink software under different environmental conditions, the results show
that the MPPT controller tracks the reference voltage in 80 ms, and exhibits good
performance at brusque variation of temperature and irradiation, in addition the
accuracy of the proposed method was further justified against previous work.

Keywords MPPT · SEPIC · Double diode model · ANN · Sliding mode

1 Introduction

The photovoltaic energy is the principal source of renewable energy, which gener-
ates electricity by using photovoltaic modules that convert sun energy into elec-
trical power. In fact solar power is more cost-effective than standard electricity.
However, the power supply output depends on the load demand, which evokes
researchers to propose several algorithms that force the photovoltaic panel to works
at its maximum power. The most commonly employed is perturb and observe [1–3]
due to its simplicity of implementation, but it’s unstable and suffers from oscilla-
tions around the maximum power point. So, to overcome this problem, the non-linear
commands are employed such as sliding mode [4, 5], backstepping [6, 7], and fuzzy
logic [8, 9]. So as to increase the accuracy of the tracking, authors in [10], propose
a hybrid method which integrates perturb and observe (PO) loop in sliding mode

K. Chennoufi (B) · M. Ferfra

Mohammadia School of Engineers, Mohammed V University in Rabat, Rabat, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1159
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_106
1160 K. Chennoufi and M. Ferfra

controller (SMC), the latter was conceived to track reference voltage generated with
PO algorithm. The proposed approach illustrates good results. However, there are
still oscillations around the maximum power point.
So as to surmount the fluctuations issue around the maximum power, the present
paper proposes an MPPT controller which combines Artificial Neural Network
(ANN) and sliding mode controller (SMC) using SEPIC converter. The ANN is
conceived to estimate the optimum voltage that generates the maximum power and
the sliding mode is designed to track the desired voltage by changing the duty cycle
of the converter (Fig. 1).
Unlike the previous works which use the one-diode model for PV modules and a
traditional DC-DC converter, the present work adopts the double diode model for an
accurate result, and a SEPIC converter, because it offers additional advantages than
the traditional DC-DC converter.
This paper is structured as follows. The extraction of the module parameters and
the modelling of the converter are presented in the Sect. 2. The Sect. 3 is devoted to
the MPPT controller design. The Sect. 4 is dedicated to the simulation results and
discussions.

2 Photovoltaic System Design

The solar power system considered in this work is a PV panel, coupled to a resistive
load with a SEPIC as illustrated in the Fig. 1. The specifications of the PV panel KD
245 are illustrated in Table 1.

Fig. 1 Proposed
photovoltaic system
Maximum Power Point Tracking Using SEPIC Converter … 1161

Table 1 Polycrystalline PV
Parameter Value
panel specifications
Isc (A) 8.91
Voc (V) 36.9
Im (A) 8.23
Vm (V) 29.8
Ns 60

Fig. 2 Equivalent circuit of

the SEPIC converter,
(a) switch is ON, (b) switch
is OFF

2.1 Analyse and Modelling of the SEPIC

The single ended primary inductor converter is used since it can reach maximum
power with feeble current’s undulation [11], and it is able to work in a buck or a
boost mode [12]. The mathematical modelling of the converter is carried out based
on these states [13, 14] as illustrated in the Fig. 2.
Figure 2-a shows the equivalent circuit of the SEPIC when the switch is on (u
= 1). In this case the output is supplied from the capacitor C3 , and the amount of
current stored in the inductors raises. The input and the capacitor C2 supply energy to
the inductor L1 and to the inductor L2 respectively. By using capacitor and inductor
equations the representation of state space is as follows:
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡V ⎤
dIL1 0 0 0 0 IL1 pv

⎢
dt
dVC2 ⎢ ⎥ ⎢
⎥ ⎢0 0 C 0 ⎥⎢ V ⎥ ⎢ 0
1 ⎥ ⎢ L1
⎥
⎢ ⎥ = ⎢ −1 2 ⎥
⎣
dt
⎦ ⎣0 ⎥ ⎢ c2 ⎥ + ⎢ ⎥
dIL2
dt L2
0 0 ⎦ ⎣ IL2 ⎦ ⎣ 0 ⎦
dV0 −1
dt 0 0 0 RC 3
Vo 0

Figure 2-b shows the equivalent circuit of the SEPIC when the switch is off (u =
0). In this case the inductor L1 and L2 are discharged, which raises the energy stored
in the capacitors. By using capacitor and inductor equations the representation of
state space is as follows:
1162 K. Chennoufi and M. Ferfra

⎡ ⎤ ⎡ ⎤⎡
−1
⎤ ⎡V ⎤
dIL1 0 −10
L1 L1 IL1 pv

⎢
dt
dVC2 ⎢ ⎥ ⎢
⎥ ⎢ C1 0 0 0 ⎥ ⎢ V ⎥ ⎢ 0⎥ ⎢ L1
⎥
⎢ ⎥=⎢ 2 ⎥
⎣
dt
⎦ ⎣ 0 0 0 1 ⎥ ⎢ c2 ⎥ + ⎢ ⎥
L2 ⎦ ⎣ L2 ⎦ ⎣ 0 ⎦
dIL2
dt
I
dV0 1 −1 −1 V
dt C3
0 C3 RC3 o
0

The state space for u = 0 or u = 1 can be is expressed as follows:

⎡ ⎤ ⎡ u−1 u−1
⎤⎡ ⎤ ⎡V ⎤
dIL1 0 0 IL1 pv
L1 L1
⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
dt L1
⎢ dVC2 1−u u
⎢ ⎥=⎢
⎢
0 0 ⎥ ⎢ V C2 ⎥ ⎢ 0
⎥+⎢
⎥
1−u ⎥ ⎢ ⎥
dt C2 C2
⎣ ⎦ ⎣ 0 −u
L2 ⎦ ⎣ L2 ⎦ ⎣ 0 ⎦
dIL2
dt L2
0 I
dV0 1−u u−1 −1 V
dt C3
0 C3 RC3 o
0

Thus, the mathematical modelling can be expressed as follows:

dIL1 (u − 1)VC2 (u − 1)Vo VPV

= + + (1)
dt L1 L1 L1
dVC2 (1 − u)IL1 dIL2
= + (2)
dt C2 C2
dIL2 uVC2 (1 − u)Vo
=− + (3)
dt L2 L2
dV0 (1 − u)IL1 (u − 1)IL2 VO
= + − (4)
dt C3 C3 RC3

2.2 Photovoltaic Modelling

The single diode model is widely used for the modelling of photovoltaic modules due
to its simplicity. However, the accuracy of the model may suffer, as the absolute errors
of current and voltage are relatively high [15]. Thus, in order to have an accurate
modelling the double diode model is considered in this work, which is represented
in Fig. 3.
The output current of a solar cell can be represented by the following equation:
   
V + Rs I V + Rs I V + Rs I
I = Iph − I01 exp − 1 − I02 exp −1 − (5)
a1 V T a2 VT Rp

Where I: Current of PV module, Iph : Photocurrent, I01 and I02 saturation current
for diode 1 and diode 2 respectively, Rp : Cell parallel resistance, Rs : Cell series
resistance, a1 and a2 ideality factor for diode 1 and diode 2 respectively, V: Voltage
of PV module in voltage, VT : Thermal voltage.
Maximum Power Point Tracking Using SEPIC Converter … 1163

Fig. 3 Double-diode
equivalent circuit

Table 2 Parameters of the

Parameter Value
module KD 245GH
a1 1.005879
a2 1.994120
Iph (A) 8.931164
I01 (A) 3.906668 × 10–10
I02 (A) 1.094485 × 10–06
Rs () 0.286199
Rp () 120.495147

The modelling of the PV module is carried out using mathematical operations

using the three operating points, that are (VOC − 0), (0 − ISC ) and (Vm − Im ). The
aim of this calculus is to make I01 , I02 , Rp and Iph depending only on Rs , a1 and a2 ,
these parameters are obtained by adjusting computed and experimental powers [16].
The obtained parameters for the module KD 245 are given in Table 2.

3 Proposed Controller

In this section a sliding mode controller (SMC) combined with Artificial Neural
Network (ANN) is designed as shown in Fig. 1. The ANN provides a voltage reference
for any environmental condition that corresponds to maximum power voltage, while
the SMC tracks this reference by computing the duty cycle of the converter.

3.1 Design of the ANN

The Artificial Neural Network is a data analysis and computational algorithm which
consists of a related group of nodes inspired from biological neuron [17]. The ANN
is the most suitable algorithm for the nonlinear system, its architecture consists of
three layers, the input layer collects the external information, the second layer is the
1164 K. Chennoufi and M. Ferfra

Fig. 4 Neural network regression

hidden layer, contains hidden neurons that receive information from the input layer
and send them to the output layer [18].
In this paper, the input layer was designed with a database of 158 irradiation and
temperature cases, the temperature has been increased by a step of 5 °C, from 5 °C to
60 °C, and the irradiance was increased with a step of 100 W/m2 for each temperature
degrees, whereas the hidden layer has been conceived with 100 neurons. The output
layer has been conceived with one neuron that corresponds to the maximum power
voltage for every 158 irradiation and temperature cases. The ANN model fitting shows
the outputs for the targets for training and test sets. The R-square value was 0.99999
for all training sets, as shown in Fig. 4, which illustrates an excellent relationship
between the output and target sets.

3.2 Sliding Mode Controller

The SMC is designed in order to tracks the maximum power voltage, Vpv and Vref
are the photovoltaic and the reference voltages respectively.
The tracking error (e) is expressed as follows:

e = VPV − Vref (6)

The time derivative of Eq. (6) is:

ė = V̇PV − V̇ref (7)

The photovoltaic current is:

IPV = IC1 + IL1 (8)

Maximum Power Point Tracking Using SEPIC Converter … 1165

The derivative of photovoltaic voltage is:

IPV − IL1
V̇pv = (9)
C1

Thus:
IPV − IL1
ė = − Vref (10)
C1

The sliding surface is defined by:

S = e − λė (11)

The following dynamic sliding surface is used:

Ṡ = AS − Bsign (S) (12)

The time derivative of (11) with respect to time is:

Ṡ = ė − λë (13)

Where λ > 0.
Second derivative of (10) is given by:

İPV − İL1
ë = − V̈ref (14)
C1

By replacing Eq. (1) into (14) the Eq. (15) is obtained:


(u−1)VC1 (u−1)Vo
İPV − L1
+ L1
+ VPV
L1
ë = − V̈ref (15)
C1

The time derivative of the current can be expressed as:

dIPV dVPV
İPV = (16)
dVPV dt

By introducing the Eq. (5) in the Eq. (16) the derivative of photovoltaic current
can be expressed as:
⎡   ⎤
V+Rs IPV V+Rs IPV
I01 exp a1 VT
I02 exp a2 VT 1 ⎦
İPV = ⎣− − − V̇PV (17)
a1 VT a2 VT RP
1166 K. Chennoufi and M. Ferfra

By replacing Eq. (17) into (14) and combining (13) with (12) the command u(t)
is obtained:

       1
u = 1 − VPV − L1 C1 (AS + Bsign(S)) − λ V̇PV − V̇ref + V̈ref + İPV
VC2 + V0
(18)

4 Simulation and Discussions

The tracking accuracy was evaluated using KD 245 module connected to a 40  load
via SEPIC converter which the parameters are shown in Table 3.
Figure 5 shows the temperature and irradiation cases used in this simulation. In
the first interval [0 s, 23 s], the photovoltaic panel operates under the maximum
irradiation, whereas the temperature has been increased to 45 °C. In the next case
[23 s, 48 s], the irradiation drops suddenly to 580 W/m2 , and the temperature drops
in two successive levels to 25 °C. In the last case [48 s, 60 s], the temperature
is maintained constant at 25 °C, while the irradiation has been increased to reach
1000 W/m2 .

Table 3. Component of the converter

Component Value
Inductor 1 (L1 ) 5.1 mH
Inductor 2 (L2 ) 5.1 mH
Capacitor 1 (C1 ) 4.5 × 103 uF
Capacitor 2 (C2 ) 7.9 × 10 μF
Capacitor 3 (C3 ) 2.4 × 103 μF
Constant (λ) 106
Constant (A) 106
Constant (B) 103

(a) (b)
Irradiance (W/m²)

Temperature (K)

Time (S) Time (S)

Fig. 5 Scenarios of irradiation (a) and Scenarios of temperature (b)

Maximum Power Point Tracking Using SEPIC Converter … 1167

Vref (ANN)
Present method
Method proposed by [10]
(a) (b)
Voltage (V)

Power (W)
Time (S) Time (S)

Fig. 6 PV voltage (a) and PV power (b), at STC.

Figure 6 illustrates a comparative results in standard test conditions (irradiation

of 1000 W/m2 and a temperature of 298 K), between the proposed method and
sliding mode combined with P&O proposed by [10]. It can be seen in the figure (a)
that the present controller reaches the desired voltage in 80 ms, which guaranty the
maximum power output with slight oscillations that fade during the tracking, Figure
(b). Whereas the method mentioned before arrives at the desired voltage at 140 ms
with large oscillations which remain throughout the tracking. Therefore the proposed
method provides good and fast reaction at any environmental condition.
Figure 7 shows the voltage and the current of the PV module using the proposed
method under the different weather scenarios. The results show that the Artificial
Neural Network generates the reference voltage at any the temperature and the irra-
diance conditions, and the sliding mode tracks this voltage with high efficiency which
ensure a smooth voltage and current curves.
Figure 8 illustrates the current and voltage of the converter using the proposed
method. It can be seen that the system reacts with good efficiency and that it is able
to operate under all environmental conditions. In addition, the SEPIC operates in this
case as a voltage boost, which is suitable for equipment that requires high voltage.

(a) (b)
Current (A)
Voltage (V)

Time (S) Time (S)

Fig. 7 PV voltage (a) and PV current (b)

1168 K. Chennoufi and M. Ferfra

(a) (b)
Voltage (V)

Current (A)
Time (S) Time (S)

Fig. 8 Output voltage (a) and output current (b)

5 Conclusion

This paper presents a sliding mode control combined to the Artificial Neural Network
for Maximum Power Point Tracking of a PV system, which consists of a photovoltaic
module, modelled with double diode model and a SEPIC converter connected to a
resistive load. The assessment of the efficiency of the proposed method has been
carried out by simulating the proposed controller in Matlab/Simulink software under
different weather conditions. In one hand the simulation demonstrates its efficiency
and stability against the previous work, in standard test condition. In the other hand,
the proposed approach exhibits a fast reaction to track the reference voltage with good
performance even when the environmental condition changes. In addition there is no
oscillation in the obtained curves and the algorithm can easily implemented, which
justify the interest behind its use.

References

1. Thakran S, Singh J, Garg R, Mahajan, P (2018) Implementation of P&O algorithm for MPPT in
SPV system. In: International conference on power energy, environment and intelligent control
(PEEIC), Greater Noida, India
2. Ishaque K, Salam Z, Lauss G (2014) The performance of perturb and observe and incremental
conductance maximum power point tracking method under dynamic weather conditions. Appl
Energy 119:228–236
3. Yong Y, Zhaob FP (2011) Adaptive perturb and observe MPPT technique for grid connected
photovoltaic inverters. Procedia Eng 23:468–473
4. Islam S, Liu XP (2011) Robust sliding mode control for robot manipulators. IEEE Trans Ind
Electron 58(6):2444–2453
5. Lachtar S, Bouraiou A, Djaafri O, Maouedj R (2019) Smooth sliding mode-based MPPT algo-
rithm for photovoltaic applications. In: 1st global power, energy and communication conference
(IEEE GPECOM 2019), Cappadocia, Turkey
6. Parsa A, Kalhor A, Atashgah MAA (2016) Backstepping control performance enhance-
ment using close loop identification for quadrotor trajectory tracking. Modares Mech Eng
16(11):224–234
7. Zhang Y, Fidan B, Ioannou P (2003) Backstepping control of linear time-varying systems with
known and unknown parameters. IEEE Trans Autom Control 48(11):1908–1925
Maximum Power Point Tracking Using SEPIC Converter … 1169

8. Belcher R, Fattahi J, Hinzer K, Schriemer H (2020) Fuzzy logic based module-level power
electronics for mitigation of rapid cloud shading in photovoltaic systems. In: IEEE Applied
Power Electronics Conference and Exposition (APEC), Hioki, USA
9. Ziane A, Necaibia A, Mostfaoui M, Bouraiou A, Sahouane N, Dabou R (2018) A fuzzy logic
MPPT for three-phase grid-connected PV inverter. In: Twentieth international middle east
power systems conference (MEPCON), Cairo, Egypt
10. Mostafa MR, Saad NH, El-sattar AA (2020) Tracking the maximum power point of PV array
by sliding mode control method. Ain Shams Eng J 11:119–131. https://doi.org/10.1016/j.asej.
2019.09.003
11. Efendi MZ, Murdianto FD, Setiawan RE (2017) Modeling and simulation of MPPT sepic
converter using modified PSO to overcome partial shading impact on DC microgrid system.
In: International electronics symposium on engineering technology and applications, Surabaya,
Indonesia
12. Duran E, Sidrach-de-Cardona M, Galan J, Andujar JM (2008) Comparative analysis of buck-
boost converters used to obtain I-V characteristic curves of photovoltaic modules. In: 2008
IEEE power electronics specialists conference, Rhodes, Greece
13. Kashyap AR, Ahmadi R, Kimball JW (2013) Input voltage control of sepic for maximum
power point tracking. In: Power and energy conference at Illinois (PECI), pp. 30–35. IEEE
14. Erickson RW (2000) Fundamentals of power electronics, 2nd edn. Springer, Heidelberg, p 910
15. Chennoufi K, Ferfra M (2020) Parameters extraction of photovoltaic modules using a combined
analytical - numerical method. In: International conference on cloud computing technologies
and applications (CloudTech), Marrakech, Morocco
16. Chennoufi K, Ferfra M, Mokhlis M (2021) An accurate modelling of photovoltaic modules
based on two-diode model. Renew Energy 167:294–305
17. Nambiar N, Palackal RS, Greeshma KV, Chitra A (2015) PV fed MLI with ANN based
MPPT. In: 2015 international conference on computation of power, energy, information and
communication (ICCPEIC), Chennai
18. Saadi A, Moussi A (2003) Neural network use in the MPPT of photovoltaic pumping system.
Rev Renew Energies ICPWE 4:39–45
Global Maximum Power Point Tracking
Using Genetic Algorithm Combined
with PSO Tuned PID Controller

Mohammed Salah Bouakkaz, Ahcene Boukadoum, Omar Boudebbouz,

Abdel Djabar Bouchaala, Nadir Boutasseta, Issam Attoui, Ahmed Bouraiou,
and Saad Motahhir

Abstract Several techniques have been proposed to track the global maximum
power point (GMPP) in the presence of multiple local peaks introduced when the
PV generator is affected by partial shading. Such methods include scanning methods
and advanced optimization-based strategies. In this paper, a Genetic Algorithm based
MPPT, combined with PID controller tuned using PSO, is proposed to track the MPP.
The GA algorithm is applied to give the reference voltage and measured PV voltage
to the PID controller that its parameters were tuned using PSO. The obtained results
give a good tracking performance of the GMPP under STC conditions and partial
shading scenarios.

Keywords Maximum power point tracking (MPPT) · Genetic algorithm (GA) ·

PID controller · Particle Swarm Optimization (PSO)

1 Introduction

Photovoltaic (PV) solar energy is one of the most important renewable energies for the
following reasons: clean and non-polluting energy, economical and maintainable, the
availability of solar radiation in most countries of the world [1, 2]. The photovoltaic
solar system consists of a group of elements which include: photovoltaic solar panels,
charge controller (including MPPT), storage batteries, and the power inverter (for
AC loads). In general; solar photovoltaic energy in terms of load connection into two

M. S. Bouakkaz (B) · A. Boukadoum · O. Boudebbouz · A. D. Bouchaala

Department of Electrical Engineering, Laboratoire d’électrotechnique de Skikda «LES»,
Université du 20 Août 1955, 21000 Skikda, Algeria
N. Boutasseta · I. Attoui
Research Center in Industrial Technologies CRTI, P.O. Box 64, Cheraga, Algiers, Algeria
A. Bouraiou
Unité de Recherche en Energies Renouvelables en Milieu Saharien (URERMS),
Centre de Développement des Energies Renouvelables (CDER), 01000 Adrar, Algeria
S. Motahhir
ENSA, SMBA University, Fez, Morocco
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1171
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_107
1172 M. S. Bouakkaz et al.

types: isolated from the grid (which is a simplified system consisting of solar PV
that feeds a load), or connected to the grid (which is a complex system that includes
solar PV, local load and grid) [3].
The operation of solar photovoltaic modules is dependent on climatic conditions
lake solar radiation and temperature of the cell, which is also considered to have
a non-linear current-voltage characteristic curve. The output power of PV arrays
depends on the characteristic curves (I-V and P-V) which are nonlinear and change
with environmental conditions [4]. The maximum power is calculated by the product
of current and voltage, this product is known as the maximum power point (MPP)
which is not constant. For this purpose, the maximum power point tracking technique
(MPPT) is used to determine and extract the MPP value [5]. The strategy to approach
MPPT varies from method to method which is classified into two classes, conven-
tional MPPT and artificial intelligence-based approaches [6–8]. Conventional MPPT
technique such as the incremental conductance technique (IC), perturb and observe
(P&O) [9], open-circuit voltage (OCV), and short circuit current [10]. On the other
hand, artificial intelligence techniques are different and developed using various
algorithms such as fuzzy logic (FL) [11], Genetic algorithm (GA) [12], Adaptive
Neuro-Fuzzy Inference System (ANFIS) [13–16], Artificial Neural Network (ANN)
[17] and Particle Swarm Optimization (PSO) [18, 19].
In this paper, the proposed method is based on GA and PSO tuned PID
controller, which combines two techniques of artificial intelligence approaches that
are applied to MPPT to track global MPP in solar PV systems operating in different
climatic conditions. Simulations in Matlab/Simulink are carried out to implement
the proposed method through which results are obtained. In the following section,
the PV energy generation system which combines the PV array and power converter
are presented. The proposed MPPT strategy in this work is illustrated in Sect. 3.
Section 4 presents simulation results and gives discussions. Conclusions are given
in the end.

1.1 Photovoltaic Array Model

The PV cell has been the subject of modeling studies resulting in the three known
and most used models depending on the application. When Maximum Power Point

Fig. 1 Bloc representation

of a PV system
Global Maximum power Point Tracking Using Genetic Algorithm … 1173

Tracking is selected as a target application as in [20], the single-diode model is

generally used for simulation purposes. The main characteristics of the single-diode
model are the simpler calculations and acceptable accuracy when compared with
the two and three diodes models which require advanced parameters identification
techniques as in [21]. In the consequent work, the representations shown in Fig. 1
are used for the PV array energy generation system.
The output current I pv of the PV array composed of N pp panels in parallel and N ss
panels in the series, which is formulated as follows:

v pv + I pv Rs N SS
I pv = I ph − Id − (1)
R p N SS

Where V t represents the thermal voltage of the PV array (V t = N s kT/q), a is

a diode constant, with N s series cells, q is the electron charge, k the Boltzmann
constant. Rsh is the shunt resistance. Rs is the series resistance that depends on the
material applied to realize the PV cell.

1.2 Step-Up Converter

A step-up converter is a converter that produces a high voltage at its output terminal
than its input voltage. The boost DC-DC converter a type of switched-mode power
supply consisting of a minimum of two switching devices and one energy storage
element, a capacitor, an inductor, and a filter is generally inserted at the output to
reduce the ripple of VR . The boost converter is illustrated in Fig. 2.

Fig. 2 Step-up converter

1174 M. S. Bouakkaz et al.

Fig. 3 Photovoltaic system developed with GA-based MPPT and PSO tuned PID controller
(proposed)

1.3 Association PV Array with the Step-Up Converter

The association PV array/step-up converter shown in Fig. 3 is connected to a load

and has three main inputs, namely: Temperature (T), Solar Irradiation (E), and PWM
control signal of the Switch (S). The role of the MPPT technique is to generate
adequate control signals to track the optimal operating point in the presence of
variation in T and E.

2 Proposed MPPT Strategy

The proposed MPPT strategy illustrated in Fig. 3 is applied to find the MPP under
regular and PS weather conditions.

2.1 GA Based MPPT

The GA is applied to find the GMPP of the PV system under different operating
conditions. Figure 4 presented the flowchart of the consequent steps of the GA-based
MPPT.
Global Maximum power Point Tracking Using Genetic Algorithm … 1175

Fig. 4 Flowchart of GA
based MPPT

2.2 PID Controller Tuning Using the PSO Algorithm

The PID is a commonly used controller in industrial control systems [22–24]. A

PID controller gives continuously a correction value based on its PID terms on a
calculated error between a reference voltage (V ref ) and a measured value (V PV ). In
this work, a PID controller is used to track the reference operating point of the GMPP
generated by the GA. The PID is characterized by the parameters (kc, Ti, Td) that
are trained using a Particle Swarm Optimization (PSO) technique [25] as presented
in Fig. 5.
The PSO tuning algorithm uses reference voltage and measured PV voltage to
tune the parameters of the PID that output the duty ratio, which will adjust the actual
operating point using the switch S. The position update of the swarm is illustrated in
Fig. 6.

Fig. 5 Tuning of PID controller parameters using PSO

1176 M. S. Bouakkaz et al.

Fig. 6 Particle Swarm

Optimization position update
representation

Fig. 7 Solar irradiation and temperature patterns for simulation scenarios

3 Simulation Results and Discussions

The proposed GA combined with the PSO-PID controller is designed and tested under
MATLAB/Simulink using the Kyocera KC200GT PV panel. In this work, changes
in solar radiation and temperature are applied to verify and assess the robustness
of the proposed controller. Irradiation and temperature patterns that are used in the
consequent simulations are shown in Fig. 7.
As demonstrated in Fig. 8, which PV tracking performance when using the genetic
algorithm-based MPPT algorithm combined with PSO tuned PID controller, the
Global Maximum power Point Tracking Using Genetic Algorithm … 1177

Fig. 8 The PV array’s response to a variable scenario: (a) STC, (b) P1, (c) P2, (d) P3
1178 M. S. Bouakkaz et al.

proposed method tracks the global optimal operating point of the PV system in
every scenario efficiently with less tracking time. It is also characterized by fast
convergence speed, high PV tracking efficiency, and low power loss.

4 Conclusions

In this paper, a global MPPT strategy to improve the efficiency of the PV system
under regular and partial shading environments has been suggested. In this work,
the impacts of various partial shading patterns on the output of the PV system is
also investigated. The proposed strategy is developed to track the optimum operating
point of the PV system. The controller is employed in the voltage control loop
to accurately regulate the optimum operating voltage. The proposed MPPT control
procedure provides excellent performances in the tracking of the generated reference
from the genetic algorithm, which is verified by different simulation results presented
in the paper. The present work may be extended by considering the integration of the
proposed method with a PV inverter to build the solar photovoltaic grid-connected
power generation system.

References

1. Bouraiou A, Necaibia A, Boutasseta N, Mekhilef S, Dabou R, Ziane A, Sahouane N, Attoui I,

Mostefaoui M, Touaba O (2020) Status of renewable energy potential and utilization in Algeria.
J Clean Prod 246:119011. https://doi.org/10.1016/j.jclepro.2019.119011
2. Blal M, Khelifi S, Dabou R, Sahouane N, Slimani A, Rouabhia A, Ziane A, Neçaibia A,
Bouraiou A, Tidjar B (2020) A prediction models for estimating global solar radiation and
evaluation meteorological effect on solar radiation potential under several weather conditions
at the surface of Adrar environment. Measurement 152:107348. https://doi.org/10.1016/J.MEA
SUREMENT.2019.107348
3. Sahouane N, Dabou R, Ziane A, Neçaibia A, Bouraiou A, Rouabhia A, Mohammed B
(2019) Energy and economic efficiency performance assessment of a 28 kWp photovoltaic
grid-connected system under desertic weather conditions in Algerian Sahara. Renew Energy
143:1318–1330. https://doi.org/10.1016/j.renene.2019.05.086
4. Bouraiou A, Hamouda M, Chaker A, Mostefaoui M, Lachtar S, Sadok M, Boutasseta N,
Othmani M, Issam A (2015) Analysis and evaluation of the impact of climatic conditions
on the photovoltaic modules performance in the desert environment. Energy Convers Manag
106:1345–1355. https://doi.org/10.1016/j.enconman.2015.10.073
5. Lachtar S, Bouraiou A, Diaafri O, Maouedj R (2019) Smooth sliding mode-based MPPT
algorithm for photovoltaic applications. https://doi.org/10.1109/gpecom.2019.8778484
6. Motahhir S, El Hammoumi A, El Ghzizal A (2020) The most used MPPT algorithms: review
and the suitable low-cost embedded board for each algorithm. https://doi.org/10.1016/j.jclepro.
2019.118983
7. Motahhir S, Chouder A, El Hammoumi A, Benyoucef AS, El Ghzizal A, Kichou S, Kara
K, Sanjeevikumar P, Silvestre S (2020) Optimal energy harvesting from a multistrings PV
generator based on artificial bee colony algorithm. IEEE Syst J 1–8. https://doi.org/10.1109/
jsyst.2020.2997744
Global Maximum power Point Tracking Using Genetic Algorithm … 1179

8. Mirza AF, Mansoor M, Ling Q (2020) A novel MPPT technique based on Henry gas solubility
optimization. Energy Convers Manag 225:113409. https://doi.org/10.1016/j.enconman.2020.
113409
9. Motahhir S, Chalh A, El Ghzizal A, Derouich A (2018) Development of a low-cost PV system
using an improved INC algorithm and a PV panel proteus model. J Clean Prod 204:355–365.
https://doi.org/10.1016/j.jclepro.2018.08.246
10. Motahhir S, El Hammoumi A, El Ghzizal A (2018) Photovoltaic system with quantitative
comparative between an improved MPPT and existing INC and P&O methods under fast
varying of solar irradiation. Energy Rep 4:341–350. https://doi.org/10.1016/j.egyr.2018.04.003
11. Bouakkaz MS, Boukadoum A, Boudebbouz O, Attoui I, Boutasseta N, Bouraiou A (2020)
Fuzzy logic based adaptive step hill climbing MPPT algorithm for PV energy generation
systems. In: 2020 International conference on computing and information technology (ICCIT-
1441), pp 1–5. IEEE. https://doi.org/10.1109/ICCIT-144147971.2020.9213737
12. Daraban S, Petreus D, Morel C (2014) A novel MPPT (maximum power point tracking) algo-
rithm based on a modified genetic algorithm specialized on tracking the global maximum power
point in photovoltaic systems affected by partial shading. Energy 74:374–388. https://doi.org/
10.1016/j.energy.2014.07.001
13. Attoui I, Boudiaf A, Fergani N, Oudjani B, Boutasseta N, Deliou A (2015) Vibration-based
gearbox fault diagnosis by DWPT and PCA approaches and an adaptive neuro-fuzzy inference
system. In: 2015 16th international conference on sciences and techniques of automatic control
and computer engineering (STA), pp 234–239. IEEE. https://doi.org/10.1109/STA.2015.750
5177
14. Attoui I, Oudjani B, Boutasseta N, Fergani N, Bouakkaz M-S, Bouraiou A (2020) Novel
predictive features using a wrapper model for rolling bearing fault diagnosis based on vibration
signal analysis. Int J Adv Manuf Technol 106:3409–3435. https://doi.org/10.1007/s00170-019-
04729-4
15. Attoui I, Boutasseta N, Fergani N, Oudjani B, Deliou A (2015) Vibration-based bearing fault
diagnosis by an integrated DWT-FFT approach and an adaptive neuro-fuzzy inference system.
In: 3rd international conference on control, engineering and information technology, CEIT
2015. https://doi.org/10.1109/CEIT.2015.7233098
16. Lasheen M, Abdel-Salam M (2018) Maximum power point tracking using hill climbing and
ANFIS techniques for PV applications: a review and a novel hybrid approach. Energy Convers
Manag 171:1002–1019. https://doi.org/10.1016/j.enconman.2018.06.003
17. Bouakkaz MS, Boukadoum A, Boudebbouz O, Bouraiou A, Boutasseta N, Attoui I (2020)
ANN based MPPT algorithm design using real operating climatic condition. In: 2020 2nd
international conference on mathematics and information technology (ICMIT), pp 159–163.
IEEE. https://doi.org/10.1109/ICMIT47780.2020.9046972
18. Boutasseta N, Ramdani M, Mekhilef S (2018) Fault-tolerant power extraction strategy for
photovoltaic energy systems. Solar Energy 169:594–606. https://doi.org/10.1016/j.solener.
2018.05.031
19. Boutasseta N, Bouakkaz MS, Bouraiou A, Necaibia A, Attoui I, Fergani N (2020) Practical
implementation of computational algorithms for efficient power conversion in photovoltaic
energy generation systems. In: 2020 international conference on computing and information
technology (ICCIT-1441), pp 1–5. IEEE. https://doi.org/10.1109/ICCIT-144147971.2020.921
3761
20. Shams I, Saad M, Kok Soon T (2020) Improved team game optimization algorithm based
solar MPPT with fast convergence speed and fast response to load variations. IEEE Trans Ind
Electron 1. https://doi.org/10.1109/tie.2020.3001798
21. Qais MH, Hasanien HM, Alghuwainem S (2020) Parameters extraction of three-diode photo-
voltaic model using computation and Harris Hawks optimization. Energy 195:117040. https://
doi.org/10.1016/j.energy.2020.117040
22. Al-Dhaifallah M, Nassef AM, Rezk H, Nisar KS (2018) Optimal parameter design of fractional
order control based INC-MPPT for PV system. Solar Energy 159:650–664. https://doi.org/10.
1016/j.solener.2017.11.040
1180 M. S. Bouakkaz et al.

23. Attoui I, Omeiri A (2014) Contribution to the fault diagnosis of a doubly fed induction generator
for a closed-loop controlled wind turbine system associated with a two-level energy storage
system. Electr Power Compon Syst 42:1727–1742. https://doi.org/10.1080/15325008.2014.
950361
24. Attoui I, Omeiri A (2015) Fault diagnosis of an induction generator in a wind energy conversion
system using signal processing techniques. Electr Power Compon Syst 43:2262–2275. https://
doi.org/10.1080/15325008.2015.1082161
25. Bouakkaz MS, Boukadoum A, Boudebbouz O, Fergani N, Boutasseta N, Attoui I, Bouraiou A,
Necaibia A (2020) Dynamic performance evaluation and improvement of PV energy generation
systems using moth flame optimization with combined fractional order PID and sliding mode
controller. Solar Energy 199:411–424. https://doi.org/10.1016/j.solener.2020.02.055
Experimental Validation
of a Neuro-Fuzzy Model of Open Circuit
Voltage and Maximum Power Voltage
Case of Amorphous Technology

Brahim Bouachrine, Mustapha Kourchi, Driss Yousfi, Kaoutar Dahmane,

Mhand Oubella, and Mohamed Ajaamoum

Abstract The objective of this work is to develop an adaptive neuro-fuzzy inference

system (ANFIS) for the determination of the open circuit voltage and the maximum
power voltage of a photovoltaic generator of the Amorphous/Microcrystalline type.
In order to evaluate the performance of the proposed ANFIS system, we compared
the results obtained using the ANFIS system to the results obtained using a system of
analytical equations developed by smoothing the experimental measurements. For
the experimental validation of our research work, we used an experimental database
from the station located at Green Energie Park in Bengrire Morocco. The compar-
ison results show that the proposed ANFIS model is more accurate than the analyt-
ical model and allows to better emulate the electrical characteristics of the studied
photovoltaic generator.

Keywords ANFIS model · Photovoltaic emulator · Analytical model ·

Amorphous/microcrystalline

1 Introduction

The photovoltaic (PV) emulator is an electronic system able to reproduce the current-
voltage output characteristics of photovoltaic generators. It facilitates laboratory
testing and operational evaluation of a photovoltaic system by applying the same
real-world conditions such as radiation and shading to achieve results similar to
reality [1]. The proper functioning of these emulators depends on the quality of

B. Bouachrine (B) · M. Kourchi · K. Dahmane · M. Oubella · M. Ajaamoum

LASIME, National School of Applied Sciences, Ibn Zohr University, Agadir, Morocco
e-mail: [email protected]
M. Kourchi
e-mail: [email protected]
D. Yousfi
Department of Electrical Engineering, National School of Applied Sciences, Mohammed First
University, Oujda, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1181
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_108
1182 B. Bouachrine et al.

the mathematical models used to generate photovoltaic characteristics. Indeed, the

choice of these models directly affects the static and dynamic performance of the
photovoltaic emulator [2, 2].
The aim of this work is to develop an adaptive neuro-fuzzy inference system
(ANFIS) for the determination of open-circuit voltage (V oc ) and maximum power
voltage (V mp ). This system should allow us to better emulate the electrical charac-
teristics of the studied photovoltaic generator. In order to evaluate the quality of the
proposed ANFIS model, the results obtained using the proposed ANFIS system will
be compared with the results generated by exploiting a system of analytical equa-
tions obtained by smoothing the experimental measurements. For the experimental
validation of our research work, we have used an experimental database from the
Green Energie Park in Bengrire. The technology used in this experiment is the amor-
phous/microcrystalline technology (case of the solar module SHARP NS-F135G5)
[4].
This paper is organized as follows: after the introduction, in the second part, we
present the studied mathematical model of the PV module. In this part, we also present
the systems of analytical equations developed for the calculation of the I(V ) char-
acteristic at the points of open circuit V oc , maximum power (V mp , I mp ), and short
circuit I sc . In the third section, we describe the adaptive neuro-fuzzy inference system
ANFIS elaborated to compute the V mp and V oc voltages which allows increasing the
accuracy of the photovoltaic model. In the fourth part, we will present the experimen-
tally validated results followed by discussions and comparisons. Finally, a conclusion
that encompasses the work done in this paper.

2 Proposed Photovoltaic Model

2.1 Mathematical Model for the Calculation

of the Characteristic I(V)

The proposed model is obtained from the semiconductors theory. It represents the
electrical behavior of a PV module by a mathematical model with a one exponential
[5]:
    
V
I = Isc 1 − C1 exp −1 (1)
C2 Voc

This model allows an explicit computation of the electrical characteristic I(V )

of the PV module regardless of the lighting and temperature conditions. The two
parameters C 1 and C 2 are determined as a function of the currents I sc , I mp and the
voltages V mp , V oc for a given illumination and temperature:
Experimental Validation of a Neuro-Fuzzy Model … 1183
   
Imp − Vmp
C1 = 1 − exp (2)
Isc C2 Voc
Vmp
Voc
−1
C2 =   (3)
I
log 1 − Imp
sc

2.2 Analytical Model for the Calculation of Currents Isc, Imp

For the computation of the currents Isc, Imp as a function of the illuminance G, and
the temperature T of the PV source we use the system of analytical Eqs. (4). The
system of equations is obtained by smoothing the experimental measurements using
the Levenberg-Marquardt algorithm [6, 7].
⎧  
⎨ Isc = Iscn −0.00534 + 0.9768 G + 0.0002441 (T − Tn )
 G n  (4)
⎩ Imp = Impn −0.005973 + 0.9712 G + 0.000476 (T − Tn )
Gn

With:
Tn : Temperature (°C) under standard conditions.
Gn : Sunlight (W/m2 ) under standard conditions.

2.3 Analytical Model for the Calculation of Voltages Vmp

and Voc

The system of analytical Eqs. (5) allows for the calculation of the voltages Vmp, Voc
as a function of the illumination G, and the temperature T of the PV source. This
system of equations is obtained by smoothing the experimental measurements using
the Levenberg-Marquardt algorithm [6].
⎧  
⎨ Voc = Vocn + 1.067 Vt log G − 0.3017 (T − Tn )
Gn  (5)
⎩ Vmp = Vmpn + 0.5284 Vt log G − 0.222 (T − Tn )
Gn

With:
Vocn : Open circuit voltage in (V) under standard conditions.
Vmpn : Voltage at maximum power in (V) in standard conditions.
V t = N s kT /q: Thermodynamic potential of Ns cells connected in series.
1184 B. Bouachrine et al.

3 ANFIS Model for Calculating Voltages Vmp et Voc

In order to improve the accuracy of the photovoltaic model, we have opted for the
neuro-fuzzy approach for the identification of Voc and Vmp. Neuro-fuzzy systems
that combine fuzzy logic and neural networks have proven their efficiency in a variety
of industrial problems. They are useful for the identification of non-linear, complex,
and difficult to model systems [7, 8].
We have consequently chosen to use adaptive neuro-fuzzy inference systems
(Anfis), which does not require a particular form of modeling. This is a Perceptron-
type neural network with 5 layers, where each layer corresponds to the realization
of one step of a fuzzy inference system (SIF) of the Takagi Sugeno type. A Matlab
code is developed for the generation of the ANFIS model according to the flowchart
in Fig. 1.
After several learning tests, we have come up with the simplest and most efficient
architecture of the ANFIS model illustrated in Fig. 2.
It has two inputs and a polynomial output of order 1, each input is represented by
two fuzzy sets of the Gaussian type. The final output is calculated as the average of
the outputs of each rule weighted by the degrees of activation.
Inputs x and y are expressed as a function of the irradiance G and temperature T
of the PV module according to the system of Eq. (6).

Fig. 1 The flowchart of the

developed code Database :
Inputs : x, y
Output : Voc or Vmp

Generating the initial SIF:

genfis1

Stopping Criterion:
Max. number of iterations
Tolerance

Learning SIF :
Anfis
Hybrid optimization

Validate the results

Experimental Validation of a Neuro-Fuzzy Model … 1185

x y
w1
A1 N
w2
x A2 N
Voc or Vmp
y B1 w3
N

w4
B N

Fig. 2 ANFIS architecture used


x= G
+ T − Tn
Gn   (6)
y = (T + 273.15) log G
(T +273.15) G n

4 Results and Experimental Validation

4.1 Experimental Bench

For the validation of the proposed ANFIS model, we used an experimental database
from Green Energy Park in Bengrire [6]. The database consists of experimental I(V)
electrical characteristics captured with a sampling pitch of 10 min along the day.
In this work, we used a set of measurement samples with 195 I(V) characteristics.
The PV technology studied is of the Amorphous/Microcrystalline type (case of the
SHARP solar module NS-F135G5 [3]). Its technical specifications are described in
Table 1.
To evaluate the accuracy of the estimated parameters and the quality of the
smoothing, we have calculated the coefficient of determination Rd and the root mean
square error RMSE [6]. To simulate the different models, we chose the Matlab 2016a
environment.
In order to estimate the speed of execution of each of the models studied, we
evaluated the average calculation time tCPU, necessary for the calculation of the
I(V) characteristics of the set of measures considered. For the processing carried out

Table 1 Electrical parameters of the solar modules used

PV Module Pmax Ns Isc Imp Vmp Voc Ksc Koc Kmp NOCT
(W) (A) (A) (V) (V) (%/°C) (°C)
SHARP 135 – 3.41 2.88 47 61.3 0.07%/°C −0.3%/°C −0.24%/°C –
NS-F135G5
Amorphe/
Microcrystalline
1186 B. Bouachrine et al.

during our investigations, we used a PC-type computer equipped with the Intel(R)
Core (TM) i5-4310U CPU @ 2.00 GHz, 2601 MHz, 2 core(s), 4 logic processor(s),
with a RAM memory capacity of 4 GB.

4.2 Results of the Calculation of the Voltages Voc and Vmp

In Fig. 3 the values predicted by the ANFIS model and the analytical model
(5) are compared with the corresponding experimental values of the Amor-
phous/Microcrystalline technology studied.

Amorphous Technology

Fig. 3. Comparison of simulated and measured voltages for the studied technologies.

Table 2 Comparison of the ANFIS model and the analytical model

Technology Parameter PV model with ANFIS PV model with analytic (5)
Rd RMSE Rd RMSE
SHARP NS-F135G5 Voc 0.9825 0.1709 0 1.4509
Amorphous Vmp 0.9670 0.2010 0.0390 1.0854
Experimental Validation of a Neuro-Fuzzy Model … 1187

Table 2 summarizes the calculated values of the determination coefficient Rd and

the root mean square error RMSE corresponding to the calculation of the voltages
Voc and Vmp by the ANFIS model and the analytical model (5).
Examination of Fig. 3 shows that the results obtained by the ANFIS model are
satisfying for all the measures, however, it can be seen that the Analytic model
is not consistent with much of the experimental data set. The different compar-
isons (see Table 2) confirm the performance and the good concordance between the
measurement and the modeling by the ANFIS neuro-fuzzy system. This shows us
the efficiency and identification capacity of this model.

4.3 Calculation Results for Characteristic I(V)

This part presents some comparison results of the experimental and simulated I(V)
characteristics obtained in the two cases where the voltages Voc and Vmp are
computed by the Analytical model (5) and by the ANFIS model.
Figure 4 shows an example of application results of the studied PV model for the
studied technology Microcrystalline amorphous, and for the following sunlight G
and temperature T values:
• G (W/m2 ) = [566, 600, 635, 669, 701]
• T (°C) = [36.1, 35.9, 35.8, 36.6, 36.9]
The Fig. 5 illustrate an example of application results of the studied PV model for
the studied technology Microcrystalline amorphous, and for the following irradiation
G and temperature T values:

Fig. 4 Experimental and simulated I(V) characteristics

1188 B. Bouachrine et al.

Fig. 5 Experimental and simulated I(V) characteristics

• G(W/m2 ) = [516, 564, 607, 648, 73]

• T (°C) = [19.4, 22.6, 23.2, 24.1, 20.3]
The Fig. 4 demonstrates that the first sample of measurements is consistent with
the values simulated by both the ANFIS and analytical models. Although, it can
be seen (Fig. 5) that the simulation results do not agree with the second sample of
measurements, especially in the proximity of the open circuit point. We note that the
use of the analytical model, for the amorphous microcrystalline technology, does not
make it possible to reproduce the experimental measurements. This is predictable
because the precision of the analytical model is not sufficient for the estimation of
the voltages Vmp and Voc for part of the experimental data set.

4.4 Results Summary

For the experimental database as a whole, results comparing the performance of the
models used in this study are grouped in Table 3.
Analysis of the results in Table 2 shows that the ANFIS model achieves a very
good trade-off in terms of speed of execution.

Table 3 Summary of
Technology Criteria PV model with PV model with
performance and computing
ANFIS analytic (5)
time tCPU
SHARP Rd 0.9974 0.9407
NS-F135G5 RMSE 0.0472 0.2247
Amorphous t CPU (s) 0.0316 0.0082
Experimental Validation of a Neuro-Fuzzy Model … 1189

5 Conclusion

During our investigations, we look for a model that reproduces the characteristics of
the PV panels as well as possible. To achieve this, we compared two types of models,
namely the neuro-fuzzy ANFIS model and an analytical model. For the calculation of
the Voc and Vmp voltages we successfully used the ANFIS neuro-fuzzy system. The
study results show that the proposed model is more accurate than the analytical model
and allows to better emulate the electrical characteristics of the studied photovoltaic
generator.

Acknowledgement This work is funded and supported by the Moroccan Research Institute for
Solar Energy and New Energies (IRESEN).

References

1. Chen C-C, Chang H-C, Kuo C-C, Lin C-C (2013) Programmable energy source emulator for
photovoltaic panels considering partial shadow effect. Energy 54:174–183
2. Ayop R, Tan CW (2017) A comprehensive review on photovoltaic emulator. Renew Sustain
Energy Rev 80:430–452
3. Ram JP, Manghani H, Pillai DS, Babu TS, Miyatake M, Rajasekar N (2018) Analysis on solar
PV emulators: a review. Renew Sustain Energy Rev 81:149–160
4. Fiche technique de 135W/ 125W NS-F135G5/NS-F125G5 Amorphous Silicon/Microcrystalline
Silicon SHARP MIDDLE EAST FZE (SMEF)
5. Rekioua D, Matagne E (2012) Optimization of photovoltaic power systems, modelization,
simulation and control. Springer, Heidelberg. ISBN 978-1-4471-2348-4
6. Idadoub H, Kourchi M, Ajaamoum M, Yousfi D, Rachdy A (2019) Comparison and experimental
validation of three photovoltaic models of four technology types. Int J Tech Phys Probl Eng
11(41):1–10
7. Shabaan S, El-Sebah MIA, Bekhit P (2018) Maximum power point tracking for photovoltaic
solar pump based on ANFIS tuning system. J Electr Syst Inf Technol 5:11–22
8. Kharb RK, Shimi SL, Chatterji S, Ansari MF (2014) Modeling of solar PV module and maximum
power point tracking using ANFIS. Renew Sustain Energy Rev 33:602–612
Simulation of Natural Ventilation
on Building with Solar Chimney Under
Climatic Conditions of Errachidia
Morocco Zone

Herouane Aboubakr, Thami Ait Taleb, and Mourad Taha Janan

Abstract In this work we intend to present a numerical simulation by the ANSYS

FLUENT software of a solar chimney used for natural ventilation, and therefore the
cooling of a cubic room, under the actual climatic conditions of Errachidia Morocco
zone.
The study focuses on the variation of the diameter of the chimney with a fixed
length, compared to a square window. The air velocity and turbulence of kinetic
energy in different zones of the room were simulated for the three considered cases
of the solar chimney.
The results of the study clearly show the important role of the solar chimney in
favoring natural ventilation in building, as well as the optimum width of the solar
chimney recommended in the climatic zone of Errachidia Morocco.

Keywords Natural ventilation · Solar chimney · Energy efficiency

Nomenclature

Ui Velocity magnitude (m/s)

X, Y, Z Cartesian coordinates (m)
ν Kinematic viscosity (m2 /s)
λ Thermal conductivity (W · m−1 · K−1 )
Cp Heat capacity (J · kg−1 · K−1 )
ρ Air density (kg/m3 )
k Pressure loss coefficient
g Gravitational acceleration (m/s2 )
T Temperature (°C)

H. Aboubakr (B) · T. A. Taleb

ERME, Department of Physics-Chemistry, Poly Disciplinary Faculty of Ouarzazate,
Ibn Zohr University, Agadir, Morocco
M. T. Janan
Advanced Technical Teachers Training School, Mohammed V University, Rabat, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1191
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_109
1192 H. Aboubakr et al.

P Pressure (Pa)
Pm Pressure in reference state of the fluid (ρ0 , T0 ) (Pa)
β Coefficient of thermal expansion (K− 1 )
K Thermal diffusivity of the fluid (m2 · s−1 )
δ Kronecker symbol
Z Height (m)

1 Introduction

The building sector is the largest energy consumer in the world with a 34% share,
ahead of industry (27%) and transports (28%). It is also responsible for about a
third of greenhouse gas emissions. The electricity consumed in the building sector
represent for 53% of the total electricity consumption in the world [1].
In Morocco the building sector has a share of about 36% of the country’s total
energy consumption, with 29% reserved for the residential sector and the rest for the
tertiary sector, with an increase of 41% in 8 years [2].
To reduce energy consumption and CO2 emissions in the building sector, several
countries are adopting increasingly strict thermal regulations. In France, for example,
by 2020, all new buildings must be positive energy [3]. The Thermal Regulation of
Buildings in Morocco (RTBM) aims essentially to improve the following thermal
performances [2]:
• Reduce the heating and air conditioning needs of buildings;
• Improve the comfort of non-air-conditioned buildings;
• Reduce the power of heating and air conditioning equipment to be installed;
• Encouraging architects, engineers and supervisors to use efficient thermal design
approaches to the building envelope;
• Provide builders, public decision-makers and funders with a tool to improve the
productivity of their investments;
• Assist in the realization of energy diagnoses of existing buildings.
To improve thermal comfort in buildings and reduce environmental impact, it’s
necessary to use passive solutions. So integrating natural ventilation into the building
can be an efficient way to reverse the national energy trend in the building.
According to ADEREE, the zoning map of Morocco includes six climatic zones
that are represented by the following cities:
Zone 1: Agadir;
Zone 2: Tangier;
Zone 3: Fez;
Zone 4: Ifrane;
Zone 5: Marrakech;
Zone 6: Errachidia.
Simulation of Natural Ventilation … 1193

This work presents a simulation study of the effect of the natural ventilation of a
room using a solar chimney with different widths under the climatic conditions of the
climatic zone of Errachidia Morocco. The overall goal is to simulate the air velocity
and turbulence of kinetic energy in each room position relative to each width of the
solar chimney, in order to determine the optimum width of the solar chimney to be
used in this climatic zone.

2 Review on Natural Ventilation with Solar Chimney

In a study presented by Sajjad et al. [4] the authors have simulated by ANSYS
software the effect of the solar chimney in the natural ventilation of buildings. They
examined the effect of the inclination of the chimney with different angles (30°,
45°, 60°, 90°) and different widths (0.2 m, 0.4 m, 0.6 m, 0.8 m) and the height of the
chimney. The results of this study show that the optimum chimney width to maximize
natural ventilation should be 0.4 m and the length should be greater than 2 m or equal
to half of the length of the room, this for different inclinations except for 90°, Thus
the width of the window must be more than 1 m.
By their study on the thermo-aeraulic phenomena in a solar chimney, Saifi et al.
[5] showed that the temperature gradient between the absorber and the glass varies
according to the incident solar flux. The design adopted has resulted in relatively
high air flows at the outlet, which are favorable for use in natural ventilation. The
variation in the thickness of the air gap (chimney width) plays a very important role
and significantly increases the volume flow rate of the air. For optimal thermal draft,
the angle of inclination must be 45°.
Kumar et al. [6] carried out a comparative study between a vertical and horizontal
solar chimney, their results show that vertical chimney enhanced ventilation rate
much better in comparison to Horizontal chimney as much as 275% enhancement,
and vertical Solar Chimney enhanced air flow stream velocity up to 22 times.
Sudheer et al. [7] studied the dynamic and thermal behaviour inside a solar
chimney for three different chimney widths (0.1 m, 0.2 m, 0.3 m). The study showed
that for a width of 0.1 m, the air velocity and temperature inside the chimney is
higher than the other two cases. Thus in case 1 (0.3 m), some amount of stagnant
air is observed at the air inlet near the absorber. By decreasing the air gap from
0.3 m to 0.1 m nearly 44.18% of outlet velocity increase has been observed. So they
concluded that when the chimney width increases there is a decrease in air velocity.
In their study on passive cooling by a solar chimney in southern Algeria,
Belfuguais et al. [8] determined that the solar ray is the most important param-
eter influencing solar chimney performance. Thus an optimal design between the
chimney width and the absorber must be taken in consideration to increase the air
flow rate in the room.
Charvat et al. [9] simultaneously installed two solar chimneys: one with an
absorber to store heat, the other without absorber, to compare their performances. The
study showed that the air velocity inside the solar chimney containing the absorber
1194 H. Aboubakr et al.

is 25% higher than that of the chimney without absorber this is because the absorber
gives up its heat to overheat the air in order to escape more quickly to the outside.
Kasim et al. [10] investigated the effect of different opening positions on natural
ventilation. The study is based on nine different positions of two openings located on
two opposite walls, the result of this study shows that the air flow rate is maximum
when both openings are located at the top of the air flow level.
Mekkawi et al. [11] simulated natural ventilation through a solar chimney in
Egypt. Their study showed that the installation of the solar chimney of size of (0.4
× 1.5 m) can decrease the temperature from 29.45 °C to 28.64 °C in the room and
increase the air velocity with 50%.
Long Shi [12] carried out a study of the impacts of wind on solar chimney perfor-
mance in a building, his study shows that a higher wind velocity does not represent
a better solar chimney performance, and the effects from the window area keep
decreasing under a higher wind velocity.
Based on the results of all these studies, it can be concluded that there are several
parameters that can affect the air velocity at the outlet of the solar chimney, such as
the intensity of the solar radiation, the width of the chimney, the type of glass used
in the chimney, the type of absorber used, the entrance surface of the chimney, the
high of the chimney and the inclination of the solar chimney.

3 Methodology

3.1 Principle

The principle of operation of the solar chimney, as schematized in Fig. 1, is to increase

the air temperature in the chimney so that it escapes to the outward. Under the effect
of the density difference, the hot air rises upwards unlike the cold air, so to create this

Fig. 1 Diagram of the solar chimney in a room

Simulation of Natural Ventilation … 1195

density difference we use a chimney with glass faces. The glass allows transmitting
the solar rays to the air that passes from the chimney, the air heats up and rejects to the
outside, the air of the room replaces the air of the chimney and the cycle continues.
With this method the room will refresh in a natural and continuous way and without
any energy consumption or noise, as opposed to mechanical cooling systems.

3.2 Physical Model

The simulations were performed using ANSYS FLUENT commercial software.

The first step on Ansys fluent is the creation of the geometry, and then generates
the mesh and finally the configuration for which we choose the model, the materials,
the boundary conditions, and the type of solution, then start of calculation.
The physical model considered is shown in Fig. 2. It is a room with sides equal
to 4 m, with a solar chimney of 1 m of high and variable width (0.3 m, 0.4 m, and
0.5 m). The mesh model used is shown in Fig. 3.
For the simplification of the problem, the following assumptions are taken a count:
• The fluid is Newtonian and incompressible.
• The flow is unsteady.
• The inlet pressure equal to the atmospheric pressure
• The outside temperature is constant at 313 k.
• The air flow is turbulent.
• Normal direct solar irradiation is 1423 W/m2

Fig. 2 Physical model of geometry

1196 H. Aboubakr et al.

Fig. 3 Mesh of the model

3.3 Mathematical Model

• Conservation of mass (Continuity):

∂ ui
=0 (1)
∂ xi

Where ui are velocity components in x, y, z directions

• Momentum equation (Navier – Stokes):
Considering the linear variation of the density of the Boussinesq approximation,
we obtain the following momentum equation:
    
∂ ui ∂ ui u j 1 ∂ pm ∂ ∂ ui ∂u j
+ =− + v + + gβ(T − T 0 )δ i z (2)
∂t ∂x j ρ0 ∂ x i ∂x j ∂x j ∂ xi

With:
• Pm = p + ρ0 gz: the pressure corresponding to the reference state of the fluid (ρ0 ,
T0 ), ν the kinematic viscosity and g the gravitational field.
β: the coefficient of thermal expansion at constant pressure. It describes the change
of volume at constant pressure as a function of temperature.
Energy equation:
This equation is deduced from the first principle of thermodynamics by neglecting
the effect of pressure in front of temperature variation. It is then expressed by:
Simulation of Natural Ventilation … 1197
    
∂(T ) ∂ u j T ∂ ∂T
+ = k (3)
∂t ∂x j ∂x j ∂x j

Where:
k = ρCλ is the thermal diffusivity of the fluid (m2 ·s−1 ), Cp is the heat capacity
p

(J.kg−1 ·K−1 ) and λ is the thermal conductivity (W·m−1 ·K−1 ) [13].

In this study, we used a tetrahedron-type mesh with a high number of mesh nodes,
access to 98,344 nodes, as it is shown in Fig. 2. After having carried out several
attempts and mesh tests in order to choose the optimal mesh which should give a
good precision on the speed of the air and on the flow of air, we opted for an optimal
choice which is composed of 556,478 elements.

4 Results and Discussion

Air velocities and turbulence of kinetic energy in different plane of the room for the
0.3 m width of the solar chimney (Figs. 4, 5, 6 and 7).
• Air velocities and turbulence of kinetic energy in different plane of the room for
the 0.4 m width of the solar chimney.
• Air velocities and turbulence of kinetic energy in different plane of the room for
the 0.5 m width of the solar chimney
• The results of the study show that the air velocity in the room is higher in the
chimney of 0.5 m of width (Fig. 14), 2,74 m/s in the middle of the plane taken
at 0.2 m (Fig. 13), and this velocity decreases as we move up and down the level

Fig. 4 Air velocity in the 0.15 m plane

1198 H. Aboubakr et al.

Fig. 5 Air velocity in the 2 m plane

Fig. 6 Air velocity in the room

of the window plane. The lowest velocity is for the chimney of 0.4 m in width,
(Fig. 10), 2.12 m/s in the middle of the plane taken at 0.2 m, (Fig. 9). For the
distribution of the air velocity in different zones of the room, it is noted that it is
more important in the case of the chimney of 0.5 m of width (Fig. 14). Thus the
ventilation is present in all zones of the room for the three cases studied (Figs. 6,
10 and 14), this clearly shows the importance of the solar chimney in the natural
ventilation. The results also show that the air velocity in the room is not uniform,
it is higher in the middle of the room for the two chimneys with a width of 0.4 m
(Fig. 9) and 0.5 m (Fig. 14), while for the chimney of 0.3 m of width the air
velocity is higher at the solar chimney (Fig. 4), this is because the air heats up
Simulation of Natural Ventilation … 1199

Fig. 7 Turbulence of kinetic energy in the 2 m plane

Fig. 8 Air velocity in the 0.2 m plane

faster and flows out into this chimney. For the turbulence of the kinetic energy it
is higher in the plane taken at 2 m according to the variations of the wind speed in
the room for the three chimneys, (Figs. 7, 11 and 15); and it decreases by going
up and down of the window plane. In all three cases, the absence of a reverse
flow of air in the solar chimney was reported; this refers to the phenomenon of
the difference in density between inside and outside air (Figs. 4, 8 and 12).
1200 H. Aboubakr et al.

Fig. 9 Air velocity in the 2 m plane

Fig. 10 Air velocity in the room

Simulation of Natural Ventilation … 1201

Fig. 11 Turbulence of kinetic energy in the 2 m plane

Fig. 12 Air velocity in the 0.25 plane

1202 H. Aboubakr et al.

Fig. 13 Air velocity in the 2 m plane

Fig. 14 Air velocity in the room

Simulation of Natural Ventilation … 1203

Fig. 15 Turbulence of kinetic energy in the 2 m plane

5 Conclusion

This work presents a numerical modeling and simulation of natural ventilation by

a solar chimney. The three-dimensional simulation of a room with a solar chimney
was carried out using the Ansys Fluent commercial software.
The results of the study show the importance of the solar chimney in natural
ventilation by increasing the air velocity in the room. For the three types of the
solar chimney simulated, significant air velocities are observed in the different room
positions and in the chimney. Thus there is a large distribution of air through in
all zones of the room, which allows for continuous passive cooling. The air flow is
normal, as it passes through the opening (inlet) and out through the chimney (outlet),
no reverse air flow through the chimney. For the climatic conditions of Errachidia
Morocco zone, according to this study, the optimum chimney to be installed in the
buildings is that of 0.5 m of width.

References

1. International Energy Agency (2013) Key world energy statistics

2. ADEREE (2011) Les éléments techniques du projet de la réglementation thermique du bâtiment
3. Faggianelli GA (2014) Rafraichissement par la ventilation naturelle traversante des bâtiments
en climat méditerranéen. Thèse de doctorat en Génie des procédés. Université PascalPaoli
4. Sajjad A, Badshah S, Chohan GY (2014) Modeling and simulation of natural ventilation of
building using solar chimney. World Appl Sci J 32(5):741–746
5. Saifi N, Settou N, Dokkar B, Negrou B, Chennouf N (2012) Experimental study and simulation
of air flow in solar chimneys. Energy Procedia 18:1289–1298
1204 H. Aboubakr et al.

6. Kumar J, Raj A, Sharma HM (2017) Enhancement of natural ventilation using SolarChimney;

a numerical investigation. Int J Adv Eng Res Sci (IJAERS) 4(3):237105
7. Sudheer V, Sridhar B (2017) Computational fluid dynamics analysis of a building ventilation
solar chimney. Int J Adv Eng Res Sci (IJAERS) 2(2)
8. Belfuguais B, Larbi S (2011) Passive ventilation system analysis using solar chimney in South
of Algeria. Int J Mech Aerosp Ind Mech Manufact Eng 5(10)
9. Charvat P, Jicha M, Stetina J (2004) Solar chimneys for ventilation and passive cooling. World
Renewable Energy Congress VIII, Denver, Colorado, USA
10. Kasim NFM, Zaki SA, Ali MSM, Ikegaya N, Razak AA (2016) Computational study on the
influence of different opening position on wind- induced natural ventilation in urban building
of cubical array. Procedia Eng 169:256–263
11. Mekkawi G, Elgendy R (2016) Solar chimney for enhanced natural ventilation based on CFD-
simulation for a housing prototype in Alexandria, Egypt. Int J Advances Mech Civil Eng 3(5).
ISSN 2394-2827
12. Long S (2019) Impacts of wind on solar chimney performance in a building. Energy 185:55–67
13. Brangeon B (2012) Contribution à l’étude numérique de la ventilation naturelle dans des cavités
ouvertes par la simulation des grandes échelles, Application au rafraichissement passif des
bâtiments. PhD thesis in Mechanics and Building Sciences, University of Reunion
Investigating the Effect of the Air Gap
of a Solar Air Heater Intended
for an Indirect Drying System

Mourad Salhi, Dounia Chaatouf, Benyounes Raillani, Samir Amraqui,

and Ahmed Mezrhab

Abstract As the world’s population increases, the demand for vegetarian products
increases, therefore, in order to preserve the food products throughout the year, it
was necessary to find a way that can preserve almost all the nutritional and vitamin
value. In this term, we find that the most efficient method is indirect solar drying.
This latter is used for drying, because of its many advantages, such as being covered
from direct sunlight, which protects the enzymes in the product, while preserving the
original color and reduce the drying time. However, with all these advantages, this
method had problems with air uniformity along the trays. Therefore, in this work,
we will study the dimensional parameters of the system in order to improve the air
and temperature distribution inside the drying chamber, and to be more precise the
study is mainly based on the optimization of the air inlet which is the air gap of the
solar air heater for a newly designed drying chamber. The ANSYS software is used to
simulate the system, which usually consists of a drying chamber and a solar collector
under the meteorological data of Oujda city. As a result, we have found that the air
distribution in the drying chamber depends directly on the size of the air gap of the
solar air heater.

Keywords Inlet · Solar air heater · Solar dryer

1 Introduction

Solar energy is one of the most important clean and inexhaustible energies on earth,
as it represents an energy that can be exploited without limit as long as the sun is
present. This energy can be converted into different forms that can be used in daily
life such as electricity, heating, and cooling. Since there is an enormous increase in
the world’s population it becomes necessary to satisfy the need for energy and food,
therefore solar energy has been exploited in other systems related to the conservation
and storage of the alimentary product, such as solar drying, Which is an ancient and

M. Salhi · D. Chaatouf · B. Raillani · S. Amraqui (B) · A. Mezrhab

Laboratory of Mechanics and Energetics, Faculty of Sciences, Mohammed First University,
60000 Oujda, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1205
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_110
1206 M. Salhi et al.

widely used method of food conservation, as a preventive against the growth of

microorganisms and bacteria and fermentation that lead to food spoilage [1].
Currently, different drying methods are available, such as conventional drying,
which is a traditional method that has been used for a very long time to preserve
the alimentary product, where the products are dried by direct exposure to both sun
and wind. This drying method is not that effective because of the time required
to dry the product and because of the disadvantage of direct exposure to weather
conditions, especially aerosol particles, rain, and insects. This is why other drying
systems were invented, intending to improve the quality of the dried product through
a clean and safe method. In this term, there are three main and best-known ways to
rely on the energy of the sun as a source for the drying process, direct, indirect, and
mixed mode. The indirect solar dryer is one of the most efficient ways to dry the
product and improves its quality by avoiding some of the disadvantages of direct and
traditional solar dryers. This system is mainly composed of a solar collector where
the air is heated by direct contact with the absorber, as well as by the greenhouse
effect created by the glass. Afterward, the air rises into the drying chamber to dry
the products that are distributed on the trays before exiting from the outlet.
In recent years, several works and studies have been carried out to improve the
performance of the indirect solar dryer. Among these researches is the use of a system
composed of two double-pass solar air heaters [2] the result that has been found has
proved that the efficiency of this system in terms of color, flavor and drying time
of the product, compared to an open solar drying system. In another study, we have
found that they have completely changed the previous concept of air heating, where
they use a heat exchanger inside the drying chamber [3], and as a result, they showed
that it can improve and increase the temperature of the drying chamber. Research
has not only focused on the solar air heater. We also found researchers who put
their expertise into developing geometry to distribute the air evenly throughout the
chamber and to improve the flow of air in and out, by studying the flow path. In this
regard, [4] have studied the chimney effect by inducing a thermosiphon to improve
the efficiency of the drying, via accelerating the process of evacuating the humid air
from the drying chamber, the results show that the humidity level decreases from
86% to 8.12% in 9 h and 20 min.
All this research studies are focused on general studies and does not take into
consideration the optimization of certain internal parameters to improve the effi-
ciency of the drying system. For this purpose and to improve the air and temperature
distribution inside the drying chamber, we will study the dimensional parameters
of the air gap of the solar air heater which is the inlet of the drying chamber for a
newly designed drying chamber using ANSYS FLUENT under the meteorological
conditions of Oujda city (eastern of Morocco).
Investigating the Effect of the Air Gap of a Solar Air Heater … 1207

Air outlet

The door of
Solar radiation
the chamber

Air inlet

Fig. 1 Schematic diagram of the solar dryer studied

2 Methodology

2.1 Geometry and Basic Governing Equations

The indirect solar dryer studied shown in Fig. 1 composed of a 1.5 m2 solar air heater
that composed of an aluminum absorber painted in black and a simple glazing, the
chamber is actually composed of two parts the first one contains three trays where the
products are distributed and the second one is equipped with solar air heater to make
the air pass horizontally in the trays, so the air distribution will be more uniform [5].
The governing equations are based on the Navier-Stokes equations, while the
turbulent pattern inside the solar dryer is modeled using the K-epsilon model,
which characterized with two equations, the turbulence kinetic energy k (1) and
the dissipation rate of turbulent kinetic ε (2).

∂k ∂(ku) ∂(kv) μt ∂ 2 k ∂ 2k
ρ( + + ) = (μ + )( 2 + 2 ) + G k − ρε (1)
∂t ∂x ∂y σk ∂ x ∂y

∂ε ∂(εu) ∂(εv) μt ∂ 2 ε ∂ 2ε ε
ρ( + + ) = (μ + )( 2 + 2 ) + (c1 G k − c2 ρε) (2)
∂t ∂x ∂y σε ∂ x ∂y k

2.2 Climate Data and Boundary Conditions

The simulation was made under the metrological conditions of Oujda city, the data
used in the simulation where collected from the metrological station installed in
university Mohammed first in Oujda presented in figure below. The climate data of
the chosen day which are the ambient temperature and the solar radiation where
transformed into equations as presented in Eqs. (3) and (4) before inserting it in
ANSYS FLUENT as a UDF subroutine (Fig. 2).
1208 M. Salhi et al.

Fig. 2 Meteorological station used to collect the data

π
Tamb (t) = 25 + 6 cos( (t − 14)) (3)
12
t −6
G sun (t) = 965 sin( π ), 6 < t < 20 (4)
14
The studied solar dryer is working under naturel convection and for that the
bossinessq approximation is used, and the inlet as well the outlet where pressure
type, regarding the walls we have convective heat losses in all walls including the
solar air heater while in the glass cover we have a mixed mode of convective and
radiation losses.
The trays are modeled as a porous medium with 50% porosity, in order to model
the airflow resistance created by the trays.

2.3 Validation Result

In order to make sure that the simulations made by ANSYS FLUENT is right, two
validations with experimental and numerical results where made the first article
[6] studied a cabinet dryer equipped with a chimney and a solar air heater, the
outlet temperature of this latter was investigated for different lengths, the comparison
between our results and the authors is presented in Fig. 3.a, the other validation
concerns a wooden chamber equipped with a solar chimney [7] the air velocity at
the outlet is presented in Fig. 3.b in those two validations our results showed a better
values than the numerical ones found by the same authors.
Investigating the Effect of the Air Gap of a Solar Air Heater … 1209

Fig. 3 Comparison between our results and the reference [] (a) and [] (b)

3 Results

Before going ahead to the simulations, a mesh independency test is necessary to

make sure that the results are independent with the mesh density, and for that, four
different grids, 5532, 12130, 16606, and 23820 triangular cells were studied, and the
results show that a mesh grid with 12130 cells is quite enough to study the geometry.
In this part, we studied the effect the inlet size of the drying chamber, which is the
air gap of the solar collector on the air behavior inside the drying chamber. For this
purpose, we have changed the size of the inlet into different values and analyzed its
effect on the temperature distribution and the air velocity inside the drying chamber.
The inlet size values studied were ranged between 1 cm to 14 cm, which is
presented in Fig. 4 in term of the mean air velocity and temperature at the inlet of
the chamber, specifically, at the outlet of the solar collector. The results show that in
terms of temperature, the optimal size value of the inlet, where we get the highest
temperature which is 373 k is 3 cm. From this value, the temperature started to
decrease with the increase of the inlet size.
In terms of air velocity, the graph shows that velocity value increases with the
increase in the inlet size until reaching the highest value which is 0.1 m/s at 4 cm
and then it started to decrease with the increase of inlet size.
Among all the studied inlet sizes, the three and four centimeters sizes showed
a better performance for the air temperature and velocity respectively. But, finding
the optimal air gap size to get a high temperature and high velocity at the inlet of
our drying system is not enough in this case, because we are more interested in the
distribution of air flow in the drying trays where the products are placed, That’s why
we have plotted in Fig. 5, the curves that shows the average air temperature in the
drying chamber and the mass flow rate at the inlet of the chamber for different air
gap of the solar air heater. And they both increase with the increase in the air gap,
the temperature increase in the chamber due to the increase in the mass flow rate
which increased the heat gained by the flowing air through the solar air heater but
the temperature remains almost stable beyond 7 cm.
1210 M. Salhi et al.

Fig. 4 The mean air velocity and temperature at the inlet of the chamber for different inlet size at
1:00 pm

Fig. 5. a The average air temperature in the drying chamber, and b the mass flow rate at the inlet
of the chamber in function of the air gap size at 1:00 pm.

To see more clearly the air velocity and temperature along the three trays, we
have presented in Fig. 6 the average air velocity and temperature in the trays for
different air gap sizes, and the results show that the air flow distribution along the
trays is better for an inlet size equal to 7 cm. although for this size, the velocity and
temperature at the inlet is low compared to 4 cm. This is due to the amount of the
airflow entering the drying chamber which becomes high for the inlet size 7 cm.
From Figs. 4, 5 and 6, it can be concluded that the increase of the inlet size
provokes an increase in the air mass flowrate as a results the efficiency of the solar
drying system.
Investigating the Effect of the Air Gap of a Solar Air Heater … 1211

Fig. 6. The mean air velocity (a) and temperature (b) in the three trays for an inlet size from 3 to
7 cm at 1:00 pm.

4 Conclusion

Food preservation is a necessity nowadays, which is why it has become a priority to

develop ecological systems that offer a very good drying quality. For this reason, an
indirect solar dryer represents the best choice; in this study we investigate the effect
of the air gap of solar air heater which is the inlet of the chamber as well. In order to
improve the distribution of the airflow along the trays by making it more uniform in
terms of temperature and velocity. The results obtained showed that the size of the
air gap of solar air heater that gives a high air velocity and temperature at the outlet
is 4 and 3 cm respectively. In terms of the drying system, the mass flow rate and the
temperature inside the drying chamber increase with the increase in the air gap but
it remains almost stable beyond 7 cm.

References

1. Bourdoux S, Li D, Rajkovic A, Devlieghere F, Uyttendaele M (2016) Performance of drying

technologies to ensure microbial safety of dried fruits and vegetables. Compr Rev Food Sci Food
Saf 15(6):1056–1066
2. Sharma A, Sharma N (2012) Construction and performance analysis of an indirect solar dryer
integrated with solar air heater. Procedia Eng 38:3260–3269
3. Reddy RM, Reddy ES, Maheswari CU, Reddy KK (June 2018) CFD and experimental analysis
of solar crop dryer with waste heat recovery system of exhaust gas from diesel engine. In: IOP
conference series: earth and environmental science, vol 164, no 1. IOP Publishing, p 012010
4. Musembi MN, Kiptoo KS, Yuichi N (2016) Design and analysis of solar dryer for mid-latitude
region. Energy Procedia 100(1):98–110
5. Chaatouf D, Salhi M, Raillani B, Amraqui S, Mezrhab A (2021) Applying CFD for the opti-
mization of the drying chamber of an indirect solar dryer. In: Proceedings of the 6th international
conference on wireless technologies, embedded and intelligent systems (wits-2020)
1212 M. Salhi et al.

6. Ghaffari A, Mehdi R (2015) Modeling and improving the performance of cabinet solar dryer
using computational fluid dynamics. Int J Food Eng 11(2):157–172
7. Jyotirmay M, Sanjay MA (2006) Summer performance of inclined roof solar chimney for natural
ventilation. Energy Build 38:1156–1163
Experimental Investigation of Efficiency
and Dynamic Losses for a Constructed
Solar Boost Converter

Weam El Merrassi, Abdelouahed Abounada, and Mohamed Ramzi

Abstract Power converters are used in numerous industrial applications; especially

for renewable energy, they are vital. However, the use of a conventional power
converter involves high switching losses, high Electro-Magnetic Interference (EMI),
and increase junction temperature of the switches. Therefore, to overwhelm the previ-
ously mentioned drawbacks, soft techniques are required to enhance the performance
of the power converters. Among several varieties of snubber cells, passive snubbers
are epitome due to their easiness and vigorousness. In this paper, a soft switching
solar DC-DC static Boost converter is presented using RID and RCD passive snub-
bers. The proposed snubber cell intends not only to provide zero soft-switching and
stabilize the temperature of the main switch but also to enhance the power conver-
sion efficiency of the converter. The effectiveness of the model is evidenced via
experimental results using a 1.6 kW prototype to validate the proposal.

Keywords Boost converter · Renewable energy sources · Passive Snubber · Zero

current switching (ZCS) · Zero voltage switching (ZVS)

1 Introduction

Up to now, power converters have been ushering in a new kind of industrial revolution.
They have been extensively indispensable in manifold industrial applications, notably
for the progress of renewable energy resources [1, 2].
Therefore, the DC-DC converter has been spotlighted in multiples studies for
diverse intentions. It provides good operability, simple and affordable design
to implement. However, the widespread shortcoming of these converters is the
switching losses that arise in the hard switching operating mode [3]. Above and
beyond, higher switching frequency gets the higher switching losses along with elec-
tromagnetic interface. Yet, renewable energy applications are one of the principal

W. El Merrassi (B) · A. Abounada · M. Ramzi

Laboratory of Automatic, Energy Conversion and Microelectronics (LACEM),
Department of Electrical Engineering, Faculty of Science and Technology, University of Sultan
Moulay Slimane, Beni Mellal, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1213
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_111
1214 W. El Merrassi et al.

applications that rely on the converter, however these class of resources known by
their significantly lower yield. Accordingly, this type of applications involve specific
design criteria to be adopted such as diminished EMI, minimized power losses,
and stabilized junction temperature. In order to break this bottleneck, and enhance
the efficiency of the converter, soft switching strategies are incited as an alterna-
tive to hard-switching operating mode. Several snubbers have indorsed for different
converter to eschew the high losses, and limited the rate change in voltage/current
and overvoltage during the turn on/off. The snubbers presented in the literature can be
classified into active and passive snubbers. As presented in [4, 5] the active snubber
are used to reduce the turn-on switching losses. The auxiliary circuit proposed for
the converters use an extra switch that includes an intermediate control circuit. Some
of the problems that persist for the active snubber is the synchronicity issue of the
control for the different switches involved in the converter, which rise the complexity
of the control procedure, the extend size of the circuit, and the cost [6]. There-
fore, the passive snubber have been a worthwhile solution that offer a low cost,
easiness of control and implementation. Yet, the passive snubber presented in the
literature has unveiled some downsides highlighted under the inability to promise a
perfect turn-off zero voltage switching, regardless of the switching turn-on at zero
current [7]. Accordingly in [2, 6, 8], the both models was not able to eradicate
the switching losses. In addition, in [9–11] the increase number of passif compo-
nents in snubber cells engender extra conduction losses, intricated applications, and
expensive. Furthermore, certain passive snubbers are not qualified for the family of
fundamental DC-DC converter, they are notably provided for their own converter
structure [12]. Consequently, the original idea for proposing new snubbers and high-
performant solar converters primarily originates from fixing the shortcomings and
improving the efficiency of previous approaches snubbers and converters. Conse-
quently, the snubbers circuits and soft-switching converter ought to achieve an equi-
librium amid the reliability and easiness of integration. In this treatise, the dynamical
losses for a conventional solar converter is explored, to promote the system’s perfor-
mance by lessening the total losses of the solar DC-DC boost converter. A novel
approach to design RID and RCD snubber cells had been proposed. The snubber
assured a zero voltage-switching in the switch-on periods side to a zero current-
switching in switch-off periods, which affords a perfect soft-switching for the main
interruptor in the both transition periods. The proposed snubber cells entail merely
few passive components; it has a simple structure with a minor size, easiness of drive
and low cost. The structure of this paper presented as follow. In the first hand, the
fundamental operational principles of the suggested snubber are proved in depth,
furthermore the experimental test is provided to verify the results’s efficiency above
to a comparison of the effictiveness of the system face to other proposed models in
literature.
Experimental Investigation of Efficiency and Dynamic Losses ... 1215

2 Power Losses Analysis in Boost Converter

2.1 Conduction Losses

The power losses in nearly all converters are leading out of two major bases: Conduc-
tion losses and dynamic losses. The conduction losses in a boost converter take place
when the switch is entirely in conduction. The resistance elements persisted in the
different component of the boost converter such as diode, inductors, capacitors, and
switch are the origin of these losses. The conduction losses is outlined as [13]:

Pcond = Rgen IL.eff

2
+ Resl IL.eff
2
+ Ron. ITr.eff
2
+ Rd Id.e
2
f f + Vd Id.av + R esr.c IC.e f f (1)
2

Where, Rgen is the generator resistor, Rgsl is the induction resistor, Ron is the switch
resistor, Rd is the diode resistor, Resrc is the capacitor resitor, IL.eff is the effective
current that cross the generator and inductance, I.Tr.eff is the effective current that cross
the main switch, I.d.eff is the effective current that cross the diode, I.d.av is the average
current of the diode, I.c.eff is the effective current that cross the output capacitor of
the converter.

2.2 Dynamic Losses

Dynamic losses in the boost converter include three sources of losses represented
as: choke core, Switch device, and diode. However, the losses in Switch depends
majorlyon losses in gate, output capacitance and switch losses during transition
period. The mathematical expression of these losses are introduced below:

Psg = Csin Vg2 f sw (2)

1
Psc = Cs V 2 f sw (3)
2 out ds

Psw = A.(tr .Imax + t f .Imin ).Vout f sw (4)

Pdiode = Vd .(trr .Imin + Q r d ). f sw (5)

Pcor = a.B j Vcor f sw

i
(6)

Where, Psg , Psc are the losses in the switch, Psw is losses in switching periods,
Pdiode is the losses produced by the diode, Pcor is the losses produced by the choke
core.
1216 W. El Merrassi et al.

Fig. 1 The voltage and

current variation during hard
switching transients

The total dynamic losses are equavilent to the summation of individual switching
losses of converter elements and they can be described as follows:

Pdynamic = Pcor + Pdiode + Psw + Psc + Psg (7)

In spite of everything, the power losses in the core choke, diode, gate control and
the output capacitance can be marginal against the losses occur the switching periods.
Which link directly the dynamic losses in the converter to the switching losses in
the switch. Therefore,in the boost converter, the Semiconductors represent the most
element presentinglosses in the circuit. Theswitchon/off periodsoftheswitch is in the
order of a few tens of nanoseconds to multifold microseconds. At these switching
transitions moment, very large power losses can occur in the semiconductor device.
Despite, the switching time of semiconductor elements is very brief, the average
power loss can be substantive. Dynamic losses are less depending on power load but
instantly depending on switching frequency as shown in the Eq. (2)–(4). As Fig. 1
unveils, the dynamic losses in the switch are predominant compared to the conduction
losses. For this reason, reducing the switch losses will gain less power loss by the
circuit.

3 Proposed Converter System Configuration

The suggested passive snubber is served to conserve the main switch’s high perfor-
mance of the converter as shown in Fig. 2. The snubber contains two cells, the first
one RID is composed with a diode and resistor with a parallel inductor implemented
in series to the switch, while the second RCD cell comprises a parallel diode and
resistor in series with a capacitor [13].
The capacitor C and the inductor L delay the time at which the switch reaches
the final value of voltage and current. In return, the value of the capacitor and the
inductor usually optimized in order to minimize the power dissipated in the switch.
However, it may not be the best value of the snubber elements from the reliability
Experimental Investigation of Efficiency and Dynamic Losses ... 1217

Fig. 2 Solar converter with the proposed snubber

point of view. The stress in switch reduced but the losses persist. Therefore, in the
proposed snubber circuit, the elements are selected based on maximum reliability
considerations for the set of the switch and its snubber. To ensure as results a soft-
switching and a minimum power loss for the switch. As shown in Fig. 2b, in this
circuit when the interruptor proceeds turning-off, load current charges the capacitor
C up to source voltage V. Consequently, it delays the time at which the capacitor
(alsoswitch) voltage waveform reaches the final value V, the snubber capacitor allows
the voltage to rise and reach full value V after the switch current falls to zero. Besides,
the same scenario achieved by the snubber inductor that allows the current to rise
and reach full value I after the switch voltage falls to zero.

4 Hardware Implementation Results

The performance characteristics of both the conventional and the proposed boost
converter are verified by the prototype of 1.6 kw and 20 Khz switching operation
using the same components and the same conditions [13]. The prototype itself is
summarized, with the aforementioned design considerations. The converter operated
with a source simulating PV as a source.

4.1 Galvanic-Isolation Circuit

The signal gate is given to the devices through an intermediate interference circuit
that provides isolation from the hardware power circuit and the control circuit, as
shown in Fig. 3(b). This circuit integrated a 6N137 Optocoupler to isolate, and an
IR2112 Driver, which had as role amplifying the output signals, delivered from the
control card. From a low voltage to reach the minimum gate threshold voltage of
1218 W. El Merrassi et al.

Fig. 3 Galvanic-isolation circuit a Hardware set-up of the boost converter with the galvanic-
isolation circuit. b the switch signal in the output of the 6N137 Optocoupler and IR2112 driver

the MOSFET. The Fig. 3(b) Channel 2 Shows the digital output waveform of gate
pulse for the boost converter Mosfet switch generated by the DSP, which offers a
signal variation of 0–3.3 V, the switching frequency is 20 kHz equals to the imposed
frequency 20 kHz and the duty cycle is about 50%. The output signal of Optocoupler
is amplified from 3 V to 12 V and the others parameters of switching frequency and
duty cycle remain the same as before as it can be seen in channel 1 in the Fig. 3(b)
that illustrates the output of IR2112 Driver for boost converter [13].

4.2 Experimental Results

In the proposed implementation, we tested our prototype using a low power supply.
Prior to experiments, the Fig. 4, 5 and 6 present the experimental results of the
active switch’s performance for a solar boost converter operated in both mode hard-
switching and soft-switching conditions. As illustrated in Fig. 4(a)–(c) the drain-
source voltage in the active switch for the conventional circuit endows a high voltage
and current values along with high spike during the device’s turn-off. And, that
is linked to the fast changing of the voltage dv/dt. The proposed circuit achieved
a soft switching with a zero voltage switch at turn-off. Due to the control of Rs
Cs that improved theperformance significantly by reducing the changingrate dv/dt,
which raised the switch’s voltage gradually from0 toV0, as shows in Fig. 4(b)–(d).
The overlapped area between the voltage and current switch is almost omitted. As
results, the main switch in the proposed circuit had zero switching point at the turn off
and turn on. Consequently, it exhibits zero power losses during turn-on and turn-off.
On this ground, the switching loss of the advanced circuit is reduced to a lower loss
compared to the one in the conventional boost converter with hard switching operating
mode, owing to the deduction of the dv/dt and di/dt. The lessen losses reach 65.55%
Experimental Investigation of Efficiency and Dynamic Losses ... 1219

Fig. 4 The voltage and current waveforms for the active switch in a, c the hard switching operating
mode, b, d the soft switching operating mode

27.0
60
i

i
Temperature (°C)

i 50 i 24.0

40
21.0
i

i 30
i 18.0
i 20

i 10 i 15.0
i 83 i 166 i 245 i 332 i 415 i 498 i 100 i 200 i 300 i 400 i 500
Time (s) (a) Time (s)
(b)
i
i

Fig. 5 The Switch junction temperatures measured for a the conventional boost and, b the proposed
boost

in the switch. From the Fig. 5(a), it can be observed that hard-switching operating
mode increases the temperature in the Switch up to 58 °C due to the presence of
the spikes, ripples as was seen in Fig. 4(c) and the high-power losses, which impact
the switch’s performance and accelerates aging of the device. Comparing to that, the
proposed circuit has unveiled a advanced dynamic behavior of the temperature by
1220 W. El Merrassi et al.

reducing it into 44.44% as given in Fig. 5(b). Owing to the fewer power losses and
the good waveforms of the voltage and current above the switch.
To analyze further the effectiveness of the proposed converter, and examine this
issue with the previous proposed models, an electrical characteristics analysis of the
principal components is adopted for the different converters. As seen in Table 1, the
major drawback of zero voltage switching of boost converter in [9] is the persistent
of the ripple and peak in the main switch current. This issue would raises the turn-off
capacitance losses, which enlarge the global losses in the system. Further, to the need
of an additional control circuit for the snubber circuit that expand the size and the cost
of the converter. Alongside in [6, 11] the zero voltage transition converter includes a
high number of extra component, which diminished the efficiency, complicated the
implementation and increased the cost. In addition, the high voltage rate of power
main switch enlarge the on-state resistance Rds(on). As result, in the converter [14]
the conduction losses of the main switch get increased permanently. However In
[15], the current and voltage stress of the main switch is substantially decreased.
On the ground of this, in this converter, the switch-on capacitive losses are atomic.
Compared totheseconverters,theproposed model provided a good voltage and current
quality waveform, by assuring less of stresson the switch, higher power efficiency,
ease of control and implementation and a very low cost.

Table 1 Comparativeresultsforthedifferentconverter
Converter Converter Converter Converter Converter Converter Proposed
[9] [6] [14] [15] [11] [10] converter
Number of 6 9 5 5 7 5 5
extra
component
Type of Active Passive Passive Passive Passive Active Passive
snubber
Voltage stress VH 1.6VH 1.2 VH VH VH 1.3 VH VH
of the main
switch
The main 0 1.5Vds 0.2 Vds 0 0 5.5 Vds 0
switch peak
voltage (V)
The main 2 Ids,rms 0 5 Ids,rms 0 1.2 Ids,rms 1.8 Ids,rms 0
switch peak
current (A)
Additional Yes No No No No Yes No
control circuit
Efficiency (%) 95% 92.8% 95.6% 94.2% 94% 93.3% 97.3%
Cost High High Medium Medium High Medium Low
Implementation Most Most Simple Complex Complex Simple Simplest
complexity complex complex
Experimental Investigation of Efficiency and Dynamic Losses ... 1221

97
Ref[15]
Efficiency (%)
92
Proposed
87
Ref[14]
82 Ref[11]

77 Ref[9]
20 40 60 80 100
Conventional
Output Power (%)

Fig. 6 Efficiency curves of the proposed converter under hard switching and soft switching,
compared with the soft converter proposed by [9, 11, 14, 15]

The efficiency curves of the conventional solar hard-switching boost converter and
the proposed soft boost converter are illustrated in Fig. 6. In total, the soft switching
converter presented an extremely high power efficiency than that of the hard switching
converter for the different power range. Overall, the value of the efficiency attained is
97.03%, while the conventional converter reached 90.17%. Therefore, the proposed
circuit enhanced, and increased the efficiency up to 7%. On another hand, the figure
presented as well the power efficiency for the previous model converter presented in
the literature [9, 11]. These models presented different value of the power efficiency.
For the [6] model the efficiency gained for full load is 95.3%, while in [9, 11, 15] it
ranges among 92% and 94.2%. Compared to these results, the proposed converter still
achieves higher efficiency, the improvement reached over 6.83% for light load and
more than 4.51% for heavy load. As advantage, the converter assured high efficiency
for light loads as for heavy one.

5 Conclusion

The power conversion efficiency can be increased by achieving soft-switching

strategy in DC-DC static converters. In this paper, a dynamic losses investigation was
treated for a solar boost behaviour under hard switching and soft-switching operating
mode. The proposed snubber model validates the effectiveness using experimental
test. According to these results, the main switch assured a zero voltage switching
and a zero current switching, beside the power losses was diminished by 65%, which
entitled a stabilization of the temperature junction and assured a reduction up to
44.5%. It also reveals that the proposed circuit has no extra current/voltage stress on
the switch. Further, the efficiency attained 97%. As results, it is proven that the Solar
DC-DC static boost converter with the proposed passive snubber provides better
performance than the conventional boost for the renewable energy applications [16].
1222 W. El Merrassi et al.

References

1. Solatialkaran D, Zare F, Saha TK, Sharma R (2020) A Novel approach in filter design for grid-
connected inverters used in renewable energy systems. IEEE Trans Sustain Energy 11:154–164
2. Shan Y, Hu J, Chan KW, Fu Q, Guerrero JM (2019) Model predictive control of bidirectional
DC–DC converters and AC/DC interlinking converters—a new control method for PV-wind-
battery microgrids. IEEE Trans Sustain Energy 10:1823–1833
3. EL Merrassi W, Abounada A, Ramzi M (2019) Design of a PWM sliding mode voltage
controller of a DC-DC boost converter in CCM at variable conditions. In: Advances in smart
technologies applications and case studies. Springer, Cham, pp 263–270
4. Jabbari M, Mokhtari M (2020) High-Frequency Resonant ZVS boost converter with grounded
switches and continuous input current. IEEE Trans Industr Electron 67:1059–1067
5. Mohammadi MR, Peyman H, Yazdani MR, Mirtalaei SMM (2020) A ZVT bidirectional
converter with coupled-filter-inductor and elimination of input current notches. IEEE Trans
Industr Electron 67:7461–7469
6. Zhang Z, Yang X, Zheng TQ, Zhang J (2020) A Passive soft-switching snubber with energy
active recovery circuit for PWM inverters. IEEE Access 8:100031–100043
7. Chen Y-T, Li Z-M, Liang R-H (2018) A Novel Soft-Switching interleaved coupled-inductor
boost converter with only single auxiliary circuit. IEEE Trans Power Electron 33:2267–2281
8. Lin H, Hu B, Li F, Chen J, Si L, Zhou X, Li Y, Chen J, Yan M, Dong Y (2018) A fault-tolerant
two-permanent magnet synchronous motor drive with field-oriented control scheme. In: 2018
2nd IEEE advanced information management,communicates,electronic and automation control
conference (IMCEC), pp 1029–1033
9. Zhang Y, Cheng X-F, Yin C, Cheng S (2018) Analysis and research of a soft-switching
bidirectional DC–DC converter without auxiliary switches. IEEE Trans Industr Electron
65:1196–1204
10. Muchina EG, Masike L, Njoroge Gitau M (2019) High boost-ratio bidirectional converter
for interfacing low-voltage battery energy storage system to a DC bus. IET Power Electron
12:2372–2381
11. Do H-L (2011) Nonisolated bidirectional zero-voltage-switching DC–DC converter. IEEE
Trans Power Electron 26:2563–2569
12. Meng T, Ben H, Li C (2019) An input-series flyback converter with the coupled-inductor-based
passive snubber for high-input voltage multiple-output applications. IEEE Trans Ind Electron
66:4344–4355
13. El Merrassi W, Abounada A, Ramzi M (2020) Switching losses analysis of a constructed solar
DC-DC static boost converter. Adv Electr Electron Eng 18:132–141
14. Lin J-Y, Lee S-Y, Ting C-Y, Syu F-C (2019) Active-clamp forward converter with lossless-
snubber on secondary-side. IEEE Trans Power Electron 34:7650–7661
15. Zhang Y, Cheng X-F, Yin C, Cheng S (2018) A soft-switching bidirectional DC–DC converter
for the battery super-capacitor hybrid energy storage system. IEEE Trans Industr Electron
65:7856–7865
16. El Merrassi W, Abounada A, Ramzi M (2021) Advanced speed sensorless control strategy for
induction machine based on neuro-MRAS observer. Mater Today: Proc
Improvement of the Non-linear Control
Strategy of a Wind Turbine
by a High-Gain Observer

Azeddine Loulijat, Najib Ababssi, and Mohamed Makhad

Abstract The production of a wind turbine depends mainly on the energy of the
wind, the power graph of the turbine, and the response of the generator to changes
in the wind. This article suggests a robust sliding mode of second-order nonlinear
control technique based on the concept of the super-twisting algorithm for the doubly-
fed induction generator (DFIG) of a wind turbine that provides optimal power produc-
tion (efficiency). For this task, an observer with a high-gain was added for Torque esti-
mation delivered from a wind turbine with a sliding mode of second-order non-linear
control. The suggested overall method was evaluated using MATLAB/SIMULINK
software with a 2 MW three-blade wind turbine.

Keywords Wind turbine · DFIG · Non-linear control · Sliding mode of

second-order · High-gain observer · Super-twisting algorithm

1 Introduction

Much other research on wind turbines has already been achieved. As result wind
turbines operate with variable speeds and have pitch control in recent models. We
are capable of modifying the speed and angle of pitch of each blade, Which makes
it possible for us to increase wind turbine production [1]. However, it must intro-
duce more intelligence into the operation of wind turbines. The main issues of this
article are proposed control of the g enerator which can improve the production of
the wind turbine, including energy production quality and energy efficiency. But
also Mechanical exhaustion reduction, which would create it viable to construction
lighter aero-turbines, Thus enhances productivity [1]. Therefore, control must take
the system’s behavior as a global system into account. Also, wind disturbances must
be taken into account. The studied wind turbine is a three-blade wind turbine with

A. Loulijat (B) · N. Ababssi

IMII Laboratory FST, Hassan 1 University, Settat, Morocco
M. Makhad
ERERA Laboratory ENSET, Mohammed V University, Rabat, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1223
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_112
1224 A. Loulijat et al.

Fig. 1 Global synoptic of the system

a horizontal axis and pitch regulation operating a doubly-fed induction generator

(DFIG). Indeed, Cost reduction is the biggest objective, the power circulating in This
design of the converters is lower compared to a cage generator [2, 3] or synchronous
generator [2–4]. The proposed control strategy schematically illustrated in Fig. 1, is
divided into two main blocks:1) This block represents the high-gain observer and
sliding mode of second-order control (SMSOC) defining the torque reference that
the generator must offer to optimize performance compared to proportional-integral
control (PIC) [5] and sensorless vector control (SVC) [6], which use considerations
and approximations to define the torque reference as a result, the performance is low,
and 2) the block controls by a sliding mode of second-order (SMSOC) that controls
the DFIG through the converters so that the measurements tend to their references
promptly.

2 Mathematical Model of the Wind Turbine

The turbine modeling is inspired from [1–4, 7]. The disposable power of the wind
passing through an Sv surface is defined by:

1
Pwind = ρ Sv v 3 (1)
2
The wind turbine then aims to capture power following:

1
Pa = ρπ R 2 C p (λ)v 3 (2)
2
Improvement of the Non-linear Control Strategy … 1225

The power coefficient is determined by the power ratio of the wind turbine:

Pcap
Cp = (3)
Pwind

C p is a function defined by the specific speed:

Rω
λ= (4)
v
The power captured by the wind turbine can be expressed by [1–4, 7]:

Pcap = Ta ω (5)

From the above equations, The aerodynamic torque expression can be described
by:

1 C p (λ)
Ta = ρπ R 3 Cq (λ)v 2 with Cq (λ) = (6)
2 λ
The dynamic equation of the turbine is written by [1–4, 7, 8]:

dω
J = Ta − f ω − Tem (7)
dt

3 Optimization of the Produced Power

The torque of reference generated by the block (1) has to answer two problems:
Maximizing wind turbine power and areas operational management [7]. The ratio
among the powers obtained from the wind and the available total power presents a
maximum known by the Betz limit. This maximum has never really occurred, And
the power coefficient of each turbine is determined by its relative speed, representing
the ratio of the speed from the tip of the turbine blades to the wind speed. Figure 2
gives an example of a power coefficient with a fixed pitch angle β for a wind turbine
changing on the tip speed ratio λ [9, 10].
More efficiently extract the wind power while keep the regime safe,the wind
turbine also must be usable following the three zones, linking wind speed, maximum
rotor speed acceptable, and the power like to see as shown in Fig. 3. In zone I,
the wind turbine is out of service because the wind is not intense to make energy
production viable about operating energy [10, 11]. In zone II, The partially loaded
wind generator works.Here the main goal is to optimize performance energy. So,
the rotation speed varies with wind speed to maintain itself at the point of operation
with maximum efficiency of aerodynamic. The aim is to ensure an always optimum
1226 A. Loulijat et al.

Fig. 2 Power coefficient curves

Fig. 3 Typical wind turbine power curves derived by wind speed

power coefficient [10, 11]. Zone III is a full loading activity (Intense wind). The
power must limit in order not to deteriorate the system [10, 11].
The maximum efficiency of the turbine is ensured by keeping the maximum power
coefficient [10–12]. From Eq. (6), we can write the expression of the aerodynamic
torque:

1 C p max
Ta = αω2 & α = ρπ R 5 3 (8)
2 λopt

Where the λopt specific speed helps us to maximize the power extracted [8]. The
aim of The approach used in the following for Ta to converge on Topt while Tg = Topt
is required by normal laws. This facilitation consists of ignoring The impact of
mechanical power transmission causing a power loss. A high-gain observer [13–15]
is used for estimating the aerodynamic torque for that the control is appropriate.
Compared to an observer in sliding mode [13], it allows us to limit the chattering
phenomenon in practice [14].
Improvement of the Non-linear Control Strategy … 1227

3.1 Observer Synthesis

From Eq. (7), we can write:

dω Ta fω Tem
= − − (9)
dt J J J
By choosing the following state variables:
 
Ta T
x = [x1 x2 ] = ω
T
; Y = x1 ; U = Tem
J

So the state model is written:


Ẋ = AX + ϕ(X, U ) + ε(t)
(10)
Y = CX

The state model of a high-gain observer [13–18] candidate to the system written
by the state Eq. (10) is:

X̂˙ = A X̂ + ϕ( X̂ , U ) − θ
−1 −1 −1
θ S C C( X̂ − X ) (11)
Ŷ = C X̂

Where
θ be the diagonal matrix and S is the only solution of Lyapunov algebraic
Eq. (12) [13–18]:

S + AT S + S A − C T C = 0 (12)

By defining the error of observation by [19]:

x̃ =
θ (x̂ − x) (13)

Deriving from it, we have:

x̂˙ = θ (A − S −1 C T C)x̃ +
θ (ϕ(x̂) − ϕ(x)) −
θ ε(t) (14)

3.2 Determination of the Conditions for Convergence.

We consider the following function Lyapunov in order to discover the convergence

[18]:
1228 A. Loulijat et al.

V (x̃) = x̃ T S x̃ (15)

V̇ (x̃) = 2 x̃ T S x̃˙ = −θ V − x̃ T C T C x̃ + 2 x̃ T S
θ (ϕ(x̂) − ϕ(x)) − 2 x̃ T S
θ ε̃(t)
(16)

By Eq. (12), therefore:

 
V̇ (x) ≤ −θ V + 2x̃σmax (S)(
θ (ϕ(x̂) − ϕ(x)) + 
θ ε̃(t)) (17)

Now we can assume (ϕ is a function considered lipchitzian and is bounded) that:

 

θ (ϕ(x̂) − ϕ(x)) ≤ τ x̃ &  f (t) ≤ ρ

Where τ = k
J
and ρ = sup t≥0 ε(t). So we have:

ρ
V̇ (x̃) ≤ −θ V + 2x̃2 σmax (S)τ + 2x̃σmax (S) (18)
θ
From the result of the mathematical Lemma 2.1. Assume that system (10) satisfies
the above assumptions. Then [18]:



We have x = T a ⇒ T = J x , conclude that:

 

2 J a 2

T̃a = T̂a − Ta = J (x̂2 − x2 ) ≤ J [θ σ e(−ηθ t) x̃(0) + θ ρ] (19)

θ is decreasing when θ is increasing, allowing a practical estimate of the

aerodynamic torque [8].

3.3 Determination of the Law of Control

The sliding mode of second-order is the control law used in this part is founded on
the algorithm of super-twisting so that the estimated aerodynamic torque converges
to the optimal torque in the short amount of time [9]. Pursuit error is defined by:

eT = Topt − T̂a (20)

The dynamics of this error is giving by its derivative for time.

ėT = 2αopt ω (T̂a − f ω − Tem ) − T̂˙a (21)

Improvement of the Non-linear Control Strategy … 1229

We develop the last expression:

ėT = −2αopt ωTem + [2αopt ω(T̂a − f ω) − T̂˙a ] (22)

We pose: W = 2αopt ω (T̂a − f ω) − T̂˙a . Then:

ëT = −2αopt ω Ṫem + Ẇ (23)

The control law is given by [20–22]:

Tem = U + β1 |eT |0.5 sgn(eT ) with U̇ = β2 sgn(eT ) (24)

The constants β1 and β2 satisfy the following inequalities.

  
2 Ẇ  < 2
β1 ≥ μm
42 μ M (l1 +1 ) &
β22 ≥ μ3m (l1 −1 ) 0 < μm ≤ μ M

Thus, we can say that there is a finite time tc for that: t ≥ tc ⇒ Ta = Topt .
The equivalent control law can obtain by annulling the Eq. (22):

[2αopt ω(T̂a − f ω) − T̂˙a ]

1
Tem eq = (25)
2αopt ω

4 Electric Generator Model

A doubly-feed induction generator model is written in [13–23] in the d-q frame of

park. Moreover, the various article assumes that stator resistance is ignored since
the DFIG used in a wind energy device still reflects a reasonable assumption. From
these considerations, we arrive at a system of Eq. (26).
⎧ dI
⎪
⎨ dt = σ L r (Vr d − Rr Ir d + gωs σ L r Irq − L s dt )
rd 1 M dφsd
d Irq
= σ 1L (Vrq − Rr Irq − gωs σ L r Ir d − gωs LMs φsd ) (26)
⎪ dt
⎩ T = − pr M φ I
em L s sd rq

Where σ = 1 − LMs L r and g = ωsω−ω

2

s
.
The reactive power reference value will be kept at zero to maintain a unity power
factor on the stator side [24]. Within d-q coordinates, the reactive power is defined
by:

Q s = Vsq Isd − Vsd Isq (27)

1230 A. Loulijat et al.

Adapting this equation to our assumptions simplifiers gives:

Vs φs Vs M
Qs = − Ir d (28)
Ls Ls

The desire reactive power is: Q s = 0

Using these equations, we get the reference value.

Vs Vs
Idr _r e f = ; φs = (29)
ωs M ωs

The proposed control strategy also uses the algorithm that says super-twisting. In
this perspective, we consider the following errors.

s Ir d = Ir d − Ir d_r e f
(30)
sTem = Tem − Tem_r e f

The dynamics of these errors is expressing by:

⎧ 
⎨ ṡ I = 1
Vr d − Rr Ir d + gωs σ L r Irq − LMs dφdtsd − I˙r d_r e f
rd σ Lr  (31)
⎩ ṡT = − p σ L1s L r φs Vrq − Rr Irq − gωs σ L r Ir d − gωs LMs φsd − Ṫem_r e f
em

By posing G 1 and G 2 such that:

1 1 M M
s̈ Ir d = V̇r d + Ġ 1 − Rr Ir d & s̈Tem = − p φs V̇rq + Ġ 2 + p φs Rr Irq (32)
σ Lr σ Lr σ L s Lr σ L s Lr

Let’s now consider the following control [19–21]:

  0.5
Vr d = z 1 − β3 s Ir d  Sgn(s Ir d ) + σ 1L r Rr Ir d with ż 1 = −β4 Sgn(s Ir d )
 0.5
Vrq = z 2 + β5 sTem  Sgn(sTem ) − p σML r φs Rr Irq with ż 2 = β6 Sgn(sTem )
(33)

With positive constants β3 , β4 , β5 , β6 , 3 and 4 that satisfy the following

inequalities.
 
Ġ < 3 ; β3 > 3 & β 2 ≥ 4 3 (β3 +3 )
 1 σ Lr 4 σ 2 L r2 (β3 −3 )
Ġ 2 < 4 ; β5 > p M
φs 4 & β 2 ≥ 4 4 (β5 +4 )
σ Ls Lr 6 2 2 σ L r (β5 −5 )
Improvement of the Non-linear Control Strategy … 1231

5 Discussion of Simulation Results

The control strategy suggested in this article was validated on the 2-MW DFIG-based
WT using the MATLAB/Simulink environment. The following table illustrates the
values of some technical parameters used in the simulation (Table 1).
As can be seen in Fig. 4, we can estimate the aerodynamic torque precisely, which
allows us to control it. As concerns, the DFIG control, the performances obtained are
very satisfactory as they show the trajectory tracking of the desired electromagnetic
torque (Fig. 5a). Also, Fig. 5a shows that the torque on the shaft is not subject to the
phenomenon of chatter, thus limiting the mechanical stress on the entire transmission
of the wind turbine and Fig. 5b shows a very rapid convergence to the rotor reference
current.
To demonstrate the interest of the proposed strategy, its performances had been
comparing with control strategies such as proportional-integral control (PIC) [5]. The
PIC uses active power as a reference. This presumes that the active power conforms
to the electromagnetic power.

Ls
Ps = Pr e f ⇒ Irq_r e f = − Pr ė f (34)
Vs M

Table 1 The wind turbine

Parameter Value
Number of blades 3
Length of blade (m) 39
Swept area Sv (m2 ) 5027
Wind speed at nominal speed and at Cp-max : v(m/s) 11

Fig. 4 Torque estimated (green), Torque optimal (blue), Torque real (red)
1232 A. Loulijat et al.

Fig. 5 Electromagnetic torque with SMSOC (a), Present efficiency of Ird tracking (b)

Fig. 6 Electromagnetic torque with PIC (a) and SVC (b)

This approximation causes problems of convergence towards the desired couple

(Fig. 6a). A second SVC is to use the following equation [6]:

Ls
Irq_r e f = − Tem_r e f (35)
pMφs

The many simplifications made to obtain the expression of the electromagnetic

torque of (35) and also the fact that the stator flux is considered constant do not allow
the desired torque to be followed precisely (Fig. 6b). Compared to these two controls,
the strategy proposed allows a very good pursuit of the reference. This undoubtedly
improves performance. The simulations were realized with the following gains: θ =
30, β1 = β5 = 1.5, β2 = β6 = 50, β3 = 200, β4 = 1000, αopt = 161240, 1 =
3 = 5.103 , 2 = 4 = 5 = 1.105 .
Improvement of the Non-linear Control Strategy … 1233

6 Conclusion

A linear control method such as PIC and SVC is standard practice for A linear control
method such as PIC and SVC is standard practice for dealing with generator control
in wind turbines. However, these control methods come at the costs of decreased
system performances awaited due to random operation and process uncertainties
considered. As a consequence, a strong non-linear control method should be used
to address these issues. The method has been provided and discussed in this article.
Firstly a high-gain observer to define the reference torque in combination with the
sliding mode of second-order control, secondly a sliding mode of second-order is
selected based on a super- twisting algorithm to control the doubly-fed induction
generator, achieves the desired goals. It is strong and so durable, increases energy
efficiency, and limits mechanical stress on the drive shaft thanks to the low chatter it
produces. In future work, this strategy will be combined with the method of passive
protection in case of network disturbance in the form of a voltage dip.

References

1. Baala Y, Bri S (2020) DFIG-based wind turbine control using high-gain observer. In: 2020 1st
Interanational Conference Innovation Research Application Science Enginnering Technology,
IRASET 2020 (2) https://doi.org/10.1109/IRASET48871.2020.9092280
2. Goudarzi N, Zhu WD (2013) A review on the development of wind turbine generators across
the world. Int J Dyn Control 1(2):192–202. https://doi.org/10.1007/s40435-013-0016-y
3. Lin FJ, Tan KH, Fang DY (2015) Squirrel-cage induction generator system using hybrid wavelet
fuzzy neural network control for wind power applications. Neural Comput Appl 26(4):911–928.
https://doi.org/10.1007/s00521-014-1759-x
4. Mustafaev RI, Gasanova LG (2015) An investigation of the dynamics of a wind power unit
equipped with synchronous generators with permanent magnets. Russ Electr Eng 86(5):258–
263. https://doi.org/10.3103/S1068371215050077
5. Poitiers F, Bouaouiche T, Machmoum M (2009) Advanced control of a doubly-fed induction
generator for wind energy conversion. Electr Power Syst Res 79(7):1085–1096. https://doi.org/
10.1016/j.epsr.2009.01.007
6. Cárdenas R, Peña R (2004) Sensorless vector control of induction machines for variable-speed
wind energy applications. IEEE Trans Energy Convers 19(1):196–205. https://doi.org/10.1109/
TEC.2003.821863
7. Beltran B, Ahmed-ali T, El M, Benbouzid H, Member S (2008) Sliding mode power control
of variable-speed wind energy conversion systems 23(2):551–558 (2008)
8. Touati A, Abdelmounim E, Aboulfatah M, Moutabir A, Majdoul R (2015) Improved strategy
of an MPPT based on the torque estimator for variable speed turbines. Int Rev Model Simul
8(6):620–631. https://doi.org/10.15866/iremos.v8i6.7122
9. Abd I, Abdelmounim E, Aboulfatah M, Majdoul R, Moutabir A (2015) Design of an Mppt
based on the torque estimator for variable speed turbines (1):166–171
10. Beltran B, Benbouzid MEH, Ahmed-Ali T (2020) A combined high gain observer and high-
order sliding mode controller for a DFIG-based wind turbine. In: 2010 IEEE International
Energy Conference Exhibilish EnergyCon 2010 (1):322–327 (2010). https://doi.org/10.1109/
ENERGYCON.2010.5771699
1234 A. Loulijat et al.

11. Zamzoum O, Derouich A, Motahhir S, El Mourabit Y, El Ghzizal A (2020) Performance

analysis of a robust adaptive fuzzy logic controller for wind turbine power limitation. J Clean
Prod 265:121659. https://doi.org/10.1016/j.jclepro.2020.121659
12. Adapa H, Pushpak V, Kumar MV (2020) Nakandhrakumar RS: Modeling and simulation of
pitch controlled FRP material based horizontal axis wind turbine system to extract maximum
power. Mater. Today Proceeding 2020. https://doi.org/10.1016/j.matpr.2020.06.131
13. Deshpande AS, Patil SL (2020) Robust observer-based sliding mode control for maximum
power point tracking. JControl Autom Electr Syst. https://doi.org/10.1007/s40313-020-006
17-5
14. Khalil HK (2008) High-gain observers in nonlinear feedback control. In: 2008 International
Conference Control Automation System, ICCAS 2008 (2008). https://doi.org/10.1109/ICCAS.
2008.4694705
15. Adekanle OS, Guisser M, Abdelmounim E, Aboulfatah M (2018) Nonlinear controller with
rotor crowbar and DC-chopper fault ride through technique for grid-connected doubly-fed
induction generator. Int Rev Autom Control 11(6):281–292. https://doi.org/10.15866/ireaco.
v11i6.13496
16. Zareian M, Shafiei MH (2020) A modification in the structure of low-power high-gain observers
to improve the performance in the presence of disturbances and measurement noise. Eur J
Controlhttps://doi.org/10.1016/j.ejcon.2020.07.009
17. Kadrine A, Tir Z, Malik OP, Hamida MA, Reatti A, Houari A (2020) Adaptive non-linear high
gain observer based sensorless speed estimation of an induction motor. J Franklin Inst.https://
doi.org/10.1016/j.jfranklin.2020.06.013
18. Farza M, M’Saad M, Rossignol L (2004) Observer design for a class of MIMO nonlinear
systems. Automatica 40(1):135–143. https://doi.org/10.1016/j.automatica.2003.08.008
19. Ahmed-Ali T, Cherrier E, M’Saad M (2009) Cascade high gain observers for nonlinear systems
with delayed output measurement. Proc IEEE Conf Decis Control 4:8226–8231. https://doi.
org/10.1109/CDC.2009.5400398
20. Levant A, Alelishvili L (2007) Integral high-order sliding modes. IEEE Trans Automat Contr
52(7):1278–1282. https://doi.org/10.1109/TAC.2007.900830
21. Moussa O, Abdessemed R, Benaggoune S (2019) Super-twisting sliding mode control for
brushless doubly fed induction generator based on WECS. Int J Syst Assur Eng Manag
10(5):1145–1157. https://doi.org/10.1007/s13198-019-00844-3
22. Zhang C, Gutierrez SV, Plestan F, De León-Morales J (2019) Adaptive super-twisting control of
floating wind turbines with collective blade pitch control. IFAC-PapersOnLine 52(4):117–122.
https://doi.org/10.1016/j.ifacol.2019.08.165
23. Beltra B (2010) Contribution à la commande robuste des éoliennes à base de génératrices
asynchrones double alimentation: Du mode glissant classique au mode glissant d’ordre
supérieur
24. Zouggar EO, Chaouch S, Abdeslam DO, Abdelhamid AL (2019) Sliding control with fuzzy
type-2 controller of wind energy system based on doubly fed induction generator. Instrum.
Mes. Metrol. 18(2):137–146. https://doi.org/10.18280/i2m.180207
Optimization of DFIG Wind Turbine
Power Quality Through Adaptive Fuzzy
Control
Hamid Chojaa, Aziz Derouich, Mohammed Taoussi, Othmane Zamzoum,
and Mourad Yessef

Abstract The control algorithms to optimize power extraction in wind energy

conversion systems are often designed by making some approximations and taking
standard parameter values into consideration. This approach has proven inefficient as
parameter variations and disturbances often degrade the robustness and power quality
of the system. This paper proposes a simple-to-implement adaptive techniques to
upgrade the efficiency and power quality of the system. Simulation results obtained
on a 1.5 MW Doubly Fed Induction Generator (DFIG) are given and comparison is
made against the performance of a PI-controller.

Keywords Doubly Fed Induction Generator · Adaptive control · Power quality ·

Maximum power point tracking (MPPT) · Fuzzy logic · Disturbance compensation

1 Introduction

Due to the fact that the real nature of wind speed is stochastic, the grid-connected
DFIG is provided with automatic control strategies for optimal power extraction and
efficient exchange of energy between the turbine and the grid.
The aim of the MPPT strategy, when the turbine is on partial load is to continually
adjust the operating point of the wind turbine in order to convert the maximum power
possible. A technique using relative velocity [1] has been proposed in the literature.
Meanwhile, wind speed acquisition from a single anemometer leads to erroneous
measure of wind speed which does not the effective wind speed incident on the
blades. Thus, a faulty measurement of the incident wind speed can cause the deterio-
ration of the extracted power. Wind speed estimation with two different Kalman filter
versions were proposed and analyzed by the authors in [7]. The complexity of Kalman

H. Chojaa (B) · A. Derouich · M. Taoussi · O. Zamzoum

Laboratory of Technologies and Industrial Services, Higher School of Technology, SMBA
University, Fez, Morocco
e-mail: [email protected]
M. Yessef
Laboratory of Engineering, Modeling and Systems Analysis, SMBA University, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1235
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_113
1236 H. Chojaa et al.

filters and the high number of tuning parameters required are the major drawbacks
of this method. A simple method based on Fuzzy Logic were synthesized in [3] to
estimate wind speed and track the DFIG variables. Easy to implement Perturbation
and Observation technique usually employed in photovoltaic systems is proposed in
[5]. It can be implemented easily, not expensive and does not require accurate param-
eters of the turbine and wind speed information. However, a compromise between
convergence speed and performance is needed. Neural Network, Gradient methods
have been proposed to give very good performance but are yet to be easy to imple-
ment [1]. The disadvantage of the Optimal Torque Control technique presented in
[9] is that characterization tests are required off-line to design the look-up table. This
paper proposes a scheme based on Fuzzy Logic, which is unrestrained by wind speed
measurement, knowledge of turbine parameters and characterization experiment to
search the maximum power point. It indirectly compensates for the different power
loss and requires only the acquisition of DFIG speed and output power signals.
Vector control schemes and PI-controllers have been employed hand-in-hand in
the literature to regulate DFIG velocity, power, and currents. However, grid voltage
dips, model incorrectness and unexpected conditions like temperature increase dete-
riorate the effectiveness of this control technique. To improve on these inconve-
niences, several other controllers like Sliding Mode [9, 10], Backstepping [2, 8],
Feedback Linearization [6], Neural Predictive Control [4] have been studied and
evaluated in the literature. The drawbacks of these approaches range from chattering
to implementation complexity.
In this paper, we design an adaptive control technique in line with Lyapunov
stability theorem to regulate speed and current. Estimating and compensating uncer-
tainty in real time, it is characterized by fast response time, good precision, better
robustness against disturbances.

2 System Modelling Under Stator Voltage Orientation

Figure 1 illustrates the set-up of grid-connected turbine comprising a DFIG connected

directly to the grid through its stator and indirectly via the rotor windings with back-
to-back power converters as intermediary. The converters are denoted Rotor Side
Power Converter (RSC) and Grid Side Power Converter (GSC).
This section presents the mathematical model of all the components of the turbine
system needed for the control and observation of the system.
For the independent control of active and reactive power to be achieved, Stator
Voltage Orientation Scheme is employed to align the d-axis of the reference frame
to the stator voltage position. Thus, it is directly obtained that vsq ≡ 0 and vsd equals
to the magnitude of the grid voltage. RSC mathematical models can then become
(1)–(4). This consequently means that direct ir d and quadrature irq components of
rotor current control generator speed ωg and stator reactive power Q s respectively
[10].
Optimization of DFIG Wind Turbine Power Quality … 1237

Fig. 1 DFIG connection to the grid

dir d vr d Rr ir d L m ωr ϕsq
= − + ωr irq + + θr d (1)
dt σ Lr σ Lr σ Lr Ls
dirq vrq Rr irq
= − − ωr ir d + θrq (2)
dt σ Lr σ Lr
dωg 3 Lm
J = Tm − f v ωg − p ϕsq ir d + θω (3)
dt 2 Ls

Ps = − 23 LLms vsd ir d
  (4)
Q s = − 23 vLsds ϕsq − L m irq

3 Controller Design

Three controllers are designed in for the work presented in this paper: pitch angle,
RSC and GSC controllers. The pitch angle controller has for objective to mechan-
ically prevent blade rotation speed from exceeding nominal power and speed. The
aim of the RSC controller is to adaptively drive and set the system on the maximum
power point and track desired stator reactive power. The PI corrector, designed but
not presented in this paper to control the GSC, is aimed at maintaining DC voltage
constant and annulling filter reactive power.
1238 H. Chojaa et al.

Fig. 2 Turbine power

variation in function of
generator speed for different
wind speed

3.1 Fuzzy Logic MPPT

Figure 2 presents the curve of mechanical power in function of generator speed for
different wind speed values. Popt represents the maximum power point (MPP) of the
turbine system. Since the wind speed is random and not fix, to obtain the optimum
output power of the wind turbine, an adaptive controller is needed to track the optimal
power curve through the MPPT strategy.
The MPPT strategy in this paper adopts the fuzzy technique described in Fig. 2(b),
well known as an adaptive and robust tool for complex nonlinear systems. It is
employed in this paper because it neither needs the information of turbine param-
eters nor wind speed signal acquisition. The FLC perturbs the generator speed and
observes the direction of variation of the output power and adjusts the speed accord-
ingly in other to place the turbine at the operating point where energy extracted is
maximal. The output is the optimal generator speed (ωopt ) that ensures the extraction
of maximum power from the turbine blades. The speed reference is then fed to a fast
controller which converges the real generator speed to the reference. The calculation
of output power from stator and filter measurement is more efficient because it avoids
the complex determination of power loss in the gearbox, the converters and the filter.
The fuzzy logic controller synoptic diagram is shown in Fig. 3, where K1, K2 and
K3 are adaptive gains.
The FLC searches the maximum point of the power-speed curve of the generator
using the rules presented in Table 1: with P and ωg as inputs, while ωgr e f
represent the output. These variables are represented by NB, NM, NS, Z, PS, PM,
PB denoting Negative Big, Negative Medium, Negative Small, Zero, Positive Small,
Positive Medium, and Positive Big respectively.
The triangular membership functions of input and output are presented in Fig. 4
and have seven fuzzy subsets. The fuzzy inference is carried out by using Sugeno’s
method and the defuzzification uses the center of gravity to determine the output of
this FLC.

3.2 PI Controller (PIC) Design

Figure 5 represents the PI controller synthesized to track reference stator power

factor and reference machine velocity. Controller parameters are deduced using the
Optimization of DFIG Wind Turbine Power Quality … 1239

Fig. 3 Fuzzy Logic MPPT controller

Table 1 MPPT fuzzy rules

ωg P
NB NM NS Z PS PM PB
NB PB PB PM Z NM NB NB
NM PB PM PS Z NS NM NB
NS PM PS PS Z NS NS NM
Z NB NM NS Z PS PM PB
PS NM NS NS Z PS PS PM
PM NB NM NS Z PS PM PB
PB NB NB NM PM PB PB

Fig. 4 Degree of membership of change in speed, power and output

Fig. 5 PI regulator for RSC control

1240 H. Chojaa et al.

pole-zero cancellation method. F(s) = 1/(σ L r s + Rr ) is the transfer function and

L ω ϕ
frq = −ωr ir d and fr d = ωr irq + σmL rrL ssq are the coupling terms.

3.3 Adaptive Backstepping Controller (ABC) Design

The most important feature of the ABC is that, unlike the PIC, it takes eventual uncer-
tainty (θω , θr d , θrq ) into account in the modelling, estimates it and sends commands to
compensates for it in real time. This is essential to mitigate the effect of model incor-
rectness, parameter variation, modelling approximations, and saturation. Current
reference value which guarantees the convergence of generator speed to its optimum
is derived in step 1. An ABC is designed to tracks this reference current in step 2 to
give the final control variable vr d .
Step 1: Eqs. 5 and (6) show speed tracking error and its integral respectively.

εω = ωg − ωopt (5)

Tm f v ωg 3 Lm
ε̇ω = − − p ϕsq ir d + θω − ω̇opt (6)
J J 2 J Ls

Lyapunov quadratic function V1 to ensure that the speed error εω convergences

and remains stable is chosen as and shown in (7). Its derivative V̇1 is given in (8),
and to guarantee that it is strictly negative, the equality in (8) has to be fulfilled.
2
1 1 θω
V1 = εω2 + (7)
2 2 mω
 
Tm f v ωg 3 Lm 

V̇1 = εω − − p ϕsq ir d + θ ω − ω̇opt = −K ω εω (8)

J J 2 J Ls

The current output ir∗d is the reference input for step 2 and is derived from (8) and
presented in (9). Equation 10 represents the estimate of the uncertainty computed in
real time.
  
2J L s Tm f 

ir∗d = kω εω + − + θ ω − ω̇opt = αr d (9)

3 pL m ϕsq J J
˙


θ ω = εω m ω (10)

Step 2: The control of stator reactive power will be included to this step. The
∗
error between the actual and the desired current can be defined as (irq = α2 ):
Optimization of DFIG Wind Turbine Power Quality … 1241

εr d = ir d − αr d
(11)
εrq = irq − αrq
 L ω ϕ
ε̇r d = σvrLdr − RσrLirrd + ωr irq + σmL rrL ssq + θr d − α̇r d
v R i (12)
ε̇rq = σ rqL r − σrLrqr − ωr ir d + θrq − α̇rq
⎧ 2
⎨ Vr d = Vω + 1 ε2 + 1 θ r d
2 rd 2 mr d
2 (13)
⎩ V = 1 ε2 + 1 θ rq
rq 2 rq 2 m rq

Second step Lyapunov functions taking into account uncertainty parameters and
their derivatives, are presented as (13) and (14) respectively.

V̇ω = −K ω εω2 − K r d εr2d
(14)
V̇rq = −K rq εrq
2

Derivatives of the estimates of the uncertainties and the control variables deduced
and given as (15) and (16) ensure that the derivative of the Lyapunov function is
strictly negative.

˙


θ r d = εr d m r d
˙
 (15)
θ rq = m r d εr d
⎧ 
⎨ vr d = σ L r −K r d εr d + 3 p εω L m ϕsq + Rr ir d + ωr irq − ωr L m ϕsq − θ r d + α̇r d


 2 J L s σ L r σ L L
r s
⎩


R i
vrq = σ L r −K rq εrq + σrLrqr + ωr ir d − θ rq + α̇rq
(16)

The control (K ω , K r d , K rq ) and adaptation (m ω , m r d , m rq ) parameters are chosen

strictly positive.

4 Simulation and Discussion

The DFIG model studied in this work is rated 1.5 MW. It is designed under
Matlab/Simulink to examine the performance of the ABC in comparison to that
of the PI-controller taking the stochastic wind speed profile presented in Fig. 6(a)
as input. Figure 6(b) shows that the PIC maintains the DC bus voltage constant for
effective power conversion. Particular interest is placed on the convergence time,
precision, and power quality under pitch angle control activation.
Figure 6(c) proves that actual speeds converge to corresponding reference gener-
ator speeds under the ABC and PIC with good precision and robustness. We however
observe discrepancies between speed evolution under both controllers. This is due
1242 H. Chojaa et al.

(a) (b)

(c) (d)

(e) (f)

(g) (h)

Fig. 6 a Wind speed b DC-link voltage c Generator speed d Power coefficient e Electromagnetic
Torque f Active and reactive power at the point of common coupling g Stator current under ABC
h Stator current under PIC

principally to the better precision and faster convergence of the ABC. This can be
explained by the higher power coefficient obtained under the ABC as illustrated in
Fig. 6(d).
The rapidity of the ABC enables the FLC to take optimum decision more quickly,
leading to better precision and more extracted power. It can also be noticed on
Optimization of DFIG Wind Turbine Power Quality … 1243

Fig. 6(e) that the electromagnetic torque is less noisy under the ABC. This conse-
quently implies less mechanical vibration and better power quality. Figure 6(f) shows
that the powers transmitted to the grid under both controllers possess smooth evolu-
tion and the reactive power converges to zero with more power extracted under
the ABC without chattering. The stator current waveforms under ABC in Fig. 6(g)
takes a better sinusoidal form than that of the PIC given in et Fig. 6(h). This conse-
quently demonstrates the better performance of the ABC. Due to the limitation of
the reference speed below 1.2 p.u, the pitch angle remains close to zero for both
controllers.

5 Conclusion

Optimizing energy conversion in DFIG wind turbine requires to, first, determine
the Maximum Power Point (MPP) and second, control rotor converters to set the
system on the MPP. Due to uncertainties and model imperfection, a Fuzzy Logic
adaptive controller is employed in this paper to track the optimum generator speed.
Comparative studies and simulation results show that the Adaptive Backstepping
rotor converter controller proved more power conversion, better robustness, faster
response time and higher precision than the conventional PI controller.

References

1. Abdullah MA, Yatim AHM, Tan CW, Saidur R (2012) A review of maximum power point
tracking algorithms for wind energy systems. Renew Sustain Energy Rev 16:3220–3227.
https://doi.org/10.1016/j.rser.2012.02.016
2. Adekanle OS, Guisser M, Abdelmounim E, Aboulfatah M (2017) Robust integral backstepping
control of doubly fed induction generator under parameter variation. Int Rev Autom Control
10. https://doi.org/10.15866/ireaco.v10i2.11163
3. Belmokhtar K, Doumbia ML, Agbossou K (2014) Novel fuzzy logic based sensorless maximum
power point tracking strategy for wind turbine systems driven DFIG (doubly-fed induction
generator). Energy 76:679–693. https://doi.org/10.1016/j.energy.2014.08.066
4. Douiri MR, Essadki A, Cherkaoui M (2018) Neural networks for stable control of nonlinear
DFIG in wind power systems. Procedia Comput Sci 127:454–463. https://doi.org/10.1016/j.
procs.2018.01.143
5. Hong C-M, Ou T-C, Lu K-H (2013) Development of intelligent MPPT (maximum power point
tracking) control for a grid-connected hybrid power generation system. Energy 50:270–279.
https://doi.org/10.1016/j.energy.2012.12.017
6. Li P, Xiong L, Wu F, Ma M, Wang J (2019) Sliding mode controller based on feedback lineariza-
tion for damping of sub-synchronous control interaction in DFIG-based wind power plants. Int
J Electr Power Energy Syst 107:239–250. https://doi.org/10.1016/j.ijepes.2018.11.020
7. Song D, Yang J, Dong M, Joo YH (2017) Kalman filter-based wind speed estimation for wind
turbine control. Int J Control Autom Syst 15:1089–1096. https://doi.org/10.1007/s12555-016-
0537-1
8. Youness EM, Aziz D, Abdelaziz EG, Jamal B, Najib EO, Othmane Z, Khalid M, Bossoufi
B (2019) Implementation and validation of backstepping control for PMSG wind turbine
1244 H. Chojaa et al.

using dSPACE controller board. Energy Reports 5:807–821. https://doi.org/10.1016/j.egyr.

2019.06.015
9. Zamzoum O, El Y, Errouha M, Derouich A, El A (2019) Active and Reactive Power Control
of Wind Turbine based on Doubly Fed Induction Generator using Adaptive Sliding Mode
Approach. Int J Adv Comput Sci Appl 10. doi: https://doi.org/10.14569/IJACSA.2019.010
0252
10. Hamid C, Derouich A, Taoussi M, Zamzoum O, Hanafi A (2020) An improved performance
variable speed wind turbine driving a doubly fed induction generator using sliding mode
strategy. In: 2020 IEEE 2nd International Conference on Electronics, Control, Optimization
and Computer Science (ICECOCS). https://doi.org/10.1109/ICECOCS50124.2020.9314629
Use of a Doubly-Fed Induction Generator
for the Conversion of Wind Energy

Wahabi Aicha, Elmoudden Abdelhadi, and Senhaji Rhazi Kaoutar

Abstract Wind energy is considered one of the most important and promising
renewable energy sources in the world. We begin this study with a description of
the current state of the world energy situation and in Morocco in particular. We
then described a medium power wind power conversion chain that we modeled
and simulated. The chain consists of the turbine, the multiplier and the dual
power induction generator (DFIG). The adopted chain model is developed using
Matlab/Simulink/Sim-Power-Systems software. We found fairly correct simulation
results which are presented and analyzed at the end of this study.

Keywords Wind energy · DFIG · Conversion chain

1 Introduction

1.1 Increase in Energy Demand in the World

Electric power is a crucial element for any socio-economic development. Given

the scale of industrialization in recent decades and the proliferation of household
appliances that consume more and more electrical energy, the demand for electrical
energy has become very important. Unfortunately, most of this growth in demand has
been met by fossil fuel energy which, in addition to its harmful effect on the ozone
layer, has a reserve which is of course limited. Faced with this situation, several
countries have resorted to nuclear power plants. However, the risks of a nuclear
accident (Chernobyl, Fukushima) have slowed down the development of nuclear

W. Aicha (B) · S. R. Kaoutar

Network, IT, Telecommunications and Multimedia Research Laboratory in the Superior School of
Technology, University Hassan 2, Casablanca, Morocco
E. Abdelhadi
Computer Science, Renewable Systems and Energies Laboratory, The National School
of Electricity and Mechanics, University Hassan 2, Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1245
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_114
1246 W. Aicha et al.

energy, which encouraged several countries like Morocco to develop other forms of
energy: renewable energies.

1.2 Renewable Energy in Morocco

In Morocco, the development of renewable energies is a national priority which

calls for a diversification of energy sources by 2020 with 42% of the total installed
electrical energy supplied by green energies [6].
The energy strategy of our country is based on the development of energy effi-
ciency and renewable energies: the objective is to save 12% of energy consumption
by 2020 and 15% by 2030. With this in mind, action plans for energy efficiency were
implemented in all key sectors, particularly transport, industry and construction [6].
Figure 1 shows the installed power in Morocco in 2018.
The city of Casablanca, which is the economic capital of morocco, has a fairly
large wind farm. Figure 2 gives an idea of the average wind speeds for each month.
Before the realization of any project, stages of study and design on computer
are essential. A wind farm project goes through several stages. The modeling and
simulation of the project constitutes a very important step for the success of the
project. The study we are going to carry out concerns a medium power aero-generator
project.

2 Description of a Wind Energy Conversion System Piloted

by a DFIG

In the conversion of wind energy, the DFIG is widely used because it has several
ad-vantages: the generation of energy at variable speed (±30% around the speed of

Fig. 1 Power installed in Morocco (MW) in 2018

Use of a Doubly-Fed Induction Generator … 1247

Fig. 2 The wind speed averages for Casablanca

Fig. 3 Conversion of wind energy

synchronism) which allows to obtain the maximum possible power for each wind
speed: MPPT. The wind chain, as shown in the following figure, includes the turbine,
the multiplier, the DFIG whose stator circuit is directly connected to the electrical
network. A second circuit placed at the rotor is also connected to the network, but via
power converters. The machine side power converter is called “Rotor Side Converter”
(RSC) and the line side power converter is called “Grid Side Converter” (GSC).
The machine-side power converter is used to control the active and reactive power
produced by the machine. The mains side converter controls the DC bus voltage and
the mains side power factor (Fig. 3).
1248 W. Aicha et al.

3 Modeling of the Asynchronous Generator

We remember that the asynchronous generator DFIG is modeled in the reference of

Park according to the following equations [4, 5]:

d∅ds
vds = Rs ids + − ωs ∅qs (1)
dt
d∅qs
vqs = Rs iqs + ωs ∅ds (2)
dt
d∅dr
vdr = Rr idr + − ωr ∅qr (3)
dt
d∅qr
vqr = Rr iqr + + ωr ∅dr (4)
dt
⎧
⎪
⎪ ∅ds = Ls ids + Midr
⎨
∅qs = Ls iqs + Miqr
(5)
⎪ ∅dr
⎪ = Lr idr + Mids
⎩
∅qr = Lr iqr + Miqs

With: ωr = ωs − p and: Ls = ls – Ms; Lr = lr + Mr.

Active and reactive power are written

P = vds ids + vqs iqs (6)

Q = vqs ids − vds iqs (7)

4 Modeling and Simulation of the Complete Conversion

Chain

4.1 Schematic Diagram

After having studied each element separately, we will now be interested in the system
in its totality: model it and simulate it. The complete open loop system can be
represented by the following block diagram (Fig. 4).
Use of a Doubly-Fed Induction Generator … 1249

Fig. 4 Diagram of the complete green-loop system (without regulation)

4.2 Results and Comments

In Fig. 5, we show the voltage across the DC bus.

We note that the voltage across the capacitor is (apart from fluctuations) continuous
and equal to ≈ 86.4 V. The small voltage variations (±1 V), are probably due to the
fact that we worked in open loop.
The Fig. 6 shows the stator voltage delivered by the generator.
The phase-to-phase voltage across the generator stator is sinusoidal, with an effec-
tive value ≈ 380 V and a period T ≈ 0.02 s: the frequency is 50 hz, we can say that
the machine operates in optimal conditions (nominal V = 380 V, nominal f = 50 hz).
The following figure shows the three currents feeding the balanced three-phase
resistive load (Fig. 7).
1250 W. Aicha et al.

Fig. 5 Voltage at DC bus terminals

Fig. 6 Stator voltage supplied by the generator

Note that the three-phase system of currents (is1, is2, is3) is balanced with the
same pulsation as the voltages. Imax ≈ 6 A.
The Fig. 8 shows the total active power supplied by the generator.
Use of a Doubly-Fed Induction Generator … 1251

Fig. 7 Stator currents (A)

Fig. 8 Total active power supplied by the wind chain to the load

It can be seen that the wind chain provides a fairly high active power: average P =
2000 W. (The minus sign indicates of course that it is a supplied power). We note the
presence of fluctuations (of the same frequency as the voltage) on the power plot which
are due to the use of power converters and to the fact that we worked without power
regulation.
The last figure shows the total reactive power supplied by the chain (Fig. 9).
The average value of the reactive power is zero since the load we have placed is
of a resistive nature.
1252 W. Aicha et al.

Fig. 9 Reactive power Q in VAR

5 Conclusion and Prospects

In the light of the results given previously, we can conclude that the studied conversion
chain was modeled correctly. The simulation of the design carried out using the
software are quite valid: the currents, the voltages as well as the delivered powers
correspond to the desired results.
In order to minimize the observed parasitic fluctuations, we plan to insert filters
at the terminals of the load.

References

1. Boyette A (2006) Control command of a dual power asynchronous generator with storage system.
Doctoral thesis from Poincaré University, Nancy
2. Wahabi A, El Moudden A, Bounifli F (2014) Modeling and Simulation of a Dual Power
Asynchronous Generator for Wind Energy Control. Moroccan thermal association (2014)
3. Belmokhtar J, Doumbia K, Agbossou K (2010) Modeling and control of a wind power system
based on a dual power asynchronous machine for the supply of power to the electricity grid. J
Sci Res 2
4. Khojet el khil S (2006) Vector control of a double-fed asynchronous machine
(MADA)/Optimization of losses in converters 2006
Use of a Doubly-Fed Induction Generator … 1253

5. Wahabi A, El Moudden A, Bounifli F, Aarib A (2015) Improved performance of a wind energy

conversion chain driven by an asynchronous generator doubly-fed. Int J Electron Commun
Comput Eng 6(2): 255–262. ISSN 2249–071X
6. Azeroual M, El Makrini M, El Moussaoui H, El Markhi H (2018) Renewable energy potential
and available capacity for wind and solar power in morocco towards 2030. J Eng Sci Technol.
Rev 11(1): 189–198
Improvement of the Wind Power System
Based on the PMSG

Btissam Majout, Badre Bossoufi, Mohammed Karim, and Chakib El Bekkali

Abstract Given the role of PMSG in the wind power system, several commands
are proposed to improve the quality of energies supplied to the electrical grid. In this
work, two control strategies are presented and applied to the static converter in order
to control the electromagnetic torque and the active and reactive power exchanged
between the PMSG and grid. Firstly, in order to extract a maximum power from
the wind a speed control strategy (MPPT) has been presented, followed by a PMSG
model in reference d-q. Then, the conception of the Field-Oriented Control (FOC)
and the sliding mode control (SMC) are developed and tested with Matlab/Similink,
where these controls make it possible to operate the system in the best performances
in dynamic regimes. And finally, we ended by a discussion about the interpretation
of the simulation results in order to compare the advantages and the performances
of each control.

Keywords Wind energy conversion system (WECS) · Permanent magnet

synchronous generator (PMSG) · Field-Oriented Control (FOC) · Sliding mode
control (SMC)

1 Introduction

Recently, the development and the exploitation of renewable energy have experi-
enced a strong progress, due to its different advantages (technical, economical and
environmental) compared to conventional energies, including fossil energies (coal,
gas, petrol). There are many types of renewable energy resources (wind, solar, hydro,
geothermal and biomass) [1–3]. Where the wind energy occupies a good place, it is
one of the cleanest, the safest and fastest sectors to implement, it is supposed to be
the best in terms of quality and price [4–6].
There are several research studies about the wind turbine. In particular, those
equipped with Permanent magnet synchronous machine. This type of machine is

B. Majout (B) · B. Bossoufi · M. Karim · C. El Bekkali

Laboratory of Engineering Modeling and Systems Analysis, SMBA University Fez, Fez, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1255
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_115
1256 B. Majout et al.

Fig. 1 General Model of WECS

characterized by reliable operation without gearboxes, lower losses and a reduced

size compared to a doubly-fed induction generator (DFIG) of the same power [7]. The
non-linear nature of the wind system requires robust controllers to mitigate effects
of internal parametric variations and external disturbances [7–10].
In order to control this machine, two control techniques proposed in this paper
include both classical vector control based on PI regulators and non-linear Sliding
mode control based on Lyapunov theory [16]. Recently, many researches are turned
towards the nonlinear techniques SMC to increase the robustness and the preci-
sion of the systems, where this control is characterized by its high precision, rapid
dynamic response, stability, simplicity [16], compared to linear command FOC which
is characterized by its sensitivity to parametric variations [7–16].
This paper is organized as follows: in Sect. 2 a Wind System Model is presented;
Sect. 3 is dedicated to the control Synthesis of SMC and FOC; the simulation results
and discussion are presented in Sect. 4; the paper finishes with a conclusion in Sect. 5.
All simulations performed in the MATLAB/Simulink environment. The general
structure of the wind system is given by (Fig. 1) [7].

2 Wind System Modelling

2.1 Wind Turbine Modeling

The turbine model is presented by the following equations [7–11]:

1
Pwind = .ρ.S.vw3 (1)
2
1
Pturb = Cp (A, β).Pwind = .ϕ.S.Cp (A, β).v3 (2)
2
Improvement of the Wind Power System Based on the PMSG 1257

Table 1 Wind turbine

Parameters Coefficients
aerodynamic constants
C1 0.5176
C2 116
C3 0.4
C4 5
C5 21
C6 0.0068

t .R
λ= (3)
v
16
p (λ, β) =
Cmax = 0.5930 (4)
27
   −C5
C p (A, β) = C1 CA2 − C3 β − C4 e A + C6 .λ
(5)
1
A
= λ+0.08.β
1
− 1+β
0.035
3

Ptur b 1 1
Ttur b = = .ϕ.S.C p (A, β).v 3 . (6)
t 2 t

Jtot = Jtur b + Jg (7)

dmec
Jtot = Tmec = Ttur b − Tem − f.mec (8)
dt
The parameters C1, C2, C3, C4, C5 and C6 are constants dependent on the turbine
geometry (Table 1) [7].
According to Fig. 2, the plot of the Cp follows the variation of the Beta angle
inversely and reaches its maximum 0.426, for λopt = 8 and B = 0°. Given that the
wind turbine can’t convert more than 59% of the kinetic energy contained in the wind
into mechanical energy according to Betz’s Law [7–9].

Fig. 2 Curve of Cp
1258 B. Majout et al.

Fig. 3 MPPT strategy without speed control

2.2 Extraction of Maximum Power (MPPT Control)

The MPPT control strategy consists in controlling the electromagnetic torque in

order to adjust the mechanical speed in a way to maximize the power of the incident
energy of the wind turbine (see Fig. 3) [8–11]. Because of erroneous measures of
the wind speed delivered by the anemometer, most wind turbines are controlled
without speed control. This technique is based on the hypothesis that the wind speed
varies very slowly in permanent regime compared to the electrical time constant
of the system [9]. In this case, the electromagnetic torque control is deduced from
the dynamic equation of the turbine [7–9]. By neglecting the torque due to viscous
friction (f.mec ) the equation of the reference electromagnetic torque becomes:
 
Ptur b−r e f ρ.π.R 5 .C p_max λopt .2mec
Tem_r e f = = (9)
t 2.λ3opt

2.3 PMSG Model System

The electrical equations of the PMSG in q-d reference are as follows (see Fig. 4) [10,
11]:

dsd
V sd = Rs .i sd + − wr .sq (10)
dt
dsq
V sq = Rs .i sq + + wr .sd (11)
dt

sd = L d .i sd + ∅ f ; sq = L q .i sq (12)

wr = p.mec (13)

The electromagnetic torque expression can be described as follows [7–11].

Improvement of the Wind Power System Based on the PMSG 1259

Fig. 4 d-q model of PMSG in synchronous reference frame

3   
Tem = . p L d − L q Isd .Isq + Isq .∅ f (14)
2
The active and reactive power are expressed by:
  
Pgen = 23 Vsd.isd + Vsq.isq 
(15)
Qgen = 23 Vsq.isd − Vsd.isq

3 Synthesis of Controls

In order to guarantee a high performance of a wind energy system based on the PMSG
with smooth poles, where the inductances Ld and Lq are equal L s = L d = L q , The
frequency and the voltage of the generator connected to the distribution grid must be
adjusted when the wind speed varies [7].
We use the following PMSG model [7–12].

d Isd 1  
I˙ds = =− Rs .i sd − p.mec .L q .i sq − V sd (16)
dt Ld
d Isq 1  
I˙qs = =− Rs .i sq + p.mec .L d .i sd + p.mec .∅ f − V sq (17)
dt Lq
1 3 p    f
˙ mec =
 .Ttur b − . . L d − L q Isd .Isq + Isq .∅ f − .mec (18)
Jtot 2 Jtot Jtot

The PMSG model represented by differential Eqs. (10) and (11) indicate that Id
and Iq are coupled together by nonlinear terms mec .L d .i sd and mec .L dq .i sq [7].
To solve this problem, two controls have been developed in this paper in order to
control the machine-side converter:
1) Field-Oriented Control (FOC).
2) Sliding Mode Control (SMC).
1260 B. Majout et al.

3.1 Field Oriented Control (FOC)

The basic principle of vector control was described by BALSCHKE and HASSE in
the early ’70s [10–16]. the Vector control consists in eliminating the coupling between
the inductor and the armature by decomposing the stator current into two parts (Isd,
Isq), where the inductive current controls the magnetic flux and the armature current
controls the electromagnetic torque. The objective of this command is to establish
a simple model of PMSG which makes it similar to a DC machine with separate
excitation [13, 14]. The vector control is generally based on Proportional-Integral
regulators PI to regulate the error between the measured input and desired output.
According to the compensation method, the expression of the reference voltages
after removing the terms of compensation −wr .Lq .isq and wr (Ld .isd + ∅f ) become:
[13–17]:

V sdref = (Rs + S.Ld )isd
(19)
V sqref = (Rs + S.Lq )isq

In order to decouple the electromagnetic torque and control it by Isq only, the
vector control imposes Id_ref = 0. the expression of the electromagnetic torque
becomes as follows [7, 13].

3 p
Tem = . .Isq .∅f (20)
2 Jtot

The control strategy FOC is represented in the Fig. 5 below:

3.2 Sliding Mode control (SMC)

The sliding mode control technique considered one of the special cases of the VSS
variable structure system, which is characterized by a choice of the structure and
the switching logic [16]. The evolution of this system depends only on properties

Fig. 5 Field Oriented Control of the PMSG.

Improvement of the Wind Power System Based on the PMSG 1261

of the sliding surface. The principle of this method consists in forcing the system
state trajectory to attain a hyper surface in finite time and then stay there, where the
behavior of results corresponds to the desired dynamics [14–18].
The implementation of this method of control requires three complementary steps
concerning [16–18]:
• Selecting the sliding surface.
• Defining the convergence conditions based on Lyapunov functions.
• Determining the strategy of the control.

The Sliding Mode applied to PMSG

Considering the sliding surface proposed by SLOTINE [16]:
 r−1
d
S(X) = +δ ∗ e(x) (21)
dt

With,
e(x): the error in the output state e(x) = xref – x.
δ: a positive coefficient.
r: the rth order system. Where for r = 1, S(X) = e(x).
For the order system r = 1, the sliding surface for the dq-axis current and
mechanical speed generator can be presented as follows [7–18]:

S(mec ) = e(mec ) = mecref − mec (22)

S(Isd ) = e(Isd ) = Isdref − Isd (23)

 
S Isq = e(Isq ) = Isq ref − Isq (24)

Where the derivative of the sliding surface is expressed by:

f 3 p 1
˙ mec−ref +
Ṡ(mec ) =  .mec + . .Isq .∅ f − .Ttur b (25)
Jtot 2 Jtot Jtot
Rs Lq V sd
Ṡ(Isd ) = I˙sd −r e f + .i sd − p.mec . .i sq − (26)
Ld Ld Ld
  Rs Ld ∅f V sq
Ṡ Isq = I˙sq −r e f + .i sq + p.mec . .i sd + p.mec . − (27)
Lq Lq Lq Lq

The Direct and Quadrature Stator Current Controller

In order to obtain a switching around the surface and a good dynamic performance,
the controller structure entails two parts: a discontinuous control VN according to
the sign of the sliding surface and an equivalent control Veq [14–18].
1262 B. Majout et al.

Vsd,q _r e f = Vsd,q _eq + Vsd,q _N (28)

The sliding mode occurs on the sliding surface when Ṡ(x) = 0 and Vsd,q _N = 0
[16–18], We can deduce the expression of the equivalent command Veq:


Rs Lq
Vsd _eq = L d I˙sd −r e f + .i sd − p.mec . .i sq (29)
Ld Ld


Rs Ld ∅f
Vsq _eq = L q I˙sq −r e f + .i sq + p.mec . .i sd + p.mec . (30)
Lq Lq Lq

While the discontinuous control is given by the following equations:

Vsd _N = K d .sat(S(Isd ))with K d > 0 (31)

  
Vsq _N = K q .sat S Isq with K q > 0 (32)

Speed Regulator
For the speed controller, the control variable becomes the reference current of the
quadrature axis.by Following the same method used for controlling the d and q axes,
the expression of Isq _ref is given as follows [7–18].

Isq _r e f = Isq _eq + Isq _N (33)

 
Isq _eq = − 23 . p.∅
Jtot
˙ mec−r e f + f
.mec − 1
.Ttur b
f Jtot Jtot (34)
Isq _N = K mec .sat(S(mec )) with K mec > 0

4 Simulation and Results

In order to evaluate the performance of both control strategies FOC and SMC, we
tested the behavior of the wind system based on PMSG in a dynamic regime, by
a test performed in MATLAB/Simulink, where the turbine is trained by a variable
wind speed using the MPPT technique to approach reality (see Fig. 6). The wind
system parameters (Turbine and PMSG) are given in Table 3.
The following figure illustrate the results of the PMSG simulation in the Matlab
/Simulink environment:
From curves shown in Fig. 7, we can notice that the sliding mode control offers
better results compared to vector control based on the classical PI regulators. On the
Improvement of the Wind Power System Based on the PMSG 1263

Fig. 6 Wind speed profile

other hand, the references fixed for both controls are well followed by the electro-
magnetic torque and the active and reactive powers as well as the stator currents. In
addition, the reactive stator power is null which gives a unit power factor.
The electromagnetic torque varies according to the variation of the mechanical
torque. The torque ripple for SMC control is strongly reduced compared to the torque
ripple for FOC control.
The control of the rotation speed is well ensured. The Zoom 1 of the rotational
speed shows clearly the quality of the SMC control compared to FOC on the mechan-
ical rotational speed as well as the rapidity and precision. We note also that the static
and dynamic errors are infinitely low in the case of SMC.
The good quality of the stator currents in the rotating frame of reference gener-
ates sinusoidal three-phase currents in good quality. However, the frequency of the
currents changes with the change of the mechanical rotation speed which is propor-
tional to the variation of the wind speed. The ripple rate of the currents for SMC
control is low than the rate found in the case of vector control.
The Table 2 shows the time when each response starts to follow its reference.

5 Conclusion

This paper presents a comparative analysis between SMC and FOC control applied
to the machine-side static converter for a variable speed wind turbine based on the
permanent magnet synchronous generator. In the first time, We presented the vector
control then the non-linear SMC control in the second time. According to the results
shown in Fig. 7 and Table 2, we notice that the vector control offers an acceptable
result, whereas the SMC control offers a high static and dynamic performances.
Therefore, the SMC control remains the most appropriate solution to be applied for
the control of the three-phase rectifier.
1264 B. Majout et al.

FOC Technique SMC Technique

Fig. 7 Simulation results of the Field-Oriented Control and Sliding mode control
Improvement of the Wind Power System Based on the PMSG 1265

Table 2 Summarization
FOC (Tr) (ms) SMC (Tr) (ms)
table of the response time
SMC and FOC Controller  20 ms 10 ms
Ps 20 ms 12 ms
Qs 22 ms 16 ms
Tem 50 ms 30 ms

Table 3 PMSG and Wind Turbine parameters

PMSG Parameters Symbol Values Turbine parameters Symbol Values
Power Generator Pn 1,5 MW Radius of the turbine R 55 m
blade
Pole number P 75 Turbine + generator Jtot 10000 N.m
moment
Stator Resistance Rs 6.25e−3  Specific density of air ρ 1.22 kg/m3
d axis inductance Ld 4.229e−3 H Tip-speed ratio λopt 8
q axis inductance Lq 4.229e−3 H Optimal power Cpmax 0.426
coefficient
Generator flux Øf 11.1464 Wb
Generator flux Øf 11.1464 Wb
Coefficient of friction f 0.0142 N.m.s/rad

Table 4 Nomenclature
Nomenclature
Pturb Power captured by the wind turbine S Blade swept area
Vw Wind speed R Radius of the turbine blade
mec Mechanical generator speed ωr Electric pulsation
t Mechanical turbine speed Pgen Active generator power
Tem Electromagnetic generator torque Qgen Reactive generator power
Tem_ref Reference generator torque Vs Stator voltage vector
Ttur Turbine torque Rs Stator resistance
Cp(λ, β) Power coefficient L(d,q) d-q axis inductance
λ Tip-speed ratio Vs(d,q) d-q axis stator voltage
β Pitch angle is(d,q) d-q axis stator current
ρ Specific density of air (d,q) d-q axis flux
p Number of poles pairs ∅f Generator flux

References

1. Ackermann T, Soder L (2002) An overview of wind energy-status 2002. Renew Sustain Energy
Rev 6(1e2):67e127
1266 B. Majout et al.

2. Allouhi A (2019) Energetic, exergetic, economic and environmental (4 E) assessment process

of wind power generation. J Clean Prod 235:123–137
3. Jelavi M, Petrovi V, Peri N (2018) Estimation based individual pitch control of wind turbine.
Automatika 51:181–192
4. Boualouch A, Essadki A, Nasser T, Boukhriss A, Frigui A (2015) Power control of DFIG
in WECS using backstipping and sliding mode controller. Int J Power Electron Drive Syst
(IJPEDS) 9:612–618
5. Dragomir G, Şerban A, Năstase G, Brezeanu AI (2016) Wind energy in Romania: a review
from 2009 to 2016. Renew Sustain Energy Rev 64:129–143
6. Gasch R, Twele J (eds) (2011) Wind power plants: fundamentals, design, construction and
operation. Springer, Heidelberg
7. El Mourabit Y, Derouich A, El ghzizal A, Bouchnaif J, El ouanjli N, Zamzoum O, Mezioui
K, Bossoufi B (2019) Implementation and validation of backstepping control for PMSG wind
turbine using dSPACE controller board. Energy Rep J 5: 807–821
8. BossoufiB, Ionita S, Alami Arroussi H, El Ghamrasni M, Ihedrane Y (2017) Managing voltage
drops a variable speed wind turbine connected to the grid. IJAAC Int J Autom Control 11(1),
15–34
9. Bossoufi B, Karim M, Lagrioui A, Taoussi M, Derouich A (2015) Observer backstepping
control of DFIG-generators for wind turbines variable-speed: FPGA-based implementation.
Renew Energy 81:903–917
10. Seyoum D, Grantham C (2003) Terminal voltage control of a wind turbine driven isolated
induction generator using stator oriented field control. IEEE Trans Ind Appl 846–852
11. Hong CM, Chen CH, Tu CS (2013) Maximum power point tracking-based control algorithm for
PMSG wind generation system without mechanical sensors. Energy Convers Manage 69:58–67
12. El Hammouchi F, El Menzhi L, Saad A, Ihedrane Y, Bossoufi B (2019) Wind turbine doubly-
fed asynchronous machine diagnosis defects using stator and rotor currents lissajous curves.
Int J Power Electron Drive Syst (IJPEDS) 10(2):961–971
13. Blaschke F (1972) The principle of field orientation as applied to the new transvector closed-
loop system for rotating-field machines. Siemens Rev 34(3):217–220
14. Taoussi M, Karim M, Hammoumi D, Elbakkali C, Bossoufi B, El Ouanjli N (2017) Comparative
study between Backstepping adaptive and Field-oriented control of the DFIG applied to wind
turbines. Process IEEE Xplore 23
15. Slotine JJ, Li W (1991) Applied nonlinear control. Prentice-Hall, Englewood Cliffs, NJ. 461,
Google Scholar (1998)
16. Ihedrane Y, El-Bekkali C, Bossoufi B (2018) Improved performance of DFIG-generators for
wind turbines variable-speed. Int J Power Electron Drive Syst (IJPEDS) 9(4):1875–1890
17. Fantino R, Solsona J, Busada C (2016) Nonlinear observer-based control for PMSG wind
turbine. Energy 113:248–257
18. Majout B, Abrahmi D, Ihedrane YE, Bakkali C, Mohammed K, Bossoufi B (2020) Improvement
of sliding mode power control applied to wind power generation system based on DFIG. Int J
Power Electron Drive Syst (IJPEDS) 11(2):580–593
Control of Wind Water Pumping Using
Input-Output Feedback Linearization
Technique

Atarsia Loubna, Toufouti Riad, and Meziane Salima

Abstract To solve the problem of water for irrigation in isolated regions and rural
remote areas, where the expansion of the power network line is complex and she is
more costly. In this paper a control of the wind pumping systems is presented. Clas-
sical control strategy of induction motors (IMs) is very sensitive to the variation of
rotor resistance, to leads a strong coupling between the rotor flux and the electromag-
netic torque. To resolve this difficulty, we propose in this paper, a nonlinear control
of high performance applied of induction motor, to improve the tracking control, for
guarantee stability, robustness to parameter variations, perfect decoupling between
the flux and torque. The robustness of the proposed study is confirmed through
software simulation with parameter variations.

Keywords Wind energy · Induction motor · PMSG · Input output linearization ·

Centrifugal hydraulic pump · Water pumping system

1 Introduction

Water supply, livestock and domestic necessities are main issues in the progress of
isolated and Saharan regions, where the extension of the conventional power network
remains very hard to do [1, 2]. Therefore, recently it is necessary to use inexhaustible
resources to fulfill their energy requirements. [2, 3].
The demand for rural water for domestic supplies and irrigation is increasing, with
the result that surface water become increasingly scarce in the future [4, 8]. For this,
it seems that the only and good alternative to this problem is groundwater. but the
water table also decreases, making manual pumping more challenging. However, the
pumping of water mechanized will be the only dependable substitute for lifting water

A. Loubna (B)
Department of Electrical Engineering, Oum El Bouaghi University, Oum El Bouaghi, Algeria
T. Riad · M. Salima
Department of Electrical Engineering, Laboratory of Electrical and Renewable Energies LEER,
Souk Ahras University, Souk Ahras, Algeria

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1267
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_116
1268 A. Loubna et al.

from the ground. Even so, water pumping in the world usually depends on conven-
tional electricity produced by electricity or diesel. A photovoltaic and wind pumping
system has appeared as a satisfactory alternative to classical pumping systems used
in isolated regions and rural remote areas. Pumping water through wind energy mini-
mizes costly fuels, noise and air pollution [9]. The water pumping systems with wind
turbines are generally composed of a wind turbine, an electric generator, one or more
converters and a motor generally driven the centrifugal pump load [7].
A squirrel cage induction generator (SCIG), and DC Motor the Permanent Magnet
Synchronous Generator (PMSG), were generally most used in the water pumping
[7–10]. the permanent magnet synchronous generator (PMSG) has been gaining
popularity due to its high mass torque efficiency with low inertia ratio, its simple
of control, recently are good competitor of the induction motor in a wind energy
conversion [11, 12].
For this, the research has oriented to other solutions involving less expensive and
more robust actuators, namely the Induction machine, more reliable and less costly
in terms of construction and maintenance [14, 15].
Induction machine is careful as a distinctive actuator in the all industries for vari-
able speed applications since it has many advantages such as reliability, simplicity
of design, high efficiency, good self-starting, the absence of the collector brooms
system, low maintenance, and relatively reasonable budget. [2, 16], However, it
dynamic model is nonlinear and there control is very complex, due the high coupling
between the input variables, output and internal variables [16–19]. But the advance-
ment of power electronics, made a fundamental revolution which its m main objective
is to develop the recent control strategies for IM. Several control strategies have been
developed and used for IM drives such as: Direct Torque and Flux Control (DTFC),
Vector Control (VC) [13, 16].
However, this technique demands the use of a magnetic sensor based on Hall
Effect, to measure the magnitude of the rotor flux, this sensor remains very sensitive
to motor parameters variations and raises the cost of the drive equipment [16, 21, 22].
To solve this problem involved in a vector control, many researchers presented
new control strategy using the rotor flux and speed estimation without any sensor
in order to decrease the expenses and have the features of precision, elimination of
disturbance and fast torque response [22]. One among nonlinear control methods is
the nonlinear control with input output linearization, based on the transformation
of the nonlinear model into several linear system; which makes the linear feedback
control easy to be applied in respect to trajectory reference [16, 22, 23], The input
output linearization control based on differential geometry theory is used successfully
to eliminated coupling problem, resulting in a rapid transient response with decoupled
flux and torque response [16]. In recent decades, the theory of feedback control of
nonlinear seen significant developments in various industrial areas and has now
the axis of several research status, it is based on theoretical concepts of differential
geometry such that the Lie derivative and deffeomorphism, it allows diffeomorphic by
a nonlinear transformation and an uncoupled state feedback to linearize the template
and placed in the canonical form [6, 19]. In this background several research paper
have been published in various research laboratory.
Control of Wind Water Pumping … 1269

In [25] the authors present a control technique of water pumping system based on
wind turbine generator in an isolated area (in South of Algeria), Simulation results
have confirmed satisfactory performance and verified the efficiency of the suggested
pumping system. In [26] The authors provided control of a water pumping system
based on a wind turbine generator in an isolated area by using the Maximum Power
Tracking Point (MPPT), The wind energy conversion system is used to drive the
permanent magnet synchronous generator (PMSG) to feed the isolated load which
is consisted of DC motor drives a hydraulic centrifuge pump. In [27] The authors
applied the vector control of a wind power conversion system based on the doubly fed
induction machine (DFIM) connected to the grid. In [28] presented an experimental
study on the efficient field oriented control with power losses for the optimization
of the energy production of a wind energy conversion chain based on the six phase
asynchronous generator. In [29] the authors used the scalar command for controlling
a PV water pumping system, consists in driving the machine with a constant flow for
variable voltages, the simulation results showed the effectiveness and feasibility of
the suggested method. In [2] the authors presented an robust control technique based
input-output linearization control of IM coupled with hydraulic pump powered by a
PV panel. Simulation results have confirmed good performance of the proposed PV
Water pumping system.
To solve the problems of high coupling between the flux and torque in vector
control strategy and to improve the tracking control, for guarantee stability, robust-
ness to parameter variations, perfect decoupling between the flux and torque, we
propose in this paper, a nonlinear control of high performance applied of (IM). In
the proposed system, the (PMSG) is wind turbine driven. The centrifugal pump is
driven by an IM and fed by a voltage-source inverter (VSI).This article is presented as
follows; the description of wind pumping system is evolved in Sect. 2. In Sect. 3 feed-
back linearization control of IM will be presented. In Sect. 4 the performance of the
proposed control of the water pumping system is verified by numerical simulation.
Finally, some conclusions are given in Sect. 5.
The robustness of the proposed control system is verified through numerical
simulation with motor parameters variations (Figs. 2, 3, 4, 5, 6 and 7).

2 Description of Wind Pumping System

2.1 Wind Turbine

The wind turbine captures the kinetic energy of the wind, passing through the section
of the active surface swept by the turbine blades, and converts it into mechanical
energy (torque) which turns the rotor of an electric generator [26, 30]. The mechanical
power of the wind turbine is given by the following expression [4, 31]:
1270 A. Loubna et al.

Fig. 1 Representation of the wind pumping system

1
Ptur = ρ SV 3 C p (λ) (1)
2

Where: Ptur : is the extracted power, ρ: is the air density [kg/m3 ], V: is the wind
speed [m/s], S: is the area swept by the rotor blades, C p : is called the power coefficient.
λ is tip speed ratio of rotor blade which equals a tip speed to a wind speed is
written as follows [32, 33].

t R
λ= (2)
V
R: is the radius, t : is the mechanical speed of the wind turbine rotor.

2.2 Permanent Magnet Synchronous Generation Model

In the (d, q) axis the model of the PMSG used in the modeling model is given by the
following equations [4, 11, 26]:
      
ψd Ld 0 id ψv
= + (3)
ψq 0 Lq iq 0

Vd = Rs .i d + dtd ψd + ωr ψq
(4)
Vq = Rs .i q + dtd ψq − ωr ψd
Control of Wind Water Pumping … 1271

The equation of mechanical speed is given by

dr
Tem − Tr = J + f.r (5)
dt
The equation of electromagnetic torque is expressed by (6) [11]:

3p    
Tem = ψ f iq + L d − L q id iq (6)
2
ψ f : Rotor Flux.

2.3 Modeling of Motor Pump System

• Model of IM. The model of the IM in the stationary frame (α, β) is given by (7)
[2, 21, 34]:

x = f (x) + g(x).u
(7)
y1 = h 1 (x); y2 = h 2 (x)

Where
⎡ ⎤
−γ Isα + TKr φr α + pK φrβ
⎢ −γ Isβ − pK φr α + T r φrβ
K ⎥  T  t
⎢ ⎥ 1
⎢ ⎥ g1 0 000
f (x) = ⎢ T r Isα − T r φr α − pφrβ
M 1
⎥ [g] = = σ LS
⎢ ⎥ g2 0 1
σ LS 0 0 0
⎣ T r Isβ + pφr α − T r φrβ
M 1 ⎦
p J L r (φr α Isβ − φrβ Isα ) − J (TL + f )
M 1

 
x = [Isα Isβ φr α φrβ ]T : u = [u sα u sβ ]T , y =  (φr2α + φrβ
2
)

Lr M2 M Rs Rr M 2
Tr = ; σ =1− ;K = ;γ = +
Rr Ls Lr σ Ls Lr σ Ls σ L s L r2

The expression of electromagnetic torque and dynamical speed can be expressed

as follows [20]:

PM
Tem = (φr d Isq − φrq Ir d ) (8)
Lr
dr
J = Tem − TL − f r (9)
dt
1272 A. Loubna et al.

Model of Centrifugal Pump. The water pump is an essential element in water

pumping systems [2] centrifugal pump is mostly commonly used due there
advantages such as: simple working, supple modular and high efficiency [5, 6, 34].

H pump = k0 + k1 Q + k2 Q 2 (10)

Where: k 0 , k 1 , k 2 : are the constant of the pump, H[m]: the head of the pump and
Q [L/s]: the flow rate of the pump [35, 36].The IM is coupled to centrifugal pump,
by the following mechanical equation [2, 2]:

d
J = Tem − T pump (11)
dt
The torque- speed relation of this pump is given by.

T pump = k T 0 + k T 1  + k T 2 2 (12)

Where, k T0 , k T1 , k T2 are the empirical constants of the pump which can be obtained
from the measurements.

3 Input-Output Linearization Technique

The difficult case for the principle of the linearization input-output decoupling is
the choice of output variable yi. For the application of linearization methods state
feedback to control the asynchronous machine, there are two cases where we can
[15, 16]. The system to be controlled by a control law linearization must be square
type. As a result, we chose to rotor speed and magnitude of rotor flux as outputs of
the process [15].
⎡ ⎤
  1 2   2 
h 1 (x) φ
⎢ 2 rα + φ 2
rβ ⎥ φr α + φrβ
2
y= =⎢
⎣ pM 
⎥
⎦ = (13)
h 2 (x) Ce
Isβ φr α − Isα φrβ
Lr
   
h 1 (x) 
y= =  2  (14)
h 2 (x) φr α + φrβ2

To obtain the optimal non-linear law control, we determine the relative degree
of the output y1 and y2 , that is to say the number of times to derive the output to
show the input Vsα Vsβ .changes in output are selected given by the Eq. (14) such that:
[15, 15].
Control of Wind Water Pumping … 1273
⎧
⎨ h 1 (x) = y1 = 
  (15)
⎩ h 2 (x) = y2 = 1 φr2α + φrβ
2
2
Using the Lie derivatives to express the machine model by the relation input-output
[2, 23].
• Rotor speed y1 = 
⎧
⎪
⎪ y1 = L 0f h 1 (x) = h 1 (x) = 
⎪
⎨ pM   fr Cr
ẏ1 = L f h 1 (x) = φr α I Sβ − φrβ I Sα −  − (16)
⎪
⎪ J Lr J J
⎪
⎩
ÿ1 = L 2f h 1 (x) + L g1 L f h 1 (x).u sα + L g2 L f h 1 (x).u sβ

• Rotor flux y1 = φr2

⎧
⎪ y = L 0f h 2 (x) = h 2 (x) = φr2 = φr2α + φrβ 2
⎨ 2
ẏ2 = ḣ 2 (x) = L f h 2 (x) (17)
⎪
⎩
ÿ2 = ḧ 2 (x) = L 2f h 2 (x) + L g1 L f h 2 (x).u sα + L g2 L f h 2 (x).u sβ

Where
⎡  
 ⎤
Rs 1−σ 1 fr 
− + + + I Sβ .φr α − I Sα .φrβ −
pM ⎢ ⎢ σ Ls σ Tr Tr J ⎥ 2
⎥ + f  + f Cr
L 2f h 1 (x) = ⎣     ⎦ 2
J Lr pM. 2 + φ 2 − p I φ + I φ J J2
 . φrβ α sα rβ sβ r α
σ L s Lr r

PM PM
L g1 L f h 1 (X ) == − φrβ ; L g2 L f h 1 (X ) = − φr α
J σ Ls Lr J σ Ls Lr
   
2 Rs 1−σ 3M   2M 2 4  2 
L 2f h 2 (x) = − M + + Isα φrβ + Isβ φrα + + 2 φrβ + φrα
2
Tr σ Ls σ Tr Tr σ Tr L s L r
2 Tr
2pM   2M 2  2 
+ Isα φrβ − Isβ φrα + 2 Isα + Isβ
2
Tr Tr

2M 2M
L g1 L f h 2 (x) = φrβ ; L g2 L f h 2 (x) = φr α
σ L s Tr σ L s Tr

The total relative degree r = r 1 + r 2 = 4 < degree system n = 5, a one not

observable dynamics of order 1 is obtained. [15, 16] to solve this problem we choice
the variable (Z5 ) of rotor angle flux corresponding to the same approach [24] for this
the new coordinates and there derivative as follows:
1274 A. Loubna et al.

⎧
⎪ z 1 = y1 = h 1 (x) ⎧
⎪
⎪ = ẏ1 = L f h 1 (x)
⎪
⎪ z 2 = L f h 1 (x) ⎪
⎪ ż 1
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ = ÿ1 = L 2f h 1 (x) + L g L f h 1 (x)Vsα + L g L f h 1 (x)Vsβ
⎪ ⎪ ż 2
⎨ z 3 = y2 = h 2 (x)
⎪ ⎪
⎨
z 4 = L f h 2 (x) Derivative ⇒ ż 3 = ẏ2 = L f h 2 (x)
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ ż = ÿ2 = L 2f h 2 (x) + L g L f h 2 (x)Vsα + L g L f h 2 (x)Vsβ
⎪
⎪   ⎪ 4
⎪
⎪
⎪ φ ⎩
⎪
⎩ z 5 = y3 = tan −1 rβ ż 5 = ẏ3 = L f h 3 (x)
φrα
(18)

Using equation system (18). Derived outputs are given by:

     
ÿ1 (x) L 2f h 1 (x) Vsα
= + D(x). (19)
ÿ2 (x) L 2f h 2 (x) Vsβ

The determinant of D(x) is non-zero except when the engine is stopped; thus the
matrix D(x) is reversible. The optimal control is given by Eq. (20) [16, 24]:
   
Vsα −1
−L 2f h 1 (x) + v1
= [D(x)] . (20)
Vsβ −L 2f h 2 (x) + v2

If the determinant of the decoupling matrix is not zero, the non-linear control law
is defined by a relation that connects the new internal inputs (v1 , v2 ) to the physical
inputs (us1 , us2 ) (usα , usβ ) [2, 18, 24]. In order to further a reference trajectory *
and the flux φ*r 2 these variations v1 and v2 are calculated as follows:
⎧  
⎪   pM   Cr
⎪ ∗ ˙∗ ¨
⎨ v1 = −k11  −  − k12 J L φrα Isβ − φrβ Isα − J −  + ∗
⎪
r
    (21)
⎪
⎪
⎪ 2    2 
⎩ v2 = −k21 φrα
2
+ φrβ2
− φr∗2 + k22 M φrα Isβ + φrβ Isα − φrα + φrβ
2
− φr∗2 + φr∗2
Tr

The coefficients k ij are chosen such that the two polynomials s2 + sk 11 + k 12 and
s + sk 21 + k 12 have roots at negative real part.
2

4 Simulation Results and Discussions

In order to confirm the performances of the proposed control, the numerical simu-
lation was presented on a water pumping system described in section II, composed:
Wind turbine, PMSG, Rectifier, Inverter and a centrifugal pump driven by IM. The
induction motor is controlled by Input-Output Linearization Control (IOLC).the
simulation results are presented by Figs. 2, 3, 4, 5, 6 and 7.
From the simulation results we notice that the Fig. 2 shows the actual and reference
rotor speed. The rotor speed tracks quickly the reference speed.
Control of Wind Water Pumping … 1275

Fig. 2 Rotor speed of IM 100

Wr
50 Wref

Wr Wref [Rad/s]
0

-50

-100

0 0.5 1 1.5 2 2.5 3 3.5 4

Time [S]

1.4
Fig. 3 Rotor flux of IM Phir
Phiref
1.2

Phir Phiref [Wb]

0.8

0.6

0.4

0.2

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [s]

30
Fig. 4 Torque pump and Te
Torque of IM Tpump
20

10
Te Tpump [N.m]

-10

-20

-30
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [S]

40
Fig. 5 Stator current of IM
30

20
Isa [A]

10

-10

-20
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [S]
1276 A. Loubna et al.

Fig. 6 Flow of pump 2.5

1.5

Flow [m3/s]
1

0.5

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [s]

Fig. 7 Total monomer head 35

speed 30

25

20
Head [m]

15

10

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [s]

Figure 3 shows manifests that the quick response of the rotor flux in the transient
state and in steady state the rotor flux is very tracking to the reference flux and is
not affected by resistances variation. We show clearly that the decoupling control
between torque and flux.
Figure 4 shows good torques responses in steady state. It can be observed that the
motor torque is not affected by the rotor speed change and is tracks with very fast
dynamic the load torque of centrifugal pump.
show the wave form that phase stator current is sinusoidal with in acceptable
limits of nominal current of the IM, in transient state the stator current has the peak
current due low values of speed.
See Fig. 6. It is also noted that the flow of the pump are not disturbed by the vari-
ations of the wind speed. The monomeric height is well maintained for the variation
of the speed rotor see Fig. 7.
So from these results we find that the dynamics of flux and the torque is stably
maintained for the different operating regimes, which show that the decoupling is
perfect and assured by this technique. It can be concluded that this technique is
robust and guarantee a good insensitivity to parametric variations of the machine. In
addition, it has the advantage of being simple to implement in practice.
Control of Wind Water Pumping … 1277

5 Conclusion

We have presented in this article, the application of control theory by input output
linearization for a wind turbine water pumping system, shows that the good regulation
of the rotor flux and speed of IM by eliminating system’s nonlinearity and coupling
effect control.
Simulation results have shown the excellent performances of the proposed IOLC
controller. It can be inferred that this system is an ideal solution for water supply in
sparsely populated, isolated and enclosed areas.

References

1. Chikh A, Chandra A (2009) Optimization and control of a photovoltaic powered water pumping
system. In: IEEE Electrical Power & Energy Conference (EPEC), pp 1–6 (2009)
2. Meziane S, Atarsia L, Toufouti. R (2018) A global stability of linearizing control of induction
motor for PV water pumping application. Int J Syst Dyn Appl (IJSDA), 7:31–56
3. Camocardi P, Battaiotto P, Mantz R (2008) Wind generator with double stator induction
machine. Control strategy for a water pumping application.In: 43rd international universities
power engineering conference, pp 1–5 (2008)
4. Rahrah K, Rekioua D (2015) Economic study of a solar/wind water pumping system with
battery storage installed on the site of Bejaia. In: 3rd International renewable and sustainable
energy conference (IRSEC), pp 1–6
5. Bouzeria H, Fetha C, Bahi T, Abadlia I, Layate Z, Lekhchine S (2015) Fuzzy logic space
vector direct torque control of PMSM for photovoltaic water pumping system. Energy Procedia
74:760–771
6. Jaziri S, Jemli K (2013) Optimization of a photovoltaic powered water pumping system. In:
International conference on control, decision and information technologies (CoDIT), pp 422–
428
7. Hammadi S, Hidouri N, Sbita L (2011) A DTC-PMSG-PMSM drive scheme for an isolated
wind turbine water pumping system. Int J Res Rev Electr Comput Eng (IJRRECE) 1:1–6
8. Girma M, Molina M, Assefa A (2015) Feasibility study of a wind powered water pumping
system for rural Ethiopia (2015)
9. Chandel S, Naik MN, Chandel R (2015) Review of solar photovoltaic water pumping system
technology for irrigation and community drinking water supplies. Renew Sustain Energy Rev
49:1084–1099
10. Lebsir A, Bentounsi A, Benbouzid M, Mangel H (2015) Electric generators fitted to wind
turbine systems: an up-to-date comparative study
11. Mahersi E, Kheder A (2016) Adaptive backstepping control applied to wind PMSG system.
In: 7th International Renewable Energy Congress (IREC), pp 1–6
12. Tahour A, Aissaoui A, Abid M, Essounbouli N, Megherbi AC (2010) La commande de la
puissance active et réactive d’une éolienne à génératrice synchrone. In: Revue des Energies
Renouvelables SMEE 2010 Bou Ismail Tipaza, pp 327–335
13. Atarsia L, Oukaci A, Toufouti R, Meziane S (2015) Direct torque control of induction motor
and regulation speed using the adaptive control. In: International conference on automatic
control, telecommunications and signals (ICATS15), pp 1–7
14. Meziane S, Toufouti R, Benalla H (2008) Generalized nonlinear predictive control of induction
motor. Int Rev Autom Control 1:65–71
1278 A. Loubna et al.

15. Toufouti R, Meziane S, Feddaoui O, Saidia M (2016) A robust decoupling control of induc-
tion motors based adaptive rotor flux observer. In: 4th International Conference on Control
Engineering & Information Technology (CEIT 2016), Hammamet, Tunisia, 16–18 December
2016
16. Oukaci A, Toufouti R, Dib D, Atarsia L (2017) Comparison performance between sliding mode
control and nonlinear control, application to induction motor. Electr Eng 99:33–45
17. Tarbouchi M, Le Huy H (1998) Nonlinear control of an induction motor in magnetic saturation.
In: Conference record IEEE industry applications conference. Thirty-third IAS, pp 648–654
18. Pietrzak-David M, De Fornel B, Purwoadi M (1998) Nonlinear control for sensorless induction
motor drives. In: International conference on power electronic drives and energy systems for
industrial growth, Proceedings, pp 300–306
19. Bendaha Y, Mazari B (2007) Commande Adaptative Linéarisanted’un Moteur Asynchrone
20. Blaschke F (1972) The principle of field orientation as applied to the new transvector closed-
loop system for rotating-field machines. Siemens Rev 34:217–220
21. Casadei D, Profumo F, Serra G, Tani A (2002) FOC and DTC: two viable schemes for induction
motors torque control. IEEE Trans Power Electron 17:779–787
22. Zhu J, Xu B, Wang X, Feng H, Xu X (2013) The research of sensorless vector control for
permanent magnet linear synchronous motor. J Comput 8:1184–1191
23. Enev S (2007) Input-output decoupling control of induction motors with rotor resistance and
load torque identification. In: Mediterranean conference on control & automation, pp 1–5
24. Marino R, Peresada S, Tomei P (1999) Global adaptive output feedback control of induction
motors with uncertain rotor resistance. IEEE Trans Autom Control 44:967–983
25. Harrouz A, Dahbi A, Harrouz O, Benatiallah A (2014) Control of wind turbine based of PMSG
connected to water pumping system in South of Algeria. In: 3rd international symposium on
environmental friendly energies and applications, pp 1–4 (2014)
26. Dahbi A, Harrouz A. MPPT control of wind turbine for water pumping system. Relation 7:0.1
27. Charles CR, Vinod V, Jacob A (2015) Field oriented control of Doubly Fed Induction Generator
in wind power system. In: IEEE international conference on computational intelligence and
computing research (ICCIC), pp 1–5
28. Pantea A, Sivert A, Yazidi A, Betin F, Carriere S, Capolino G-A (2016) Efficient field oriented
control with power losses optimisation of a six-phase induction generator for wind turbines.
In: IECON 42nd annual conference of the IEEE industrial electronics society, pp 1912–1917
29. Rebei N, Benghanem B, Hmidet A, Hasnaoui O (2013) Study of photovoltaic water pumping
system using scalar-DVC based control. In: International conference on electrical engineering
and software applications, pp 1–8
30. Guettaf A, Bettka A, Bennis O, Benchabane F, Khaled Y, Ali A (2011) Direct field oriented
control of induction motor fed by wind turbine generator under saturation effect
31. Toual B, Mokrani L, Kouzou A, Machmoum M (2018) Control and management of a solar-wind
hybrid system for power quality improvement. J Eng Sci Technol 13:1420–1439
32. Guettaf A (2013) Optimisation floue-génétique d’un système de pompage éolien.Université
Mohamed Khider–Biskra
33. Zhang J, Qi Z, Wen J (2017) Control of water pumping driven by wind turbine based on Quasi
Z source inverter. In: Chinese automation congress (CAC), pp 7074–7077
34. Atasrsia L, Toufouti R, Meziane S (2020) Standalone photovoltaic array fed induction motor
driven water pumping system. Int J Electr Comput Eng (2088–8708) 10
35. Belarbi M (2006) Modeling and simulation of PV water pumping system. Master thesis Oran
university
36. Karampuri R, Jain S, Somasekhar V (2014) A single-stage solar PV power fed open-end
winding induction motor pump drive with MPPT. In: IEEE international conference on power
electronics, drives and energy systems (PEDES), pp 1–6
Improved Hybrid Control Strategy
of the Doubly-Fed Induction Generator
Under a Real Wind Profile

Mourad Yessef, Badre Bossoufi, Mohammed Taoussi, Ahmed Lagrioui,

and Hamid Chojaa

Abstract The design of a control technique of doubly-fed induction generator

(DFIG) for providing a high quality of energy, without harmonics accumulation,
to the electrical grid is a real challenge because of the nonlinearity of the wind
energy conversion system (WECS). In this paper, an improved robust command has
been developed and then tested in order to control the rotor side converter (RSC) of
the WECS. The main contribution of this research work is to optimize the extraction
of the wind energy in the real circumstances and to push the DFIG working prop-
erly, with the highest performance and robustness, in the different operation modes.
Therefore, the dynamic model of the turbine has been presented in this research
paper, also the hybrid Direct Power Control-Backstepping (DPC-BS) command is
highlighted. Matlab/Simulink simulations analysis, with the real parameters of the
turbine and the real wind profile of Dakhla-Morocco city, confirm the high accuracy,
the high performance and the robustness of the proposed method, which based on
the the combination between the Direct Power Control (DPC) and the Backstepping
controller. Moreover, the obtained results prove, despite the variable wind speed, the
validation of the developed command with a total harmonic distortion THD ~0.33%
and a null overshoot for the grid currents with a frequency of 50 Hz, which gives the
possibility of injection into the electrical grid.

Keywords Backstepping command · DFIG · Direct Power Control · WECS

M. Yessef (B) · B. Bossoufi

Laboratory of Engineering, Modeling and Systems Analysis, SMBA University, Fez, Morocco
e-mail: [email protected]
M. Taoussi · H. Chojaa
Laboratory of Technologies and Industrial Services, Higher School of Technology, SMBA
University, Fez, Morocco
A. Lagrioui
Department of Electrical and Computer Engineering, Higher National School of Arts and Trades,
Moulay Ismail University, Meknes, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1279
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_117
1280 M. Yessef et al.

1 Introduction

All over the planet, the demand of electrical energy is increased more and more.
However, the major energy resources of electricity generation are natural gas, fossil
fuels, petroleum [1]. Nowadays, the excessive consumption of these fossil sources
causes a serious environmental and economic problems; namely, on the environ-
mental side, planet temperature increases and global warming, water pollution, the
emission of CO2 and other toxic gases, and acid rain. For the economic impact, it
consists of the increase in the price of petrol, oil and gas, which influences the elec-
tricity tariff and, subsequently, on all economic, agricultural and service activities.
Generally, this huge use has a dangerous effects on humans and animals.
Therefore, the renewable energy resources are revealed to be the serious alter-
natives to the polluting energies [2]. There are many kinds of renewable energy
resources, that are exploited all over the world (e.g. solar, geothermal, biomass,
hydropower and wind) [3]. In the last decades, the installation of wind farms has
been developed and received a great support from the most governments, thanks to
their profitability and cleanliness. Then, to face these urgent energy challenges, the
wind energy becomes usual more and more in many countries with global power
reaching over 600 GW and it target to be 1600 GW in 2030 [Word Wind Energy
Association]. So, the optimization of the operation of the wind energy conversion
system (WECS) has become essential, given their large share of the global energy
production.
Thereby, due to their several advantages, the DFIG (Doubly-Fed Induction Gener-
ator) technology with variable speed is presented to be the important way to develop
turbine generators within other competing machines. The DFIG is characterized by
its ability to work over a wide range of wind speeds and it performs an optimal
conversion of aerodynamic power at a low cost [4]. The Kinetic power of the wind
is transformed to mechanic power by the turbine and then delivered to the electrical
machine by the gearbox in the transmission machine. The DFIG stator is directly
linked to the electrical grid, at a time, the rotor is connected by an electronic power
converter. Since the stator of the DFIG is connected directly to the electrical grid,
when high peaks of stator currents occur at the stator windings, because of the estab-
lishment of a high voltage drop, overcurrents occur at the rotor windings and then on
the Rotor Side Converter (RSC), due to the coupling between the stator and rotor [5,
6]. For this reason the DC bus voltage increases and the torque fluctuations would
take place. Therefore, in the absence of the protection circuit, the high transient
currents will damage the RSC and the DFIG [5, 7].
Parallel to the wind farms development, several algorithms and structures of
control are then proposed to confident the stability, the performance and the robust-
ness of aero-generator of WECS such as the PI (Proportional Integral) controllers
and the adaptive PI [9, 10]. Other works focused on the other command types such
as the scalar, the adaptive fuzzy logic controller, the direct power control (DPC) [5],
Improved Hybrid Control Strategy ... 1281

the sliding mode control (SMC) [11, 12], the predictive control and the lineariza-
tion control. Unfortunately, the simulations with real wind and the testing of several
commands on the same system in the same environment are lacking, in this research
fields.
In this work, an improved command has been developed, which combines between
the Direct Power Control and the Backstepping command. The utilization of real
wind conditions from a wind farm located in south of Morocco is another innovative
aspect meted in this research paper.
This research paper is presented as follow; firstly, after the introduction, the turbine
is modeled and the dynamical model of the DFIG is presented, then the general
structure of the hybrid DPC-BS command is approached. The last section for this
paper will be reserved for obtained simulation results of performance and THD tests.
All results are reported and have been analyzed to prove the validity, of the proposed
command, to be implemented in a real scenario.

2 Methods

2.1 WECS General Model

The Fig. 1 shows the structure of the studied WECS. This system is composed of a
wind turbine driving a DFIG connected to a three-phase electrical grid. This model
is characterized by the fact that the stator is directly coupled to the grid, whereas,
in order to establish a power exchange, between the machine and the grid, in sub-
synchronous speed, two converters are installed between the electrical grid and the
rotor. As the active and reactive powers of DFIG are controlled through the RSC, the
GSC operates as a rectifier [5, 7].

Fig. 1 General model of the WECS based on DFIG

1282 M. Yessef et al.

2.2 Wind Turbine Model

The wind energy is captured by the wind turbine and converted into a torque or speed.
This process is modeled by the following equations [7]:
⎧
⎪
⎪ PRotor = Pwind .C p (λ, β) ⎧
⎪
⎨ ⎨ PRotor = 0.5ρ.A.C p (λ, β).v
3
Pwind = 0.5ρ Av 3
⇒ 1 (1)
⎪
⎪ PRotor ⎩ TRotor = .ρ.S.C p (λ, β).v 3
⎪
⎩ TRotor = 2t
t
⎧
⎨ C (λ, β) = k .( k2 − k .β − k ).e( −kA5 ) + k .λ
p 1 3 4 6
where: A
⎩
k1 = 0.5872, k2 = 116, k3 = 0.4, k4 = 5, k5 = 21, k6 = 0.0085
At the rotor of the turbine, the expression of the mechanical torque is [13]:
⎧
⎨T dmec
mec =J dmec
dt ⇒J = Tg − Tem − f mec (2)
⎩ dt
Tmec = Tg − Tem − f mec

g = G.t
where the gearbox equations are given by:
TRotor = G.Tg

2.3 Dynamic Model of DFIG Wind Energy System

In the WECS, the mechanical power is transferred to electrical power by the DFIG.
In the (d,q) reference frame, the electrical equations of the stator and the rotor which
model the DFIG are given by [7]:
⎧ ⎧
⎪ dϕsd ⎪ dϕr d
⎪
⎪ V = R .I + − ω .ϕ ⎪
⎪ Vr d = Rr .Ir d + − ωr .ϕrq
⎪
⎪
sd s sd
dt
s sq
⎪
⎪ dt
⎪
⎪ ⎪
⎪
⎪
⎪ dϕsq ⎪
⎪ dϕrq
⎪
⎨ Vsq = Rs .Isq + + ωs .ϕsd ⎪
⎨ Vrq = Rr .Irq + + ωr .ϕr d
dt dt
and (3)
⎪
⎪ 1 Msr ⎪
⎪ 1 Msr
⎪
⎪ Isd = .ϕsd − .ϕr d ⎪
⎪ Ir d = .ϕr d − .ϕsd
⎪
⎪ σ.L σ.L ⎪
⎪ σ.L σ.L r .L s
⎪
⎪
s r
⎪
⎪
r
⎪
⎪ 1 Msr ⎪
⎪ 1 Msr
⎩ Isd = .ϕsd − .ϕr d ⎩ Irq = .ϕrq − .ϕsq
σ.L s σ.L s .L r σ.L r σ.L r .L s

The electromechanical torque, in the rotor, is modeled by the equation [7]:

Tem = p(ϕr d .ϕsq − ϕrq .ϕsd ) (4)

Improved Hybrid Control Strategy ... 1283
⎧
⎪ ϕr d = L r .Ir d + Msr .Isd
⎪
⎪
⎨ϕ = L r .Irq + Msr .Isq
rq
where:
⎪
⎪ ϕ = L s .Isd + Msr .Ir d
⎪
⎩
sd
ϕsq = L s .Isq + Msr .Irq
The equations which model the active and reactive powers are:

Ps = Vsd Isd + Vsq Isq
(5)
Q s = Vsq Isd − Vsd Isq

2.4 Hybrid Control Strategy of DFIG (DPC-BS)

In 1991, Kokotovic proposed the Backstepping algorithm, which leads to a global

and asymptotic stabilization of the system if and only if each Lyapunov function
is developed to converge in a function of the Lyapunov stability theory [5]. The
basic process of the Backstepping algorithm is the decomposition of the sophisti-
cated nonlinear system into several lower order subsystems, bordering on the system
orders. Then comes the stage of the elaboration of the Lyapunov functions of all the
subsystems from the lowest one back until to the global system; as the functions are
constructed, intermediate variables are often used to check that the subsystems are
working correctly. The system could be guaranteed to be strictly stable if the devel-
opment of the Lyapunov function of each subsystem satisfies the general conditions
of Lyapunov stability theory.
On the other hand, Neghouchi, in 1998, was the first scientific researcher who
developed the DPC algorithm to control directly the active and reactive powers of
the rectifiers.
The presented approach in this research paper proposes a DPC-BS algorithm,
which combines between the both algorithms DPC and Backstepping, to control the
DFIG, whose the objective is to provide active and reactive powers which chase
values without derivations [5, 7]. Therefore, after combining the both methods, the
stator active and reactive powers equations become:
⎧  
⎪ K r Vs 1
⎪
⎨ Ps = σ L ϕsq − ϕrq
s Kr
  (6)
⎪
⎪ K V 1
⎩ Qs = r s ϕsd − ϕr d
σ L s Kr

The rotor voltage equations are given by [7]:

1284 M. Yessef et al.
⎧
⎪ dϕr d Rr ϕr d Rr K s
⎪
⎨ Vr d = + − ωr ϕrq − ϕsd
dt σ Lr σ Lr
(7)
⎪
⎪ dϕrq Rr ϕrq Rr K s
⎩ V rq = + − ωr ϕr d − ϕsq
dt σ Lr σ Lr

The rotor voltage equations can be regenerated as follows [5]:

⎧
⎪ 1 d Qs 1 1 dϕsd 1
⎪
⎨ Vr d = − μ dt − α Q s + μ ωr Ps + K dt + δϕsd − K ωr ϕsq
r r
(8)
⎪
⎪ 1 d P 1 1 dϕ 1
⎩ Vrq = − s
− α Ps − ωr Q s +
sq
+ δϕsq + ωr ϕsd
μ dt μ K r dt Kr
Rr (1−K r K s )
where: μ = kr Vs
σ Ls
,α = Rr
σ Lr μ
,δ = σ L r Kr
.
Thus, the correct operation of the DFIG is expressed by the first three terms, in
the voltage equations. Contrariwise, the last three terms model the variation of flux
during the fault of the electrical grid. These equations can be written differently as
follows [7]:
⎧
⎪ d Qs μ dϕsd μωr ϕsq
⎪
⎨ = −μα Q s + ωr Ps + + μδϕsd − − μVr d
dt K r dt Kr
(9)
⎪
⎪ dP μ dϕsq μωr ϕsd
⎩ s = −μα Ps − ωr Q s + + μδϕsq + − μVrq
dt K r dt Kr

According to the latest equations, the rotor voltages will be the instructions to be
given to the RSC to control the active and reactive stator powers. Thus the reactive
stator power Qs will be controlled by the rotor voltage Vrd , and the active stator power
Ps will be controlled by the rotor voltage Vrq . On the other hand, the powers errors
eQ and eP , used in the proposed DPC-BS algorithm, are given by [5]:

e Q = Q ∗s − Q s
(10)
e P = Ps∗ − Ps

The derivatives from the above parameters are given by:

⎧ ∗
⎪
⎪ ∗ ∗ μ ϕsd μωr ϕsq
⎪
⎨Qe = Q ∗
s +μα Q s − ω P
r s − − μδϕsd + + μVr d
Kr Kr
(11)
⎪
⎪ ∗
∗
μ ϕsq μωr ϕsq
⎪
⎩ e∗P = Ps∗ +μα Ps − ωr Q s − − μδϕsq − + μVrq
Kr Kr

To reduce these errors, we use the Lyapunov candidate function:

1 2

v= e Q + e2P (12)
2
Improved Hybrid Control Strategy ... 1285

Fig. 2 Structure of the proposed hybrid DPC-BS command

The Lyapunov candidate function derivative should be negative, to ensure a good

regulation of the stator powers. Therefore, using this function and its derivatives, the
expressions of the control rotor voltages variables Vrd and Vrq can be given by [7]:
⎧   ∗  
⎪ ∗ 1 ∗ 1  ∗

⎪
⎨ Vr d = − μ Q s +k Q e Q − ωr Ps + α Qs − K ωr ϕsq + ϕsq + δϕsq
r
  ∗    
⎪
⎪ 1 1 ∗
⎩ Vrq∗ = − Ps∗ −ωr Ps + α Qs − ωr ϕsq + ϕsq + δϕsq
μ Kr
(13)

The Backstepping controller intervenes when the voltage dips occur on the stator
supply terminal, in order to calculate the stator power references to control the RSC
and push the DFIG to operate correctly and satisfactory, by imposing the return of
the voltage at its nominal value (Fig. 2).

3 Simulation Results and Discusion

3.1 Performance Test under the Real Wind Profile

To ensure the satisfactory operation of the proposed hybrid command (DPC-BS)

under the real wind profile, simulation tests based on Matlab/Simulink have been
carried out with the 1.5 kW wind turbine driven the DFIG. Technically, to perform
this test scenario, the reactive power set-point will be kept zero, to ensure a unity
power factor, so as to optimize the quality of the injected energy into the eletrical
grid. The chosen real wind profile is that of the Dakhla city, located in the south of
Morocco (Fig. 3). This choice of site was based on the studies by researcher Allouhi
1286 M. Yessef et al.

[4]. The parameters of the used machine are described in the Tables 1 and 2 (see
appendix).
This test allows to evaluate the performance, the reliability of the WECS and the
decoupling of the powers generated by the DFIG during the instantaneous variation
of the power set-points (Fig. 4, 5, 6, 7).
The analysis of the simulation results confirm, that the active power follows
perfectly its reference value, which is deduced by the MPPT strategy, with a dynamic
error of εd = 1.05%. It can be also seen, that the reactive stator power is kept null in
order to have a unit power factor obtained after the end of its transient regime. More-
over the stator currents varies according to the wind variations, which confirms its
proportional relationship with the power. Despite the wind fluctuations, this current

zoom
15

10

0 5 10 15 20

10.5

10
Zoom wind speed (m/s)

9.5

8.5

7.5
0 5 10 15 20 25
time (s)

Fig. 3 Wind profile of the Dakhla city

1500 -500
Ps Ps*
0
zoom
Actice power (w)

-500 -1000

-1000
Zoom

-1500
-1500
-2000 -2000
-2500

-3000 -2500
0 5 10 15 20 25 14.5 15 15.5
time (s) time (s)

Fig. 4 Active power

Fig. 5 Reactive power 15

Qs Qs*
Reactive power (VAR)

10
5
0
-5
-10
-15
0 5 10 15 20 25
time (s)
Improved Hybrid Control Strategy ... 1287

Fig. 6 Stator currents 80

isa isb isc

Stator currents (A)

40

-40
zoom(1)
-80
0 5 10 15 20 25
time (s)

10
8 ira irb irc
Rotor currents (A)

6
4
2
0
-2
-4
-6
-8 zoom(1)
-10
0 5 10 15 20 25
time (s)

Fig. 7 Rotor currents

remains sinusoidal with a frequency of 50 Hz, which is the frequency of the electrical
grid. The rotor currents remain sinusoidal during the DFIG operation, its frequency
varies proportionally to the mechanical speed. These currents reveal fewer ripples
compared to other control techniques existing in literature.

3.2 Spectral Analysis Test

In order to verify the effect of the hybrid command DPC-BS on the quality of the
supplied power by the machine, a spectral analysis of the stator and rotor currents
was carried out. Note that this measurement was carried out for 3 operation cycles
with a harmonic order of 70. The Fig. 8 shows respectively the currents of the stator
phase “isA ” and the rotor phase “ira ” supplied by the DFIG, as well as their harmonic
spectra.
These figures confirmed the influence of the control on the quality of the supplied
energy. The rate of harmonic distortion is equal to 0.33% at the level of the rotor
current, and 0.94% at the level of the stator current. From the depicted results in

ira (3Hz) = 1.85 , THD= 0.33%

Fig. 8 THD of injected currents into the electrical grid

1288 M. Yessef et al.

these simulations, it can be noted that the harmonic distortion rate THD has a good
quality for the both injected currents into the electrical grid and remains below the
limit of 2% which imposed by the IEEE-519 standard [13].
This THD value (0.33%) obtained, in this work, by the proposed DPC-BS hybrid
control is the lowest, compared to the value recently reached by the researchers [8]
for WECS based on DFIG which is 1.67%.

4 Conclusion

In this work, we have presented a contribution for the control of the wind energy
conversion system. Therefore we applied the improved hybrid DPC-BS command
to improve the performance of the doubly-fed induction generator. This approach
is based on the combination of the Backstepping controller and the Direct Power
Control. The simulation results analysis have shown that the hybrid DPC-BS
command gives highest dynamic performance in the reference tracking, in a test
scenario with a real wind profile, and it also ensures an optimal energy conversion.
Decidedly, this command provides, a concrete solution to the problems of robustness
and pursuit. Moreover, another essential point of this work is that, it confirmed the
injecting possibility of the supplied energy, by the studied WECS, into the electrical
grid.
For the future works, a hardware-in-the-loop simulation on dSPACE will be
realized to verify the proposed approach in the real time by experiments.

Appendix

Table 1 The DFIG

Name Values
parameters
Nominal stator voltage 220/380 V
Nominal rotor voltage 12 V
Nominal current 3.64/6.31A
Nominal power Pn = 1,5 kW
Nominal frequency fn = 50 Hz
Nominal torque Cn = 10 N.m
Power factor 0.89
Number of pole pairs P=2
Nominal speed 1440 tr/min
Stator resistance Rs = 0.435
(continued)
Improved Hybrid Control Strategy ... 1289

Table 1 (continued)
Name Values
Rotot resistance Rr = 0.816
Stator inductance coefficient Ls = 0.071H
Rotor inductance coefficient Lr = 0,070H
Mutual inductance coefficient M = 0,069H
Friction coefficient f = 6,73.10−3 kg.m2
Moment of inertia J = 0,3125 N.s/rad

Table 2 The wind turbine

Name Value
parameters
Nominal power Pn = 7,8 kW
Density of air ρ = 1,225
Blade angle β = 2°
Number of blades 3
Blade radius R=3m
Gearbox ratio G = 10,38
Conversion yield η=1
Moment of inertia J = 0,042 kg.m2
Viscous friction coefficient f = 0,024 N.s/rad

References

1. Allouhi A (2019) Energetic, exergetic, economic and environmental (4 E) assessment process

of wind power generation. J Clean Prod 235:123–137
2. Lacal Arantegui R (2015) Materials use in electricity generators in wind turbines – state-of-
the-art and future specifications. J Clean Prod 87:275–283
3. Allouhi A, Zamzoum O, Islam MR, Kousksou T, Jamil A, Saidur R, Derouich A (2017)
Evaluation of wind energy potential in Morocco’s coastal regions. Renew Sustain Energy Rev
72:311–324
4. De Luca E, Nardi C, Giuffrida LG, Krug M, Di Nucci MR (2020) Explaining factors leading
to community acceptance of wind energy. Results of an expert assessment. Energies. 13:2119
5. Mensou S, Essadki A, Nasser T et al (2020) A direct power control of a DFIG based-WECS
during symmetrical voltage dips. Prot Control Mod Power Syst 5:5
6. Bossoufi B, Karim M, Lagrioui A, Taoussi M, Derouich A (2014) Modeling and backstepping
control of DFIG generators for wide-range variable-speed wind turbines. J Electr Syst JES
10(3):317–330
7. Bossoufi B, Karim M, Taoussi M, Alami Aroussi H, Bouderbala M, Motahhir S, Camara MB
(2020) DSPACE-based implementation for observer backstepping power control of DFIG wind
turbine. IET Electric Power Appl 14(12):2395–2403
8. Zamzoum O, Derouich A, Motahhir S, El Mourabit Y, El Ghzizal A (2020) Performance
analysis of a robust adaptive fuzzy logic controller for wind turbine power limitation. J Cleaner
Prod 265:121659. https://doi.org/10.1016/j.jclepro.2020.121659
9. Taoussi M, Karim M, Lagrioui A, Bossoufi B, El Mahfoud M (2014) The fuzzy control for
rotor flux orientation of the doubly-fed asynchronous generator drive. Int J Comput Technol
13(08):4707–4722
1290 M. Yessef et al.

10. Taoussi M, Karim M, Bossoufi B, Hammoumi D, Lagrioui A (2015) Speed backstepping

control of the doubly-fed induction machine drive. J Theor Appl Inf Technol 74(2):189–199
11. Bossoufi B, Karim M, Lagrioui A, Taoussi M, Derouich A (2015) Observer backstepping
control of DFIG-generators for wind turbines variable-speed: FPGA-based implementation.
Renew Energy 81:903–917
12. Taoussi M, Karim M, Bossoufi B, Hammoumi D, Elbakkali C, El Ouanjli N (2017) Comparative
study between Backstepping adaptive and Field-oriented control of the DFIG applied to wind
turbines. In: Processing IEEE Xplore
13. Institute of Electrical and Electronics Engineers. Static Power Converter Committee (1981)
IEEE Guide for Harmonic Control and Reactive Compensation of Static Power Converters
Flow-Oriented Control Design of Wind
Power Generation System Based
on Permanent Magnet Synchronous
Generator
Nada Zine Laabidine, Chakib El Bakkali, Karim Mohammed,
and Badre Bossoufi

Abstract In this article we have studied the control of a high performance electric
machine, more precisely a permanent magnet synchronous generator using the Flux
Orientation Control (FOC), this allowing us to independently control the torque and
the flux of our machine in the same way as a DC machine with separate excita-
tion, where the inducing current controls the magnetic flux, and the induced current
controls the electromagnetic torque.
We transformed the stator instantaneous currents to two current components, one
which controls the flux (along the d axis), and the other controls the torque (along
the q axis).
The goal of our article is based on the DFOC (direct) technique, which consists
of neglecting the coupling terms, and setting up a proportional integral regulator (PI)
in order to follow the parametric variations of our machine using two types of wind
(variable and step) and independently control the active and reactive power.

Keywords Wind power generator system · Permanent magnet synchronous

generator (PMSG) · Flow-oriented control

1 Introduction

In the last few years, the use of synchronous machines in industry has grown sharply
because of their similar characteristics with DC machines [1, 2].
Two types of synchronous machines are present in the market, synchronous
machines with permanent magnets and winding rotor machines, those with perma-
nent magnets have superior performance compared to the other with excellent
efficiency, high mass power and reduced maintenance [3, 4].
Several control strategies are present in the literature and are reserved to
synchronous generator with permanent magnet, we can quotes for example:

N. Z. Laabidine · C. El Bakkali · K. Mohammed · B. Bossoufi (B)

LIMAS Laboratory, Faculty of Sciences Dhar El Mahraz, Sidi Mohamed Ben Abdellah
University, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1291
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_118
1292 N. Z. Laabidine et al.

direct torque control (DTC), vector oriented flow control (FOC) or the non-linear
Back-stepping control.
We present in this article the modeling of the synchronous permanent magnet
machine using vector-oriented flow control [5] (FOC) applied to the static converter
on the machine side [6], which allows us to control the electromagnetic torque,
magnetic flux, active and reactive powers of a wind generator in order to obtain a
simple and reliable model for numerical simulation.

2 Permanent Magnet Synchronous Generator Model

(PMSG)

The GSAP has a coiled stator but the rotor is replaced by magnets, which eliminates
rhetoric losses and is considered a power source that can provide excitation current.

2.1 PMSG Electrical Equations

The equations are related to the stator by [7]:

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
Vsa isa sa
d
⎣ Vsb ⎦ = Rs ⎣ isb ⎦ + ⎣ sb ⎦ (1)
dt
Vsc isc sc

With:
• R s :The resistance of a statoric phase ()
• V sa : Statoric phase voltage (V)
• i sa : Statoric current intensity (A)
• sa : Total flow through the stator coils (Wb)
PMSG stator voltage equations in a system of axes (d-q)

Vsd = −Rs i sd − sd − ωe sq (2)

Vsq = −Rs i sq − sq + ωe sd (3)

With:
• ωe : The electric pulsation of voltages (rad.s−1 )
• sd : Direct component of the statoric flow (Wb)
• sq : Quadratic component of the statoric flow (Wb)
Flow-Oriented Control Design of Wind Power Generation System ... 1293

Expressions of (sd , sd ) depending on the stator currents

sd = L d · i sd +  f (4)

sq = L q · i sq (5)

With: L d L q are the direct and quadratic components of the statoric inductance
(H).
Then [8]:
 
d L d · i sd +  f  
Vsd = −Rs i sd − − ωe L q · i sq (6)
dt
 
d L q · i sq  
Vsq = −Rs i sq − + ωe L d · i sd +  f (7)
dt
Ignore the stator transients and obtain:

d(i sd )
Vsd = −Rs i sd − L d . − ωe (L q · i sq ) (8)
dt
 
d i sq  
Vsq = −Rs i sq − Lq · + ωe L d · i sd +  f (9)
dt
From (8) and (9) we can deduce the equations of statoric currents

d(i sd ) −1
= (Vsd + Rs i sd + ωe (L q · i sq )) (10)
dt Ld
d(i sq ) −1  
= (Vsq + Rs i sq − ωe L d · i sd +  f ) (11)
dt Lq

Expression of electromagnetic torque

−3 
Tem = .p. sd .isq − sq· isd (12)
2
−3 
Tem = .p. (Ld − Lq )isd isq + f· isq (13)
2
−3
Tem−r e f = 2
.p.f .i sq (For a machine with plain poles(L d = L q )).
1294 N. Z. Laabidine et al.

Fig. 1 Equivalent circuit (d-q) for a PMSG

The expression of statoric voltages (using the laplace transformation) [9]

Vsd = −(R s + S.L d ).I sd − ωe (L q · Isq ) (14)

 
Vsq = −(R s + S.L q ).I sq + ωe L d .Isd +  f (15)

−1
Isd = (Vsd + ωe (L q · i sq )) (16)
(R s + S.L q )
−1  
Isq = (Vsq − ωe L d .i sd +  f ) (17)
(R s + S.L q )

Expression of active and reactive powers

3
Pgen = Tem . = . Vsd .i sd + Vsq .i sq (18)
2
3
Q gen = . Vsd .i sd − Vsq .i sq (19)
2
Equivalent circuit (d-q) for a PMSG (Fig. 1)

3 Oriented Flow Vector Control (FOC)

Vector control was first proposed by “Blaschke” in 1972 [1], after it was developed
by “Hasse” in 1979 [2], also called flow-oriented control (FOC), is a method in which
the three-phase statoric currents of an AC electric motor are transformed into two
orthogonal components which can be considered as vectors, The first produces the
electromagnetic torque and the second produces the magnetic flux. This allows us to
operate similar to a DC motor [10].
An Integral Proportional Controller (PI) was used to maintain the current at its
requested value (Id) (Fig. 2).
Flow-Oriented Control Design of Wind Power Generation System ... 1295

Fig. 2 Oriented Flow Vector

Control (FOC)

4 Wind Turbine Model

The kinetic energy of the wind is transformed into mechanical energy by the rotor
of the turbine, and then converted into electrical energy by a generator; its formula
is given by [11]:

1
Ec = .m.v 2 (20)
2
The kinetic power of the wind received by the turbine through its blades is given
by:

1
Pwind = .ρ.π.R 2 .v 3 (21)
2
The aerodynamic power Ptur appearing at the turbine rotor is written:

1
Ptur = .C p (λ, β).ρ.π.R 2 .v 3 (22)
2
The expression of the aerodynamic torque Cp of the turbine is given by the
following formula:

1
Cp = .C p (λ, β).ρ.π.R 2 .v 3 (23)
2.t

The rotor torque induced by the wind is determined by the equation:

Ptur = ω.T
 tur ; λ = V
Rω

2
Ttur = 21 ρπ R 3 V C P (λ, β)
1296 N. Z. Laabidine et al.

The coefficient of performance changes with different angles of inclination (β)

as well as the speed ratio ((λ):

Ttur _opt = K ω2
K = 21 ρπ R 5 C P−max (λopt )

• Ptur , The power captured by the wind turbine.

• Ttur _opt , The optimal turbine torque.
• Ttur , The turbine torque.
• C p , The power coefficient.
• S, The blade swept area.
• R, The radius of the turbine blade.
• ρ, the specific density of air.
• V, the wind speed.
• ω, the rotor radium.
• λ, tip-speed ratio; β: pitch angle.
• λopt The optimal tip speed ratio that maximizes the power extracted by the wind
turbine.

5 MLI Model

In order to have a constant frequency and amplitude voltage on the machine side and
to optimize the power taken from our wind energy conversion system, we will use
MLI power converters.
⎧
⎨ +1, S = −I
S= , S = a, b, c (24)
⎩
−1, S = +I

The input voltages between phases of the rectifier are:

⎧ ⎡ ⎤ ⎡ ⎤⎡ ⎤
⎨ U AB = V AO − VB O VA
1
2 −1 −1 V AO
U = VB O − VC O ; ⎣ VB ⎦ = ⎣ −1 2 −1 ⎦⎣ VB O ⎦ (25)
⎩ BC 3
UC A = VC O − V AO VC −1 −1 2 VC O
⎡ ⎤ ⎡ ⎤
V AO S
UCC ⎣ A ⎦
We have : ⎣ VB O ⎦ = SB (26)
2
VC O SC
⎡ ⎤ ⎡ ⎤⎡ ⎤
vA 2 −1 −1 SA
U
then : ⎣ v B ⎦ =
CC ⎣
−1 2 −1 ⎦⎣ S B ⎦ (27)
6
vC −1 −1 2 SC
Flow-Oriented Control Design of Wind Power Generation System ... 1297

Park Transformation
Park’s transformation is also called direct-quadrature-zero (dq0) allows to transform
the stator or rotor windings a, b, c into two windings d and q, We denote by “d” the
direct axis and “q” the quadrature axis.

5.1 Direct Park Transformation

 ⎡     ⎤⎡ ⎤
Cos() cos  − 2π
3 
cos  + 2π
3 
xA
2 ⎣ 2π ⎦ ⎣ ⎦
· −sin() −sin  − 2π3
−sin  + 3
. x B (28)
3 1 1 1
2 2 2
xC

5.2 Reverse Park Transformation

⎡ ⎤  ⎡ ⎤⎡ ⎤
xA Cos()
  −sin()
  1 xd
⎣ xB ⎦ = 2
.⎣ cos  − 2π −sin  − 2π
1 ⎦.⎣ xq ⎦ (29)
3  3   3 
xC cos  + 2π
3
−sin  + 2π 3
1 xo

6 PI Regulator

In order to regulate the stator current values i sd eti sq , we are going to use a PI regulator,
which also aims to provide better performance and robustness to our wind power
k .s+K
system. The equation of the PI regulator used to control the GSAP: p s i
With:
• k p : proportional gain of the corrector PI
• K i : proportional gain of the corrector P

d(i sd )
Vsd = −Rs Isd − L d . − ωe (L q · i sq )
dt
d(i sq )
Vsq = −Rs i sq − L q . + ωe (L d .i sd +  f )
dt

We put Aq = ωe (L q · i sq ); Ad = ωe (L d .i sd +  f )
1298 N. Z. Laabidine et al.

Then the transfer function of our GSAP becomes:

Vsd = −Rs Isd − S.L d .I sd − Aq (30)

Vsq = −Rs Isq − S.L q .I sq − Ad (31)

1 1 1
FT(s) = = . ;
Rs + L s .s Rs 1 + Te .s

with Te = L s Rs (Electric time constant)
L d = L q (Smooth pole machine stator inductance) that implies that k p = ki
(Fig. 3 and 4).
We can simplify the control loop to be similar to that of a first order system (Fig. 5).
The open-loop transfer function is given by:
  
k p .s + K i 1
FTBO(s) = (32)
s Rs + L s .s

Fig. 3 Direct stator current regulation loop

Fig. 4 Quadrature stator current regulation loop

Fig. 5 Simplified stator

current regulation loop
Flow-Oriented Control Design of Wind Power Generation System ... 1299

The closed loop transfer function:

1 1
FTBF(s) =   =   (33)
Rs +k p
.s
Ls 2
ki
+ ki
s+1 1
ωn
.s 2 + ω2ζn s + 1

• ωn : The natural pulsation of the system

• ζ : The damping coefficient of the system.

ki = ωn 2 .L s ; k p = 2.ζ.ωn .L s − Rs

7 Modeling and Simulation Result

The simulation results obtained after the introduction of models in

MATLAB/SIMULINK are illustrated in the following figures:
We have studied a mathematical model of the permanent magnet synchronous
machine in Park’s benchmark to make the GSAP machine similar to the separately
excited DC machine (Fig. 6 and 7).
We used a variable wind speed to check the decoupling between the components
of the stator current to control the GSAP then a step speed (Fig. 8, 9, 10, 11, 12, 13,
14, 15, 16, 17 and 18).

Fig. 6 Model of our permanent magnet synchronous generator

1300 N. Z. Laabidine et al.

Fig. 7 The GSAP model with the SMC command

Fig. 8 Variable wind profile; Step wind profile

Fig. 9 Active power as a function of the wind

Fig. 10 The electromagnetic torque as a function of the wind

Flow-Oriented Control Design of Wind Power Generation System ... 1301

Fig. 11 The direct stator current «i sd » as a function of the wind

Fig. 12 The quadrature stator current «i sq » as a function of the wind

Fig. 13. The stator currents produced by the PI regulator

Fig. 14 The active power as a function of the wind (scale)

Fig. 15 The active power as a function of the wind (Scale)

1302 N. Z. Laabidine et al.

Fig. 16 The direct stator

current «i sd » as a function of
the wind (Scale)

Fig. 17 The quadrature

stator current «i sq » as a
function of the wind (Scale)

Fig. 18 The stator currents

produced by the PI regulator
as function of wind (Scale)

8 Conclusion

We have studied a mathematical model of the synchronous machine with permanent

magnets in Park’s landmark in order to make the GSAP machine similar to the DC
machine with separate excitation. Two types of wind speed will be studied in this
test, one variable and the other step in order to analyze the decoupling between the
components of the stator current (to control the GSAP) as well as the operation of
our PI regulator.
The previous figures summarize the simulation results performed while applying
the MPPT power maximization algorithm with mechanical rotation speed control.
Following the results of the test we can notice that the instructions set for vector
control based on PI regulators are followed for the electromagnetic torque, the active
and reactive powers as well as the stator currents generated.
On the other hand, the stator currents show ripples due to the switching of the
electronic switches of the static converter on the machine side.
The active power generated is a direct consequence of the electromagnetic torque
of the machine. In other words, the active power is controlled by the stator current
of the quadrature axis (q), while the reactive power is a direct consequence of the
stator current of the direct axis (d).
Flow-Oriented Control Design of Wind Power Generation System ... 1303

References

1. EL Mourabit Y, Derouich A, EL Ghzizal A, Bouchnaif J, EL Ouanjli N, Bossoufi B (2019)

Implementation and validation of backstepping control for PMSG wind turbine using dSPACE
controller board. Energy Report J 5:807–821
2. EL Ouanjli N, Derouich A, EL Ghzizal A, Bouchnaif J, Taoussi M, Bossoufi B (2019) Real-
time Implementation in dSPACE of DTC-Backstepping for doubly fed induction motor. Eur
Phys J Plus 135(1):2–9
3. Bossoufi B, Karim M, Lagrioui A (2014) MATLAB & simulink simulation with FPGA Based
Implementation adaptative and not adaptative backstepping nonlinear control of a permanent
magnet synchronous machine drive. WSEAS Trans Syst Control 9:86–100
4. Bossoufi B, Karim M, Lagrioui A, Taoussi M, El Hafyani ML (2014) Backstepping control of
DFIG generators for wide-range variable-speed wind turbines. Int J Autom Control 8(2):122–
140
5. Beltran B, Ahmed-Ali T, Benbouzid MEH (2009) High-order sliding-mode control of variable-
speed wind turbines. IEEE Trans Industr Electron 56(9):3314–3321
6. Bossoufi B, Ionita S, Alami Arroussi H, EL Ghamrasni M, Ihedrane Y (2017) Managing voltage
drops a variable speed wind turbine connected to the grid. IJAAC Int J Autom Control 11(1)
7. Ihedrane Y, El Bekkali C, EL Ghamrasni M, Mensou S, Bossoufi B (2019) Improved wind
system using on-linear power control. Indonesian J Electr Eng Comput Sci 14(3):1148–1158
8. Bossoufi B, Karim M, Ionita S, Lagrioui A (2012) Nonlinear non adaptive backstepping with
sliding-mode torque control approach for PMSM motor. J Electr Syst JES 8(2):236–248
9. Bossoufi B, Karim M, Lagrioui., (2015) Observer backstepping control of DFIG-generators
for wind turbines variable-speed: FPGA-based implementation. Renew Energy 81:903–917
10. Yi Liu W, Ai BC, Chen K, Luo G (2016) Control design and experimental verification of the
brushless doubly-fed machine for stand-alone power generation applications. IET Electr Power
Appl 10(1):25–35
11. Bossoufi B, Taoussi M, Alami Aroussi H, Bouderbala M, Motahhir S, Camara MB (2020)
DSPACE-based implementation for observer backstepping power control of DFIG wind
turbine. IET Electr Power Appl 14(12):2395–2403
Households Energy Consumption
Forecasting with Echo State Network

Wadie Bendali, Mohammed Boussetta, Ikram Saber, and Youssef Mourad

Abstract Careful planning and forecasting of energy consumption not only influ-
ences a nation’s environmental and energy sustainability, as well as giving a useful
basis for policy makers to make decisions. This paper presents the results of an appro-
priate deep learning model forecasting for consumption using echo state network
(ESN). ESN is a new paradigm that offers an intuitive methodology using for time
series prediction. Basically, it is a recurrent neural network (RNN) with a vaguely
connected hidden layer, known as a reservoir, which functions in a strange way in
the existence of time-series patterns. In this contest, three types of recurrent neural
network used for comparison with ESN, aiming to evaluate the accuracy and perfor-
mance of the model. To train and test the proposed model, we used the historical data
of multivariate household consumption. ESN showed an improvement of 0.057% and
0.095% in terms of mean square error (MSE) and mean absolute error (MAE), and it is
trained faster than other models, in very short-term energy consumption forecasting.

Keywords ESN · Consumption energy · RNN models · Deep learning ·

Forecasting

1 Introduction

The role of energy in human life is absolutely essential. In one way, the evolution of
human society due to the rise of energy efficiency and the exploitation of new artificial
intelligence techniques. In parallel with economic and population dynamics, global
consumption of energy continues to increase, causing several challenges. There is
enormous pressure on the energy sectors worldwide to ensure a stable and reliable
energy supply. Over the years, the consumption of energy has grown remarkably.
Based on the statistics published in 2019 by British Petroleum, there will be a 2.8%
increase in primary energy consumption in 2018, which is higher than in the previous
year 2017 [1]. Due to the rapid growth of intelligent measuring systems and the

W. Bendali (B) · M. Boussetta · I. Saber · Y. Mourad

Laboratory of Industrial Technologies and Services, Higher School of Technology (EST) of Fez,
University of Sidi Mohammed Ben Abd Allah, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1305
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_119
1306 W. Bendali et al.

“Green Button” project [2], electricity consumption power for household users is now
measured, collected and presented. The use of Demand Response (DR) and Home
Area Networks (HAN) has put a new emphasis on individual households [3]. While
residential households contribute to DR programs as a willing-seller, it is difficult
to plan and estimate energy for these individual households prior to the program
without an energy forecast at the individual household level. Certain utilities publish
utilization or billing forecasts for residential customers to provide an innovation in
service, assisting households to plan their consumption and reduce their energy bills.
As a result, there is a growing recognition of the potential for forecasting household
energy consumption by governments and research institutes. Short- and very short-
term energy forecasting methods are very useful for managing residential energy
demand, energy efficiency, and market design of electricity prices, in addition they
are very powerful for planning in maintenance of large-scale, complex smart grids
[4–6]. Furthermore, it offers the smart grid reliability, security and protection to help
meet the increasing energy demand of residential homes.
ESN represents a family of Reservoir computing (RC) models where the recurrent
part is created randomly, and then kept fixed. Even with this strong simplification, the
recurring part of the model (the reservoir) provides a rich set of dynamic character-
istics that allow to solve a wide variety of problems. In deep learning, the echo state
network (ESN) further belongs to the general class of the reservoir computing. The
ESN has powerful non-linear time series modelling capabilities ESN has a dynamic
reservoir with a number of connected units, which makes the calculation of the data
very uncomplicated [7]. Among the specific features of the reservoir computing
methods, they have the ability to train only the readout weights by using the method
of linear regression throughout the training procedure. By condensing this character-
istic, ESN can achieve a high velocity of convergence towards an excellent overall
solution. This study aims to propose echo state network model with multivariable
input for energy consumption prediction.
The rest of this study is organized as follows. Section 2 introduces consump-
tion data description and describes preprocessing of data. Section 3 represents the
basic ESN architecture, his parameters and training process. Section 4, experimental
application used, and presentation of results. Finally, conclusion.

2 Data Description

2.1 Household Power Consumption Dataset

Generally, four time horizons can be distinguished, that are used to predict time series:
very short, short, medium and long term [8]. Very Short-term prediction concentrates
on minute time ranges, contributing to assure an uninterrupted power flow in the near
future and supporting electricity trade and price management [8, 9]. Medium and
long-term prediction concentrate over a duration of a few days to several months and
Households Energy Consumption Forecasting with Echo State Network 1307

assist in the installation and maintenance of production facilities [8, 10], transmission
and distribution system development [11], and energy plan formulation [12]. In this
work, very short term data used for train forecasting model, which is a multivariate
time series of household electricity consumption. The data used are collected from
a house located in Sceaux (7 km of Paris, France), during one year (2009), with
observations of every minute [13]. These data made up of 7 features:
Pa :: Global active household power consumption (KW).
Pr : Overall reactive power required by the household (KW).
V : Means voltage (volts).
Ig : Average current intensity (amps).
E a1 : Sub-measure of active energy for the kitchen (Wh).
E a2 : Sub-measure of active energy for laundry (Wh).
E a3 : Sub-measure of active energy for climate control systems (Wh).
The data set contains the active power, voltage and intensity current averages
and certain distribution of active energy in the household like kitchen, washroom
and air-conditioning. In order to obtain greater accuracy, we also add the rest of the
sub-measures of the active energy E a4 , subtracting the three sub-measures from the
overall active energy, by this following equation (Eq. (1)). All features are shown in
Fig.1:

E a4 = (P a × 1000/60) − (E a1 + E a2 + E a3 ) (1)

2.2 DATA Preprocessing

In this work, we use data of 1 year (2009), which contain timescale for each minute.
This data are separated like this: 473040 sample used for model train, 52560 samples
used for the test. Global active power data have strong nonlinear properties, this can
cause negative impacts on the prediction model. Before the model is fed by train
data, we pre-process input values, for raison to minimize the effect of the wide values
scale. The pre-processing process that we choose, namely the linear normalization
that changes the data in the interval between 0 and 1. Post-processing is necessary
upstream or downstream of the analysis for the performance of the forecast model.
In this case post-processing of the normalized values must be anti-normalizing to
extract the expected actual global power active and analyze the model performance.
1308 W. Bendali et al.

Fig. 1 Model input variables

Fig. 2 ESN architecture

Households Energy Consumption Forecasting with Echo State Network 1309

3 Echo State Network

3.1 Architecture

The architecture of ESN is like any neural network, has input, hidden, and output
layers, which are also having a certain number of units (neurons). However, as shown
in Fig. 2, ESN has a reservoir instead of simple hidden layer, made up of some number
of recurrent or simple units. These units are interconnected in order to achieve high
treatment capability and, consequently, obtain good forecasts [7]. Although the ESN
has been applied in many areas since its introduction in 2001 [14], this includes time
series prediction and data classification [7, 15]. Generally, ESN consists of k input
neurons that represent features or variable numbers, n units in the hidden reservoir,
which are responsible for creating internal information, and o units in the final layer,
which gives forecast time series values. This study deals with prediction problems
in a single step so we are used o = 1, in time step j, assume that the input array of
the input state is given as Eq.(2) The array of the reservoir state is Eq. (3), and the
array of the output state is Eq. (4):

U ( j) = [U1 ( j), U2 ( j), U3 ( j), · · · , Uk ( j)]T (2)

X (i) = [X 1 ( j), X 2 ( j), X 2 ( j), · · · , X n ( j)]T (3)

Y ( j) = [Y1 ( j), Y2 ( j), Y2 ( j), · · · , Yo ( j)]T (4)

Where j = 1, 2…N and N represent all training time steps.

During processing, the reservoir and output states are updated as shown in Eqs. (5)
and (6), accordingly:

X ( j + 1) = F(win U ( j + 1) + w X ( j) + wback Y ( j)) (5)

Y ( j + 1) = G(wout X ( j + 1)) (6)

In ESN, four weight matrices exist, win represent the connection between input
and reservoir state, w represent the weight between internal units of reservoir layer,
wback gives the connection between neurons of output layer and reservoir layer, and
wout as a matrix weight that link the internal units and output units. The activation
functions of the reservoir and output layers is F and G respectively. These functions
can represent a linear or nonlinear function depending on the nature of data.
1310 W. Bendali et al.

The weight matrix wout , which is from the reservoir to the output layer, can be
updated throughout the training process, whereas win , w and wback are randomly
chosen then stay unchanged [16, 17].

3.2 Hyperparametres of ESN

ESN has three critical reservoir Hyperparametres, which have a major effect on the
efficiency of the system and should be set at appropriate values.
n: The number of units in hidden reservoir, that can effect in terms of model
quality and accuracy, due to his role in dealing with large data size in the training
phase. Generally, the choose of number of units according on the nature and size of
data using for training. In this study, we use n = 60.
α: The connectivity rate, which allow to choose the linking degree between units
of hidden reservoir layers. That can aide to control the memory chance and deal with
computing complexity [16]. In general, with reference to review of experience, α is
chosen between 1%–5% [18].
ρ: The spectral radius, has a very important role in ESN model, his value represents
the max of the absolute proper value of the weight matrix w. In other to make ESN
has dependency only with input and output sample of training we concede that the
range of ρ is between 0 and 1. It is unaffected by its first state after a few epochs if
the entries are long enough. [19, 20].

3.3 The Training Algorithm of ESN

In order to achieve accurate output compared with target values, which are collected
from data training, we have to train models by using N time samples of input and
target data. There are three steeps using for training models: Firstly, it must initialize
the ESN parameter values, by choosing a pertinent number of reservoir units n, select
the best values of connectivity rate α and spectral radius ρ, and initialize randomly
weight values win ,wback , and w. After that, it is important to update the reservoir
states using the actual input, the previous values of reservoir and output states as
mentioned in Eq. (5). Finally, Computing wout by obtaining M the vector of hidden
reservoir values, which has (N − I0 + 1) × n sample, and T the target values, which
has (N − I0 + 1) × o as follows in Eq. (7). With I0 is the washout time step.
T
wout = (M −1 T ) (7)

ESN has a special training algorithm, which can make it able to deal simply and
quickly with data like consumption power of households [14, 21].
Households Energy Consumption Forecasting with Echo State Network 1311

4 Application and Results

In this work, we create our model with Python environment by using KERAS as a
deep learning library, which supports CPU and GPU [22]. We choose RNN and their
extension Long Short Term Memory (LSTM), and Gated Recurrent Unit (GRU) as
deep learning models to compare them with ESN. These models used input data that
contains eight variables represent eight units, total active power(GAP), total reactive
power, voltage, total intensity, and four Sub-measuring data of the household’s energy
(Fig. 1). The row of this data contains 525600 samples of minute, these data are used
as inputs of models training and testing. Data are split into 90% for training and 10%
for testing. The output of the models has one unit, which represent, predict global
active power consumed. For dealing with the effect of dimensionality problem of data,
we normalize data values between 1 and 0 by using normalization as a preprocessing
method. After obtaining predict values, we used anti-normalization for obtaining
the forecast GAP. The hyperparameters of ESN that we choose the number of units
n = 50, connectivity rate α = 0.02, and spectral radius ρ = 0.5. We choose these
parameter values based on Trial and Error method, aiming to achieve the smaller
error by testing various parameter values. This method is an efficient and standard
method of parameterization. Activation functions using in ESN are F = Tanh and
G = identity. Concerning the standard RNN and its extensions (LSTM, GRU), has
two hidden layer’s neurons,each one of theme has 60 neurons or units with Tanh
activation function for two hidden layers and Relu for the output layer. The error
metrics compared in this paper was mean square error (MSE) (Eq.(8)), and mean
absolute error (MAE) (Eq.(9)), as following equations:

1 
N
MSE = (P f or ecast (t) − Ptr ue (t))2 (8)
N i=1

1  
N
M AE = P f or ecast (t) − Ptr ue (t) (9)
N i=1

With N is the number of samples, P f or ecast (t) is the forecasting of global active
power at time t, and Ptr ue (t) is the true global active power at time t.
In this work, we compared the accuracy of the proposed deep learning method
ESN with simple RNN, LSTM and GRU. As we can see in Figs. 3, 4, 5, and 6, ESN
model has more accuracy than standard RNN and his extension in consuming power.
In Table 2, comparisons based on MAE and MSE of models. As shown in this table,
the LSTM and GRU models have better performance than Simple RNN method,
that is because LSTM and GRU methods characteristics of carrying initially learnt
important information over a long distance [23, 24]. However, ESN still much better
compared with them, because of his ability to deal with nonlinearity, and his special
training algorithm. In terms of training speed, we can see from Table 1, ESN is faster
1312 W. Bendali et al.

Fig. 3 True values vs. ESN forecast values of global active power consumed in 1 day

Fig. 4 True values vs. GRU forecast values of global active power consumed in 1 day

Fig. 5 True values vs. simple RNN forecast values of global active power consumed in 1 day

than RNN, which is considerably faster than LSTM, and GRU, that is because in
ESN we only train the output weight wout and it can speed up the training of neural
network.
Households Energy Consumption Forecasting with Echo State Network 1313

Fig. 6 True values vs. LSTM forecast values of global active power consumed in 1 day

Table 1 Training time of models

Models Simple RNN LSTM GRU ESN
Time training (s) 25.859375 53.703125 46.640625 15.890625

Table 2 MSE and MAE of models

Simple RNN LSTM GRU ESN
MSE 0.0011 0.00067 0.00063 0.000576
MAE 0.0203 0.0108 0.0103 0.009746

5 Conclusion

The development of an appropriate energy prediction model is required to assure

a secure and reliable power consumption. This study illustrates the efficiency and
performance of the ESN model for forecasting energy consumption. Different deep
learning methods are used to predict consumption power like RNN, which has a good
accuracy compared with other machine learning methods. In this study, we compared
ESN with simple RNN, LSTM, and GRU, to test the accuracy and the performance of
ESN for consumption hold houses forecasting. LSTM and GRU are famous model
that is used in many fields, they are good in term of efficiency specially for time
series. However, these methods have slow and cumbersome calculation, because
of their algorithm, which trait input, recurrent and output relations using gradient
descent, and makes the learning disrupted. Unlike other methods, ESN has a loosely
connected hidden layer, called “Reservoire”, which allows for rapid training, and
can easily be implemented. As a result, ESN has outperformed all other methods of
nonlinear dynamic modeling.
1314 W. Bendali et al.

References

1. World B (2018) BP Statistical Review of World Energy. https://www.bp.com/en/global/corpor

ate/energy-economics/statistical-review-ofworld-energy.html
2. Zhang XM, Grolinger K, Capretz MA, Seewald L (2018) Forecasting residential energy
consumption: single household perspective. In: 2018 17th IEEE international conference on
machine learning and applications (ICMLA), Orlando, FL, pp 110–117. https://doi.org/10.
1109/ICMLA.2018.00024
3. Pipattanasomporn M, Kuzlu M, Rahman S (2012) Demand response implementation in a home
area network: A conceptual hardware architecture. In: Proceedings of the IEEE innovative smart
grid technologies (ISGT), pp 1–8
4. Guo P, Lam JC, Li VO (2019) Drivers of domestic electricity users price responsiveness: a
novel machine learning approach’. Appl Energy 235:900–913
5. Mazzi N, Kazempour J, Pinson P (2018) Price-taker offering strategy in electricity pay-as-bid
markets. IEEE Trans Power Syst 33(2):2175–2183
6. Wang W, Chen H, Lou B, Jin N, Lou X, Yan K (2018) Data-driven intelligent maintenance
planning of smart meter reparations for largescale smart electric power grid. In: Proceed-
ings IEEE smartworld, ubiquitous intelligent computer, advance trusted computer, scalable
computer communication, cloud big data computer, internet people smart city innovation, pp
1929–1935
7. Zhong S, Xie X, Lin L et al (2017) Genetic algorithm optimized double-reservoir echo state
network for multi-regime time series prediction. Neurocomputing 238:191–204
8. de Oliveira EM, Oliveira FLC (2018) Forecasting mid-long term electric energy consumption
through bagging ARIMA and exponential smoothing methods. Energy 144:776–788
9. Arora S, Taylor JW (2018) Rule-based autoregressive moving average models for forecasting
load on special days: a case study for France. Eur JOper Res 266(1):259–268. https://doi.org/
10.1016/j.ejor.2017.08.056
10. Mason K, Duggan J, Howley E (2018) Forecasting energy demand, wind generation and carbon
dioxide emissions in Ireland using evolutionary neural networks. Energy 155:705–720
11. Al-Hamadi HM, Soliman SA (2005) Long-term/mid-term electric load forecasting based on
short-term correlation and annual growth. Electric Power Syst Res 74(3):353–361
12. Lima RM, Novais AQ, Conejo AJ (2015) Weekly self-scheduling, forward contracting, and
pool involvement for an electricity producer. An adaptive robust optimization approach. Eur. J
Oper Res 240(2):457–475
13. Berard A (2012) TELECOM ParisTech Master of Engineering Internship at EDF
R&D, Clamart, France. https://archive.ics.uci.edu/ml/datasets/individual+household+electric+
power+consumption
14. Jaeger H (2001) The, “echo state” approach to analysing and training recurrent neural networks-
with an erratum note. Bonn, Germany: German Nat Res Center Inf Technol GMD Technical
Report 148(34):13
15. Lacy SE, Smith SL, Lones MA (2018) Using echo state networks for classification: a case
study in Parkinson’s disease diagnosis. Artif Intell Med 86:53–59
16. Chouikhi N, Ammar B, Rokbani N, Alimi AM (2017) PSO-based analysis of echo state network
parameters for time series forecasting. Appl Soft Comput 55:211–225
17. Ma Q, Shen L, Chen W et al (2016) Functional echo state network for time series classification.
Inf Sci 373:1–20
18. Xu X, Niu D, Fu M et al (2015) A multi time scale wind power forecasting model of a chaotic
echo state network based on a hybrid algorithm of particle swarm optimization and tabu search.
Energies 8(11):12388–12408
19. Wang L, Lv SX, Zeng YR (2018) Effective sparse adaboost method with ESN and FOA for
industrial electricity consumption forecasting in China. Energy 155:1013–1031
20. Maiorino E, Bianchi FM, Livi L et al (2017) Data-driven detrending of nonstationary fractal
time series with echo state networks. Inf Sci 382:359–373
Households Energy Consumption Forecasting with Echo State Network 1315

21. Chen Q, Shi L, Na J et al (2018) Adaptive echo state network control for a class of pure-feedback
systems with input and output constraints. Neurocomputing 275:1370–1382
22. Dahl GE, Sainath TN, Hinton GE (2013) Improving deep neural networks for LVCSR using
rectified linear units and dropout. In: 2013 IEEE international conference on acoustics, speech
and signal processing, vancouver, BC, p 8609−8613. https://ieeexplore.ieee.org/stamp/stamp.
jsp?tp=&arnumber=6639346&isnumbe=6637585
23. Ruiz CW, Perapoch J, Castillo F, Salcedo S, Gratacós E (2006) Bessons monocoriònics afectes
de transfusió fetofetal: conseqüències a curt i llarg termini. Pediatria Catalana 66(2):53–61
24. Cho K, van Merrienboer B, Bahdanau D, Bengio Y (2015) On the properties of neural machine
translation: encoder–decoder approaches. pp 103–111. https://doi.org/10.3115/v1/w14-4012
Management of Multi-agents in a Smart
Grid Network with the Python Using
the Contract Net Protocol

Saber Ikram, El Bachtiri Rachid, Bendali Wadie, Boussetta Mohammed,

El hammoumi Karima, and Cheddadi Fatima

Abstract Smart grids are one of the best solutions for the integration of different
distributed energy sources (DES), load, and storage elements. However, optimizing
and managing these systems is a big challenge because it is only done in a distributed
way. One of the best methods of managing a smart grid is the use of artificial intelli-
gence (AI), especially multi-agent systems (MAS) approaches. This work constitutes
a demonstration of one of those methods which is the contract net protocol (CNP).
We will use the extended version of the contract net protocol(ECNP) to optimize
the movement of MASs inside a smart grid network. Our example of a smart grid
will be a grid constituted from nodes, and each node is a location of an element of
the smart grid which could be a wind turbine, a customer…, in other words, a task
or/and an agent. In the scenario each task it’s given a target which is also a node
inside the grid, and our main objective is that the agent or agents could contribute
to each task, depending on their priority order and may allow it to reach its target,
which is another node in the network far from its initial location, using the shortest
way as possible.

Keywords Microgrid · Multi-agent systems · Contract net protocol

1 Introduction

The great and fast development of renewable energy sources (RESs) [1, 2] encouraged
researchers to begin forming different size units provides as microgrids [3]. They

S. Ikram (B) · E. B. Rachid · B. Wadie · E. Karima

Laboratory: Technologies et Services Industriels, University Sidi Mohammed Ben Abd Allah,
High school of technology of fez (ESTF), Fez, Morocco
e-mail: [email protected]
B. Mohammed
Laboratory: Technologies Innovantes, University Sidi Mohammed Ben Abd Allah, High school of
technology fez (ESTF), Fez, Morocco
C. Fatima
Laboratory: Systèmes Intelligents, Géoressources et Énergies Renouvelables, Faculty of Science
and Technologies Fez (FSTF), Fez, Morocco
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1317
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_120
1318 S. Ikram et al.

were presented as a solution for the reliable integration of distributed energy resources
[4, 5], and resilience-related problems [6, 7] as shown in Fig. 1.
Developing such outstanding units of energy has so many economic and environ-
mental benefits [8, 9] as reliability because they improve consumers’ supply [10, 11],
their average interruption frequency, duration between consumers and system, and
resilience which is the capability of the system to minimize possible power outage
and returning to stable status [12].
With the start-up of microgrids, different types of control architectures have been
defended, starting with the very first ones as centralized and decentralized procedures.
That kind of approach depends on how many components the microgrid is constituted
of and its dimension, so their complexity increase despite it. Each architecture has
a particular number of levels, and every level supports unique tasks, as maintaining
stability or load sharing. However, the major challenge with microgrids is to provide
efficient management techniques with a high level of optimization results for system
efficiency, minimizing the use and processing of data, and, of course, independence in
management [13, 14]. Yet, those architectures didn’t bring much satisfaction to their
inventors, due to their high cost and management’s difficulties, the reason behind
swishing into decentralized architectures.
On account of the arguments mentioned above, added to the important number
of DERs, loads, and storing devices implemented on MGS [12, 15, 16], shifting into
distributed control methods has been motivated [17], as agent technology has given
their many advantages like using local data exchanges between neighbors, especially
the multi-agents systems (MASs) approaches [18–20], to optimize the operation of
local power production, and improve performances of microgrid using decentralized
control [11].
MASs have been defined in many ways in literature, as «a comprises of two or
more independent agents with some information to achieve a set target» [21], or
autonomous decision-making entities [22]. They are known for their capability of
reaction in the environment, intercommunication inside the network, and high level
of autonomy.

Fig. 1 Example of a modern distribution network

Management of Multi-agents in a Smart Grid Network ... 1319

Yet, to reach those advantages and gain from MASs qualities, we should overcome
technical challenges and problems, which are quite numerous [21, 22], starting with
MASs platform choices, communication languages, architecture, and design while
respecting the Foundation for Intelligent Physical Agents standards [23] shown in
Fig. 2. In other words, insuring compatibility, robustness, flexibility, extensibility
plus the capability of interacting with each other irrespective of the platform they
run on [22]. One of those techniques will be presented below.
This article is organized as follows: in the second section we will present a defi-
nition of the contract net protocol, the technique used in our work to optimize the
movement of the agents, after that, we will explicitly explain our work with different
scenarios, tables, and simulation showing the movement inside the smart grid, at last
in the third section we will end up with a conclusion and a discussion.

2 Contract Net Protocol

2.1 The Organizational Architecture

Contract net protocol (CNP) is a technique invented to solve communication prob-

lems in distributed networks, it applies to microgrid of all types and it gave very
good results on the cost side, management, and speed of data processing. It was
defined as a high-level protocol for communication among the nodes in a distributed
problem solver [24, 25]. Modified by the FIPA in 2000, to become a more satisfying,
sophisticated protocol, by adding rejection and communication acts [26].
This protocol is applied essentially to perform negotiation tasks [27] while using
exchange messages in a bidding process.
In this protocol, agents can play manager or bidder role, divided into two agent
types, the initiator and the participants:
Initiator: is the manager who is occupied by commissioning a task or demanding
a service.
Participants: are the elements of the networks which can collect manager’s calls
for proposals and make them propositions depending on the call nature.
Explication: The first message sent from the initiator to the participants is a call
for proposals, the answer to this call can be divided into two:
– To refuse, by sending a not understood message.
– To accept, by making a proposition. In this case, the initiator receives their propo-
sitions and chooses among them which to accept. After a lapse of time, the partic-
ipant which made the accepted proposition informs the initiator of the result, it
could be a collapse, and it means that the task is concluded, or that the service
asked for is executed.
The work presented in this paper constitutes an optimization of agents shifting
inside a smart-grid, as it is known, agents have to move inside the smart-grid, being a
1320 S. Ikram et al.

software or a hardware agent, it has to move between smart-grid elements to receive

tasks and provide solutions in the shape of bides.
In the next paragraph, a very detailed explication will be provided through
scenarios, the results of our program will be exposed in tables extracted from the
program and simulations.

2.2 Different Scenarios Used in Our Work and Their

Experiment Results

Before explaining our first scenario, we will clarify our program configuration and
the implementation structure:

Program Configuration
Our configuration is established of six major scripts:

– Agent script: written as an index of vocabularies respectively introducing a set

of attributes describing an agent (identification number, launching cell, priority
order, speed, leaning time, and max capacity).
– Tasks: written as an index of vocabularies respectively introducing a set of
attributes describing a task (identification number, initiating moment, target cell).
– Initiators: written as an index of vocabularies respectively introducing a set
of attributes describing an initiator (identification number, task’s identification
number)
– Smart grid network: written as an index of vocabularies respectively introducing
a set of attributes describing the smart grid network (quantity of columns rows in
which tasks and agents are placed).
– Simulation script: it contains the main instructions for all scripts simulation at
once.
– Experiment script: at this level, we use the scenarios’ arguments and conditions
to simulate our scripts.

The inputs of our system as shown in the scripts are the tasks and the outputs are
the bides provides by the agents at each scenario.

About Scenario_1_Part_AG:
The first scenario is developed to test our version of ECNP for tasks and agent’s
movement inside a smart grid network. We choose to use one agent and three tasks.
Every task comes at a specific time identification(id), as known that our agent has
a special order of priority, we spread tasks in the opposite order of that priority, the
main goal of that is to show the capability of the agent to bid and contract for a task
no matter its importance for him (Table 1).
1) Task used in the first scenario:
Management of Multi-agents in a Smart Grid Network ... 1321

As shown in the table above, every target is introduced in a special time id, and each
one has its aim that it should reach to complete its role inside the smart grid network
(Table 2).
2) The agent used in the first scenario:
The second table presents our agent’s attributes and characteristics. Depending on
those qualities the agent can bid and contract for every task and offer the best path
between the task first position and its aim position.
3) Smart grid network used in the first scenario studied:
The grid extracted from our simulation, as shown in Fig. 2, illustrates the agent and
task’s initial positions and their aims as cited in the table above. We choose a network
from three rows and three columns, the results extracted from the first pdf document
are outlined in the next table, and the simulation presents how the agent could bide
for each task without failures (Table 3 and Fig. 3):
4) About scenario_2_part_ag:
The second scenario is developed to test the use of two agents at the same time, with
three tasks.

Table 1 Task’s properties

Task identification(id) Introduction time Aim position
2 0 (2,2)
0 1 (1,1)
1 2 (0,0)

Table 2 Agent’s properties

Agent (id) Agent departure position Agent priority speed Leaning time Max capacity
0 (0,0) [2, 1, 0] 1 1 3

Fig. 2 Smart grid network

used in the first scenario
1322 S. Ikram et al.

Table 3 Agent’s bid for each

Simulation steps Tasks Agent 0 provides
task
0 0 5
1 0 5
1 1 2
2 0 5
2 1 2
2 2 10
3 0 5
3 1 2
3 2 10
4 0 5
4 2 2
4 1 10

Fig. 3 Simulation of the agent bides using the CNP

The main goal of this scenario is to demonstrate the capability of the agents to
bid and contract for the tasks, without interfering in each other job, and the contract
came at a specific time, depending on that time, one of the agents took the task and
provide bids, so it could be completed, depending on its priorities, rest time and
speed (Table 4 and Table 5, Fig. 4).
5) Task used in the SECOND scenario:

6) The agent used in the SECOND scenario:

Table 4 Task’s properties

Task Identification(id) Introduction time Aim position
0 0 (3,0)
1 1 (3,4)
2 2 (2,2)
Management of Multi-agents in a Smart Grid Network ... 1323

Table 5. Agent’s properties.

Agent (id) Agent departure position Agent priority speed Leaning time Max capacity
0 (0,0) [2,0,1] 1 1 2
1 (0,4) [2,0,1] 1 1 2

Fig. 4 Smart grid network

used in the second scenario

7) Smart grid network used in the SECOND scenario studied:

The grid extracted from the second simulation of the second scenario is wider and
more complicated than the first one. It is clear that in our first work, we choose to
optimize only the movement of the agents inside the grid network, but it is known in
every smart-grid, agents been hardware or software they need to move between the
components of that grid, so came the importance of our work which could facilitate
this job for the agents (Table 6, Fig. 5).

Table 6 Agent’s bid for each

Simulation steps Tasks Agent 0 provides Agent 1 provides
task
0 0 4 10
0 1 10 16
1 0 4 10
1 1 10 10
1 2 5 5
2 0 4 10
2 1 10 4
2 2 5 5
3 0 4 10
3 1 10 4
3 2 5 5
1324 S. Ikram et al.

Fig. 5 Simulation of the

agent bides using the CNP

3 Discussion

At is shown in the tables below, extracted as a simulation for the different bids, our
agents have contracted for the three tasks in every step time of the first and second
scenario, the ECNP allowed them to provide many movements for each task to help
it to get to its target easily and without forgetting any of them.
We chose these particular scenarios to show that MAS are perfectly capable of
managing any size of smart grid starting with three rows and columns and ending
with a much bigger number of elements.
Using his priority, resting time, and capacity, which are variable and could be
changed from one scenario to another, the agent has the capability of managing very
well movement inside a smart grid network, without interfering in the other agent
job, executing orders and tasks in the disposal that he has been given and the most
important without forgetting any.
The simulations show that the bides are very clear and separated from each other,
which explain that the speed, capacity and rest time for every agent is respected
in both scenarios, and every bides come at a specific time and don’t collapse with
another, this property is very important of good work of a smart-grid.

4 Conclusion

The application of the extended version of contract net protocol for the movement
of agents in a smart grid of different sizes, and with a more challenging number
of agents and elements, has shown the ability of agents to manage objectives and
accomplish tasks, which can be used as a technique to forecast energy production
inside a wind farm or a photovoltaic field.
In other words, MAS are capable of achieving real-time management of distributed
systems as a microgrid, creating from it a smart-grid with self-reflections every
moment of the day, using the proprieties of MAS mentioned above, it can provide
Management of Multi-agents in a Smart Grid Network ... 1325

energy for a very large number of customers, also it can maintain its secure level of
energy by using various types of RES it can acces to.

References

1. Zhou B, Lei Y, Li C, Fang B, Wu Q, Li L, Li Z (2019) Electrical LeaderRank method for

node importance evaluation of power grids considering uncertainties of renewable energy. Int
J Electr Power Energy Syst 106:45–55. https://doi.org/10.1016/j.ijepes.2018.09.034
2. Steffen B (2018) The importance of project finance for renewable energy projects. Energy Econ
69:280–294. https://doi.org/10.1016/j.eneco.2017.11.006
3. Olivares DE, Mehrizi-Sani A, Etemadi AH, Canizares CA, Iravani R, Kazerani M, Hajimiragha
AH, Gomis-Bellmunt O, Saeedifard M, Palma-Behnke R, Jimenez-Estevez GA, Hatziargyriou
ND (2014) Trends in microgrid control. IEEE Trans Smart Grid 5(4):1905–1919. https://doi.
org/10.1109/TSG.2013.2295514
4. Martin-Martínez F, Sánchez-Miralles A, Rivier M (2016) A literature review of Microgrids:
a functional layer based classification. Renew Sustain Energy Rev 62:1133–1153. https://doi.
org/10.1016/j.rser.2016.05.025
5. Hirsch A, Parag Y, Guerrero J (2018) Microgrids: a review of technologies, key drivers, and
outstanding issues. Renew Sustain Energy Rev 90(April):402–411. https://doi.org/10.1016/j.
rser.2018.03.040
6. Jirdehi MA, Tabar VS, Ghassemzadeh S, Tohidi S (2020) Different aspects of microgrid
management: a comprehensive review. J Energy Storage https://doi.org/10.1016/j.est.2020.
101457
7. Ruotolo MA (2018) A social cost-benefit analysis of community microgrid systems in New
York State. https://udspace.udel.edu/handle/19716/23942
8. Adefarati T, Bansal RC (2019) Reliability, economic and environmental analysis of a microgrid
system in the presence of renewable energy resources. Appl. Energy 236:1089–1114. https://
doi.org/10.1016/j.apenergy.2018.12.050
9. Wang S, Li Z, Wu L, Shahidehpour M, Li Z (2013) New metrics for assessing the reliability
and economics of microgrids in distribution system. IEEE Trans Power Syst 28(3):2852–2861.
https://doi.org/10.1109/TPWRS.2013.2249539
10. Khodaei A (2014) Resiliency-oriented microgrid optimal scheduling. IEEE Trans Smart Grid
5(4):1584–1591. https://doi.org/10.1109/TSG.2014.2311465
11. Logenthiran T, Srinivasan D, Khambadkone AM (2011) Multi-agent system for energy resource
scheduling of integrated microgrids in a distributed system. Electric Power Syst Res 81(1):138–
148. https://doi.org/10.1016/j.epsr.2010.07.019
12. Logenthiran T, Srinivasan D, Khambadkone AM, Aung HN (2010) Scalable Multi-Agent
System (MAS) for operation of a microgrid in islanded mode. In 2010 Joint International
Conference on Power Electronics, Drives and Energy Systems, PEDES 2010 and 2010 Power
India. https://doi.org/10.1109/PEDES.2010.5712459
13. Reilly JT (2019) From microgrids to aggregators of distributed energy resources. The microgrid
controller and distributed energy management systems. Electricity J 32(5):30–34. https://doi.
org/10.1016/j.tej.2019.05.007
14. Lasseter R, Akhil A, Marnay C, Stephens J, Dagle J, Guttromson R, Eto J et al (2002)
The CERTS microgrid concept, white paper on integration of distributed energy resources.
California Energy Commission, Office of Power Technologies-US Department of Energy,
LBNL-50829, https://Certs. Lbl. Gov, 29
15. Hatziargyriou N, Jenkins N, Strbac G, Pecas Lopes JA, Ruela J, Engler A, Amorim A et al
(2006) Microgrids - Large scale integration of microgeneration to low voltage grids. In 41st
International Conference on Large High Voltage Electric Systems 2006, CIGRE 2006 (2018)
1326 S. Ikram et al.

16. Liu G, Jiang T, Ollis TB, Zhang X, Tomsovic K (2019) Distributed energy management
for community microgrids considering network operational constraints and building thermal
dynamics. Appl Energy 239:83–95. https://doi.org/10.1016/j.apenergy.2019.01.210
17. Sermakani S, Thangaraja M (n.d.) Power demand optimization in smart grid via wireless
networks. J Electr Electron Eng (IOSR-JEEE), 18–23
18. Ghiani E, Serpi A, Pilloni V, Sias G, Simone M, Marcialis G, Armano G, Pegoraro P (2018)
A multidisciplinary approach for the development of smart distribution networks. Energies
11(10):2530. https://doi.org/10.3390/en11102530
19. Worighi I, Maach A, Hafid A, Hegazy O, Van Mierlo J (2019) Integrating renewable energy
in smart grid system: Architecture, virtualization and analysis. Sustain Energy Grids Netw
18:100226. https://doi.org/10.1016/j.segan.2019.100226
20. Kamdar R, Paliwal P, Kumar Y (2018) A state of art review on various aspects of multi-agent
system. J Circ Syst Comput 27(11):1–5. https://doi.org/10.1142/S0218126618300064
21. FIPA standards, https://www.fipa.org/Smith R (1980) Communication and control in problem
solver. IEEE Trans Comput 29(12):12. https://www.csupomona.edu/~ftang/courses/CS599-
DI/notes/papers/contractnetprotocol.pdf
22. PankiRaj JS, Yassine A, Choudhury S (2019) An auction mechanism for profit maximization of
peer-to-peer energy trading in smart grids. Procedia Comput Sci 151(2018):361–368. https://
doi.org/10.1016/j.procs.2019.04.050
23. Smith RG (1988) The contract net protocol: high-level communication and control in a
distributed problem solver. Morgan Kaufmann, pp 357–366
24. FIPA. Fipa contract net interaction protocol specification. Technical, Report SC00029H, FIPA
(2000)
25. Bicocchi N, Cabri G, Leonardi L, Salierno G (2019) A survey of the use of software agents
in digital factories. In 2019 IEEE 28th International Conference on Enabling Technologies:
Infrastructure for Collaborative Enterprises (WETICE), 3–8 https://doi.org/10.1109/wetice.
2019.00010
26. Mařík V, Kadera P, Rzevski G, Zoitl A, Anderst-Kotsis G, Tjoa AM, Khalil I (Eds) (2019)
Industrial applications of holonic and multi-agent systems. Lecture Notes in Computer Science.
https://doi.org/10.1007/978-3-030-27878-6
27. Li K, Zhou T, Liu B, hai, & Li, H. (2018) A multi-agent system for sharing distributed
manufacturing resources. Expert Syst Appl 99(32):43. https://doi.org/10.1016/j.eswa.2018.
01.027
Smart Grid Operation Using Hybrid AI
Systems

Chakib Alaoui and Hajar Saikouk

Abstract Battery storage system is an essential component for the depletion of the
microgrid, and of the smart grid on a larger scale. It acts as an interface between
the highly intermittent renewable energy resource such as solar photovoltaic or wind
energy system, the load, such as residential homes and businesses, electric vehicle
fleets and the electric power grid. In principle, battery storage is intended to store
excess electricity production from the renewable resource. The energy stored is then
typically discharged in the evening when demand is higher, or sold to the grid when
the available electricity is higher than the demand. However, the implementation of
such a system becomes more complicated when taking into account the stochastic
nature of the renewables, the variations of the load demand and the electric vehicle
availability. In this project, a fuzzy logic management system, optimized through
genetic algorithm process, is proposed. It seeks to obtain a fully functional microgrid
that minimizes electricity consumption from the utility grid, while offering ancillary
services to the grid such as peak shaving. These objectives are accomplished by
maximizing the energy input from the renewables and optimizing the utilization
of the energy storage system. Finally, the microgrid’s control main objective is to
maintain a constant voltage at the DC bus under all operating conditions. Initial
simulations and experimental data show some promising results.

Keywords Microgrid · Battery storage system · Smart grid · Artificial

intelligence · Fuzzy logic · Genetic algorithm

1 Introduction

The integration of renewable energy resources (RES) in a microgrid system constitute

an effective solution to improve reliability, resilience and the economics of the smart
grid systems [1]. However, the increasing development of RES with uncertain outputs
and increasing the number of microgrids makes the operation and control of them a

C. Alaoui (B) · H. Saikouk

INSA EuroMediterannée, Euromed University, Fez, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1327
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_121
1328 C. Alaoui and H. Saikouk

major challenge, especially with the inclusion of electric vehicles into the microgrid
[2–4]. The adjustment of a DC microgrid voltage with the use of classical controllers
in such environment is a challenging work and requires new and appropriate control
strategies [5, 6]. Moreover, the battery-based electric energy storage (EES) play a
key role for an efficient smart grid [7, 8] or microgrid management, in addition to
the integration of electric vehicles (EV) into the grid through an effective battery
management [9, 10].
Recently, fuzzy logic (FL) based controllers have been developed and used in
various industrial applications [11]. The performance of these controllers is influ-
enced by fuzzy rules and control coefficients that determine their level of control-
lability [12]. Compared to other adaptive control methods such as artificial neural
networks (ANN) [13], FL controllers have a special place in the industry and have
been used in a variety of applications such as electric machine ripple minimiza-
tion [12], energy management system in microgrid [14], automatic voltage regulator
system [15], load frequency control [16] and others [17, 18].
Several methods have been used to provide a solution to efficiently manage the
energy flow of a smart grid while reducing its overall costs. These methods vary from
rule-based controllers through optimization techniques to machine learning methods
with different objectives. The author of [19] proposed a rule-based controller to
dispatch a battery system with a PV and a grid-connected residential house. The
objective of the controller was to maximize the energy delivered from the PV. The
results showed a significant reduction from the power taken from the grid. The author
of [20] implemented Mixed Integer Linear Programming (MILP) for a similar system.
They used a weighted multi-objective optimization to minimize energy consumption,
costs and peak power demands. They achieved a reduced peak demand of 50%
and a reduced cost of 20%. While the authors of [21] used a similar system with
the objective of reducing the overall costs and CO2 emission. Genetic algorithms
have been used by [22, 23]. They achieved higher system efficiency and higher self-
consumption from PV panels. Many universities and research centers, such as in [24,
25], have used machine learning techniques for microgrid control. Their objective
was to maximize the usage of the battery bank and the energy input from a PV system.
The authors of [26] used 2-step ahead Q-learning system to manage a grid-tied home
having a battery pack and a wind turbine. The results of this study revealed that
while the use of the battery and wind turbine to supply the load improved by several
percent’s after the agent had learned, the economic benefit was less than 1%. Finally,
the author of [27] used a Q-learning parametrized by neural networks in order to
control a system consisting of PV, battery and hydrogen storage. They have been
able to reduce the costs of imports from the grid.
The aim of this project is to implement a fuzzy logic system, whose parameters
are optimized with genetic algorithm, that can control the charge scheduling of a
grid-tied battery, photovoltaic and wind energy residential system, with the objective
of taking the minimum energy from the grid, while making the grid to benefit from
ancillary services such as peak demand shaving. Two scenarios will be considered:
an electric vehicle connected to the smart grid, and disconnected from the smart grid.
Finally, the validation of the proposed algorithm was done in Morocco, where the
Smart Grid Operation Using Hybrid AI Systems 1329

Fig. 1 Block diagram of the proposed microgrid system

electricity tariff is not in hourly rate, but rather in monthly rate, where the electricity
rates vary monthly from ‘low-power’, ‘high power’ and ‘peak power’ consumption
periods [28].

2 System Description

2.1 Microgrid System

The general block diagram of the microgrid is shown in Fig. 1. The main components
of the proposed system are a PV system (PVS) with appropriate MPPT system, and a
feedback controlled boost converter used to connect the PV system to the 750 V DC
bus. In addition, a wind power system (WPS) equipped with a Permanent Magnet
Synchronous Generator (PMSG) and appropriate converters and MPPT system. A
battery based energy storage system (ESS) connected with a bidirectional DC-DC
converter to the main DC bus. An electric vehicle (EV) connected to the DC bus
through a battery charger/discharger. Finally, an AC load and a grid connection.

2.2 Genetic-fuzzy Control

Fuzzy logic systems (FLS) possess many interesting features allowing it to manage
the energy flow within a microgrid having stochastic units such as PVS and WES,
1330 C. Alaoui and H. Saikouk

Fig. 2 Fuzzy-genetic control system

in addition to a load that contains an EV. FLS is model free, rule-based approach
incorporating linguistic description of a highly non-linear system.
However, fine-tuning a FLS necessitates a trial-and-error process that may be
long, expensive and tedious. In order to establish the rules, some prior knowl-
edge or experience is needed. Sometimes, such experience or prior knowledge are
unavailable.
Genetic algorithms (GAs) are sometimes used to optimize the FLS rules.
Depending on the available information and on the system’s settings, GA can be
used to optimize the fuzzy rules and/or the fuzzy memberships. Fig. 2. shows the
block diagram of such a system.

2.3 Photovoltaic System

Conventional photovoltaic modules are typically rated between 200 and 300 W with
an open circuit voltage range of 20–30 V. For the PV-grid charging system, the
modules are arranged in series-parallel strings to attain the required working voltage,
current and power. DC-DC converters are connected as an interface between the PV
array and the DC bus. In order to transfer the maximum power to the DC bus at all
times, two converters are used as shown in Fig. 3. The first converter, referred to as
the boost DC-DC converter, tracks the MPPT of the PV panels in real time. At its
output, the maximum power is transferred, but with varying output voltage value,
which correspond to different irradiation and temperature values. This is shown in
Fig. 4. In order to maintain a constant voltage value at the output of the PV system,
a second DC-DC converter is added, and with appropriate feedback control, in order
to maintain 750 V at the output and hence connect it to the DC Bus. The switching
ON/OFF of the PV panels to the grid is decided by the second stage converter; if
Smart Grid Operation Using Hybrid AI Systems 1331

Fig. 3 PV panel connected to the DC bus through a first stage converter for the MPPT, and the
second stage to maintain 750 VDC at the output

Fig. 4 PV panel output

voltage vs. solar irradiation
and ambient temperature

the voltage input to the second inverter is below a threshold V PV,th such that the 750
V cannot be established at the output DC-DC converter and the feedback controller,
the fuzzy control system will disconnect the PV system from the DC bus.

2.4 Wind Power System

The wind power system (WPS) consists of a Permanent Magnet Synchronous Gener-
ator (PMSG) rotated by a wind turbine which is interfaced to the DC microgrid
through a proper AC-DC power converter. Eq. (1) shows the mechanical power PM
available at the shaft of the turbine.
1332 C. Alaoui and H. Saikouk

Fig. 5 Typical Cp-λ curve

for three bladed wind turbine

Fig. 6 PMSG – MPPT

control diagram

PM = 0.5ρ AVW3 C P (β, λ) (1)

where ρ is the air density (kg/m3 ), A is the cross-sectional area of the stream (m2 )
and V w is the free wind velocity (m/s). For a small wind turbine with fixed pith angle,
C P is solely a function of the tip speed ratio λ expressed in (2).

ωr ot R
λ= (2)
VW

where, R is the radius of the rotor (m) and ωrot is the rotational speed of the rotor
(rad/s). A typical C p -λ curve is shown in Fig. 5.
Incremental MPPT technique was selected for this wind turbine. Its main advan-
tage is its independence from the wind turbine parameters. This algorithm solely
requires power output measurement for operation. This can be readily achieved from
the DC link current and voltage without requiring any speed or wind measurements.
Fig. 6. shows the algorithm of the MPPT. It uses the perturb and observe principle
to increase the delivered power by increased the duty cycle of the PWM signal sent
to the boost converter of Fig. 7.

2.5 Battery-based Energy Storage System (BESS)

The battery ESS can only be charged whenever there is a surplus power from
the PVS or WTS. The state of charge (SOC) of the ESS is expressed in Eq. (3)
as [26, 29].
Smart Grid Operation Using Hybrid AI Systems 1333

Fig. 7 MPPT algorithm for

the wind turbine

Fig. 8 BESS connection to

the DC bus block diagram


n 
n
S OC(t) = S OC(0) + ηc Pch,i(t) − ηd Pdis,i(t) (3)
t=1 t=1

where ηc and ηd are the charging and discharging efficiencies of the battery pack,
Pch,i(t) and Pdis,i(t) are the power for charging and discharging the batteries at time t.
Fig. 8 shows the block diagram of the battery based energy storage system (ESS).
The DC-DC converter maintains the voltage output at 750V as long as the SOC of
the batteries allows it.
1334 C. Alaoui and H. Saikouk

Fig. 9 EV connection to the

microgrid

2.6 V2G Block Diagram

Figure 9 shows the bidirectional DC-DC converter that will act as a battery charger,
or as a battery discharger that will transfer the battery pack energy into the grid
through the DC bus.

3 Microgrid Operation

In order to establish the appropriate rules for the operation and control of the proposed
microgrid, the following guidelines were established:
i. The ESS and the EV can be recharged according to two modes: ‘NORMAL’
or ‘URGENT’
ii. Either the ESS or the EV can be charged in ‘URGENT’ mode, they cannot be
operating in this mode simultaneously
iii. The fuzzy controller reads the voltage levels at each component in order to
estimate its power and energy status
iv. The EV cannot be recharged from the ESS. This would incur important energy
losses due to the cycling efficiencies of the batteries.
v. The power demand from the load PL must be met at all times.
For the proposed microgrid, there are seven possible modes of operation, as
described below and illustrated in Fig. 10. In each mode of operation, the fist equa-
tion expresses the condition necessary to pass to the mode of operation as in Eqs. (4,
6, 8, 10, 12, 14, 16), and the second equation shows the power and hence the energy
transferred, this is shown in Eqs. (5, 7, 9, 11, 13, 15, 17). The DC link voltage is used
as the criteria for switching between the seven modes of operation.
Mode 1: The power generated by the renewable sources and the power available in
storage devices BESS and EV is higher than the power demanded by the load (SOC EV
> 50%). In addition, this mode of operation matches peak power demand from the
grid. Satisfying the Grid peak demand and load demand, EV being connected and
Smart Grid Operation Using Hybrid AI Systems 1335

discharging. The condition to pass to this mode is expressed in (4), and the power
received by the grid is (5).

Ppv + Pwp + Pess + Pev > PL (4)

PG = Ppv + Pwp + Pess + Pev − PL (5)

Mode 2: This mode of operation is similar to mode 1, except that the EV needs to
be recharged (SOC EV < 50%). Satisfying the Grid peak demand and load demand, EV
being connected and charging. The grid receives power expressed in (7) as ancillary
support.

Ppv + Pwp + Pess > PL + Pev (6)

PG = Ppv + Pwp + Pess − Pev − PL (7)

Mode 3: This mode of operation is also similar to mode 1, except that the EV is
disconnected from the microgrid. Satisfying the Grid peak demand and load demand,
EV being disconnected

Ppv + Pwp + Pess > PL (8)

PG = Ppv + Pwp + Pess − PL (9)

Mode 4: When enough power is received from the renewable sources and the
BESS is below a threshold (SOC ESS < 50%), the available energy is used to recharge
the BESS assuming that there is no peak demand from the grid. Moreover, the EV
is not connected in this mode of operation. Recharging ESS and load demand, EV
being disconnected

PL < Ppv + Pwp (10)

Pess = Ppv + Pwp − PL (11)

Mode 5: When the BESS is almost fully discharged (SOC ESS < 10%), it must
include the power from the grid in order to reach SOC ESS = 50%. Beyond 50% of
recharge, the BESS is recharged in mode4. Recharging ESS in urgent mode, EV
being disconnected

PL > Ppv + Pwp (12)

Pess = PG + Ppv + Pwp − PL (13)

1336 C. Alaoui and H. Saikouk

Table 1 Microgrid modes of

Modes Services
operation
1, 2, 3 Grid peak power demand
4, 5 Recharge of the ESS
6, 7 Recharge of EV

Fig. 10 Microgrid modes of

operation

Mode 6 : Recharging EV in normal mode

PL < Ppv + Pwp (14)

Pev = Ppv + Pwp − PL (15)

Mode 7: When the energy available in the EV is below a threshold (SOC EV <
10%) and very little to no energy input from the renewables, then the EV should be
recharged in urgent mode by using power from the grid. Recharging EV in urgent
mode

PL > Ppv + Pwp (16)

Pev = PG + Ppv + Pwp − PL (17)

Table 1 recapitulates the operating modes of the microgrid.

4 Microgrid Simulation

The system under study consists of a grid-connected house situated in the north
of Morocco. It is equipped with photovoltaic panels (PV) rated at 7 KW, a wind
Smart Grid Operation Using Hybrid AI Systems 1337

turbine generator (WTG) of 5 KW and an electric storage system (ESS) of 14 KHh

of capacity. Lastly, the house owner is supposed to possess an electric vehicle (EV)
to be connected to the system under study when needed. In order to connect these
components, it is necessary to add an inverter between the AC bus and the PV and
the battery pack. The EV has its own on-board inverter, hence it can be charged
directly from the AC bus. In order to charge the battery from the grid, it is necessary
to add a rectifier circuit. Finally, all the inverter circuits must provide electricity at
the same frequency as the grid. All conversions incur losses, however, the converters
were modeled together and their inefficiencies were assumed to be 90% and were
included in the models of the PV, wind turbine and battery systems.
Figure 11 shows the household electricity demand for one whole year. It shows
the overall trend and seasonal variations. This house is supposed to demand about
15000_KWh of electricity annually.
Figure 12 shows the output of the PV panels used in this project. Each data point
denotes the electricity provided by the panels over a ten-minute interval.
Figure 13 shows the wind data used to simulate the wind power system.
Figure 14 shows the simulation result of the DC bus voltage that was regulated to
750 V. At 0.75 and 1.5 s, switching between modes of operations happen, and spikes
in output voltage values result as a consequence.

Fig. 11 Electricity demand

data points with trend line
illustrating the seasonal
variations

Fig. 12 Energy output from

PV panels
1338 C. Alaoui and H. Saikouk

Fig. 13 Wind speed data

Fig. 14 DC bus voltage

5 Conclusion

This paper has presented the initial results of a FLC-based design of an energy
management system for smoothing the voltage level of a DC microgrid. It contains
renewable resources such as PV and Wind power systems, an EV, an energy storage
system and a grid-connected load.
The proposed approach is implemented on a microgrid simulation model to vali-
date the proposed methodology. Simulation of real-condition operation including
experimental data has been performed. The preliminary results of this study has
shown that it is possible to train an RL agent using a continuous action algorithm
to control the charging scheduling of a battery in a residential setting. Initial results
showed the feasibility of this agent, although the reward function needed further
tuning.
Smart Grid Operation Using Hybrid AI Systems 1339

References

1. Teimourzadeh S, Aminifar F, Davarpanah M, Shahidehpour M (2018) Adaptive control of

microgrid security. IEEE Trans Smart Grid 9:3909–3910
2. Sadabadi MS, Shafiee Q, Karimi A (2017) Plug-and-play voltage stabilization in inverter-
interfaced microgrids via a robust control strategy. IEEE Trans Control Syst Technol 25:781–
791
3. Sassi HB, Errahimi F, Essbai N, Alaoui C (2019) V2G and Wireless V2G concepts: state of
the art and current challenges. In: 2019 International Conference on Wireless Technologies,
Embedded and Intelligent Systems (WITS) pp 1–5 IEEE
4. Diouri O et al (2019) Modeling and design of single-phase pv inverter with mppt algorithm
applied to the boost converter using back-stepping control in standalone mode. Int J Photoenergy
2019
5. Amoateng DO, Hosani MA, Elmoursi MS, Turitsyn K, Kirtley JL (2018) Adaptive voltage and
frequency control of islanded multi-microgrids. IEEE Trans Power Syst 33:4454–4465
6. Diouri O et al (2019) Control of single phase inverter using back-stepping in stand-alone
mode. In: 2019 International Conference on Wireless Technologies, Embedded and Intelligent
Systems (WITS). IEEE
7. Alaoui C (2017) Thermal management for energy storage system for smart grid. J Energy
Storage 13:313–324
8. Chakib A, Mhamed Z, Saidou N (2017) Towards the implementation of refurbished EV lithium-
ion batteries for smart grid energy storage. In: 2017 Intelligent Systems and Computer Vision
(ISCV), pp 1-5. IEEE
9. Alaoui C, Salameh Z (2007) Electric vehicle diagnostic and rejuvenation system (EVDRS).
Int J Power Energy Syst 27(2):151
10. Alaoui C, Salameh ZM (2006) A novel system-on-chip system for diagnostic & rejuvenation
for electric & hybrid vehicles. In: International Symposium on Power Electronics, Electrical
Drives, Automation and Motion, 2006. SPEEDAM, pp 920–925. IEEE
11. Xiang X, Yu C, Lapierre L, Zhang J, Zhang Q (2018) Survey on fuzzy-logic-based guidance
and control of marine surface vehicles and underwater vehicles. Int J Fuzzy Syst 20:572–586
12. Feng G, Lai C, Kar NC (2017) A closed-loop fuzzy-logic-based current controller for PMSM
torque ripple minimization using the magnitude of speed harmonic as the feedback control
signal. IEEE Trans Ind Electron 64:2642–2653
13. Chakib A (2019) Hybrid vehicle energy management using deep learning. In: 2019 International
Conference on Intelligent Systems and Advanced Computing Sciences (ISACS), pp 1–5. IEEE
14. Arcos-Aviles D, Pascual J, Marroyo L, Sanchis P, Guinjoan F (2018) Fuzzy logic-based energy
management system design for residential grid-connected microgrids. IEEE Trans. Smart Grid
9(2):530–543. https://doi.org/10.1109/TSG.2016.2555245
15. Shayeghi H, Younesi A, Hashemi Y (2015) Optimal design of a robust discrete parallel FP +
FI + FD controller for the automatic voltage regulator system. Int J Electr Power Energy Syst
67:66–75
16. Shayeghi H, Younesi A (2017) A robust discrete FuzzyP + FuzzyI + FuzzyD load frequency
controller for multi-source power system in restructuring environment. J Oper Autom Power
Eng 5:61–74
17. Mirjalili S, Gandomi AH, Mirjalili SZ, Saremi S, Faris H, Mirjalili SM (2017) Salp swarm
algorithm: A bio-inspired optimizer for engineering design problems. Adv Eng Softw 114:163–
191
18. Mahat P, Chen Z, Bak-Jensen B (2011) Control and operation of distributed generation in
distribution systems. Electr Power Syst Res 81:495–502
19. Bozchalui MC, Hashmi SA, Hassen H, Canizares CA, Bhattacharya K (2012) Optimal opera-
tion of residential energy hubs in smart grids. IEEE Trans. Smart Grid 3(4):1755–1766. https://
doi.org/10.1109/TSG.2012.2212032
20. Ren H et al (2016) Optimal operation of a grid-connected hybrid PV/fuel cell/battery energy
system for residential applications. Energy 113:702–712
1340 C. Alaoui and H. Saikouk

21. Pena-Bello A et al (2017) Optimizing PV and grid charging in combined applications to improve
the profitability of residential batteries. J Energy Storage 13:58–72
22. Matallanas E et al (2012) Neural network controller for active demand side management with
PV energy in the residential sector. Appl Energy 91(1):90–97
23. Glavic M, Fonteneau R, Ernst D (2017) Reinforcement learning for electric power system
decision and control: past considerations and perspectives. IFAC-Papers On Line 50(1):6918–
6927
24. Leo R, Milton RS, Sibi S (2014) Reinforcement learning for optimal energy management
of a solar microgrid. In: Global Humanitarian Technology Conference-South Asia Satellite
(GHTC-SAS), pp 183–188. IEEE
25. Kuznetsova E et al (2013) Reinforcement learning for microgrid energy management. Energy
59:133–146
26. François-Lavet V, et al (2016) Deep reinforcement learning solutions for energy microgrids
management. In: European Workshop on Reinforcement Learning (EWRL 2016) (2016)
27. Schaarschmidt M, Kuhnle A, Fricke K (2017) Tensor force: a tensor flow library for applied
reinforcement learning. Web page https://github.com/reinforceio/tensorforce
28. Website: electricity tariff in morocco https://www.invest.gov.ma/index.php?Id=34503&lan
g=en. Accessed 19 Jul 2020
29. Sassi HB, Errahimi F, Essbai N, Alaoui C (2018) A comparative study of ANN and
Kalman Filtering-based observer for SOC estimation. In: IOP Conference Series: Earth and
Environmental Science vol 161, p 012022
Economic Comparison Between Two
Hybrid Systems (Wind-Hydrogen)
and (Wind-Hydroelectric) for Electricity
Production in Socotra, Yemen

Saif Serag, Outhman Elbakkali, and Adil Echchelh

Abstract Renewable energy sources are one of the main sources of energy produc-
tion, Therefore, all researchers interesting in these sources, and consider it as a
primary source to cover a country’s needs, especially in developing countries such
as Yemen, which suffer from a large deficit to cover region’s need for electric
energy, including remote islands such as Socotra. In this paper, we will present
an economic study for electricity production by wind turbines in Socotra Island, and
an economic comparison between two means of energy storage, which is energy
storage by hydrogen production or by hydroelectric, and we accurately estimate the
energy unit price (kWh) to know the economic cost of two storage methods.

Keywords Wind energy · Hydrogen · Hydroelectric · KWh cost

1 Introduction

Renewable energy sources have grown significantly in recent times due to the prob-
lems that fossil fuels cause to environment such as air pollution, climate change, and
other problems, which have made all researchers and scientists a trend to seriously
study renewable energy and classify it as an energy that ensures a balance between
energy security and economic development [1]. Globally, The annual growth rate
of wind energy is about 30% annually, this is an indication that many countries
are interested in this source type to energy production and reducing dependence on
fossil energy [2]. In remote areas that are far from the power grid of main station, the
demand for energy sources increases, especially wind energy, and since wind energy
changes seasonally, this energy must be stored, one of the best storage methods is
the method that clean, renewable, and hybrid to energy produce at need time, peak
energy use, and use Part of the excess wind energy, especially in off-peak, to store
it as hydroelectric energy, which has been greatly distinguished recently due to its
high capacity and ease of use [3], or to produce hydrogen to be used when needed to

S. Serag (B) · O. Elbakkali · A. Echchelh

Laboratory of Energetic Engineering and Materials, Ibn Tofail University, Kenitra, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1341
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_122
1342 S. Serag et al.

Fig. 1 Socotra Island

electricity production by fuel cells, which has proven high efficiency to produce elec-
tricity better than other types of fuel [4], and it will noted in this study the economic
difference between these two methods according to the knowledge of energy unit
price for each of storage methods separately [2].
Yemeni population is considered to have the least access to electrical energy, as
the population rate benefiting from electricity is 51.7% of total population number
[5], due to Yemen circumstances which have passed from 2011 until today, total
capacity of energy production has decreased to 70%, which has negatively affected
of Yemeni population significantly [6]. All this makes it necessary to plans study
to energy production by renewable sources, specifically wind energy, as the average
wind speed for coastal areas in Yemen is 8 m/s [7], which is sufficient to energy
production with high efficiency. This study has been applied in Socotra island as a
field of study, which is one of the largest Yemeni islands, and contains a population
density of approximately 80,000 people, which increase their need for electricity
energy recently [8], and it will check the economic cost of producing and storing
the energy by renewable energy sources, and compare it in cost, which has great
importance, as it gives a future vision of possibility of renewable energy using,
storing it by a clean, environmentally friendly methods, and even at lowest costs, as
the economic comparison enables to know the best methods of electricity production,
least cost, and best in efficiency.

2 Site Data

Socotra is Yemen island, and considered within the Hadramawt governorate, located
250 km east of the Africa Horn and 350 km south of Arabian Peninsula [8]. It has an
area of 3,600 m3 and height above sea level is 257 m. All remote islands suffer from
Economic Comparison Between Two Hybrid Systems … 1343

a great shortage of generating electricity and depends on generating electricity there

by diesel or gasoline generators. Because of high prices of these exporters, electricity
prices are high in the region. On other hand, Socotra Island possesses wind energy
averaging between 6–12 m/s, which enables it to generate electricity at a lower cost,
more sustainable, and safer for environment [9] (Fig. 1).

3 Methodology

3.1 Wind Energy Cost

Wind energy cost is calculated by knowing several factors, turbine type used, its
capacity factor, tower height, annual amount of energy produced according to wind
speed, its distribution during year, and lifetime for energy production, then applying
the following equations [1, 10]:
      
1+i 1+i t 1+i t
P V C = Cinv + Comr × × 1− − Cs × (1)
r −i 1+r 1+r

Where Cinv investment cost, Comr Operation and maintenance costs, Cs Additional
costs, i interest rate = 1.5%, t Turbine lifetime = 20 years, r inflation rate = 2%.

PV C
(kWh cost) = . (2)
PE
Where PE = energy production in a rated lifetime.
This study will be applied by turbines (Nordex 1500 kW) with height towers
between 65–80 m and cost approximately 1220 $/kW [11].

3.2 Hydrogen Production Cost

To calculate the cost of electrical energy production by hydrogen producing and

storing, it must calculate the cost for each production stages, and hydrogen conver-
sion. Figure 2 shows the stages for electricity production from hydrogen, and it note
that the alternating current must be converted into continuous for use in water anal-
ysis to its basic components (hydrogen and oxygen), which is followed by hydrogen
storage process, to be used at peak electricity demand to cover the shortage for this
period or when wind speed is weak.

LCOE. (Levelized cost of electricity generation) which is equal to total cost divided
by total energy produced during the lifetime, for renewable energies calculated by
formula [12–14]:
1344 S. Serag et al.
n It +Mt +Ft
t=1 (1+r )t
LC O E = n Et
(3)
t=1 (1+r )t

Where It = investment expenditures in year t, Mt operations and maintenance

expenditures in t, Ft = fuel expenditures, Et electricity generation in the year t, r =
discount rate (wind = 6.5 to 10%), (Hydro = 3.75 to 10%) [15], and n = system
economic life = 20 years.
Now we should explain each of the stages shown in Fig. 2 to clarify the method
that it uses in this economic study:
Electrolysis. This will be studied with taking into account analyzing process of
hydrogen production by proton exchange membrane [PEM], and it is one of best
ways to analyze water into oxygen and hydrogen. It is noted that energy needed to
produce 1 kg of hydrogen consumes 50 kWh of electrical energy. Table 1 shows the
costs per kW in dollars with rate of operation, maintenance, and capacity factor [16]:

Storage. There are several methods for hydrogen store according to hydrogen state
(solid, liquid, gas) and each method has its advantages and disadvantages, but the
best economic way is to store hydrogen in a gaseous state under a pressure of 700 bar,
it is less expensive, simpler than other methods. We will note in Table 2 Storage cost
per kg of hydrogen, percentage of costs allocated to compression, operation, and
maintenance [17].

Electricity Production (Fuel Cell). When the amount of energy generated from
turbines decreases due to peak use or low winds in region, hydrogen stock must
be used to electricity production to cover the electricity need. Fuel cells are used to
convert hydrogen into electricity in oxygen presence and the reaction result is electric
current and water. Table 3 shows the price of 1 kilowatt of fuel cell, its efficiency,
percentage of operation and maintenance during its lifetime [19].

Transport Cost. In event that hydrogen needs to be transported from the generation
place to the consumption place, there are many ways to transport hydrogen, the best
way is tube trucks, and the transportation costs will appear in Table 4, [20].

Other Cost. (Fuel expenditures, water treatment, management, taxes…) = 5% kWh.

Table 1 PEM analysis cost

Description Amount Unit
Investment PEM 1000 $/kW
Capacity factor 50 %
O&M 3 %
Life time 20 Years
Economic Comparison Between Two Hybrid Systems … 1345

Fig. 2 Hydrogen production and storage stages for Using to electricity production

Table 2 Hydrogen storage

Description Amount Unit
cost (gas state)
1kg storage cost 500 $/kg
Compression cost 700 20 %
O&M 6 %
Life period 20 Years
Note: 1 kg of hydrogen produces approximately 39.4 kWh [18]

Table 3 Fuel cell cost

Description Amount Unit
1 kW fuel cell cost 1400 $/kW
Capacity factor 70 %
O&M 2 %
Lifetime 20 Years

Table 4 Truck transport cost

Description Amount Unit
Cost of 180 kg transport (tube truck) 100,000 $
Lifetime 20 Years

3.3 Hydroelectric Cost

Figure 3 shows the stages of electric power generation by hydroelectricity, where

surplus energy from wind turbines is used to store a quantity of water by pumping
it from the sea into an elevated reservoir, and then using stored water to generate
electricity as fallen and rushing water from reservoir to drive a turbine to generate
electricity at peak use. On the economic side, investment costs for hydroelectric
1346 S. Serag et al.

Fig. 3 Hydroelectric stages for electricity storage and production

Table 5 Hydroelectric cost

Description Amount Unit
Hydroelectric installed with storage costs 5000 $/kW
O&M 3 %
Capacity factor 70 %
Lifetime 20 Years

stations with storage range from 1050 $ to 7650 $ per kW, depending on the efficiency
and use period of the station. Table 5 shows the potential price of hydroelectric
stations with storage, along with cost of maintenance, operation, and lifetime [14].

4 Results and Discussion

4.1 Wind Data

Figure 4 shows the monthly average wind speeds for nine years from 2006 to 2014,
and clearly shows a large amount of wind speed, especially in June, July, August,
and September, while average of wind speed value was in January and December,
slightly lower values in remaining months, the lowest value for that speed is in April
4.37 m/s and the largest value in July is 17.22 m/s. As for the mean of wind speed for
9 years as a whole, it reached 9.17 m/s, a value that emphasizes a large and sufficient
amount to use for energy production.
When using a simulation program such as ALWIN, and by adding wind speed
data for area and type of turbine used (Nordex 80−1500 kW), the results are shown
in Fig. 5.
Economic Comparison Between Two Hybrid Systems … 1347

17.22 15.91
20.00

Wind Speed
13.99
11.04
8.37 7.91
10.00 6.79 5.63 6.64 5.65 6.54
4.37

0.00

Nov.

dec.
Apr.

Aug.
Jan.

May.

Jun.

Jul.

Oct.
Feb.

Mar.

Sep.
Month

Fig. 4 Monthly mean wind speed for 9 years

Fig. 5 AlWIN simulation application for (Nordex80-1500) turbine

It is noted that all variables related to wind speed, such as Weibull parameters (a
= 13.55 m/s, c = 1.42), average speed at a height of 80 m (vm = 12.34 m/s), as well
as the variables for turbine, which are average power (Pm = 635.6, 859.1 and 677.8)
kW, annual energy generated by the turbine (5567.7, 7525.8, and 5937.7) MWh, and
capacity factor (42, 56.9, and 45) %, were calculated in three cases: Measured values,
Rayleigh distribution, Weibull distribution.
By taking average annual energy for three values, measured, Weibull and Rayleigh
will be 6343 MWh, knowing the island’s population, and amount of energy consumed
by Yemeni citizen, we can know the number of turbines needed to cover the region’s
need for electricity, as shown in Table 6.
Thus, the number of turbines needed to cover island’s energy need is 3 turbines
with an annual excess of energy of 1429 MWh, equivalent to 3915 kWh per day, and
this excess energy contributes in storage process for use when needed.
1348 S. Serag et al.

Table 6 Number of turbines

Description Amount Unit
needed for Socotra Island
Turbine annual energy 6343 MWh
Energy per Capita 220* kWh
Population 80000 Person
Total consumption 17600 MWh
Turbine no 3 Turbines
* data.worldbank.org

Table 7 Wind energy cost in

Description Amount Unit
a lifetime
PVS 3.24 × 107 $
PE 3.8 × 108 kWh
kWh cost $ 0.085 $/kWh

Table 8 Costs for all

Input Cost Unit
hydrogen production stages
Electrolysis cost 0.021 $/kWh
Storage cost 0.018 $/kWh
Fuel cell cost 0.012 $/kWh
Other cost 0.003 $/kWh

4.2 Wind Energy Cost

By applying Eqs. (1), (2), calculating values of (Cinv , Comr., Cs ), and compensating
for parameters ( i = 1.5%, t = 20 years, r = 2%), it get the results as in Table 7,
where it was found that the total cost of energy producing by wind turbines is (3.24
× 107 )$ and energy produced from those turbines during their lifetime is (3.8 × 108 )
kWh, Thus, we can see the cost of kWh equal to (0.085) $.

4.3 Hydrogen Production Cost

By knowing cost and values , which appear in Tables 1, 2, 3, and 4, and applying the
equation of total cost to total energy production in a lifetime, we can get the results
shown as follows:
Table 8 shows the cost for each stage of hydrogen production in dollars per kWh,
and Table 9 shows total costs with or without transportation (without transportation
in the case of direct electricity production so that the hydrogen production plant and
energy generator plant are adjacent and do not need to transport the hydrogen from
one place to another), as it is shown that the largest costs are for the production
Economic Comparison Between Two Hybrid Systems … 1349

Table 9 Total cost in tow

Transport cost 0.014 $/kWh
methods
Total cost without transport 0.054 $/kWh
Total cost with transport 0.068 $/kWh

Table 10 Hydroelectric cost

Input Cost Unit
Investment costs 0.0408 $/kWh
O&M 0.0012 $/kWh
Other cost’s 0.0008 $/kWh
Total 0.0428 $/kWh

and storage of hydrogen because of the large energy used for production or storage
under pressure of not less than 700 bar. While the cost of transportation occupies a
relatively large percentage of 21% for same reason, (energy required for pressure,
storage, and cost for truck tanks and pipes, which are used for transportation).

4.4 Hydroelectric Cost

Table 10 shows the costs of generating electricity by hydroelectric stations, which

with total 0.0428 $/kWh, and we note that the costs of electricity producing in it are
less than cost of electricity producing by fuel cells because hydroelectric stations do
not need pressure and great security, as a case in storing hydrogen, Hydroelectric
just need a water pump, dam or tank, and a turbine to electricity generating, it is an
easy and safer method than others.

4.5 Total Cost and Hybrid Systems (Wind—Hydrogen)

and (Wind—Hydroelectric)

After studying the costs of electric energy generating from wind turbines, storage
methods, and generation of electricity by (hydrogen production and storage) or
by (hydroelectric storage), we ought to study the total costs of hybrid systems
that link (wind-hydrogen) or (wind-hydroelectric) energy, or even the use of three
systems Wind, hydrogen and hydroelectric. Table 11 shows the total costs for
(wind-hydrogen), (wind-hydroelectric), and (wind-hydrogen-hydroelectric).
It is noticeable the total costs shown in previous tables are low compared to
electricity prices resulting from oil derivatives, especially with high prices and even
lack of in recent times in Yemen. Therefore, electricity production in these ways is
lower in cost, safer, more sustainable, and it is recommended to be used.
1350 S. Serag et al.

Table 11 Total cost for

Input Cost Unit
hybrid system
Wind-Hydrogen 0.14 $/kWh
Wind-Hydrogen (transport) 0.15 $/kWh
Wind-Hydroelectric 0.13 $/kWh
Wind-Hydrogen-Hydroelectric 0.18 $/kWh

5 Conclusion

Renewable energy sources are the best and cheapest way to electricity production,
including wind energy, but because of the instability of these sources, energy must be
stored for use when needed, so this study was applied on the Socotra island, Yemen
for electric energy production by wind energy and store this energy in two methods,
first is produce, store hydrogen, and second is hydroelectric, these methods have
been economically compared and we get the following:
1. Socotra Island has a mean wind speed of 9.17 m/s. This result of the average
winds for nine years, which is a great speed that makes this island able to produce
electricity by turbines with high efficiency.
2. When using turbines Nordex 1500 kW with tower height equal 80 m, we find
an annual average energy produced from this turbine is 6,343 MWh, since the
annual consumption of the island is 17,600 MWh, Island needs three turbines
to cover regions by electricity with energy excess approximately 1429 MWh.
3. The unit cost of energy produced from wind turbines, according to the parame-
ters concluded from wind distribution and its probability is approximately 0.085
$/kWh.
4. Electricity producing process from hydrogen goes through several stages, the
most important of which is hydrogen production and storage, cost has been
calculated for all of those stages and it has been found that the total cost is equal
to 0.054 $/kWh to electricity production from hydrogen without transportation
and 0.068 $/kWh by transportation.
5. Hydroelectricity is the easiest and safest method for electricity production, the
costs of electricity producing by it are 0.043 $/kWh, which is less expensive
compared to other storage and production methods.
6. When applying a hybrid system (wind-hydrogen) and (wind-hydroelectric) or
(wind—hydrogen and hydroelectric), it is found that the costs are ((0.14 or 0.15),
0.13, 0.18) $/kWh respectively, which are low costs compared to the methods
of producing electricity with fossil fuels, especially in recent times when fuel
value has increased.
Economic Comparison Between Two Hybrid Systems … 1351

References

1. Ohunakin OS, Akinnawonu OO (2012) Assessment of wind energy potential and the economics
of wind power generation in Jos, Plateau State, Nigeria. Energy Sustain Dev 16(1):78–83
2. Shawon MJ, El Chaar L, Lamont LA (2013) Overview of wind energy and its cost in the Middle
East. Sustain Energy Technol Assess 2:1–11
3. Kumar K, Ansari MA (2013) Design and development of hybrid wind-hydro power generation
system. In: 2013 International Conference on Energy Efficient Technologies for Sustainability,
Nagercoil
4. Garland NL, Papageorgopoulos DC, Stanford JM (2012) Hydrogen and fuel cell technology:
progress, challenges, and future directions. Energy Procedia
5. Alkholidi AG (2013) renewable energy solution for electrical power sector in Yemen. Int J
Renew Energy Res (IJRER), 9
6. Saleh Qasem AQ (2018) Applications of renewable energy in Yemen. J Fundam Renew Energy
Appl 08(02) (2018)
7. Hashim Alkipsy EI, Raju V, Kumar H (2020) A review of the challenges of energy sector and
prospects of renewable energy utilization in Yemen. Glob J Manag Bus
8. Fritz M, Okal EA (2008) Socotra Island, Yemen: field survey of the 2004 Indian Ocean tsunami.
Nat Hazards 46(1):107–117
9. Saqqaf A, Alkaf, Muqbel MA, Bin SM (2010) World renewable energy congress XI cost
analysis of renewable energy for power supply for remote area in Yemen: a case study for
Socotra Island
10. Soulouknga MH, Oyedepo SO, Doka SY, Kofane TC (2020) Evaluation of the cost of producing
wind-generated electricity in Chad. Int J Energy Environ Eng 11(2):275–287
11. ‘Renewable Energy Technologies Cost Analysis Series Wind Power’ International Renewable
Energy Agency IRENA June 2012
12. Kornbluth K, Greenwood J, Jordan E, McCaffrey Z, Erickson PA (2012) Economic feasibility
of hydrogen enrichment for reducing NOx emissions from landfill gas power generation alter-
natives: a comparison of the levelized cost of electricity with present strategies. Energy Policy
41:333–339. https://doi.org/10.1016/j.enpol.2011.10.054
13. Shea RP, Ramgolam YK (2019) Applied levelized cost of electricity for energy technologies
in a small island developing state: a case study in Mauritius. Renew Energy 132:1415–1424.
https://doi.org/10.1016/j.renene.2018.09.021
14. Gielen D (2012) Irena renewable energy technologies cost analysis series hydropower
15. A Grant Thornton and Clean Energy Pipeline initiative (2018) Renewable energy discount rate
survey a grant thornton and clean energy pipeline initiative, p 36
16. Chan A, Hart D, Lehner F, Madden B, Standen E (2014) Fuel cells and hydrogen Joint
undertaking, “Development of Water Electrolysis in the European Union”, February 2014
17. Riis T, Hagen EF, Preben, Vie DJS, Ulleberg Ø (2005) International energy agency hydrogen
production and storage © OECDIE
18. Touili, S Alami Merrouni A, Azouzoute A, El Hassouani Y, Amrani A (2018) A technical and
economical assessment of hydrogen production potential from solar energy in Morocco. Int J
Hydrog Energy 43(51):22777–22796
19. Fuel Cell Technologies (2016) Office of energy efficiency and renewable energy - section 3.4
20. Amos WA (1999) Costs of Storing and Transporting Hydrogen NREL/TP-570–25106, ON:
DE00006574, 6574
A New Hybrid System Used to Connect
PV to a Microgrid

Alexandru Dusa, Petru Livinti, Daniel Ciprian Balanuta,

and Gelu Gurguiatu

Abstract This paper proposes a hybrid system by connection the photovoltaic panels
to a microgrid through an active power filter (APF). This new hybrid system enables
the limitation of the power generated by photovoltaic panels and follow the load
demand without supplying power to the microgrid. This whole unified hybrid system
will allow the attenuation of the harmonic injected to the microgrid and the compen-
sation of the reactive power in the case of lack of power from the PV system. The
Matlab/Simulink simulation environment is used to demonstrate the proper func-
tioning of this hybrid system. A modified perturb and observe algorithm is used to
control the boost converter and indirect method for the APF control.

Keywords PV · Microgrid · Boost converter · Shunt active power filter · Power

limitation

1 Introduction

The increase of clean energy demand involves the need of using renewable sources
like PV panels. The storage of energy through batteries for later uses leads to an
increased investment costs, maintenance costs and chemical pollution associated
with batteries, which makes this option less feasible. Injecting energy directly into
the microgrid ensures the maximum use of clean power generated by renewable
energy sources [1]. The integration of renewable energy systems is the most current
trend in the field of energy research. The increase in renewable energy production
and injection into electric its grids has begun to affect the security and stability of
the operation of power systems [2]. Obtaining the maximum power from renewable
sources without affecting the grid management and the attenuation of harmonics in

A. Dusa (B) · D. C. Balanuta · G. Gurguiatu

“Dunarea de Jos”, University of Galati, Galat, i, Romania
e-mail: [email protected]
P. Livinti
“Vasile Alecsandri” University of Bacau, Bacău, Romania

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1353
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_123
1354 A. Dusa et al.

the distribution grid is the main objectives of this paper. The maximum power point
(MPP) principle and supply of nonlinear load from PV is presented and demonstrated
in the papers [3–6].
In this paper it is proposed to interconnect the PVs to the low power grid through
the shunt active power filter (APF) and injecting only the amount of power needed
by the electrical loads. The APF consists of a three-phase inverter with three levels
in Voltage Source Inverter (VSI) with two capacitors on the DC side, with the aim
of attenuating the harmonics introduced in the grid from nonlinear tasks as well as
compensating the power factor. The control algorithm of the APF is performed using
the indirect control method. The connection of the PVs with APF is made through a
boost converter. The connection between APF and boost converter is made on the DC
side. Extraction of the maximum power from the PVs and its limitation is performed
using a modified Perturb and Observe (P&O) control algorithm.

2 The Hybrid System

Figure 1 presents in detail the proposed diagram for this new hybrid system that
allows supply the nonlinear load, PVs output power limitation at load level and
harmonics attenuation as well as correction of the power factor.

Fig. 1 Detailed electrical diagram of the hybrid system

A New Hybrid System Used to Connect PV to a Microgrid 1355

2.1 Boost Convertor

The boost converter allows the output voltage to be raised. The wiring diagram of a
boost converter is present in Fig. 2 For designing this type of converter it is necessary
to know the following values: minimum input voltage U in_min = 345.4 V, output
voltage U out = 750 V, intensity of the load current I out = 9.25 A, switching frequency
F sw = 25000 Hz, power P = 7300 W, effectiveness η = 95%, the percentage from
the output voltage for calculating the ripple voltage is equal to 1 [7].
The design elements of the boost converter are:
The filling factor D for the PWM control signal for IGBT in the boost converter
circuit is determined by:
  
Uin_min · η
D = 1− % (1)
Uout

The output current is calculated with the relation:

P
Iout = (2)
Uout

The value of the input current I in results from the formula: P = I in · U in = I out · U out

P
Iin = (3)
U1n

The value of the input current 1 is equal to the current rate in the coil. The ripple
current of the inductance L will be equal to 10% of the input current.

I L = Iin · 10% (4)

The inductance L is determined through the relation:

Uin · D
L= (5)
I L · Fsw

The value of the ripple voltage U out will be equal to 1% of the value of the output
voltage:

Fig. 2 Boost converter

1356 A. Dusa et al.

Uout = Uout · 1% (6)

The determination of the output capacity is performed with the relation:

Iout · D
C= (7)
Fsw · Uout

The load resistance is determined by the relation:

Uout
R= (8)
Iout

2.2 Maximum Power Point Tracking Algorithm

The modified P&O control method [8] allows the extraction of the maximum power
from the PVs and then compares it with the load need. If the power value of the PVs
is higher than the load need then the P&O control algorithm increases the duty cycle,
otherwise is decreases the duty cycle. The diagram of the modified P&O control
algorithm is present in Fig. 3.

Fig. 3 Diagram of the modified P&O control algorithm

A New Hybrid System Used to Connect PV to a Microgrid 1357

2.3 Active Power Filter Control

Indirect control is a strategy that does not require knowledge of electrical current
spectrum absorbed by the nonlinear load. This strategy imposes that the grid current
to be sinusoidal and in phase with the voltage [9].

i sa (t) = i La (t) + i f a (t) (9)

The electric current on phase a absorbed by the nonlinear load is:


i La (t) = i La
1
(t) + i Lak (t) + i Laq (t) (10)
k

1
where: i La (t) - fundamental active component of load current;

i Lak (t) - harmonics sum of load current;
k
i Laq (t) - reactive component of load current.
The electric current through the APF on phase a will be:

i f a (t) = i 1f a (t) + ĩ f a (t) (11)

where: i 1f a (t) - fundamental active component of APF current;

ĩ f a (t) - harmonics sum of APF current.
The electric current absorbed by the grid must be sinusoidal and must have the
same phase as the voltage. The component to be compensated by the APF is given
by:

ĩ(t) = ĩ f a (t) + i Lak (t) + i Laq (t) (12)
k

From (1) ÷ (4) it can be writing:

i sa = i La
1
+ i 1f a + ĩ (13)

The signal is generated for the charging input of the current regulator of phase a
of the power supply:
va
i a∗ (t) = ε DC √ = ε DC sin ωt (14)
2V

where: V is the effective value of the phase voltage of the grid and εDC the output of
the DC voltage regulator.
The above charge is compared with the measured value of the electric current
absorbed from the isa grid, resulting for the control of the APF on phase a:
1358 A. Dusa et al.

  
u ∗ca = k i a∗ − i sa = k i a∗ − ĩ − i La
1
− i 1f a (15)

where: k is the amplifier of the regulator.

The regulator is linear, the sinusoidal components of the load and the APF are
found, in the sinusoidal imposition of the regulator:

i a∗ = i La
1
+ i 1f a (16)

This remark results in the imposition for phase a of the APF:



u ∗ca = −k ĩ = −k ĩ f a (t) + i Lak (t) + i Laq (t) (17)
k

proportional to the polluting component. If it is correctly designed in steady state,

the controller cancels the steady state error



ĩ f a (t) = − i Lak (t) + i Laq (t) (18)
k

so, the APF will generate the polluting component itself necessary for the nonlinear
load.
The assessment εDC from the regulator for charging capacitor C is converted to
current reference, as follows:
⎧ ∗
⎪ i = ε · sin ωt
⎨ a   
i b∗ = ε · sin ωt − 2π 3 (19)
⎪
⎩ ∗   
i c = ε · sin ωt − 4π 3

Sinusoidal signal in phase with fundamental are obtained with Phase-Locked Loop
(PLL), which allows the synchronization of the compensation signal and the voltage
of the power supply system.
Figure 4 presents the control scheme of the APF with the indirect method. Where
we can see the PLL block for determining the grid phase (ωt) and the DC regulator
∗
for maintaining the voltage at the set value (Vdc ).

3 Simulation Results

The simulations were done in the Matlab/Simulink programming environment

considering two hypotheses:
A New Hybrid System Used to Connect PV to a Microgrid 1359

Fig. 4 The control schemes

Table 1. Values used to stimulate the hybrid system

Equipment Data Value Unit
Microgrid Line voltage (U L ) 400 V
Apparent power (S) 25000 VA
Nonlinear load Resistance (R) 21/96 
Inductance (L) 0.001 H
Boost converter Inductance (L) 0.0471 H
Capacitance (C) 2.5·10−5 F
Switching frequency IGBT (F sw ) 25000 Hz
Photovoltaic panel Peak Power (Pmax ) 365 W
(LG365Q1C-A5) Open-circuit Voltage (V oc ) 42.8 V
Voltage at Pmax (V mp ) 36.7 V
Short-circuit current (I sc ) 10.8 A
Current at Pmax (I mp ) 9.95 A
Series/Parallel 10/2 Panels
Irradiances (Irr) 1000 W/m2
Temperature (T ) 25 °C
Active power filter Inductance (L) 0.0038 H
Capacitance (C) 2·0.001 F
Switching frequency IGBT (F sw ) 30000 Hz
Continuous voltage of the APF (V dc ) 750 V
Proportional and integral (PI) The proportional coefficient (kp) 1
controllers The integral coefficient (ki) 20

• PVs output power was less than the nonlinear load need and the deficit was
supplemented from the microgrid.
• PVs output power was limited at load consumption level.
In Table 1 are presented the values used for the simulation of the proposed hybrid
system.
1360 A. Dusa et al.

Fig. 5 Boost converter results a) power, b) voltage, c) current

Figure 5 presents (a) power, (b) voltage and (c) current from the boost converter,
were:
• The first hypothesis is presented in the range 0.22–0.25 s:
– PVs produce the maximum power, PPV = 6700 W (due to losses on the boost
converter and APF);
– nonlinear load power PL = 13700 W;
– the power deficit was supplemented from the microgrid.
• The second hypothesis in the range 0.25–0.32 s:
– the PVs output power was limited to the nonlinear load need (PL = 3000 W).
In Figs. 6 and 7 is presented the voltage and current waveforms taking into account
the two hypotheses. Here is presented for the first hypothesis the waveform of the
microgrid voltage (Fig. 6(a)), the current generated by the PVs through the APF
(Fig. 7(a)) and injected in the common connection point (CCP) to supply the nonlinear
load (Fig. 7(b)) and attenuate the harmonics in the microgrid. And the current from
the microgrid (Fig. 6(b)) supplements the supply of the nonlinear load.
In the second hypothesis it is notice that the current from the microgrid is 0 (zero)
(Fig. 6(b)). This is due to the limitation of the power generated by the PVs, injecting
the power needed by the nonlinear load.
A New Hybrid System Used to Connect PV to a Microgrid 1361

Fig. 6 Current and voltage results a) microgrid voltage, b) microgrid current

Fig. 7 Current results a) PV + APF current, b) nonlinear load current

Also, here we can observe the transient regime at the moment t = 0.25 s, at the
change of the power of the nonlinear load.
Figure 8 present the Fast Fourier Transform (FFT) analysis for the source voltage
waveforms on phase 1. Figure 8(a) presents the harmonic level of the source voltage
for the first hypothesis, THDu = 4.01%. Figure 8(b) presents the harmonic level of
the source voltage for hypothesis two, THDu = 4.83%.
Figure 9 present the FFT analysis for current waveforms for phase 1. Figure 9(a)
presents the harmonic level of the nonlinear load current, THDi = 29.24%.
Figure 9(b) presents the harmonic level of the microgrid current, THDi = 4.88%.
1362 A. Dusa et al.

Fig. 8 FFT analysis of the voltage waveform a) in the first hypothesis, b) in hypothesis two

Fig. 9 FFT analysis of the current waveform a) the nonlinear load, b) the microgrid current

4 Conclusions

The aim of this paper was to create a new hybrid system were to connect the PVs to the
microgrid using a shunt APF. Which will allow the maximum power to be extracted
from the PVs or to limit the output PVs power, without introducing the excess power
into the microgrid. At the same time, this hybrid system allows the attenuation of the
harmonic level in the microgrid and the compensation of the reactive power.
For this paper, simulations were performed in Matlab/Simulink, which demon-
strate the advantages of this proposed system.
A New Hybrid System Used to Connect PV to a Microgrid 1363

References

1. Ali Q, Al-Shetwi MA, Hannan KPJ, Mansur M, Mahlia TMI (2020) Grid-connected renewable
energy sources: review of the recent integration requirements and control methods. J Clean Prod
253(119831):1–17
2. Sriranjani R, Jayalalitha S (2018) PV interconnected shunt active filter for power balancing and
harmonic mitigation. Int J Pure Appl Math 118(18):2341–2354
3. Djeghader Y, Chelli Z, Rehaimia S (2018) Harmonic mitigation in electrical grid connected
renewable energy sources. ACTA Electrotehnica 59(4):287–291
4. Manimegalai V, Senthilnathan N, Sabarimuthu M (2017) Grid interfaced renewable energy
source to improve power quality using SAPF. Int Res J Eng Technol 4(4):834–839
5. Chaithanakulwat A (2019) Multi-functionality control, power filtering single-phase grid-
connected photovoltaic system. Am J Electr Power Energy Syst 8(2):62–70
6. Bouzelata Y, Kurt E, Chenni R, Altın N (2015) Design and simulation of a unified power quality
conditioner fed by solar energy. Int J Hydrogen Energy 40(44):15267–15277
7. Livinti P, Ghandour M (2020) Fuzzy logic system for controlling of DC/DC boost converter
developed in MATLAB SIMULINK. Int Res J Eng Technol 7(9):1629–1636
8. Zala MP, Pandya MH, Odedra KN, Patel DP (2017) Active power control of PV system in MPPT
and CPG mode. Kalpa Publ Eng 1:270–277
9. Gurguiatu G Munteanu T (2012) Indirect control in active power filters. In: International
Conference on Optimization of Electrical and Electronic Equipment (OPTIM) 2012 - 13th,
ResearchGate
Coordinated Control and Optimization
Dispatch of a Hybrid Microgrid in Grid
Connected Mode

Marouane Lagouir, Abdelmajid Badri, and Yassine Sayouti

Abstract This paper presents a novel daily energy management system for coordi-
nated control and optimization dispatch of a grid connected hybrid microgrid (HMG).
The main objective is to minimize both operating cost and emissions level, taking
into account ensuring continued power balance, maximize the generation of renew-
able energy sources, optimal charging/discharging of battery storage system, and the
system constraints. The economic and emission dispatch problem has been solved
using grey wolf optimizer (GWO) algorithm, under four case studies. Then a compar-
ison with other heuristic algorithms is considered. The obtained results confirm the
potential of GWO in terms of stability, robustness, and convergence speed. Therefore,
it can be used for real time control and optimization dispatch of HMG.

Keywords Energy management system · Optimization dispatch · Hybrid

microgrid · Operating cost · Emissions level · Renewable energy sources · Grey
wolf optimizer

1 Introduction

In the last years, with the growing electricity demand and the development of
distributed generation (DG) technology. The integration of microgrid (MG) into
the existing electricity power system is considered as a promising alternative that
provides more reliable operation, considerable cost saving and reduction in pollutant
gas emissions due to excessive energy dependency on fossil fuel [1]. However, the
intermittent generation of renewable energy sources (RESs) and difficulty of predic-
tion of load demand caused a significant mismatch between the generated power and
the required demand [2]. Therefore, an energy management system (EMS) plays a
vital role in optimizing dispatch problem and coordinated control in MG system.
Several studies have been conducted in area of MG energy management and
control.

M. Lagouir (B) · A. Badri · Y. Sayouti

Department of Electrical Engineering, EEA&TI Laboratory, Faculty of Science and Technology
(FSTM), Hassan II University of Casablanca, BP 146, 20650 Mohammedia, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1365
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_124
1366 M. Lagouir et al.

In [3], the authors proposed a novel optimization model to minimize both opera-
tional and pollution cost in a MG, then an advanced dynamic programming (ADP)
method is employed to solve the optimization dispatch problem. To evaluate the
system performance two case studies were considered in the presence and without
a battery storage system. In [4], the authors discuss the application of ant lion opti-
mizer (ALO) algorithm for handling the optimal economic load dispatch problem.
Considering challenging constraints such as spinning reserve constraint, ramp rate
limits, prohibited operating zones of generating units. Three case studies have been
investigated, then the results are compared with other heuristic methods to prove the
stability and robustness of ALO method. Many heuristic methods are implemented
for optimal energy management in MG, including particle swarm optimization (PSO)
algorithm [5], antlion optimizer algorithm [4, 6], artificial immune system (AIS) [6],
grey wolf optimizer (GWO) algorithm [7], and others.
The main contribution of this work is the development of a novel energy manage-
ment system, with the aim to minimize both operating cost function and emissions
level of pollutant. While taking into account system power balance, high level pene-
tration of RESs, optimal charging/discharging of battery storage system (BSS), moni-
toring state of charge (SoC) of the battery, and the system’s components power supply
constraint. The grey wolf optimizer heuristic algorithm is used in order to solve the
optimization dispatch problem. And a comparison from applying GWO algorithm
and other known heuristic methods, including PSO, novel bat algorithm (NBA) [8],
whale optimization algorithm (WOA) [9], has been performed, to show the validity
and effectiveness of the proposed strategy.
The rest of the paper is organized as follows: The system description is given in
Sect. 2. Section 3 outlines the proposed energy management approach. The results
and discussion are presented in Sect. 4. Finally, Sect. 5 concludes the work.

2 Hybrid Microgrid System

The proposed hybrid microgrid (HMG) system is running under grid connected
mode. Figure 1 shows the architecture of the studied HMG. The actual load demand
is primarily supplied by renewable energy sources represented by photovoltaic
panels (PV) and wind turbine (WT) unit. Three conventional energy sources (CESs)
composed of diesel generator, fuel cell unit and microturbine unit (DG, FC and MT)
are settled and used as backup to supply the load whenever the combination output
power from RESs and the energy storage system (ESS) represented by battery is not
sufficient to supply the required demand. Moreover, the HMG can exchange power
with the main grid through purchasing or selling power. Finally, all the system’s
components use various conversion devices for connection to the common bus AC.
Thus they are connected to a centralized control system dedicated to dealing with
the coordinated control and optimization dispatch problem in HMG.
Coordinated Control and Optimization Dispatch … 1367

Fig. 1. Studied hybrid microgrid architecture

3 Problem Statement

In this work, two contradictory objective functions are considered for minimizing:
The first objective function is the operating cost and the second objective function
is the pollutant gas emissions of Nitrogen oxide, Sulfur dioxide and Carbon dioxide
(NOx , SO2 and CO2 ), respectively. The problem can be mathematically formulated
by [10]:

min O FH M G = ω · C1 + (1 − ω) · C2 (1)

Where ω is weight coefficient, and it is always a positive value. Finally, C1 and C2

represent the operating cost and emissions objective function, respectively.

3.1 Operating Cost Function

The first objective function is the operating cost of generating units, it combines four
terms representing the fuel cost, operation and maintenance cost, start-up cost and
the exchanged power cost with the main grid, expressed as [11, 12]:
1368 M. Lagouir et al.


N
C1 = {C Fi (Pi ) + O Mi (Pi ) + ST Ci + (C P E i − RS E i )} (2)
i=1

The fuel or gas cost C Fi consumed by DG or FC, MT unit is expressed as a quadratic

function of the produced power Pi [12, 13]:

C Fi (Pi (t)) = αi + βi Pi (t) + γi Pi2 (t) (3)

Where αi , βi, γi are cost coefficients of each dispatchable source i.

The second term represents the operation and maintenance cost; it can be
represented mathematically as [11, 13]:


N
O Mi = K O Mi Pi (4)
i=1

The start-up cost of dispatchable generator i is expressed in term of the unit has been
off [11, 13]:
  
To f f,i
ST Ci = σi + δi 1 − exp − (5)
τi

According to the RESs supply limitation and the actual required demand, the HMG
may import or export power from or to the main grid. The exchanged power cost can
be represented mathematically as [11]:

C P E i = C p × max(PD − Pi , 0)
RS E i = Cs × max(Pi − PD , 0) (6)

Where C p and Cs represent the price of power purchased or sold to the main grid in
[$/Kwh].

3.2 Pollutant Gas Emissions Cost

The pollutant gas emissions cost of (NOx , SO2 and CO2 ) is expressed as a linear
function of the generated power Pi from dispatchable source i and the power provided
by main grid PGrid [12, 13]:


N 
M 
M
C2 = χk (E Fik Pi ) + χk (E FGridk PGrid ) (7)
i=1 k=1 k=1
Coordinated Control and Optimization Dispatch … 1369

Where χk is externality costs of emission type k. E Fik and E FGridk represent the
emission factors of generator i and main grid, respectively.

3.3 System Constraints

In the operation of the HMG system, the power balance should be satisfied with each
time step. It is expressed as follows:

PD (t) = PGrid (t) + PC E Ss (t) + PE SS (t) + PR E Ss (t) (8)

The limit of output power of each distributed generator i is given by:

0 ≤ PP V ≤ PP VM ppt
0 ≤ PW T ≤ PW TM ppt
PE SSmin ≤ PE SS ≤ PE SSmax
Pi Min ≤ Pi ≤ Pi Max
PGrid Min ≤ PGrid ≤ PGrid Max (9)

4 Energy Management System

This work aims to design and implement a two-layer energy management system,
for daily coordinated control and optimal dispatch of HMG system. The upper
layer combines a fuzzy logic control system (FLS), used to define the optimal
charging/discharging power of BSS, the decision to turn on or off the dispatch-
able sources and the interactive mode between the HMG and the main grid. On
the other hand, the lower layer consisting of grey wolf optimizer algorithm dedi-
cated to solve the economic and emission dispatch problem in case of using the
CESs. Figure 2 shows the structure of the proposed energy management system. The
following subsection briefly describes the mathematical models of GWO algorithm.

4.1 Overview of Grey Wolf Optimizer Algorithm

The grey wolf optimizer algorithm was proposed by Mirjalili in 2014 [14]. The basic
idea of this algorithm was inspired by hunting technique and social hierarchy of grey
wolves. To simulate the leadership hierarchy, the grey wolves living in the same pack
are divided into alpha, beta, delta and omega. The alpha is responsible for decision-
making about hunting, sleeping place and so on. In the second level of hierarchy, we
1370 M. Lagouir et al.

Fig. 2. Proposed energy management system

found beta wolves with the mission of helping the alpha in decision-making or any
other activities related to the pack. The third level in the grey wolves’ hierarchy is
delta, in this level the wolves are responsible for watching boundaries, guarantee the
safety and taking care for the weakest of the pack. And in the last level, there are
the omega wolves that assumed to follow the wolves of the highest levels in GWO
algorithm.
Mathematically, the position of wolf is modeled as follows [7, 14]:
→ −
− → − → → 
−
D =  C · X P (t) − X (t) (10)

−
→ −
→ −
→ − →
X (t + 1) = X P (t) − A · D (11)

−
→ −
→
Where X and X p represent the position vector of a grey wolf and the prey,
respectively. t is the current iteration.
−
→ −
→
Finally, A and C indicate coefficient vectors, and they are calculated using the
following formulas:
−
→
A = 2−
→ r1 − −
a ·−
→ →
a (12)

−
→
C =2·−
→
r2 (13)
Coordinated Control and Optimization Dispatch … 1371

Where, the components of a decreased linearly from 2 to 0. During iteration process,

and −
→r1 , −
→
r2 are random vector in [0, 1].
In the GWO algorithm, the α wolf is considered as fittest solution, following by
β and δ as the second and the third best solutions, respectively. Finally, the rest of
wolves, including the
are supposed to move according to the position of wolves
of higher levels. This is mathematically modelled as follows:
     
 α = C1 · X α (t) − X (t), D
D  β = C2 · X β (t) − X (t), D
 δ = C 3 · X δ (t) − X (t)
(14)

X 1 = X α − A1 · D
 α , X 2 = X β − A2 · D
 β , X 3 = X δ − A3 · D
δ (15)

X 1 + X 2 + X 3
X (t + 1) = (16)
3

Where X α (t), X β (t) and X δ (t) represent the position of α, β, and δ, respectively.

5 Simulation and Analysis Results

First, it is assumed that PV and WT units generate power without any emissions,
and with a generation limits equal to 25 kw and 20 kw, respectively. The maximum
and minimum output powers of BSS are set to 30 kw and −30 kw, respectively. And
the limit on power exchanged with the main grid is considered as 30 kw. Finally, the
maximum capacity of each dispatchable generator is set to 30 kw. Figure 3 shows
the actual output power from both RESs and optimal output power of the BSS based
on a fuzzy logic control system. As it is observed, the combination output power
of RESs and BSS is not enough to serve the load demand, therefore the fuzzy logic

Fig. 3. Output power of RESs and the battery storage system

1372 M. Lagouir et al.

Fig. 4. Optimal output power of conventional generators and main grid

Fig. 5. Contribution of system’s components in power management

controller incorporates the dispatchable sources (DG, FC, and MT) in the energy
management, while permits importing power from the main grid. Figure 4 reveals
optimal output power of CESs and main grid. And Fig. 5 shows the contribution
rates of each power generation source. As we can see, the best choice to minimize
both operating cost and emissions level is to make use of generated power of FC and
MT unit, respectively. Hence, the DG is dispatched only when the demand becomes
very high. Finally, the total system balance for a daily simulation time is shown in
Fig. 6. It is to be noted that the previously described figures are plotted with a weight
coefficient ω = 0.16, and a time step of 15 min.
To evaluate the impact of the selected weight coefficient on the formulated objec-
tive function, the operating cost and emissions level are calculated for different values
of ω, as is shown in Fig. 7. It is clearly seen that as long as ω is increased from 0
to 1, more the priority is given to minimizing the operating cost, therefore the emis-
sions level becoming higher (Fig. 8). Similarly, when the value of ω is decreased,
the emissions level are first considered to minimized. According, to the results better
compromise between the two contradictory objective functions is achieved when the
value of ω = 0.16. Finally, Table 1 summarizes the obtained results considering four
Coordinated Control and Optimization Dispatch … 1373

Fig. 6. Total system power balance

Fig. 7. Evolution of operating cost as a function of emissions level

Fig. 8. Evolution of the pollutant gas emission as a function of ω

case studies when ω = 1; ω = 0.5; ω = 0.16; and ω = 0.01, respectively, then a

comparison of GWO algorithm with other heuristic methods has been evaluated.
From the results, it is clear that the proposed GWO algorithm is able to achieve
better solutions. In addition, the results show the advantage of GWO in terms of
stability, and good convergence to the optimum solution in fewer iterations. Which
confirms its potential to deal better with the optimization dispatch problem.
1374 M. Lagouir et al.

Table 1. Comparison of obtained simulation results under 4 different case studies

Daily Daily Daily Daily Daily Daily Daily
operating power power power exported emissions emissions
cost ($/h) generated generated generated power cost ($/h) level
by DG by FC by MT from main (kg/h)
(kwh) (kwh) (kwh) grid (kwh)
Case GWO 1864.6752 464.1969 2053.0748 886.1505 1975.1393 79.9978 15.3247
1 PSO 1864.9310 461.4069 2065.4071 933.3225 1918.4876 71.8993 15.1015
NBA 1868.2241 444.5014 2102.1870 968.4948 1863.3782 69.6505 14.7345
WOA 1868.6513 463.8059 2148.5225 918.3371 1847.8960 70.8982 14.8524
Case GWO 1877.5217 199.2115 2250.0304 1296.4861 1632.8335 46.2096 11.4882
2 PSO 1879.2765 203.4923 2197.7221 1229.9017 1747.8690 48.3720 11.9399
NBA 1885.3066 199.5290 1991.6969 1419.1773 1768.1582 48.6768 12.0558
WOA 1890.4623 198.9263 2141.8636 1347.8002 1689.9714 47.2162 11.7234
Case GWO 1919.3090 199.2432 2554.6371 1515.0981 1109.5831 37.8433 9.5989
3 PSO 1919.7558 199.1626 2522.3023 1516.2146 1141.5502 38.3704 9.7200
NBA 1922.2369 199.0307 2311.2919 1523.0586 1345.1803 41.7383 10.4924
WOA 1939.7665 199.0106 1976.3342 1776.4579 1426.7588 43.5054 10.9081
Case GWO 2317.5969 199.2226 2350.9518 1827.6136 1000.7734 36.5137 9.3184
4 PSO 2390.9327 198.9273 1961.9595 1842.5563 1375.1183 42.7414 10.7396
NBA 2385.0831 199.0805 2032.3525 1800.0990 1347.0365 42.2211 10.6168
WOA 2385.0619 199.0102 2000.2704 1832.5740 1346.7069 42.2600 10.6286

For all case studies, the population size is set to N = 30, and the maximum
iteration I termax is considered as 100.

6 Conclusion

In this paper, a novel approach based on combining the fuzzy logic control system
with a GWO heuristic method, is introduced to deal with the control and optimization
dispatch problem in hybrid microgrid system. Four case studies are considered, then
the system’s performance is compared with other algorithms, including PSO, NBA,
and WOA. Simulation results prove that GWO is most effective in determining
the global optimal solution in less time, which confirms the validity of the proposed
approach to efficiently manage the generated power of each component of the studied
HMG system, considering technical, economic and environmental criteria.
Coordinated Control and Optimization Dispatch … 1375

References

1. Murty Vallem VVSN, Kumar A (2020) Optimal energy dispatch in microgrids with renewable
energy sources and demand response. Int Trans Electr Energy Syst 30(5):1–27
2. Ren Sh, Wang J, Ma M (2019) Multi-objective optimal control of microgrid based on economic
model predictive control. In: Chinese Control Conference (CCC), pp 1–6. IEEE, Guangzhou
3. Moradi H, Esfahanian M, Abtahi A, Zilouchian A (2018) Optimization and energy management
of a standalone hybrid microgrid in the presence of battery storage system. Energy 147:226–238
4. Van TP, Snasel V, Nguyen TT (2020) Antlion optimization algorithm for optimal non-smooth
economic load dispatch. Int J Electr Comput Eng 10(2):1187–1199
5. Radosavljevic J, Jevtic M, Klimenta D (2015) Energy and operation management of a microgrid
using particle swarm optimization. Eng Opt 48(5):811–830
6. Hatata AY, Hafez AA (2019) Ant lion optimizer versus particle swarm and artificial immune
system for economical and eco-friendly power system operation. Int Trans Electr Energy Syst
29(4)
7. Cherkaoui N, Belfqih A, El Mariami F, Boukherouaa J, Berdai A (2020) Active power output
optimization for wind farms and thermal units by minimizing the operating cost and emissions.
Int J Electr Comput Eng 10(4):3412–3422
8. Meng XB, Gao XZ, Liu Y, Zhang H (2015) A novel bat algorithm with habitat selection and
doppler effect in echoes for optimization. Expert Syst Appl 42(17–18):6350–6364
9. Mirjalili S, Lewisa A (2016) The Whale optimization algorithm. Adv Eng Softw 95:51–67
10. Nwulu NI, Xia X (2015) Multi-objective dynamic economic emission dispatch of electric power
generation integrated with game theory based demand response programs. Energy Convers
Manag 89:947–963
11. Mohamed FA, Koivo HN (2012) Online management genetic algorithms of microgrid for
residential application. Energy Convers Manag 64:562–568
12. Nemati M, Braun M, Tenbohlen S (2018) Optimization of unit commitment and economic
dispatch in microgrids based on genetic algorithm and mixed integer linear programming.
Appl Energy 210:944–963
13. Alvarado-Barrios L, Rodríguez del Nozal A, Tapia A, Martínez-Ramos JL, Reina DG (2019)
An Evolutionary computational approach for the problem of unit commitment and economic
dispatch in microgrids under several operation modes. Energies 12
14. Mirjalili S, Mirjalili SM, Lewis A (2014) Grey Wolf Optimizer. Adv Eng Softw 69:46–61
Smart Energy Management System:
SCIM Diagnosis and Failure
Classification and Prediction Using
Energy Consumption Data

Oussama Laayati, Mostafa Bouzi, and Ahmed Chebak

Abstract This paper represents the impact of smart energy management system
in industrial microgrid on improving the key performance indicators of the mainte-
nance, the case study will be on the squirrel cage induction motors, the idea is to
finding correlations between energy consumption data and mechanical failures of
these machines, this work shows a first version of a test bench acquiring vibration,
voltage, current and speed while the eccentricity fault is present, after test bench
preparation, installation, configuration, gathering the data, monitoring it, and see the
harmonics of current and vibration in frequential domain and see the first results, the
idea is to store the data and trying to implement machine learning algorithms, which
gives an accuracy between 0.71 and 0.96 between model features which are current
voltage and vibration, this work is a primary work of a big idea that is already devel-
oped and mention by a lot of scientists and researchers, that can be implemented in
a real case of study in mining industry.

Keywords SCIM failure · Energy management · Machine learning

O. Laayati (B) · M. Bouzi

Computer Science, Mechanical, Electronics and Telecommunication Laboratory, LMIET, FSTS,
University Hassan 1, Settat, Morocco
e-mail: [email protected]
O. Laayati
Innovation Lab for Operations, University Mohamed VI Polytechnic, Benguerir, Morocco
A. Chebak
Green Tech Institute, University Mohamed VI Polytechnic, Benguerir, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1377
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_125
1378 O. Laayati et al.

1 Introduction

1.1 Problematic Description

In smart micro grids energy consumption and production management is a necessary

element that can improve the health and the performance of the grid and its compo-
nents, sources, and loads [1]. Squirrel Cage Induction Machines (SCIM), invented
by Nicolas Tesla with 3 phase and Galileo Ferraris with 2 phase, is one of the most
popular load in industry in the whole world represents 90% of industrial motors in the
world, and 50% of global electrical energy consumption is due to induction motors,
Researchers proposed efficiency monitoring of SCIM motors for maintenance
decision making and decrease the CO2 emission [1] others proposed analytical
solutions to optimize the energy consumption for induction motors [2].
Mainly when an electrical load is in a faulty condition, this abnormal behavior
increases the energy consumption, the purpose of this work is to know if we can
detect the abnormal activities of SCIM motors using only it is energy consumption
data [3]. There were a lot of methodologies that can identify and classify mechanical
faults in the SCIM, and are based on vibration, bearing temperature, moisture [4].
The mining industry is one of the most industries that uses SCIM motors in
conveyors, reclaimers, Stackers, draglines and jaw crushers because of its efficiency
and high performance, the approach is detecting faulty loads in an industrial micro
grid using only specific energy consumption of these loads.

1.2 Proposed Solution

This work represents new methods for classifying SCIM faulty behaviors, by only
using current and voltage data, that means specific energy consumption data which is
acquired by a smart grid monitoring system, the goal is using new machine learning
techniques and algorithms that help finding correlations between vibration data, and
classifies failures [5], such as an unsupervised and general algorithm that can classify
an anomaly while a voltage unbalance is detected, and supervising the oscillations
of the current third harmonic.

2 Methodology

2.1 Introduction

The idea is to prepare a test bench to test the data acquisition and monitoring system,
installing and testing the current, voltage, speed and vibration sensors on the test
Smart Energy Management System … 1379

bench, monitoring the results using a data acquisition software, monitoring these
data and store it then apply classification and prediction algorithms to know if there
is some correlations between data.
The next step is to implement an industrial solution in the mining industry to test
the efficiency and measure the performance of the smart prediction and classification
algorithms and control its impact on the maintenance and production.

2.2 Failure Types

The main defects and failures of a SCIM showed in the state of art [6, 7] are:
The rotor eccentricity which is the most common problem in the mining industry
in conveyors and draglines, this fault is due to an offset between the axis of symmetry
and the rotation axis, we can see a lot of eccentric bearing and eccentric armature
motor.
The breakage of rotor bars and end-bearings this type of failures is classified
as critical failures in SCIM motors and can generates damage and more energy
consumption due to friction and mal functioning of the bearing lubrication.
The rotor bows this fault is due to the heating or cooling of the rotor, it causes a
local thermal asymmetric distribution. And other failures after the damage of the rotor
that can increase the energy consumption and impact voltage and current behavior.

2.3 Failures Analysis Techniques Using Energy

Consumption Data

There are many failures analysis techniques that helps to detect or to predict defects
of SCIM motors, we mention here some popular one such as acoustic emission, Air-
gap torque, stator current, electromagnetic field monitoring, induced voltage, instan-
taneous angular speed, instantaneous power, motor circuit analysis, surge testing,
vibration, and voltage.
In this work we will focus on electrical energy consumption data, which is current,
voltage, and time, and correlate these data with faulty SCIM motors vibration data
regarding torque and speed.
The voltage unbalance and electrical signature is a fault can be detected by current
signal [8] it can be applicable in framework of 3 different signal domain, time, domain
knowledge and hybrid time frequency data [9], the spectral analysis of the current of
SCIM motor using the spectrogram background theory short time fourier transform,
the auto winger distribution and the wavelet decomposition [10], for the detection
of air gap eccentricity the most popular method is the stator current monitoring, by
showing shape deviations in park’s vector which is a circular pattern normally [11],
the comparing method by combining harmonics of electrical components, power
1380 O. Laayati et al.

supply frequency and harmonics of cage frequency which is the rotation of the rotor,
getting spectral amplitude and comparing the slopes show very efficient results to
classify and detect failures [12].

2.4 Machine Learning Algorithms

Introducing machine learning and artificial intelligence to the classification of defects

and finding some not explored correlations between energy consumption data and
mechanical data, many researchers start implementing algorithms to classify defects
by focusing only on an expert system and tried to make new defects predictions
algorithms by using convolutional neural network (CNN), linear regression, support
vector machine, k-means clustering for unsupervised SCIM failures classification.
The idea is comparing healthy and faulty condition of the SCIM motor, gathering
the energy consumption data, vibration, and rotation per minutes, in variable rota-
tion speed and variable torque, these gathered data will feed the classification and
prediction algorithm.
This scheme Fig. 1 describes the data flow from vibration, current, voltage torque
and speed sensors, from the SCIM under test, the database contain all different tests
data, these data will be prepared to feed the clustering algorithm, the maintenance
expert must analyze the result of the algorithm and give a score of the anomaly, the
machine learning algorithm using reinforcement learning and visualize the result.
The adaptative gradient optimizer (ADG) based on deep (dCNN) algorithms for
faulty bearing with ball defect, with outer and inner race defect and rotor with one
broken bar [3]. This algorithm shows some good results that can be enhanced and
implemented in a mining industry machine for example reclaimer, dragline, or jaw
crushers.

Fig. 1 General scheme of test data mining, preparation, clustering, and visualization
Smart Energy Management System … 1381

3 Case Study

3.1 Introduction

This case study will discuss the feasibility of this idea and the general acquisition
system architecture, with monitored data and results, we will propose the data mining
steps data acquisition, introducing machine learning classification and prediction
models and the key performance indicators that will be monitored.

3.2 General Schematic

In this figure we are proposing a scheme to gather data using National Instruments
Modules of vibration, current voltage and rotation speed, the gathered data using
LABVIEW and stored into a PostgreSQL database, introducing python code to
extract data from POSTGRESQL database to check correlations between data. This
scheme is based on the intelligent power quality monitoring test bench using Internet
of Things monitoring [11].

Fig. 2 Simulation hardware and software data flow

1382 O. Laayati et al.

Fig. 3 Simulation test vibration, STFT and RMS values

3.3 Simulation Steps

This test bench has options to create and to simulate defects and failure, in this paper,
we will focus on the eccentricity fault detection, the acquired data, of voltage of the
stator in each line [13], the current, the root mean square value of each current line
and its Short Time Fourier Transform STFT [6], the vibration and its FFT in 4 points
axial and vertical, and the rotation per minutes.
The data is being recorded into the database, using database connection toolbox
for historical data with timestamp, extracted by python code, preprocessed, and
recalculated using predefined machine learning algorithm to find and explore visual
correlations to choose effectively good clustering, classification, and prediction
algorithms.
The Fig. 3 graph shows the RMS and the vibration temporal values with STFT,
the current, signal, RMS and the STFT in the Fig. 4, we are using these data to see if
there is any correlations and implementing a multilinear regression and see the score
of each, the Fig. 5 shows the voltage signal to detect voltage unbalance.

3.4 Machine Learning Results

As mentioned, to choose a good machine learning model for classification and

prediction, first all the acquired signals are being plotted to show the correlations.
The plots in Figs. 6 shows the correlation between STFT signals in a healthy
condition, while the Fig. 7 shows the faulty condition with the eccentricity fault
[14], here it is identified that there are some strong correlations between gathered
data of STFT vibration and current, for the machine learning algorithm we can use in
first step the multi Linear Regression algorithm with a gradient descent cost function
the results shown in Fig. 8 that represent the accuracy of this machine learning
multilinear regression between each feature measured data.
Smart Energy Management System … 1383

Fig. 4 Current data and STFT with RMS values

Fig. 5 Simulation Voltage

values.

4 Conclusion

This work identifies correlations between STFT Current and vibration data, in the
test bench simulation and data gathering of a faulty and healthy conditions, there are
some differences in the data plots shown in the test, the voltage unbalance is not well
shown in this test bench simulation, because of the condition of the SCIM motor
which is in a good health, the machine learning basic linear regression algorithms
shows good results between STFT vibration and current of each current line, which
1384 O. Laayati et al.

Fig. 6 Correlations plots between STFT vibration and STFT current healthy condition.

Fig. 7 Correlations plots between STFT vibration and STFT Current faulty condition.

is a good indicator to continue working on this idea and implementing new machine
learning algorithms, to classify and to predict other failures and SCIM electrical and
mechanical defects and finding more correlations between all faults.
Smart Energy Management System … 1385

Fig. 8 Accuracy of linear regression model between STFT Signal of current and vibration.

This paper shows the importance of gathering energy consumption data in main-
tenance of SCIM motors, this idea can be implemented the smart micro grid in the
mining industry to control every single SCIM with current and voltage and save a
lot of time of maintenance and increase production and the profitability.

5 Future Work

In the future work we propose to continue on gathering more data, and testing all
mechanical and electrical defects of the SCIM motors in the hardware, in software
we are proposing to implement machine learning algorithms, such as neural network,
Support vector machine, dCCN, and trying to image data to better understand the
behavior of the faulty and health SCIM machine, the next step is to develop an applica-
tion and implement it in a real case study of jaw crusher in mining industry, gathering
data, and classifying its faults and problems, while controlling the Maintenance KPIs.

References

1. Singh G, Anil Kumar TCh, Naikan VNA (2018) Efficiency monitoring as a strategy for
cost effective maintenance of induction motors for minimizing carbon emission and energy
consumption. Reliab Eng Syst Safety. https://doi.org/10.1016/j.ress.2018.02.015
2. Abdelati R, Mimouni MF (2011) Analytical solution of optimized energy consumption of
induction motor operating in transient regime. Eur J Control 4:397–411 © 2011 EUCA. https://
doi.org/10.3166/EJC.17.397-411
3. Kumar P (2020) Ananda Shankar Hati: deep convolutional neural network based on adaptive
gradient optimizer for fault detection in SCIM. ISA Trans. https://doi.org/10.1016/j.isatra.2020.
10.052
4. Raja Singh R, Yash SM, Shubham SC, Indragandhi V, Vijayakumar V, Saravanan PC, Subra-
maniyaswamy V (2020) IoT embedded cloud-based intelligent power quality monitoring
system for industrial drive application. Future Gener Comput Syst 112:884–898
5. Dorrell DG, Thomson WT (1997) Analysis of airgap flux, current, and vibration signals as
a function of the combination of static and dynamic airgap eccentricity in 3-phase induction
motors. IEEE Trans Ind Appl 33:24–34
1386 O. Laayati et al.

6. Frosini L (2019) Novel diagnostic techniques for rotating electrical machines—a review. Elec-
trical machines. Design, control and diagnosis (WEMDCD), Athens, Greece, 22–23 April
2019
7. Mehrjou MR, Mariun N, Marhaban MH, Misron N (2011) Rotor fault condition monitoring
techniques for squirrel-cage induction machine—a review. Mech Syst Signal Process 25:2827–
2848
8. Ebrahimi BM, Roshtkhari MJ, Faiz J, Khatam SV (2014) Advanced eccentricity fault recog-
nition in permanent magnet synchronous motors using stator current signature analysis. IEEE
Trans Ind Electron 61:2041–2052
9. Hendrickx K, Meert W, Mollet Y, Gyselinck J, Gryllias BCK, Davis J (2020) A general
anomaly detection framework for fleet-based condition monitoring of machines. Mech Syst
Signal Process 139:106585
10. Burnett R, Watson JF, Elder S (1996) The application of modern signal processing techniques
for use in rotor fault detection and location within three-phase induction motors. Signal Process
49:57–70
11. Faiz J, Moosavi SMM (2016) Eccentricity fault detection – from induction machines to DFIG—
a review. Renew Sustain Energy Rev 55:169–179
12. Han Q, Ding Z, Xueping X, Wang T, Chu F (2019) Stator current model for detecting rolling
bearing faults in induction motors using magnetic equivalent circuits. Mech Syst Signal Process
131:554–575
13. Nandi S, Lee SB (2011) Detection of eccentricity faults in induction machines based on
nameplate parameters. IEEE Trans Ind Electron 58(5)
14. Park Y, Choi H, Shin J Park J, Lee SB, Jo H (2020) Airgap flux-based detection and classification
of induction motor rotor and load defects during the starting transient. IEEE Trans Ind Electron
67:10075–10084
Intelligent Technique Proposed
for Nonlinear Inductor Modelling
for DC/DC Converters

Rafika El idrissi, Ahmed Abbou, Abderrahim Taouni, and Mohcine Mokhlis

Abstract The inductor is one of the main components for converting energy within
power converters. The inductance value does not always be constant as it decreases
with the current, particularly in the saturation region where the current’s value
increases. Thus, the inductor decrease affects the dynamic of the power converter.
To exploit its nonlinear behavior and to achieve a precise dynamic behavior of the
converter, an accurate inductor model must, therefore, be considered. Several models
has been dedicated to analyze these characteristics, but they are complicated and diffi-
cult to be implemented. In this paper, the Adaptive Neuro-Fuzzy Inference system
(ANFIS) model is used to solve the nonlinear behavior of the inductor. The proposed
ANFIS model is a clever neuro-fuzzy technique, which has proved to model and
control diverse processes that are nonlinear. The main feature of this model is that
the evolution of the flux vs current curve behavioral analysis has been taken into
account using an experimentally characterized set of data. The study is set up on the
DC/DC buck converter, using Matlab/Simulink. The results obtained by simulation
provide the inductor current and voltage prediction, which have good agreement with
the experimental measurements with relative error about 25%.

Keywords DC/DC converters · Nonlinear inductor modeling · Adaptive

Neuro-Fuzzy Inference System (ANFIS) · Magnetic saturation

1 Introduction

In many applications, energy converters are commonly used, particularly when

energy produced from renewable energy sources [1, 2]. These converters stock energy
in the inductor, which is driven by voltage source and able to transmit this energy

R. El idrissi (B) · A. Abbou · M. Mokhlis

Department of Electrical Engineering, Mohammadia School of Engineers,
Mohammed V University, Rabat, Morocco
A. Taouni
Laboratory of Electrical Systems and Control Engineering, Faculty of Sciences Aïn-Chock,
Hassan II University, Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1387
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_126
1388 R. El idrissi et al.

to the load maintained by a capacitor at constant voltage. Most of the theoretical

studies suppose that the inductor can be perfectly described by linear characteristics,
a magnetic flux dipole relative to the current. In fact, the inductance is characterized
by a nonlinear behaviour (saturation, hysteresis and power losses) which must be
included in circuit simulators to correctly simulate such power converters [3]. The
induction saturation theory triggers a renewed interest nowadays, primarily because
of its favorable effect on power density increase and cost reduction. In order to
avoid the inductor saturation, the converter should be operated at a higher switching
frequency. This leads to increased power converters and magnetic material switching
losses. A strong ripple current results when the inductance is decreased. The current
during off-on switching varies from low to maximum, so that the poor saturation
area is integrated. In order to properly design the converters, an understanding of
magnetic material behavior must be improved while the current changes. Thus, new
models must be created.
Many papers deal with the saturation of the inductance in literature [4–6]. In [7]
a finite element analysis (FEA) to model a dynamic core is presented. Reference [8]
proposes an analytical model of magnetically saturated inductance using lambert W
function. The inductor model is described in [9] based on a polynomial curve fitting
the variance in the inductance. In addition, some analytical nonlinear models are
presented in [10, 11].
In this work, the technique suggested was extended to model an inductor, with
a saturation effect of the magnetic core, using as simple and easy model based on
the adaptive neuro fuzzy inference (ANFIS) model. The model was trained based
on experimental flux vs current curve data. The ANFIS technique was chosen due
to its potential for creating fuzzy models with strong prediction skills [12, 13]. In a
basic buck converter, the nonlinear model of inductance reveals the accuracy of small
ripple and large ripple inductor current waveforms prediction. The proposed model
is simulated using Matlab/Simulink and the obtained results prove a high agreement
between the simulation and experimental inductor current waveforms. Therefore,
experimental results confirm the proposed model accuracy.
The rest of the paper is systemically as follows: the model of the buck converter
including the nonlinear inductor model is described in Sect. 2, Sect. 3 deals with
the inductor characterization, Sect. 4 is devoted to the inductor modelling, Sect. 5
presents the experimental and the simulation results, and conclusion is presented in
Sect. 6.

2 DC/DC Buck Converter with Nonlinear Inductor

To test the proposed model for the nonlinear inductor taking into account the satura-
tion, a buck converter, working in CCM, is used. The inductor is excited by a square
waveform provided by the converter. The power inductor of all switching converters
conducts DC and triangular ripple currents, in a similar manner. The buck converter
is also a good way to check and illustrate how saturating conditions usually meet
Intelligent Technique Proposed for Nonlinear Inductor Modelling … 1389

in switched-inducer circuits, particularly because in point of load applications it is

often operated at high currents. In CCM, when the switch is on the inductor under-
goes a constant voltage and, if it is linear, its current increases with inclination of
(Vin − Vout )/L, in contrast when the switch is off the current of the inductor moves
through the diode with an inclination of −Vout /L (Fig. 1).
If the average inductor current rises or the ripples rise, saturation is achieved; the
maximum current is consequently higher when compared to the linear case and the
form varies depending on the change in the induction Fig. 2 shows the details.

Fig. 1 Buck converter scheme with nonlinear inductor

Fig. 2 The buck converter

inductor current vs time
waveforms in case of a linear
inductor b nonlinear inductor
1390 R. El idrissi et al.

3 The Inductor Characterization

3.1 Linear Behaviour

The relation between magnetic flux and current is widely considered to characterize
an inductor:

 = Li (1)

The current vs voltage ratio is obtained by the Faraday law for the linear inductor,
which considers inductance to be constant:
d di
VL = =L (2)
dt dt

3.2 Nonlinear Behaviour

In most studies assume that the inductance L is constant. Indeed, in practice the
inductance varies with respect to current in the inductor. In particular, the inductor
core magnetic flux is a nonlinear function of the current as follow:

 = (i) (3)

Where (i) represents the inductor core’s magnetic characteristic. Figure 3a

depicts this function.
According to (3), the inductor voltage becomes as follow:

(a) (b)

Fig. 3 Flux value against current (a) Inductance value against current (b)
Intelligent Technique Proposed for Nonlinear Inductor Modelling … 1391

d d(i L ) d(i L ) di L di L
VL = = = = L(i L ) (4)
dt dt di L dt dt

Where:
∂(i L )
L(i L ) = (5)
∂i L

L (i L ) is a nonlinear function of the magnetic core saturation. Figure 3b demon-

strates an example of such function characterized experimentally, the values relating
to this are LR = 3.5 mH, L deepsat = 2.3 mH, L70% = 3.14 mH, L30% = 3.14 mH. An
iron ring core’s inductor with an inductive rate of 2.5 mH and a current rated of 3A
has been analyzed. It may be found that for low currents the highest value is shown.
The greater the current, the lower the inductance. Three operating regions can be
described in the inductance versus current curve:
– The region of low saturation where inductance decreases slightly with L R to
approximately 70% of L R .
– The roll-off region in which the inductance ranges from near L R to a near
saturation inductance value, the value is (Ldeepsat).
– Deep saturation, where the inductance is well below 30% and the value of Ldeepsat
is approaching.

4 The Inductor Modeling

When using an inductor, some non-ideal characteristics can be encountered. The

main one is the magnetic saturation. There exist different approaches to model the
saturation behavior of an inductor. The method has been proposed here is based on
the ANFIS model using the flux versus inductor current curve’s data. The advantage
of this method that does not require any mathematic formula and using ANFIS
model does not need a large amount of data to predict and compute the output
variables. Figure 4 shows the structure of the nonlinear inductance. It includes a
fuzzy logic controller block, which enforces a Simulink fuzzy inference system

Fig. 4 The proposed nonlinear model

1392 R. El idrissi et al.

(FIS), an integrator to get the flux, and an enforced current source, which transforms
the input signal into a corresponding current source.

4.1 Anfis Structure

A Neuro-Fuzzy architecture is includes a fuzzy logic and neural networks combi-

nation. It is a fuzzy sugeno paradigm, which has been developed as an integrated
framework for learning facilitation and adaptation. The main advantages of ANFIS
are to work with language phrases that human experts comprehend (if then) rules
and the capacity to receive training with samples of input output data. In [12, 13] are
given the ANFIS architecture and implementation examples for modeling a nonlinear
function and dynamic system recognition.
In this work, the ANFIS that is taken into consideration is based on an experimental
data set of the flux and the current of an inductor core working saturation. The
magnetic flux (i) is considered as input and the current i L as output. The data set
of the curve of the (F vs I) in Fig. 3a was taken into consideration.
Six membership functions were considered of Gaussian function type, which was
found suitable in this work as it is more accumulated with modelling behaviour, were
considered to generate the FIS model. The grid partition was chosen to define these
membership functions. Figure 5a demonstrates the required ANFIS architecture. It
consists of a five layers.
Once the data is loaded and the FIS model is generated, the FIS training step comes.
The hybrid algorithm was chosen, in this work, to train the input output data in 400
epochs and 0.0005 error tolerances. Figure 5b illustrates the curve of ANFIS training
error. After the ANFIS is trained, the system can be tested to check the system accu-
racy. The following figures illustrate the results obtained using the learned ANFIS.
The ANFIS generated surface is depicted in Fig. 6a. Figure 6b shows the ANFIS
rule viewer. The first column displays the member function of the magnetic flux

(a) (b)

Fig. 5 ANFIS structure (a) ANFIS training error (b)

Intelligent Technique Proposed for Nonlinear Inductor Modelling … 1393

(a) (b)
Fig. 6 ANFIS surface viewer (a) ANFIS rule viewer (b)

input parameter. The second column is the output variable’s membership function,
the inductor current. The rule Viewer simultaneously interprets the complete fuzzy
inference mechanism.
Figure 7 validates the precision of the current-flux curve modelling using ANFIS
model. It contrasts the real value of the inductor with the values expected by the
ANFIS model. It is obvious from Fig. 7 that the ANFIS predicted values well
represent the real values of the entire set of test data.

Fig. 7 Comparison of the inductor current values predicted by ANFIS models and the actual values
of data set and the error between them
1394 R. El idrissi et al.

Fig. 8 Simulink model of

ANFIS inductor current
predictor

The ANFIS MATLAB/Simulink model of inductor current predictor is illustrated

in Fig. 8. The developed ANFIS model has an extremely uncomplicated structure,
a short workout time, rapid computational speed and robustness properties. It has
proven that is suitable for nonlinear modelling of the inductor. The method is very
simple and easy to implement.

5 Simulation Results

The DC/DC buck converter was introduced in Matlab/Simulink taking into account
the inductor’s parasite effects and not the use of a linear inductor; the proposed model
in Fig. 5 was taken into account. In this simulation, the load has been changed at
0.008 s and the frequency has been lowered to 6 kHz. Figures 9a, b and c show the
inductor current, the magnetic flux, and the variation of the inductance value in the
step response. The step response is obtained by changing the load resistance from
70  to 13 . It is obvious that when the current increases, the inductance declines
and the current ripples rise. Magnetic flux also increases according to the inductor
current.
Figure 9d shows the comparison of the proposed model with linear inductance
model at load change. After load changes, the current in the proposed model, for
the inductor core saturation effect, is rising at a much higher rate. Inductance value
significantly decreases. Hence, the current ripples rise. In contrast, the linear model
does not consider saturation. Figure 9d gives the difference between the nonlinear
and linear inductor model.

6 Experimental Results

The circuit was also hardware designed for testing the simulation efficiency. B82615-
B2302M EPCOS inductor is used. Figure 11 illustrates the experimental system photo
(Fig. 10).
The model is tested by comparing the current and the voltage of the inductor.
Figure 11 reflects the current form and voltage that passes through the inductor while
the frequency of the switch is high and low. In Fig. 11a, the switching frequency is
20 kHz with a supply voltage 30 V and a duty cycle equals to 50%. It can be observed
Intelligent Technique Proposed for Nonlinear Inductor Modelling … 1395

(a)
(b)

(c) (d)

Fig. 9 The buck converter’s simulation results based on the proposed inductor model

Fig. 10 Tested buck

converter model picture

from this figure that the shape of the current is almost triangular. In Fig. 11b, the
switch frequency is decreased to 5 kHz and the voltage of supply is raised to 60 V
while the duty cycle is kept at 50%; it can be noted from the figure that the current
ripples are increased, the peaks on the current, and some fluctuations on the converter
output voltage.
Figure 12a displays the experimental and simulation current and Fig. 12b indicates
the relative error. It is observed that in the in the worst case, the loss is smaller than
15% in a switching period.
The simulation of the proposed method, the linear, and experimental waveforms
of the current are illustrated in Fig. 12c. Figure 12d shows the zoom of the last figure.
1396 R. El idrissi et al.

(a) (b)

Fig. 11 The buck converter assessed high frequency (a) and low frequency (b) experimental
waveforms: output voltage (blue), the inductor current (red)

(a) (b)

(c) (d)

Fig. 12 The inductor current simulation and experimental details comparison

It is observed in Fig. 12c that the inductor current of experimental circuit and the
proposed model are much closed and the difference is largest between the proposed
and the linear models. Therefore, the proposed model performs well to keep up the
current waveforms. Figure 13 shows waveforms of the inductor voltage obtained in
experimental and simulation studies.
Intelligent Technique Proposed for Nonlinear Inductor Modelling … 1397

Fig. 13 Simulation and experimental inductive voltages

7 Conclusion

A new analytical model, considering the nonlinear behaviour of inductor core, has
been proposed in order to exploit the inductor saturation. The model is based on
the ANFIS technique. The ANFIS model is trained using a separate collection of
measured data of magnetic flux vs inductor current. The proposed technique is char-
acterized by its simplicity and does not require a complicated mathematical equation.
In a Buck converter, the nonlinear model for the inductor has been implemented in
Matlab/Simulink and experimentally evaluated. The study of performances indicates
a very close coordination between inductor current ripple waveforms of the proposed
model and the experimental waveforms, which proves the proposed model prediction
accuracy.

References

1. Palma L, Todorovic MH, Enjeti P (2006) Design considerations for a fuel cell powered DC-DC
converter for portable applications. In: 21st IEEE APEC 2006, p 6
2. Vazquez MJV, Marquez JMA, Manzano FS (2008) A methodology for optimizing stand-alone
PV-system size using parallel connected DC/DC converters. IEEE Trans Ind Elect 7:2664
3. Fadil HEl, Giri F, Magueri OEl, Chaoui FZ (2009) Control of DC–DC power converters in the
presence of coil magnetic saturation. Cont Eng Pract 17:849
4. Salas RA, Pleite J (2011) Simulation of the saturation and air-gap effects in a POT ferrite core
with a 2-D finite element model. IEEE Trans Magn 47:4135
1398 R. El idrissi et al.

5. Salas RA, Pleite J (2013) Equivalent Electrical model of a ferrite core inductor excited by
a square waveform including saturation and power losses for circuit simulation. IEEE Trans
Magn 49:4257
6. Stadler A, Stolzke T, Gulden C (2012) Nonlinear power inductors for large current crest factors.
Proc PCIM 1548
7. Salas R, Pleite J, Olias E, Barrado A (2006) Nonlinear saturation modelling of magnetic
components for circuit simulation. IEEE. INTERMAG, 993
8. Gurleyen H, Erkan M (2017) Analytical modeling of magnetically saturated inductance by
lambert W function. J Magn 22:369
9. Lullo G, Scirè D, Vitale G (2017) Non-linear inductor modelling for a DC/DC buck converter.
Ren En Pow Qual J
10. You BG, Lee SW, Choi GB, Yoo DW, Lee BK (2011) Comparison of simulation and
experimental results for Mega Flux inductors in hybrid electric vehicles. In: 8th IPCE
11. You BG, Lee BK, Lee SW, Choi GB, Yoo DW (2011) Experimental comparison of mega-flux
and JNEX inductors in high power DC-DC converter of hybrid electric vehicles. IEEE. VPPC
1
12. Susitra D, Paramasivam S (2014) Non-linear flux linkage modeling of switched reluctance
machine using MVNLR and ANFIS. J Int Fuz Syst 26:759
13. Bassoui M, Ferfra M, Chraygane M (2016) ANFIS modeling of nonlinear inductance. In: 2nd
ICEIT
14. Jang JSR, Sun CT (1995) Neuro-fuzzy modeling and control. Proc IEEE 83:378
15. Jang JSR (1993) ANFIS: adaptive-network-based fuzzy Inference Systems. IEEE Trans Syst
23:665
High Order Sliding Mode Power Control
of Doubly Fed Induction Generator
Based on Wind Turbine System

Yahya Dbaghi, Sadik Farhat, Mohamed Mediouni, Hassan Essakhi,

and Aicha Elmoudden

Abstract This paper discusses the control of the active and reactive stator power of
a doubly fed induction generator (DFIG) based on a wind energy conversion system
using super-twisting sliding mode control (STSMC) control. The objective of this
work is to evaluate the robustness of STSMC control against the variation of internal
parameters of the DFIG, and also to reduce the power ripple resulting from the
chattering phenomenon which is the major drawback of the conventional SMC. The
modeling of the DFIG and the implementation of the proposed control strategy are
performed and tested using the numerical simulation environment Matlab/Simulink.

Keywords DFIG · Sliding mode control (SMC) · Super-twisting · Wind turbine ·

MPPT · Tip speed ratio · Vector control

1 Introduction

The use of renewable energies is nowadays becoming an indispensable reality.

Whether onshore or offshore, today, wind energy is taking its rightful place in the
market of electricity production. They are clean, safe, and less toxic and, above all,
they are renewable.
The overall studied wind energy system that is shown in Fig. 1 consists of the
following devices: a wind turbine with a gearbox, a doubly fed induction generator,
and two bi-directional static converters.
The implementation of DFIG in wind energy systems has been the subject of
several pieces of research, due to its ability to operate over a wide range of wind
speeds in addition to the reduced sizing of the associated converters (about 30%
of the nominal power) [1]. The control of the DFIG and its operation has been
intensively addressed by the researchers. In the literature in recent years, several
control approaches have been proposed, notably the linear PI control, the non-linear

Y. Dbaghi (B) · S. Farhat · M. Mediouni · H. Essakhi · A. Elmoudden

Laboratoire Des Sciences de L’Ingénieur et Management de L’Energie(LSIME), ESTA Ibn Zohr
University, BP 33/S, 80000 Agadir, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1399
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_127
1400 Y. Dbaghi et al.

Fig. 1 Synoptic diagram of the studied system

controls such as backstepping and sliding mode control, and the intelligent controllers
based on fuzzy logic and neural networks.
Boubzizi, S. et al. have demonstrated in [2] that the PI controller is limited in
the linear functioning of the system otherwise it diverges. Kelkoul B. et al. have
concluded that the use of the conventional sliding mode offers acceptable robustness
and a fast response time except that it leads to an undesirable chattering phenomenon
[3].
In our paper, the objective is to design a high order sliding mode controller
that allows both the minimization of the chattering phenomenon and ensures high
robustness and rapid response against unpredictable internal disturbances of the
machine.
This article is organized as follows: in Sect. 2 we present the modeling of the wind
turbine. Section 3 will be devoted to the modeling and vector control of the DFIG.
The design of the proposed control is described in Sect. 4. The 5th section presents
the simulation results and their Interpretations.

2 Modeling of Wind Turbine

The following formula presents the power produced by the wind turbine as [4]:

1
Paer = Cp (λ, β).ρ.π.R2 .V3 (1)
2
High Order Sliding Mode Power Control of Doubly Fed Induction Generator ... 1401

Where: ρ is the air density, R is the radius of the blades, V is the wind speed, β is
the pitch angle.
λ is the speed ratio defined by [5]:

R.t
λ= (2)
V
t : The angular speed in the shaft. of turbine.
Cp(λ, β) is the power coefficient represent the efficiency of the wind energy
conversion system, which given by [6]:

Cp (λ, β) = C1 (C2 × A − C3 β − C4 )e−c5 ×A + C6 λ
(3)
A = λ+0.08β
1
− 0.035
β3 +1

With C1 = 0.5109, C2 = 116, C3 = 0.4, C4 = 5, C5 = 21, C6 = 0.0068.

For a MPPT operating with β = 0◦ , the optimal value of power coefficient is
C Pmax ≈ 0.475, and the optimal speed ration λopt ≈ 8.1 (Fig. 2).
The rotary speed and torque at the generator shaft are defined as [7]:

mec = G.t
(4)
T g = TGaer

Where, mec is the mechanical speed of the generator, Tg is the generator torque,
and Tem is the electromagnetic torque.
Newton’s second law of motion on the generator shaft gives [8]:

Fig. 2 Power coefficient curves in terms of β and λ

1402 Y. Dbaghi et al.

Fig. 3 Block diagram model of wind turbine system with TSR MPPT algorithm


Tg − Tem − f  = J ddtmec
(5)
J = GJt + Jg

With: J is the total inertia, Jg is the generator inertia and, f is the viscosity
coefficient, and Jt is the turbine inertia.
The MPPT algorithm used is based on TSR (Tip Speed Ratio).the optimal
electromagnetic torque expression for the maximum power is defined as [8] (Fig. 3):

C pmax (λ, β) ρ.π.R 5 2

Temopt = × × mec (6)
λopt
3 2 G3

3 Modeling and Vector Control of the DFIG

The stator and rotor DFIG’s voltages in the park frame are expressed as [9]:
⎧
⎪
⎪ Vds = Rs Ids + d
ϕ
dt ds
− ωs ϕqs
⎨
Vqs = Rs Iqs + d
ϕ
dt qs
+ ωs ϕds
(7)
⎪
⎪ V = Rr Idr + d
ϕ − ωr ϕqr
⎩ dr dt dr
Vqr = Rr Iqr + d
ϕ
dt qr
+ ωr ϕdr

With: Rr and Rs are respectively the rotor and stator resistances per phase.
The stator and rotor flux of DFIG are defined by:
High Order Sliding Mode Power Control of Doubly Fed Induction Generator ... 1403
⎧
⎪
⎪ ϕds = L s Ids + M Idr
⎨
ϕqs = L s Iqs + M Iqr
(8)
⎪
⎪ ϕ = L r Idr + M Ids
⎩ dr
ϕqr = L r Iqr + M Iqs

Where: Idr , Iqr , Ids , Iqs , are respectively the rotor and stator currents in dq-axis. Lr ,
Ls are the rotor and stator inductances. M is the DFIG mutual inductance.
The expression of the stator active and reactive powers is defined by [10]:

Ps = Vds Ids + Vqs Iqs
(9)
Q s = Vqs Ids − Vds Iqs

Electromagnetic torque’s expression is given as:

pM
Tem = (Idr ϕqs − Iqr ϕds ) (10)
Ls

p is the pair pole number.

By orientation of the stator flux in d-axis direction [11] we obtain:

ϕds = ϕs , ϕqs = 0 (11)

The value of the stator resistance is small as that can be neglected, then stator
voltages can be rewritten as:

Vds = 0
(12)
Vqs = Vs = ωs ϕs

From Eq. (8), Eq. (9) and Eq. (12), the stator active and reactive powers can be
written as function of rotor currents as:

Ps = Vs LMs Iqr
2 (13)
Qs = ωVssLs Vs LMs Idr

By combining the previous equations, the rotor voltages (d, q) in terms of the
rotor currents (d, q) are expressed as:
⎧ 
⎨ Vdr = Rr Idr + Lr − M2 dIdr + eq
 Ls dt
(14)
⎩ Vqr = Rr Iqr + Lr − M2 dIqr + ed
Ls dt
⎧ 
⎨ eq = −gω Lr − M Iqr2

s L s
(15)
⎩ ed = gωs Lr − M2 Idr + gMVs
Ls Ls
1404 Y. Dbaghi et al.

Where: g is the slip, and ed , ed represent the coupling terms between the dq-rotor
voltages.

4 Super-Twisting Sliding Mode Control of the DFIG

The sliding mode is one of the most powerful nonlinear control techniques that is
based on the stability of the Lyapunov. This technique was developed by V. I. Utkin in
[12], its main operating principle lies in its ability to force a system’s state variables to
slide in a specific area called the sliding surface around a desired state of equilibrium.
This is achieved while respecting stability in the sense of Lyapunov. Its weak point
is the phenomenon of chattering caused by the discontinuous part of the control, this
phenomenon degrades and causes undesirable functioning of the system. The use
of a super-twisting sliding mode control presented a solution to this problem by its
ability to reduce this phenomenon to minimum as possible.
The STSMC The control law is given by [12]:

V(t) = VC (t) + Vst (t) (16)

This control law consist of a continuous control VC (t), and a super-twisting control
Vst (t) that is defined by [13]:

Vst (t) = v1 (t) + v2 (t) (17)

With:
 √
v1 (t) = −α |S|.sat (S)
(18)
v2 (t) = −γ .sat (S)

Where α and γ are positive gains and are calculated as follows:


ϕ
α > τmin
4ϕ τmax (α+ϕ) (19)
γ ≥ τmin 2 τmin (α−ϕ)
2

With: ϕ, τmin , τmax are positive constants.

The sliding surfaces are chosen as follows:

S(Ps ) = Ps∗ − Ps
(20)
S(Qs ) = Qs∗ − Qs

By deriving Eq. (20) and using Eqs. (13), and (14) we get:
High Order Sliding Mode Power Control of Doubly Fed Induction Generator ... 1405


Ṡ(Ps ) = Ṗs∗ + σL
Vs M
s Lr

Vqr − Rr Iqr − ed
 (21)
Ṡ(Qs ) = Q̇s∗ + σLVs M
s Lr
Vdr − Rr Idr − eq

M2
With: σ = 1 − Lr Ls
.
In the sliding mode we have S(Ps , Qs ) = Ṡ(Ps , Qs ) = 0. Then we can express the
continuous controls of the dq rotor voltages by:

Vqr (C) = −Ṗs∗ σLs Lr
Vs M
+ Rr Iqr + ed
∗ σLs Lr (22)
Vdr (C) = −Q̇s . Vs M + Rr Idr + eq

The super-twisting controls of the dq rotor voltages are given by:

 √
Vqr(st) = −α1 |S(Ps )|.sat[S(Ps )] − γ1 .sat[S(Ps )]
√ (23)
Vdr(st) = −α2 |S(Qs )|.sat[S(Qs )] − γ2 .sat[S(Qs )]

To ensure the stability of system control we should respect the Lyapunov condi-
tion S(Ps , Qs ) × Ṡ(Ps , Qs ) < 0, then we should have α1 , α2 , γ1 and γ2 as positive
constants.

5 Results and Discussion

5.1 Simulation of the MPPT Operating

The results of this section show the MPPT operation of the wind turbine based on
the DFIG. We applied a random wind profile (Fig. 4 (a)) of a mean value of 9 (m/s)
to the wind turbine. The results obtained show a successful operation while ensuring
an optimal tracking of the maximum power (Fig. 4 (b)). This is verified by the power
coefficient maintained at its maximum value according to Fig. 5.

Fig. 4 (a) Wind turbine profile (b) MPPT stator active power response
1406 Y. Dbaghi et al.

Fig. 5 Power coefficient Cp response

5.2 Simulation Results in the Robustness Test

In order to test and validate the performance of our proposed control method, we have
attacked the rotor side converter of the DFIG with two profiles representing active
and reactive power references such as illustrated by the blue traces in Fig. 6. The
simulation results shown in Fig. 6 describe the responses of the active and reactive
stator powers of the generator under two tests, a normal operation and an operation
disturbed by an increase of the stator resistance and the stator inductance of the DFIG
by 50% of their nominal values.
The simulation results show a great tracking of the active and reactive power
references and the chattering phenomenon is almost negligible for the normal test
(red trace) and also for the robustness test (pink trace). We can also notice that the
response time is quite fast: tr5% = 1.5(ms) in the normal test. The robustness test
has only shown a degradation of the response time, which has increased to 5.2(ms)
that remains just as fast. It is also observed that the decoupling of the q-axis and
the d-axis has been perfectly achieved, allowing an independent control of the stator
active and reactive power.
Figure 7 shows the evolution of the stator and rotor currents of DFIG. It is clear
that they vary according to the stator powers references.
By comparing our results with others published in the literature, our approach
shows a very fast response time with a negligible chattering phenomenon in compar-
ison to the results obtained by the conventional SMC proposed by Y. Bekakra et al.
In [14]. For the backstepping control proposed by L. Saihi et al. in [15], our method
offers a perfect decoupling with respect to their controller, which presents a poor sepa-
ration between the dq-axis of Park, which have caused of intense and undesirable
peaks in the active and reactive power responses.
High Order Sliding Mode Power Control of Doubly Fed Induction Generator ... 1407

Fig. 6 (a) Stator active power response (b) Stator reactive power response

Fig. 7 (a) Stator currents response (b) Park frame rotor currents response

6 Conclusion

This paper presents the control of the active and reactive stator power of the DFIG by
Super-twisting SMC according to the DFIG model with stator flux orientation. The
simulation results have largely satisfied the required objectives, they have shown a
high robustness against variations of the machine’s internal parameters, a fast enough
response time, and the chattering phenomenon has now been nearly neglected. From
these results it can be concluded that the use of a high order sliding mode controller
can significantly enhance the performance of a DFIG based wind energy system.

Appendix

DFIG Nominal Parameters

Pn = 7.5 kW, Vs = 380/220 V, f = 50 Hz, p = 2, Rr = 0.19 , Rs = 0.455 , Ls
= 0.07H, Lr = 0.0213H, M = 0.034H.
Wind Turbine Parameters
Pt = 7.5 kW, ρ = 1.225 kg.m−3 , R = 3 m, G = 5.4, J = 0,042 kg.m−1 , f =
0,024 N.m.s.rad−1 .
1408 Y. Dbaghi et al.

References

1. Tahir A, EL-Mukhtar M, EL-Faituri A, Mohamed F (2020) Grid connected wind energy system
through a back-to-back converter. Comput Electr Eng (85):106660
2. Boubzizi S, Abid H, El Hajjaji A, Chaabane M (2018) Comparative study of three types of
controllers for DFIG in wind energy conversion system. Prot Control Mod Power Syst 3
3. Kelkoul B, Boumediene A (2019) Nonlinear sliding mode control of DFIG based on wind
turbines. ICAIRES 2018, LNNS 62, pp 206–215
4. Mensou S, Essadki A, Nasser T, Bououlid Idrissi B, Ben TL (2020) Dspace DS1104 imple-
mentation of a robust nonlinear controller applied for DFIG driven by wind turbine. Renew
Energy 147:1759–1771
5. Nahilia H, Boudour M, Cardenas A, Agbossou K, Doumbia M (2019) Doubly fed induction
wind generators model and field orientation vector control design and implementation on
FPGA. Int J Dyn Control. 7:1005–1014
6. Farhat S, Alaoui R, Kahaji A, Bouhouch L (2018) Wind turbine MPPT strategy with DFIG
vector control. Int Rev Model Simul 11(6):406
7. Sabiri Z, Machkour N, Rabbah N, Nahid M, Kheddioui E (2017) Command of a doubly-fed
induction generator with a backstepping controller for a wind turbine application. Int Rev
Autom Control 10(1):56–62
8. Dbaghi Y, Farhat S, Mediouni M (2020) First order sliding mode and super-twisting sliding
mode control of doubly fed induction generator driven by wind turbine system: a comparative
study. J Adv Res Dyn Control Syst 12(04):1656–1667
9. Moradi H, Vossoughi G (2015) Robust control of the variable speed wind turbines in the
presence of uncertainties: a comparison between H∞ and PID controllers. Energy 90:1508–
1521
10. El Azzaoui M, Mahmoudi H (2017) Fuzzy control of a doubly fed induction generator based
wind power system : Int. J. Wseas Trans Power Syst (12)
11. Kiruthiga B (2015) Implementation of first order sliding mode control of active and reactive
power for DFIG based wind turbine. Int J Inf Futur Res 2(8):2487–2497
12. Utkin VI (1973) Sliding modes in multidimensional systems with variable structure. In Confer-
ence on Decision and Control including the 12th Symposium on Adaptive Processes, pp
727–727, IEEE ,San Diego, CA, USA
13. Han Y, Ma R, Pan W, Wang C (2019) A novel super-twisting algorithm-based direct power
control strategy for doubly fed induction generator. In 12th Asian Control Conference (ASCC),
pp 1619–1624, Kitakyushu-shi, Japan
14. Chalanga A, Kamal S, Fridman LM, Bandyopadhyay B, Moreno JA (2016) Implementa-
tion of super-twisting control: super-twisting and higher order sliding-mode observer-based
approaches. IEEE Trans Ind Electron 63(6):3677–3685
15. Bekakra Y, Ben Attous D, Bennadji H (2019) Sliding mode control of DFIG driven by wind
turbine with SVM inverter. In Hatti M. (eds) Renewable Energy for Smart and Sustainable
Cities. ICAIRES 2018. LNNS, (62). Springer, Cham
16. Saihi L, Bakou Y, Harrouz A, Colak I, Kayisli K, Bayindir R (2019) A comparative study
between robust control sliding mode and backstepping of a dfig integrated to wind power
system. In 7th International Conference on Smart Grid (icSmartGrid), IEEE
Robust Voltage Control for Four-Phase
Interleaved DC-DC Boost Converter
for Electric Vehicle Application

Mohamed Barara, M. R. Barakat, Nabil Elhaj, and Ghania Belkacem

Abstract The electric vehicle and the quality of power electric are currently the most
requested areas of research in power electronics. In this context this paper presents
a robust control strategy for 4 phase interleaved bidirectional DC/DC for HEV/HV
application. The interlaced topology is used to reduce the current and voltage ripples
and to limit in current rating the inductor switch. Indeed, this paper aims to improve
the dynamics of the output voltage of this converter, two control strategies are
analyzed and conformed by simulation and experimental work. Advanced control
technique based on RST controller and PI controller. The work presented in this
article shows the advantage of the RST controller method over the PI controller, the
measurements correspond well to the simulated results.

Keywords Electric vehicles · Power quality · Four phase interleaved boost

converter

1 Introduction

The Electric Vehicles (EVs) have advantages over existing cars until now, these cars
contribute to reducing CO2 emissions and earth warming and improving public health
and reducing ecological damage, which slows down climate change. The advent of
electric cars has reduced the intensive exploitation of natural resources as petroleum

M. Barara (B) · G. Belkacem

ESME, SUDRIA, Paris, France
M. R. Barakat
Faculty of Engineering at Helwan, Helwan University, 1 Sherif Street, Helwan, Cairo P.O. 11792,
Egypt
N. Elhaj
Materials and Subatomic Physics Laboratory, Faculty of Sciences, Ibn Tofail University, Kénitra,
Morocco
M. Barara
Lusac Laboratory, University of Caen-Base Normandie, Cherbourg-Octeville, France

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1409
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_128
1410 M. Barara et al.

Fig. 1 Block diagram for the power system configuration

and gas. EVs have turned to other alternative sources of electrical energy which has
improved the environment.
The EVs are essentially based on the bidirectional DC-DC converter which
connects the inverter to the energy storage system. During, the buck mode the energy
is sent back to the energy storage system, while in boost mode power is supplied
from the energy storage system to the drive system [1] as shown in Fig. 1.
In the EV system, the input voltage of the DC-DC converter boost is the battery
voltage and the current flowing through this system can flow in either direction
depending on the mode of operation of the vehicle. This is the bidirectional converter.
The boost converter is used in EVs to maintain the motor current at the desired
value in order to ensure motor voltage regulation. Thus it allows the regulation of the
speed of the engine and the torque and consequently of the vehicle in traction and also
in electric braking. The bidirectional converter is a DC-DC converter that allows us to
obtain from a fixed DC voltage source a DC voltage source with an adjustable value
of upper or lower value (boost or buck of voltage). It essentially consists of power
switches (transistors, diodes…) and passive components (inductors, capacitors …).
The DC-DC converter provides power to the vehicle’s onboard network (lighting,
computers, air conditioning…) and recharges the low voltage battery from the high
voltage battery.
Among the topologies of the DC-DC converters existing in the literature, the
choice was based on the interlacing topology because it is very suitable for the system
studied in this research work given the minimization of the nominal current of the
inductance switch and reduction of current and voltage ripple, smaller component
size, improved output impedance properties. Several control techniques have been
proposed in the literature to control converters in order to solve the problem of the
load variation. Among the classic converter control techniques there is the PI, PID,
fuzzy logic controller or a mixed controller between them [1].
The interleaving technique allowed to reduce input/output voltage ripples, and
minimize power loss to provide higher efficiency as well as reduce the weight and
size of passive converter components while reduction can also be performed by
increasing the switching frequency.
In reference [2], the mode of operation of the converter with the input current
ripple lowest compared to regular inductance, and then determine the appropriate
Robust Voltage Control for Four-Phase Interleaved DC-DC Boost Converter ... 1411

range of duty cycle. As the reference [3] proposes for the control of the DC-DC
converter the controller in sliding mode in order to reduce the ripples in the output
voltage of the converter and the input currents. Finally, reference [4] also proposes
a digital control based on the synchronization and the phase shift operation. While
this paper focuses on improving the dynamics of the output voltage of a DC-DC
converter, within this framework two control strategies are analyzed in simulation
and tested on an experimental step for one phase.
In order to improve the regulation of the output voltage of the DC-DC converter
influenced by any disturbance, we propose an advanced control technique based on
RST controller which has proved its robustness in terms of stability, response fast and
dynamic and also in terms of simplicity to implement. The RST controller is tested
experimentally with a single-arm DC-DC converter using the dSPACE controller;
the results once again confirm the robustness of the control technique proposed in
this work.
PI controller coefficients are designed to compare the behavior of this classic
controller with advanced RST controller. The simulation results showed that the
RST controller gave transient rapid response with less startup undulations stable
than the PI controller.
This paper is organized in 5 sections. The Sect. 2 is dedicated of the studied system
description. The advanced RST controller is presented in the Sect 3. In the Sect 4 the
design of the RST controller is presented. The last section is devoted to the results
and discussions.

2 Studied System Description

To ensure correct operation of the step-up converter, a synchronization circuit is

required in order to generate the signal necessary for the control of this converter [4].
Figure 2 shows the digital control system used for the boost converter. A controller
based on the error signal of the external voltage loop generates the reference current.
The inductance current in each converter is compared to the reference current.
The generation of the PWM signals can be accomplished by programming general
purpose timers with a switching frequency period and initializing four different
counters registers [5].
The control strategy of the proposed bidirectional DC-DC converter is shown in
Fig. 2, its composed of a cascade voltage controller with current controllers, which
are used to regulate the output voltage with a maximum preset current and minimal
[6]. The Control technology based on RST controller essentially consists of a pole
placement method based on zero gonpole cancellations and  solving
  aDiophantine 
equation [8]. The RST controller consists of polynomials R z−1 , S z−1 and T z−1 ,
provides the control law u(k) with the following z-transform:
1412 M. Barara et al.

Fig. 2 Control mechanism

Fig. 3 RST controlled discrete-time system

Robust Voltage Control for Four-Phase Interleaved DC-DC Boost Converter ... 1413
   
 −1  T z−1  −1  R z−1  −1 
U z =  −1  Yr z −  −1  Y z (1)
S z S z

Where Yr (k) is the reference input. The following figure shows the block diagram
of the RST controller [7] (Fig. 3).

3 Simulation Results

The simulation work has been carried using Matlab/Simulink. The model of the
studied system presented in Fig. 1 is simulated by applying the parameters of the
Table 1. In this section, the studied system has been simulated using a four arms
converter.
Figure 2 shows the control scheme that has been designed in the case where the
DC-DC converter is powered by an ideal source on the input side, and a variable
load on the output side. This scheme has been tested in several configurations. The
parameters of the converter support are given in Table 1.

3.1 Variation of Load and Reference Voltage

We varied the load (between 0 and 0.8 s and from 0.8 s to 1.4 s and the reference
voltage between 1.4 s and 1.8 s) to see the response and the behavior of the studied
system especially when we have applying two controls type on this system.
Figure 4 shows the output voltage of the Four-phase converter with PI controller.

3.1.1 PI Controller

A smalls peaks in the output voltage is observed in Fig. 4 due to sudden variation
of load and reference voltage, but it recovers quickly by the corrective action of
the PI controller. It is observed that the value of the output voltage is maintained

Table 1 Simulation
Parameter Value
parameters
Input voltage 200 V
L1,L2,L3 and L4 2 mH
C 26000 µF
R1 32
R2 16
F 20 kH
1414 M. Barara et al.

Fig. 4 Output voltage of four phase converter with PI controller

at a constant value of reference even with variations and variation of reference DC

voltage.
We can clearly see that after varying the load, the output voltage follows with
delay and appearance of the ripples with different amplitudes, which proves the
ineffectiveness of the PI controller.

3.1.2 Advanced Controller

Now, the RST controller is applied. We can see the improvement carried by the
advanced controller by reducing the ripples of the output voltage also at the level of
the pursuit of the reference voltage.
Figure 5 shows the output voltage of the system studied after applying the
advanced control. The output voltage precisely pursues its reference without ripples
when the load is varied (between 0 and 0.8 s; from 0.8 s and 1.4 s and reference voltage
between 1.4 s and 1.8 s), which proves the efficiency of the proposed controller.

Fig. 5 Output voltage of for phase converter with advanced controller

Robust Voltage Control for Four-Phase Interleaved DC-DC Boost Converter ... 1415

Finally, it’s clear that even if varied when the reference voltage and load the
output voltage perfectly follows its reference. This result proves the effectiveness of
the proposed method of control in the simulations works.

4 Experimental Results

In this section we present the results of the experimental setup. We limited the
experimental setup to one arm of the converter using a dSPACE card.
Figure 6 shows the experimental setup of the studied system with the different
devices used in work.

4.1 Variation of Reference Voltage

Figure 7 show the experimental performance of the control scheme for the output
voltage under variable reference voltage to prove the efficiency of the proposed
control method, by applying the classic PI controller.

4.1.1 PI Controller

We varied the reference voltage while the PI controller is applying, it s clear that the
output voltage follows its reference with a fairly long delay time (Fig. 7).

4.1.2 Advanced Controller

This time we applied the advanced RST controller, it is obvious that the output voltage
follows precisely its reference with a faster response time, a reduced overshoot
(Fig. 8).

Fig. 6 Test stand of one

phase interleaved boost
converter
1416 M. Barara et al.

(a) Output voltage under sudden variation of reference voltage with PI controller

(b) Zoom of Output voltage under sudden variation of reference voltage with PI controller

Fig. 7 Variation of reference voltage with PI controller

4.2 Variation of the Load

4.2.1 PI Controller

After varying the load, we can clearly see that we observe ripples and small peaks
(Fig. 9).

4.2.2 Advanced Controller RST

This figure shows that the experimental work confirms the simulation results, after
applying the advanced RST controller and even if when the load is changed the value
of the output voltage remains stable and this voltage follows its reference perfectly
without ripple and small peaks (Fig. 10).
Robust Voltage Control for Four-Phase Interleaved DC-DC Boost Converter ... 1417

(a) Output voltage under sudden variation of reference voltage with advanced controller

(b) Zoom of output voltage under sudden variation of reference voltage advanced controller

Fig. 8 Output voltage after applied the advanced controller

5 Conclusion

The study portrayed in this paper purposes to improve the quality both output of
current and voltage of the DC-DC converter connecter to an electric vehicle. To
overcome this problem, a boost converter was adopted introducing a new controller
design for regulating a DC bus voltage. The proposed design of the controller is based
on RST controller. Theoretical study of the RST approach was dressed and verified
through computer simulations performed under MATLAB/Simulink software. The
simulation results and experimental works are organized in subsections. The objective
was to prove the effectiveness of the RST controller when compared to PI controller.
The comparative study and analysis are demonstrated the superiority of the RST
controller which it s presents a good regulation of voltage, a faster response time, a
reduced overshoot, and a lower static error.
1418 M. Barara et al.

(a) Output voltage under sudden variation of load with PI controller

(b) Zoom of output voltage under sudden variation of load with PI controller
Fig. 9 Output voltage with PI controller
Robust Voltage Control for Four-Phase Interleaved DC-DC Boost Converter ... 1419

(a) Output voltage under sudden variation of load with advanced controller

(b) Zoom of output voltage under sudden variation of load with advanced controller

Fig. 10 Output voltage with controller REST

References

1. Schumacher D, Magne P, Preindl M, Bilgin B, Emadi A (2016) Closed loop control of a six
phase interleaved bidirectional DC-DC boost converter for an EV/HEV application. In IEEE
Transportation Electrification Conference and Expo (ITEC), pp 1–7. Dearborn
2. Kascak S, Prazenica M, Jarabicova M, Konarik R (2018) Four phase interleaved boost converter-
analysis and verification. Acta Electrotechnica et Informatica 18(1)
3. Ayoubi Y, Elsied M, Oukaour A, Chaoui H, Slamani Y, Gualous H (2016) Four-phase interleaved
DC-DC boost converter interfaces for super-capacitors in electric vehicle application based on
advanced sliding mode control design. Electric Power Syst Res 134:186–196
4. Nagulapati K, Rangavalli V, Vanajakshi B (2014) Modelling, analysis and simulation of 4- phase
boost converter. Int J Electr Comput Energ Electron Commun Eng 8(9)
5. Nisak K, Aida SA, Zainul AZ (2018) Interleaved 4-Phases DC-DC converter for electric
scooter to improve efficiency. In: MATEC Web Conference Volume 225, 2018 UTP-UMP-VIT
Symposium on Energy Systems
1420 M. Barara et al.

6. Elsied M, et al (2015) Efficient power-electronic converters for electric vehicle applications. In

2015 IEEE Vehicle Power and Propulsion Conference (VPPC), Montreal, QC, pp 1-6
7. Prasanth P, DRS (2016) Investigation of four phase interleaved boost converter under open loop
and closed loop control schemes for battery charging applications. Int J Adv Mater Sci Eng
(IJAMSE) 5(1)
8. Godoy E, Ostertag E (2006) RST-controller design : a rational teaching method based on two
diophantine equations. IFAC Proc Volumes 39(6):541–546
MPPT and Pitch Angle Based on Neural
Network Control of Wind Turbine
Equipped with DFIG for All Operating
Wind Speed Regions

Moussa Reddak, Ayoub Nouaiti, Anass Gourma, and Abdelmajid Berdai

Abstract The wind turbine coupled with a doubly fed induction generator (DFIG)
demonstrates good robustness and energy performances against the intermittent
nature of the wind, which allows many control strategies driving this system to
control it in the right conditions. Besides, the wind energy is completely depending
on the wind flow which is variable and can overpass the rated value of the wind speed.
In this case, the wind turbine should be sheltered versus load stress and the poten-
tial risk of destruction. In this paper, to profit from wide wind speed and exploiting
the totality of wind energy, an artificial neural network is designed to implement a
maximum power point tracking MPPT-pitch angle control strategy for controlling the
wind turbine. This strategy lets the DFIG extract the maximum power and protects
the wind turbine system from the overload of the electromagnetic torque when the
wind speed is higher than its rated value. The simulation results are performed using
the Matlab/Simulink environment and show well control performances under a wide
wind speed range.

Keywords Wind turbine · DFIG · Neural network · MPPT · Pitch angle control

1 Introduction

Nowadays, to face the challenges of increased energy demand and climate changes,
several renewable energy sources are committed to meeting the expectations of manu-
facturers and consumers. Among these renewable energy sources, wind energy is
considered a vital energy source that relies on the kinetic energy extracted from the
wind.

M. Reddak (B) · A. Nouaiti · A. Gourma · A. Berdai

Laboratory of Energy and Electrical Systems ENSEM, Hassan II University, Casablanca, Morocco
A. Nouaiti
Laboratory of Computer Science, Applied Math and Electrical Engineering, Moulay Ismail
University, IEVIA Team, EST, Meknès, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1421
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_129
1422 M. Reddak et al.

Due to the intermittent nature of the wind, new wind turbines are equipped
with a robust energy conversion system [1, 2] for achieving maximum energy effi-
ciency. Among the components of this system, a generator is used for converting
mechanical energy to electrical one. Currently, (DFIG) [3, 4] is extensively used
as an electromechanical structure in wind energy conversion systems because of its
advantages.
To benefit of full range of the wind speed, control strategies are designed to reach
the maximum power once the wind speed is average, and also to preserve the power
of the wind turbine at its maximale value while the wind speed is upper relying on
the pitch control [5].
In this context, several robust control strategies, for example, nonlinear controls
(sliding mode, backstepping) [6] and even fuzzy logic control [7], are used to drive
the wind turbine in different wind regions. Artificial neural network (ANN) control
also meets this need since it has given good results in the areas of complex problem
solving and robust control [8–10].
The objective of this paper is the control of the wind turbine in large wind speed
by adopting the pitch control. The pitch angle is generated by a neural network while
the wind speed is upper to the limit value. Also the system control strategy aims to
warrant the taking out of the maximum power once the wind speed is less.
This paper is planned as follow: in Sect. 2, the model of the system is developed.
In Sect. 3, the pitch angle control using neural network is described. The strategy
control of the system will be introduced in Sect. 4 and before conclusion, some
simulation results will be discussed in Sect. 5.

2 Model of the Wind Turbine System

The wind turbine equipped with the DFIG system, shown in Fig. 1, contains three
elements: the generator (DFIG), the turbine and the power converters. These later are
divided into two converters, a rotor side converters (RSC) which controls the powers
delivered by the generator, and a grid side converter that controls the DC-bus. The
control of the grid side converter is not concerned in this paper.

2.1 Turbine Model

The wind speed profile could be characterized by Eq. 1 [11, 12]:

Vw (t) = 12.8 + 0.869 sin(3t) + 0.74 sin(5t)−

0.624 sin(10t) + 0.24 sin(30t) + 0.49 sin(50t)+ (1)
0.24 sin(100t)
MPPT and Pitch Angle Based on Neural Network Control of Wind Turbine Equipped ... 1423

Fig. 1 DFIG system piloted by wind turbine

The power taken from kinetic energy, according to Betz’s theory, is defined as
follows:
1
Pa = ρπ R 2 v 3 C p (λ, β) (2)
2
According to this theory, the extraction of the power produced by the wind flow
through the wind turbine cannot be fully achieved. But this power is fractionated by
the coincidence effect of the wind with the blades. To define this limit, the power
coefficient C p that depends on this aerodynamic power is defined in Eq. 3:

150.9 − 18.4
C p (λ, β) = 0.46( − 0.579β − 0.00201β 2.14 − 13.21)(e λi ) (3)
λi

λi is expressed in Eq. 4:

1 1
λi = + 3 (4)
λ + 0.002β β +1

The ratio of tip speed is given by Eq. 5:

Rt
λ= (5)
v
Figure 2 gives the curve of C p of different pitch angle values.
And Eq. 6 gives the obtained aerodynamic torque:

1 Cp
Ta = ρπR3 v 2 (6)
2 λ
1424 M. Reddak et al.

Fig. 2 Power coefficient

across λ for different pitch
angle values

2.2 DFIG Model

With simplified notation and by relating the Park reference to the stator field, the
expression of the DFIG voltages and fluxes in dq reference are given by Eq. 7.
⎧
⎪
⎪ v = Rs ids + dϕdtds − ωs ϕqs
⎪ ds
⎨ dϕ
vqs = Rs iqs + dtqs + ωs ϕds
(7)
⎪
⎪ vdr = Rr idr + dϕdtdr − ωr ϕqr
⎪
⎩ v = R i + dϕqr + ω ϕ
qr r qr dt r dr

ϕds = L s i ds + Mi dr = ϕs
(8)
ϕqs = L s i qs + Mi qr = 0

The stator and rotor current are related as follow:


i sd = Lϕss − LMs ir d
(9)
i sq = − LMs irq

The expression of the powers is given in Eq. 10:


Ps = −vs LMs irq
(10)
Q s = vs Lϕss − vs LMs ir d

The equation of the electromagnetic torque is shown in Eq. 11:

M
Tem = − pϕs irq = K T irq (11)
Ls

And the dynamic of the rotor currents is expressed in Eq. 12:

MPPT and Pitch Angle Based on Neural Network Control of Wind Turbine Equipped ... 1425

di dr
dt
= σ 1L r (vr d − Rr ir d + σ L r ωr irq )
dirq (12)
dt
= σ 1L r (vrq − Rr irq − σ L r ωr ir d − ωr LMs ϕs )

3 Neural Network

A neural network is based on the working of biological neurons and materializes

in a computer as an algorithm. The neural network can adjust itself according to
the results of its actions, which tolerates learning and problem solving without an
algorithm and conventional programming.
The neuron mathematical model, in the lth layer, is assumed by the following
equation:


nl
y lj = f ( wlji xi + bil ) (13)
i

With
xi : Outputs of the previous neurons.
wlji : The weight of synapses binding neurons.
bil : Coefficient for adjusting the neural network.
The function f can be presented by hyperbolic, threshold, sigmoid or radial
functions.
This ensemble of interconnected formal neurons allows the resolution of complex
problems by adjusting the weight coefficients using a learning algorithm.
In our case, the ANN bloc, that control the pitch angle, is a static multi-layer
perceptron MLP. The architecture of this later is revealed in Fig. 3.
According to Fig. 3, the neural network takes as input the ratio of the tip speed λ
and C p , and as output the pitch angle β.
The performance training of the proposed neural network is depicted in Fig. 4.

4 Wind Turbine Control Strategy

Figure 5 illustrates the curve of the wind turbine power across the wind speed; there
are three operation modes: In the first region, before the cut-in speed, there is no
production of the power. Then in second region, that matches to the average wind
speed, it is necessary to control the DFIG to benefit from the maximum power. In
the last region, where the wind speed exceeds the passage value which is 12 m/s, it
is necessary to ensure the control of the pitch angle to conserve the rated power.
1426 M. Reddak et al.

Fig. 3 The proposed MLP neural network for pitch angle control

Fig. 4 Performance training of the MLP neural network

Figure 6 shows the wind turbine coupled with the DFIG system and the used
control strategy. The control system is separated to two parts: the control at the
turbine side and the control at the DFIG side.
The first part maintains the power captured at its nominal value by acting on the
pitch angle control in region 3, it also ensures the speed control of the generator
according to the measured speed of the wind by optimizing the power extracted in
region 2 regarding this equation.
MPPT and Pitch Angle Based on Neural Network Control of Wind Turbine Equipped ... 1427

Fig. 5 Operating the wind turbine regions

Fig. 6 Control strategy of the wind turbine driven a DFIG using an ANN

1
Popt = ρπ R 2 v 3 C p_max (λopt , βmin ) (14)
2
Therefore, this part gives the reference torque necessary for the active power
control by generating the rotor current reference irq r e f .
On the side of DFIG control, the converter of the machine-side controls the
generated active and reactive powers.
1428 M. Reddak et al.

Parameters of DFIG and PI controllers are given by [13].

5 Simulation Results

To confirm the control strategy approach of the wind turbine equipped with a DFIG,
a wind profile, described in Eq. 1, has been adopted and potted as depicted in Fig. 7.
Indeed, it can be verified, in Fig. 8, that the pitch angle produced by the artificial
neural network, varies according to this wind profile. It is at minimum value (0°)
when the wind speed is under the transition value (12 m/s). Whereas in region 3,
which reflects the wind speed exceeding the pass through value, the pitch angle
takes the corresponding value other than the minimum one. With these strategies,
the MPPT and pitch control has been achieved.
Also it can be shown in Fig. 9 that the Cp is maintaining in the maximum value,
that is around 0.45, when the wind speed reflects the region 2. Although, the Cp is less
when the wind speed surpasses the transition value, which endorse the robustness of
the control under neural network.
The electromagnetic torque shown in Fig. 10 is limited to protect the system from
damaging and load fatigue while the wind speed is upper to the pass trough value.
Besides, as depicted in Fig. 11, the stator power is limited while the wind speed
is higher than the pass through value to give the nominal value that is 2e6 W, and
when the wind speed reveals the region 2, the regulator forces the DFIG to extract
the maximum power.

Fig. 7 The used wind speed profile

MPPT and Pitch Angle Based on Neural Network Control of Wind Turbine Equipped ... 1429

Fig. 8 Pitch angle given by the neural network

Fig. 9 The variation of the power coefficient

1430 M. Reddak et al.

Fig. 10 The reference and response of the electromagnetic torque

Fig. 11 Electromagnetic power delivered by the DFIG

MPPT and Pitch Angle Based on Neural Network Control of Wind Turbine Equipped ... 1431

6 Conclusion

An intelligent ANN control, which is used for the MPPT and the pitch angle strategy,
has been suggested in this paper, to profit from the power delivered by a wind turbine
equipped by DFIG under full range of wind speed. For the purpose of protecting
the wind turbine from the fatigue load when the speed surpasses the rated value, the
electromagnetic torque has been limited. In this case, the pitch angle, generating using
an ANN control block, and power coefficient are set by producing a corresponding
value. However, when the wind speed is under the pass through value, and the pitch
angle given by ANN is at a minimum value to extract minimum value to ensure
maximum power extracting. The control strategy potency using ANN of wind turbine
equipped by DFIG has been shown in the simulation results under Matlab/Simulink
environment .

Appendix

(Table 1)

Table 1 DFIG parameters

Nomenclature Designation Value
Ps Stator power 2e6 W
Vs Stator voltage 690 V
Rs Stator resistor 2.6e–3 
Rr Rotor resistor 2.9e–3 
Lm Mutual inductance 2.5 mH
Ls Stator inductance 2.587 mH
Lr Rotor inductance 2.587 mH
1432 M. Reddak et al.

References

1. Tripathi SM, Tiwari AN, et Singh D (2016) Optimum design of proportional-integral controllers
in grid-integrated PMSG-based wind energy conversion system: optimum design of PI
controllers. Int Trans Electr Energy Syst 26(5):1006−1031
2. Singh M, Khadkikar V, Chandra A (2011) Grid synchronisation with harmonics and reactive
power compensation capability of a permanent magnet synchronous generator-based variable
speed wind energy conversion system. IET Power Electron 4(1):122. https://doi.org/10.1049/
iet-pel.2009.0132
3. Bekakra Y, Attous DB (2014) Optimal tuning of PI controller using PSO optimization for
indirect power control for DFIG based wind turbine with MPPT. Int J Syst Assur Eng Manag
5(3):219−229
4. Boubzizi S, Abid H, El hajjaji A, Chaabane M (2018) Comparative study of three types of
controllers for DFIG in wind energy conversion system. Prot Control Mod Power Syst 3(1):21
5. Reddak M, Berdai A, Nouaiti A, Vlasenko V (2018) Collaboration of nonlinear control strategy
and pitch angle control of DFIG equipped wind turbine during all operating regions. Int J
Comput Appl 179(25):16−21
6. Reddak M, Berdai A, Gourma A, Boukherouaa J, Belfiqih A (2016) Enhanced sliding mode
MPPT and power control for wind turbine systems driven DFIG (doubly-fed induction
generator). Int Rev Autom Control IREACO 9(4):207
7. Karimi-Davijani H, Sheikholeslami A, Livani H, Karimi-Davijani M (2009) Fuzzy logic control
of doubly fed induction generator wind turbine. p 11
8. Zhang Z-Y, Wang K-S (2014) Wind turbine fault detection based on SCADA data analysis
using ANN. Adv Manuf 2(1):70−78
9. Fazelpour F, Tarashkar N, Rosen MA (2016) Short-term wind speed forecasting using artificial
neural networks for Tehran. Iran Int J Energy Environ Eng 7(4):377−390
10. Marugán AP, Márquez FPG, Perez JMP, Ruiz-Hernández D (2018) A survey of artificial neural
network in wind energy systems. Appl Energy 228:1822−1836
11. Abdelli A, Sareni B, Roboam X (2007) Optimization of a small passive wind turbine generator
with multiobjective genetic algorithms. Int J Appl Electromagn Mech 26(3−4):175−182
12. Mirecki A, Roboam X, Richardeau F (2007) Architecture complexity and energy efficiency of
small wind turbines. IEEE Trans Ind Electron 54(1):660−670
13. Abad G, Iwanski G (2014) Properties and control of a doubly fed induction machine. In Power
electronics for renewable energy systems, transportation and industrial applications, H. Abu-
Rub, M. Malinowski, et K. Al-Haddad, Éd. Chichester, UK: John Wiley & Sons, Ltd, pp
270−318
Cairo Metro Power Analysis
and Harmonic Mitigation

Ahmed Abd Elsadek, O. A.Monem, Zaki Matar, Mokhtar Hussien,

and Yasser Elsayed

Abstract In Egypt, Traction became one of the most important transportation means
so many electrical projects related to the metro have been established. Metro electrical
network is supplied by 12-pulse converter rectifier substations. These rectifier stations
are of course nonlinear loads which result in several orders of harmonics to the
electrical grid. Moreover, some problems occurred due to the negative effect of
harmonics, for example, cables and transformers insulation failure due to fatigue
resulted from the high frequency voltage harmonic content.
This paper proposes a passive filter for this harmonic reduction. IEEE519-2014
recommendation sets the limits of current harmonics, and the resulted voltage distor-
tion at the point of common coupling. This work uses the measured current and
voltage for estimating the significant harmonic orders. A suitable filter can be
designed, and implement according to these measured harmonics. The effect of trans-
former impedance and changing the configuration of motor traction from DC series
motor into AC induction motor with speed drive is illustrated in the analysis section.
To overcome the harmonic distortion problem and to define the proper mitigation
method, recommendations have been concluded.

Keywords Point of common coupling · Traction · Cairo metro · Total harmonic

distortion · Nonlinear loads · 99th percentile · Rectifier station · High voltage
station

A. Abd Elsadek (B)

Egyptian Company for Cairo Metro Management and Operation, Cairo, Egypt
O. A.Monem
Higher Institute of Engineering and Technology Alobour,
Cairo-Belbeis Desert Road, Cairo, Egypt
Z. Matar · Y. Elsayed
Electrical Power Department, ALAZhar University, Cairo, Egypt
e-mail: [email protected]
M. Hussien
North Cairo of Electrical Distribution Company, Cairo, Egypt

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1433
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_130
1434 A. A. Elsadek et al.

1 Introduction

Rectifier substations (RS’s) are employed to convert AC to DC voltage to metro trains.

The twelve-pulse converter has a non-linearity nature. This non-linearity distorts the
drawn rectifier station source current. Distorted source current will have high levels
of harmonic contents. These harmonics have a negative impact in different ways; like
transformer heating, error in metering, false tripping of breakers, insulation damage,
etc. Due to these problems, the need to assess and evaluate the harmonic level in
Egypt metro network arises.
The applicable recommended standard IEEE 519-2014 is the most common in
power quality analysis [1]. Different researches use it to evaluate the power quality
[2–4].

2 System Configurations

Cairo metro network is composed of two lines, line 1 and line 2, two high voltage
stations (HVS) that feed the two lines. HVS of Tura El Ballad 66/20 kV that supplies
17 rectifier stations in line 1. The second HVS is at Ramses 220 kV /20 kV which
feeds the rest of line 1 rectifier stations, and all rectifier stations of line 2.
In line 1, RS consists of three parts, two rectifier groups, each group has one power
transformer with voltage 20 kV/1228 v and 2.6 MW connected ϒ/ − ϒ to achieve
phase shift of 30° to cancel 5th and 7th harmonic order from source current. The
third transformer is an auxiliary for feeding the rectifier lighting loads. The output
DC voltage is 1500 V.
On the other hand, the RS’s in line 2 have the same configuration, however, it
contains one traction group, and the output dc voltage is 750 vdc. For line 1, there
are two types of traction motors; the first one is induction motor which was working
since 2015 and the second oldest one is DC series motors. Figure 1 describes the
rectifier station configuration which represents the building block of metro electrical.

3 Measurements and Results

The measurements were carried out using the power quality analyzer, Hioki 3196
used for recording source current, and voltage waveforms, harmonic content, power
factor, and other power quality indices which are required for harmonic assessment.
All data are collected and analyzed using statistical analysis to consider the full
period of measurements in terms of 99th percentile and maximum values according
to the recommendation of IEEE 519-2014 [5].
The first location of measurement is point “A” considering one group of traction
loads. The second measurement has been taken aFt location “B” to consider the two
Cairo Metro Power Analysis and Harmonic Mitigation 1435

Fig. 1 Configuration system

groups of the substation and the auxiliary transformer which supply the substation
loads (Lighting, air condition, etc.).
The third measurement location is at point “C” which considers the effect of the
cable length. On the other hand, the last measurement location is at point “D” which
considers groups of rectifier substations.
Measurements of point “B” are indicated in the following figures as a sample.
Figure 2 indicates source voltage and current waveforms. The instantaneous value
of voltage THDv and THDi versus time for 24 h period compared to the predefined
limit are described in Fig. 3, and Fig. 4 respectively showing actual value exceeds
the standard limit.

4 Measurement Summary

Table 1 illustrates a summary of the power quality indices measured for THDv, and
THDi at different locations of points A, B, C, and D.
• It is noticed that the distortion rate increases in case of an increase of load current,
conversely at the no-load condition the voltage distortion is under the critical
1436 A. A. Elsadek et al.

Fig. 2 Source current and voltage waveform of point (B)

Fig. 3 Measured THDv of point (B)

Cairo Metro Power Analysis and Harmonic Mitigation 1437

Fig. 4 Measured THDi of point (B)

Table 1 Measurement results of THDV and THDi

limit, a combination between the non-load period and the start of moving the first
train indicates the difference in Fig. 5.
• The voltage individuals of orders 11th , 13th appeared at all measured points
as a result of the behavior of 12-pulse rectifiers in load condition, it’s called
characteristic harmonics.
• Despite of these results, an old study related to measurement analysis attempt
only in line 1 of the metro network in 2010. It is showed that the current distortion
is slightly out of specs, and consequently, it doesn’t have an effect on voltage
distortion to exceed the limits. By comparing both analyses, resulted that the
reason for increasing the harmonic distortion is related to adding new trains in
2015 to the operating system. These trains have induction motors driven by PWM
drives, while the oldest trains have dc series motors controlled by resistances, also
1438 A. A. Elsadek et al.

Fig. 5 Instantaneous values of THDv at the first hour of operating

changing impedance, power of the new AC trains are important factors for high
harmonic value results.
• The distortion of line two is much more than line one, this is because of increasing
the value of the impedance of the traction transformer of line two than line one
[6], moreover, all trains are AC induction machine, and driven by PWM inverters.
• Common feeding of line 1 and line 2 through the point (D) has major effecting
on each other through a distorted voltage supplied. Up till modernizing line 1 it
is recommended to separate loads of line 2 on HVS Ramses from loads of line 1
to decrease the voltage distortion effect, moreover design proper filter on AC side
to line 2.
• Auxiliary transformers of 20 kv/400 v, voltage transformers (VT), and current
transformers (CT) are affected rapidly more than traction transformers 20 kv
/1228 v, which may be due to different values of K-factor.
• Slightly impact of auxiliary transformer this is due to the light consumed loads,
although having significant non-linear loads such as (led lamps, chargers and
inverters …), this result got by connecting and disconnecting the auxiliary
transformer on service at operating period and after the non-traction period.
• Reference to system configuration, it’s noticed that the resonant dc filter has a
slight effect as the measurements on AC side, maybe it has an effective rule on
dc side, this result was according to measuring at some (RSs) that have a filter on
dc side and the others haven’t a filter.
• It is noted that sometimes the transformers overloaded with about 1.5 of its
primary rate, due to changing some of the old trains with higher power air-
conditioned trains without increasing the power of the rectifier stations, and
headway increasing.
• The harmonic currents associated with high losses and temperature will decrease
the expected lifetime of transformers. Recommendation of resizing substation
supplied power transformer depends on many factors included in reference [7].
• In 2019, a new RS was put in service according to the extension plan of line 1;
the dc filter needs to be checked according to proper design for new trains.
Cairo Metro Power Analysis and Harmonic Mitigation 1439

Fig. 6 PF measured at point D

• Also, the presence of audible noise in transformers is due to high frequencies

current harmonics, it is an important indication of losses [8].
• Note that some of the individual’s currents orders decreased gradually from
measured point “A” up to location “D”, it has high values, but it’s under the
limit, maybe due to shifting cancellation and the effect of impedance of cable and
network [9, 10].
• Due to harmonic distortion, poor power factor results at point D ranged between
0.66 to 0.88 according to the load current values, hence has high reactive power
due to inductive load as shown in Fig. 6.
• Practical measurements on 220 kV are not permissible, but it is expected that
more voltage and current distortion and decreasing of power factor based on the
following equation:

Vpcc = Vs − Ls dls /dt (1)

Where Vpcc : the voltage at the point of common coupling,

Vs : the supply voltage,
Ls : the source inductance and (Is ) is the source current.

The harmonic current components multiplied by the source impedance at this

frequency represent a voltage drop and cause a voltage distortion for the supply
voltage affecting other consumers at the point of common coupling [11].

5 Filter Design and a Comparative Study

As per the violation identified by the recorded current THDI , the voltage at the point
of common coupling THDv, distortion factor, and total power factor, the need to
mitigate the harmonics is arising. Few of the most prevailing methods used today
1440 A. A. Elsadek et al.

to reduce harmonics for example (Delta-Delta and Delta-Wye Transformers, Multi-

pulse converter (12-pulse, 18-pulse, and pulse multiplication 36-pulse), Isolation
Transformers, Use of Reactors, Passive Harmonic Filters (or Line Harmonic Filters),
and Active filters. The selection of harmonic mitigation method for a specific appli-
cation is an important decision, especially from cost point of view. In Research [12]
co-phase traction transformer decreases the current distortion from 38.3 to 4.7%,
although this is an effective solution, it is not easy to change 32 power transformers
along the line from a cost and reliability point of view. On the other hand, using multi
pulse techniques of 18 pulses as mentioned in [13] to decrease the voltage distortion
from 8.9 to 6% a has high cost and is not applicable for unbalanced loads. Passive
filters are cheaper, simple, and reliable. Moreover, it supports tackle specific orders
harmonic, and plug and play.

5.1 Passive Filter Design

A passive filter is combined of capacitors and inductors that are tuned to resonate at
a single frequency, or through a band of frequencies.
Passive filters work by exhibiting different impedance values at the resonant
frequency. A filter connected in series should present high impedance to the harmonic
frequency that needs to be blocked. Although a series configuration is possible, it
is more common to connect filters in parallel. Such a shunt configuration diverts
harmonic currents to ground, and simultaneously provide reactive power, which
may be used to correct the power factor. As such, passive shunt filters are designed
to be capacitive at the fundamental frequency.
Passive filter design is conducted to mitigate 11th and 13th harmonic order [4, 14,
15]. According to IEEE design recommendation, filter design will be 10.5th instead
of 11th order, and 12.5th instead of 13th order [16].

– Single Tuned for the 11th Order

The value of filter components can be evaluated through the equations from (2)
to (7) as the following:

V2 (KV)
Xeff = (2)
Q rate of filter(MVAR)

(20)2
Xeff = = 31746.03 
0.0126
  
XC = h2 / h2 − 1 ∗ Xeff (3)

 
10.52
XC = ∗ 31746.03 = 32036.6 
10.52 − 1
Cairo Metro Power Analysis and Harmonic Mitigation 1441

1
C11 = (4)
2Π f XC
1
C11 = = 0.099 μF
2π ∗ 50 ∗ 32036.6
Where Xeff : effective reactance of harmonic filter
Qeff : effective reactive power of harmonic filter, this value from the measure-
ment
VLL : System line to line voltage
XC : Capacitive reactance of harmonic filter capacitor at f tuned
XL : Capacitive reactance of harmonic filter reactor at f tuned
h: harmonic number.
Xc 32036.6
XL = = = 290.58 (5)
h2 (10.5)2
XL 290
L11 = = = 924 mH (6)
2πF 2π ∗ 50
The resistance of filter value (R) is calculated according to the following equation
by assuming the quality factor Q is 50 as in [11]

XL
R = (7)
Q
290
R= = 5.8 
50

– Single Tuned for the 13th Order

Similarly, as the single tuned 11th equations from Eqs. (2) to (7), the final values
of the 13th order filter can be designed as the following:

C13 = 0.099 μF L13 = 651 mH R13 = 4.1 

5.2 PQI’s Improvement Achieved by Inserting Passive Filter

• Some individual currents exceed the limits; moreover, it does not affect its voltage
distortion, so there is no need for a high pass filter to eliminate from 23th to 50th
as the most significant effect of current is at 11th and 13th as in Eqs. (8), and (9).

( (I2)2 +(I3)2 + (I4)2 + . . . . . . . . . . . . In2 .
THDi = (8)
I1
1442 A. A. Elsadek et al.

THDi at point (B) is 78% and by eliminating the value of 11th and 13th of individual
current, the new value of THDi = 19% as considering filter compensation

( (V2)2 +(V3)2 + (V4)2 + . . . . . . . . . . . . Vn2 .
THDv = (9)
V1
THDv at point (B) is 8.2% and by eliminating the value of 11th and 13th
of individual voltage, the new value of THDv = .07% as considering filter
compensation.
• The optimum filter location and best area space is at point (B) of the incoming of
the rectifier substation.

6 Conclusion

As the nature of time-varying of traction loads, and geographical location (from

a station to another according to the population), the drawn load current and the
associated harmonic currents amplitude and phase angle will be different accordingly.
This load variation nature of traction is the reason behind the unsuitability of passive
filter to mitigate the harmonics from the system in this load type. This is a clear
drawback that can be added to the conventional drawbacks of passive filters. In
the case of no-load, it will be a source of reactive power. On the other hand, the
pulse multiplication technique is difficult to be applied due to the reliability factor
which can’t be overcome due to the importance of the metro as sensitive and critical
transportation means in Egypt that transports millions of people per day.

7 Future Work

Reference to the previous argument, the recommended solution for metro harmonic
mitigation is the active filter. The active filter has the advantage of not only mitigate
the harmonics but also compensates the reactive power to improve both distortion
factor and displacement power factor. The active filter becomes a common industrial
solution; however, the design of it is a critical job. The location and the rating of
the active filter should be carefully selected to achieve the mitigation target to meet
the international standard with better harmonics handling performance and overall
system efficiency. This will be discussed in future work as a part of the metro continual
improvement strategy.
Cairo Metro Power Analysis and Harmonic Mitigation 1443

References

1. Smith JC, Hensley G, Ray L (2009) IEEE recommended practice for monitoring electric power
quality. IEEE Std 1159–2009 (Revision of IEEE Std 1159–1995), pp c1–81
2. Gunavardhini N, Chandrasekaran M, Thirumoorthy A (2012) A study on harmonics in Indian
railway traction. IOSR Journal of Electrical and Electronics Engineering 2:1–4
3. Navaneethakrishnan G, Muthial C (2017) Power quality in railway traction and compensation
by combining shunt hybrid filter and TCR. Automatika 57:610–616
4. Popescu M, Bitoleanu A, Dobriceanu M (2008) Harmonic current reduction in railway systems.
WSEAS Transactions on Systems 7:689–698
5. Sabin DD (2002) Analysis of harmonic measurement data. IEEE power engineering society
summer meeting 2:941–945
6. Nhavira S, Polimac V (2007) Quality of supply issues arising from DC traction loads on a
metro system
7. EN 50329 (2003) Railway applications – Fixed installations – Traction transformers
8. Lukic L, Djapic M, Lukic D, Petrovic A (2001) Aspects of design of power transformers for
noise reduction. Parameters 60076:10
9. Elhenawy, AAE-M., Sayed, MM, Gilany MI (2018) Harmonic cancellation in residential build-
ings. In: 2018 twentieth international middle east power systems conference (MEPCON), pp
346–351
10. Zhongming Y, Pong M, Lo W, Yuen K (1999) Harmonic evaluation of traction system by Monte
Carlo simulation. In: 30th annual IEEE power electronics specialists conference. Record (Cat.
No. 99CH36321), pp 424–429
11. Monem O (2019) Harmonic mitigation for power rectifier using passive filter combination. In:
IOP conference series: materials science and engineering, p 012013
12. Fathima F, Karthikeyan SP (2016) Harmonic analysis on various traction transformers in co-
phase traction system. Ain Shams Eng J 7:627–638
13. Persson J (2014) Comparing harmonics mitigation techniques. In: Comsys AB, Lund, p 6
14. Vasanthi V, Ashok S (2011) Harmonic filter for electric traction system. In: ISGT2011-India
15. Shah N (2013) Harmonics in power systems causes, effects and control. Whitepaper design
engineering low-voltage drives
16. IEEE std 1531 (2003) IEEE guide for application and specifications of harmonic filter
Dual Fuzzy Direct Power Control
for Doubly-Fed Induction Generator:
Wind Power System Application

Hala Alami Aroussi, El Mostafa Ziani, Manale Bouderbala,

and Badre Bossoufi

Abstract In this work, a control strategy based on fuzzy logic controllers is devel-
oped for the control of an asynchronous machine. The proposed control consists of
active and reactive powers estimation, two fuzzy controllers, and tracking based on
the theory of direct torque control. The originality of our work has been in the imple-
mentation of the different switching states in different fuzzy controllers by rules that
describe the operation of each magnitude controlled, ensured by the simplicity of
implementation and the robustness offered by the fuzzy logic.
As an approach, a fuzzy controller will be studied and implemented for the purpose
of improving the static and dynamic performance of a dual classical direct power
control applied to a wind power system. In this context, hysteresis controllers will
be replaced by two fuzzy controllers and the simulation results of the generator-side
and grid-side controls will be presented to prove the e of the control.

Keywords DTC · Dual DPC · Fuzzy logic · Asynchronous machine · Hysteresis

controllers

Nomenclature

Symbols Parameters
Vsα , Vsβ Stator voltages in (α, β) reference
Vgα , Vgβ Grid voltages in (α, β) reference
i sα , i sβ Stator currents in (α, β) reference
ir α , irβ Rotor currents in (α, β) reference

H. Alami Aroussi (B) · E. M. Ziani

Laboratory of Electrical and Maintenance Engineering (LGEM), Ecole Supérieure de
Technologie, Mohammed Premier University, Oujda, Morocco
e-mail: [email protected]
M. Bouderbala · B. Bossoufi
LISTA Laboratory, Faculty of Sciences Dhar El Mahraz, Sidi Mohammed Ben Abdellah
University, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1445
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_131
1446 H. Alami Aroussi et al.

i gα , i gβ Grid currents in (α, β) reference

ψsα , ψsβ Stator fluxes in (α, β) reference
ψr α , ψrβ Rotor fluxes in (α, β) reference
ωm Rotating speed
Ps , Q s Stator active & reactive powers
Pg , Q g Grid active & reactive powers
S Surface swept by the pales
ρ Air density
v Wind speed
C p (λ, β) Power coefficient

1 Introduction

In many countries, wind power has become a pillar of their energy systems since it
is a reliable and inexpensive source of electricity.
Indeed, the advancement of using clean energies in recent years lead researchers
to develop wind energy controllers to exploit the maximum of wind energy extracted
and to ensure its quality.
Nowadays, the wind power system (WPS) based on a doubly-fed induction gener-
ator (DFIG) is the most implemented configuration. The latter has numerous advan-
tages that can be summarized mainly in the fact that the machine can drive under
variable speed, operates in both sub and hyper-synchronism modes, offers better
energy capture, reduces mechanical stress, and costs fewer thanks to the emerged
power converters [1–3].
The schematic of the wind power system configuration that we will use in the
simulation part is illustrated in Fig. 1. The DC-Link is linked to the back-to-back
converter. The rotor side converter (RSC) is connected to the rotor to control the
stator active and reactive powers while the grid side converter (GSC) is connected to
the grid to maintain the voltage of the DC-Link fixed [4].
As a first step, we propose a model of the wind power system, described before,
including the dynamic model of both the machine and the wind turbine. Then, we
choose the control that will be applied to run the whole system and will generate the
results displayed using the software Matlab/Simulink.
In our case, many techniques were developed by researchers, considering that
the system is linear or not, such as Field Oriented Control, Direct Power Control,
Backstepping, Sliding mode, Predictive Control…
Indeed, Vector Control is the most implemented one in wind farms. However, its
main drawback is the fact that this control is sensible to parameters variations of
the system such as resistance and inductance variations [5]. To overcome this, the
Direct Power Control, which is initially based on the Direct Torque Control, offers
direct control of the powers using a predefined lookup table based on estimation and
a perfect decoupling between reactive and active powers [6]. Though, the non-linear
Dual Fuzzy Direct Power Control for Doubly-Fed Induction Generator … 1447

Fig. 1 Dual DPC applied to DFIG-WPS synopsis

behaviours of the WPS parts and the non-constant switching frequency lead us to
use non-linear techniques such as the Fussy Logic.
As a result, the major emphasis of the work is targeted at the application of Fuzzy
logic controllers as the replacement of the hysteresis controllers in the direct power
control scheme (Control level 1). The dual direct power control applied to the back-
to-back converter is carried under variable speed and the results are displayed through
simulations.

2 Dynamic Model of the System

The dynamic model of the wind power system in the reference αβ can be expressed
by [7]:

2.1 Stator and Rotor Voltages

The voltages of the doubly-fed induction machine, in space vector notation, are
written as:
1448 H. Alami Aroussi et al.

−
→ −
→ d ψ s Vsα = Rs i sα + dψsα
Vs = Rs i s + ⇒ dt
dψsβ (1)
dt
Vsβ = Rs i sβ + dt

−
→ −
→ d ψr −
→ Vr α = Rr ir α + dψr α
+ ωm ψrβ
Vr = Rr ir + − jωm ψr ⇒ dt
dψrβ (2)
dt
Vrβ = Rr irβ + dt
− ωm ψr α

2.2 Stator and Grid Powers

The stator and grid powers (active and reactive) are given by the following equations
[4]:
−→ − →
Ps = 23 Re{Vs , i s∗ } ⇒ Ps = 23 (Vsα i sα + Vsβ i sβ ) (3)

−→ − →
Q s = 23 I m {Vs , i s∗ } ⇒ Q s = 23 (Vsβ i sα − Vsα i sβ ) (4)

−→ − →
Pg = 23 Re{Vg , i g∗ } ⇒ Pg = 23 (Vgα i gα + Vgβ i gβ ) (5)

−→ − →
Q g = 23 I m {Vg , i g∗ } ⇒ Q g = 23 (Vgβ i gα − Vgα i gβ ) (6)

2.3 Aerodynamic Power

Theoretical wind power corresponding to a random wind crossing a surface S without

a decrease in speed v is [8, 9]:

1
Ptheor = .ρ.S.v 3 (7)
2
Though, the aerodynamic power appearing on the shaft of the turbine is expressed
as:
1
Paeor = C p .Ptheor = .ρ.S.C p (λ, β).v 3 (8)
2
The power coefficient C p (λ, β) will characterize the aerodynamic efficiency of
the turbine.
The previous equations will help us to elaborate a model of the main parts of our
wind power system. In the next task, we will explain the direct power control and
how to apply it to the machine part as well to the grid part.
Dual Fuzzy Direct Power Control for Doubly-Fed Induction Generator … 1449

3 Dual Fuzzy Direct Power Control

The purpose of this technique (see Fig. 2) is to control directly the powers of our
system by applying different voltages to the converters. These vectors determine
their states [10, 11].
The two magnitudes controlled are the powers that are generally controlled by
hysteresis regulators.
The principle is to keep the magnitudes of the powers within these bands of
hysteresis. The performance of these comparators defines the optimum voltage to be
injected into the inverter for every switch [6].
The use of this type of regulator assumes the existence of a variable switching
frequency that needs a very minor calculation step. Furthermore, it is preferable
to operate with high frequency to decrease oscillations generated by this kind of
comparator.
Assuming that, in classic DPC, two types of power controllers are used: hysteresis
controllers that have a property based on the command and its history or strip
controllers where the command history does not play any role. They are based on
the number of power error bands [12].
In this approach, we will replace the classic regulators with two fuzzy controllers.
Here are all the rules describing the operating principle for this algorithm [13–15]:
⎧
⎪
⎪ If eQ s P then HQ I ncr ease
⎪
⎪
⎪
⎨ f
I eQ s N then HQ Decr ease
If e Ps P then H P I ncr ease (9)
⎪
⎪
⎪If
⎪ e Ps Z then H P is kept
⎪
⎩If e Ps N then H P Decr ease

These fuzzy controllers (see Fig. 2) have the following properties:

Fig. 2 Synopsis of the fuzzy DPC applied to DFIG

1450 H. Alami Aroussi et al.

• The fuzzy controller is a Sugeno type.

• Inputs fuzzification e Ps (stator active power error), eQ s (stator reactive power
error)
• The control strategy depends mainly on inferences (rules). The inference is based
on Mamdani’s method that is associated with the min-max decision. To simplify
the description of inferences, we use an inference table.
• For defuzzification, the center of gravity is the method used.
The same control will be applied to the grid side converter by changing the stator
powers by grid powers.

4 Results and Simulations

In this part, the dual fuzzy direct power control applied to a wind power system of
1.5 kW was tested under the Matlab/Simulink environment with the presence of a
balanced three-phase load (Rload ) of 300 W.
The system, shown in Fig. 1, was driven under subsynchronous and hypersyn-
chronous speed (steady and transient state) regarding the following conditions:
• The generator operates at variable speed (see wind profile);
• The DC bus voltage: Vbus = 600 V.
The simulation results of the power fuzzy logic applied to dual DPC are given by
the following figures (Fig. 3):
Under the circumstances mentioned previously, we observe that the stator active
power follows its reference. It has a very fast dynamic and fewer harmonics.
Indeed, the figure above shows a very good response of the stator active and
reactive powers (around 0.28 s), where these follow their references perfectly with
almost zero static error except for the peak that appears at 3.9 s.
On the other hand, the grid active power presents a clear distortion beginning at t
= 3 s, while the grid reactive power transient is almost not seen due to a reasonably
decoupled control performance.
It should also be remembered that the reactive power setpoint is imposed zero to
obtain a unit power factor on both rotor and grid sides.
Also, the patterns of the stator, rotor, and grid currents show small harmonic rates.
These currents show also fewer ripples but keep their sinusoidal shapes.
Figure 4 shows the spectrums of the stator (2.05%), rotor (3.43%), and grid
(9.01%) currents. Notice that the Total Harmonic Distortion (THD) of the currents
is less than 10% recommended by the standards IEC 61727 and IEEE 1547.
Dual Fuzzy Direct Power Control for Doubly-Fed Induction Generator … 1451

Wind Profile DC Voltage

800
Vbus
12
Vbusref

10 600
8

Vbus (V)
V (m/s)

6 400

4
200
2
0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s) Time (s)

(a) (b)
Stator Currents Zoom Stator Currents

Zoom Stator Currents (A)

4 3
Stator Currents (A)

2
2
1

0 0

-1
-2
-2

-4 -3
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 3.7 3.71 3.72 3.73 3.74 3.75 3.76 3.77 3.78 3.79 3.8
Time (s) Time (s)
(c)
Rotor Currents Zoom Rotor Currents
Zoom Rotor Currents (A)

6 6
Rotor Currents (A)

4 4
2 2
0 0
-2 -2
-4 -4
-6 -6
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 2.6 2.8 3 3.2 3.4 3.6 3.8 4 4.2
Time (s) Time (s)

(d)
Stator Active Power Stator Reactive Power
Stator Reactive Power (Var)

1000 2000
Stator Active Power (W)

Ps Qs
Psref Qsref
500 X: 0.2862
Y: -68.77 -1000 1000
0 -1200
X: 0.2998
Y: 0
-1400
-500 0
3.8 4 4.2
-1000
-1000
-1500

-2000 -2000
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s) Time (s)
(e) (f)

Fig. 3 Simulation results: (a) wind profile (b) DC voltage (c) abc stator currents (d) abc rotor
currents (e) stator active power (f) stator reactive power (g) abc grid currents (h) grid active power
(i) grid reactive power
1452 H. Alami Aroussi et al.

Grid Currents Zoom Grid Currents

10 5

Zoom Grid Currents (A)

Grid Currents (A)

0 0

-5

-10 -5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 3.7 3.71 3.72 3.73 3.74 3.75 3.76 3.77 3.78 3.79 3.8
Time (s) Time (s)

(g)
Grid Active Power Grid Reactive Power
1500

Grid Reactive Power (Var)

4000
Grid Active Power (W)

Pg Qg
Pgref Qgref
3000 1000

2000 500
1000
0
0
-500
-1000
-2000 -1000
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s) Time (s)

(h) (i)

Fig. 3 (continued)

a b c
Fundamental (50Hz) = 2.129 , THD= 2.05% Fundamental (5Hz) = 7.76 , THD= 3.43% Fundamental (50Hz) = 4.333 , THD= 9.01%
Isa Ira 6 Iga
0.4 5
5
4
0.3 4
3
0.2 3
2
2
0.1
1 1

0 0 0
0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000
Frequency (Hz) Frequency (Hz) Frequency (Hz)

Fig. 4 THD: (a) stator current Isa (b) rotor current Ira (c) grid current Iga

5 Conclusion

The work performed is a digital simulation of the direct control of the powers of a
wind power system based on the doubly-fed induction generator. We have proposed
a method to improve the dynamic performance of direct power control by intro-
ducing a fuzzy type regulator while detailing its main modules such as Fuzzification,
Inference, and Deffuzification.
Dual Fuzzy Direct Power Control for Doubly-Fed Induction Generator … 1453

Table 1 DFIG & WT parameters

DFIG parameters Value WT parameters Value
Nominal power Pn = 1.5 KW Rated power 1.5 KW
Stator voltage Vs = 380V No. of blades 3
Stator resistance Rs = 1.18  Radius R=1m
Stator leakage Ls = 0.4 H Gearbox G=2
Rotor resistance Rr = 1.66  Inertia J = 1000 Kg.m2
Rotor leakage Lr = 0.18 H Friction f = 0.007 N.m.s−1
No. of poles P=2

The simulation results on the machine and grid sides showed acceptable robustness
of these regulators, especially from the point of view of decoupling the active and
reactive powers, and the speed variation. The next work will be a comparison between
dual direct power control using a different type of comparators and its validation using
a testbench.

Appendix

(See Table 1).

References

1. Stiebler M (2008) Wind energy systems for electric power generation. Springer
2. Wu B, Lang Y, Zargari N, Kouro S (2011) Power conversion and control of wind energy
systems. Wiley, Hoboken
3. Derbel N, Zhu Q (eds) (2019) Modeling, identification and control methods in renewable energy
systems
4. Abad G, Lopez J, Rodríguez MÁ, Marroyo L, Iwanski G (2011) Doubly fed induction machine:
modeling and control for wind energy generation. Wiley-IEEE Press
5. Boubzizi S, Abid H, El Hajjaji A et al (2018) Comparative study of three types of controllers
for DFIG in wind energy conversion system. Prot Control Mod Power Syst 3:21
6. Alami Aroussi H, Ziani EM, Bouderbala M, Bossoufi B (2020) Improvement of the direct
torque control of the doubly fed induction motor under variable speed. Int J Power Electron
Drive Syst 11(1):97–106
7. Dehong X, Frede B, Wenjie C, Nan Z (2018) Advanced control of doubly fed induction
generator for wind power systems. The Institute of Electrical and Electronics Engineers. Wiley
8. Aissaoui AG, Tahour A (eds) (2016) Wind turbines - design, control and applications.
IntechOpen, no 4133
9. Bouderbala M, Bossoufi B, Alami Aroussi H, Taoussi M, Lagrioui A, Livinti P (2020) Dead-
beat control applied to wind power system. In: Proceedings the 5th international confer-
ence on renewable energy in developing countries, IEEE-REDEC’2020, March 24–26, 2020,
Marrakech, Morroco
1454 H. Alami Aroussi et al.

10. Alami Aroussi H, Ziani EM, Bossoufi B, Bouderbala M (2020) DPC & DNPC applied to wind
energy converter system. In: Proceedings of the 5th international conference on renewable
energy in developing countries, IEEE-REDEC’2020, March 24–26, 2020, Marrakech, Morroco
11. Lee Y-S, Martin HL (2018) Chow in power electronics handbook, 4th edn. Elsevier
12. Bossoufi B, Alami Aroussi H, Bouderbala M (2020) Direct power control of wind power
systems based on DFIG-generator (WECS). In: Proceedings the 12th international conference
on electronics, computers and artificial intelligence, IEEE-ECAI’2020, June 25–June 27, 2020
13. Sudheer H, Kodad SF, Sarvesh B (2018) Improvements in direct torque control of induction
motor for wide range of speed operation using fuzzy logic. J Electr Syst Inf Technol 5(3):813–
828
14. Chikouche TM, Hartani K, Bouzar S, Bouarfa B (2020) New direct power control based on
fuzzy logic for three-phase PWM rectifier. In: Hatti M (ed) Smart energy empowerment in
smart and resilient cities. ICAIRES 2019. Lecture Notes in Networks and Systems, vol 102.
Springer
15. Aissa O, Moulahoum S, Kabache N, Houassine H (2014) Fuzzy logic based direct power control
for PWM three-phase rectifier. In: 22nd mediterranean conference on control and automation,
Palermo, pp 79–84
DFIG-WECS Non Linear Power
Controls Application

Manale Bouderbala, Badre Bossoufi, Hala Alami Aroussi,

Mohammed Taoussi, and Ahmed Lagriou

Abstract This study proposes two different controlling approaches of Wind Energy
Conversion System (WECS). The first one deals with Field/Oriented/Control/ (FOC)
and the second with Direct Power Control (DPC). Both approaches were applied on
Doubly-Fed-Induction-Generator-(DFIG).
DPC seems to give good tracking and better efficiency compare to FOC. In fact,
by comparing Stator active and reactive powers, DPC gives much less oscillations.
Regarding rotor and stator currents, DPC needs less time to achieve steady-state.

Keywords DFIG · FOC · DPC · WECS

1 Introduction

The field of electric machine control has known a great evolution in industrial sector.
Used for several centuries (windmills), wind energy has been developed considerably.
This significant development of this energy is going to be essential to attain the aims
set for renewable energies. A wind system is usually composed of turbine, gearbox,
and generator. There are two varieties of wind turbines: vertical axis wind turbines
and horizontal. The latter is the most used because of its great efficiency. Regarding
the generator, DFIG is often used thanks to its several advantages: the rotor power
reaches around 30% of the total power. As result, the dimensions of converters among
the conversion system is reduced [1].

M. Bouderbala (B) · B. Bossoufi · M. Taoussi

LISTA Laboratory, Faculty of Sciences Dhar El Mahraz, Sidi Mohammed Ben Abdellah
University, Fez, Morocco
e-mail: [email protected]
H. Alami Aroussi
Laboratory of Electrical and Maintenance Engineering (LGEM), Ecole Supérieur de Technologie,
Mohamed Premier University, Oujda, Morocco
A. Lagriou
Department of Electrical and Computer Engineering, The Higher National School of Arts and
Trades, Moulay Ismail University, Meknes, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1455
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_132
1456 M. Bouderbala et al.

For many years, power control has been the topic of many researches. Many
control strategies have been proposed. Among these strategies, “vector controls”
gives dynamic performance identical to that acquired by the-direct-current (DC)
machine. Nowadays, the improvement of modern digital signal processing strategies
has prompted extensively, to give a rise to “Direct Control of the Powers” [2].
Blaschke was the first one to develop the primary theory of the field-oriented
techniques in 1970. Ten years later, its effective applications were adopted. Indeed,
it requires calculations of Park transform, evaluation of trigonometric functions and
regulations, which could not be done in an analogical approach. Vector control
is designed to provide the same control performance as a separately excited DC
machine, which is characterized by a natural decoupling of the variable flux and
torque [3].
In the middle of the 1980s, TAKASHI initiated the Direct control methods of
asynchronous machines as concurrent with classical controls depending on Pulse-
Width-Modulation (PWM) which uses field’s orientation to guarantee the decoupling
between flux and torque. This study deals with direct power control (DPC) that lies
on the principle of direct torque control (DTC) [4].
These two control strategies discussed previously both aims to control
the/active/and/reactive/powers/separately.
The comparison between FOC and DPC evaluate the difference between two
different regulators: PI and hysteresis. However, each control lies on a different
approach. Vector control is based on PI controllers, which are highly dependent
on the machine parameters, while DPC control is based on hysteresis/comparators,
which do not depend on the machine parameters.
In order to compare between DPC and FOC control strategies, this work is orga-
nized as follows: First, In Sect. 2 the turbine and DFIG was modelled to model
the WECS. The Field Oriented Control and Direct Power Control are described
in Sect. 3 and Sect. 4 respectively. The results simulations are obtained using
Matlab/Simulink then, an extensive results discussions are presented in (Sect. 5).
Finally, the conclusion is presented.

2 The modeling of the WECS

Figure 1 illustrates the synopsis of the studied system. The/wind/turbine drives the
DFIG at a variable rotational speed through a multiplier. The DFIG stator is linked
to the network directly. On the other hand, the rotor is linked to the network via
a/back-to-back/converter and transformer [1].
First of all, it is necessary to model the system before proceeding to control.
DFIG-WECS Non Linear Power Controls Application 1457

Stator Power

Grid 50Hz
Wind
DFIG
turbine
Rotor

Controls

Fig. 1 WECS model

2.1 Turbine Model

The aerodynamic/power (Paer ) is the product of power coefficient (Cp ) and wind
power (Pv ). This power is calculated from wind speed (V), blades swept area (S),
and air density (ρ).
The power coefficient presents efficiency as a function of beta angle (β) and the
lambda ratio (λ). The ratio depends on the wind speed (V), blade-radius (R), and
turbine speed () t [2].
The following equations describe wind turbine system.

Paer = Cp .(λ, β).Pv (1)

ρ.S.V3
Pv = (2)
2
R.t
λ= (3)
V
The multiplier is described by the equations cited below:

Ct
Cg = (4)
G
mec
t = (5)
G
With, Cg, mec and G are generator/torque, mechanical/speed of the generator,
and multiplier/gain respectively.
The dynamic equation of the turbine is given by:
1458 M. Bouderbala et al.

dmec
J = Cmec = C g − Cem − C f .mec (6)
dt
With, Cem, Cmec, Cf, and J: electromagnetic torque, mechanical torque, and
viscous friction torque, total inertial respectively.

2.2 Doubly Fed Induction Generator Model

There are three types of equations organize the dynamic modeling of DFIG in the
park frame (d.q): Electrical equations, Flux equations and electromagnetic torque
[5].
Electrical equations:

dϕsd
V sd = Rs.I sd + − ϕsq.ωs (7)
dt
dϕsq
V sq = Rs.I sq + + ϕsd.ωs
dt
dϕr d
V r d = Rr.I r d + − ϕrq.ωr
dt
dϕrq
V rq = Rr.I rq + + ϕr d.ωr
dt
Flux equations:

ϕsd = Ls.I sd + Lm.I r d (8)

ϕsq = Ls.I sq + Lm.I rq

ϕr d = Lr.I r d + Lm.I sd

ϕrq = Ls.I rq + Lm.I sq

Electromagnetic torque:

M  
Cem = p. . ϕsq .Ir d − ϕsd .Irq (9)
Ls
Vs (d.q), ϕs (d.q), Is (d.q): voltages, flux, and currents statoric in PARK reference
Vr (d.q), ϕr (d.q), Ir (d.q): voltages, flux, and currents rotoric in PARK reference
Rs, Rr: stator and rotor resistances
DFIG-WECS Non Linear Power Controls Application 1459

Ls, Lr: cyclic stator and rotor Inductances

M: mutual inductance.

ωr = ωs − ω (10)

ω = p. (11)

p: machine’s pairs of pole

ωs, ωr: stator and rotor angular speed.

3 Field Oriented Control (FOC)

The direct current machine is utterly adjusted to variable speed applications. By

decoupling the electromagnetic torque and the flux. Consequently, FOC command
makes DFIG same as DC machine. This method lies on the transformation of the
electrical variables of the machine to a reference frame that rotates with the flux
vector.
The flux may be oriented according to the stator (SFOC) or the rotor (RFOC),
i.e. the dq frame can be fixed to the stator or the rotor. In this study the SFOC has
been used. The following equation describe the orientation of the flux according to
the axis d therefore:

ϕsq = 0 ϕsd = ϕs (12)

As result, the DFIG can easly be controlled. Active power is controlled separately
from reactive power [6, 7].
It is assumed that the stator resistance is negligible as well as the flux is constant.
Equations 7 becomes then:

Vsd = 0 (13)

Vsq = ϕsd.ωs

Injecting Eqs. 12 and 13, in Eqs. 8 and we obtain the following equations of stator
currents and stator flux respectively:

ϕs M
Isd = . .Ir d (14)
Ls Ls
M
Isq = − .Irq
Ls
1460 M. Bouderbala et al.

ϕsd = ϕs = Ls.I sd + M.I r d (15)

ϕsd = 0 = Ls.I sq + M.I rq

The following equations are obtained using stator currents (Eqs. 14) and stator
voltages (Eqs. 13).

3 M
Ps = − . .V s.I rq (16)
2 ϕsd
 
3 ϕsd M
Qs = .V s − .I r d
2 Ls Ls

M 2 d Ir d M2
Vr d = Rr .Ir d + (L r − ). − g.ωs .(L r − ).Iqr
Ls dt Ls
(17)
M 2 d Irq M2 M.Vs
Vrq = Rr .Irq + (L r − ). − g.ωs .(L r − ).Idr + g.
Ls dt Ls Ls

The equations above shows clearly that:

• Vr depends on Ir
• Ps depends on Iqr
• Qs depends only on Idr.
In order to establish this command, the PI controllers will be used.
Based on the Eqs. (17), the following Synopsis Fig. 2 presents the vector control
modelling.

Fig. 2 FOC modeling

DFIG-WECS Non Linear Power Controls Application 1461

4 Direct Power Control (DPC)

Based on direct torque control (DTC) approach, which is specifically designed to

control torque and flux. Direct Power Control (DPC) has been developed to control
directly machine’s power outputs [8–10].
As mentioned above, the DPC control is based on hysteresis controllers.
However, the obtained errors between the reference and measured power values
are inserted in hysteresis controllers. Those controllers define the switching state
of the semiconductors, depending on the switching table and voltage vectors values
[11, 12].

4.1 Switching Table

The switching table presented in this section divides the plan into six sectors ((1)–(6)).
Other tables divide the plan into twelve sectors to improve accuracy.
When the voltage vector is aligned with the chosen rotating reference frame,
the power is related to the direct component of the output current and the power
instantaneous reactivation to the quadrature current [13].
The selection of the inverter switching mode is imposed by two hysteresis bands
Hp, Hq in order to maintain the errors between the power reference values (Pref and
Qref) and their measured values in these bands.
To accomplish this, the errors of the instantaneous active and reactive powers are
processed by the hysteresis comparators. Their outputs are set to 1 to increase the
control variable (P or Q) and to 0 when the variables controlled are unchanged or
must decrease [14, 15].

4.2 Hysteresis Comparators

The hysteresis controllers are used for the purpose to adjust the active and reactive
power settings.
One is influenced by the error ep = Pref − P for active power. The other is
influenced by the error eq = Qref − Q for reactive power.
The hysteresis band widths affect the efficiency of the inverter, especially, the
distortion of the harmonic current, and the frequency average switching time [16, 17].
The following Synopsis Fig. 3 presents the DPC modelling through the hysteresis
comparators then, vector selection after that the pulse generation as indicated above:
1462 M. Bouderbala et al.

5 Results

To approve these controls, Matlab/Simulink was used. The results identified with a
DFIG of 2MW are given as follows.

Fig. 3 DPC modeling

10
6 / / 10
6
Stator Reactive Power ( Var)

2 2
Stator Active Power ( w)

1
1

0
0

-1
-1

-2
-2
1 2 3 4 5
1 2 3 4 5
times(s)
times(s)

a Stator Active Power .b Stator Reactive Power

4000
Stator Currents(A)

5000 2000
Rotor Currents (A)

0
0
-2000

-5000
-4000
0 1 2 3 4 5

1 2 3 4 5 times(s)
times(s) 4000
Stator Currents (A)

c Rotor Currents 2000

-2000

-4000
4.3 4.4 4.5 4.6
times(s)

d Stator Currents

Fig. 4 a, b, c, d FOC control responses

DFIG-WECS Non Linear Power Controls Application 1463

6 6
10 10
2 2

Stator Reactive Power ( Var)

Stator Active Power ( w)

1 1

0 0

-1 -1

-2 -2

1 2 3 4 5 1 2 3 4 5
times(s) times(s)

a Stator Active Power b Stator Reactive Power

4000

5000

Stator Currents (A)

Rotor Currents(A)

2000

0 0

-2000
-5000

-4000
1 2 3 4 5 1 2 3 4 5
times(s) times(s)

c Rotor Currents
Stator Currents(A)

2000

-2000

4.3 4.35 4.4 4.45 4.5 4.55

times(s)

d Stator Currents

Fig. 5 a, b, c, d DPC control responses

Both FOC and DPC control strategies are simulated, tested, and compared to their
powers reference. Due to this reason, steps have been applied as references.
The active power step varies from 0 to −2 MW at times t = 4 s; while the reactive
power step varies also from 0 to −2 MVAR at times t = 25 s. The simulation results
are presented in Fig. 4 for the FOC strategy and Fig. 5 for the DPC strategy; as
follows (a): active power Ps, (b): reactive power Qs, (c): rotor current Ir, (d): stator
current Is.
Firstly, based on Fig. 4a, b and Fig. 5a, b we can clearly conclude that; active and
reactive powers are decoupled. Also, the variables follow their references for both
types of controls. At the instant t = 2 and t = 4, the currents Fig. 4c, d and Fig. 5c,
d change values with the variations of powers Fig. 4a, b and Fig. 5a, b respectively.
By comparing Stator active Fig. 4a and Fig. 5a, and reactive powers Fig. 4b and
Fig. 5b, DPC gives much less oscillations. Regarding rotor and stator currents, DPC
needs less time to achieve steady-state.
From the results, it can be deduced that with hysteresis comparators, there are
less oscillations even the response time is reduced compared to the PI controllers.
1464 M. Bouderbala et al.

Table 1 FOC & DPC

FOC DPC
differences
Coordinates d-q reference frame α-β reference
reference frame frame
Controllers PI controllers Hysteresis
controllers
Switching frequency Constant Variable
Active and reactive Indirectly controlled Directly
power control by rotor currents controlled
PWM Required Not required
Robustness Low Medium
Transitory response Medium Good

The Table 1 summarizes the principal differences between the FOC and the DPC
strategy.

6 Conclusion

This work was dedicated to a comparative study between field-oriented control (FOC)
and direct power control (DPC) applied to a wind energy conversion system (WECS).
The configuration and principle of the FOC and the DPC has been detailed.
The FOC is based on the PI regulator. This regulator has some drawbacks among
them it depends greatly on machine parameters. Though, The DPC is based on
hysteresis comparators and independent of parameter variations.
In this way, the authors can attest that the utilization of the hysteresis comparators
in DPC control has different focal points, for example, diminishing the exchanging
recurrence of intensity switches and improving the waveforms of the yield factors of
the machine, instead of the utilization of the corresponding fundamental regulator
(PI) in the FOC.
Thus, the authors can affirm that the use of the hysteresis comparators in DPC
control has various advantages such as: decreasing the switching frequency of power
switches and improving the waveforms of the output variables of the machine, rather
than the use of the proportional-integral controller (PI) in the FOC.
The results achieved prove the effectiveness of the DPC strategy employed for
DFIG-based system control and attest to the expected performance.
After obtaining these results in the Matlab software, as a perspective, we envisage
applying these commands on the test bench.

Funding The authors declare that they have funding for research from CNRST.

Conflict of Interest The authors declare that they have no conflict of interest.
DFIG-WECS Non Linear Power Controls Application 1465

References

1. Gonzalo A, Jesus L, Miguel R, Luis M, Grzegorz I (2011) Doubly fed induction machine:
modeling and control for wind energy generation. Wiley, Hoboken
2. Bossoufi B, Aroussi H, Bouderbala M (2020) Direct power control of wind power systems
based on DFIG-generator (WECS). In: Proceedings of the 12th international conference on
electronics, computers and artificial intelligence, IEEE-ECAI’2020, June 25–June 27, 2020
3. Li S, Wang H, Tian Y, Aitouch A, Klein J (2016) Direct power control of DFIG wind turbine
systems based on an intelligent proportional-integral sliding mode control. ISA Trans 64:431–
439
4. Alami Aroussi H, Ziani EM, Bouderbala M, Bossoufi B (2020) DPC & DNPC applied to wind
energy converter system. In: Proceedings of the 5th IEEE-REDEC’2020, March 24–26, 2020,
Marrakech, Morroco
5. Bouderbala M, Bossoufi B, Aroussi H, Lagrioui A, Taoussi M, Livinti P (2020) DEAD-
BEAT control applied to wind power system. In: Proceedings of the 5th international confer-
ence on renewable energy in developing countries, IEEE-REDEC’2020, March 24–26, 2020,
Marrakech, Morroco
6. Siegfried H (2006) Grid integration of wind energy conversion systems. Wiley, Chichester
7. Muller S, Deicke M, De Doncker RW (2003) Doudly fed induction genertor systems for wind
turbines. IEEE Ind Appl Mag 8:26–33
8. Alami AH, Ziani EM, Bouderbala M, Bossoufi B (2020) Improvement of the direct torque
control of the doubly fed induction motor under variable speed. Int J Power Electron Drive
Syst 11(1):97–106
9. Noguchi T, Tomiki H, Kondo S, Takahashi I (1998) Direct power control of PWM converter
without power-source voltage sensors. IEEE Trans Ind Appl 34(3):473–479. https://doi.org/
10.1109/28.673716
10. Kendouli F, Boulahia A, Nabti K, Benalla H (2010) Modeling of a variable speed wind turbine
driving PMSG using hysteresis band pulse width modulation (PWM). Int Rev Model Simul
3(6):1196–1201
11. Kumar DK, Ram Das GT (2018) Simulation and analysis of modified DTC of PMSM. Int J
Electr Comput Eng 8(5):2894–2903
12. Alami AH, Ziani EM, Bouderbala M, Bossoufi B (2020) Enhancement of the direct power
control applied to DFIG-WECS. Int J Electr Comput Eng (IJECE) 10(1):35–46
13. Tamalouzt S, Idjdarene K, Rekioua T, Abdessemed R (2016) Direct torque control of wind
turbine driven doubly fed induction generator. Rev. Roum. Sci. Techn.– Électrotechn. et Énerg.
61(3):244–249
14. Daoud A, Derbel N (2019) Direct power control of DFIG using sliding mode control approach.
In: Derbel N, Zhu Q (eds) Modeling, identification and control methods in renewable energy
systems. Green Energy and Technology. Springer, Singapore
15. Kairous D, Belmadani B (2015) Robust fuzzy-second order sliding mode based direct power
control for voltage source converter. Int J Adv Comput Sci Appl 6:167–175
16. Yousefi-Talouki A, Zalzar S, Pouresmaeil E (2019) Direct power control of matrix converter-fed
DFIG with fixed switching frequency. Sustainability 11:2604
17. Monmasson E, Cirstea M (2007) FPGA design methodology for industrial control systems – a
review. IEEE Trans Ind. Electron. 54(4):1824–1842
Behaviour of Concrete and Steel
for the Construction of Storage Tanks:
Comparative Study

Salma El Aamery, El Hassan Achouyab, and Hassan Ouajji

Abstract The demographic growth and economic development are two main factors
which create more demand for energy reserve production in petrochemical industry.
Under such circumstances, an increase in the volume and quantity of storage tanks has
become must due to the their importance in preventing harms and risks by storing
large volumes of flammable, explosive, toxic and harmful materials, that is why
many researchers are orientating themselves in research projects of materials for the
construction of storage tanks with good quality and with the cheapest. In this article,
we show the different thermo-mechanical behaviors of the two materials which are
concrete and steel so that the constructors can make the most adequate choice for
their needs (storage conditions, temperature, pressure, etc…), especially since each
of these materials has specific characteristics that can influence the types of tanks as
well as the storage modes.

Keywords Storage tanks · Concrete · Steel · Material behaviour · Comparative

study

1 Introduction

In recent years, several events have led to the development of the construction of
storage tanks, to store the various products encountered in the petrochemical and
chemical industry, which is mainly carried out in metal tanks, which are welded
and installed freely in the air (overhead tanks) [1]. Storage has been the subject of
much attention because it offers the possibility of energy savings and efficient use
[2]. For this reason, manufacturers use various types of storage tanks, for example,
aboveground tanks, flat-bottomed tanks, concrete silos or steel for the storage of coal,

S. El Aamery (B) · H. Ouajji

Laboratory SSDIA, ENSET, Mohammedia, Hassan II University
of Casablanca, Casablanca, Morocco
E. H. Achouyab
Laboratory of Condensed Matter Physics and Renewable Energies FST, Mohammedia,
Hassan II University of Casablanca, Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1467
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_133
1468 S. El Aamery et al.

coke, grain, etc., tanks spherical (pressure vessels) for the storage of water and oil, of
high-pressure liquefied water and underground tanks for the storage. This requires
thinking about tanks with larger volumes [3], therefore, it is useful to do research,
with the aim of comparing the properties, and the thermo-mechanical behaviour of
the materials, in order to select the best to build the storage tanks, the most efficient
at the lowest cost [4].
Numerous researchers are studying the physical and chemical properties of
concrete, as concrete is a material widely used all over the world and in different
fields [5]. It is composed of four basic ingredients: water, cement, gravel as coarse
aggregate, and sand as fine aggregate. On the other hand, advanced energy storage
systems are being actively pursued worldwide to respond to the rapid development of
sustainable energy [6, 7]. Aggregates are the world’s most mined mineral resources,
and the global construction market consumed the equivalent of $360 billion in 2018
alone [7]. In this context, this article also discusses concrete as a building material
for storage tanks where it will be compared with steel.
For a better presentation of the results and the different characteristics of these
materials, our article is organized as follows: A review of the literature on tanks
and storage methods is described in Sect. 2. Section 3 unveils a comparative study
between concrete and steel. A discussion is presented in Sect. 4; finally, a conclusion
will be presented in the last section.

2 Literature Review

2.1 Types of Tanks

There are several structures in the industrial sector, among the best known of these
structure, there are storage tanks, which are of different types: vertical cylindrical
tanks, horizontal cylindrical tanks, spherical tanks and spheroidal tanks, each of these
tanks has its own specific characteristics as they can be used for different purposes
in industrial projects [8].
Storage tanks Liquid, as special structures, behave differently from ordinary struc-
tures in terms of dynamic characteristics [9]. They are most often used in distribution
systems, industries and nuclear power plants for the storage of drinking water, toxic
and flammable liquids and nuclear waste respectively. Because earthquake damage
to liquid storage tanks can cause secondary hazards, they are strategically important
structures [10], ich are built in two different shapes, either cylindrical or rectangular.
Figure 1 shows the cylindrical tank with the chordal system (r; 8; z).
Behaviour of Concrete and Steel for the Construction of Storage Tanks … 1469

Fig. 1 The cylindrical tank

and the coordinate system [9]

2.2 Storage Modes

The dimensions of the tank and its material depend on the fluids to be stored and
the type of storage. Oil storage generally consists of a steel structure for a shell and
annular reinforced concrete beams that transfer the load to the ground. The storage
mode can be classified according to the pressure and the operating temperature of
the stored liquid, taking into account the relationship between these two parameters.
Figure 2 shows the three storage modes for liquids with a normal boiling point
(STP) below room temperature.

Fig. 2 The three storage modes for liquids with a normal boiling point (STP) below room
temperature [1]
1470 S. El Aamery et al.

Fig. 3 Concrete Sand

composition Aggregate
[POURCENT [POURCENT
AGE] AGE]

cement
[POURCENT
AGE]
water
[POURCENT
AGE]

2.3 Concrete Composition

Concrete is a generic term for a composite construction material manufactured (see

Fig. 3) from aggregates (sand, gravel) agglomerated by a binder (cement) [11]. Due
to the characteristic difference between the aggregates (rigidity, strength, coefficient
of expansion), the heterogeneity of concrete undergoes significant deformations and
mechanical stresses that can actually locally exceed the elastic limit. The change in
volume of the concrete would therefore be limited to a certain extent. Consequently,
the internal tensile stress developed. Adding that cracking would occur when the
tensile stress in the concrete exceeds the tensile strength, which can lead to a series of
problems concerning the mechanical properties, durability and aesthetics of structural
concrete [12].
Several experiments [13] show that concrete resists better to compression than to
traction, its tensile behavior is strongly influenced by temperature, above 600 °C, most
concretes show a low expansion and sometimes a slight shrinkage due to chemical
decomposition of their various components [14].

2.4 Steel Composition

Steel is a material consisting mainly of iron and some carbon, which are extracted
from natural raw materials from underground (iron and coalmines). Carbon is only
involved in the composition of the material for a very small part (generally less than
1%) [15]. In Table 1, we will show the symbols and values of the steel properties.
On the other hand, electrical steels used for motors, transformers and generators
are usually covered with an insulating coating to improve the performance of the
material in terms of reduced power loss, punching and welding characteristics and
corrosion resistance [16].
Behaviour of Concrete and Steel for the Construction of Storage Tanks … 1471

Table 1 Steel properties

Steel properties Symbol Value
Young’s module Ea 200 × 106 KNm−2
Poisson Coefficient νa 0.3
Elastic limit stress σay 400000 KNm−2
Ultimate Constraint σau 500000 KNm-2
Hardening module Ha 10 × 106 KNm-2
Softening module Sa −30 × 106 KNm-2

3 Comparative Study Between the Behaviour of Concrete

and Steel Tanks

3.1 Behaviour of Concrete Tanks

In general, the tensile strength and the flexural strength of concrete are respectively
of the order of ~10 and ~15% of the compressive strength. In addition, compressive
strength is often considered as a main quality of concrete because it is directly related
to the structure of the hydrated cement paste [17]. On the other hand, concrete is
classified in several strength classes based on compressive strength in most European
standards [18].
As an example, fiber reinforced concrete (FRC) is a high performance material
for the construction and reinforcement of structural elements exposed to impacts
and other types of extreme loads. Many studies for prediction were a reference to
the advantages of incorporating fibers in concretes exposed to explosions [19, 20],
ballistic shocks [21], cyclic loads, and different types of impacts [22]. When fibers
are homogeneously distributed and used in appropriate quantities inside the concrete,
they reduce strength, toughness, ductility, durability and improve other mechanical
properties [23].

3.2 Behaviour of Steel Tanks

Several authors have studied the dynamic behavior of steel storage tanks and their
contents using simplified models capable of predicting the seismic response of these
facilities [24]. These tanks are generally used in petrochemical plants.
Vertical cylindrical steel storage tanks are responsible facilities that are filled
to their maximum capacity for most of their operational life. They often contain
toxic, flammable, and/or expensive products and must therefore be treated with care
and responsibility [25]. They can also be stored above ground using sandbags or
tires as chocks [26]. Steel structures must have sufficient toughness to withstand
various operating temperatures [27], that is why several tests were done to reveal its
1472 S. El Aamery et al.

Fig. 4 Diagrammed
contrainte/deformation

mechanical characteristics, the most important one is the tensile test that leads to the
following diagram (stress/strain) [28] (see Fig. 4).
• OA: rectilinear zone where there is proportionality between the deformation and
the applied stress. It is a reversible elastic zone
• AA’: horizontal bearing, reflecting elongation under constant load. There is
material flow, this is the plastic zone.
• A’B: the load increases again with the elongations up to point B. If the specimen
is unloaded in the plastic zone A’B a remanent elongation is observed; if it is
reloaded, an elastic behavior up to the previous load is observed: the yield strength
has been increased, the metal has been hardened.
• BC: Elongation continues, although the load is decreasing, until point C where it
breaks. In this phase, there is striction: the plastic deformation is localized in a
small portion of the specimen and is therefore no longer homogeneous.

3.3 Comparative the Behaviour of Concrete and Steel Tanks

When tensile tests were done on a steel beam and a concrete beam, with two different
temperature conditions, to compare the displacements of these two materials, we find
that:
At a temperature of 100 °C: we notice that the displacements of these two materials
increase (see Fig. 5), also the displacements of steel are a little bit larger than concrete
of the order of 0.02, same thing we notice at a temperature of 500 °C (see Fig. 6).
Thus, the displacements of either concrete or steel are higher at 500 °C than at
100 °C. Thus, we can conclude that the displacement increases with the increase in
temperature.
Behaviour of Concrete and Steel for the Construction of Storage Tanks … 1473

0.6

0.5

0.4

0.3

0.2

0.1

0
0 100 200 300 400 500 600

steel concrete

Fig. 5 Travel u(x) at temperature 100 °C

2.5

1.5

0.5

0
0 100 200 300 400 500 600

steel concrete

Fig. 6 Travel u(x) at temperature 500 °C

4 Discussions

In the world of energy storage, we are always looking for the evolution and use of
materials and storage conditions, in this context that always researchers try and carry
out studies for new materials and compare them in order to make the best choice in
terms of price quality.
In this part, we will compare two materials that are Steel S355 and Concrete B30, to
know their thermo-mechanical behavior and study the advantages and disadvantages,
to build storage tanks. Table 2 and Table 3 below show the different properties of
strengths and stiffnesses in MP.
1474 S. El Aamery et al.

Table 2 Strength properties

Steel S355 Concrete 30
in MPa
Flexion 270 18
Axial traction 270 2,4
Perpendicular traction 270 2,4
Axial compression 270 18
Perpendicular compression 270 18
Shear 213 1,8

Table 3 Stiffness properties

Steel S355 Concrete 30
in MPa
Average axial module 210000 11500
Mean Transversal Module 210000 11500
Average shear modulus 8100 4800

Compared to concrete tanks, steel tanks have many advantages which are:

Main Advantages
• The high tensile strength of steel.
• The good resistance to earthquakes.
• Easy transportation due to the low weight which allows to transport far.
• Architectural possibilities, much more extensive than concrete [29].
On the other hand, this material also has a number of disadvantages, of which one
finds.

Main Drawbacks:
• Lower compressive strength than concrete.
• Poor fire resistance, requiring expensive protective measures.
• Need for regular maintenance against corrosion.

5 Conclusion

Based on this comparative study between concrete and steel, we can conclude that
concrete is a good material, which resists compression unlike steel, which is more
rigid in traction. Therefore, to better build a storage tank, we recommended using a
composite material of concrete and steel, or using a high performance concrete, espe-
cially if it is used with high temperatures and high pressures, that is to say, physical
properties that are particularly suitable for municipal and industrial applications.
As a research perspective, we are thinking of modeling by the finite element
method the storage tanks built by any material in order to have the thermo-mechanical
behaviors.
Behaviour of Concrete and Steel for the Construction of Storage Tanks … 1475

References

1. Chamayou R (1997) Réservoirs métalliques: stockage des liquides. Généralités, p 7

2. Wu S, Yan T, Kuai Z, Pan W (2020) Thermal conductivity enhancement on phase change
materials for thermal energy storage: a review. Energy Storage Mater 25:251–295. https://doi.
org/10.1016/j.ensm.2019.10.010
3. Guo X, Ji J, Khan F, Ding L (2020) Fuzzy Bayesian network based on an improved similarity
aggregation method for risk assessment of storage tank accident. Process Saf Environ Prot
144:242–252. https://doi.org/10.1016/j.psep.2020.07.030
4. Chen HC et al (2019) Synthesis of amorphous nickel–cobalt–manganese hydroxides for
supercapacitor-battery hybrid energy storage system. Energy Storage Mater 17:194–203.
https://doi.org/10.1016/j.ensm.2018.07.018
5. del Rey CE, Almesfer N, Saggi O, Ingham JM (2020) Light-weight concrete with artificial
aggregate manufactured from plastic waste. Constr Build Mater 265:120199. https://doi.org/
10.1016/j.conbuildmat.2020.120199
6. Zhang LL, Zhao XS (2009) Carbon-based materials as supercapacitor electrodes. Chem Soc
Rev 38(9):2520–2531. https://doi.org/10.1039/B813846J
7. Zheng S et al (2017) Graphene-based materials for high-voltage and high-energy asymmetric
supercapacitors. Energy Storage Mater 6:70–97. https://doi.org/10.1016/j.ensm.2016.10.003
8. Télécharger norme api 650 français pdf api 650 version francaise, Réservoirs de stockage :
Méthodologie de - Revues et Congrès PDF. PdfCours.com. https://pdfcours.com/Exercices_
PDF_Corriges_Cours_1.php?Exercices_Cours_Corriges_Gratuit=7807&PDF_Corriges=
norme_api_650_fran%C3%A7ais_pdf. Accessed 23 Sep 2020
9. Moradi R, Behnamfar F, Hashemi S (2018) Mechanical model for cylindrical flexible concrete
tanks undergoing lateral excitation. Soil Dyn Earthq Eng 106:148–162. https://doi.org/10.1016/
j.soildyn.2017.12.008
10. Soni DP, Mistry BB, Panchal VR (2011) Double variable frequency pendulum isolator for
seismic isolation of liquid storage tanks. Nucl Eng Des 241(3):700–713. https://doi.org/10.
1016/j.nucengdes.2011.01.012
11. Ec B (2005) Calcul des structures en béton, p 27
12. Igarashi SI, Bentur A, Kovler K (2000) Autogenous shrinkage and induced restraining stresses
in high-strength concretes. Cem Concr Res 30(11):1701–1707. https://doi.org/10.1016/S0008-
8846(00)00399-9
13. Hager I (2020) Comportement a haute temperature des betons a haute performance - evolution
des principales proprietes mecaniques. https://www.academia.edu/18524207/Comportement_
a_haute_temperature_des_betons_a_haute_performance_evolution_des_principales_propri
etes_mecaniques. Accessed 15 Sep 2020
14. Noumowe A (1995) Effet de hautes températures (20-600°C) sur le béton: cas particulier du
béton a hautes performances. These de doctorat, Lyon, INSA
15. Zrak A, Koňár R, Jankejech P (2015) Influence of chemical composition in steel on laser
cutting stability. Manuf Technol 15(4):748–752. https://doi.org/10.21062/ujep/x.2015/a/1213-
2489/MT/15/4/748
16. Lindenmo M, Coombs A, Snell D (2000) Advantages, properties and types of coatings on
non-oriented electrical steels. J Magn Magn Mater 215:79–82. https://doi.org/10.1016/S0304-
8853(00)00071-8
17. Adenot F (1992) Durabilite du beton: caracterisation et modelisation des processus physiques
et chimiques de degradation du ciment. These de doctorat, Orléans
18. NORME NF EN 206/CN: Présentation Générale. Infociments. https://www.infociments.fr/bet
ons/norme-nf-en-206-cn-presentation-generale. Accessed 22 Sep 2020
19. Drdlová M, Buchar J, Krátký J, R̆ídký R (2015) Blast resistance characteristics of concrete
with different types of fibre reinforcement. Struct Concr 16(4):508–517
20. Luccioni B et al (2018) Experimental and numerical analysis of blast response of high strength
fiber reinforced concrete slabs. Eng Struct 175:113–122. https://doi.org/10.1016/j.engstruct.
2018.08.016
1476 S. El Aamery et al.

21. Almusallam TH, Siddiqui NA, Iqbal RA, Abbas H (2013) Response of hybrid-fiber reinforced
concrete slabs to hard projectile impact. Int J Impact Eng 58:17–30. https://doi.org/10.1016/j.
ijimpeng.2013.02.005
22. Yahaghi J, Muda ZC, Beddu SB (2016) Impact resistance of oil palm shells concrete reinforced
with polypropylene fibre. Constr Build Mater 123:394–403. https://doi.org/10.1016/j.conbui
ldmat.2016.07.026
23. Mehta PK, Monteiro PJM (2001) Concrete microstructure, proprietis and materials, 20 Oct
2001
24. Caprinozzi S, Paolacci F, Dolšek M (2020) Seismic risk assessment of liquid overtopping in a
steel storage tank equipped with a single deck floating roof. J Loss Prev Process Ind 67:104269.
https://doi.org/10.1016/j.jlp.2020.104269
25. Elsevier Enhanced Reader. https://reader.elsevier.com/reader/sd/pii/S2452321620300913?
token=BAA4AC4C9E8559789F2173771B34EB6C910AFAAD131CA7A1AB38CBBBD
7B9982BA183C696752D3A40375F3FBFE9D06392. Accessed 25 Sep 2020
26. Cheremisinoff NP, Cheremisinoff PN (1995) Steel and fiberglass construction for below ground
storage tanks. In: Fiberglass reinforced plastics. Elsevier, pp 158–180
27. Diagnosis of the microstructural and mechanical properties of over century-old steel railway
bridge components. Elsevier Enhanced Reader (2020). https://reader.elsevier.com/reader/
sd/pii/S1350630719311501?token=406588265883F448006A629FB71499E9200E0552B3
C37465EB2E1FAE3EBF2B8B895A70028F5D6F5C811518653F7C41AF. Accessed 08 Oct
2020
28. Les essais sur le matériau acier. https://notech.franceserv.com/materiau-acier-essais.html.
Accessed 09 Oct 2020
29. Tomek R (2017) Advantages of precast concrete in highway infrastructure construction.
Procedia Eng 196:176–180. https://doi.org/10.1016/j.proeng.2017.07.188
Sensorless Control of DC-DC Converter
Using Integral State Feedback Controller
and Luenberger Observer

Djamel Taibi , Toufik Amieur , Mohcene Bechouat ,

Moussa Sedraoui , and Sami Kahla

Abstract This paper presents a design of linear state feedback control of DC-DC
Boost converters, in order to achieve a particular desired behavior. To guarantee a
zero steady state error, we introduce an integral action, which will work out this
problem by assuring that the steady state error will end up to zero. If it is supposed
that both the voltage and current are measured, so much more sensors are needed
then and consequently causing a high cost, so that to estimate the voltage and current
with a low cost and less complexity it is preferred to introduce a state observer. An
observer or estimator is a dynamic system that uses the available information on a
real system, according to the inputs and outputs of the real process and estimate the
system state. Simulations results demonstrate the robustness and effectiveness of the
proposed control scheme.

Keywords Boost converter · Observer · Static error · Feedback control

D. Taibi (B)
Département de Génie Électrique, Université de Kasdi Merbah, Ouargla, Algeria
T. Amieur
Département de Génie Électrique, Université de Larbi Tebessi, Tebessa, Algeria
e-mail: [email protected]
T. Amieur · M. Bechouat · M. Sedraoui
Laboratoires des Télécommunications LT, Université 8 mai 1945, Guelma, Algeria
e-mail: [email protected]
M. Bechouat
Faculté des Sciences et Technologie, Université de Ghardaïa, Noumirat BP 455,
Route Ouargla Ghardaia, 47000 Ghardaia, Algeria
S. Kahla
Centre de Recherche en Technologies Industrielles CRTI, 16014 Cheraga Alger, Algeria

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1477
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_134
1478 D. Taibi et al.

1 Introduction

The development of semiconductors in recent years resulting from advances in

physics led to the advent of a new discipline in electrical engineering. Under the
name of power electronics, there is a part of electrical engineering that treats the
static conversion of electrical energy from one form to another, adapted to the need
of the user. The systems responsible for handling electrical energy are the converters
that allow the electrical energy source to be adapted to a given receiver by converting
between network and load. These transformations appear in four forms for which
four types of converters are associated: rectifier, inverter, dimmer, chopper [1, 2].
The use of power converters is becoming more and more important. Recent tech-
nological applications require a high level of precision and performance, so DC-DC
converters have a very important role in systems requiring conversion and adapta-
tion of the energy level. In this work; we are interested in modelling of DC-DC
converters and in developing control laws to allow ensuring stability and a certain
level of performance by taking into account the problem and their application in an
industrial environment. The objective is therefore: to improve the performance of the
converters and to reduce the static error. For this reason, several control strategies are
tested in this work in order to regulate the output voltage of DC-DC Boost converters.
Finally, in order to economically reduce the number of sensors, we have introduced
the estimation of the converter coil current by using an observer.
The control technique consists of using a linearized model of the system with
uncertainties proposed in [3]. The LQR controller used in [4–6], robust state feedback
has been designed in [7] using passivity and using LMI in [8–11].
In this work, we investigate the application of linear state feedback control tech-
niques to improve the dynamic behaviour of DC-DC converters. Section 2 is reserved
for the modelling step of the boost. Sections 3.A and 3.B present the necessary theo-
retical elements for the use of these techniques, namely, the state feedback with
integral action. Section 4, presents the state feedback with integral action and the
state feedback with observer. Section 5, presents the application of these techniques
to boost converters which is modelled in the second section. Finally, the whole of
this paper is achieved by a conclusion.

2 Modeling of Boost Converter

The boost converter has earned its name due to its ability of producing a DC output
voltage greater in magnitude than the DC input voltage (step-up). The Circuit diagram
of a Boost converter is as shown in Fig. 1.
While the transistor is passing current, the diode is reverse biased. Hence, the
source E is not connected to a load R.
While the transistor is blocking current, the diode is forward biased and
conductive, which makes the source E connected to a load R [12–14].
Sensorless Control of DC-DC Converter … 1479

Fig. 1 Circuit diagram of a Boost converter

Fig. 2 Boost converter switch turned ON

Fig. 3 Boost converter switch turned OFF

2.1 Switched Converter Model

When the ideal switch is ON u = 1 as shown in the Fig. 2

The following dynamic is obtained:
 di
L dt =E
(1)
C dv
dt
= − Rv

When the switching function is u = 0 (Fig. 3):

The following dynamic is obtained:
 di
L dt = E −v
(2)
C dv
dt
= i − Rv
1480 D. Taibi et al.

From the two Eqs. (1) and (2), we can obtain a single unified model, which is
u ∈ {0, 1}. So the dynamic of the converter is as follows [12]:

L di
dt
= E − v(1 − u)
(3)
C dt = − Rv + i(1 − u)
dv

2.2 Linearization

The linearization of the average model is found to be given by [12]:

   E 
0 − DL1
DL
ė = e+ D 2 eu (4)
1
DC
− RC
1
− RC
 
i −i
With e = and eu = u av − u av .
v−v

3 State Feedback Controller with Integral Action

3.1 State Feedback Controller

The State feedback controller consists in considering the model of the process
described by the following equation [15]:

ẋ(t) = Ax(t) + Bu(t)
(5)
y(t) = C x(t)

It assumes xi (t) of the state vector x(t) are accessible for measurement.
A possible state-feedback control law can be given as [16]:

u(t) = r − K c x(t) (6)

Hence, the matrix K c of (1 × n) dimension is called the gain matrix.

If the process is controllable, the choice of components of K c allows you to place
n poles of the closed loop system as desired.
The associated block diagram is given below (Fig. 4).
Referring back to the precedent equations and schematic, we see that by employing
a law u(t) = r − K c x(t), the state-space equations become the following [17]:
Sensorless Control of DC-DC Converter … 1481

Fig. 4 Closed loop system with u(t) = −K c x(t)

Fig. 5 State feedback with integral action


ẋ(t) = (A − B K c )x(t) + Br (t)
(7)
y(t) = C x(t)

3.2 State Feedback Controller with Integral Action

As we have seen before, the status feedback command moderates the poles of the
system in closed loop. However, the latter does not ensure a position error. One
possibility to solve the problem is to add an integrator in the direct chain as shown
in Fig. 5.
We introduce an integral action: a noted state variable xi placed at the exit of the
integrator with [18].
 
xi = (r (t) − y(t)) = (r (t) − C x(t)) (8)
1482 D. Taibi et al.

Where v(t) is an external disturbance which acts on the states of the system
through the matrix B0 . The command designed before the screen cannot cancel the
effect of this disturbance on monitoring the reference which is given by:
⎧
⎨ ẋ(t) = Ax(t) + Bu(t) + B0 v(t)
ẋi (t) = −C x(t) + r (9)
⎩
y(t) = C x(t)

If we define the augmented state vector xaT = x xi , we can write the system in
the following increased form [18]:

ẋa (t) = Aa xa (t) + Ba u(t) + B0a v(t)
(10)
y(t) = Ca xa (t)

With the matrices

     
A 0 B B

Aa = , Ba = , B0a = , Ca = C 0
−C 0 0 0

The control law is given by

 

x(t)
u(t) = −K a xa (t) = − K K i (11)
xi (t)

u(t) = −K x(t) − K i xi (t) (12)

Which gives a closed loop equation as follows:

ẋa (t) = [Aa − Ba K a ]xa (t) + Ba u(t) + B0a v(t) (13)

Finally the gain vector is equal to the desired polynomial

 
 s I − (A − B K ) B K i 

|s I − (Aa − Ba K a )| =   (14)
C s 

Based on linearization (4)

     
ė1 e1 eu1
=A +B (15)
ė2 e2 eu2

The integral action introduced by e3 = e2 therefore ė3 = e2 .
The increased Eq. (15) becomes:
Sensorless Control of DC-DC Converter … 1483
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
ė1 e1 eu1
⎣ ė2 ⎦ = Aa ⎣ e2 ⎦ + Ba ⎣ eu2 ⎦ (16)
ė3 e3 eu3

With
 
A0
Aa = (17)
C0

4 Luenberger Observer

In this part, it is considered that all the state variables are available for the feedback.
However, practically, some variables may not be available. Therefore, the estimate
phase of those variables is definitely an essential step, means, that we must estimate
the non-measurable state variables so as to elaborate the control signals.
The observer is given by [19, 20]:
  
ˆ = A x̂(t) + Bu(t) + L y − ŷ
ẋ(t)
(18)
ŷ(t) = C x̂(t)

L: gain, L T = l1 l2 . . . ln .
Error is defined as

e(t) = x(t) − x̂(t) (19)

Differentiating Eq. (19) gives

˙
ė(t) = ẋ(t) − x̂(t) (20)

and by substituting Eqs. (5), (18) into (20) we get

 

ė(t) = Ax(t) + Bu(t) − A x̂(t) + Bu(t) + L y(t) − ŷ(t) (21)

   
ė(t) = A x(t) − x̂(t) + L y(t) − ŷ(t) (22)

With (5) and (18)

   
ė(t) = A x(t) − x̂(t) + LC x(t) − x̂(t) (23)

ė(t) = (A − LC)e(t) (24)

1484 D. Taibi et al.

Solution of Eq. (24) is given by

e(t) = e( A−LC) e(0) (25)

The eigenvalues of the matrix (A − LC) can be made arbitrary by appropriate of

the choice gain L (Fig. 6).

5 Simulation and Results

5.1 Open Loop

The converter is controlled by the PWM technique (carrier and modulating) linked
to an open loop IGBT switch (Figs. 7, 8 and 9).

Discussion
– The controlled system in open loop is a blind system.
– The open loop simulation gives a short rise time, but a long response time with
oscillations and in addition a significant error and this is not good for the system
(Figs. 10, 11, 12 and 13).

5.2 Integral State Feedback Controller Integral with Observer

Discussion
– State feedback improves system stability and eliminates overflow and static error
of the system that appears in the open loop.
– The integral action compensates the error that the variation of the load creates.

Fig. 6 Schematic diagram of the Luenberger observer

Sensorless Control of DC-DC Converter … 1485

Fig. 7 Open loop boost converter

Fig. 8 Current measurement 1.4

in the open loop 1.2

Current Measurement

0.8

0.6

0.4

0.2

0
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05
Time (sec.)

Fig. 9 Voltage measurement 30

in the open loop 25

Voltage Measurement

20

15

10

-5
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05
Time (sec.)

– According to the gotten results from the above graphs, the convergence of the
estimated current and voltage towards its true value is perfectly achieved.
1486 D. Taibi et al.

Fig. 10 Boost converter controlled by state feedback integral with an observer

Fig. 11 Control signals 0.9

0.8

0.7
Control signals

0.6

0.5

0.4

0.3

0.2

0.1
0 0.02 0.04 0.06 0.0 8 0.1 0.12 0.14 0.16 0.18 0.2
Time (sec.)

Fig. 12 Current 1
I Estimate
0.9 I Measurement

0.8

0.7

0.6
Current

0.5

0.4

0.3

0.2

0.1

0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
Time (sec.)
Sensorless Control of DC-DC Converter … 1487

Fig. 13 Voltage 25

20

15

Voltage
10

5
V Estimate
V Measurement
Reference
0

-5
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
Time (sec.)

6 Conclusion

This paper has demonstrated the control design methods for Boost converter. The
state feedback control for with an integrator, underling the clear advantages of state
feedback, that has a positive effect on response settling time, reducing the undesirable
peak overshoots and serve having a less oscillated performance, referring that this
approach doesn’t provide a zero static error, this latter has been solved by adding an
integral action to state feedback control, that has proven its efficiency when working
out the steady-state error then the use of state estimation technique with Luenberger
observer was discussed and simulated, that comes out with the advantages of the
observer in estimating those non measurable state variables.

References

1. Hsu CF, Lee TT, Wen YW, Ding FS (2006) Intelligent control for DC-DC power converter
with recurrent fuzzy neural network approach. In: 2006 IEEE international conference on fuzzy
systems, pp 457–462
2. Xia C, Geng Q, Gu X, Shi T, Song Z (2012) Input–output feedback linearization and speed
control of a surface permanent-magnet synchronous wind generator with the boost-chopper
converter. IEEE Trans Ind Electron 59(9):3489–3500
3. Ortiz-Lopez MG, Leyva-Ramos J, Carbajal-Gutierrez EE, Morales-Saldana JA (2008)
Modelling and analysis of switch-mode cascade converters with a single active switch. IET
Power Electron 1(4):478–487
4. Woo YJ, Le HP, Cho GH, Cho GH, Kim SI (2008) Load-independent control of switching DC-
DC converters with freewheeling current feedback. IEEE J Solid-State Circuits 43(12):2798–
2808
5. Jaen C, Pou J, Pindado R, Sala V, Zaragoza J (November 2006) A linear-quadratic regulator
with integral action applied to pwm dc-dc converters. In: IECON 2006–32nd annual conference
on IEEE industrial electronics. IEEE, pp 2280–2285
1488 D. Taibi et al.

6. Olalla C, Queinnec I, Leyva R, El Aroudi A (2011) Optimal state-feedback control of bilinear

DC–DC converters with guaranteed regions of stability. IEEE Trans Ind Electron 59(10):3868–
3880
7. Linares-Flores J, Reger J, Sira-Ramérez H (2006) A time-varying linear state feedback tracking
controller for a boost-converter driven dc motor. IFAC Proc Vol 39(16):926–931
8. Doliya D, Bhandari M (November 2017) An LMI approach for robust lqr control of PWM buck
converter with parasitics. In: 2017 7th international conference on communication systems and
network technologies (CSNT). IEEE, pp 103–108
9. Olalla C, Leyva R, El Aroudi A, Garces P, Queinnec I (2010) Lmi robust control design for
boost pwm converters. IET Power Electron 3(1):75–85
10. Olalla C, Queinnec I, Leyva R, El Aroudi A (2011) Robust optimal control of bilinear DC–DC
converters. Control Eng Pract 19(7):688–699
11. Olalla C, Leyva R, Queinnec I, Maksimovic D (2012) Robust gain-scheduled control of
switched-mode dc–dc converters. IEEE Trans Power Electron 27(6):3006–3019
12. Nisso N, Raïdandi D, Djongyang N, Menga FD (2018) Modeling and analysis of boost converter
in small-signals applied to the wind energy conversion system using Matlab/Simulink. Revue
des Energies Renouvelables 21(4):635–649
13. Ounis F, Goléa N (May 2015) µ-synthesis based robust voltage control for cascade boost
power converter. In: 2015 3rd international conference on control, engineering & information
technology (CEIT). IEEE, pp 1–6
14. Yazdani A, Iravani R (2010) Voltage-sourced converters in power systems: modeling, control,
and applications. Wiley, Hoboken
15. Shuai DX (November 2014) State feedback exact linearization control of Buck-Boost converter.
In: 2014 international power electronics and application conference and exposition. IEEE, pp
1490–1494
16. Gkizas G, Amanatidis C, Yfoulis C, Stergiopoulos F, Giaouris D, Ziogou C, Voitekis S,
Papadopoulou S (June 2016) State-feedback control of an interleaved DC-DC boost converter.
In: 2016 24th mediterranean conference on control and automation (MED). IEEE, pp. 931–936
17. Shen M, Yang GH (2012) H 2 state feedback controller design for continuous Markov jump
linear systems with partly known information. Int J Syst Sci 43(4):786–796
18. Chowdhury A, Debnath D (2013) Performance comparison between PID controller and state-
feedback controller with integral action in position control of DC motor. In: Applied mechanics
and materials, vol 367. Trans Tech Publications Ltd., pp 188–193
19. Vinodh KE, Jovitha J, Ayyappan S (2013) Comparison of four state observer design algorithms
for MIMO system. Arch Control Sci 23:131–144
20. Kwon TS, Shin MH, Hyun DS (2005) Speed sensorless stator flux-oriented control of induction
motor in the field weakening region using Luenberger observer. IEEE Trans Power Electron
20(4):864–869
Optimization and Improvement
of the Efficiency of a Drying System
Based on the Dimensional Parameters
Dounia Chaatouf, Benyounes Raillani, Mourad Salhi, Samir Amraqui,
and Ahmed Mezrhab

Abstract Nowadays, modern drying methods have replaced traditional drying that
consists of exposing the product to direct sunlight. The indirect solar dryer is one
of the best-known systems used since this revolt. For the latter, there are different
concepts and equipment to build it, and although there are differences between them,
they all have the same goal, which is preserving food by reducing its moisture content.
These methods use heat under appropriate conditions in terms of temperature, relative
humidity, and air velocity, which results in a decrease in the moisture content of the
dried product, in the extent to prevent enzymatic activities as well as all biological
organisms that can be the cause of a corruption of the vegetal products. Therefore, it
is very important to ensure a good dried product, for that several parameter must be
adjusted, the objective is to optimize the distribution of the airflow along the trays
and to improve the efficiency of the system based on the dimensional parameters and
the quantity of the product that can be dried, for that, the optimal number of trays
and the distance between the trays is investigated using ANSYS FLUENT software
under the oriental region climate of Morocco. The results show that a drying chamber
with 3 trays found to be the optimal number with a distance of 5 cm between the
trays.

Keywords ANSYS FLUENT · Indirect solar dryer · Trays

1 Introduction

There have been many ways of preserving organic products in different parts of the
world thousands of years ago, and these methods differ depending on the product to be
dried and the regional traditions. Some of the most knowns methods of preservation
are drying, canning, freezing, processing (smoking or salting), and pickling.
Morocco has considerable solar potential that present for most of the year, which
makes the solar drying the most widely used method for preserving products. But,

D. Chaatouf · B. Raillani · M. Salhi · S. Amraqui (B) · A. Mezrhab

Laboratory of Mechanics and Energetics, Faculty of Sciences, Mohammed First University,
60000 Oujda, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1489
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_135
1490 D. Chaatouf et al.

after the failure of traditional drying [1] that consists of drying the products in the open
air, extensive research has been carried out to design and develop the performance
of solar dryers, some of them are presented in this section.
The general design of a drying system is mainly based on the dimensions and the
construction concept of the most important components such as the drying chamber,
the solar air heater, and other several parameters that influence the drying quality
of the system. However, the chamber geometry and the solar air heater have not
interested some developers who have chosen to study how to encourage airflow
through intake fans driven by photo-voltaic (PV) solar panels [2, 3]. In addition,
several studies were made on the hybridization of the indirect solar dryer, like the
one made by S. Dhanushkodi et al. [4] in which they use biomass and solar energy as
sources. The study of the dryer was also investigated on different modes, to idealize
and analyze the behavior of the system. There are many modes of operation to
evaluate the efficiency of the dryer, I. Montero et al. [5] conducts a comparison study
between indirect, passive, mixed, active, forced convection, natural convection and
hybrid modes of operation, and it has been revealed that the forced hybrid is the most
efficient mode of operation, followed by passive and active modes.
In the indirect natural convection solar dryer, the most common problem with it
is the non-uniform distribution of air inside the chamber, which affects the quality
of the dried products. Therefore, in this study, we will focus on the geometric study
and optimization of the dimensional parameters specifically, the distance between
the trays and the number of it which is a crucial parameter in terms of the uniformity
of air [6]. The chosen geometry in this paper is an electric solar dryer studied by
Y. Amanlou et al. [7] that characterize with a side mounted plenum chamber, which
affects positively on the air distribution inside the chamber, but in our case which is
natural convection solar dryer equipped with a solar air heater, the distribution of air
is going to be totally different, that we need a new study to ensure good uniformity
of air along the trays.

2 Methodology

2.1 Solar Dryer Description

In this study, the drying system presents an indirect natural convection solar dryer
(Fig. 1) consists with a drying chamber where the figs are distributed on trays as
a product to be dried, equipped with a trapezoidal plenum chamber and a solar air
heater of a 1.5 m2 that composed of an aluminium absorber and simple glazing.
Optimization and Improvement of the Efficiency of a Drying System … 1491

Fig. 1 Schematic diagram

of the solar dryer studied Air outlet

Solar radiation

Air inlet

Fig. 2 Meteorological
station used to collect the
data

2.2 Boundary Conditions

The meteorological data used in this paper are collected from the high precision
station shown in Fig. 2 installed in the university Mohamed first Oujda. The main
data provided for this study are the ambient temperature and the solar irradiation
presented in Eqs. (1) and (2) respectively, that have been used in the subroutine UDF
(User Defined Function) that we wrote to integrate it in ANSYS FLUENT so we can
study the solar dryer more realistically.
π 
Tamb (t) = 25 + 6 cos (t − 14) (1)
12
 
t −6
G sun (t) = 965 sin π , 6 < t < 20 (2)
14
1492 D. Chaatouf et al.

Fig. 3 Comparison of the average outlet temperature and velocity with those of [8]

The inlet boundary condition is a pressure type with atmospheric pressure because
the dryer studied works under natural convection, as well as the outlet type. Regarding
the convective losses in the walls a heat transfer coefficient is defined (3)

h 0 = 2.8 + 3Vw (3)

Where V w = 3 m.s−1 is the wind velocity.

To simulate the resistance to the airflow by the trays that are filled with the fruit of
the fig, we modeled them as a porous medium with 50% porosity using the power-law
model.

3 Results

3.1 Validation

To validate our work, we compare our results with certain experimental and theo-
retical results available in the literature as Jyotirmay et al. [8] that studied a cubical
wooden chamber, similar to a cabinet dryer equipped with an inclined solar chimney
for natural ventilation. Our results presented in Fig. 3 shows a good agreement. The
difference rate for temperature and velocity does not exceed 1.47% and 16% respec-
tively. Noting that our results are closer to the experimental results than the numerical
ones calculated by the authors.

3.2 The Optimal Distance Between the Trays and Its Number

In order to obtain an optimum drying chamber, that is capable of carrying out the
drying process with high quality in a short period of time, the optimization of the
dynamic behavior is necessary for the chosen geometry. For the mesh independency
Optimization and Improvement of the Efficiency of a Drying System … 1493

test, four different grids, 5532, 12130, 16606, and 23820 triangular cells were studied,
and the results show that a mesh grid with 12130 cells is quite enough to study the
geometry. We will study the effect of adding trays on the dynamic behavior of the
existing ones. The cases studied are generally two to five trays, and the distance
between the trays is also studied. The objective is to create a high-performance
drying system with an optimal number of trays for high-quality drying. Concerning
the distance between the trays, three different distances were studied between the two
first trays and the results are presented in Fig. 4 in term of the average temperature of
the trays in the sunshine hours, which shows that as long as the distance between the
two trays increases, the temperature decreases along the lower tray, while the first
one remains constant, this is due to the hot air that rises directly when entering the
drying chamber after being heated by the solar air heater.
Furthermore, the Fig. 5 shows the average temperature of the trays in the sunshine
hours for 4 different cases, where the number of trays is the parameter on which the
study is based.
As you can see, the higher the number of trays, the more T between the additional
tray and the first one. The big differences can affect the products, that’s why the
temperature in the trays must be kept at the same level as uniform as possible. It has
been shown in the figure that the temperature in the case of two trays is higher than
the others; and the difference in temperature is very remarkable, after every addition
of a tray compared to the initial case which has an important and significant effect
on the temperature of the existing ones.
In the cases of two and three trays, it can be noticed that the average temperature in
the trays is relatively equal in each case, but when the number of trays is greater than
three, the temperature differences between the plates of each case are getting bigger.
In particular, by decreasing the average temperature in these trays. The difference
can reach more than 10° in some cases, which is not good for the uniformity of
drying and the product will take a longer time to get it dried, especially the product
that was put in the lower trays where the temperature is lower compared to the high
(first) one. In terms of these qualifications (drying uniformity and drying time), the
two-trays case is better than the others, followed by the three-trays case, which is
slightly better than the others, with a temperature difference of around 2°, besides
this, adding a third tray will carry more product, and thus the optimal number of
trays in our case is three trays (Fig. 6).

Fig. 4 The average 316 TRAY 1

temperature of the two trays TRAY 2
in the sunshine hours of the 315
chosen day 314
T(K)

313
312
311
310
d= 5 cm d= 7,5 cm d= 10 cm
1494 D. Chaatouf et al.

318
316
314
312 Tray 1
310
Tray 2
308
T(K)

306 Tray 3
304
Tray 4
302
300 Tray 5
298
296
5 trays 4 trays 3 trays 2 trays

Fig. 5 The average temperature of trays in the 4 cases in the sunshine hours of the chosen day

V (m/s)
0.00 0.009 0.018 0.027 0.036 0.042 0.048 0.054 0.069 0.075 0.087 0.093 0.1

T(K)
290 296 301 310 316 321 327 333 338 347 353 358 364 370 378 385

(a)

(b)

10a.m 1p.m 5p.m

Fig. 6 Distribution of velocity (a) and temperature (b) inside the chamber at 3 different hours in
the chosen day

The figure above shows the air distribution inside the optimal geometry of the
drying chamber that we studied (the solar dryer was cut off in the pictures). Where
we can see the drying chamber getting hot through the day and the temperature distri-
bution become more uniform after 1 p.m. where the irradiation is at its maximum.
The hot air rises directly when entering the drying chamber after being heated by
the solar air heater and then moves horizontally in the trays before exiting from the
outlet. The air with high velocity passes through the three trays which is not the case
in more than three ones because it crosses just the high trays due to the lower density
of the hot air.
Optimization and Improvement of the Efficiency of a Drying System … 1495

4 Conclusion

In the present work, the optimization of an indirect solar dryer composed of a solar
air heater and a wooden chamber is studied in 2D using ANSYS FLUENT software
under the meteorological conditions of Oujda city (eastern of morocco), the distance
between the trays and its number is a crucial parameter in term of the uniformity of
air inside the drying chamber and according to the results, a drying chamber with 3
trays found to be the optimal number with a distance of 5 cm between the trays.

References

1. Essalhi H, Tadili R, Bargach MN (2017) Conception of a solar air collector for an indirect solar
dryer. Pear drying test. Energy Procedia 141:29–33
2. Goud M, Reddy MVV, Chandramohan VP, Suresh S (2019) A novel indirect solar dryer with
inlet fans powered by solar PV panels: drying kinetics of Capsicum Annum and Abelmoschus
esculentus with dryer performance. Sol Energy 194:871–885
3. Saini V, Tiwari S, Tiwari GN (2017) Environ economic analysis of various types of photovoltaic
technologies integrated with greenhouse solar drying system. J Clean Prod 156:30–40
4. Dhanushkodi S, Wilson VH, Sudhakar K (2017) Mathematical modeling of drying behavior of
cashew in a solar biomass hybrid dryer. Resour-Effic Technol 3(4):359–364
5. Montero I, Blanco J, Miranda T, Rojas S, Celma AR (2010) Design, construction and
performance testing of a solar dryer for agroindustrial by-products. Energy Convers Manag
51(7):1510–1521
6. Chaatouf D, et al (2021) Trays effect on the dynamic and thermal behavior of an indirect solar
dryer using CFD method. In: Hajji B, Mellit A, Marco Tina G, Rabhi A, Launay J, Naimi S (eds)
Proceedings of the 2nd international conference on electronic engineering and renewable energy
systems, ICEERE 2020. Lecture Notes in Electrical Engineering, vol 681. Springer, Singapore.
https://doi.org/10.1007/978-981-15-6259-4_72
7. Amanlou Y, Zomorodian A (2010) Applying CFD for designing a new fruit cabinet dryer. J
Food Eng 101(1):8–15
8. Mathur J, Mathur S (2006) Summer-performance of inclined roof solar chimney for natural
ventilation. Energy Build 38(10):1156–1163
Monitoring and Control System
of a Hybrid Micro-CSP/Biomass Boiler
System for Traditional Baths

Said Lamghari, Hassan Hamdi, Mehdi Khaldoun, Mickael Benhaim,

Fatima Ait Nouh, and Abdelkader Outzourhit

Abstract Traditional hammams are big consumers of wood-energy and water. This
work presents a hybrid micro-csp/biomass system for space heating and hot water
production in traditional baths that enables to offset the wood that is inefficient used
in such baths. In addition, energy efficiency of these baths is enhanced through an
optimal control system. Furthermore, the use of local agricultural residues instead
of wood all contribute to reducing the stress on forests. The system to monitor and
control the consumption of hot water and energy is presented. The data can be locally
stored and transferred through internet, together with system faults. Preliminary test
of the micro-csp showed that the maximum temperature achieved at the outlet of
the parabolic trough system was 41.7 °C while at the bottom of the storage tank it
was 34.4 °C corresponding to a 8.7 °C increase even though water was withdrawn
by users. Space heating and hot water production are decoupled which allows to
stabilize the hot water temperature to meet the needs when the boiler is off.

Keywords Control system · Monitoring · Parabolic trough · Micro-CSP · Hybrid

systems · Biomass boiler · Underfloor heating · Solar energy · Hot water ·
Traditional hammam

S. Lamghari (B) · A. Outzourhit

LaMEE, Faculty of Science Semlalia, Cadi Ayyad University, Marrakech, Morocco
e-mail: [email protected]
H. Hamdi
LMFE, Faculty of Science Semlalia, Cadi Ayyad University, Marrakech, Morocco
H. Hamdi · F. Ait Nouh
EnR2E, National Center for Studies and Research on Water and Energy (CNEREE),
Cadi Ayyad University, Marrakech, Morocco
M. Khaldoun
BuildIT, Marrakech, Morocco
M. Benhaim
Water Solar Maroc, Marrakech, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1497
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_136
1498 S. Lamghari et al.

1 Introduction

Traditional baths (hammams) are a part of the Moroccan culture. Actually, there are
about 10 000 hammams in Morocco, including 300 located in the city of Marrakech.
Since, on average, a traditional hammam can consume about 1.5 tons of wood energy
per day, they are regarded as one of the most important consumers of wood energy.
This is mainly a consequence of using the traditional method of heating, where the
water is heated by wood combustion and the energy of the combustion gases is used
to heat the floors of the hammam rooms.
Several studies were performed on traditional baths (hammams) in Egypt, Turkey,
Morocco, Syria and Algeria [1]. They found that three-quarters of the total thermal
energy consumption of a traditional baths was due to water heating. In addition, insu-
lated roofs could reduce the demand for space heating by around 20%. Sibley and
Sibely [2] evaluated the initiatives undertaken in Morocco to reduce wood consump-
tion in hammams, in order to improve their energy efficiency and reduce deforesta-
tion. Sobhy et al. [3], on the other hand, performed a study on an individual hammam
located in Marrakech. Temperatures between 30 °C and 37 °C can be reached for
sunny days in the winter by using a solar heating system without thermal storage.
Most of the studies that target the hammams focused on the different part used in the
installed system. For example, Badran and Hamadan [4] compared the use of solar
collectors and solar ponds as a supply of heat for an underfloor heating system. The
found that the solar collectors are more efficient and break even in shorter time than
solar ponds. For these systems to operate continuously (as in the case of hammam)
and late in the evenings either large systems with storage are required. Another option
is to couple the solar heating systems with another source (ex. Biomass) to remedy
to the problem of the intermittence and to ensure a continuous supply of hot water
and heat, which is the object of this work.
The main objective of the present work is to present a hybrid system, which
consists of parabolic troughs (micro-csp) and biomass boilers that was developed
for heating water and space of a traditional hammam. We focus our attention of the
control and monitoring system. Previous studies have focused on the modeling in
the system in order to evaluate study the energetic performance. It was shown that
there is no significant difference in the temperature of the rooms when copper or
PEX tubing are used in floor heating [5]. In the same aspect there are several studies
on the performance of the underfloor hating [4, 6–10]. An experimental study of the
system, which is the subject of the present work, was carried out to test and improve
the control system used.
Monitoring and Control System of a Hybrid Micro-CSP/Biomass Boiler System ... 1499

2 Description of the Hammam and the Hybrid System

The hammam is located in Mhamid 4, Marrakech, (Morocco). The hammam is north-

oriented and consists of two bathhouses, one for women, and the other for men. Every
hammam consists of four bathing rooms: a hot one, two warm ones and last one is
not heated, in addition to a dressing room.

2.1 Description of the Studied Hybrid System

Figure 1 represents the hybrid micro-cps/biomass heating system under investiga-

tion destined for space heating and the production of hot water. The Micro CSP
plant using an external heat exchanger preheats the first hot water tank. When the
set point temperature (60 °C) is not attained, the biomass boilers will automatically
provide enough heat to the water in the other two tanks via external heat exchangers

Fig. 1 Schematic diagram of the hybrid biomass-micro-CSP system

1500 S. Lamghari et al.

by activating the corresponding pumps (circulators) installed to the supply manifold.

Space heating; on the other hand, is performed by biomass boilers through under-
floor heating systems and radiators, which are independently operated. In addition,
the corresponding circulating pumps are installed on the supply manifold and are
independently controlled by temperature controllers.

2.2 The Micro-CSP Plant

The installation consists of six branches of collectors in parallel. Each branch has
four parabolic troughs with a total of 8.24 m collector length and 8.25 m2 -aperture
area yielding a peak power of 6 kW under a DNI of 1000 W/m2 . The total peak
power of the solar field is therefore 36 kW. A single-axis east-west tracking systems
was used in our case.

2.3 Biomass Boiler

The boiler used is designed to heat water in a closed loop to a maximum temperature
of 90 °C and uses only biomass feed in granular form. Local agricultural residues
such as argan shells, olive pomace and nuts shell can be used as the biomass feed,
thus allowing to avoid the extensive use of wood energy. The boiler is equipped
with a controller allowing to control the biomass feed rate by adjusting the speed of
rotation of a worm shaft and flow of the air leading to an efficiency as high as 90%.

2.4 The Monitoring and Controller System

The role of the monitoring and control is to ensure the operation of the various parts
of system within the specified temperature set points, the protection of personnel and
material, and to help in detecting system faults. It is also designed to shut down a part
of the system without affecting the other parts, which allows to perform preventative
or corrective maintenance without stopping the whole system. A schematic diagram
of the control/monitoring system is shown in Fig. 2:
The main controller uses a simple algorithm. First, the temperatures are measured
through the sensors located in various places of the hammam and the system compo-
nents (boiler, storage reservoirs, rooms, inlet and outlet of the micro-csp plant…).
The measured temperatures are then compared with the corresponding set points
(TCSP, TBoiler…). When difference is higher than the 6 °C for a given sensor, the
corresponding circulating pump is activated until the difference is lower than 4 °C.
Monitoring and Control System of a Hybrid Micro-CSP/Biomass Boiler System ... 1501

3 Results and Discussion

The evolution of the outdoor air temperature and irradiance during the testing period
of the hybrid micro-csp/biomass boiler system is given in Fig. 3.
In addition, the temperature of the water at the outlet the micro-csp plant and
the bottom of the solar tank were continuously monitored. Figure 4 shows typical
temperatures profiles at these points. It should be noted that hot water is always
drawn from the storage tanks. At night, the temperature of water in the bottom of the
solar tank is the temperature of feed water from well which is around 25.7 °C. For a
sunny day, the maximum temperature achieved at the outlet of micro-csp plant was
41.7 °C and in the bottom of the tank is 34.4 °C, corresponding to increase by 8.7 °C
due to the fact that water is withdrawal by users from the tanks.

Fig. 2 Overall block diagram of the monitoring system

Fig. 3 The evolution of the outdoor air temperature and irradiation

1502 S. Lamghari et al.

Fig. 4 The evolution of the temperature of water in the bottom of the solar tank and at outlet of
parabolic trough system

For a cloudy day, the maximum temperature achieved was 35.6 °C at the outlet
of micro-csp plant and 28.5 °C in the bottom of the solar tank, corresponding to an
increase of 2.8 °C as result of the water withdrawal by users from the tanks.
The evolution of the temperatures of the biomass boiler and of the two tanks
heated by the boiler is shown in Fig. 5, in addition to the control signals (on/off) of
their respective circulator pumps.
When the temperature is higher than the set-point temperature (60 °C), as clearly
seen for example in the case of the second biomass-heated tank, the corresponding
pump is stopped as it can be seen in Fig. 5a (red curves). However, it can be seen
that the water temperature at the bottom of this tank keeps increasing beyond 60 °C
as result of water coming from the top of the first tank since water is drawn from this
the system.
Figure 5b shows the evolution of the temperatures of the biomass boiler and of
the two tanks heated by the boiler when the underfloor heating system is on. It can
be seen that their respective pumps are stopped when the temperature of the boiler
had dropped below or is within 4 °C of the temperatures of the tank. The pumps are
activated when the temperature of the corresponding tank is less than 60 °C and the
boiler temperature is 6 °C higher this value.
Monitoring and Control System of a Hybrid Micro-CSP/Biomass Boiler System ... 1503

(a) (b)

Fig. 5 Control scenarios of the tanks heated by the biomass boiler (a) first one (b) second one

4 Conclusion

In this work we described a hybrid solar-biomass system for space and water heating
in a traditional Hammam. The focus was made on the system control. A simple
algorithm was used to control the system and regulate the temperatures at various
points. The results show that the control system enabled to avoid the cooling of the
hot water when the water in the biomass boiler loop is colder than the water in storage
tanks when the underfloor heating is activated. This work will allow us to develop
an energy management system that will enable to optimize the energy consumption.

Acknowledgment This study is carried out thanks to the funding from the Solar and New Energy
Research Institute of Morocco (IRESEN) under the InnoTerm/InnoBiomass 2014 program.

References

1. Kristina Orehounig AM (2011) Energy performance of traditional bath buildings. Energy Build
43(9):2442–2448. https://doi.org/10.1016/J.ENBUILD.2011.05.032
1504 S. Lamghari et al.

2. Sibley M, Sibley M (2015) Hybrid transitions: combining biomass and solar energy for water
heating in public bathhouses. Energy Procedia 83:525–532. https://doi.org/10.1016/j.egypro.
2015.12.172
3. Sobhy I, Brakez A, Benhamou B (2017) Energy performance and economic study of a solar
floor heating system for a hammam. Energy Build 141:247–261. https://doi.org/10.1016/j.enb
uild.2017.02.044
4. Badran AA, Hamdan MA (2004) Comparative study for under-floor heating using solar collec-
tors or solar ponds. Appl Energy 77(1):107–117. https://doi.org/10.1016/S0306-2619(03)000
12-6
5. Krarouch M, Hamdi H, Lamghari S, Outzourhit (2018) A simulation of floor heating in a
combined solar-biomass system integrated in a public bathhouse located in Marrakech. In: IOP
conference series: materials science and engineering, vol 353, no 1. https://doi.org/10.1088/
1757-899X/353/1/012005
6. Mokhtari A, Kazeoui H, Boukezzi Y, Achard G (1998) Utilisation d’un Circuit Hydraulique
dans un Plancher pour le Chauffage et le Rafraîchissement des Locaux. Revue Energies
renouvelables 1:17–27
7. Oudrane A, Aour B, Benhamou MHM (2016) Méthodologie pour la détermination de l’ écarte-
ment optimal de la chaîne tubulaire d’ une dalle chauffante. Revue Energies Renouvelables
19:11–19
8. Belkacem N, Loukarfi L, Khelil A, Naji H, Braikia M, Missoum M (2015) Simulation des
charges thermiques dans une habitation pilote à plancher réversible. Nature Technologie
9. Amir M, Lacroix M, Galanis N (1999) Thermal analysis of electric heating floor panels with
daily heat storage. Int J Therm Sci 38(2):121–131. https://doi.org/10.1016/S1290-0729(99)800
49-0
10. El Mays A et al (2017) Using phase change material in under floor heating. Energy Procedia
119:806–811. https://doi.org/10.1016/j.egypro.2017.07.101
Low Capacity Diffusion Absorption
Refrigeration: Experimental
Investigations and Thermal
Characterization

Ikram Saâfi, Ahmed Taieb, and Ahmed Bellagi

Abstract A low capacity commercial absorption-diffusion refrigerator (DAR) is

tested under various operating conditions, in transient and steady-state mode. A series
of twenty-three experiments in unsteady-state mode were performed to characterize
the thermal exchange between refrigerator cabin and ambient air and inside the
cabin, between its content and the evaporator. Steady-state tests are also realized by
varying the electrical power supplied to the generator from 15 W to 63 W in order to
identify the optimal operating conditions of the refrigerator. All the experiments are
performed in an air-conditioned room at 26 °C. It is found that the optimal coefficient
of performance is 0.15 for a cooling capacity of 7 W.

Keywords Refrigeration · Absorption-diffusion · Experimental study ·

Steady-state and dynamic mode

1 Introduction

The DAR, subject of our present investigations, was invented in 1928 by two Swedish
engineers, von Platen and Munters [1]. It is a clean thermally-powered refrigeration
technology that can readily be activated by low- to medium-grade renewable heat. It
has been recognized as one of the most encouraging sustainable technologies for cold
production. The machine’s cycle operates at constant total pressure. It uses ammonia
as refrigerant, water as absorber and hydrogen (or less frequently helium) as a non-
absorbable inert gas, necessary to reduce the partial pressure of the refrigerant in the
evaporator to allow the evaporation process to occur in the uniform pressure device.

I. Saâfi (B) · A. Taieb · A. Bellagi

Thermal and Thermodynamic Research Laboratory of Industrial Processes, National Engineering
School of Monastir, University of Monastir, Monastir, Tunisia
A. Bellagi
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1505
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_137
1506 I. Saâfi et al.

The main feature of this machine is its good reliability because it doesn’t have
any moving parts. The circulation of the aqueous ammonia solution is driven by a
bubble pump and that of the gases between absorber and evaporator by gravity.
Numerous theoretical and experimental studies are reported in the literature
dealing with the analysis of the performance of this machine, or part of it, operated
with different energy sources and using various mixtures of working fluids [2–8].
Chen et al. [9] designed a modified generator including a heat exchanger that reuses
the heat released in the rectifier to preheat the rich solution leaving the absorber.
The new cycle configuration showed a significant improvement in cooling coeffi-
cient of performance, COP, up to 50% compared to the original cycle for the same
cooling capacity. Taieb et al. [10, 11] proposed a simulation model for a DAR using
ammonia/water/hydrogen as working fluids. The refrigerator was supposed cooled
with ambient air and activated with solar hot water at 200 °C. An optimal result of
COP of 0.126 associated to a cooling capacity of 22.3 W was found. Freeman et al.
[11] established a thermodynamic model to simulate the performances of a DAR
using ammonia-water-hydrogen as working fluid and integrating solar heat pipes as
a source of heat under the climate conditions of Chennai, India. They found that for a
charge pressure of 14 bar and 200 W heat input, a maximum COP of 0.25 was found
for cooling temperature of 5 ºC.
We report in this paper on our experimental investigations of a low-power commer-
cial refrigerator with a capacity of 25 L, powered by an electrical resistance and
operating according to the von Platen and Munters cycle. The refrigerator uses
ammonia/water as refrigerant/absorbent pair and hydrogen, as pressure equalizer.
All its elements are made of steel.
The objective of the study is to assess the performance of the machine under
various energy input conditions, in steady state as well as in transient operation mode.
A key parameter for the evaluation of the cooling capacity of the refrigerator is the
overall heat exchange coefficient that characterizes the thermal exchange between
cabin and ambient air, i.e. that allows the estimation of heat infiltration from outside in
the isolated cabin. This crucial parameter is generally determined by tests consisting
of electric-resistance heating of the refrigerator cabin in steady-state mode. In the
present study however we deduce this key parameter from a series of unsteady state
temperature measurements during the startup of the refrigerator. To this purpose,
a thermal model of the cabin is developed and a nonlinear regression procedure is
applied to retrieve the overall heat exchange coefficient from the performed transient
tests.
Low Capacity Diffusion Absorption Refrigeration … 1507

2 Experimental Setup

2.1 Description of the Cycle

We refer in the following description to the schematic representation of the DAR in

Fig. 1. The machine is energized when the electrical resistance begins to heat the
ammonia-rich aqueous solution (1). The thermal energy, denoted Q̇ G , is supplied
in the lower part of the combined generator/bubble pump ensemble. When enough
heat is sullied, the liquid solution inside the vertical tube dealing as heat pump starts
boiling. The generated vapor is composed of ammonia, the more volatile compound,
and a small quantity of steam. The vapor forms first small bubbles that during their
ascendant movement coalesce to form larger bubbles occupying the entire section
of the tube. These bubbles rise up pushing the liquid above and carrying in their tail
some liquid to the upper part of the heat pump (2). The pumped liquid is discharged
by gravity into the annular space of the coaxial tube of the generator which is in direct
contact with the electrical resistance. This heated and boiling solution generates an
additional amount of refrigerant vapor.
The ammonia-poor solution (13) exiting the generator is gravity-fed to the
absorber via a solution exchanger. The vapor, on the other side, mounts to the
condenser after purification in the rectifier where most of the accompanying steam
is condensed. The condensed water returns to the annulus of the coaxial tube of
the generator to rejoin the ammonia-poor solution moving to the absorber (14). The
purified ammonia vapor liquefies in the air-cooled condenser by rejecting the thermal
power Q̇ C D . The exiting liquid (5) is subcooled in the gas heat-exchanger (GHX)
before accessing the evaporator at (7). GHX and evaporator form an assembly of
two coaxial tubes: the low-density hydrogen-rich gas (12) leaving the absorber flows
into the central tube upwards and the denser gaseous mixture of cold hydrogen and
refrigerant vapor leaving the evaporator at (9), downwards in the annular space. The
density difference between the two branches of the gaseous loop is at the origin of
the circulation of the gases between absorber and evaporator. The small diameter
tube carrying the liquid refrigerant from the condenser (5) to the evaporator entrance
(7) is placed in direct contact with the outer tube of GHX. At the evaporator entrance
the partial pressure of incoming liquid ammonia is suddenly so reduced that it starts
evaporating at low temperature. The evaporation continues during its journey through
the evaporator and is completed in the GHX. The energy Q̇ E V required for the evap-
oration of the refrigerant is supplied by the content of the refrigerator cabin where
the evaporator is placed, which in its turn cools.
The gaseous mixture entering the absorber (10) is thus ammonia-rich. Inside the
absorber, the ammonia vapor in the gas mixture is absorbed by the lean solution
(14) coming back from the generator. The liberated absorption heat is rejected to
environment at the rate of Q̇ AB . The remaining gas, hydrogen-rich and ammonia-
free (12) becomes lighter and rises towards the evaporator, and the ammonia–rich
liquid solution from (15) starts moving back to the bubble pump through the solution
heat exchanger. And the cycle starts again.
1508 I. Saâfi et al.

Fig. 1 Schematics of the

DAR refrigerator

2.2 Measuring Devices

To follow the operation of the refrigerator, we began by equipping it with adequate

measuring devices as shown in Fig. 2. Eighteen type K thermocouples were fixed on
the outer surfaces of the various elements of the refrigerator and connected to a data
acquisition system.

Fig. 2 Experimental set-up

Low Capacity Diffusion Absorption Refrigeration … 1509

3 Tests and Results

We began our tests by determining the minimum input power required to start the
machine and ensure the production of cold in the cabin. Increasing gradually the
supplied power with each test and monitoring the evolution of the cabin temperature
as well as the temperatures of various locations in the machine, it was found that a
minimum of 22 W heating power is necessary to activate the DAR. For lower energy
inputs (e.g. 20 W) no refrigerant vapor is generated in the boiler and hence, no cold
can be produced, because the solution inside the generator is not hot enough to start
boiling (Figs. 3 and 4). Temperatures of 205 °C and higher are required to initiate
the boiling.
For low heating power in the range 22–35 W, large amplitudes oscillations of the
temperatures (up to 40 °C), particularly in generator (Fig. 3) and evaporator (Fig. 4)
are observed. It was also noticed that by increasing the heating power, the frequency
of the temperature oscillations is augmented and their amplitude reduced. These
oscillations disappear completely for energy input of 43W and larger (Fig. 5).

Fig. 3 Evolution of
evaporator temperature for
increasing heat supply: 20,
22, 24 and 30 W

Fig. 4 Evolution of
generator temperature for
increasing heat supply: 20,
22, 24 and 30 W
1510 I. Saâfi et al.

4 Thermal Characterization of the Machine

To evaluate the DAR performance, i.e. the rate of cold production in evaporator Q̇ E V ,
the thermal exchange characteristics of the refrigerator cabin has to be known. In
fact, in stationary mode operation, the cold production in evaporator compensates
exactly the heat leakage Q̇ ext from outside,

Q̇ ext = Q̇ E V = (U A)ext (T AM B − TC AB )

TC AB and T AM B denote the mean temperature inside the cabin and that of ambient air,
respectively. (U A)ext , the product of overall heat transfer coefficient and exchange
surface area, characterizes the thermal exchange between the DAR cabin and its
environment. The knowledge of (U A)ext is thus necessary for the evaluation of the
cooling capacity Q̇ E V and the coefficient of performance of the refrigerator, C O P

Q̇ E V
COP =
Q̇ G

We propose in this paper to deduce the coefficient (U A)ext from experiments

performed in unsteady-state operation mode of the DAR. To this purpose twenty three
tests in dynamic mode under various operating conditions were performed: 16 tests
were run with empty cabin and the rest with the cabin loaded with different quantities
of water. A simplified model of the thermal exchange between cabin, ambient air
and evaporator is established. Similar to (U A)ext , an overall heat transfer coefficient
(U A)int is defined for the heat exchange rate Q̇ int inside the cabin, between cabin
content and the cold source, the evaporator. Both heat transfer coefficients are to be
inferred from the measurements in dynamic mode operation of the DAR.
The energy balance of the cabin writes

dUcab
= Q̇ ext − Q̇ int (1)
dt

Fig. 5 Evolution of
generator and evaporator
temperatures for heat supply
Q̇ G ≥ 43 W
Low Capacity Diffusion Absorption Refrigeration … 1511

where Ucab is the internal energy of the cabin and

Q̇ ext = (U A)ext (T AM B − TC AB ) (2)

Q̇ int = (U A)int (TC AB − TE V ) (3)

TE V is the mean temperature of the evaporator. In steady-state operation, we have

obviously

Q̇ ext = Q̇ int = Q̇ E V

Considering Eqs. (2) and (3), the energy balance becomes

dUC AB
= (U A)ext (T AM B − TC AB ) − (U A)int (TC AB − TE V ) (4)
dt
In the case of empty cabin (no water bottles inside), we can write

dUC AB dTC AB
= C p,C AB (5)
dt dt
with C p,C AB denoting the heat capacity of the cabin (hardware + inside air). When
a mass m W of water is added, the internal energy of system is increased, and the lhs
of Eq. (4) is correspondently modified

dUC AB dUw dUC AB   dTW

+ = + mC p W (6)
dt dt dt dt
 
where mC p W is the heat capacity of the loaded water and TW , its temperature.
Setting
 
(U A)ext (U A)int mC p W
α= ;β = ;γ =
C p,C AB C p,C AB C p,C AB

the energy balance equation becomes

dTC AB dTW
= α(T AM B − TC AB ) − β(TC AB − TE V ) − γ (7)
dt dt
And in the case of the empty cabin (m W = 0),

dTC AB
= α(T AM B − TC AB ) − β(TC AB − TE V ) (8)
dt
1512 I. Saâfi et al.

Basing on unsteady-state measurements during the start up of refrigerator of the

temperatures T AM B , TC AB and TE V , and TW in the case of water loaded cabin, the
model parameters α and β in Eq. (8) or α, β and γ in model (7) are deduced by
applying a nonlinear regression procedure. The calculations are performed using the
algebra software Mathematica©. Figures 6, 7 and 8 illustrate the applied proce-
dure in the case of empty refrigerator cabin. Fig. 6 depicts an example of measured
temperatures of ambiant air, cabin and evaporator.
To avoid the problems associated with differentation of numerical data series,
because of the derivates dTdtC AB , dTdtW in the model Eqs. (7) and (8), the measured
temperatures TC AB and TW are fitted to the empirical model

T (t) = T0 (1 + Atexp[−Bt] + Cexp[−Dt] + Eexp[−Ft]) (9)

where T0 , A, B, C, D, E and F are smoothing parameters. As Fig. 7 shows, this

model reproduces faithfully the experimental data. Finally, as means of validation of
the thermal model Fig. 8 compares the derivative dTdtC AB ex p deduced from Eq. (9)
 
and dTdtC AB cal obtained from Eq. (8) with the experimental data of TC AB and TE V
and the regressed values of α and β. As can be noted, the experimental findings are
well reproduced by the model.
Table 1 gives the obtained values of the parameters α, β and γ as well as their
average values for all 23 tests.
With the knowledge of the amount of water in the cabin, the mean value of γ
allows to determine the heat capacity of the cabin C p,C AB . Once this parameter
fixed, (U A)ext and (U A)int are respectively deduced from the average values of α
and β. The obtained final results are

C p,C AB = 1.91 kJ/K

(U A)ext = 0.53 W/K

Fig. 6 Typical experimental

temperature measurements
(empty cabin) in dynamic
mode for the determination
of α and β
Low Capacity Diffusion Absorption Refrigeration … 1513

Fig. 7 Comparison of
measured and calculated
(Eq. 9) cabin temperature

Fig. 8 Thermal model

validation: Comparison of
 calculated
measured and 
derivative dTdt
C AB

5 Dar Performance in Steady-State Mode

Now, refrigerator tests are performed in stationary mode and for each heating power
input, the cooling capacity evaluated using (U A)ext . The performances of the inves-
tigated refrigerator are graphically depicted in Figs. 9 and 10. As previously noted,
a minimum heat supply is required to start production of cold by the refrigerator. To
ensure sustainable operation however, a larger activation power, in the order of 36 W,
is needed, producing a cooling capacity of 5 W. Beyond 43 W, the cooling capacity
Q̇ E V stabilizes and the C O P begins to decrease.
As the heat supplied to the generator is progressively increased, the flow rates
of pumped solution and generated vapor in boiler increase too. Raising the flow
rate of the refrigerant vapor in boiler means introducing more liquid ammonia in
the evaporator and hence, larger cooling capacity. Further, an augmented flow rate
of the pumped solution allows absorbing a larger amount of refrigerant vapor in
the absorber, corresponding to an enhancement of the refrigerator performance. This
explains the observed ascending trend of C O P and Q̇ E V in Figs. 9 and 10. Increasing
the heat supply to generator however gives rise to an opposite effect on performance.
1514 I. Saâfi et al.

Table 1 Model parameters

Test # α (h−1 ) β (h−1 ) γ (−)
α, β and γ (*: repeated test)
1* 1.0586 0.4605
2 1.0586 0.4605
3* 0.9456 0.4428
4 0.9099 0.4027
5 0.9456 0.4428
6 0.9099 0.4027
7* 1.0667 0.4862
8 1.0666 0.4862
9 0.9166 0.3555
10 0.8865 0.3992
11* 0.9838 0.4468
12 0.9838 0.4468
13* 0.9026 0.3498
14 1.0579 0.4103
15 1.0579 0.4103
16 1.0579 0.4103
17 0.9549 0.4023 0.7890
18 0.9033 0.3917 0.7889
19* 0.9169 0.3847 0.9812
20 0.9169 0.3847 0.9812
21 1.1142 0.4963 0.9560
22* 1.2240 0.5685 0.9556
23 1.2234 0.5685 0.9556
Mean values 1.0027 0.4355 0.9154

Fig. 9 Experimental C O P
vs. supplied heat to generator
(steady state operation)
Low Capacity Diffusion Absorption Refrigeration … 1515

Fig. 10 Cooling capacity vs.

supplied heat to generator
(steady state operation)

In fact, due to the raising temperature in the boiler, more vapor is produced but
containing more steam. The purification of ammonia by condensation of the excess
water vapor in the rectifier implies wasting more of the supplied energy in the boiler
without producing more refrigerant vapor as Fig. 10 shows. Even if more refrigerant
vapor is generated, the low value of the overall heat transfer coefficient (U A)int
limits the heat supply to the evaporator and hence, the evaporation of more liquid
refrigerant. The consequence is the declining refrigerator performance observed in
Fig. 9 and the stagnation of the cooling capacity for Q̇ G larger than 43 W. Additional
heating at the generator beyond the conditions for maximum C O P are not necessary
and should be avoided.

6 Conclusion

In this paper, experimental investigations of a low-power commercial refrigerator

with a capacity of 25 L and powered by an electrical energy are presented. It is
observed that this hermetic, uniform pressure refrigerator exhibits a stable oscillatory
behavior for low energy input between 22–37 W. The frequency of the temperature
oscillations increases with increasing heating power and the amplitudes reduce. For
higher power input, the oscillations completely disappear. To evaluate the perfor-
mance of the refrigerator transient as well steady-state tests are performed. A simple
thermal model is established in order to deduce the thermal exchange characteristics
of the machine cabin from the unsteady-state experiments. In particular, the overall
heat transfer coefficient between the cabin and the ambient is determined. It was
found that.

(U A)ext = 0.53 W/K

This parameter allows the evaluation of the cooling capacity Q̇ E V of the refrigerator
and the coefficient of performance C O P in stable steady-state operation mode.
1516 I. Saâfi et al.

Following optimal performances for ambient air temperatures of 23–25 °C are found

Q̇ E V = 7.5W

C O P = 0.15

References

1. Von Platen BC, Munters CG (1928) Refrigerator, US Patent 1, 685–764

2. Schmid F, Bierling B, Spindler K (2018) Development of a solar-driven diffusion absorption
chiller. Sol Energy 177:483–493
3. Mansouri R, Boukholdaa I, Bourouis M, Bellagi A (2015) Modelling and testing the perfor-
mance of a commercial ammonia/water absorption chiller using Aspen-Plus platform. Energy
93:2374–2383
4. Pérez-García V, Rodríguez-Muñoz JL, Ramírez-Minguela JJ (2019) Theoretical modeling
and experimental validation of a small capacity diffusion-absorption refrigerator. Int J Refrig
104:302–310
5. Chaves FD, Moreira MFS, Cortez MFB (2019) Experimental study and modeling within
validation of a diffusion absorption refrigerator. Int J Refrig 101:136–147
6. Ben Jemaa R, Mansouri R., Bellagi A (2017) Experimental characterization and performance
study of an ammonia–water–hydrogen refrigerator. Int J Hydrogen Energy 42:8594–8601
7. Rattner S, Garimella S (2016) Low-source-temperature diffusion absorption refrigeration.
Experiments and Part II: Model assessment. Int J Refrig 65:312–329
8. Benhmidene A, Hidouri K, Chaouachi B, Gabsi S, Bourouis M (2016) Experimental investi-
gation on the flow behaviour in a bubble pump of diffusion absorption refrigeration systems.
Case Stud Thermal Eng 8:1–9
9. Chen J, Kim KJ, Herold KE (1996) Performance enhancement of a diffusion–absorption
refrigerator. Int J Refrig 19:208–218
10. Taieb A, Mejbri K, Bellagi A (2016) Theoretical analysis of a diffusion-absorption refrigerator.
Int J Hydrogen Energy 41:14293–14301
11. Taieb A, Mejbri K, Bellagi A (2016) Detailed thermodynamic analysis of a diffusion-absorption
refrigeration cycle. Energy 115:418–434
12. Freeman J, Najjaran A, Edwards R, Reid M, Hall R, Ramos A, Markides CN (2017) Testing
and simulation of a solar diffusion-absorption refrigeration system for low-cost solar cooling
in India. ISES Solar World Congress
The Synchronized Electrical Charge
Extraction Regulator for Harvesting
Energy Using Piezoelectric Materials

Youssef El Hmamsy, Chouaib Ennawaoui, Ikrame Najihi,

and Abdelowahed Hajjaji

Abstract The aim of this paper is studied the Elements that affect the circuit
SECE for piezoelectric energy harvesting and the influences on piezoelectric energy
harvesting performance. In addition, the energy harvesting system based on piezo-
electric energy plays an important role in the conversion of vibration energy from
the environment into electrical energy, which can be used by low-power elec-
tronic devices. Concerning the interface circuit, the SECER (Synchronized Electrical
Charge Extraction Regulator) circuit is usually required to rectify the alternating
current (AC) signal into a direct current (DC) signal. In this paper, the properties
of the SECER circuit and the influences on the energy recovery performance are
analyzed and presented by power visualization simulation in the case of low load
and high source frequency, as well as in the case of high load and low frequency.
and whose characteristics can significantly influence the energy harvesting. It can
be seen that the harvesting energy has a close relationship with the characteris-
tics of the SECER circuit components. This study discloses the SECE influences
on piezoelectric energy harvesting performance. In addition, the results show that
the nonlinear SECER technique is 20% more efficient than the standard circuit, in
terms of maximum power and bandwidth, for generators characterized by a moderate
electromechanical coupling coefficient.

Keywords Circuit SECER · Piezoelectric materials · Energy harvesting · LT1764

circuit

1 Introduction

Recent advances in microelectronics and ambient energy harvesting now make it

possible to envisage the design of completely autonomous electronic systems [1–5].
This possibility, combined with a strong demand for autonomous sensors from the
industrial and biomedical sectors [6–8].

Y. El Hmamsy (B) · C. Ennawaoui · I. Najihi · A. Hajjaji

Laboratory of Engineering Sciences for Energy (LabSIPE), National School of Applied Sciences
EL Jadida, Chouaib Doukkali University, El Jadida, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1517
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_138
1518 Y. El Hmamsy et al.

Fig. 1 SECE circuit

Non-linear energy extraction techniques have been developed to simply optimize

the energy transfer between a piezoelectric generator and the load to be powered
[9–13]. It is then a question of valorizing the extracted energy by associating a
circuit allowing the shaping and storage of the mechanical energy converted into
electrical energy [14–18]. The common point between these approaches is a non-
linear treatment of the piezoelectric voltage carried out synchronously with the
vibration [19–21]. As well as approaches proposed by other research teams [21–
23]. SSHI (Synchronized Switch Harvesting on Inductor), SMFE (Synchro-
nized Magnetic Flux Extraction), OSECE (Optimized Synchronized Electrical
Charge Extraction) and FTSECE (Frequency Tuning Synchronized Electrical
Charge Extraction).
The SECE circuit, shown in Fig. 1, extracts all of the electrical charges generated
at each voltage end at the terminals of the piezoelectric element [11]. It improves the
power generated in the case of weakly coupled structures [17].
The SECE technique is a proposal to improve the behavior of the nonlinear energy
extraction circuits in the case of highly coupled structures.
In this paper, we will focus on the simulation of the energy harvesting by the
SECER technique based on an integrated circuit LT1764, and the comparison with a
simple rectifier bridge. It is important to specify that the performance for an optimal
load. The simulation of the simple rectifier could thus give much poorer results than
the SECER performance because of the high dependence of the energy harvesting
on the electrical load, which implies the necessary use of a SECER circuit is much
more tolerant to load variations, which makes these practical performances near the
theoretical simulations.
This paper is organized as follows. Section 2 introduces the theoretical model
of the technique SECE. Section 3 provides a circuit SECER study. Then, Sect. 3
piezoelectric modeling and simulations the circuit SECE. Finally, conclusions from
this study are presented in discussions.
The Synchronized Electrical Charge Extraction Regulator … 1519

2 The SECE Technique

The energy extracted per cycle in the SECE technique corresponds to the electrostatic
energy available on the piezoelectric element, which equals (01), with VM being the
voltage at the instant of switching (Fig. 1). Due to the losses in the SECE converter,
the energy recovered by switching is given by (02), with γC the efficiency of the
converter (considered constant) [11].

  1
Welectr ostatique max
= C0 VM2 (1)
2
1
(W S EC E )max = γC C0 VM2 (2)
2
The calculation of the voltage VM is done considering that the voltage of the piezo-
electric element is reduced to zero when the displacement is minimum or maximum.
The piezoelectric insert then being in open circuit until the next voltage extremum,
VM is therefore expressed according to relation (3). The expression of the energy
recovered by switching is then given by (4), and we therefore deduce the power
recovered by the SECE, the value of which is given by relation (5). These two quan-
tities are independent of the rectified voltage and the load, and 4γC times greater
than the maximum energy or power recovered with the standard technique.
α
VM = 2 uM (3)
C0

α2 2
W S EC E = 2γC u (4)
C0 M

α2 2
P S EC E = 4f0 γC u (5)
C0 M

3 Circuit SECER (Synchronized Electrical Charge

Extraction Regulator)

3.1 Piezoelectric Modeling

We consider the piezoelectric as an alternating current generator whose amplitude is

proportional to the strain rate. The piezoelectric has a significant dielectric constancy
that we take into account by the parallelization of a CPZ capacitor. The model is
therefore the following.
1520 Y. El Hmamsy et al.

Q = α piezo u (6)

dU
I = α piezo = α piezo V (7)
dt
With Q the displaced charge, I the generated current, u the displacement of the
piezoelectric membrane, V the speed of the membrane and α electromechanical
constant [24].
We consider the piezoelectric as an alternating current generator whose ampli-
tude is proportional to the strain rate. The piezoelectric has a significant dielectric
constancy that we take into account by the parallelization of a CPZ capacitor. The
model is therefore the following.

3.2 Description of LT1764 Circuit

The LT1764 is a low dropout regulator optimized for fast transient response. The
device is capable of supplying 3 A of output current with a dropout voltage of 340 mV.
Operating quiescent current is 1 mA, dropping to <1 μA in shutdown. Quiescent
current is well controlled; it does not rise in dropout as it does with many other
regulators. In addition to fast transient response, the LT1764 has very low output
voltage noise which makes the device ideal for sensitive RF supply applications.
Output voltage range is from 1.21 V to 20 V.
The LT1764 regulators are stable with output capacitors as low as 10 μF. Internal
protection circuitry includes reverse battery protection, current limiting, thermal
limiting and reverse current protection. The device is available in fixed output voltages
of 1.5 V, 1.8 V, 2.5 V, 3.3 V and as an adjustable device with a 1.21 V reference
voltage (Fig. 2).
The LT1764 series are low loss 3 A voltage regulators optimized for fast transient
response. The devices are capable of to supply 3 A at a drop voltage of 340 mV. The

Fig. 2 Block LT1764 circuit in LtSpice

The Synchronized Electrical Charge Extraction Regulator … 1521

Fig. 3 SECER circuit built in LtSpice

Table 1. Value of circuit

Parameter Value/model
proposed components
Q3 2N2905A
Q4 2N2222
R3 2.08 M
R7 200 K
R8 and R9 1 K
Diode 1N5817/BAT41
C3 33.74 μF
C4 1 μF
C5 2 μF
C6 33.74
L 2.8 mH
R6 600 K

bottom the operating quiescent current (1 mA) drops to less than 1 μA at standstill.
In addition to the low quiescent current.
The LT1764 controllers incorporate several protection functions making them
ideal for use in battery powered systems. The units are protected against the two
inverted inputs. and reverse the output voltages. In backup battery applications where
the output can be maintained by a backup battery when the inlet is pulled down to
earth, the LT1764-X acts as a it has a diode in series with its output and prevents
inversion current flow. In addition, in dual-powered applications where the regulator
load is returned to a negative supply, the output can be pulled under the floor up to
20 V and always let the unit start and run.
1522 Y. El Hmamsy et al.

Fig. 4 Power graph of SECER circuit and STD circuit (Standard) as a function of time

3.3 Simulation Model

In this part, we will study the SECER circuit making it possible to recover the energy
delivered by the piezoelectric (Fig. 3).
Instead a piezoelectric element we will replace a sine current generator of
maximum current I = 1 mA of frequency F = 105.3 Hz, resistance RL = 600 K
and the filter capacitor is 1 μF [13]. Using the LtSipce software, we simulate the
SECER circuit for 1 s. All used parameters are described in the Table 1 below.
The Fig. 4 shows the evolution of power in terms of time. The power increases
up to a maximum value of 400 μW for 0.5 s. In addition, SECER circuit power is
much greater than STD circuit power during this period.

3.4 Power Variation as a Function of Frequency

The Fig. 5 shows the evolution of the power proportionally to a frequency, at during
10 Hz the dissipated power (Pdi) exceeds the output power (Pout), at 50 Hz we notice
the convergence of the powers, and at 100 Hz the capacities become almost equal,
this explains the approach of the resonant frequency.
The Synchronized Electrical Charge Extraction Regulator … 1523

8.0E-02

7.0E-02

6.0E-02 Pout Pdi Pin

5.0E-02
Power (W)

4.0E-02

3.0E-02

2.0E-02

1.0E-02

0.0E+00
10 Hz 50 Hz 100 Hz 150 Hz
Frequency (Hz)

Fig. 5 Power variations as a function of frequency

3.5 Power Variation as a Function of Load

Figure 6 shows the power variation as a function of the load in two cases (50 Hz,
100 Hz), it also shows that the power is almost stable. Therefore, the SECER circuit
improves the power generated in the case of weak structures, the power of the SECER
technique remains almost constant, but the energy extracted and recovered has the
advantage of remaining constant whatever the load.
G (SECER): the gain of a SECER circuit
G (STD): the gain of a standard circuit.
Distribution of the powers recovered and extracted in the case of vibrations of
constant amplitude at the optimum load, normalized against the maximum power in
the standard technique is the SECER technique. Note also that, despite the fact that
the STD circuit has the same parameters of the input of the SECER circuit. The gain
of the STD circuit is lower when the circuit connected to the optimum load (Fig. 7).
However, when the circuits are operating in the 150 Hz frequency range, the STD
circuit gain increases relatively to the SECER circuit gain.

3.6 Power Variation as a Function of Load and Frequency

Through the curve it is found that the power changes by changing the frequency and
on the other hand, while the power remains stable despite the change of resistance.
The SECER technique also has the great advantage that it is always adapted whatever
the load resistance, allowing the recovery of non-negligible energy even on relatively
low resistances (Fig. 8).
1524 Y. El Hmamsy et al.

1.2E-03
100 Hz
100 Hz

Power (W) 8.0E-04

4.0E-04

0.0E+00
20 Kom 50 Kom 100 Kom 150 Kom 200 Kom
Load (KΩ)
2.8E-06
50 Hz
50 Hz

2.1E-06
Power (W)

1.4E-06

7.0E-07

0.0E+00
20 Kom 50 Kom 100 Kom 150 Kom 200 Kom
Load (KΩ)

Fig. 6 Power variations as a function of load

20% 19%
18% 16%
15%
16%
14%
12%
10%
7% 7% 7% 8%
8%
6% 5%
4%
2%
0%
10 Hz 50 Hz 100 Hz 150 Hz

G(SECER) G(STD)

Fig. 7 Gain of output power

The Synchronized Electrical Charge Extraction Regulator … 1525

1.6E-02
1.4E-02
1.2E-02
Power (w)

1.0E-02
8.0E-03
6.0E-03
1000 Kom
4.0E-03 800 Kom
2.0E-03 600 Kom
0.0E+00 400 Kom
10 Hz 100 Kom
50 Hz
100 Hz
150 Hz
Frequency

Fig. 8 Power variation as a function of load and frequency

Therefore, SECER circuit allows improving the power generated in the case of
weak structures, to improve the bandwidth, and to provide a power theoretically
independent of the electrical load (independent of VDC ).

4 Conclusions

In this paper, a SECER circuit structure is studied and used as interface for piezoelec-
tric system to enhance the performance of the piezoelectric system. The basic working
principles of the piezoelectric structure and several frequently used interfaces, the
influences of resistive load and source frequency,
In the case of SECER, the ripple rate is a priori independent of the parameters
of the active element, but strongly depends on the inductance of the converter. The
harvesting power of SECER is determined by γC , with the intermediate reversal
factor being mainly determined by Cint and L.
The SECER technique presents a better harvesting performance with a load-
independent, the maximum powers are obtained on load resistances distributed
around the optimum load as standard. On the other hand, although exhibiting a
lower power gain, allows energy to be recovered regardless of the load connected to
the micro generator.
In addition, the results show that the SECER non-linear technique is more efficient
than the use of a simple rectifier, in terms of maximum power and bandwidth, for
generators characterized by an electromechanical coupling coefficient.
1526 Y. El Hmamsy et al.

References

1. Zhang Y, Bowen CR, Ghosh SK, Mandal D, Khanbareh H, Arafa M, Wan C (2019) Ferroelectret
materials and devices for energy harvesting applications. Nano Energy 57:118–140
2. Ennawaoui C, Lifi H, Hajjaji A, Elballouti A, Laasri S, Azim A (2019) Mathematical modeling
of mass spring’s system: hybrid speed bumps application for mechanical energy harvesting.
Eng Solid Mechanics 7(1):47–58
3. Ennawaoui C, Lifi H, Hajjaji A, Samuel C, Rguiti M, Touhtouh S, Courtois C (2019) Dielectric
and mechanical optimization properties of porous poly (ethylene-co-vinyl acetate) copolymer
films for pseudo-piezoelectric effect. Polym Eng Sci 59(7):1455–1461
4. Cao Y, Li W, Figueroa J, Wang T, Torres D, Wang C, Sepúlveda N (2018) Impact-activated
programming of electro-mechanical resonators through ferroelectret nanogenerator (FENG)
and vanadium dioxide. Nano Energy 43:278–284
5. Wilson SA, Jourdain RP, Zhang Q, Dorey RA, Bowen CR, Willander M, Johansson C (2007)
New materials for micro-scale sensors and actuators: an engineering review. Mater Sci Eng R
Rep 56(1–6):1–129
6. Malki Z, Ennawaoui C, Hajjaji A, Najihi I, Eljouad M, Boughaleb Y (2020) Pedestrian crossing
system for the mechanical energy harvesting using piezoelectric materials. In: IOP conference
series: materials science and engineering, vol 948, no 1, p 012030. IOP Publishing, November
7. Zhukov S, von Seggern H, Zhang X, Xue Y, Ben Dali O, Pondrom P, Kupnik M (2020)
Microenergy harvesters based on fluorinated ethylene propylene piezotubes. Adv Eng Mater
1901399
8. Hajjaji A, Guyomar D, Pruvost S, Touhtouh S, Yuse K, Boughaleb Y (2010) Tempera-
ture/electric field scaling in Ferroelectrics. Physica B 405(13):2757–2761
9. Rjafallah A, Hajjaji A, Guyomar D, Kandoussi K, Belhora F, Boughaleb Y (2019) Modeling
of polyurethane/lead zirconate titanate composites for vibration energy harvesting. J Compos
Mater 53(5):613–623
10. Hajjaji A, Guyomar D, Touhtouh S, Pruvost S, Boughaleb Y, Rguiti M, Benkhouja K (2010)
Nonlinearity and scaling behavior in a soft lead zirconate titanate piezoceramic. J Appl Phys
108(6):
11. El hmamsy Y, Ennawaoui C, Hajjaji A, Boughaleb Y (2020) Theoretical study and Simulation
method for optimizing the performance of advanced Energy Harvesting techniques. In: IOP
conference series: materials science and engineering, vol 948, no 1, p 012014. IOP Publishing,
November
12. Guyomar D, Lallart M (2011) Recent progress in piezoelectric conversion and energy
harvesting using nonlinear electronic interfaces and issues in small scale implementation.
Micromachines 2(2):274–294
13. Lallart M, Garbuio L, Petit L, Richard C, Guyomar D (2008) Double synchronized switch
harvesting (DSSH): A new energy harvesting scheme for efficient energy extraction. IEEE
Trans Ultrason Ferroelectr Freq Control 55(10):2119–2130
14. Nechibvute A, Chawanda A, Luhanga P, Akande AR (2012) Piezoelectric energy harvesting
using synchronized switching techniques. Int J Eng Technol 2(6):936–946
15. Garbuio L, Lallart M, Guyomar D, Richard C, Audigier D (2009) Mechanical energy harvester
with ultralow threshold rectification based on SSHI nonlinear technique. IEEE Trans Industr
Electron 56(4):1048–1056
16. Erturk A, Inman DJ (2008) Issues in mathematical modeling of piezoelectric energy harvesters.
Smart Mater Struct 17(6):065016
17. Lallart M, Anton SR, Inman DJ (2010) Frequency self-tuning scheme for broadband vibration
energy harvesting. J Intell Mater Syst Struct 21(9):897–906
18. Lallart M, Guyomar D (2008) An optimized self-powered switching circuit for non-linear
energy harvesting with low voltage output. Smart Mater Struct 17(3):035030
19. Ennawaoui C, Hajjaji A, Azim A, Boughaleb Y (2016) Theoretical modeling of power harvested
by piezo-cellular polymers. Mol Cryst Liq Cryst 628(1):49–54
The Synchronized Electrical Charge Extraction Regulator … 1527

20. Rjafallah A, Hajjaji A, Belhora F, El Ballouti A, Touhtouh S, Guyomar D, Boughaleb Y (2020)

PZT ceramic particles/polyurethane composites formalism for mechanical energy harvesting.
Eur Phys J Appl Phys 89(3):30901
21. Belhora F, Cottinet PJ, Hajjaji A, Guyomar D, Mazroui MH, Lebrun L, Boughaleb Y (2013)
Mechano-electrical conversion for harvesting energy with hybridization of electrostrictive
polymers and electrets. Sens Actuators A 201:58–65
22. Lefeuvre E, Audigier D, Richard C, Guyomar D (2007) Buck-boost converter for sensorless
power optimization of piezoelectric energy harvester. IEEE Trans Power Electron 22(5):2018–
2025
23. Belhora F, Hajjaji A, Mazroui MH, El Fatnani FZ, Rjafallah A, Guyomar D (2015) Energy
harvesting using hybridization of dielectric nanocomposites and electrets. Polym Adv Technol
26(6):569–573
24. El hmamsy Y, Ennawaoui C, Hajjaji A (2021) Study and design the circuit for piezoelectric
vibration energy harvester to charge a datalogger. Int J Eng Appl Phys 1(1):18–25
Calculation of the Magnetic Field
in the Vicinity of the Overhead
Transmission Lines

S. Houicher and R. Djekidel

Abstract The purpose of this study used image method and superposition theorem
for plot the lateral profile of the magnetic fields around three-phase overhead power
lines based on current return on ground, the analytique results we obtained by
MATLAB numerical software which makes it possible to better analyze and to repre-
sent the 2D transverse profile of the magnetic induction under overhead power lines.
the research works show that strength of the magnetic field was in the middle phase
center of circuit line is very important and decrease rapidly of increase the lateral
distance of conductors. The analysis results can be used to determine the magnetic
field intensity generated by high-voltage power lines and evaluate the factors effects
in the magnetic induction.

Keywords Overhead transmission line · High voltage · Magnetic field

1 Introduction

Overhead transmission power lines high voltage, travel long distances and generally
cross regions that are very different in their satisfaction (topography of the land,
altitude, etc.), their climate (temperature, pressure, wind, etc.), their environment
(industrial zones, coastal regions, etc.). As a result, utility of electrical power trans-
mission is exposed to various forms and degrees of stress [1]. The increase in the
human population in the world has resulted the increase demand for the consumption
of electrical energy, which requires the construction of new transmission lines (HV
and VHV) of electricity. The latter has become an essential partner in our daily life,
our technological and economic development [2]. Very low frequency electromag-
netic fields generated by power transmission lines have raised serious questions about

S. Houicher (B) · R. Djekidel

Department of Electrical Engineering, LACoSERE Laboratory, University of Ammar Telidji,
03000 Laghouat, Algeria
R. Djekidel
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1529
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_139
1530 S. Houicher and R. Djekidel

the possible biological effects on human health and the environment associated with
the resulting high levels of the intensity of these fields. HV and VHV overhead trans-
mission lines generates magnetic fields which can reach an intensity of 0.4 μT at a
distance of 100 m (400 kV) or at 30 m in (225 kV). The precise evaluation of electric
and magnetic fields produced by high voltage overhead power lines is very important
in several research areas, necessary in many applications in the design, maintenance
and operation of electrical and electronic equipment [3, 4]. There are various exper-
imental and simulation projects studies on electromagnetic characteristics of the
overhead transmission line. The majority of work is based on the assessment and
comparison of the product field under high voltage power lines with international
standards, there are some research studies have also been conducted on decreasing the
field intensity of electric power lines by the technical solutions. A very good modeling
of the circuit lines relies on several parameters, such as: the characteristics of the
electric cables, the height of conductors, the geometric and structural characteristics
of pylons, the electric frequency, the distance between phase conductors and the envi-
ronmental characteristics [5]. Exposure of the magnetic fields intensity generated by
high voltage transmission lines were estimated to raising cancer. For around twenty
years, the environment has been the subject of social concerns, whose exposure
to the electromagnetic fields (EMF) is perhaps the most significant fear expressed
by the human population. From view point of environmental impact, the magnetic
flux density around overhead power line must be measured. It is well understood the
geometric properties of extra high voltage lines (EHV) influences the magnetic induc-
tion and profiles. With the increase power demand, the power transmission utilities
increase the transmission voltage levels for effective magnitude power transmission.
The interference of these low frequency (50 Hz) electromagnetic fields with human
organisms (working personnel and general public) may cause a menace. The subject
of this assumption is discussed in several research projects over the last few decades
[6]. In this work, the MATLAB program for magnetic field calculation was created,
this program was designed to handle various line configurations. For this, a model
including the partitioning of two conductors in each phase, and the effect of catenary
described by the line conductors between towers was considered. Our aim is simu-
lation the lateral distribution of the magnetic field intensity under overhead power
line (400 kV) horizontal configuration in 1 m at the ground level by application
of Garson method, in order show the factors effecting the magnetic induction such
as the loading current, the height of conductors, separation distance between phase
conductors, observed point height above the ground and arrangement of the circuit
line. This paper organizes four parts, the first is introduction. the second defines theo-
rical formulas for calculation of power line magnetic fields based in Garson method
and superposition principle. the third provides the simulation results and discussion;
the fourth is conclusion.
Calculation of the Magnetic Field in the Vicinity … 1531

2 Calculation Method of the Magnetic Flux Density Under

Three-Phase Overhead Power Line

We have proposed in the part of work the calculations performed on an overhead

power line horizontal configuration, at height of 1 m above the ground. the circuit
line is equipped with bundled conductors two per phase, and two ground wires. the
phase conductors and ground wires are assumed to be parallel to a ground plane.
the sag is taken into account by using an average height of conductors. Due to low
frequency, it is possible to use the quasi-static methods independently to evaluate the
electric and magnetic fields. [7, 8]. Magnetic induction is calculated by using image
method and superposition theorem. the magnetic permeability of free space is taken
to be the value: 4 π.10−7 H/m. the magnetic field above ground is equal to the line
source magnetic field plus a “complex image”, A simple application of the Ampere’s
law calculates the value of magnetic field intensity around a single conductor, and
then we apply the superposition theorem of the three partial fields for get the total
magnetic field. To calculate the magnetic flux density near high voltage transmission
line, we consider a long overhead power line with electric current and the conductors
are cylindrical, the radius of conductors is weak compared to its length and its height
above earth.
As a shown in Fig. 1, electric current intensity (I) in the point (xi, yi) created
at observation point p(x, y) of space a magnetic field, the magnetic flux density is
defined by the following relation:

μ0 .I
B = μ0 .H = (1)
2.π.r
Where: μ0 is the magnetic permeability of vacuum (N/A2).
We can use the theory of conductor images in taking into account the depth of
penetration. The magnetic field produced by the line induces return currents in the
earth, thus creating a magnetic field composed to that generated by the line. Precise

Fig. 1 Magnetic intensity

generated by real current and
its image flowing in the
conductor
1532 S. Houicher and R. Djekidel

calculation of magnetic induction requires the use of Garson method, which is based
on the notion of return currents through the ground [9]. We consider a fictitious plane
of return currents, parallel to the ground plane and placed at the penetration depth
at a complex distance equal to α (see Fig. 1). Its relation is given by the equation
below:
√
α= 2.δ.e− jπ/4 (2)
√
Where, δ = ρ/μ0 .π. f : depth of penetration, and ρ: the electrical resistivity
of ground (.m), and f: the frequency in Hz. In the simulation, the distribution of
magnetic fields under overhead transmission line is obtained for a balanced system
in which the currents in phase conductors are symmetric and in direct succession at
a phase angle ϕ of 120° from one to the other. the electric current module of the line
is 1000 A, the current of guard cable is neglected and the earth it is assumed that to
be homogeneous with a resistivity of 100 .m. the frequency is f = 50 Hz.
The horizontal and vertical compenents of the magnetic field are calculated by
using the following equations:
 ⎫
μ0 y − hn y + hn + α ⎪⎪
Bxn = − .In 2
− 2 ⎪
⎬
2.π rcn rin
  (3)
μ0 x − dn x − dn ⎪⎪
B yn = .In − ⎪
⎭
2.π 2
rcn 2
rin

Where, (x, y) are the positions of the observation point above ground and (dn , hn )
are the locations of the phase conductors respectively, rcn is the distance between
each conductor and observed point above ground, rin is the distance between each
image conductor and observed point.
The components of magnetic field intensity due to conductor can be written
by adding the two components vertical and horizontal of the given field for each
conductor are given in the expression below:
⎫
N =3
 ⎪
⎪
Bx = Bxn ⎪
⎪
⎪
⎬
n=1
(4)
N =3
 ⎪
⎪
⎪
By = B yn ⎪
⎪
⎭
n=1

The resultant magnetic flux density is calculated according to this equation:

Bt = Bx2 + B y2 (5)
Calculation of the Magnetic Field in the Vicinity … 1533

Fig. 2 Single circuit very

high voltage transmission
line

Table 1 Parameters used for

Parameter Values
overhead power circuit line
Height of conductor 13 m
Radius of conductor 1.34 cm
Distance of phase 10 m
Height of guard wire 21 m
Radius of guard wire 0.7 cm

We consider a three-phase high voltage overhead power line having the arrange-
ment and the geometric coordinates shown in the figure presented below, the hori-
zontal configuration it is a type of circuit lines, the phase conductors are arranged
on the same horizontal plane, that is to say the same height above the ground. it
is usually used for electric power transmission lines. We have three phases, the
horizontal bundle of two conductors contains in each phase are separated by 0.4 m
between them, the radius of conductor is 1.34 cm. According to the coordinates
system presented in Fig. 2, the overhead power line conductors are horizontally posi-
tioned at −10 m, 0 m, +10 m. Conductors vertical separations (heights of conductors)
are 13 m constant between phase conductors and ground, the radius of guard cable
is 0.7 cm positioned at −7 m and +7 m horizontal, and 21 m above the ground,
The span of the circuit line is of 300 m. The schema of overhead transmission line
400 kV used in this study, with the arrangement and geometric is shown in Fig. 2.
Its parameters are presented in Table 1.

3 Results and Discussion

For the simulation of magnetic field intensity we found the following results are illus-
trated the lateral profile of highest magnetic induction under three-phase overhead
transmission line. Figures 3(a, b) show the lateral profile of the magnetic induction
1534 S. Houicher and R. Djekidel

Fig. 3 Lateral profile of the 15

Bx
magnetic induction By
calculated in 1 m from the Bt
ground level under 400 kV

Magnetic induction[µT]
overhead transmission lines 10
used by following: a Garson
method, b Biot-savart law.

0
-80 -60 -40 -20 0 20 40 60 80
Lateral distance[m]

(a)
14
Bx
By
12
Bt
Magnetic induction[µT]

10

0
-80 -60 -40 -20 0 20 40 60 80
Lateral distance[m]

(b)

distribution in 1 m from the ground level, On observe in two Figures the lateral
distribution of horizontal, vertical and total components (Bx, By, Bt) of the magnetic
induction, and also the magnetic flux density increases in symmetrically contin-
uous manner for significative decreases of the lateral distance of the high voltage
overhead power line.
For Garson method (see Fig. 3a) the maximum total value directly under middle
phase conductor(about 12.44 μT) register at the center distance(x = 0 m), accounting
for only 12.44% of the exposure limit established by ICNIRP for the general public
100 μT, then reduces rapidly in symmetric way for increase the lateral distance
when one moves away from the conductors, to reach at distance 75 m, far from the
conductors a value is about 0.61 μT almost negligible, On the other hand, For the
comparison results by Biot-savart law (see Fig. 3b) the maximum total value directly
under middle phase conductor (about 12.46 μT) register at the center distance (x =
0 m), accounting for only 12.46% of the exposure limit established by ICNIRP for
the general public 100 μT, then reduces rapidly in symmetric way when one moves
Calculation of the Magnetic Field in the Vicinity … 1535

away from the conductors, to reach at distance 75 m, far from the conductors a value
is about 0.51 μT almost negligible, The calculated magnetic field strength is within
the acceptable range right of way in the corridor high voltage three-phase overhead
power line.
In Fig. 4 plot the lateral profile of magnetic field intensity as a function for different
values of loading current of phase conductors 500, 1000, 1500, and 2000 A. It is clear
that as this curve is symmetric in the two sides positive and negative of the circuit
line, it is show the currents increase, the magnetic induction increase until to reach
a maximum values in all this four values of current in the center distance under
the middle phase conductors, then decrease rapidly with significative increase of the
lateral distance, to reach the low values when one moves away from the conductors,
a very far for the circuit line, this is linearity relationship between them more the

Fig. 4 The magnetic field 30

intensity as a function of I=500A
I=1000A
various values of load current 25 I=1500A
I=2000A
Magnetic induction[µT]

20

15

10

0
-80 -60 -40 -20 0 20 40 60 80
Lateral distance[m]

Fig. 5 Variation of the 15

H=12m
lateral profile of magnetic H=13m
induction for various heights H=15m
Magnetic induction[µT]

H of conductor H=18m
H=20m
10

0
-80 -60 -40 -20 0 20 40 60 80
Laterale distance[m]
1536 S. Houicher and R. Djekidel

current is higher, the magnetic induction intense but on observe in the Fig. 5 the
lateral profile of magnetic field intensity as a function for different heights H of the
phase conductors, on illustrate the height H of conductors increases, the magnetic
induction decreases in symmetric manner in the two sides positive and negative of
the circuit line, for significative increase of the lateral distance, to reach the low
values when one moves away from the conductors, a very far for the circuit line,
this is opposite relationship between the conductors height and the magnetic field.
Increasing the line height is the most effective parameter in line design, which is
decrease the field strenght under the high voltage overhead power line.
In Fig. 6 plot the magnetic induction lateral distribution as a function various
observed point heights 0, 1, 1.5, 1.8 and 2 m above the ground, it is clear that as the
observation point heights increase, the magnetic induction increases for the lateral
distance of the circuit line increase in all types of configuration, this is symmetrically
manner linearity relation between them in the right of way of the overhead power line.
the magnetic induction values in the vicinity of single circuit high-voltage power line
depending on the distance to several levels starting with level 0 m, on observe a two
dimensional plot of the magnetic induction lateral distribution for this five levels.

Fig. 6 Variation of the 15

z=0m
lateral profile of magnetic z=1m
induction for various z=1.5m
Magnetic induction[µT]

z=1.8m
observation point heights
z=2m
above earth 10

0
-80 -60 -40 -20 0 20 40 60 80
Lateral distance[m]

Figure 7 illustrates the magnetic induction lateral profile in four different phase
configurations: horizontal, vertical, triangular, and inverted triangular circuit line in
1 m from the ground level. By comparing the values obtained, it can be observed that
the magnetic induction values are higher for the horizontal configuration in all points
of the transmission line corridor then decrease rapidly with the distance increase, and
the magnetic induction values for the inverted triangular configuration are lower in
comparison with the magnetic induction values with other configurations (vertical,
triangular). It is very clear that the horizontal circuit line configuration causes the high
values of magnetic induction with a maximum value of 12.98 μT, while the inverted
triangular circuit line configuration obtained the lower values for the magnitude of
the magnetic induction with a maximum value of 6.69 μT in the center distance
of the configuration for the same load current amplitude of 1000 A. This is due to
Calculation of the Magnetic Field in the Vicinity … 1537

Fig. 7 Lateral profile of 14

Horizontal line
magnetic induction Vertical line
calculated in 1 m from the 12 Triangular line
Inverted triangular line
ground level for different

Magnetic induction[µT]
configurations of overhead 10

transmission lines
8

0
-80 -60 -40 -20 0 20 40 60 80
Lateral distance[m]

difference in the height of phase conductors above the ground in each circuit line
configuration.

4 Conclusions

In this article used a method for calculation 50 Hz lateral profile of the magnetic
induction produced by high voltage transmission line. This method enables for a
precise evaluation of all currents in the power line, is based on the theory of images
but where the plan of null potential is located at a complex depth, a program for
calculation the magnetic induction was applied. It can handle any line geometry;
accurate results on the 2D quasi-static analysis of the magnetic flux density generated
by EHV transmission line has been obtained. From the simulation results, it is clear
that the magnetic induction is maximum in the center distance of the power line, and
then decreases with increase in the lateral distance of conductors, It has also been
found that the magnetic field depends on several factors, such as the load current,
the height of conductors, the distance between two adjacent phase conductors, the
observation point height, the configuration of the circuit line: horizontal, vertical,
triangular or inverted triangular line, the horizontal configuration creates the high
magnetic induction values due to all conductors being near to the ground level, on
the other hand the inverted triangular configuration produces the lowest magnetic
field values. The magnetic induction values underneath 400 kV overhead power
lines with same amplitude current of 1000 A does not exceed the value 20 μT ,which
is five times lower of the exposure limit established by ICNIRP for the general public
100 μT.
1538 S. Houicher and R. Djekidel

References

1. Ahmadi H, Mohseni S, Shayegani A (2010) Electromagnetic fields near transmission lines

problems and solutions. J Environ Health Sci Eng 7(2):181–188. https://doi.org/10.1186/1735-
2746-181-188
2. Tourab W, Babouri A, Nemamcha M (2012) Characterization of high voltage power lines as
source of electromagnetic disturbance. In: Proceedings of the 16th international conference and
exhibition on electromagnetic compatibility CEM, Rouen, France, April
3. Radwan RM, Mahdy AM, Abdel-Salam M, Samy MM (2013) Electric field mitigation under
extra high voltage power lines. IEEE Trans Dielectr Electr Ins
4. Lindell IV, Hänninen JJ, Pirjola R (2000) Wait’s complex-image principle generalized to arbitrary
sources. IEEE Trans Antennas Propag 48:1618–1624
5. Amiri HH, Marich M (2006) The influence of sag in the Electric Field calculation around High
Voltage Overhead transmission lines. 2006 annual report conference on electrical insulation and
dielectric phenomena
6. Ismail H (2009) Characteristics of the magnetic field under hybrid ac/dc high voltage
transmission line. Electric Power Syst Res
7. Vujević S, Lovrić D, Sarajčev P (2011) Comparison of 2D algorithms for the computation of
power line electric and magnetic fields. Eur Trans Electr Power
8. Brandao Faria JA IEEE Trans Power Deliv, 4/2007 Publication
9. Ruiz J-RR, Espinosa AG (2009) Magnetic field generated by sagging conductors of overhead
power lines. Comput Appl Eng Educ, May
Implementation of the Vehicle Speed
and Location Monitoring System
to Minimize the Risk of Accident
Mohammed Karrouchi, Ismail Nasri, Hajar Snoussi, Abdelhafid Messaoudi,
and Kamal Kassmi

Abstract Every year, millions of people around the world die from road accidents,
excessive or unsuitable speed is one of the main factors responsible for accidents, and
it is implicated in a third of deaths. The main objective of this paper is to follow the
driver from the speed of the vehicle as well as to know his behavior and his physical
state on the roads to reduce the risk of accidents. In this paper, we monitor the
instantaneous speed of the vehicle using a control system that can be mounted in any
vehicle equipped with the OBD-2 protocol (on-board diagnostics). A monitoring
chain based on a PIC microcontroller will be mounted in the vehicle as well as
other peripherals such as an SD memory card module, a GPS geographic positioning
sensor, a shock sensor and a global system for mobile communications GSM module.
Unauthorized speed and the location data of the vehicle will be recorded in the SD
memory card with accurate date and clock. In the event of an accident, an emergency
message containing the latest speed and position data will be sent through the GSM
module to the nearby medical center and local police as well as to the operator or
owner (in case of car rental).

Keywords Data logging system · Speed excessive · PIC microcontroller · OBD-2

protocol · CAN bus

1 Introduction

Many electric units are integrated into modern cars as well as relevant technologies
such as the anti-lock braking system (ABS), electronic stability program (ESP) and
the forward collision warning, to improve the safety of driver and passengers, and

M. Karrouchi (B) · I. Nasri · H. Snoussi · K. Kassmi

Electrical Engineering and Maintenance Laboratory, High School of Technology,
Mohammed First University, BP. 473, Oujda, Morocco
e-mail: [email protected]
A. Messaoudi
Energy, Embedded Systems and Information Processing Laboratory, National School of Applied
Sciences, Mohammed First University, Oujda, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1539
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_140
1540 M. Karrouchi et al.

also to make life easier for the driver. Despite this change, accident statistics world-
wide keep growing due to bad behavior and driver negligence. The major causes of
accidents is unsuitable or excessive speed [1], it is among the most frequent human
causes, is present in more than one fatal accident in 4. This is shown by the analysis
of the causes of fatal accidents, according to the ministry of transport equipment, that
occurred outside the urban area in Morocco, is presented below in Fig. 1. In order
to reverse this tendency, the government is integrating actions having an immediate
effect and added value on improving road safety, It reflects on the installation of
several devices in vehicles, among them a data logger or a black box, it is a device
that facilitates the provision of information on the accident and also records the last
events before the coincidence to know-how was to create the accident.
Several university researchers have proposed different strategies to reduce the
proportions of accidents. In this paper, we have designed an intelligent node dedicated
to the vehicle, this chain behaves like an electrical control unit (ECU), which receives
data of unauthorized or excessive speed data with location, and records them in real
time on an SD memory card as a punctual speed history. In case of a collision, the
latest events of the car will be collected and sent through a telephone message to
emergency numbers using both GPS and GSM modules. Speed data is collected using
the OBD-2 port in the tests, which are performed on a DACIA Dokker. A hardware
and software design has been validated by simulation and practical experience.
The rest of the paper is constructed as follows: the second section recalls the
concepts of CAN bus and the OBD-2 protocol, on which our work is based. The
third section describes some existing techniques for recording vehicle parameters
before the accident. The fourth section presents our proposed system (Hardware and
software). The fifth section details our practical and simulation results to monitor the
driver and the vehicle. Finally, the sixth section concludes this paper.

% of accident causes
40
excessive speed
30
non respect of priority
20
alcohol drink
10
redness
0

Fig. 1 The main causes of fatal accidents in % outside urban areas (according to ministry of
transport equipment)
Implementation of the Vehicle Speed and Location Monitoring System … 1541

Fig. 2 The CAN bus

structure in the vehicle.

2 Relevant Background

2.1 Controller Area Network (CAN)

CAN (Controller Area Network) bus is a local area network developed firstly for
the field of automotive and industrial automation control. Under the name of multi-
plexing, the CAN is a broadcast serial bus standard for connecting electronic control
units (ECU) peripherals, or microcontroller devices [2]. When a unit fails, it does
not affect others. Each node is able to send and receive messages. Figure 2 shows us
the topology of the CAN module.

2.2 On-board diagnostics (OBD-2)

As shown in Fig. 1, the standard OBD-2 port is connected to the CAN bus, this
protocol is implemented in most of the vehicles which are manufactured lately. The
OBD-2 standard covers only the exchange of diagnostic data. Figure 3 shows the
OBD-2 connector, The CAN bus pins are pin 6 (CAN high) and pin 14 (CAN low).
The OBD-2 standard defines 10 diagnostic services, the first is used to determine
current information from certain vehicle sensors via the request/response approach
(for example temperature, engine speed, vehicle speed…). The OBD-2 message is
the same useful width as CAN protocol, it supports 8 bytes maximum. The format
of the OBD2 message is shown in Fig. 4.
In mode 01 which defines the request for current diagnostic data, there are several
PID codes defining the data that can be acquired. Table 1 shows some PID codes.
The A and B signifying the fourth and the fifth bytes respectively on the response
frame.

Fig. 3 The OBD-2

connector
1542 M. Karrouchi et al.

Identifier Number of
Data Data Data Data Not
Request: 7DF additional Mode PID used
A B C D
Response: 7E8 bytes

Fig. 4 OBD-2 data frame format

Table 1 Some common

Description PID code (Hex) Formula
OBD-2 PIDS codes
Engine speed 0×0C RPM = (256 A +
(Tr/min) B)/4
Vehicle speed 0×0D V=A
(km/h)
Engine oil 0×5C T = A-40
temperature (°C)

3 Related Works

Abhirup Das et al. [3] have shown different causes generating fatal road accidents,
effective solutions to fight against them. With the support of various useful sensors,
they react to the speed of the vehicle and define their location. The first step is the
monitoring of the blinking of the driver’s eyes using an infrared sensor, in the event
of sleep, an alarm is triggered. In the second phase, they used an alcohol sensor to
detect the drunken state of the driver. The third step is the use of a proximity sensor
to automatically control the speed of the vehicle, if the distance between the vehicle
and their neighbor lowered below the threshold, the sensor would activate the motor
control circuit and a speed limit occurs. Using a GSM module, the last phase consists
of sending an accident aid message to the nearest station.
Kumar et al. [4] have demonstrated a solution to ensure vehicle safety using IoT
technology. An automatic warning system is implemented for the caution of the
driver, they monitored the performance of the vehicle using various sensors such
as a breath analyzer, an ultrasonic sensor to measure the distance from surrounding
vehicles and an accelerometer. During vigorous consumption of alcohol by the driver,
a message will be sent to the owner or the vehicle authority. In the event of an accident,
the acceleration (tilt) and location data of the vehicle will be sent to the local police
and the hospital.
In this paper [5], Reeja S and Mr. Jayaraj V S have proposed a data recording
system before and during the accident, they integrated a drowsiness detection system
linked to an alarm to wake up the driver, a camera records the vehicle during the road
for analysis in the event of collisions, an automatic light dimmer for night driving so
as not to blind driver on the opposite path, radio frequency identification (RFID) as
well as a GPS and GSM to share the facts for alerting purposes.
Rekha S and Hithaishi B S [6] have realized a monitoring system dedicated
to the vehicle based on several devices such as flame sensor, Temperature sensor,
Accelerometer sensor and Pressure sensor. In the event that one of the magnitudes
Implementation of the Vehicle Speed and Location Monitoring System … 1543

measured by the sensors exceeds the limits, a message containing the exact location
of the car will be sent to the persons concerned using GPS and GSM.
S Sethuraman and S Santhanalakshmi [7] discussed and produced an Internet
of Things (IoT) model of the vehicle black box system, which mainly consists of
the raspberry pi, to communicate between all devices, such as a driver recognition
camera, a gas sensor to detect if someone is intoxicated, pressure sensor, MEMS tilt
sensor (accelerometer) to determine the tilt position of the vehicle. If the threshold
quantity reaches a certain level of any sensor, using GPS and GSM modules, vehicle
location, sensor data and images will be submitted to the owner.

4 Proposed System

In this paper, we have made our own vehicle speed monitoring system, based on
electronic components that are inexpensive and consume less energy. The speed data
is collected using the OBD-2 protocol to validate our tests, which saves us from using
an additional speed sensor and reducing loads. Our solutions and very efficient to
monitor driver behavior on the roads and reduce collisions due to increased speed.
Our system allows highway police, rental car owners and managers to check the
vehicle’s speed history and location. If the speed exceeds the allowable limits, A
message will be sent to the owner contains instantaneous speed data and location of
the vehicle. In the event of an accident, the nearby police and hospital station will be
informed.

4.1 System Block Diagram

The system consists of a PIC18f4580 microcontroller to act and supervise the commu-
nications between the peripherals, a transceiver to adapt the signals on the bus, an
SD memory card for recording speed data, Sound alarm to warn the driver, a GPS
module to locate the vehicle, an shock sensor to detect collisions, a GSM module
to connect to the telephone network. The use of this series of pic microcontroller
has several advantages such as the integration of a CAN controller to manage the
transmission and reception rules, an analog-to-digital converter to translate an analog
magnitude into a digital value. Figure 5 describes the system architecture proposed
in this paper.
The CAN circuit (the PIC and the transceiver) extracts the speed data through the
OBD2 port, and the location data using the GPS module through the UART commu-
nication (Universal Asynchronous Receiver Transmitter), then broadcasts these
messages to the SD memory card, using serial communication SPI (Serial Peripheral
Interface).The shock sensor is used to determine if an accident has occurred during
the speed boost. The transmission rate on the CAN bus is 500 kb/s, and the frequency
value of the quartz connected to the microcontroller used is 20 MHz.
1544 M. Karrouchi et al.

Fig. 5 Block Diagram of proposed system

4.2 Software Structure

Our program implemented in the microcontroller can be presented in two phases, the
first contains a program whose role is to collect the speed data exceeding the defined
limits, to display them on the LCD screen as well as the spatial coordinates of the
vehicle (Latitude and longitude). This information will be recorded in real-time on
the SD memory card as well as sent to the owner or the rental manager. The second
task is to check if a collision has occurred. In this case, an emergency message will
be broadcast to the emergency number, to the nearby police and hospital stations.
The program was written in the MicroC pro for Pic compiler and loaded into the
PIC using the WinPic software. Figure 6 summarizes the flowchart followed.

5 Implementation and Results

When the vehicle battery turns on, using a voltage regulator (12 V to 5 V) inte-
grated in the electronic circuit of the transceiver MCP255, our main development
board electronic managed by the microcontroller starts to work. The LEDs shows the
state of the voltage generated on the board. Printed circuit boards are designed using
Altium Designer software. Figure 7 describes the 3D design of electronic develop-
ment boards. After several tests were carried out to ensure the correct functioning of
the system, we were able to determine and display the instantaneous speed data of
the vehicle. Figure 8 shows the materials and the work done in this step.
According to our system and using an oscilloscope, we were able to extract
and visualize the electrical signals defining the CAN transmission frame. It mainly
consists of 11-bit identifier (ID = 0x7DF), 64-bit data field, contains mode 01,
PID 0x0C and data (A, B, C, D/F: not used). Figure 9 shows the electrical signals
characterize the speed data request message.
Implementation of the Vehicle Speed and Location Monitoring System … 1545

Fig. 6 Flowchart followed in this work

a b

Fig. 7 a 3D design of PIC18f4580 microcontroller board; b 3D design of transceiver MCP2551

board

Due to the collisions of the many frames circulating on the CAN bus, it is hard to
collect the desirable reception frame, relating to the speed data using the oscilloscope,
since the latter is connected directly to the physical bus of the vehicle.
After this phase, the processing of comparison between the speed data and the
defined limits begins in the microcontroller. The second step is based on a simulation
of the system, which contains the SD card, the GSM and GPS modules connected to
the PIC18F4580 microcontroller on Proteus. It is a software that allows the design
of a virtual electronic system and to simulate it, designed for industrial and educa-
tional use [8]. Proteus integrates the simulation of animated components, circuits and
1546 M. Karrouchi et al.

Fig. 8 Working model

Fig. 9 Capture of data transmission signals on the vehicle’s CAN bus using an oscilloscope

models of microcontrollers or microprocessors, to facilitate the modeling, develop-

ment and testing of projects before creating them in a physical prototyping. Figure 10
shows the simulated system. The first phase consists of formatting the SD card, then
extracting and saving the coordinates (Latitude and longitude) on it. GSM module
sends these data to the specified number.
In this way, we were able to save the excessive speed data on the SD card, locate
the car using the GPS module by determining the attitude and longitude, that we can
use them in Google map and define the location exactly, as well as simulate the GSM
module to send an emergency message.
After these two stages, the speed and location data of the vehicle are recorded in
the memory card. In the event of excitement, an audible alarm will sound to alert
the driver, and information is transmitted automatically to the owner. In the event of
an accident, the shock sensor informs the microcontroller so that it must distribute
the emergency message to the nearby police and hospital stations. The hardware
implementation of the proposed system is shown in Fig. 11. After testing the system,
Implementation of the Vehicle Speed and Location Monitoring System … 1547

Fig. 10 The fundamental scheme of the system and the result of the simulation in ISIS Proteus

Fig. 11 Global hardware

implementation of the
proposed system

we noticed that it continually updates and works fine with the change of data. The
excessive speed data taken as an example is 100 km/h.

6 Conclusion and Future Scope

Exaggerated speed is one of the most common causes of road accidents, this factor
concerns the driver and his behavior. This paper focuses on the design of a surveil-
lance system that determines where the speed of the vehicle is enormous. We used
a hardware platform composed of a PIC18f4580 microcontroller and anmcp2551
transceiver for CAN communication with car, a 12 V to 5 V voltage regulator to
power the system from the vehicle battery, a GPS module to locate the car, a memory
card that records speed and location data, a crash sensor to determine if a collision
1548 M. Karrouchi et al.

has occurred, GSM module to send an important message to the people concerned, a
sound alert to alarm the driver in the event of excitement and an LCD to display the
results. The whole control system has high reliability and can translate into a small
volume. An efficient solution to monitor and know the behavior of the driver, as well
as minimize the risk of accident.
Our work can be improved by the implementation of several devices such as
an alcohol sensor, vibration sensor, we also want to realize a dialogue between the
present control system and a system which monitors the drowsiness of the driver
using a camera device.

References

1. Weigel H, Cramer H, Wanielik G, Polychronopoulost A Saroldit A (2006) Accurate road geom-

etry estimation for a safe speed application, intelligent vehicles symposium, Tokyo, Japan, 13–15
June 2006
2. ISO 11898-1:2015 Road vehicles – Controller area network (CAN) – Part 1: Data link layer and
physical signaling
3. Das A, Ray A, Ghosh A, Bhattacharyya S, Mukherjee D, Rana TK (2017) Vehicle acci-
dent prevent cum location monitoring system, In: 2017 8th Annual Industrial Automation and
Electromechanical Engineering Conference (IEMECON), Bangkok, pp 101–105
4. Anil Kumar M, Venkata Suman M, Misra Y, Geetha Pratyusha M (2018) Intelligent vehicle
black box using IoT. In: International Journal of Engineering & Technology (UAE)-SCOPUS 7
March 2018, no 2.7, pp 215–218
5. Reeja S, Jayaraj VS (2017) An embedded system for traffic rule violation and vehicle crash
analysis using black box, In: International conference on Innovations in information embedded
and communication systems (ICIIECS). IEEE
6. Rekha S, Hithaishi BS (2017) Car surveillance and driver; assistance using blackbox with the
help of GSM and GPS technology. In: International conference on recent advances in electronics
and communication technology (2017)
7. Sethuraman S, Santhanalakshmi S (2020) Implementing vehicle black box system by IoT based
approach. In: Proceedings of the fourth international conference on trends in electronics and
informatics (ICOEI 2020), IEEE Xplore Part Number: CFP20J32-ART. ISBN 978-1-7281-5518-
0
8. Su B, Wang L (2010) Application of Proteus virtual system modeling (VSM) in teaching of
microcontroller. In: 2010 international conference on E-health networking digital ecosystems
and technologies (EDT) (Shenzhen), pp 375–378
Comparison Study of the Resonant
Inductive Power Transfer for Recharging
Electric Vehicles
Naima Oumidou, Ali Elkhatiri, Sara Khalil, Moussa Labbadi,
and Mohamed Cherkaoui

Abstract Inductive power transfer (IPT) systems are designed to efficiently deliver
power from a primary source to one or more secondary loads via magnetic coupling
over relatively large air gaps. The traditional IPT system is improved by adding a
compensation system to compensate for the reactive power due to leakage fluxes
between the coils, thus improving the system efficiency and increasing the transfer
distance. There are different compensation topologies, but the two most suitable for
the electric vehicle charging system are the SS and SP topology. In this paper, a
comparison between Series-Series and Series-Parallel topology is conducted. This
comparison will help to determine the topology that will ensure maximum transfer
efficiency even over long distances. Simulation results are given to demonstrate the
best topology to improve transfer.

Keywords Electric vehicle · Wireless charging · Inductive power transfer ·

Compensation topologies

N. Oumidou (B) · A. Elkhatiri · S. Khalil · M. Labbadi · M. Cherkaoui

Engineering for Smart and Sustainable Systems Research Center, Mohammadia School
of Engineers (EMI), Mohammed V University in Rabat, Rabat, Morocco
e-mail: [email protected]
A. Elkhatiri
e-mail: [email protected]
S. Khalil
e-mail: [email protected]
M. Labbadi
e-mail: [email protected]
M. Cherkaoui
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1549
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_141
1550 N. Oumidou et al.

1 Introduction

Growing environmental concerns such as global warming and air pollution make
it necessary to find a solution to the problem of greenhouse gas (GHG) emissions,
which is worsening the situation. So as one of the major consumers of fossil fuels,
the transportation sector is responsible for a large part of these emissions. Hence
the need to minimize this contribution by opting for electric vehicles as a solution
because of the advantages they offer in terms of performance, emissions, and safety.
However, the acceptability of EVs is linked to their autonomy, purchase cost, avail-
ability, and performance of the charging infrastructure. The objective is, therefore,
to improve EV autonomy without having to increase battery volume, and to have a
high-performance, available recharging infrastructure. It is in this context that the
idea of using wireless recharging as a solution to achieve the desired objectives was
born.
Wireless charging is one of the three methods proposed as alternatives for charging
electric vehicles in addition to battery exchange and conductive charging. The battery
exchange allows us to replace the empty battery with a full one in a fast way and
allows a flexible charging time, but it only minimizes the time. As for the conductive
charge that is currently used, it is cost-effective and feasible but has the disadvantage
of a long charge time and also shows some safety problems related to the use of the
cables.
For the wireless charge, it is in the spotlight thanks to its advantages over the
conductive charge. It is one of the five key technologies mentioned in [1] that will
enable a transition to electric mobility and will disrupt the transportation industry
while eliminating the need for cables and the problems of independence anxiety.
Authors in [1] also highlights the importance of integrating sustainable electricity
sources such as solar and wind power into the charging process to make electric
vehicles truly green.
The wireless power transfer technology (WPT) is then a feasible solution that
will boost the range of the vehicle while minimizing the volume of the battery and
consequently the cost. There are different vectors for wireless energy transfer which
are classified as far-field WPT and near-field WPT as in [2]. For far-field WPT it
allows the transfer of power by electromagnetic radiation (laser, microwave, and
radio waves), it is a technique that allows the transfer of high power over long
distances. While WPT in the near field allows power transfer over short distances
using inductive coupling or capacitive coupling.
The two most common wireless transfer techniques are inductive transfer (IPT)
and capacitive transfer (CPT), which have common applications such as charging
some mobile devices and low-power biomedical devices [3].
Currently, this capacitive coupling is used to transfer low powers (a few tens of
watts) over small distances. However, work is underway to improve this technique
and extend it to the recharging of electric vehicles [4]. So, for the EV application
inductive transfer prevails. It is currently the most mature technology for contactless
recharging of electric vehicles. It allows the transfer of several tens of kilowatts over
Comparison Study of the Resonant Inductive Power Transfer … 1551

short distances (a few tens of centimeters) with good efficiency and without danger to
the human body, if a correctly dimensioned magnetic shield is present. Technique is
attractive for EV applications because of its advantages in terms of maturity, transfer
capacity, transfer distance and security. Some of the advantages of IPT are listed
below [5, 6]:
• The simplicity of connection thanks to the absence of cables to handle.
• Safe and reliable system.
• Low maintenance compared to the current system (conductive load).
• Safe load whatever the climatic conditions (system unaffected by dirt, rain, snow,
etc.).
• System with long product life.
But despite these benefits, it should be noted that there are some health and safety
concerns regarding human exposure to EMF and various research is being conducted
to find exposure limits [7]. Researchers in [8] indicates that the two most common
exposure limits are those published by the International Commission on Non-Ionizing
Radiation Protection (ICNIRPT) and the Institute of Electrical and Electronics Engi-
neers (IEEE). Also, the Ref [1] mention that a standard is under development as part
of SAE-J2594, this standard aims to standardize a small frequency range for light
EV load.
In an IPT system, there are two electrically insulated sides; the transmitter side
called the primary side which is usually fixed or integrated to the ground and the
receiver side called the secondary side located under the vehicle. The power transfer
is achieved via magnetic induction between these two components and the charging
process begins when the transmitter coil and receiver coil are fully or partially aligned.
It can be used in static, dynamic, or quasi-dynamic recharging mode. In the case
of stationary or static charging (SWC), the vehicle is charged while stationary during
long-term parking.
The Dynamic Load Control (DWC) allows the vehicle to be loaded while in
motion. In this mode, a set of primary coils follow one another to form a “continuous
strip” on a track or a portion of the road that will continuously power an electric
vehicle while it is moving. Figure 1 shows an example of one of the commercially
available applications of the DWC; the On-Line EV (OLEV) developed at the Korea
Advanced Institute of Science and Technology (KAIST) discussed in [9], which
confirms that the DWC system is significantly beneficial for battery viewing time

Fig. 1 Illustration of the OLEV power channel unit [9]

1552 N. Oumidou et al.

and can reduce the size of the battery, but points out that the battery could suffer
serious degradation if it is too small.
Another model of the DWC system is presented in [10], but this time with a single
longer transmission coil. This system will allow loading several VEs at the same time.
This mode, also called on-road power supply, is the future-oriented development for
charging electric vehicles and will allow increasing the number of EVs [1]. Different
possible architectures for the DWC can also be found in [11].
For the quasi-dynamic charging mode (QWC) it is beneficial in the case of tran-
sient stops where vehicles stop at regular intervals such as traffic lights, cab ranks or
bus stops [2].
In this paper, the main contribution is to make a comparison between serial-
to-serial and serial-to-parallel topology in order to choose the best compensation
topology. The comparison consists in varying the frequency and the gap and noting
the evolution of the efficiency of the system and then comparing the two systems to
determine the topology that ensures a better transfer efficiency.
The rest of this paper is organized as follows. Section 2 describes the resonant
system and the inductive coupler as a key element of the system and mentions the two
parameters influencing the transfer efficiency. Section 3 presents the two topologies
that are the subject of the comparison and gives expressions of the efficiencies of
each system found from equivalent electrical circuits. Results and their discussion
are provided in Sect. 4. Finally, Sect. 5 ends this paper with conclusions.

2 System Description

It is simply the traditional inductive power transfer system to which a compensation

system has been added in the form of series and/or parallel connection of capacitors
with the winding. This compensation system aims at compensating the reactive power
due to the leakage fluxes between the coils and consequently maximizes the transfer,
improves the efficiency and increases the transfer distance (Fig. 2).

(a) (b)

Fig. 2 Illustration of the traditional inductive power system a [5], resonant inductive power system
b [12] for electric vehicle
Comparison Study of the Resonant Inductive Power Transfer … 1553

Fig. 3 The four basic topologies of compensation

Table 1 Expressions of compensation capacities

SS SP PP PS
Cp 1  1   1  1
ω0 2 L p ω0 2 L p − M
2
ω0 2 L p − M
2 ω0 2 L p
Ls Ls

1 1
Cs ω0 2 L s ω0 2 L s  1   1 
2 2
ω0 2 L s − M
Lp ω0 2 L s − M
Lp

Most of the work cites the four basic compensation circuits as follows: Series-
Series (SS), Series-Parallel (SP), Parallel-Parallel (PP) and Parallel-Series (PS) which
are illustrated in Fig. 3 with the relationships for calculating the primary capacity Cp
and the secondary capacity Cs in Table 1 [7, 8, 13–16].
However, the Refs [7, 8] and [17] present other compensation systems or networks,
namely: Serial-parallel combination, LCL compensation and LCC compensation.
The authors of [18] present an analysis of another topology: the topology of the
double resonant LCCL network.
In what follows, we will content ourselves with our study of the two topologies;
the SS and SP topology, which are more suitable for variable load conditions because
resonance is guaranteed (according to the table of different topologies) [2].
In an IPT system, the inductive coupler constitutes the key element of the wireless
transfer system. It is composed of transmitter coil and receiver coil separated by an
air gap. It is a transformer characterized by the absence of the magnetic circuit;
air transformer. In addition to the two coils a ferromagnetic shielding consisting of
ferrite and aluminum is used in the design of the coupler. This shielding protects
passengers from the risks associated with exposure to magnetic radiation (Fig. 4).
Several geometries can be used to form the inductive coupler. The basic geome-
tries described in the literature use circular or rectangular coils. However, in order
to improve the coupling between primary and secondary, various authors propose
combinations of several windings at the primary and secondary levels. For the wires
used in the design of these geometries, and in order to reduce losses in the windings,
multi-stranded yarns and more precisely Litz wire are used.
1554 N. Oumidou et al.

Fig. 4 Inductive coupler

The transmission efficiency of the inductive wireless power transfer system is

influenced by the mutual inductance M which describes the influence of one magnetic
circuit on another and the coupling coefficient K which represents the percentage of
coupling between the two coils.
The expression of the mutual inductance can be deduced from this; it is Neumann’s
formula [19] which is written in the form:

(1)

With µ0 is the permeability of the vacuum, dl1 and dl2 are respectively the sections
of circuits C1 and C2 and r12 is the distance between sections dl1 and dl2.
In the case of two windings of N1 and N2 turns, the mutual inductance between
the two circuits is written:

μ0 πr1 2
M1−2 = N1 N2 . r2 2 (2)
2 r 2 + Z 2  23
1

μ0 πr2 2
M2−1 = N1 N2 . r1 2 (3)
2 r 2 + Z 2  23
2

Where r1 and r2 represent the winding radius and Z the distance between the two
windings (air gap).
It is related to the coupling factor by the following relationship:

M M
K = =√ (4)
Mmax L1 L2

Figures 5 and 6 show the influence of the increase of the air gap between the primary
and secondary windings on the mutual and the coupling factor.
According to the curves in Figs. 5 and 6, it is clear that the closer the two windings
are to each other, the higher the coupling factor, i.e. the lower the flux losses and the
higher the efficiency of the system. Also, it can be seen that the greater the transfer
distance, the lower the mutual inductance M, which also influences the efficiency of
the system.
Comparison Study of the Resonant Inductive Power Transfer … 1555

Fig. 5 Variation of the mutual inductance M as a function of the air gap

Fig. 6 Variation of the coupling factor K as a function of the air gap

3 Comparison Between SS and SP Compensation Topology

The various works on the WPT for EV have mentioned that the addition of a compen-
sation system improves the efficiency of the system as well as the transfer distance.
In this section we will compare the two so-called adequate compensation topologies;
the SS and SP topology shown in Fig. 7.
From the equivalent circuit of the system Fig. 7 and as in [20], the efficiency of
the system in SS structure is

RL
η SS =   2  (5)
[R2 + R L ] + (ωM)
R1
2 (R L + R 2 ) 2
+ L 2ω − 1
C2 ω

And the efficiency of the system in SP structure as in [20] is:

1556 N. Oumidou et al.

(a) (b)

Fig. 7 (a) WPT system with SS compensation and (b) WPT system with SP compensation

Table 2 System parameters

Parameters Values
L2 38.6 µH
R1 = R2 0, 34 
RL 40 
C2 0.31 µF

RL
ηS P = 
 2 
R1 2
R L + R2 (1 + (R L C2 ω) ) +
2 R L + R2 − R L C2 L 2 ω2 + ω2 (L 2 + R L C2 R2 )
(ωM)2
(6)

In this comparison we will first vary the frequency for a fixed air gap and then we will
vary the air gap for a fixed frequency and we will note the influence of this variation
on the efficiency. Parameters values used in this comparison are shown in Table 2.
With L 2 represent the self-inductance, R1 and R2 the wire resistance and R L the
load.

4 Results and Discussion

In this section, the simulation results are conducted on MATLAB software to show
the variation of efficiency as a function of frequency and air gap. The curves obtained
are shown in Figs. 8 and 9.
According to the curves obtained, the more the transfer distance increases, the
lower the yield. Also, it is noted that the SP topology allows to have a good efficiency
for high air gaps 80% at 40 cm while for the same distance the efficiency of a system
with SS compensation is 46%.
Regarding the influence of the frequency variation, we notice that for the system
with SP compensation by increasing the frequency the efficiency of the system
increases until it reaches a maximum value of 83% for a frequency of 45 kHz and
then begins to decrease.
Comparison Study of the Resonant Inductive Power Transfer … 1557

Fig. 8 Variation in efficiency as a function of frequency

Fig. 9 Variation in efficiency as a function of air gap

While for the system with SS compensation the efficiency increases by increasing
the frequency up to a maximum value of 94% and then becomes stable.
Then we can conclude that according to the results obtained theoretically the SP
topology is the most adequate. This result remains to be verified experimentally.

5 Conclusion

Electric vehicles (EVs) have the potential to achieve environmentally-friendly trans-

portation that reduces the transportation sector’s contribution to greenhouse gases.
However, the acceptability of electric EVs requires finding solutions to overcome
certain problems such as range, purchase cost, availability, and performance of the
recharging infrastructure.
Wireless inductive recharging is a solution to overcome these problems and will
replace the troublesome conductive charge. Traditional IPT systems have a good
1558 N. Oumidou et al.

transfer efficiency but the problem of leakage flux between the coil influences the
transfer quality. So, to minimize these leaks, a compensation system is used in the
form of series and/or parallel connection of capacitors with the winding. Different
topologies can provide the compensation but there are four simple basic topolo-
gies, two of which prevail over the others, the SS and SP topology. The theoretical
comparison between the two has shown that the SP topology ensures good transfer
efficiency even for large distances. This result remains to be confirmed or refuted
later by experimentation.

References

1. Mouli GRC, Venugopal P (2017) Future of electric vehicle charging. In: 19th international
symposium power electronics Ee 2017, Novi Sad, Serbia, 19–21 October 2017
2. Ahmad A (2018) A comprehensive review of wireless charging technologies for electric
vehicles. IEEE Trans Transp Electrif 4(1). https://doi.org/10.1109/TTE.2017.2771619
3. Dai J (2015) A survey of wireless power transfer and a critical comparison of inductive and
capacitive coupling for small gap applications. IEEE Trans Power Electr 30(11). https://doi.
org/10.1109/TPEL.2015.2415253
4. Li L (2020) A family of compensation topologies for capacitive power transfer converters for
wireless electric vehicle charger. Appl Energy 260:114156
5. Ferrieux J-P (2018) Conception d’un système de transfert d’énergie sans contact bidirectionnel
par induction. In: Symposium De Genie Electrique (SGE 2018), Nancy, France, 3–5 Juillet 2018
6. Khutwad SR, Gaur S (2016) Wireless charging system for electric vehicle. In: International
conference on signal processing, communication, power and embedded system (SCOPES)-
2016
7. Patil D (2018) Wireless power transfer for vehicular applications: overview and challenges.
IEEE Trans Transp Electrif 4(1). https://doi.org/10.1109/TTE.2017.2780627
8. Bi Z (2016) A review of wireless power transfer for electric vehicles: prospects to enhance
sustainable mobility. Appl Energy 179:413–425
9. Jeong S (2019) Charging automation for electric vehicles: is a smaller battery good for the
wireless charging electric vehicles? IEEE Trans Autom Sci Eng 16(1). https://doi.org/10.1109/
TASE.2018.2827954
10. Tan L, Guo J Efficiency optimization of on-road charging for electrical vehicles. In: 7th Asia
pacific international symposium on electromagnetic compatibility
11. Gori P-A, Sadarnac D (2017) Sensorless inductive power transfer system for electric vehicles:
strategy and control for automatic dynamic operation. In: 19th European conference on power
electronics and applications (EPE 2017 ECCE Europe)
12. Kalwar KA (2015) Inductively coupled power transfer (ICPT) for electric vehicle charging – a
review. Renew Sustain Energy Rev 47:462–475
13. Aziz AFA Review of inductively coupled power transfer for electric vehicle charging. IET
Power Electr
14. Hammoud A (2017) Charge par induction de véhicules électriques : Analyse du potentiel,
limitations, nouveaux concepts. Le 24 novembre 2017
15. Panchal C (2018) Review of static and dynamic wireless electric vehicle charging system. Eng
Sci Technol Int J 21:922–937
16. Machura P (2019) A critical review on wireless charging for electric vehicles. Renew Sustain
Energy Rev 104:209–234
17. Al-Saadi M, Ibrahim A Analysis and comparison of resonance topologies in 6.6kW induc-
tive wireless charging for electric vehicles batteries. In: The 12th international conference
interdisciplinarity in engineering
Comparison Study of the Resonant Inductive Power Transfer … 1559

18. Jin X (2019) Calculation and analysis of optimal design for wireless power transfer. Comput
Electr Eng 80:106470
19. Raju S (2014) Modeling of mutual coupling between planar inductors in wireless power
applications 29(1).https://doi.org/10.1109/TPEL.2013.2253334
20. Hasanzadeh S (2013) Efficiency analysis of contactless electrical power transmission systems.
Energy Conv Manage 65:487–496
Magnetic Chargers in Electrical Models:
Operating Principle and Efficiency
Analysis of an Inductively Coupled
Power Transfer System

Ali Elkhatiri, Naima Oumidou, Moussa Labbadi, Mouna Lhayani,

and Mohamed Cherkaoui

Abstract Wireless charging can automate the charging of electrical devices due to its
simplicity operation, robust, and safety. The basic principle of wireless power transfer
(WPT) system is based on the induction coupling between the primary (transmitter)
and secondary (receiver) coils. This paper proposes and describes a new efficient
wireless recharging system which will maximize the transfer of energy to the usage
part. Initially, define the operating principle and the way to charge electric vehicle
(EV). (i) two electric recharging models are described and studied; (ii) the equations
and the efficiency of each recharging model are quoted and discussed. On the other
hand, a typical WPT system consists of a circular transmitter coil, two capacitors
for leakage compensation, one in the primary side and one in the secondary side
connected either in series or in parallel has been constructed in this work. Finally,
several parameters have an effect on the energy transfer efficiency such as the value
of the added capacitors, the value of the operating frequency, the choice of the
compensation topology (serie-serie: serie-parallel: parallel-serie: parallel-parallel)
are discussed in this paper.

Keywords Electric vehicle (EV) · Wireless power transfer (WPT) · Coil ·

Compensation · Capacitors · Topology · Efficiency

A. Elkhatiri (B) · N. Oumidou · M. Labbadi · M. Lhayani · M. Cherkaoui

Department of Electrical Engineering, Mohammadia School of Engineering (EMI),
Mohammed V University in Rabat, Rabat, Morocco
e-mail: [email protected]
N. Oumidou
e-mail: [email protected]
M. Labbadi
e-mail: [email protected]
M. Lhayani
e-mail: [email protected]
M. Cherkaoui
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1561
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_142
1562 A. Elkhatiri et al.

1 Introduction

During the last few years, the electrification sector of transport, such as tramways,
trains and electric vehicles, has become one of the most attractive areas of develop-
ment. In a special way the electric vehicle is attracting more and more attention of
the industrialists because of economic and environmental aspects. [1] the use of the
connecting cable for recharging electric vehicles (VE) presents risks for the user, to
solve these problems, the EV can be recharged by using a wireless power transfer
system, in particular the inductive coupler power transfer mechanism (ICPT). In the
existing literature, some approaches have has been proposed for Rechargeable Elec-
tric Vehicles. The implementation of a wireless charging system for electric vehicles
based on inductively coupled power transfer technology, to transfer electrical energy
from the source to the vehicle battery in order to solve the problems of EV charging
[2]. In the paper [3], the authors compare the shapes of the power coils with the intro-
duction of a ferrite core to design an efficient power supply for wireless charging of
EVs. Resonance frequency is a key factor for improving transfer efficiency in a Wire-
less Power Transfer (WPT) system which has a wide application because the wired
media is eliminated. In this paper, an adaptive frequency measurement method is
presented and the influence on the transfer efficiency in the WPT system is explored
in [4]. In this paper [5], the magnetic equivalent circuits and analytical models of
non-contact electrical power transfer systems are developed and evaluated through
experiments. Overall coupling the coefficient and overall efficiency are introduced as
a means of evaluating the performance of the systems. Compensation of capacitors
on the primary and secondary sides of the systems improves the overall coupling
coefficient and overall efficiency. [6] In this paper, a new method of impedance
matching and power distribution using impedance inverters only on the receiver side
is proposed. The mathematical equations of the proposed method are then generalized
for an arbitrary number of receivers and an arbitrary number of repeaters. This paper
[7] presents a state-of-the-art literature review on recent advances in load coupling
design. This document [8] presents the design and implementation of a wireless
energy transfer system for the movement of electric vehicles, as well as an example
of an On-Line Electric Vehicle (OLEV) system. Electric vehicles are charged on
the road using wireless energy transfer technology. The interoperability between the
two coil topologies is studied. The proposed transformer structures with aluminum
shielding meet the human exposure regulations of the International Commission on
Non-Ionizing Radiation Protection (ICNIRP) guidelines as a basis. Finally, exper-
iments have been carried out in [9] to validate the analyses. In [10], an overview
of a typical WPT system has been given. Simple design equations were given to
calculate inductance, capacitance, power, quality factor and coupling coefficient to
optimize the coil design for the electric vehicle application. In addition, a comparison
was made between the most common coil shapes and the effects of changing coil
parameters such as number of turns, pitch and inner and outer radius on coil effi-
ciency were studied. This paper presents a study on wireless charging technology.
The absence of the magnetic core in the magnetic coupler and the transfer operates
Magnetic Chargers in Electrical Models … 1563

with high air gap values, causing leakage and losses that will minimize the system’s
performance. For this reason, we will consider adding two capacitors, one in the
primary part and the other in the secondary part, and increase the frequency of the
system. There are four types of compensation (serie-serie, serie-parallel, parallel-
serie, parallel-parallel). The objective of this study is to select the most appropriate
type of compensation for the wireless charging system by setting the value of the
operating frequency to the frequency of the secondary resonance system and finding
the value of the primary capacitance to minimize the apparent input power and mini-
mize the reactive power of the power supply to maximize the transfer of energy to
the vehicle battery. The paper is organized as follows: Sect. 2 presents an overview
of the principle of wireless recharging. The evaluation of the efficiency of a tradi-
tional magnetic coupler is discussed in Sect. 3. Description of a resonant magnetic
coupler recharging system, and the improvement of their transfer efficiency is given
in Sect. 4. Section 5 discusses this work and the end of it is concluded in the last
section.

2 Overview of Wireless Recharging Principle

The electric vehicle essentially consists of a battery system that acts as a fuel reservoir
and an electric motor. Recharging the battery of the electric car is done according to
the following procedures [11]: Charging station on the road, near shopping centers or
car parks. A recharging cable is necessary which makes the operation tedious since
the recharging is done in a repetitive way and the cable is always in contact with
the ground and can contribute to get your hands dirty and easy to forget in case of
emergency. So, induction recharging came as a solution to these problems. That is,
we’re going to use a wireless charger to charge the vehicle. Wireless charging allows
electric cars to be charged without any physical contact between the primary coil
located on the ground and a secondary coil located above the vehicle chassis. This

Fig. 1 Basic principle of inductive power Transfer

1564 A. Elkhatiri et al.

solution offers ease of use and good robustness. A general diagram for a wireless
magnetic coupler is shown in (Fig. 1).

3 A Traditional Magnetic Coupler

A traditional magnetic coupler is a wireless charging system that works with two
electrical coils, placed at a distance called the air gap (Fig. 2):
The electric model of a traditional system:

V 1 = R1 i 1 + L 1 ddit1 − M ddit2 (1)

0 = R2 i 2 + L 2 ddit2 − M ddit1 + R L i 2 (2)

with: L 1 , L 2 are respectively the primary inductance and the secondary inductance.
R1 , R2 are the primary wire resistance, the secondary wire resistance respectively
and M is the mutual inductance between the primary and secondary circuits.

3.1 Mutual Inductance

The expression of the mutual inductance M12 for two circular circuits will be:

M 12 = N 2 N 1 μ20 
R12 
3 π R22 (3)
 2
R12 +Z 21

with: N2 = 7 spires are the Number of secondary windings, N1 = 7 spires are the
Number of primary windings, μ0 = 4 π * 10−7 H.m−1 is the Vacuum permeability,
 
R1 = 40 cm et R2 = 40 cm are respectively the radius of the primary spiral and the
secondary spiral. Hence the expression of mutual inductance becomes:
 
N 2 = N 1 = N : R2 = R1 = R : Z1 = Z2 = Z (4)

Fig. 2 Traditional Magnetic coupler

Magnetic Chargers in Electrical Models … 1565

Fig. 3 The variation of the mutual according to the airgap

M 12 = M 21 = N 2 μ20 R4
3 π (5)
(R 2
+Z 2 ) 2

When Z tends towards 0:

M 12 = M 21 = N 2 μ20 Rπ (6)

But, physically speaking, when the two windings coincide, almost all the field lines
emitted by the first conductor pass through the second conductor. This means that,
theoretically, the value of the mutual inductance tends towards the value of the proper
inductance of the coil (Fig. 3).
μ0
L1 = L2 = N 2 Rπ = 38, 68 μH
2
the mutual decreases rapidly if the air gap in the system increases. Therefore, the
magnetic coupling weakens when the distance d exceeds 1 m.

3.2 Efficiency of the Traditional System

According to the electrical model of a complex traditional system (Fig. 4):

V 1 = R1 I 1 + j wL 1 I 1 − j w M I 2 (7)

0 = (R L + R2 )I 2 + j wL 2 I 2 − j w M I 1 (8)

Hence:

ηt = RL
(9)
R1
w2 M 2
((R2 +R L )2 +(L 2 w)2 )+R2 +R L
1566 A. Elkhatiri et al.

Fig. 4 The variation of the efficiency as a function of the airgap

With: R1 = R2 = 0.5 ohm; ηt : Efficiency of the traditional system; w = 2πf:

System Pulse with f = 50 Hz; R L = 37.87 ohm, wire cross-section: S = 40 mm2 .
The increased space between the two coils increases the leakage inductance, which is
why the efficiency of a traditional system decreases rapidly. As a result, the magnetic
coupling between them weakens.

4 A Resonant Magnetic Coupler

A resonant charging system is one of the most well-known and advanced versions
of a traditional charging system, in terms of power electronics design and wireless
transformer coils. In addition, compensation networks are added to the primary and
secondary windings not only to create the resonance phenomenon but also to mini-
mize additional losses. Increasing the gap between the two coils and the absence of
a magnetic core increases leakage and voltage drops in the magnetic coupler. As a
result, the magnetic field generated has a negative effect on the mutual inductance
which decreases the value of the magnetic coupling coefficient. (k) the expression
of k is formulated as follows:

K = √M
L1 L2
(10)

It can be seen that the value of k falls rapidly as a function of the air gap d. Therefore,
if we minimize flux losses, the two circuits (primary and secondary) will be strongly
coupled, the mutual inductance will be higher and vice versa.
Magnetic Chargers in Electrical Models … 1567

4.1 Compensation Technologies

Since the two circuits are separated by a large air gap which results in a low mutual
inductance M. There are four compensation topologies which are illustrated SS SP PS
PP [12]. The reflected impedance Z r can be found by dividing the reflected voltage
by the primary input current [13]:

w2 M 2
Zr = Zs
(11)

Or Z s is the secondary impedance. It depends on the chosen compensation topology.

The series compensated secondary impedance is:

Z s = j wL s + 1
j wC s
+R (12)

For series compensation, the reflexive resistance and reflexive reactance are:

w 4 M 2 C 2S R w3 M 2 C s (1−w 2 L s C s )
Zr = + j (13)
(wC s R)2 +( 1−w 2 L sCs )
2
(wC s R)2 +(1−w 2 L s C s )
2

Besides: The parallel-compensated secondary impedance is:

Z s = j wL s + 1
j wC s + R1 (14)

Similarly:

w2 M 2 R w 3 M 2 [ C s R2 (1−w 2 L s C s )−L s ]
Zr = + j (15)
R (1−w 2 L s C s ) +(wL s )2 R2 (1−w 2 L s C s ) +(wL s )2
2 2 2

Then, the total impedance of the Z t system is formulated as follows:

For a Series-Compensated Primary System

Zt = 1
j wC p
+ j wL p + Z r (16)

For a Parallel-Compensated Primary System: Admittance of the system noted Yt

is:

Yt = 1
Zt
= j wC p + 1
j wL p +Z r (17)

In the procedure for the design of a resonant Inductively Coupled Charging System,
he excites two fundamental hypotheses:—The operating frequency equals the deter-
mined nominal resonance frequency. f 0 = 2π √1C L = √1 —The input current
s s 2π CpLp
is constant. With a changing load there will be a shift in operating frequency So, the
point of secondary compensation is to increase the power transfer capacity. It is
1568 A. Elkhatiri et al.

necessary for the system to operate at the resonance frequency of the compensated
secondary system in series or parallel. The value of the resonance frequency is:
f 0 = 2π √1C L So, to minimize the apparent input power and minimize the reactive
s s
power of the power supply when the system is operated at the secondary resonance
frequency, it is necessary to add a capacitance in the primary circuit by designing the
imaginary part of Z t or Yt equal to zero at the secondary resonance frequency. The
solutions are given in Table 1 for the four topologies.

A simple approach is to select the normalized primary capacity which can be

formulated as follows [14]:
Cp
C pn = Cs Ls (18)
Lp

Depends on: Selected primary and secondary topologies, magnetic coupling coeffi-
cient k and secondary quality Qs (Fig. 5 and Table 2):
The results obtained show that: SS topology is the best topology because the
normalized primary capacitance is independent of the magnetic coupler or secondary
quality Qs.

Table 1 Expression of primary capacity for the four topologies

Topology SS SP PS PP
Cs Cs = 1
Cs = 1
Cs = 1
Cs = 1
w02 L s w02 L s w02 L s w02 L s
Cs L s
Cp Cp = Cp =
Cs L 2s
Cp = Cs L s (L p L s −M 2 )Cs L 2s
Lp
L p L s −M 2 M4 Cp = 2
RCs L s L p +L p M 4 CsLRs +(L p L s −M 2 )2

Table 2 The expression of the normalized primary capacitance for the four topologies for different
values of k
Magnetic Chargers in Electrical Models … 1569

Fig. 5 The variance of C pn as a function of k for different values of Qs.

Fig. 6 System efficiency of SS topology as a function of frequency

4.2 Efficiency of a Resonant Magnetic Coupling:

For Serie-Serie Topology: Based on the electrical model of a resonant magnetic

coupler:
 
V 1 = R1 I 1 + j wL 1 − 1
C1w
I1 − jwM I2 (19)
 
0 = (R L + R2 )I 2 + j wL 2 − 1
C2w
I2 − jwM I1 (20)

Hence the efficiency of the system in SS structure is obtained as follows:

RL
ηss =   [15] (21)
R1
w2 M 2
((R L + R2 ) + L 2 w −
2 1 2
C2w
) + (R2 + R L )
1570 A. Elkhatiri et al.

Fig. 7 Variation of SP topology efficiency as a function of frequency

with: W = 2πf: with f the system frequency. C2 = 0.31 uF: secondary capacity. The
operating frequency f varies between: [3000–1000000] Hertz.
According to the curve, the efficiency of the system increases with increasing the
frequency: At 261 kHz, for M = 1.97e−05, an air-gap d = 30 cm: the value of ηss
is of the order of 94%. Above the 261 kHz the value of ηss remains constant.

For Serie-Parallel Topology

The expression of ηsp is formulated as follow

[15] ηsp= RL
(22)
[R L +R2 (1+( R L C 2 w)2 )]+ w2 M 2 [( R L +R 2 −R L C 2 L 2 w ) +w (L 2 +R L C 2 R 2 ) }
2 2
R1 2 2

The operating frequency f varies between: [3000–1000000] Hertz. C2 = 0.31 uF:

secondary capacity.
To be noticed that the performance of the SP topology increases according to the
frequency of the system: At 45000 Hz the maximum efficiency of the system for M.
1.97e−05 with d = 30 cm k = 0.51 is ηsp = 81%. The value of ηsp drops quickly
to 0 after 45000 Hz.

5 Discussion

The principal conditions for delivering maximum power are: 1-Operation at the
secondary resonant frequency. 2-Deduct the values of the added primary capacitors.
3-Choosing a suitable topology for the wireless charging system. The design of the
system principally corresponds to the parameters necessary for resonance operation
are the inductors and capacitors: f 0 = 2π √1C L = 45.96 kHz. The values of the
s s
primary capacitors selected for the two topologies at Cs = 0.31 uf are: for example:
at M = M = 1.97 × 10−5 , d = 30 cm: C p (uF) (Serie Serie) = 0.31 and C p
(uF) (Serie-parallel) = 0.41. Of the four types of compensation, both Serie-Serie
and Serie-Parallel are suitable for building a wireless charging system: Serie-Serie
topology:—Their normalized primary capacity C pn = 1. It does not depend on
Magnetic Chargers in Electrical Models … 1571

the system settings—primary capacity C p does not depend on the load R L . Serie-
parallel topology:—The SP topology may be preferred for resonance as its normal-
ized capacity for resonance C pn increases with K and does not depend on the load
R L .—primary capacity C p depends Just on the air gap d of the system and is inde-
pendent of the load R L . The two parallel-compensated primaries: parallel-Serie and
parallel-parallel are dependent on the load and their normalized primary capacitance
decreases as a function of the magnetic coupling K. So the two topologies serie-serie
and serie-parallel are suitable for the construction of a resonant magnetic coupler
for the recharging of electric vehicle batteries since their normalized capacities are
not depending on k and the load R L . According to Figs. 6 and 7 the performance of
each topology depends on the frequency of the system: −ηss = 60% at f 0 = 45 kHz
−ηsp = 81% at f 0 = 45 kHz. The performance of the parallel-serie topology is
greater than that found in the serie-serie topology when the system is operated at
the resonance frequency. The Serie-parallel topology is convenient for the charging
system of electric vehicles as it gives us an acceptable efficiency about of 81%. Since
the Serie-Serie topology gives us just a value of 60% at the secondary resonance
frequency. The optimization of parameters such as primary capacity and operating
frequency at constant air gap is necessary to provide the maximum power to the
utilization part and to satisfy the energy requirement, solve the issues of batteries
and charging of electric vehicles.

6 Conclusion

In this paper, a new wireless recharging system based on a resonant electrical circuit
of the inductive coupler power transfer technology is studied. The mutual induc-
tance between primary and secondary coils has been calculated and plotted by using
MATLAB for different air gap values between the two circuits. The performance of
a traditional charging system as a function of the air gap is plotted and discussed.
In addition, a wireless recharging system based on a resonant magnetic coupler has
been developed by adding capacitors to compensate the reactive power consumed by
the primary and secondary coils. The performance of a resonant wireless recharging
system as a function of frequency for both types (serie-serie: serie-parallel) is traced
and discussed. The value of the resonance frequency, the values of the added capac-
itors are calculated and optimized to achieve maximum transfer efficiency. Finally,
the system presented can be a good solution for recharging EV batteries and can be
installed in parking lots, motorways to increase the use of EVs.
1572 A. Elkhatiri et al.

References

1. Elwalaty M, Jemli M (2019) A two-rectangular coils wireless charging electric vehicle with
photovoltaic generato. In: International conference on sciences and techniques of automatic
control and computer engineering (STA) (2019)
2. Elwalaty M, Jemli M, Ben Azza H (2020) Modeling, analysis, and implementation of series-
series compensated inductive coupled power transfer (ICPT) system for an electric vehicle. J
Electr Comput Eng 2020:1–10
3. Ahmad A, Alam MS (2019) Magnetic analysis of copper coil power pad with ferrite core for
wireless charging application. Trans Electr Electron Mater
4. Yang G, Liang H (2018) Adaptive frequency measurement in magnetic resonance coupling
based WPT system. J Int Measur Confeder (IMEKO)
5. Hasanzadeh S, Vaez-Zadeh S (2012) Performance analysis of contactless electrical power
transfer for Maglev. J Magnet
6. Kim Ean K, Takehiro I (2013) Impedance matching and power division using impedance
inverter for wireless power transfer via magnetic resonant coupling. IEEE Trans Ind Appl
50(3)
7. Shuo W, David D (2013) Review of wireless charging coupler for electric vehicles. In: 39th
annual conference of the IEEE industrial electronics society. https://doi.org/10.1109/IECON.
2013.6700342
8. Jaegue S, Seungyong S, Yangsu K (2013) Design and implementation of shaped magnetic
resonance based wireless power transfer system for roadway-powered moving electric vehicles.
IEEE Trans Ind Appl 50(3)
9. Wei Z, Chunting Chris M (2015) Loosely coupled transformer structure and interoperability
study for EV wireless charging systems. IEEE Trans Power Electron
10. Soham Ch, Archana I (2017) Design optimization for an efficient wireless power transfer
system for electric vehicules. In: 1st international conference on power engineering
11. La recharge des voitures électriques. https://www.automobile-propre.com/dossiers/recharge-
voitures-electriques. Accessed 09 June 2020
12. Kalwar KA, Aamir M, Mekhilef (2015) Inductively coupled power transfer (ICPT) for electric
vehicle charging. Renew Sustain Energy Rev 47:462–475
13. Wang C-S, Covic GA, Stielau OH (2001) InGeneral stability criterions for zero phase angle
controlled loosely coupled inductive power transfer systems. In: 27th annual conference of the
IEEE industrial electronics society, vol 2, pp. 1049–1054, USA
14. Wang C-S, Stielau OH, Covic GA (2005) Design considerations for a contactless electric
vehicle battery charge. IEEE Trans Ind Electron 52(5):1308–1314 (2005)
15. Hasanzadeh S, Vaez-Zadeh S (2013) Efficiency analysis of contactless electrical power
transmission systems. Energy Convers Manage 65:487–496
Improve the Energy Harvesting
Alternatives Using the Bond Graph
Approach for Powering Critical
Autonomous Devices

Souad Touairi and Mustapha Mabrouki

Abstract In this article, new modelling is proposed of both the piezoelectric

harvester dynamics and vehicular seat suspension in accordance with the Bond
Graph approach. However, a mathematical analysis has been performed to provide
insights into which piezoelectric harvester parameters that highly broaden the energy
harvesting potentiality. Furthermore, the conducted mechatronic modelling included
the change in both driver load, suspension and piezoelectric parameters of the
harvestable power magnitude, vehicle acceleration, suspension dynamic, and tire
deflection. This proposed system has been investigated under wide broad banded
road levels (parametric weight and measurement acceleration) via mechatronic esti-
mation methodology over a piezoelectric harvester system to improved harvesting
efficiencies in different road conditions. During the system simulations, a piezoelec-
tric harvester (3 degrees-of-freedom) model is connected to vehicle suspension being
excited by randomized Gaussian signal. The simulation results are in general accor-
dance with expectations in the harvestable power interval versus the input excitation
parameters, road profile rate, and piezoelectric properties.

Keywords Vibration energy · Bond graph · Chassis suspension · Piezoelectric ·

Energy harvesting

1 Introduction

Nowadays, with increasing concern about clean energy demand the vibration energy
harvesting has become great challenges as a harvestable power for the automotive
industry to power smart sensing actuators and self-powered applications. Generally,
Piezoelectric Energy Harvesting (PEH) from ambient vibration energy is nothing but
a scavenger of anybody’s motion from its mean position [1]. Mechanical vibration
can be harvested from different construction such as a vehicle suspension, vehicle
tires, and any other solid system when subjected to the motion on different road

S. Touairi (B) · M. Mabrouki

Industrial Engineering Laboratory, Faculty of Sciences and Techniques, Sultan Moulay Slimane
University, BP 523, 23000 Beni Mellal, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1573
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_143
1574 S. Touairi and M. Mabrouki

roughness [2]. However, several industrial applications such as environmental moni-

toring systems [3], military monitoring devices, medical implants [4], structural
health monitoring sensors [5], and transport applications require a low electric power
amount [6]. Currently, an intelligent piezoelectric harvester system was created to
harvest a small vibration energy value and convert it into functional electrical energy
[7]. To harvest vibration energy namely energy dissipated during vehicle displace-
ment high efficiency, there are mainly reliable piezoelectric harvesters operate with
high-speed [8]. Traditionally, the vibrations energy is a non-used energy which distur-
bance the vehicle parts exposed to different road and driving conditions [9]. The latter
decreases the level of driving comfort and increases the fuel consumption of the
vehicle [10]. While the piezoelectric vibration energy harvesters are become useful
to realize self-powering and harvest energy lost due to vibration during vehicle motion
[11]. Currently, researchers have conducted various works on the non-uniform road
profile as one major parameter affect the random vibration on vehicle suspension
systems [12]. The authors in [13, 14], represented the piezoelectric cantilever as a
simple element to include the small amount of damping coefficient in the nature of
the suspension system. Most of piezoelectric energy harvesting system is mainly
modelled using classical concepts of Lagrange’s equations [15], and standard piezo-
electric effect [16]. Such factors are playing a major role in increasing the vibration
energy and efficiency of a piezoelectric harvester device for reducing the losses of
energy in the suspension system of electric vehicles. The above literature demon-
strated that the maximum electric power harvested (RMS) by the vehicle suspension
element reach 238 mW. This paper addresses the limitations for current piezoelec-
tric energy harvesters by a new piezoelectric harvesting system concept for produce
a usable electrical energy. For this objective, mechatronic modelling is applied to
the piezoelectric harvester (PZT-7k) device through the Bond Graph technics. This
model was finally simulated under road surface conditions to estimate the system
performance and the produced energy amount. Most notably, the proposed model
results were compared to previous works (e.g. [8–10]), the examination verifies
and validates the effectiveness and great performance of the proposed approach. As
regards, Sect. 2 concentrates on the mathematical analysis of the model and selection
of the standard parameters, Sect. 3 gives the proposed modelling on the piezoelec-
tric harvester system and finally, important simulation results and discussions are
exposed in Sect. 4 conclusions and outline are presented in Sect. 4.

2 Piezoelectric Beam Scavenger (PBS) Operating Under

the Siege Suspension

In order to model the mechanism for energy harvesting by piezoelectric composite, a

Bond Graph technic is used. This latter ensures an overview of the external interaction
that causes a change in charge distribution. The piezoelectric material deformation
by the applied mechanical constraints induces a potential difference due to poling
Improve the Energy Harvesting Alternatives … 1575

Fig. 1 The system

mass-spring-damper
architecture

at both of the piezoelectric materials ends [17]. This approach deals to mount the
piezoelectric beams on a vehicle seat suspension system parallel to the driver all
seat system which maximizes the generates voltage difference distribution in the
piezoelectric material. As a result, this strain generated by the suspension defor-
mation with the road’s higher roughness ensures a wide harvestable electric power.
However, simplified system modelling was presented and improved in this section.
The mathematical analysis is considered with one degree of freedom two degrees
of freedom and a large number of degrees of freedom. Figure 1 shows the vehicle
dynamic architecture divided into sub-modules in the proposed complete BG model.

2.1 Integrated Suspension of Vehicle Siege

In this study, the vibration of the vehicle movement supposed vertical as well as we
neglect the pitch and the roll movements. The piezoelectric beam (ceramic PVDF)
integrated into the driver’s seat is under disturbance whit complex mechanical proper-
ties. As in previous works (e.g. [17, 18]), the state equations of vehicle displacement
have been calculated as Eqs. (1), (2), (3) and (4) shown. The harvester dynamics
action and piezoelectric hysteresis are supposed to be an absolute generator force
as a vertical motion of the system that gives the assimilation motion for the suspen-
sion of the complete siege. The conductor’s body is modelled with mass, spring and
damper in the proposed BG scheme that exhibits a phase transition caused by the
mass of the vehicle chassis. The calculated equations for the vehicle dynamic during
motion are as follows:

m u z u = −kt (z u − zr ) − ct (z u − zr ) + ks (z s − z u ) + cs (z s − z u ) (1)

1576 S. Touairi and M. Mabrouki

m s z s = −ks (z s − z u ) − cs (z s − z u ) + kss (z f − z s ) + css (z f − z s ) + u (2)

m f z f = −kss (z f − z s ) − css (z f − z s ) + kc (z c − z f ) + cc (z c − z f ) − u (3)

m b z b = −kc (z c − z f ) − cc (z c − z f ) (4)

Therefore, the seat suspension model is supposed composed of the driver body
model (Suspended mass ms ) directly attached to the vehicle chassis, the rolling
weight and tire assembly (Unsprang mass mu ). In 20-Sim software input, the sieges
body mass is presented by mf ; mb is the conductor body weight; zu , zs , zf , zc and
zb are the movements of the equivalent masses; the input excitation (road profile)
was noted zr , cs and ks are the suspensions damping and stiffness respectively. Also,
kt and ct are the positions of the compressibility and restraining of the pneumatic
tires, respectively; css , kss and kc are damping coefficients, stiffness parameter of
the driver’ cushion siege suspension, respectively. The factor u signifies the matrix
of the command force applied on the siege suspension. According to this proposed
model, the connection between the piezoelectric actuators PEA actuating strength
and the consequential beam hysteresis displacement becomes linear. This model
increases the harvested power and decrease the use of the battery (long-life cycle)
[13]. However, the standard state equations of the piezoelectric harvester system
were modified by a functional analysis proposed in [14]. The conversion parameters
(of the piezoelectric beam displacement) were estimated in the stiffness matrix of
electrical energy output based on a BG distributed methodology.
This latter facilitates the task by a simple causality inversion and easily solve
the state equations. Regarding the input excitation frequency scale, the piezoelectric
beam operates only with the initial resonant frequency for conductor safety and rides
comfort.

2.2 Piezoelectric Beam Modelling Integrated on the Siege

Suspension

In this section, an explicit formula of the piezoelectric beam modelling has been
proposed. The relationship between medium pressure whit maximum diversion of
homogeneous yarn-beam is analysed at the exert point load in x = 2 * Lpzt . In addi-
tion, a new graphical representation tool is used to model the interaction between flow
connections and extern efforts. In addition, the associated characteristics are calcu-
lated with the motion equations in the equilibrium conditions, through a new graph-
ical representation tool to model the interaction between flow connections and extern
efforts. This flexible model (with multiple degrees of freedom) achieves control
performing adequately for a broad spectrum of road perturbation. This model gives
an agreeable electric energy amount and decreases the road’s disturbance gain (input
Improve the Energy Harvesting Alternatives … 1577

Fig. 2 Typical piezoelectric harvester under siege suspension

signal w) for the displacement signal (output z) as seen in Eq. (1). The piezoelectric
harvester system is modelized in Fig. 2 as a function of the suspension parameters
proposed in Ref [13]; that estimate the disturbance amount to regenerate large electric
energy.
The harvested power by this model was Eelectric = 48 W during an hour for vehicle
speed equal to VSpeed = 40 km/h. The fixed connection will duplicate the open-circuit
voltage accumulated to a unique factor, and the practical capacitance will be C =
0.5  of the theoretical capacity impedance. The selection BG model utilizes the
collateral joining in order to raise the electrical power quantity. Owing to the fact
that, the analogy connection will be selected for doubling the flow com-pared to a
unique element outline and multiply the efficient electric power (Fig. 3).
The proposed model dimensions require a small beam length, a large thickness
and a minor size to harvest a wide range of vibration energy. The parameters used in
the simulation are presented as follows [13]:

Wss = Fs X + FsY = K t X (z X X − z nz X ) + K vY (z sY − z nzY )

∼ 2m (suspension+ pezo) ∗ L (5)
= W X Y (t − )
D33 V

I pt z ausn = Fsv + v ∗ Fsv + v ∗ FP ztY + w ∗ FP zt X + ctcy (−z 1 + θ1 ∗ v − z nc

f)∗v
+ ctsy (−z 1 − θ1 ∗ w − z ncY

)∗w (6)


Mvehicle z usn = −Fsv + FP zt X + FP ztY + ctsy (−z 1 + θ1 ∗ v − z ns

f)
(7)
+ctcy (−z 1 − θ1 ∗ w − z ncY

)

MncY z ncY = FP ztY + cwsY (−z 1 + θ1 ∗ v − z ncY

) + kvsY (z cY − z ncY ) (8)


MncX z ncX = FP zt X + cws X (−z 1 + θ1 ∗ w − z ncX

) + kt y (z x y − z nsY ) (9)
1578 S. Touairi and M. Mabrouki

Fig. 3 A Bond Graph model of the Piezoelectric Harvester system

Where, Mvehicle noted the car’s mass (sprung and unsprung), and Kvy is the suspen-
sion stiffness, Tire suspension stiffness is KtX , Sprung damping coefficient CwX and
Tire damping coefficient CvY .

3 Energy Harvesting with the Whole System: Numerical

Simulations

3.1 The Dynamic Equilibrium

For a precise formulation of this approach, a numerical simulation is going to be

ana-lysed and discussed in this section. The simulation results show a great success
of the proposed harvester impartial modelling since it eliminates the batteries’ use as
presented in Fig. 2. This numerical simulation takes into account all system interac-
tions to predict the harvestable electrical power from the driver’s seat vibration that
optimizes the power consumption (20%). Figure 4 exposes the large electrical power
amount by road class C (Higher vibration).
The important result of this article is the electric power regenerated (P = 5.89 W)
has been enhanced [15]. This latter is an autonomous source of electrical energy
adapted to the function environment. The parameters used in this simulation (vehicle
siege suspension, piezoelectric harvester factors) have been calculated by the func-
tional analysis according to the conductor body and vehicle acceleration. Let suppose
P = [M, {Gsj , Qsj , Ecj )}, {Vi(.)}] denote the electric power amount, where M = {1,2}
is the functional analysis set input signal. The energy harvesting strategy assumes that
Improve the Energy Harvesting Alternatives … 1579

Fig. 4 Bond Graph Inverse model of the piezoelectric harvester system

Esj , Qsj , and Gci for each beams PZTj are compact and convex sets with maximum
and minimum constraints give through the functional analysis. In addition, each
IS Pj considered and solved respectively as a transfer function in the closed solutions
intervals. In order to maximize their power production, each piezoelectric beams IS Pj
produces a power amount Esj , QoS qsj , the sum gives the total power Pcj . Formally,
the piezoelectric hysteresis resolution expressed as: H = [M, {Ksj , Qsj , Dcj },{Vi (.)}]
[19].
The solutions are deduced by the following Bond Graph Inverse resolutions
as shown in Fig. 4. This approach gives fast parameters estimation to check the
implementation of the modelling and reduce the mass matrix size. This solution is
promising to enhance the electric power when the vehicle’s mass is bigger.

3.2 Results and Discussions

In this part, the global road disturbances at 15 m/s have been applied to the piezoelec-
tric harvester by COMSOL and MATLAB/SIMULINK software as shown in Fig. 5.
In addition, the vehicle speed according to the piezoelectric harvester deformations
is analysed. The simulation results were obtained for the PZT-7k harvester substrate.
The thermal analysis in Fig. 5 shows a good cyclical waveform and lower hysteresis
level (total piezoelectric patch creeps’ distortion) about 15.5% for PZT-7k compared
with 25.44% for the classical piezoelectric harvester model. Accordingly, the stan-
dard road profile the relative vehicle movement with the highest road amplitudes
which are respectively 0.32, 0.52, and 0.99 m. The harvested power is 306, 430, and
1580 S. Touairi and M. Mabrouki

Fig. 5 A piezoelectric harvester thermal analysis and Electric vehicle input driving

470 mW for a vehicle’s suspended mass of 340 kg and unsuspended mass 180 kg
animated by the random road function for the transversal excitation during an hour
of displacement as seen in Fig. 8. This simulation exposes efficient energy harvesting
for vehicle seat suspension. In fact, when the amount of vehicle speed increases, then
the piezoelectric harvester dimensions decrease and the power bandwidth decrease
(Figs. 6 and 7).
Figures 8 and 9 show respectively: The large road roughness causing wide vertical
displacements, thus increasing the harvester system performances and the harvestable
electric power quantity. Comparing Figs. 8 and 9, it can be clearly noticed that the
proposed approach has great responses compared to the classical model for which the
harvested power and piezoelectric patch have less efficiency. The latter decreases the
vehicle fuel consumption in higher road disturbances. The piezoelectric beam had
suitable dimensions for manufacture and realization thus defined the need for energy
to consume by the functional analysis approach. Those results prove the model’s
uniqueness, following [14], the two homogeneous drivers’ body mass was used to
maximize the producer power and define the optimal weighted sum of harvester
utility functions. As a result, the Y P Z T decreases the vehicle speed H P Z T to further

Fig. 6 Road input

disturbance C-class
Improve the Energy Harvesting Alternatives … 1581

Fig. 7 Random disruption

of the road

Fig. 8 Effects of the vehicle

displacement on the electric
power amount

Fig. 9. Driver’s mass effects

on the PZT behaviours and
power amount
1582 S. Touairi and M. Mabrouki

conducting greater driver comfort for any road surfaces. Figure 5 summarizes the
characteristics of PZT-7K previously analyzed. Finnaly, note that Hpzt increases with
respect to the number of beams PZT and ISPs . The vehicle speed and the piezoelectric
deformation are presented for the first time according to the hysteresis phenomena
as exposed in the electric power nonlinear function Eqs. (5–9).

4 Conclusion

The principal goal of this article is to improve a piezoelectric system built into the
car seat to quantify and convert mechanical vibrations energy of complex equip-
ment (critical devices) and road disturbance. The system gives sustainable electric
power and improves the voltage magnitude with the purpose of increasing the piezo-
electric substances’ abilities. Moreover, this model increased the time life of the
vehicle’s Micro and Nano devices. As shown in the interpretation part the choice
of the piezoelectric harvester’s bond graphs model developed in this paper is very
important to improving the harvested energy. The time spent to reach 302 mW was
20 s on different road test references (50, 70 and 100 km/h) exploiting the beams’
sustainability and simplicity of representation.
As a conclusion, the simulation results confirm that the use of the piezoelectric
harvester attached to the siege suspension ensures excellent electrical power produc-
tion for the autonomous device alimentation. In addition, modelling this system by
the Bond Graph model tool improves the overview of vehicle dynamics. Finally,
the use of the numerical results allowed harvesting vibration energy beneficial for
the seat suspension. Furthermore, this paper gives a simple model that includes a
detailed analysis of piezoelectric hysteresis behaviours and creeps in multiple road
conditions.

References

1. Touairi S, Mabrouki M (2020) Optimization of energy harvesting system design by functional

analysis. In: 1st international conference on innovative research in applied science, engineering
and technology (IRASET), pp 1–6
2. Liu M, Gu F, Zhang Y (2017) Ride comfort optimization of in-wheel-motor electric vehicles
with in-wheel vibration absorbers. Energies 10:1647. https://doi.org/10.3390/en10101647
3. Luo ACJ, Liu C (2020) Analytical periodic motions in a discontinuous system with a switching
hyperbola. Int J Dyn Control. https://doi.org/10.1007/s40435-020-00648-5
4. Rafiq D, Bazaz MA (2020)A comprehensive scheme for reduction of nonlinear dynamical
systems. Int J Dyn Control 8:361–369. https://doi.org/10.1007/s40435-019-00589-8
5. Reddy RKK, Arunkumar MP, Bhagat V, Reddy MBSS (2020) Vibro-acoustic characteristics
of viscoelastic sandwich panel: effect of inherent damping. Int J Dyn Control. https://doi.org/
10.1007/s40435-020-00654-7
Improve the Energy Harvesting Alternatives … 1583

6. Ruttanatri P, Cole MOT, Pongvuthithum R (2020) Structural vibration control using delayed
state feedback via LMI approach: with application to chatter stability problems. Int J Dyn
Control. doi: https://doi.org/10.1007/s40435-020-00645-8
7. Shao K, Xu Z, Wang T (2020) Robust finite-time sliding mode synchronization of fractional-
order hyper-chaotic systems based on adaptive neural network and disturbances observer. Int
J Dyn Control. https://doi.org/10.1007/s40435-020-00657-4
8. Wei C, Jing X (2020) A comprehensive review on vibration energy harvesting: modelling and
realization. Renew Sustain Energy Rev 74:1–18. https://doi.org/10.1016/j.rser.2017.01.073
9. Xie X, Wang Q (2015) A mathematical model for piezoelectric ring energy harvesting tech-
nology from vehicle tires. Int J Eng Sci 94:113–127. doi: https://doi.org/10.1016/j.ijengsci.
2015.05.004
10. Yuvapriya T, Lakshmi P (2020) Numerical analysis and performance enhancement of active
suspension system using bat optimization. Int J Dyn Control. https://doi.org/10.1007/s40435-
020-00664-5
11. Zerimeche H, Houmor T, Berkane A : Combination synchronization of different dimensions
fractional-order non-autonomous chaotic systems using scaling matrix. Int J Dyn Control.
https://doi.org/10.1007/s40435-020-00660-9
12. Badfar E, Ardestani MA (2020) Design of adaptive fuzzy gain scheduling fast terminal sliding
mode to control the radius of bubble in the blood vessel with application in cardiology. Int J
Dyn Control. https://doi.org/10.1007/s40435-020-00641-y
13. Touairi S, Mabrouki M (2020) Mechatronic modeling and control of energy recovery in
motorcycle tires. In: 2020 IEEE 6th international conference on optimization and applications
(ICOA), pp 1–5
14. Gambhire SJ, Kishore DR, Londhe PS, Pawar SN (2020) Review of sliding mode based control
techniques for control system applications. Int J Dyn Control. https://doi.org/10.1007/s40435-
020-00638-7
15. Ghelichi M, Goltabar AM, Tavakoli HR, Karamodin A (2020) The cost-effective and multi-
objective optimal positioning of active actuators based on tug of war method in seismically
excited benchmark highway bridge. Int J Dyn Control. doi: https://doi.org/10.1007/s40435-
020-00655-6
16. Gopala Rao LVV, Narayanan S (2020) Optimal response of half car vehicle model with sky-
hook damper based on LQR control. Int J Dyn Control 8:488–496. https://doi.org/10.1007/s40
435-019-00588-9
17. Hasan S, Dhingra AK (2020) 8 Degrees of freedom human lower extremity kinematic and
dynamic model development and control for exoskeleton robot based physical therapy. Int J
Dyn Control 8:867–886. https://doi.org/10.1007/s40435-020-00620-3
18. Iggidr A, Oumoun M (2020) Simultaneous stabilization of single-input nonlinear systems with
bounded controls. Int J Dyn Control. https://doi.org/10.1007/s40435-020-00658-3
19. Touairi S, Mabrouki M (2020) Optimization of Harvester System in embedded vehicle
systems via Bond Graph modeling algorithm. In: 2020 IEEE 6th international conference
on optimization and applications (ICOA). pp 1–6
Online Energy Management Strategy
of FC/UC Hybrid Electric Vehicle Based
on Neural Network

Nada Rifai, Jalal Sabor, and Chakib Alaoui

Abstract Fuel cell (FC) vehicles seem to be the solution for clean and efficient
transportation. However, the coupling of the fuel cell with another power source
remains essential to overcome its low power density and preserve its lifetime as long
as possible. Thanks to its fast response and high power density, the ultra-capacitor
(UC) can meet a fast and short-duration engine power demand. This paper presents a
power management strategy between the FC and the UC, based on a neural network
(NN). The proposed NN model is trained with wavelet decomposition outputs. To
ensure that both sources operate optimally, the fuel cell power limits as well as the
state of charge (SOC) of the UC are considered.

Keywords Fuel cell · Ultracapacitor · Energy management strategy · Neural

network · Wavelet transform

1 Introduction

Fuel cells have been attracting a lot of attention from researchers and major car
manufacturers thanks to their zero emission of greenhouse gases. A major advan-
tage of fuel cells is that they are small and lightweight. In addition to that, they can
generate electrical energy as long as the fuel is supplied [1]. However, two impor-
tant disadvantages may limit the mass commercialization of fuel cell vehicles. First,
the high cost, which is estimated to be 5 times higher than conventional vehicles,
and second, its durability, which does not exceed 5,000 h, which is equivalent to

N. Rifai (B) · J. Sabor

LSMI, ENSAM, Moulay Ismail University, Meknes, Morocco
e-mail: [email protected]
J. Sabor
e-mail: [email protected]
N. Rifai · C. Alaoui
Insa EuroMéditerranée, Euromed University, Fez, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1585
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_144
1586 N. Rifai et al.

150,000 miles in optimal operation [1]. Hybridization of fuel cell with peak power
sources has proven good results to solve these problems. Batteries and Ultra capac-
itors are the most suitable sources for this type of application. An energy manage-
ment strategy (EMS) is therefore necessary to ensure the optimal distribution of
energy between the main source which is the FC and the secondary source which
is the UC. EMSs as reported in the literature can be classified into 4 classes: Rule
based methods, Controller methods, a time-frequency filter methods and optimiza-
tion methods [2]. For rule based methods, we can find EMSs based on fuzzy logic
[3, 4], or Neural network [5] or a combination of both of them such as in ref [3].
For controllers methods, different methods are investigated such as PI controllers,
differential flatness control [7]. The time frequency filter method is achieved using
low pass filters [8], wavelet transform [9, 10], or Fast Fourier transform [11]. Opti-
mization methods are efficient to solve complex management goals such as reducing
cost, minimizing fuel consumption, increasing efficiency or energy sources lifespan.
The problem of these methods is their complexity and computation burden which
is time-consuming. Multiple methods are used for optimization, the most recurrent
ones are the dynamic programming [12, 13] the adaptive optimal control method
[14]. Pontryagin’s Minimum principle approach [15, 16].
In this paper, the fuel cell was hybridized with a UC, since this architecture seems
to have better performances. In fact, the supercapacitor can more effectively support
the FC to meet transient power demand as the UC can be charged or discharged at
a high current, in which the battery is not able to operate. Moreover, highcurrent
charging and discharging of the battery will also reduce its lifetime [17]. The aim
of this paper is to present an energy management strategy based on time-frequency
filter method using Wavelet Transform and a FC/UC limitation check algorithm
which avoid that the fuel cell works out of its comfort zone and control the UC’s
SOC. The outputs of these are used as the database to train a Neural Network. The
purpose of replacing the two algorithms with a neural network is to reduce calculation
time, and also to be able to use the block for online energy control.
The combination of Wavelet Transform and Neural Network has already been used
in literature. However, the energy system architectures and the use of the Wavelet
transform and neural networks are different from their application in this article.
For example, in [18] Authors worked on battery/ultracapacitor energy system, They
used a nonlinear auto regressive NN to predict was used to predict the used the
wavelet transform to predict the vehicle’s power demand, the wavelet transform is
then applied, and the low frequency output is considered as the reference for the
power battery and high frequency output, the reference of UC power. In [19], a
combination of wavelet transform, neural network and fuzzy logic is adopted. Only
the low frequency output of the wavelet transform is used to train the neural network.
The online control is composed of an adaptive linear neural network and fuzzy logic
controller. This EMS is also applied for a UC/battery energy system.
This paper is organized as follows: Sect. 2 presents the energy management
strategy proposed in this paper. Section 3, the simulation of this EMS and the results
are discussed, and finally a conclusion and perspectives are presented in Sect. 4.
Online Energy Management Strategy … 1587

2 Energy Management Strategy

In this study, we propose a real time energy management strategy for a fuel cell
and ultracapacitor energy storage system. The EMS presented in this article is based
on a neural network. The wavelet decomposition combined with a checking limits
algorithm is used offline to train the neural network.

2.1 Wavelet Decomposition

As the fuel cell is unable to respond to fast and abrupt power demands, the main
objective of the EMS would be to avoid rapid demands on the fuel cell but at the
same time to guarantee the traction of the vehicle. As a signal processing method,
wavelet transform is known for its efficiency in dealing with non-stationary signals.
For this, it will be used for sharing power demand between the two energy sources by
decomposing the load power demand to a low frequency signal and a high frequency
signal. The Mexican Hat wavelet function is used. This function is one of the most
used wavelet function for detecting and localizing transients. The complexity of this
function is not critical, since its efficiency has been proved and it used in offline to
prepare the datasets for the neural network.
Continuous wavelet transform (CWT) is defined as below:
 +∞
1
Wa,b (x) = s(t). √ .a,b (x).dx (1)
−∞ a

Where  is the wavelet function defined by:

 
x2 x2
(x) = 1 − 2 .e− 2.a2 (2)
a

And a is the width of the wavelet function, which will be variable during the
transformation and will be calculated for each unique spectral component. The signal
s(x) is reconstructed from the coefficients W (a, b) using continuous inverse wavelet
transform given by (3).
 +∞  +∞  
1 x − b dadb
s(x) = W (a, b) √  (3)
0 −∞ a a a2

The detailed description about the proposed wavelet algorithm can be found in
ref [10].
The low-frequency component of the signal is considered as the reference for the
power to be requested from the fuel cell, P f cr e f . The high-frequency component in
1588 N. Rifai et al.

addition to recovered the braking energyare considered for the power as a reference
of the power to be requested from the ultracapacitor, Pscr e f .

2.2 Limit Check Algorithm

Based on the outputs of the Wavelet Decomposition, the limit check algorithm will
allow the fuel cell to operate in its comfort zone which is the linear opertaing region,
and will control the state of charge of the ultracapacitor. Through the respect of
following constraints:

P f cmin < P f c < P f cmax (4)

d Pf c
P f cmin < < P f cmax (5)
dt

S OC scmin < S OC uc < S OC scmax (6)

If the P f cr e f exceed the P f cmax , FC delivers its maximum power, P f cmax, UC

supplies less power. Likewise, If the P f cr e f is less than P f cmax , FC delivers its
minimum power, P f cmin and UC delivers more power. The FC current variation step
is limited by the manufacturer. This limitation also imposes boundaries for the power
variation as in (4). If abs(P f cr e f (k) − P f c (k − 1)) > P f cmax , FC will deliver less
power and UC more and vice versa. If the SOC is around the desired level, FC
supplies the reference power signal value as well as the UC. However, if SOC is less
than the desired value (here 0.2), the FC delivers more power than its reference, and
and as a result the UC will deliver less power that what is expected. Similarly, if SOC
is higher than the desired value (here 0.75), then the FC provides less power.

2.3 Feed Forward Neural Network

The power demand of the vehicle is calculated based on four driving cycles. First ones
are the Atremis Driving cycles which are based on a statistical study done in Europe
within the so called Artemis project. ArtRoad cycle that simulates driving on a rural
road and ArtUrban that simulates driving on a rural road were used. The Highway
Fuel Economy Driving Schedule (HWFET) was also used. This cycle represents
highway driving conditions under 100 mph. Finally, The EPA Urban Dynamometer
Driving Schedule (UDDS) which represents city driving conditions. It is used for
light duty vehicle testing. This is a certified driving cycle for verifying emissions
after a cold start.
Online Energy Management Strategy … 1589

Based on the power demand profiles required for these four driving cycles, the
power needed from the FC and the UC are generated by the offline EMS. 70% of
these data are used as datasets for the neural network training. 30% is used for model
validation.
The neural network has three inputs, the first is the load power demand, Second
is a binary coefficient to indicate whether it is in motor or generator mode. Finally,
the SOC of the UC.
As neural networks work best with features that are at a relatively similar scale
and close to the normal distribution, and in order to reduce the training time of the
neural network. The power features (input power, FC and UC power Targets) have
been normalized by subtracting the mean and then scaling to unit variance, which
means dividing all values by the standard deviation:

x − mean
x= (6)
stand deviation
The outputs of the NN will then be rescaled by multiplying with the same standard
deviation and adding the same average used during training.
In this article, a multiple layer neural network was applied. This network is used
as a general function approximator. It can approach any function with a finite number
of discontinuities arbitrarily well, given a sufficient number of neurons in the hidden
layer. Its simple structure makes it a good solution for an online EMS, since it can be
easily implemented on hardware. For this application, The NN has two hidden layers
of 40 and 20 neurons respectively. Hyperbolic tangent sigmoid transfer function
(tansig) is used to allow the network to learn nonlinear relationships between input
and output vectors. These two layers are then followed by an output layer of linear
neurons, purelin transfer function is employed for nonlinear regression.
Weights and bias were updated using BP (back-propagation) which is the most
powerful method to minimize the performance function [20]. In this paper, Trainlm
network training function based on Levenberg-Marquardt optimization is used. It is
the fastest backpropagation algorithm in Matlab toolbox. The NN performance was
evaluated using the Mean square error.
In the next section, the results of the wavelet decomposition coupled with the limit
check algorithm are presented. Then, the results of the training and validation of the
neural network are provided.

3 Simulation and Results

The EMS based on wavelet transform coupled with the check limit algorithm was
tested EPA Urban Dynamometer Driving Schedule (UDDS) presented in Fig. 1. The
vehicle on which this study was conducted is a Citroen Berlingo. As shown in Fig. 1.
this vehicle needs an energy system of about 40 kW to propel it. Considering the
maximum power of the fuel cell studied (1.2 kW), we have reduced the scale of the
1590 N. Rifai et al.

Fig. 1 UDDS drive cycle (blue) and power needed for traction of the Citroen Berlingo vehcile

Fig. 2 First plot: Motor Power (red) FC power (black) UC power (Bleu), Second Plot (SOC of
UC)

power profile required to 12.5 of the real scale. The offline EMS results are presented
in Fig. 2. As can be observed, based on the EMS offline, the power demand is covered,
in addition the power delivered by the FC is slow and of low frequency, on the other
hand it does not exceed its limits which in this case are 200 W at the lower bound and
Online Energy Management Strategy … 1591

12 kW at the upper bound. Then, for UC it can be clearly noticed that its dynamic
is much faster, which makes it possible to efficiently respond to a fast transition of
power demand. And finally, the second plot in Fig. 2 represents the SOC of the UC.
This one is as desired between 0.2 and 0.75 which proves the efficiency of the EMS
in controlling the state of charge of the UC.
Based on the results of the four driving cycles ArtUrban, ArtRoad, HWFET and
UDDS. 70% of the data is standardized and used for NN training. Figure 3 presents
the outputs of NN during training. The comparison between the output of the neural
network and the target powers of the FC and UC, which are the outputs of the offline
EMS, led to a root mean square error (RMSE) of 0.040.
30% of the data is used for NN validation. Three inputs are applied. The Power
demanded by the load which is normalized, SOC of the UC and finally alpha the
binary coefficient. At the output of the NN, the powers are rescaled. The result is

Fig. 3. Training Neural network

Fig. 4 Test and validation of the neural network

1592 N. Rifai et al.

shown in Fig. 4. First plot, In blue, the FC power Target output of the Offline EMS
and in red, the power proposed by the NN. In the second plot, in blue the Target power
UC and in black the power proposed by the NN.As can be noticed the feed forward
NN employed is capable of reproducing the work of the wavelet decomposition
algorithms with the verification algorithm. The error between the two (Targets and
NN outputs) is small and does not influence the global rules of the EMS.

4 Conclusion

In this article, we replace a management strategy based on wavelet transform by a feed

forward neural network. The mean square error between the results of EMS based
on WT and the neural network outputs is of 0.0720 which proves the efficiency of
the neural network to predict the power of both energy sources. The neural network
used in this paper has a simple architecture which make easily implementable on
hardware and suitable for energy management applications in vehicles.
The next steps would be to simulate the complete energy storage system, and
validate it experimentally.

References

1. Pollet BG, Staffell I, Shang JL (2012) Current status of hybrid, battery and fuel cell electric
vehicles: from electrochemistry to market prospects. Electrochimica Acta 84:235–249
2. Djerioui A et al (2019) Energy management strategy of Supercapacitor/Fuel Cell energy storage
devices for vehicle applications. Int J Hydrogen Energy 44(41):23416–23428
3. Gao D, Jin Z, Lu Q (2008) Energy management strategy based on fuzzy logic for a fuel cell
hybrid bus. J. Power Sources 185(1):311–317
4. Rodatz P, Paganelli G, Sciarretta A, Guzzella L (2005) Optimal power management of an
experimental fuel cell/supercapacitor- powered hybrid vehicle. Control Eng. Pract
5. Zhou D, Al-Durra A, Gao F, Ravey A, Matraji I, Godoy Simões M (2017) Neural network. J
Power Sources 366, 278–291
6. Zhang R, Tao J, Zhou H (2019) Fuzzy optimal energy management for fuel cell and super-
capacitor systems using neural network based driving pattern recognition. IEEE Trans Fuzzy
Syst 27(1):45–57
7. Thounthong P, Pierfederici S, Martin JP, Hinaje M, Davat B (2010) Modeling and control of
fuel cell/supercapacitor hybrid source based on differential flatness control. IEEE Trans Veh
Technol 59(6):2700–2710
8. Oukkacha I, Mamadou BC, Dakyo B (2018) Energy management in electric vehicle based on
frequency sharing approach, using fuel cells, lithium batteries and Supercapacitors. France,
Paris
9. Masih-Tehrani M, Ha’Iri Yazdi MR, Esfahanian V, Dahmardeh M, Nehzati H (2019) Wavelet-
based power management for hybrid energy storage system. J Mod Power Syst Clean Energy
7(4):779–790
10. Rifai N, Sabor J, Alaoui C (2021) Energy management strategy of a fuel-cell electric vehicle
based on wavelet transform. In: Lecture Notes in Networks and Systems, vol 144, Springer, pp
220–235
Online Energy Management Strategy … 1593

11. MacKie DM, Jahnke JP, Benyamin MS, Sumner JJ (2016) Simple, fast, and accurate method-
ology for quantitative analysis using Fourier transform infrared spectroscopy, with bio-hybrid
fuel cell examples. MethodsX 3:128–138
12. Chen Z, Mi CC, Xu J, Gong X, You C (2014) Energy management for a power-split plug-in
hybrid electric vehicle based on dynamic programming and neural networks. IEEE Trans Veh
Technol 63(4):1567–1580
13. Ansarey M, Shariat Panahi M, Ziarati H, Mahjoob M (2014) Optimal energy management in a
dual-storage fuel-cell hybrid vehicle using multi-dimensional dynamic programming. J. Power
Sources 250:359–371
14. Lin WS, Zheng CH (2011) Energy management of a fuel cell/ultracapacitor hybrid power
system using an adaptive optimal-control method. J. Power Sources 196(6):3280–3289
15. Hemi H, Ghouili J, Cheriti A (2014) An optimal control solved by Pontryagin’s minimum
principle approach for a fuel cell/supercapacitor vehicle. In: Proceedings - 2014 Electrical
Power and Energy Conference, EPEC 2014, pp 87–92
16. Ettihir K, Boulon L, Agbossou K (2016) Optimization-based energy management strategy for
a fuel cell/battery hybrid power system. Appl. Energy 163:142–153
17. Thounthong P, Chunkag V, Sethakul P, Davat B, Hinaje M (2009) Comparative study of fuel-cell
vehicle hybridization with battery or supercapacitor storage device. IEEE Trans Veh Technol
58(8):3892–3904
18. Ibrahim M, Jemei S, Wimmer G, Hissel D (2016) Nonlinear autoregressive neural network
in an energy management strategy for battery/ultra-capacitor hybrid electrical vehicles. Electr
Power Syst Res 136:262–269
19. Zhang Q, Wang L, Li G, Liu Y (2020) A real-time energy management control strategy for
battery and supercapacitor hybrid energy storage systems of pure electric vehicles. J. Energy
Storage 31
20. An N, Zhao W, Wang J, Shang D, Zhao E (2013) Using multi-output feedforward neural network
with empirical mode decomposition based signal filtering for electricity demand forecasting.
Energy 49(1):279–288
The Potential Outcomes of Artificial
Intelligence Applied to the Powered
Two-Wheel Vehicle: Analytical Review

F. Jalti, B. Hajji, and A. Mbarki

Abstract Modern modes of transport are experiencing notable progress in terms of

the inclusion of driver assistance systems, which the distribution is however unfair
in relation to the criticality of each mode of transport. Indeed, two-wheeled vehicles
benefit the least from technological progress despite their vulnerability.
With the aim of defining an inventory for our research work, this paper describes
the potential of AI (Artificial Intelligence) in driver assistance systems. Through a
bibliographic and technological analysis, we examined the state of the art around the
PTW (Powered Two-Wheeled) to make the link between the potentialities and the
key indicators related to this mode of transport.
This study was carried out on numerous articles published on international plat-
forms and journals, distributed between 1979 to 2020, of which 12% include the use
of artificial intelligence. The study found that by applying ITS (Intelligent Transport
Systems) to motorcycles, road accidents can be reduced and mobility would be more
efficient. In addition, this study also suggested a new use of technology by involving
AI to improve the management of problems with loss of control, speed, wearing of
helmets, visibility and alertness.

Keywords Intelligent transport systems (ITS) · Artificial intelligence (AI) ·

Powered two-wheeled (PTW). Advanced rider assistance systems for powered
two-wheeler (ARAS-PTW) · Machine/Deep learning

1 Introduction

In recent years, the number of motorized two-wheeler accidents has increased signif-
icantly, several causes are singled out for serious motorcycle accidents: Speeding

F. Jalti (B) · B. Hajji

Laboratory of Renewable Energy, Embedded System and Information Processing,
National School of Applied Sciences, Mohammed First University, 60000 Oujda, Morocco
e-mail: [email protected]
A. Mbarki
National School of Applied Sciences, Mohammed First University BP665, 60000 Oujda, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1595
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_145
1596 F. Jalti et al.

or unsuitable for the conditions, Loss of control, Wrong turn/maneuver, Wrong

estimation of the trajectory or speed of another road user, Negligent behavior or
reckless.
The first works which had tried to draw up a development framework began in
2009 with the SAFERIDER project [1], then came the first paper which characterizes
the Advanced Rider Assistance Systems for Powered Two-Wheeler “ARAS-PTW”
[2]. From this point on, research has diversified. Notably in terms of dynamics and
stability, if we want to define the operating rules according to a complete mathemat-
ical model [3] we must be able to measure the huge number of parameters, so until
then the best solution was the use of estimation and observation [4, 5], these conven-
tional methods have shown their limit given the low development which is mainly
due to the level of uncertainty, the complexity and the inability of these methods to
predict the danger.
In this paper, the application of artificial intelligence (AI) has been seen closely.
By comparing the capacities that machine learning algorithms offer today with the
weak aspects of PTWs, we established an inventory allowing to build a basis for our
research work. Indeed, several interesting projects have arisen and in particular by
exploiting object recognition by AI: Detection of sufficient space for a safe maneuver
[6], Obstacle detection [7], prediction and pattern detection [8]. We observed that
the trend has been shifted from the physical to data acquisition performance.
Thus, the organization of the paper will be as follows: Sect. 2 presents an oppor-
tunities and weaknesses study, Sect. 3 details the state of the art around PTW linked
with AI, Sect. 4 opens the horizon on the perspectives by discussion of results and
finally Sect. 5 concludes on the results and sets out the research perspectives.

2 Opportunities and Weaknesses Inventory of PTW Vehicle

2.1 Accident Causality Analysis

According to the WHO report in 2017 [9]. Although PTW vehicles are important
for the mobility of goods and people, they also lead to an increased risk of traffic
accidents. The significant number of deaths and injuries associated with this mode
of transport can be explained by several factors, including the proliferation of these
vehicles and the vulnerability of their users. In Fig. 1, by consulting the statistics
relating to fatalities linked to each mode of transport, we can observe, worldwide,
the percentage of PTW users having died in a road accident is the second highest
after cars [9].
With around 516 million PTWs vehicles registered worldwide in 2013, accounting
for 29% of all registered vehicles, the actual number in circulation will only be greater
as a significant proportion of vehicles go unrecorded. Adding to this, according to
WHO’s Global status report on road safety 2015, that 88% of all registered PTWs
vehicles globally in 2013 were in low- and middle-income countries.
The Potential Outcomes of Artificial Intelligence ... 1597

Fig. 1 Distribution of road fatalities according to the mode of transport involved

By linking the factors, multiple studies have been conducted to reveal causalities
and correlations in order to identify the most appropriate remedies and actions [10].
Our analysis showed that since the beginning of the 2000s, accident studies have
tended to increase. From the batch of papers, we reviewed, 26 scientific articles
(representing 22%) have dealt with PTW accidentology. Although, the scientific
research has differed on how accidentological aspects have been treated, the common
factor remains human error in the estimation of the danger.

2.2 PTW Ecosystem Potentiality Analysis

With the increasing popularity of this transport mode and increased casualty levels,
new national attention is currently being given to this area.
Indeed, it is no coincidence that the number of users is increasing (Fig. 2), but it
is rather due to the growth of emerging countries, where we can see the incentive
to this mode of transport given its accessibility with a price of purchase and cost of
ownership quite low, and also the flexibility to overcome traffic jams. Adding to this
that despite the pollution of two motorized wheels being pointed out, PTWs remain
a low-polluting alternative if we associate the track occupancy and consumption
parameters (close to 1l/100 Km for some models) [11].
In terms of safety, the technological advance of the driving assistance systems
initially developed for cars, offers an opportunity for PTWs if we succeed in popu-
larizing these systems at the level of the motorcycle industry. At the same time, the
application of artificial intelligence in this field is very promising.
1598 F. Jalti et al.

Fig. 2 The world according to PTW ownership. From www.worldmapper.org

3 State of Art Analysis

Given the weaknesses of the PTW ecosystem, the challenges are faced on several
levels (public policies, social, economic, R&D, etc.). In terms of scientific research,
more than 80% of the batch of papers we reviewed were conducted after 2010,
reflecting the growing interest in the subject.

3.1 Intelligent System Characterization

Before reviewing the state of the art related to the use of AI in PTWs, it’s necessary
to characterize the system made up of the {Driver, Vehicle, Environment}. The rider
continuously participates in the acquisition of information relating to the environment
as well as his role of Navigation and guidance through which he interacts with road
traffic. As for the machine, it mainly participates in the action.
In such a system we can assign three roles (Navigation, guidance and stabilization)
and we can designate 3 activity levels (Perception, Decision, Action). So, one way
to reflect “the level of intelligence” (or the level of automation) of the machine is to
position the system on the chart of Fig. 3.

3.2 Scientific Research Trend Around PTW

The work around PTW Vehicles was examined, through a classification according
to field and theme as well as the focus provided by the research concerned.
Concerning the themes, we observe in Fig. 4 that more than 80% of the
articles focus on (in order of importance): ITS (35%), driving behavior (19%),
Accidentology/safety (16%) and finally modeling (15%).
The Potential Outcomes of Artificial Intelligence ... 1599

Fig. 3 Tasks, Activities and

Automation when driving
(Godthelp, 1992)

By analyzing the statistics emerging from our bibliographic inventory, we can

confirm the particularity of PTWs given that 70% of the work has affected the driver.
Indeed through accident studies, driving behavior studies or studies on intelligent
systems we address mainly the rider scope.
By examining these documents, we noted that the studies, the models, the tech-
niques or the perspectives drawn up tend to reinforce the automation of the system
which today is not at the same level as that of the four wheels. The difficulty in
achieving the same level of autonomy for automotive systems is due to several factors,
the main ones being: The complexity of the dynamics of PTWs, the driver’s direct
contact with the environment, the weakness of the development framework in soft-
ware and hardware term, the cost of development which does not allow generalization
on all ranges of the market.
However, the popularization of ADAS systems has already crossed borders to
reach PTWs, also several AI applications have started to be used in this sector.

3.3 The Application of AI in Relation to ITS Field

The field of intelligent transport systems (ITS) brings together all the subjects that
are interested in driving assistance, optimizing the use of transport infrastructures,
improving safety and security as well as the use of artificial intelligence.
Concerning machine learning (ML) techniques, they have been used with success
in a wide variety of fields ranging from computer vision tasks, industry, autonomous
vehicles, etc. [29]. We now recognize two branches in the field of artificial intelligence
which depend on the type of learning, supervised and unsupervised learning [30]. In
supervised learning, models are designed to learn based on the labeled training data.
These trained models are used to make predictions about invisible data. The idea is
to design models capable of learning the patterns, behaviors and characteristics of
the data seen and to apply this knowledge while making decisions in the invisible
data. In unsupervised learning, models are designed to identify relationships and
1600 F. Jalti et al.

60

QTY of PAPERS
50
40
30
20
10
0

DOMAINE

Fig. 4 Scope Distribution of the papers around PTWs

Fig. 5 Trend of papers

interested on ITS and AI for
PTW - Qty of papers per year

correlations within unlabeled data. Depending on the level of precision desired,

various ML algorithms are available to perform prediction tasks. we can roughly
consider such an algorithm a learning function f which maps the input variables X
(x1, x2…, xn) to output variables Y (y1, y2…, yn).

Y = f (X) (1)

Since the shape of the function is usually unknown, the task is to evaluate different
ML algorithms to determine which one is best to approximate an underlying func-
tion. The learning model differs depending on whether it learns from a fixed set of
parameters, regardless of the amount of data fed, so we will call it parametric ML. Or
if it doesn’t make strong assumptions about the shape of the mapping function, in this
case it is a nonparametric ML model. Parametric models are simple, quick to learn,
and work well even if the data is not perfect. However, they tend to be unsuitable and
heavily constrained to a specific shape. On the other hand, nonparametric techniques
can adapt to many functional forms and can lead to more efficient prediction models.
They require a large amount of data to estimate the underlying mapping function,
take longer to practice, and can lead to overfitting.
Besides, the trend on the importance given by scientific research to Artificial
intelligence in the riding assistance for PTWs is at least recent as can be seen in
graph of Fig. 5. In fact, before 2000 we have almost no project interested in the ITS
related to PTWs. But since 2010 we have witnessed the increase in works that have
focused on intelligent systems applied on PTWs. Also, we had to wait until 2015
The Potential Outcomes of Artificial Intelligence ... 1601

to notice a remarkable shift towards AI with 14 articles out of 17 that address the
concept of AI.
In our analysis, on the 57 articles concerning ITS, we classified our batch
according to the affected area in 29 groupings. Where it was found that the percentage
occupied by artificial intelligence represents 28% distributed according to the table
below (Table 1):

3.4 Technology Advancement Review

The technological advance in terms of intelligent systems that equip PTWs is natu-
rally led by the manufacturing giants. Since the associated intellectual property is
required, we will cite the main areas of development that the sector knows:

Table 1 AI and Two-Wheeled vehicles - inventory of papers

Title Year Scope
Automatic Braking System for Two-Wheeler with Object 2020 Autonomous braking
Detection and Depth Perception [14]
Learning a Curve Guardian for Motorcycles [16] 2019 Curve characterization
Estimation of Mental Workload during Motorcycle Operation 2015 Driver behavior
[18]
Machine Learning Algorithm in Two wheelers fuel Prediction 2019 Fuel consumption
[17]
Helmet Use Detection of Tracked Motorcycles Using 2019 Helmet use
CNN-Based Multi-Task Learning [15]
Automated Vehicle Classification with Image Processing and 2017 Image recognition
Computational Intelligence [19]
Detecting motorcycle helmet use with deep learning [20] 2020 Image recognition
Performance evaluation of HMM and neural network in 2011 Maintenance
motorbike fault detection system [12]
Applying knowledge structure to the usable fault diagnosis 2005 Maintenance
assistance system: A case study of motorcycle maintenance in
Taiwan [21]
Road Targets Recognition Based on Deep Learning and 2018 Obstacle detection
Micro-Doppler Features [7]
Powered Two-Wheeler Riding Pattern Recognition Using a 2015 Pattern recognition
Machine-Learning Framework [8]
Riding patterns recognition for Powered two-wheelers users’ 2013 Pattern recognition
behaviors analysis
Automated Traffic Monitoring Using Image Vision [13] 2018 Traffic regulation
DeepWiTraffic_Low Cost WiFi-Based Traffic Monitoring 2019 Traffic regulation
System Using Deep Learning [22]
1602 F. Jalti et al.

a) Instrument Development and Safety: technological development has allowed

motorized two-wheelers to evolve in terms of performance and safety, namely:
– Electronic injection: equipping a good part of the fleet, it allows consumption
of less than one liter per 100 km for certain models and better acceleration
performance [23]
– Automatic or double-clutch transmissions: If the continuously variable model
(CVT), is still widely used, some manufacturers have developed much more
complex systems, these allow a shift automated reports managed by electronics
[24].
– Ride-by-Wire (TCS, driving modes, anti-wheeling): The electronic accelerator
(Ride-by-Wire) constitutes one of the main advances of recent years. Appeared
around 2000s, this system replaces traditional mechanical cables with an elec-
tronic interface between the handle and the motor. It’s managed by an Electronic
Control Unit (ECU) [25].
– Braking: In the EU, the obligation of ABS for cylinder capacity >125 cm3 is
legislated (European regulation COM 2010/542). However, this is not the case
for smaller displacements and is not necessarily applicable outside Europe.
– Helmets: After the initial development of the modular helmet in 1980, it is in
terms of comfort that the evolution has been felt the most. Today, expectations are
raised at the level of the arrival of head-up display technologies which will allow
speed and navigation information to be found directly on its visor.
– Protective equipment: Anti-torsion and anti-reversal clothes had benefit of the
advances of materials. e.g.: the airbag vest but accessible only for high-end
vehicles [26].
b) ARAS and AI Applications: The development of ARAS has always been a real
challenge for actors in the field. Given the complexity of the dynamics of this type
of vehicle, the cost and integration constraints. In other hand, the applications AI
technics are mainly seen at the level of image processing, this is partly explained by
the opportunity offered by Databases to constitute a repository for the machine.
In this context we have listed the following scopes:
– Connectivity: today is manifested by the connection that certain systems offer
with smart phones. These systems are accessible for medium ranges of vehicles;
besides, projects regarding inter-vehicle communication are in progress to ensure
communication between vehicles to signal the approach of a vehicle, using wifi
or 5G waves [22].
– Stability control: developed primarily for high-end vehicles, the MSC is practi-
cally the best-known system in this field. However, a compressed gas propulsion
system is being developed to overcome the fall in a maximum tilt situation [27].
– Blind spot warning: still for a relatively small proportion of the motorcycle fleet,
retro-vision blind spot assistance is available today.
– The electric motorcycle: several manufacturers have embarked on this path with
technological variability, however what characterizes the electric motorcycle is
its low noise level, which contributes to the sound comfort of the rider
The Potential Outcomes of Artificial Intelligence ... 1603

– Co-pilot: This project under development consists of driving assistance through

dialogue with the machine.
– Motobot: Promising example in the attainment of a level of skills by the machine
equal to that of the man. Base on AI this machine succeeded the imitation, by a
rate superior to 70%, of the Number 1 of pilots in 2018 [28]

4 Discussion

The analysis of the state of the art around the PTW aims to constitute an image of
the current situation of the ecosystem as a whole and in relation to the potential of
AI. The accident side for two-wheeled vehicles was addressed to identify areas of
weakness for this mode of transport. After the picture is completed by the synthesis
of the various intelligent transport systems, the objective is to gauge the level of
technological advance experienced by this transport sector.
Our system being a triplet {Driver, Vehicle, Environment}, the driver influences
the dynamics of the motorcycle by his actions (torque, roll and angle of inclination),
the interaction with the infrastructure is manifested by the tire/road contact. However,
the driver’s interaction with the environment is amplified [2], hence the need to
improve the level of perception of the system by acting on the reliability and the
quality of the information captured (which until today is mainly captured by the
driver).
Nowadays, ITS for PTWs are only accessible for a small percentage of the fleet,
even for an instrument like ABS which must be fitted to practically all cars manu-
factured, the latter is not necessarily accessible for most motorcycles put on the
market.
By examining topics addressing AI, although we can require an investment in
graphics processing, but the use of image processing associated with ML made
it possible to achieve a fairly high performance (latency <= 0.3 s) [14], also for
articles [7, 16, 18] we have successively dealt with: the problem of lateral dynamics,
cognitive load when driving, obstacle detection. These projects were developed to
be embedded and used image processing associated with machine / Deep learning
algorithms,
Also, the machines equipped with AI have been able to achieve very remarkable
performances, indeed if we are able to reproduce the biker robot “Motobot”, which
has learned through machine learning to achieve almost the same performances
of a Human being, we can think of generalizing learning with exchange between
machines. the idea is not to produce a robot for each machine but rather to consider
the machines as sensors and exploit the potential of connection technologies to ensure
complementarity between the different vehicles. In this sense we know an emerging
field, it is that of AI.
This leads us to assert the perspectives that Artificial Intelligence can play in
the development of driver assistance, particularly to improve the management of
problems with loss of control, speed, wearing of helmets, visibility and alertness.
1604 F. Jalti et al.

5 Conclusion and Perspectives

Given that the software and hardware development framework is nowadays depen-
dent on the manufacturer’s own design, it’s difficult to achieve interesting perfor-
mances and the price will always be exorbitant. But if we had a common frame-
work, we could hope for faster and more interesting development. Adding to this
that connection technologies will allow access to a larger database for an artificial
learning.
Joining the general trend in this field, the use of AI by object recognition remains
the main orientation of the projects developed for the benefit of PTW. Is this the only
lead? definitely no. In fact, the acquisition of data by sensors (whether PTW or other
means of transport) in different forms (image, sound, electrical or electromagnetic
signal,…) can constitute a real-time database.
In our context, the perspectives that we have drawn up will be focused on the
measurement of the roll angle for stability in curves as well as the detection of
obstacles.

References

1. Bekiaris ED (2009) SAFERIDER Project: New safety and comfort in Powered Two Wheelers.
Humans System Interaction, IEEE
2. Kuschefski A (2011) Advanced rider assistance systems for powered two-wheelers (ARAS-
PTW). In: 22nd international technical conference on the enhanced safety of vehicles (ESV),
Germany Paper Number 11–0123, NHTSA
3. Cossalter V (2006) Motorcycle dynamics. Second edition, Copyright by Vittore Cossalter
4. Damon P-M (2017) Lateral & steering dynamics estimation for single track vehicle:
experimental tests. IFAC, pp 2405–8963
5. Slimi H (2012) Système d’assistance a la conduite pour vehicules a deux-roues motorises.
Automatique/Robotique. Universitéd’Evry-Val d’Essonne
6. Manussawee P (2018) Detection of moving vehicle spacing. In: The 18th International
Symposium on Communications and Information Technologies ISCIT, IEEE
7. Wang J (2018) Road targets recognition based on deep learning and micro-doppler features.
In: International Conference on Sensor Networks and Signal Processing (SNSP), IEEE
8. Attal F (2015) Powered two-wheeler riding pattern recognition using a machine-learning frame-
work. IEEE Trans Intell Trans Syst 16(1):475–487. https://doi.org/10.1109/TITS.2014.234
6243
9. Krug E, Billingsley S, Watson B, Irigoyen JL (2017) Powered two- and three-wheelers safety: a
road safety manual for decision-makers and practitioners. 3.0 IGO, ISBN 978–92–4–251192–5,
World Health Organization
10. Sachin K (2018) Responsible factors of powered two wheeler ACCIDENT _ a review. In:
International Conference RICE, IEEE
11. Haworth N (2012) Powered two wheelers in a changing world - Challenges and opportunities.
Accident Analysis and Prevention, Elsevier
12. Nair Pravin R (2011) Performance evaluation of HMM and neural network in motorbike fault
detection system. In: International Conference ICRTIT, IEEE
13. Krishnamoorthy R (2018) Automated traffic monitoring using image vision. In: International
Conference ICICCT, IEEE
The Potential Outcomes of Artificial Intelligence ... 1605

14. Khandelwal Y (2020) Automatic braking system for two wheeler with object detection and
depth perception. In: International Conference ICCES, IEEE
15. Hanhe L (2019) Helmet use detection of tracked motorcycles using CNN-based multi-task
learning. Creative Commons Attribution 4.0 License, IEEE
16. Simon H (2019) Learning a curve guardian for motorcycles. In: Intelligent Transportation
Systems Conference (ITSC), IEEE
17. Ranjana P (2019) Machine learning algorithm in two wheelers fuel prediction. In: International
Conference on Innovations in Information and Communication Technology, IEEE
18. Ryoichi O (2015) Estimation of mental workload during motorcycle operation. In: 6th
International conference on applied human factors and ergonomics AHFE, ELSEVIER
19. Sarikan SS (2017) Automated vehicle classification with image processing and computational
intelligence. In: Complex Adaptive Systems Conference, ELSEVIER
20. Felix Wilhelm S (2020) Detecting motorcycle helmet use with deep learning. ELSEVIER
(2020)
21. Kuo-Wei S (2006) Applying knowledge structure to the usable fault diag-nosis assistance
system: a case study of motorcycle maintenance in Taiwan. Expert Systems with Applications,
ELSEVIER
22. Won M (2019) DeepWiTraffic_low cost wifi-based traffic monitoring system using deep
learning. In: International conference on mobile ad hoc and sensor systems, IEEE
23. Konrad JW (2017) The impact of the modernization of the injection-ignition system on
the parameters of motion of the motorcycle. In: XXI international polish–slovak conference
machine modeling and simulations 2016, ELSEVIER
24. Kuen-Bao S (2005) Design and implementation of a novel hybrid-electric-motorcycle
transmission. ELSEVIER
25. Daniel A-H (2020) Influence of catalyst, exhaust systems and ECU configurations on the
motorcycle pollutant emissions. ELSEVIER
26. Lindsay EN (2019) Do riders who wear an air jacket in equestrian eventing have reduced injury
risk in falls? A retrospective data analysis. J Sci Med Sport ELSEVIER
27. Patrick S (2010) Perspectives for motorcycle stability control systems. ELSEVIER
28. Bobtsov, AA (2013) Control approaches for complicated self-unstable plants with applications
for two-wheel mobile system. In: Conference on Manufacturing Modelling, Management, and
Control, IFAC
29. Jordan MI, Mitchell TM (2015) Machine learning: trends, perspectives, and prospects. Science
349(6245):255–260
30. Mitchell TM, et al (1997) Machine learning. Burr Ridge, IL: McGraw Hill 45(37):870–877
The Design of an EEG Bio-Amplifier
for the Primary Motor Cortex Zone
in 180 nm CMOS Technology

Younes Laababid , Karim El Khadiri , Hassan Qjidaa ,

Ahmed Lakhssassi, and Ahmed Tahiri

Abstract This article presents an amplification circuit for Electroencephalogram

(EEG) signals related to the intention of movements in the primary motor cortex.
The proposed design consists of a differential amplifier designed from an operational
amplifier in 180 nm CMOS technology. It amplifies signals at low amplitudes 20 μV
and includes a protection circuit as well as a bandpass filter making it possible to
target the frequencies associated with the signal of interest. The system offers a
differential gain of 60 dB, a Common Mode Rejection Ratio (CMRR) of 140 dB,
and the equivalent noise referred to its input (IRN) is 96 nV at 14 Hz.

Keywords EEG · AOP · CMOS · CMRR · IRN · The primary motor cortex

1 Introduction

Neuromuscular pathologies affect the nerves associated with the voluntary movement
of muscles. In the case of motor dysfunctions, there is a disruption of the neural
pathways between the primary motor cortex, where movements are initiated, the
spine and muscles. This ensures that the desired voluntary movement is not translated
into actual movement of the limbs [1].
A system for amplifying and extracting signals from the primary motor cortex
is proving to be very interesting for converting the movement intentions into the

Y. Laababid (B) · K. El Khadiri · A. Tahiri

Laboratory of Computer Science and Interdisciplinary Physics (LIPI), Normal Superior School
Fez (ENSF), Sidi Mohamed Ben Abdellah University, Fez, Morocco
e-mail: [email protected]
H. Qjidaa
Laboratory of Computer Science, Signals, Automation and Cognitivism (LISAC),
Department of Physics, Faculty of Sciences Dhar El Mahraz, Sidi Mohamed Ben Abdellah
University, Fez, Morocco
A. Lakhssassi
Department of Computer Science and Engineering, University of Quebec in Outaouais (UQO),
B-2014, Pavillon Lucien-Brault, Gatineau, Canada

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1607
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_146
1608 Y. Laababid et al.

real movement of a prosthesis, for example. To do this, several types of signals can
be used: EEG, ECOG and action potential. Unlike ECOG devices and intercellular
action potentials, EEG systems have the advantage of being non-invasive, as they do
not require surgery and the electrodes are not prone to gliosis, an immune process
that results in scarring on the surface of the electrode over time [2]. In contrast, EEG
systems must acquire signals through the skull and the scalp, thereby limiting spatial
and temporal resolution. This makes signals more difficult to acquire and process to
obtain the desired information, that is, the will to move.
In this sense, this paper proposes an EEG bio amplifier designed for signals from
the area of the primary motor cortex.
The remainder of this work is organized as follows. Section 2 gives an overview of
the proposed EEG acquisition System design as well as the schematics of the building
blocks. Simulation results and comparisons are presented in Sect. 3. Finally, Sect. 4
describes the conclusion and future work.

2 Design of Amplification System

2.1 Protection Circuit

When acquiring EEG signals, pulses of much greater amplitude than that of the
interest signals can damage the amplifier. These voltage spikes can be caused by
static discharge upon contact between a charged person and the acquisition system
input.
The electrical protection of the patient is also a concern to have. Thus, a protective
circuit is important to prevent damage to the electrical components while ensuring
user safety.
The protection circuit consists of two parallel diodes to limit the amplitude of
the voltage. The diodes chosen are suitable for low amplitude signals (<200 mA).
Since the transition from non-conductive to conductive state is not instantaneous,
signal distortion may be visible from 300 mV [3]. Since the signals of interest are
of amplitude from 25 to 100 μV, they will not be altered by the protection circuit
illustrated in Fig. 1.
When there is no presence of high amplitude unwanted pulses, the diodes do not
conduct and do not affect the EEG signal. On the other hand, when a pulse of more
than 200 mV is present, the diodes start to conduct the current, so that the system
acts as a short circuit in series with a constant voltage source [3].
Resistances are presented at the input to limit the current and thus provide protec-
tion to the patient against the defects of the amplifier. The Association for the
Advancement of Medical Instrumentation (AAMI) has developed a national standard
in the United States on safe current limits for electromedical devices. The leakage
current limit to the patient is 50 μA [4]. If a short circuit occurs between the VDD =
1.8 V power supply of one of the amplifiers and the electrode input, the resistances
The Design of an EEG Bio-Amplifier for the Primary Motor Cortex Zone … 1609

Fig. 1 Protection circuit

Table 1 Transistors
Transistors Type W/L (μm)
dimensioning
M1, M2 PMOS 1/0.18
M3, M4 NMOS 0.45/1
M5, M9 PMOS 2/0.18
M6 NMOS 0.5/1
M7 PMOS 3/0.18
M8 NMOS 0.45/1
M10 PMOS 1/0.18

will protect the patient.

U max 1.8 V
R= = = 45 K (1)
If 40 μA

2.2 Operational Amplifier

The proposed Operational Amplifier (AOP) was designed with CMOS 180 nm tech-
nology, shown in Fig. 2. It consists of a right bias circuit, a differential stage in the
center and an output stage on the left. To improve the performance of the AOP, a
PMOS cascade current mirror has been added to a standard two-stage configuration.
This arrangement provides high gain, high bandwidth, high slew rate, high stability
and high input impedance [5].
Since noise is a major concern for an EEG signal acquisition system, PMOS
transistors have been chosen as input. This type of transistors generates less 1/f noise
than NMOS [6]. The dimensioning of the AOP transistors is presented in Table 1. The
1610 Y. Laababid et al.

Fig. 2 Operational amplifier (AOP)

choice of the dimensions of the transistors has been made with particular attention
to the equivalent noise referred to the input to minimize it [7].

2.3 Instrumentation Amplifier

The proposed instrumentation amplifier consists of three operational amplifiers

according to the configuration (see Fig. 3).
The first part acts as a buffer for each of the inputs while the second part allows
for differential amplification. This configuration has a higher CMRR, a low offset
voltage, a higher input impedance and a higher gain. The gain of the instrumentation
amplifier is determined by the resistors R5 and R6 (see Fig. 4).

2.4 Filters

To eliminate interference from electromagnetic noise of 60 Hz and to isolate the

signal of interest as much as possible, a 14–30 Hz Bandpass filter has been designed.
It is shown in Fig. 4, to the right of the Instrumentation Amplifier.

2.5 Subservience to the Right Leg

A right leg subservience system has been integrated into the circuit to reduce
common-mode interference. The human body is a good conductor, it acts as an
antenna which can pick up the surrounding electromagnetic noise and thus interfere
with EEG signals. Since the signals of interest are very weak, this interference greatly
affects the quality of the signal. The subservience circuit of the right leg illustrated in
The Design of an EEG Bio-Amplifier for the Primary Motor Cortex Zone … 1611

Fig. 3 Configuration of the instrumentation amplifier

Fig. 4 The complete circuit of the EEG acquisition system

the lower part of Fig. 4 makes it possible to cancel this interference and thus obtain
a clean signal at this level. The common-mode voltage of the body is first perceived
and averaged by the resistors R5 and R6. It is then amplified and inverted before
being returned to the body by the E3 electrode. This negative feedback loop reduces
common-mode voltage [3].
The cables connecting the signal sources and the amplification system are often
shielded to reduce the effects of ambient interference. Active shielding consists of
bringing the shielding of cables to the same potential as that of their internal conductor
using a “shield driver” circuit [8]. The latter consists of a follower amplifier allowing
the common mode to be extracted from the input signal and the shielding of the cables
to this potential [9]. This active shielding is represented in burgundy (see Fig. 4).
1612 Y. Laababid et al.

3 Simulation Results

3.1 Protection Circuit

The Protection Circuit has been simulated to verify that high amplitude pulses are
limited in voltage before entering the instrumentation amplifier. The protection circuit
is used to clip the input signal.
Figure 5a illustrates how the number of diodes used for clipping the input signals
of 600 mV and 800 mV is insufficient for this protection operation.
Figure 5b illustrates how when by acting on the number of diodes used, the input
signals of 600 mV and 800 mV are both limited (clipped) to a value of around 450 mV
by the protection circuit.
A protection system makes it possible to limit the amplitude of pulses to protect
the circuit. Additional protection has been added to the system to limit the leakage
current to the patient to 40 μV to comply with safety standards for electromedical
devices.

Fig. 5 a, b. Effect of the protection circuit on the 600 mV and 800 mV input signals
The Design of an EEG Bio-Amplifier for the Primary Motor Cortex Zone … 1613

3.2 Operational Amplifier

AC Simulation. Is mainly used to obtain the frequency response of the circuit. This
simulation allows us to determine the differential gain, Phase Margin and CMRR of
the amplifier. The results are visible in Figs. 6, 7 and 8.
The Miller Compensation technique is essential to achieve stability in a closed
loop. For this reason, a compensation capacitor Cc has been added in series with a
resistance Rz to maintain good stability [10].
 
CL 1
Rz = 1 + × (2)
Cc gm6

The gain in Differential Mode of the 2-stage AOP.

Common Mode Gain.

Fig. 6 AC simulation

Fig. 7 Differential gain simulation

1614 Y. Laababid et al.

Fig. 8 Common mode simulation

3.3 Performance of the EEG Acquisition System

The proposed EEG acquisition system has been tested using simulations with
Cadence software.

Instrumentation Amplifier. Figure 9 illustrates how the system amplifies amplitude

signals of 20 μV with a gain of 1000. The transient simulation results of the system
by viewing a single input.
In Fig. 10, we will see how the instrumentation system operates without the
involvement of a protective circuit towards the patient.
In Fig. 11, we will see how the instrumentation system acts with patient protection.

AC Simulation of the Instrumentation Amplifier. The Ac response of the system

can be observed in the following Figs. 12 and 13 showing the behavior of the system
in common mode, the CMRR as well as the IRN.
The system has a common gain of −87 dB, so the CMRR becomes 146 dB
according to the following relation. The IRN of 96 nV at 14 Hz.

AD
CMRR = 20Log (3)
AC

Fig. 9 Transient response from the instrumentation AOP

The Design of an EEG Bio-Amplifier for the Primary Motor Cortex Zone … 1615

Fig. 10 Transient response without the protection of the patient

Fig. 11 Transient response with the protection of the patient

Fig. 12 Common mode simulation of the system

Filter
The filter has also been tested to verify that the bandwidth is between 14 Hz and
30 Hz, the frequency range of the signals of interest in the primary motor cortex (see
Fig. 14 and Table 2).
1616 Y. Laababid et al.

Fig. 13 IRN simulation of the system

Fig. 14 Bandpass filter AC response

Table 2 Measured performance summary and comparison

Parameters This work [11] [12] [13] [14] [15] [16]
Technology 180 nm 180 nm 180 nm 180 nm 180 nm 50 nm 65 nm
Supply voltage ±0.9 V ±1.8 V ±1.8 V ±1.5 V NA NA ±1 V
Gain 60.96 dB 64 dB 38.4 dB 47.6 dB 67 dB 45 dB 69.8 dB
Gain bandwidth 2.65 MHz 1.98 MHz 5.1 KHz NA 1.75 Hz 1.1 MHz 100 Hz
Phase margin 63.24° 61° NA NA NA NA NA
CMRR 146 dB 110 dB 70 dB 105 dB 92 dB 75 dB 84 dB
IRN 96 ηV 1.38 μV 5.01 μV 1.22 μV 89 μV 22 μV 0.81 μV

4 Conclusion

An instrumentation amplifier was built with three operational amplifiers designed

with transistors in CMOS 180 nm technology. The system can amplify signals as
weak as 20 μV with a differential gain of 60 dB, CMRR of 146 dB and The IRN is
96 nV.
The Design of an EEG Bio-Amplifier for the Primary Motor Cortex Zone … 1617

The proposed acquisition system could be improved with the use of dynamic
offset cancellation techniques such as automatic zeroing or chopping. This would
reduce the output signal offset with the system proposed in this article. Another
possible improvement would be to reduce the size of the operational amplifiers. To
minimize the noise of the transistors, relatively large dimensions have been chosen,
which means that the overall size of the EEG acquisition system is larger.

References

1. Lena HT, Hillel JC, Randy D (2015) Neuromechanical principles underlying movement
modularity and their implications for rehabilitation. Cell Press 86:38–53
2. Polikov VS, Tresco PA, Reichert WM (2005) Response of brain tissue to chronically implanted
neural electrodes. J Neurosci Methods 148(1):1–18
3. Webster JG (2010) Medical instrumentation. Application and design, 4th edn. Wiley
4. Olson WH (1978) Electrical safety. Med Instrum Appl Des 4:713
5. Gupta P, Mehra R (2016) Low power three-stage operational transconductance amplifier design.
Int J Adv Res Comput Commun Eng 5(7):314–320
6. Allen PE, Holberg DR (2002) CMOS analog circuit design. 2nd edn. Oxford University Press
7. Akhil G, Shweta S (2018) Design of two-stage CMOS Op-Amp with high slew rate and high
gain in 180nm. In: IEEE proceedings of the second international conference on I-SMAC (IoT
in social, mobile, analytics and cloud), pp 341–345
8. Al Nasser E (2014) The stability analysis of biopotential measurement system equipped with
driven-right-leg and shield-driver circuits. IEEE Trans Instrum Meas 63(7):1731–1738
9. Spinelli EM, Reverter F (2010) On the stability of shield-driver circuits. IEEE Trans Instrum
Meas 59(2):458–462
10. Jacob R (2010) CMOS circuit design layout and simulation. 3rd edn. IEEE Press
11. Janaki K, Beno A (2017) Noise immune bio potential amplifier for neural amplifying
applications. Int J Adv Res Electr Electron Instrum Eng 6(5):3310–3316
12. Gubta L, Kumar A (2018) Comparision on low-noise neural signal amplifiers and design of
current mirror OTA for EEG. Int J Pure Appl Math 119(12):14769–14784
13. Chung-Jae L, Jong-In S (2019) A chopper-stabilized amplifier with a tunable bandwidth for
EEG acquisition applications. IEEE Access J 7:73165–73171
14. Akshay G, Gurmohan S (2013) Novel high gain low noise CMOS instrumentation amplifier
for biomedical applications. In: International conference on machine intelligence research and
advancement, vol 83, pp 392–396
15. Manish G, Smriti K (2011) DC suppressed high gain active CMOS instrumentation amplifier
for biomedical application. In: IEEE proceedings of Icetect, pp 747–751
16. Chung-Yu W, Chia-Shiung H (2015) An 8-channel chopper-stabilized analog front-end ampli-
fier for EEG acquisition in 65-nm CMOS. In: IEEE Asian solid-state circuits conference, vol
19, no 5 (2015)
Design of a Non-inverting Buck-Boost
Converter Controlled by Voltage-Mode
PWM in TSMC 180 nm CMOS
Technology

Anas Boutaghlaline , Karim El Khadiri , Hassan Qjidaa ,

and Ahmed Tahiri

Abstract This paper presents a design of a DC-DC non-inverting Buck-Boost

converter controlled by a voltage-mode pulse width modulation (PWM) in TSMC
180 nm CMOS technology for a solar energy management system. Whatever the
weather conditions or the sunlight incident angle, the presented power converter
delivers to the load, a regulated DC voltage. Therefore, this system converts the input
voltage into higher or lower voltages. Since the conventional Buck-Boost converter
does not serve the needs of applications where a positive polarity is required. A four
switches topology is implemented in this work to keep the polarity unchanged. A
voltage-mode PWM control ensures the output voltage regulation by generating the
appropriate PWM signal duty cycle following the input voltage value. The designed
power converter delivers an output voltage in the range of [423.2—526.38 mV]
for different input voltages ranging from 200 to 800 mV. The output current range
is [41.92—52.43 mA]. The efficiency of the presented work ranges from 70.99 to
83.08%. The design area is 0.043 mm2 .

Keywords Battery management system · Non-inverting Buck-Boost converter ·

Solar energy · Voltage-mode PWM control

1 Introduction

Renewable energies, such as solar, wind and others, are promising solutions for
ensuring access to clean energy [1]. In several applications related to the production
of electrical power from a renewable source and battery management systems [2], the
DC-DC converters are required to remedy the instability of the produced energy by

A. Boutaghlaline (B) · K. El Khadiri · A. Tahiri

Laboratory of Computer Science and Interdisciplinary Physics, ENS, Sidi Mohamed Ben
Abdellah University, Fez, Morocco
e-mail: [email protected]
H. Qjidaa
Laboratory of Computer Science, Signals, Automation and Cognitivism (LISAC),
Department of Physics, FSDM, Sidi Mohamed Ben Abdellah University, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1619
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_147
1620 A. Boutaghlaline et al.

these conversion devices. In the case of solar energy, the photovoltaic cell produces
a DC voltage which varies according to weather conditions and the sunlight incident
angle, to regulate the produced voltage, a Buck-Boost converter is required to satisfy
the need of application [3].
As shown in Fig. 1, a Buck-Boost converter in his conventional topology, uses a
switch and a diode to deliver a DC output voltage with inverted polarity [4], lower or
higher than its input voltage, depending on the switching signal duty cycle controlling
the switch state [5].
To overcome the problem of output voltage inverted polarity where a conventional
Buck-Boost converter is used, another Buck-Boost architecture that combines a Buck
converter and a Boost converter to keep the output voltage polarity same as the input
polarity [5].
As shown in Fig. 2, there are three operation modes where the converter can
operate: Buck, Boost and Buck-Boost. In the case where the input voltage Vi is
greater than the output voltage Vo , the converter operates in Buck mode. When Vi is
lower than Vo , the converter operates in Boost mode. This converter can also work
in Buck-Boost mode when the value of the voltage Vi is the same as the voltage
Vo at the output [3]. A non-inverting Buck-Boost converter can work in continuous
conduction mode (CCM). The current through the inductor flows continuously in
this mode, as shown in Fig. 3 (a), only the duty cycle affects the output voltage.
In discontinuous conduction mode (DCM), the current through the inductor falls to
zero during the off time of the switching signal period, as shown in Fig. 3 (b), so the
output voltage depends on some factors other than the duty cycle, such as the output
current average, the switching frequency and the inductor value [6].

Fig. 1 Conventional Buck-Boost

Design of a Non-inverting Buck-Boost Converter Controlled … 1621

Fig. 2 Buck-Boost operation modes (a) Boost mode (b) Buck-Boost mode (c) Buck mode

Fig. 3 Inductor current (a) inductor current in CCM (b) inductor current in DCM

The limit between these two conduction modes is when the current flowing
through the inductor is canceled just at the time of switching. The chosen value
of the inductor defines the conduction mode. By calculating the critical value of the
inductor where the converter operates within the limit between the two conduction
modes, by choosing a higher inductor value than the calculated value, the converter
operates in CCM, which induces a strong load current. If the value of the inductor
is chosen to be lower than the critical value, the non-inverting Buck-Boost converter
works in DCM, causing a weak load current.
As already mentioned, the duty cycle defines the converter operating mode. There-
fore by setting the appropriate value of the duty cycle, the converter output voltage
reaches the target voltage. In this sense, two control modes exist, voltage-mode
control and current-mode control [7]. Current-mode control takes into account the
output voltage delivered by the non-inverting Buck-Boost converter and the current
flowing through the inductor to control the switches. As for voltage-mode control,
only the converter output voltage is taken into account. Current-mode control has
several advantages in terms of improved stability in closed-loop and rapid tran-
sient response than control in voltage-mode. The advantages of the voltage-mode
control boil down to the simplicity of the feedback loop design and analysis, and a
1622 A. Boutaghlaline et al.

large amplitude of the ramp signal provides noise immunity for a stable modulation
process [6]. Among the voltage-mode control techniques, Pulse Width Modulation
(PWM), which is a control technique that consists of varying the clock pulse width,
which controls the switches. The operation principle of this technique based on a
DC voltage generation representing the error between the voltage at the output of
the Buck-Boost converter and the reference voltage. This error is compared with a
ramp signal to produce a clock signal with a sufficient duty cycle to correct the output
voltage deviation from the target voltage [8]. Another control technique exists, called
Pulse Frequency Modulation (PFM). This method is widely used to improve the effi-
ciency of Buck-Boost converters for low load currents. The PFM control technique
is based on the idea of varying the frequency of the clock signal controlling the
switches, so the power transmitted to the load is linked to the switching frequency
[9]. The PWM control technique is used for the high load current. In the case of
low load current, this technique use affects the efficiency of the converter due to
switching losses, which involves the PFM technique use to increase significantly the
converter efficiency in the low load current case [10].
The objective of this work is to design a non-inverting Buck-Boost converter
controlled by the voltage-mode PWM technique, in TSMC 180 nm CMOS tech-
nology. The input of this converter is derived from a photovoltaic cell. The methods
leading to the realization of this work are presented in Sect. 2. The completed results
are presented and discussed in Sect. 3. Section 4 presents the layout of the designed
Buck-Boost converter. Finally, this work is concluded in Sect. 5.

2 Methods

Figure 4 presents the block diagram of this work. The non-inverting Buck-Boost
converter is designed assuming the CCM. The converter input voltage is derived
from a photovoltaic cell. The clock signal generated by the control circuit, with a
1.8 V amplitude, and a 10 MHz frequency, to control the MOSFET switches. This
study is carried out and simulated with Cadence Virtuoso tool, using TSMC 180 nm
CMOS technology.

2.1 Non-inverting Buck-Boost Converter

As shown in Fig. 5, this converter includes four MOSFET switches. During the ON
time, Q1 and Q2 switches are simultaneously closed, then the inductor starts storing
energy, so Q3 and Q4 switches are open. During the OFF time, Q3 and Q4 switches
are closed, the load capacitor is charged by the energy stored in the inductor, so the
switches Q1 and Q2 are open. The presented converter is designed according to the
following specifications (Table 1):
Design of a Non-inverting Buck-Boost Converter Controlled … 1623

Fig. 4 Block diagram of the proposed study

Fig. 5 Non-inverting
Buck-Boost

Table 1 Buck-Boost design

Specifications Value
specifications
Input voltage range 200—800 mV
Output voltage 500 mV
Output current 50 mA
Switching frequency 10 MHz

The resistive load R0 value is defined from design specifications, then:

Ro = 10  (1)
1624 A. Boutaghlaline et al.

The converter output voltage is given by:

Dut y cycle
Vo = × Vi (2)
1 − Dut y C ycle

The switching frequency is 10 MHz. The critical value of the inductance is

obtained as follows:
Vi × Dut y cycle
L crit =   (3)
2 × f sw × Vo × 1 + VVoi

The output voltage ripple is fixed to be 1% of the output voltage.

The capacitor minimum value is calculated by:

Vo × Dut y cycle
Cmin = (4)
Ro × f sw × Vo

The converter efficiency is defined by:

Pout Vo × i o 
η= = (5)
Pin Vi × i i 

2.2 PWM Control Circuitry

Voltage-mode PWM control is chosen in this work. Figure 6 shows the PWM
control circuit. The main characteristic of this architecture is the simplicity of its
implementation.

Fig. 6 PWM control circuit

Design of a Non-inverting Buck-Boost Converter Controlled … 1625

Fig. 7 PWM signal generation

The feedback is provided to the control circuitry from the converter output by
a voltage divider. This scaled voltage is compared with the reference voltage by a
Schmitt trigger based on a two-stage operational amplifier, to generate a DC voltage
qualified as the output voltage deviation from the reference voltage. The obtained
error is compared with a ramp signal using a comparator to produce a clock signal
with an appropriate duty cycle and a switching period same as the ramp signal period
to correct the detected error, as shown in Fig. 7.
Two non-overlapping clock signals are required to control the non-inverting Buck-
Boost converter switches. These two clock signals must not switch at the same time,
to avoid the shoot-through problem over the converter switches. Due to the high input
capacitance of the MOSFET switches, a driver is required to control these switches.

3 Results

This work is implemented and simulated with Cadence Virtuoso tool. The power
converter is designed to deliver an output voltage of 500 mV for input voltages
ranging from 200 to 800 mV.
Figure 8 illustrates the transient response of the designed converter with voltage-
mode PWM control in Buck-Boost mode. The applied input voltage is 500 mV. The
output voltage reaches 501.54 mV and its ripple voltage is 3.53 mV with a 10  load
resistor. The output current average is 50.01 mA. The rise time and the settling time
are 5.46 µs and 6.15 µs respectively. The converter achieves in this mode 83.08%
efficiency.
When the non-inverting Buck-Boost converter operates in Buck mode, with
800 mV as input voltage, the converter provides 526.38 mV with an output voltage
ripple of 4.93 mV, for a resistive load of 10 , as shown. The output current average
is 52.43 mA. The efficiency is 82.90%, with a rise time of 4.66 µs, the settling time is
6.02 µs. The converter output voltage transient response in Buck mode is represented
in Fig. 9.
1626 A. Boutaghlaline et al.

Fig. 8 Output voltage transient response for Vi = 500 mV

Fig. 9 Output voltage transient response for Vi = 800 mV

For an input voltage of 200 mV, the designed converter operates in Boost mode.
The output voltage reaches 423.2 mV and its ripple voltage is 2 mV, for a resistive
load of 10 , as shown in Fig. 10. The output current average is 41.92 mA. The
settling time is less than 15.91 µs, and the rise time is 13.69 µs. The converter
efficiency is 70.99%.
The results shows the output voltage delivered by the Buck-Boost converter from
different input voltages (Table 2). In 200 mV and 800 mV as input voltages cases,
the designed converter approached the desired output voltage due to the designed
voltage-mode PWM controller. The controller generates a PWM signal with a closest
duty cycle value to the appropriate one, calculated using Eq. (2). For 200 mV input
voltage, the generated PWM signal duty cycle is 67.91%. For 800 mV input voltage,
the generated PWM signal duty cycle is 39.68%. Using Eq. (2), the appropriate duty
Design of a Non-inverting Buck-Boost Converter Controlled … 1627

Fig. 10 Output voltage transient response for Vi = 200 mV

Table 2 Designed converter simulation results summary

Parameter Value
Input voltage range 200—800 mV
Output ripple 2—4.93 mV
Output voltage 423.2—526.38 mV
Load current 41.92—52.43 mA
Efficiency 70.99—83.08% (@ Po = 25 mW)

cycle for 200 mV is 71.43%. In the case of 800 mV input voltage, the required duty
cycle is 38.46%.

4 Layout

The designed DC-DC non-inverting buck-boost converter and its voltage-mode PWM
control circuitry are implemented in TSMC 180 nm CMOS technology. Figure 11
shows the layout of the design presented in this paper. The layout area of this work
is 0.043 mm2 . The four NMOS switches occupy 65.43% of the total layout area. The
DC-DC power converter layout is done symmetrically, especially for the switching
clocks nets to avoid clock skew. The differential pairs and current mirrors layout is
done respectively with common centroid and inter-digitized matching techniques.
The power nets are routed with a larger width to minimize metal resistivity.
1628 A. Boutaghlaline et al.

Fig. 11 Layout of the presented work

5 Conclusion

In this paper, a non-inverting Buck-Boost converter with voltage-mode PWM control

is designed in TSMC 180 nm CMOS technology, implemented and simulated under
Cadence Virtuoso tool. The designed converter regulates the voltage provided by
a photovoltaic cell ranging from 200 to 800 mV, in continuous conduction mode.
The output voltage regulation is done with a voltage-mode PWM controller. The
simulations results shows that the designed converter delivers 423.2 mV in the case
where Vi is 200 mV (Boost mode) with 70.99% efficiency. When Vi is 800 mV (Buck
mode), the regulated output voltage reaches 526.54 mV with 82.90% efficiency. For
Vi of 500 mV (Buck-boost mode), the presented design provides 501.54 mV at its
output with 83.08% efficiency.
Design of a Non-inverting Buck-Boost Converter Controlled … 1629

References

1. Ogudo KA, Umenne P (2019) Design of a PV based power supply with a non inverting buck-
boost converter. In: IEEE PES/IAS PowerAfrica conference: power economics and energy
innovation in Africa, PowerAfrica 2019. Institute of Electrical and Electronics Engineers Inc.,
pp 545–549. https://doi.org/10.1109/PowerAfrica.2019.8928656
2. Ouremchi M, El Alaoui M, Farah F, El Khadiri K, Qjidaa H, Lakhssassi A, Tahiri A (2018)
Li-ion battery charger based on LDO regulator for portable device power management. In:
Proceedings of 2018 6th international renewable and sustainable energy conference, IRSEC
2018. Institute of Electrical and Electronics Engineers Inc. https://doi.org/10.1109/IRSEC.
2018.8702961
3. Ma J, Zhu M, Chen Y, Wang H, Cai X (2019) LPV-based cascade control for three mode
non-inverting buck-boost converter. In: 2019 IEEE 4th international future energy electronics
conference, IFEEC 2019. Institute of Electrical and Electronics Engineers Inc. https://doi.org/
10.1109/IFEEC47410.2019.9015003
4. El Khadiri K, Akhmal H, Qjidaa H (2017) Li-ion battery charging with a buck-boost DC-DC
converter for a portable device power management. J Low Power Electron 13:263–270. https://
doi.org/10.1166/jolpe.2017.1479
5. Bhat S, Nagaraja HN (2015) A novel DSP based non inverting buck-boost converter for PV
system. In: 2015 IEEE international conference on signal processing, informatics, communi-
cation and energy systems, SPICES 2015. Institute of Electrical and Electronics Engineers Inc.
https://doi.org/10.1109/SPICES.2015.7091475
6. Chen WW, Chen JF (2018) Review of the PWM control circuits for power converters. In: Power
Systems. Springer, pp 37–79. https://doi.org/10.1007/978-981-10-7004-4_2
7. El Khadiri K, Qjidaa H (2014) Inverting buck-boost DC-DC converter design challenges. J
Embed Syst Appl 4. https://doi.org/10.5121/ijesa.2014.4101
8. Huang J (2016) A new method of generate the PWM wave using analog circuits. https://doi.
org/10.2991/icsnce-16.2016.23
9. Nagarjun S, Debnath D, Chakraborty C (2018) Buck-boost buck CCM-DCM converter for PV
based DC standalone system. In: Proceedings of 2018 IEEE international conference on power
electronics, drives and energy systems, PEDES 2018. Institute of Electrical and Electronics
Engineers Inc. https://doi.org/10.1109/PEDES.2018.8707723
10. Chen HM, Chang RC, Lin KH (2013) A high-efficiency monolithic DC-DC PFM boost
converter with parallel power MOS technique. VLSI Des 2013:643293. https://doi.org/10.
1155/2013/643293
Low-Power High-Speed On-Chip 5 to 1
Serializer in 180 nm Technology

Aicha Menssouri , Karim El Khadiri , Hassan Qjidaa ,

Ahmed Lakhssassi , and Ahmed Tahiri

Abstract Nowadays, Traditional parallel communication cannot meet the standard

for high-speed links for inter-Integrated Circuits (IC) data transmission. On-Chip
SerDes Transceiver is a promising solution which can reduce the number of inter-
connects and offers remarkable benefits in context with power consumption, area
congestion and crosstalk. This article presents a description of the design of Low-
Power High-Speed On-Chip 5 to 1 Serializer architecture in 180 nm Technology that
is used to perform the data serialization function in a SerDes system. This architecture
is designed with a clock frequency of 160 MHz and the overall power consumption
is estimated as 2.02 mW. The proposed serializer was designed and simulated in
Cadence using TSMC 180 nm technology with a supply voltage of 1.8 V.

Keywords Low-power · High-speed · On-chip · SerDes

1 Introduction

The development of System-on-Chip (SoC) technologies allows the number of

modules integrated on a single chip to grow faster, requiring a large number of
parallel links and buses for interconnections and data communications between these
modules [1].

A. Menssouri (B) · K. El Khadiri · A. Tahiri

Laboratory of Computer Science and Interdisciplinary Physics (LIPI), Normal Superior School
Fez (ENSF), Sidi Mohamed Ben Abdellah University, Fez, Morocco
e-mail: [email protected]
H. Qjidaa
Laboratory of Computer Science, Signals, Automation and Cognitivism (LISAC),
Department of Physics, Faculty of Sciences Dhar El Mahraz, Sidi Mohamed Ben Abdellah
University, Fez, Morocco
A. Lakhssassi
Department of Computer Science and Engineering, University of Quebec in Outaouais (UQO),
Pavillon Lucien-Brault, Gatineau B-2014, Canada

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1631
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_148
1632 A. Menssouri et al.

However, traditional parallel data communications cannot meet the standard of

high-speed links between integrated circuits (ICs) for data transmission [2]. On the
one hand, they appear to be inefficient in terms of power supply problems. On the
other hand, the difference in arrival time of simultaneously transmitted data is delayed
due to the increased operating frequency of high-speed data links. Also, there is area
congestion and crosstalk that interferes between adjacent parallel data links [3].
A promising solution to overcome these problems is to replace parallel buses with
a serial link by using a SerDes transceiver, which allows incoming parallel data to
be mapped to a single data stream using a serializer prior to transmission. Serial data
are transmitted over the channel with a built-in clock. The data are received at the
other end of the channel and deserialized to generate a parallel data output [4].
The organization of this paper is as follows: Sect. 2 presents the architecture of
SerDes systems, Sect. 3 mainly focuses on the design of the proposed serializer. And
finally, Sect. 4 gives a conclusion.

2 The Architecture of SerDes Systems

Figure 1 shows the underlying architecture of a high-speed SerDes system, which

consists of two main entities: a transmitter and a receiver. The transmitter that sends
the data and which includes a serializer that serializes the n-bit parallel data, an
equalizer (FFE) ensures the excellent signal integrity of this serialized data, which
will subsequently be pushed into a differential signal pair. The receiver consists of a
differential receiver receiving the serial data, which is afterward fed into a clock data
recovery circuit (CDR) and an equalizer (DFE). The data is restored on a parallel

EQUALIZER
SERIALIZER FFE DRIVER

CHANNEL

DESERIALIZER EQUALIZER DRIVER

CDR
DFE

Fig. 1 Block diagram of the basic high-speed SerDes systems

Low-Power High-Speed On-Chip 5 to 1 Serializer in 180 nm Technology 1633

data bus through a deserialization circuit on the receiver block. In this article, we
focus on the serializer [5, 6].

3 The Design of the Proposed Serializer

3.1 Serializer Architecture

The serializer is used for the parallel to serial conversion of data. There are several
architectures to design it [7]. Figure 2 shows the basic schematic of a 2N:1 serializer
that allows serializing a pair number of data. We may think that a multiplexer can
perform the serialization operation. Still, in practice, we do not get the proper serial-
ization, because different synchronization requirements (setup time, hold time, other
types of delay) impose strict limits on the design. The primary internal blocks are
shown in Fig. 3. Two D flip-flops that transfer input data to the output with a delay
when there is a rising edge of the clock, a D-latch and a multiplexer are required to
design the latter [8].
Our proposed serializer is based on the architecture shown in Fig. 4. It is designed
using D and D-Latch flip-flops but replacing the multiplexer with parallel tri-state
inverters.

D1

2:1 Serializer
D0
Q0
CLK
2:1 Serializer
D3 Q1 CLK P0
2:1 Serializer OUT
D2 2:1 Serializer
Pn CLK
Q n-1
CLK

2:1 Serializer
Dn-1 Qn CLK

Dn 2:1 Serializer

Fig. 2 The basic schematic of a 2N:1 serializer

1634 A. Menssouri et al.

D1
D flip-flop out
2:1
mux

D0
D flip-flop
D-latch

CLK

Fig. 3 A 2:1 serializer circuit

Usually the multiplexer has pair number bits (2X --> 1) but in our work we need
an odd-number multiplexer, for this we propose to use a parallel tri-state (0, 1, HZ)
inverter, instead of a simple multiplexer in order to realize a serializer (5 --> 1).

3.2 Design of a Tri-State Inverter

Many logic gates require a tri-state output: high, low, and high impedance. The high
impedance state is also called the high Z state and is used to connect many gates
outputs to a single line, such as a data bus or address line. A potential conflict would
exist if more than one door output tried to control the bus line simultaneously. A
controllable high impedance state circuit solves this problem [9]. Figure 5.a shows
the schematic of a tri-state inverter proposed in our architecture.
When the level of the EN input is high (EN = 1), the component behaves like a
typical inverter, its output always gives a level complementary to that applied to the
“IN” input. On the other hand, when the level of the EN input is low (EN = 0), the
“out” output goes to the state called “high impedance”.
Figure 5.b shows the simulation obtained for the three-state inverter such that this
result appears that the output signal “out” (the third waveform from top) is the inverse
of the input signal “IN” (the second waveform from top). When the transmission gate
‘EN = 0’ (the first waveform from top) the output gives a state called high impedance.

3.3 D-Latch Design

A D-Latch primarily generates a delay to the input data. It is based on two tri-state
inverters. When CLK is high, the first tri-state inverter is active and the second is
blocked so that our output Q takes the value of the input D, which is said to be
transparent. If the clock CLK is low the first tri-state changeover switch is blocked
Low-Power High-Speed On-Chip 5 to 1 Serializer in 180 nm Technology 1635

C1

D1
D flip-flop
C0

C2

D2
D flip-flop
C0

C3
out
D3
D flip-flop
C0

C4

D4
D flip-flop
C0

C0

D0
D flip-flop
D-Latch
C0

D flip-flop D flip-flop D flip-flop D flip-flop D flip-flop

C0 C4 C2 C1
C3

CLK

Fig. 4 A 5:1 serializer proposed circuit

and the second one is active then the output Q, in this case, takes the previous state.
Figure 6.a shows the cadence virtuoso implementation of a D-Latch based on two
tri-state [10].
Figure 6.b verifies the logical behavior of D-latch with 0.18 µm technology. We
notice that if the clock “CLK” (the first waveform from top) takes the value “1”
the output “OUT” (the third waveform from top) takes the value of input “D” (the
second waveform from top) and if the clock is low “0” the output is memorizing the
chairman value.
1636 A. Menssouri et al.

VDD

IN OUT

EN

(a)

(b)

Fig. 5 A schematic (a) and simulation (b) of tri-state inverter

3.4 D Flip-Flop Design

D flip-flops are also called “delay flip-flops” or “data flip-flops”. They are used to
store 1-bit binary data. D flip-flops are one of the most commonly used flip-flops as
essential memory and zero-order elements.
Low-Power High-Speed On-Chip 5 to 1 Serializer in 180 nm Technology 1637

CLK

D OUT

CLK

D-Latch

(a)

(b)

Fig. 6 A schematic (a) and simulation (b) of a D-Latch based on two tri-state inverters

Our D flip-flop used are designed by connecting two series D-latches. When the
clock signal is high (rising edge) and if the D input is high, then the output will be
a delayed copy of the input signal, and it will also be high, and if the D input is
low, then the output will become small. If we do not apply any clock input to the D
flip-flop or during the falling edge of the clock signal, there will be no change in the
output. It will keep its previous value at output Q. The output Q, therefore, follows
the input D in the presence of a clock signal [11].
Figure 7.a shows a schematic of the D flip-flop that responds to the clock on its
rising edge.
1638 A. Menssouri et al.

CLK

D D-Latch D-Latch OUT

D flip-flop

(a)

(b)

Fig. 7 A schematic (a) and simulation (b) of a D flip-flop based on two D-latches

Figure 7.b shows the analog simulation waveform of flip-flop D. This analog
simulation shows the output voltage “OUT” (the third waveform from top), the input
voltage “D” (the second waveform from top) and clock voltage “CLK” (the first
waveform from top) as a function of time. If the clock voltage on the rising edge,
the output voltage copies the value of the input voltage otherwise the output voltage
stores the previous value.

3.5 Operation of the Proposed Serializer

Our proposed serializer shown in Fig. 8 was designed under Cadence virtuoso. It
consists of ten D flip-flop, five parallel tri-state inverters and a D-latch. The serial-
ization is done by selection signals generated by a register with ST and RST to rest
Low-Power High-Speed On-Chip 5 to 1 Serializer in 180 nm Technology 1639

Fig. 8 The proposed serializer implemented under cadence

circuit, one serial input and parallel circularly shifted outputs. Thus, the first output of
this register is used as the clock signal of the D flip-flop which shift the parallel data
to the parallel tri-state inverters with this high-speed clock signal. In turn, the parallel
tri-state inverters take these shifted data for serialization. The addition of a D-latch
is essential because the data obtained at the output of the D flip-flop is shifted by
small delays which results in the crosstalk of the data. However, the D-latch allows
generating a signal with a large delay to overcome the crosstalk problem. Figure 9
shows simulation result of the serializer block (Tables 1 and 2).

Fig. 9 Output of the proposed 5:1 serializer

1640 A. Menssouri et al.

Table 1 Design
Parameter Specification Unit
specifications for proposed
5:1 serializer Supply voltage 1.8 V
Data bus width 5 Bit
Clock frequency 160 MHz

Table 2 Performances
Performance [6] [5] This work
comparison between a
previous and this work CMOS technology 65 nm 180 nm 180 nm
Supply voltage (V) 1.2 1.8 1.8
Type of serializer 4:1 8:1 5:1
Power consumption (mW) 80 1216 2.02

4 Conclusion

This work presents a 5:1 serializer implemented in CMOS 180 nm technology is

designed, created, and simulated using Cadence Virtuoso, in this article we discussed
the reasons why we switched to use serial buses instead of parallel buses. The result
obtained can confirm that it worked well, also the comparison of our serializer power
consumption, which almost 2.02 mw with the previous work shows the efficiency of
this serializer.

References

1. Nivedita J, Radheshyam G (2015) Design of a new serializer and deserializer architecture for
on-chip SerDes transceivers. Sci Res 06(03):81
2. Jerry Y (2015) System-level design, simulation and measurement for high-speed data links, pp
1–3
3. Abdallah A, Mohamed A (2016) High speed serial link design (SERDES) introduction,
architecture, and application. academia.edu
4. Mohammed H, Maher A, Fawnizu A, Israel Y (2012) Design and FPGA implementation of
PLL-based quarter-rate clock and data recovery circuit. In: 4th international conference on
intelligent and advanced systems (ICIAS2012)
5. Vinod K, Shalini C, Monika G (2016) Design of high speed serializer for interchip data
communication with phase frequency detector. Int J Eng Res Appl 6(5):30–40
6. Huang K, Deng L, Ziqiang W, Xuqiang Z, Fule L, Chun Z, Zhihua W (2015) A 190mW 40Gbps
SerDes transmitter and receiver chipset in 65nm CMOS technology. IEEE
7. Tongsung K (2016) High-speed low-power transmitter using tree-type serializer in 28-nm
CMOS
8. Arunthathi G, Umamaheswari K, Vijeyakumar K (2018) Design of CMOS serialiser. J Netw
Commun Emerg Technol (JNCET) 8(6)
9. Jaume S, Charles F (2004) CMOS electronics: how it works, how it fails. IEEE Solid-State
Circuit Society Newsletter, Canada
Low-Power High-Speed On-Chip 5 to 1 Serializer in 180 nm Technology 1641

10. Kiran K, Fazal N (2017) Design of low voltage D-flip flop using MOS current mode logic
(MCML) For high frequency applications with EDA tool. IOSR J VLSI Signal Process (IOSR-
JVSP) 7(4):9–14
11. Casper L, Sorin C, Stamatis V (2004) Single electron encoded latches and flip-flops. IEEE
Trans Nanotechnol 3(2):237–248
Computing and Analyzing Through
Silicon Via (TSV) Noise Coupling
in 3D-IC Structures
Khaoula Ait Belaid , Hassan Belahrach , Hassan Ayad,
and Fatima Ez-zaki

Abstract In three-dimensional structures based on Through Silicon Via (TSV), an

important design consideration is coupling noise between TSVs in addition to TSV
and Active Circuits. Indeed, it is very important to estimate the mathematical form in
the design of reliable 3D-structures. This paper proposes a technique to compute the
TSV-TSV noise coupling in addition to the TSV-active circuit noise coupling. First
both models are presented, then, the used Finite Difference Time Domain (FDTD)
method is explained in detail for TSV-TSV noise, TSV-Active circuit noise, and TSV-
TSV noise with interconnect line. To validate the proposed mathematical models,
Matlab and Pspice simulation tools are used. The results given by the FDTD method
and Pspice tool are presented. These results show that the calculation of the FDTD
method and Pspice are in a good agreement.

Keywords 3D-ICs · TSV-TSV noise coupling · TSV-active circuit noise

coupling · Interconnect line · FDTD

1 Introduction

In the current and next decade, a major trend of microelectronics is to grow the inte-
gration density by utilizing three-dimensional (3-D) integration based on packages.
Presently, 3D structures are achieved by stacking several dies of integrated circuits
(IC), all these dies communicate with other sub-modules via vertical 3-D intercon-
nections. Hence, 3D interconnection is a good solution that helps 3D integration to
achieve the desired performance [1]. Higher power consumption and increased delay
caused by the large resistance and capacitance of the wires are the main problems

K. Ait Belaid (B) · H. Belahrach · H. Ayad · F. Ez-zaki

Department of Applied Physics, Electrical Systems, Energetic Efficiency
and Telecommunications Laboratory, Cady Ayyad University, Marrakesh, Morocco
H. Belahrach
Electrical Engineering Department, Royal School of Aeronautics (ERA), Marrakesh, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1643
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_149
1644 K. Ait Belaid et al.

that prevent the improvement of the integrated circuits quality [2, 3]. Indeed, 3D-
integration based on TSVs is a good solution for great integration permitting higher
system speed and lower power consumption [4]. 3-D integrated architecture based
on TSVs is an almost new technique and widely used. But these architectures have
some difficulties in their modeling and analysis. Therefore, a precise model of TSV
is crucial to improve 3D design performance. Several papers have proposed models
of TSV interconnections embedded in a substrate. In [5, 6], authors proposed a tech-
nique based on RF characterizations and simulations, which conduct to an analytical
model based on frequency and comprising MOS effect of high aspect ratio of TSV.
Others in [7] gave an equivalent lumped element model for different multi-TSV
structures and introduce closed form equations for resistive, inductive and capacitive
coupling structures. The closed form equations depend of physical dimensions and
material properties. An efficient method was detailed in [8] to model TSV inter-
connections. This procedure occurs an equivalent network parameter which covers
the combined effect of silicon substrate, insulator, and conductors. The method uses
Maxwell’s equation in integral form. Another equivalent circuit model of the chip-
to-chip vertical via, used in high frequency, and based on physical configuration of
the structure was presented in [9, 10]. The parameters of the model were extracted
and analyzed using S-parameters measurement plying a vector network analyzer up
to 20 GHz range.
One of the problems found in the modeling of 3-D technology is the modeling of
the substrate in high frequencies. Because of substrate distribution nature, it cannot
be converted into a dense analytical model. In general, the substrate coupling models
could be extracted from a full 3-D numerical simulation, based on an appropriate
discretization such as Maxwell’ equations [11]. In [12, 13], the box integration could
be used to derive a distributed RC network of the substrate. In this technique, an
equivalent circuit representation of the substrate is extracted using a 3-D rectangular
RC mesh. In [14, 15], to solve the obtained Maxwell’s equations the Boundary
Element Method is applied as an alternate approach.
In relation to the topic, authors in [16–18] proposed a TSV-TSV, and TSV-Active
circuit noise coupling models. Both of these noise coupling paths models were
extracted and analyzed using a 3-D transmission line matrix approach. A model for
coupling noise that takes into account substrate contacts between TSV in bulk CMOS
technologies is proposed in [19]. The suggested model is dense but has reasonable
accuracy for the compact substrate contacts. To study 3-D integrated circuit struc-
tures, based on TSVs, a computation of noise coupling is required [20–23]. Hence,
this work consists to compute this noise in time-domain. First, both models of TSV-
TSV noise coupling and TSV-active circuit noise coupling are presented. Then, a
method to calculate these coupling noises in the time-domain is given. The tech-
nique is the FDTD-1D. The rest of this paper is organized as follows. The TSV-TSV
and TSV-active circuit noise coupling models are presented in Sect. 2. Then, the
computing technique is explained in the same section. The results and simulations
are presented and examined in Sect. 3. The conclusions are drawn in the fourth
section.
Computing and Analyzing Through Silicon Via (TSV) Noise Coupling … 1645

2 Modeling and Computing TSV-TSV and TSV-Active

Circuit Coupling Noise in the Time-Domain Using FDTD

In general, the TSV is surrounded by an insulation layer. Although, this insulation

layer cannot avoid the appearance of coupling noise in high frequencies. Signal TSV
might be coupled to the silicon substrate, due to the presence of capacitance among
TSV metal. The noise in 3-D ICs structures can be coupled through different paths:
TSV to substrate, and TSV to another TSV. Both coupling paths are presented in this
second part of our work.

2.1 Noise Coupling Between Two TSVs

The 3-D ICs contain several TSVs; therefore, TSV-TSV coupling noise is a notable
problem. To compute the TSV-TSV coupling noise, a basic structure was used. This
structure contains two signal TSVs and two ground TSVs, that are linked by a ground
line, as depicted in Fig. 1 [16]. The TSV-TSV coupling noise might be presented
as an RLGC equivalent circuit as explained in [17]. Indeed, the structure in Fig. 1
can be presented by the lumped circuit given in Fig. 2. In this figure, the TSVs
are represented by Ctsv-total, and the substrate among the TSVs is presented by a
capacitance and resistance, except that between two ground TSVs. The ground TSVs
are linked using a metal line.
The lumped circuit model in Fig. 2 could be further reduced to the simplified
equivalent circuit model in Fig. 3. This simplified circuit contains only three essen-
tial elements: substrate capacitance, substrate resistance, and the equivalent TSV
capacitance. This equivalent circuit was studied with several terminations in Sect. 3.

Fig. 1 Coupling noise between two TSVs basic structure

1646 K. Ait Belaid et al.

Fig. 2 The equivalent lumped circuit model for coupling noise between two TSVs

Fig. 3 The reduced

equivalent lumped circuit
model for coupling noise
between two TSVs

2.2 TSV-Active Circuit Coupling Noise

In 3D-ICs, different silicon chips; with several MOS; are stacked using TSVs.
Hence, the coupling noise among the active circuit and the TSV must be consid-
ered and modeled. Some difficulties appear, when modeling the TSV-active circuit
noise coupling. An active circuit is composed of many CMOS devices and intercon-
nections. Each CMOS device is divided into substrate contacts well contacts, n-well,
deep n-well and gate metal. In this case, coupling between the substrate contact
and TSV is significant, due to the direct connection via the bulk silicon substrate.
Indeed, a substrate contact can be thought as an active circuit. Therefore, coupling
noise among active circuit and the TSV might be modeled as the coupling among the
contact and the TSV. TSV-active circuit coupling noise structure is depicted in Fig. 4.
In similar manner with the coupling noise between two TSVs, the basic structure
in Fig. 4 can be modeled by a lumped circuit model given by Fig. 5, then, reduced
to the simplified circuit in Fig. 6, which consists of three elements: the substrate
capacitance, the substrate resistance, and the total equivalent TSV capacitance [17].
Computing and Analyzing Through Silicon Via (TSV) Noise Coupling … 1647

Fig. 4 TSV-Active circuit coupling noise basic structure

Fig. 5 The equivalent lumped circuit model for TSV-Active circuit coupling noise

Fig. 6 The reduced

equivalent lumped circuit
model for TSV-Active circuit
coupling noise

2.3 TSV Noise Coupling Using FDTD

In general, TSVs are terminated by the Input/output drivers that must be considered.
Consequently, the circuit in Fig. 3 becomes as illustrated in Fig. 7.
The technique used to compute the TSV coupling noise is the FDTD method. By
observing Fig. 7 Eqs. (1)–(9) could be written.
1648 K. Ait Belaid et al.

Fig. 7 The equivalent lumped circuit model for coupling noise between two TSVs with I/O drivers

i(t) = i 1 (t) + i 2 (t) (1)

i 2 (t) = i 21 (t) + i 22 (t) = i 3 (t) + i 4 (t) (2)

d V1 (t)
i 3 (t) = C4 (3)
dt
V1 (t)
i 4 (t) = (4)
R3
d V2 (t)
i 1 (t) = C2 (5)
dt
V (t)
i 21 (t) = (6)
R2

V in(t) = R1 i(t) + V2 (t) (7)

V2 (t) = 2Vc1 (t) + V (t) + V1 (t) (8)


1
V (t) = i 22 (t)dt (9)
C3

Using Eqs. (1)–(9) and replacing each other, then, applying the FDTD-1D method,
Eqs. (10)–(12) are found.

Vin (n) = Rt

1 C2
(V2 (n) − V2 (n − 1))
(10)
+ t (V1 (n) − V1 (n − 1)) + RR13 V1 (n) + V2 (n)
R1 C 4

V2 (n) = V2 (n − 1) + AV1 (n) − BV1 (n − 1) + V (n) − V (n − 1) (11)

Computing and Analyzing Through Silicon Via (TSV) Noise Coupling … 1649

 
V (n − 1) + DV1 (n)
V (n) = C (12)
− DV1 (n − 1) + E V 1 (n)

Replacing Eqs. (12) and (10) in Eq. (11), the TSV-TSV noise coupling V1 is
obtained as Eq. (13).
 
− Vin (n) + aV2 (n − 1)
V1 (n) = −H (13)
+ bV1 (n − 1) + cV (n − 1)

Where: t is the sampling step which must be chosen carefully, A, B, C, D, E, H,

a, b, and c depend on the circuit parameters. In the similar manner, the TSV-active
circuit noise coupling in Fig. 6 was calculated. In this work, other circuit was studied.
This circuit consists to add an interconnect line next to signal TSV2. The interconnect
line is represented by its characteristic impedance Rc, and propagation delay τ. The
equivalent lumped circuit of the novel structure is depicted in Fig. 8, where:

E i (t) = V1 (t − τ ) + Rc i 4 (0, t − τ ) (14)

Er (t) = Vc4 (t − τ ) − Rc i 4 (l, t − τ ) (15)

Using Eqs. (1)–(9) in addition to (14)–(17) and applying the FDTD-1D method
the noise Vc4 (n) is found as Eq. (18).

Vc4 (t) = −Rc i 4 (l, t) + E i (t) (16)

D B C
V1 (n) = Vin (n) + Vc1 (n − 1) + Vc4 (n − m)
A A A (17)
G E F
+ Vc4 (n − m − 1) + Vc2 (n − 1) − V (n − 1)
A A A

Fig. 8 The equivalent lumped circuit model for TSV-TSV coupling noise with interconnect line
1650 K. Ait Belaid et al.

Vc4 (n) = −H Vc4 (n − 1) + I V (n − m) − J V (n − m − 1) + K V1 (n − 1) (18)

With: l is the length line, m represents the numerical delay, A, B, C, D, E, F, G,

H, I, J and K according to the circuit parameters.

3 Results and Discussions

In this section, the proposed method (FDTD) is verified using Matlab and Pspice
simulation tools. The objective is to realize the circuits of the TSV coupling on
Pspice and to program the equations found by the FDTD on Matlab. The proposed
FDTD method is based on discretization, indeed a good choice of the discretization
step is necessary. The circuit’s parameters used during the simulations are presented
by Tables 1 and 2.
A trapezoidal signal switching source from 0 to 1.8 V with rising/falling time of
50 ps at frequency 1 GHz was used. The simulated waveforms of each studied circuit
are depicted in Fig. 9, Fig. 10 and Fig. 11. Based on the results, one can see that the
proposed method is in good agreement with Pspice. Hence, the proposed method is
valid, and it can be used for the studied technology. Observing Fig. 9 and Fig. 10
respectively, the peak-to-peak coupling noise in the first case is about 0.15 V. For the
second case, the peak-to-peak coupling noise is about 0.045 V. These results show
that the noise coupling of both paths, the coupling noise between two TSVs and
TSV-Active circuit coupling noise, are significant, must be considered, and reduced.
Observing Fig. 11, the peak-to-peak coupling noise is about 0.15 V. The shapes in
addition to the peak-to-peak coupling noise are similar to those presented in case 1.
The difference is in the disturbances added by the interconnect line. The results show

Table 1 Lumped circuit

Component Value
elements of coupling noise
between two TSVs Ctsv-equi = C1 201.3 fF
Rsub-equi = R2 928.5 
Csub-equiv = C3 11.2 fF
R3 50 
C2 10 fF
C4 10 fF
l 1 mm

Table 2 Lumped circuit

Component Value
elements of coupling noise
between and an active Ctsv-equiv 817.5 fF
Rsub-equiv 879.5 
Csub-equiv 12 fF
Computing and Analyzing Through Silicon Via (TSV) Noise Coupling … 1651

Fig. 9 TSV-TSV coupling 0.1

noise FDTD
Pspice
0.05

Noise Coupling (V)

0

-0.05

-0.1
0 0.2 0.4 0.6 0.8 1
-9
Time (s) x 10

Fig. 10 TSV-Active circuit 0.03

coupling noise FDTD
Pspice
0.02
Noise coupling (V)

0.01

-0.01

-0.02
0 0.2 0.4 0.6 0.8 1
-9
Time (s) x 10

Fig. 11 TSV-TSV noise 0.15

coupling with interconnect FDTD
line 0.1 Pspice
Noise coupling (V)

0.05

-0.05

-0.1

-0.15

-0.2
0 0.2 0.4 0.6 0.8 1
-9
Time (s) x 10
1652 K. Ait Belaid et al.

that the interconnect line influences the noise coupling by adding some disturbances
as depicted in Fig. 11. The interconnect line can also add reflections that can threaten
the normal functioning of the circuits. The method proposed in this work shows good
results similar to the one proposed and detailed in [20]. This is why it can be adopted
for the modeling of TSV couplings in the time domain.

4 Conclusions

A lot of papers in the literature treat TSV noise in the frequency domain, while,
noise coupling modeling is more important in the time domain. Modeling in the time
domain is more significant and helps to understand the effects of the noise. Hence, for
a good design and construction of 3-D structures based on TSVs, modeling methods
in the time domain must be proposed. In this paper, a technique to model the time-
domain coupling noise in 3D-IC structures is explained in detail. The proposed
method, FDTD-1D, was verified and validated using Matlab and Pspice tools. Then,
a time domain analysis was done for all studied circuits. Based on the found results,
it could be seen that the noise in all studied cases exceeds 40 mV, which means
that TSV-TSV coupling noise and TSV-Active circuit, are significant and must be
reduced, and that the interconnect lines influence the shape of waveforms simulated.

References

1. Han K (2009) Electromagnetic modeling of interconnection in three-dimensional integration.

School of Electrical and Computer Engineering Georgia Institute of Technology
2. Erasekera R, Grange M, Pamunuwa D, Tenhunen H, Zheng L (2009) Compact modeling of
through-silicon vias (TSVs) in three dimensional (3-D) integrated circuit. In: Proceedings of
IEEE international conference on 3D system integration (3D IC), San Francico, USA
3. Kim D, Mukhopadhyay S, Lim S (2009) TSV aware interconnect length and power prediction
for 3D stacked ICs. In: IEEE international interconnect technology conference, pp 26–28
4. Salah K (2015) A TSV to TSV, A TSV to metal interconnects, and A TSV to active device
coupling capacitance: analysis and recommendation. In: 10th international conference on
design & technology systems in nanoscale era, pp 1–2
5. Cadix L, Fuchs C, Rousseau M, Leduc P, Chaabouni H, Thuaire A, Sillon N (2010) Integration
and frequency dependent parametric modeling of through silicon via involved in high density
3D chip stacking. ECS Trans 33(12):1–21
6. Bandyopadhyay T, Han KJ, Chung D, Chatterjee R, Swaminathan M, Tummala R (2011)
Rigorous electrical modeling of through silicon vias (TSVs) with MOS capacitance effects.
IEEE Trans Compon Packag Manuf Technol 1(6):893–903
7. Salah K, Ragai H, Ismail Y, El Rouby A (2011) Equivalent lumped element models for
various n-port through silicon vias networks. In: 16th Asia and South Pacific design automation
conference. IEEE, pp 176–181
8. Han KJ, Swaminathan M, Bandyopadhyay T (2010) Electromagnetic modeling of through-
silicon via (TSV) interconnections using cylindrical model basis functions. IEEE Trans Adv
Packag 33(4):804–817
Computing and Analyzing Through Silicon Via (TSV) Noise Coupling … 1653

9. Ryu C, Chung D, Lee J, Lee K, Oh T, Kim J (2005) High frequency electrical circuit model
of chip-to-chip vertical via interconnection for 3-D chip stacking package. In: 14th topical
meeting on electrical performance of electronic packaging. IEEE, pp 151–154
10. Qu C, Ding R, Zhu Z (2017) High-frequency electrical modeling and characterization
of differential TSVs for 3-D integration applications. IEEE Microw Wirel Compon Lett
27(8):721–723
11. Afzali-Kusha A, Nagata M, Verghese NK, Allstot DJ (2006) Substrate noise coupling in SoC
design: modeling, avoidance, and validation. Proc IEEE 94(12):2109–2138
12. Kerns KJ, Wemple IL, Yang AT (1995) Stable and efficient reduction of substrate model
networks using congruence transforms. In: Proceedings of international conference on
computer aided design. IEEE, pp 207–214
13. Verghese N, Allstot DJ, Masui S (1993) Rapid simulation of substrate coupling effects in
mixed-mode ICs. In: Proceedings of custom integrated circuits conference. IEEE, p 18.3.1–
18.3.4
14. Sedes T (1993) Substrate resistance extraction for physics-based layout verification. In:
Proceeding of workshop circuits systems signal. IEEE/PRORISC, pp 101–106
15. Costa JP, Chou M, Silveria LM (1999) Efficient techniques for accurate modeling and simula-
tion of substrate coupling in mixed-signal IC’s. IEEE Trans Comput-Aided Des Integr Circuits
Syst 18(5):597–607
16. Lee M, Pa JS, Kim J (2014) Electrical design of through silicon via. Springer, Heidelberg
17. Cho J (2011) Modeling and analysis of TSV noise coupling and suppressing using a guard
ring. IEEE Trans Compon Packag Manuf Technol 1:220–233
18. Lim J (2018) Modeling and analysis of TSV noise coupling effects on RFLC-VCO and shielding
structures in 3D-IC. IEEE Trans Electromagn Compat 60:1–9
19. Watanabe M, Karel R, Niioka N, Kobayashi T, Fukase MA, Imai M, Kurokawa A (2014)
Effect of substrate contacts on reducing crosstalk noise between TSVs. In: IEEE Asia Pacific
conference on circuits and systems. IEEE, pp 763–766
20. Ait Belaid K, Belahrach H, Ayad H (2019) Numerical Laplace inversion method for through-
silicon via (TSV) noise coupling in 3D-IC design. Electronics 8(9):1010
21. Xu C, Kourkoulos V, Suaya R, Banerjee K (2011) A fully analytical model for the series
impedance of TSV with consideration of substrate effects and coupling with horizontal
interconnects. IEEE Trans Electron Devices 58:3529–3540
22. Attarzadeh H, Lim SK, Ytterdal T (2016) Design and analysis of a stochastic flash analog-
to-digital converter in 3D-IC technology for integration with ultrasound transducer array.
Microelectron J 48:39–49
23. Beanato G, Gharibdoust K, Cevrero A, De Micheli G, Leblebici Y (2016) Design and analysis
of jitter-aware low-power and high-speed TSV link for 3D-ICs. Microelectron J 48:50–59
Design of a CMOS Bandgap Reference
Voltage Using the OP AMP in 180 nm
Process
Ahmed Rahali , Karim El Khadiri , Zakia Lakhliai, Hassan Qjidaa ,
and Ahmed Tahiri

Abstract This article proposes the implementation and design of a first-order CMOS
Bandgap reference using an operational amplifier with negative feedback to improve
the power supply rejection ratio (PSRR) and reduce the temperature coefficient (TC).
The circuit is designed in 180 nm CMOS process technology and provides a reference
output voltage of 1.2 V over an extended temperature range from −40 °C to 120 °C
with a measured temperature coefficient of 54 ppm/°C. The BGR chip uses a 1.8 V
supply.

Keywords Bandgap reference (BGR) · Complementary to absolute temperature

(CTAT) · Proportional to absolute temperature (PTAT) · Temperature coefficient
(TC) · PSRR

1 Introduction

An essential part in the design of most analog and digital electronic systems is
to produce reliable and accurate voltage and current references. A reference in a
circuit provides a stable point which is employed by other subcircuits to achieve
predictable and repeatable results. It is important that this reference point should
be unchanged significantly under different operating conditions. Analog-to-digital
converter, AC-DC or DC-DC converter, and linear regulators are some examples of
circuit applications where voltage reference are intrinsically required.
The Bandgap is an advantageous solution for generating a stable and precise
voltage reference regardless of temperature variations, power supply, and technology
fluctuations. this circuit operates by combining two voltages having a temperature
dependence of opposite signs Fig. 1. The first is the base-emitter voltage VB E of a

A. Rahali (B) · K. El Khadiri · Z. Lakhliai · A. Tahiri

Laboratory of Computer Science and Interdisciplinary Physics, ENS, Sidi Mohamed Ben
Abdellah University, Fez, Morocco
e-mail: [email protected]
H. Qjidaa
Faculty of Sciences Dhar el Mahraz, Sidi Mohamed Ben Abdellah University, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1655
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_150
1656 A. Rahali et al.

Fig. 1 The principle of BGR

bipolar junction transistor (BJT), this voltage is a decreasing function in tempera-

ture; it is qualified as CTAT (Complementary to Absolute Temperature). The second
voltage is an increasing function, called PTAT (Proportional to Absolut Tempera-
ture), and extracted from the difference between base-emitter voltages of two bipolar
transistors traversed by different current densities [1, 2]. The reference voltage is then
the sum of the PTAT and CTAT voltage with a factor K.
This article presents a CMOS BGR circuit using an operational amplifier with
negative feedback. the document is organized as follows: the principle of bandgap
reference generated by a conventional architecture is presented in Sect. 2. Section 3
illustrates the proposed methodology and structure using OP-AMP. the results of the
simulations are shown and discussed in Sect. 4. Finally, Sect. 5 is devoted to the
conclusion of the overall document.

2 Conventional Bandgap Voltage Reference Structure

Due to its temperature-dependence characteristic, a diode-connected bipolar tran-

sistor (BJT) is used as a basic device in most voltage reference architectures. Fig. 2
[3] shows a conventional first-order bandgap reference using a current mirror. When
the BJTs are biased in the forward active region, the base-emitter voltage VB E can
be developed as [4]

   V −V (T )+(η−χ )VTr   (η−χ )V    

V B E = VG0 + (η − χ )VTr − G0 B E r T− Tr
T ln TT − T + Tr
Tr Tr r
(1)

Where VG0 is the bandgap voltage, η is a process-dependent constant, χ refers

to the temperature dependence of the current forced through the collector, VTr is the
Design of a CMOS Bandgap Reference Voltage Using the OP AMP ... 1657

Fig. 2 Conventional first-order Bandgap reference

 
thermal voltage at the reference temperature Tr , VG0 − VB E (Tr ) + (η − χ )VTr TTr
is a linear first- order term in temperature. To have a voltage reference, it is necessary
to cancel or reduce the temperature dependence of this term by using VT with a
judiciously chosen multiplicative coefficient.
The voltage Vr e f is written as:

Vr e f = VB E + K VT (2)

3 Study of the Proposed Voltage Reference

The voltage reference circuit designed is composed of two bipolar transistors

connected in diodes as illustrated in Fig. 3. The transistors MP2, MP3, and MP4
form the startup circuit, which plays the role of avoiding the undesirable zero current
state when the power is applied. This block supplies or absorbs a low intensity current
to ensure the normal operating state with non-zero current. Resistors R1 and R2 are
chosen equal to have I1 = I2 .
The operational amplifier in the BGR circuit is a typical two-stage operational
amplifier as illustrated in Fig. 4, and imposes voltages V A and VB to be equal.
Therefore, the current through resistor R0 can be expressed by:

VB E0 − VB E1 VT
I1 = = ln N (3)
RO RO
1658 A. Rahali et al.

Finally, the voltage reference Vr e f is written:


R1
Vr e f = VB E1 + 1 + VT ln N (4)
R0

Fig. 3 Implementation of proposed Bandgap circuit

Fig. 4 Implementation Of two-stages operational amplifier

Design of a CMOS Bandgap Reference Voltage Using the OP AMP ... 1659

4 Simulation Results

4.1 Variation of Reference Voltage with Respect

to Temperature

Figure 5 illustrates the variation of output reference voltage as a function of tempera-

ture over the range from −40 °C to 120 °C. This variation shows a maximum voltage
of 1.2073338 V at −40 °C, and a minimum of 1.2015818 V at 65 °C, which gives a
temperature coefficient of 54 ppm/°C.
Figure 6 shows the simulation results of the reference voltage given at different
process corners FF, FS, SF, and SS which respectively represent fast-fast, fast-slow,
slow-fast, and slow-slow over the same temperature range.

Fig. 5 Reference voltage Vref as function of temperature

Fig. 6 Reference voltage Vref at different process corners

1660 A. Rahali et al.

4.2 Variation of Reference Voltage Depending on Power

Supply

The reference voltage as function on power supply variation is represented in Fig. 7.

It represents an almost constant reference with a maximum variation of 9.8 mV
compared to the nominal voltage at 1.8 V.
The PSRR simulated in Fig. 8 indicated a value of −35.6 dB for frequencies below
100 kHz.
Figure 9 shows the parametric simulation of Vref versus VDD for different
temperature values from −40 °C to 120 °C.
Figure 10 illustrated the transient response of bandgap and indicated that the time
taken to achieves the value 1.14 V (95% of Vr e f ) is 2.6 us, which proves the BGR
responds quickly.

Fig. 7 Reference voltage Vref versus power supply VDD

Fig. 8 PSRR of Bandgap

Design of a CMOS Bandgap Reference Voltage Using the OP AMP ... 1661

Fig. 9 Parametric simulation of Vref versus VDD from −40 °C to 120 °C

Fig. 10 Transient response of Bandgap

Table 1 Comparison with other published results

Parameter This work [3] [5] [6]
Supply voltage (V) 1.8 1.8 3.3 1.2
Reference voltage (V) 1.2 1.12 1.124 0.82
Temperature range (°C) −40–120 −40–125 −25–125 −40–100
TC (ppm/°C) 54 152 65 250
PSRR (dB) −35.6 – −49.2@ 1 kHz −47@ 40 kHz
CMOS Process (μm) 0.18 0.18 0.35 0.18

5 Conclusion

A first-order CMOS bandgap reference circuit using an operational amplifier is

presented in this article, to generate a stable output voltage reference of 1.2 V with
a calculated temperature coefficient of 54 ppm/°C over a temperature range varying
of −40 °C to 120 °C. The circuit exhibits a PSRR of −35 dB at 100 kHz and uses a
1.8 V supply.
1662 A. Rahali et al.

References

1. Jacob Baker R (2010) CMOS Circuit Design, Layout, and Simulation. Third Edition. Wiley-
IEEE Press
2. Razavi B (2001) Design of Analog CMOS Integrated Circuits. McGraw-Hill Higher Education,
international Edition
3. Akshaya R, Yellampali Siva S (2017) Design of an improved Bandgap Reference in 180
nm CMOS process technology. In 2nd IEEE International Conference on Recent Trends in
Electronics Information & Communication Technology (RTEICT), pp 521–524. India
4. Rincon-Mora GA (2002) voltage reference from diodes to precision high-order Bandgap circuits.
Wiley-IEEE Press
5. Zhao Y, Yue S, Bian Q (2008) A novel low temperature coefficient Band-gap reference without
resistors. In International Multiconference of Engineers and Computer Scientists 2008, IMECS,
vol II, Hong Kong
6. Shrivastava A, Kaur A, Sarkar M (2017) A 1.2 V, 33 ppm/°C, 40 nW, regeneration based BGR
circuit for nanowatt CMOS. In International SoC Design Conference (ISOCC), pp 111–112,
South Korea
Mechatronic, Robotic and Control System
Analytical Study of the CMOS Active
Inductor
Imane Halkhams, Wafae El Hamdani, Said Mazer, Moulhime El Bekkali,
and Mohammed Fattah

Abstract Active inductors, widely used in filtering and amplification functions,

and based on frequency filtering selection, are the subject of several investigations.
Through this paper, we propose an analytical study of the active inductor to under-
stand its frequency functioning. To do this, a small transistor signal study is carried
out first in order to extract the intrinsic elements that will be used to study the active
inductor, by calculating H and Y parameters.

Keywords Active inductor · CMOS inductor · Small signal analysis · H

parameters · Y parameters

1 Introduction

Frequency filtering is a key element in electronic circuits, which explains the number
of studies carried out on filtering operations [1, 2]. In the literature, different topolo-
gies of active inductors have been presented. Widely used in modern telecommu-
nication systems such as the mobile phone network, wireless and satellites. Their
manufacture has undergone a vertiginous evolution. Indeed, integrable filters are
produced in tens of millions each year and are present in all daily telecommunication
equipment (smartphones, TVs, ADSL, etc.).
Active inductor-based filters reported in the literature [3, 4], have many advantages
such as frequency tuning ease, high-quality factors, and on-chip integration, however,
the transistor’s parasitic effects impact the linearity and stability of systems which

I. Halkhams (B)
SED, Research, Systems and Sustainable Environments Laboratory, The Private University
of Fez, Fez, Morocco
e-mail: [email protected]
W. El Hamdani · S. Mazer · M. El Bekkali
IASSE Laboratory, Sidi Mohamed Ben Abdellah University, Fez, Morocco
M. Fattah
IMAGE, Moulay Ismail University, Meknes, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1665
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_151
1666 I. Halkhams et al.

calls for accurate noise and linearization studies. In this context, we propose a study
of the intrinsic parameters of the CMOS transistor to optimize them and to provide
adequate power to the active inductor circuit. We also propose an analytical study of
the latter to highlight its role in the frequency filter chain.
We propose a study of the active inductor described in [5], to facilitate the design
of bandpass filters and LNAs. This article starts with the extraction of the intrinsic
parameters of the transistor via the h and y matrices. In the third section, we present
the analytical study of the active inductor, using the parameters calculated in Sect. 1.

2 Small Signal Study and Extraction of the Intrinsic

Parameters of the Transistor

The need for a high inductance value at the resonance frequency requires that the
transconductances used in the active inductor design be properly configured. To do
this, a small signal study from the transistor is fundamental. The term small signal
is dedicated to signals small enough that the characteristics of the components vary
slightly and can remain in a linear approximation. We used a CMOS transistor whose
five-element equivalent model is given in Fig. 1.
In this model, the components electrically characterize the transistor’s frequency
behaviour. The channel is described by its gds conductance. The gm transconduc-
tance of the transistor determines the gains. The gate-channel interaction is mainly
capacitive because of the existence of the MOS conjunction.

2.1 H Parameters

H parameters are hybrid parameters that connect the input voltage VGS to the input
current IG and the output voltage VDS to the output current ID. The equations
according to the H matrix are written:

Fig. 1 Five elements transistor small signal model

Analytical Study of the CMOS Active Inductor 1667

V gs = H 11.I G + H 12.V ds
(1)
I D = H 21.I G + H 22.V ds

Parameter H11 represents an impedance in Ω, parameters H12 and H21 are

without unit, where H21 represents the current gain and parameter H22 is an
admittance in Siemens (S).
By short-circuiting the output of the transistor and injecting a voltage at its input,
IG can be measured:

H 11 = VI Ggs
(2)
H 11 = j.ω.(Cgs+Cgd)
1

Parameter H12 represents the coupling of the output to the input. To calculate it,
this time we inject a voltage at the output and we measure the voltage at the input of
the circuit.
V gs
H12 = =0 (3)
V ds
Parameter H21 is the current gain of the transistor. To calculate it we inject a
voltage at the input while short-circuiting the output voltage, then we calculate the
currents IG and ID.

H 21 = II GD
(4)
gm
H 21 = j.ω.(Cgs+Cgd)

At the transition frequency, the current gain of the transistor H21 is equal to the
unit.

|H21| = 1 (5)

2.2 Y Parameters

To know the functionality of a quadruple, one method consists of studying its voltage-
current transformation Y (matrix admittance), defined by the Eq. (6):
⎧
⎪ Y 11 = |V 2 =0
I1
⎪
⎨ V1
Y 12 = I1
V2
|V 1 =0
(6)
⎪
⎪ Y 21 = I2
|V 2 =0
⎩ V1
Y 22 = I2
V2
|V 1 =0
1668 I. Halkhams et al.

To do this, we proceed by making each time short circuits for the admittance
matrix. However, the realization of short circuits becomes almost impossible when we
exceed 100 MHz, because of the parasitic capacitors and inductors, hence the interest
to use the S matrix, which is very effective for high-frequency implementations and
which is measurable by loading the input and the output with a load of 50 Ω.
By studying the small-signal circuit of the transistor, the latter’s equations using
the Y representations are given by:

I G = Y 11.V gs + Y 12.V ds
(7)
I D = Y 21.V gs + Y 22.V ds

When the voltage VDS is zero, Y11 is expressed:

IG
Y 11 = (8)
V gs

Y 11 = j.ω.(Cgs + Cgd) (9)

When the VGS voltage is zero, Y12 is expressed:

IG
Y 12 = (10)
V ds

Y 12 = − j.ω.Cgd (11)

When the Vds voltage is zero, Y21 is expressed:

ID
Y 21 = (12)
V ds

Y 21 = gm − j.ω.Cgd (13)

When the VGS voltage is zero, Y22 is expressed:

ID
Y 22 = (14)
V ds

Y 22 = gds + j.ω.(Cds + Cgd) (15)

Finally, from (9, 11, 13 et 15) we can extract the intrinsic parameters of the
transistor:
Analytical Study of the CMOS Active Inductor 1669

(16)

Transconductance gm represents the slope of the ID characteristics as a function

of Vgs for a certain channel length.

∂Id(Vgs, Vds)
gm = (17)
∂Vgs

The output conductance gds represents the resistance of the channel. The greater
the width of the area under the gate, the smaller the conductance. The space load
zone is an area in the vicinity of the metal-semiconductor interface. It is less rich in
electrons than in the volume of the semiconductor.
∂ I d(V gs, V ds)
gds = (18)
∂ V ds
The Cgd capacitor is related to the variation of the load stored under the gate
according to the gate-drain voltage when the gate-source voltage is fixed:

∂ Qg(Vgs, Vgd)
Cgd = (19)
∂Vgd

The Cgs capacitor is linked to the variation of the load stored under the gate
according to the source-gate voltage when the gate-drain voltage is fixed.

∂ Qg(V gs, V gd)

Cgs = (20)
∂ V gs

The Cds capacitor takes into account the electrostatic coupling between the heavily
doped regions below the source and drain contacts.
The transition frequency is a characteristic of the transistor’s operating speed.
Because the higher the frequency, the more the transistor can achieve a high transcon-
ductance for a low capacity. The transition frequency can be calculated from the
current gain of the transistor.
gm
fT ≈ (21)
(Cgs + Cgd)

The calculated parameters will be used in the study of the active inductor in the
following paragraph.
1670 I. Halkhams et al.

3 Analytical Study of the Active Inductor

3.1 Gyrator Circuit

The gyrator circuit was invented by Bernard D. H. Tellegen (1900–1990) [6]. Its
symbol is given in Fig. 2. It was first designed for low-frequency applications
where transconductance operational amplifiers were used to simulate active inductors
connected to the mass or floating using external integrated capacitors.
The gyrator considered a 2-port component, when closed to an impedance (Fig. 2),
is described by the following impedance matrix:
   
V1 I1 Z11 Z 12 I1
= [Z] = (22)
V2 I2 Z 21 Z 22 I 2

Thus, it transforms the impedance according to the following relation:

  
V1 0 −Rg I1
= (23)
V2 Rg 0 I2

Where Rg is the gyration resistance. Rewriting this equation yields:


V1 = − Rg.I2
(24)
V2 = Rg.I1

Thus:

V1.I1 + V2.I2 = 0 (25)

This equation states that there is no energy generated, stored, or dissipated since
the input power is equal to the output power.

Fig. 2 Symbol of a gyrator closed on an impedance

Analytical Study of the CMOS Active Inductor 1671

When the impedance is capacitive, the input impedance is as follows:

 2
Z e = Rg
Zc (26)
Z e = Rg2 .C p

This equation means that when the ideal gyrator is closed on a capacitor, it
simulates the effect of an inductive impedance of value:

L = Rg2 .C (27)

Equation (27) illustrates the possibility of using ideal gyrators in combination

with capacitors to simulate inductors, hence the designation of the active inductor.

3.2 Active Inductor Circuit

The active inductor consists of a common source transistor, in feedback with another
transistor mounted in common gate. The two transistors form a gyrator that trans-
forms the internal Cgs1 capacitor (between gate and source) of the first transistor
into an inductor.
The small equivalent signal model of the inductor used in [7] is given in Fig. 3.
The Yin input admittance calculated from Fig. 3(a) is given in (28):

gm1.gm2
+ gm2.gds1 + gds1.gds2 + R −1 (gm1
 + gds1 + gds2)
+p. Cgs2(gm1 + gds1 + gds2) + Cgs1 gds1 + R −1
+p2(Cgs1.Cgs2)
Yin = (28)
gds1 + R −1 + p.Cgs2

Using the small-signal model of the active inductor and taking into account the
parasitic capacities, input admittance calculated from Fig. 3(b) is given in (29).

gm1.gm2 + gm2.gds1 + gds1.gds2 + R −1 (gm1 + gds1 + gds2)

+p.[C2(gm1 +
 gds1 + gds2) + C1(gds1 + R −1 )
+Cgd2 gds2 − R −1 + gm2 + gds1
+Cgs2(gm1 + gds1 + gds2)]
+p2[C1.C2 + C1.Cgd2 + C1.Cgs2 + Cgd2.C2 + Cgd2.Cgs2]
Yin = (29)
gds1 + R −1 + p.Cgs2

With C1 = Cgs1 + Cgb1; C2 = Cds1 + Cbd1; C3 = Cgs2 + Cgb2; C4 = Cds2

+ Cbd2.
The two admittances (28) and (29) can be simplified and give the expression (30):
1672 I. Halkhams et al.

(a)

(b)

Fig. 3 Active inductor small signal equivalent model (a) simplified (b) with capacitors

1
Yin = gm1 + j.w.Cgs1 + gds1+R −1 Cgs2
(30)
gm1.gm2
+ j.w. gm1.gm2

The equation is a second-order transfer function whose terms constitute a parallel

RLC resonator (31), containing an inductor with a serial resistor, in parallel with a
capacitor and a resistor.

1 1
Yin = + Cp + (31)
R Rs + L
Analytical Study of the CMOS Active Inductor 1673

By investigating Eq. (30) and by identifying between (30) and (31), four terms
can be deduced (32) [6]:
⎧ Cgs2
⎪
⎪ L = gm1.gm2
⎪
⎨ −1
Rs = gds1+R
gm1.gm2 (32)
⎪
⎪ Rp = gm1
1
⎪
⎩
C p = Cgs1

This proves that the active inductor is equivalent to a resonator, which can be used
for band-pass filtering applications as well as for oscillator design.
The resonator circuit may be considered as an inductor if the operating frequency
is less than the resonance frequency. The two parallel and serial resistors are the
sources of resonator circuit losses. To remedy this, a resistor loss compensation
circuit shall be added to the active inductor circuit.

4 Conclusion

A CMOS active inductor analytical study was presented in this paper to understand
its functioning. H and Y parameters were used to extract the transistor’s intrinsic
elements.
Active transistor-based inductances only and using internal capacity have been
a giant step towards systems integration and in the implementation of preselection
filters, bandpass filters, LNA, and oscillators. All these implementations exploit the
inductive effect and the selectivity around the resonance frequency.
This analytical study demonstrates the functioning of active inductors using
CMOS transistors. The simplicity of their configuration makes them indispensable
in the implementation of bandpass filters that exploit resonance to select a frequency
band [8]. However, the selection of a restricted frequency band or even a single
frequency requires a very narrow bandwidth, characterized by a very high-quality
factor, which leads to selectivity and loss study to make bandwidth as narrow as
possible. This study may be used in the design of bandpass filters and LNAs [9–11].

References

1. Erokhin VV et al (2019) Active tunable bandpass filter with voltage gain control in SiGe
BiCMOS 130 nm. In International Siberian Conference on Control and Communications
(SIBCON)
2. Pantoli L, et al (2017) Design considerations and effects of class-AB polarization in active filters
realized by means of active inductors. In: 47th European Microwave Conference (EuMC), pp
37–40
1674 I. Halkhams et al.

3. Wu Y, Ismail M (2003) RF bandpass filter design based on CMOS active inductors. IEEE Trans
Circuits Syst 50(12):942–949
4. Leoni A, et al (2017) Bandpass filter design with active inductor by means of wave digital
approach. In 13th International Conference on Advanced Technologies, Systems and Services
in Telecommunications (TELSIKS), pp 339–342
5. Halkhams I et al (2017) A Selective active filter for the 5G in the mm-wave band in pHEMT
technology. Contemporary Eng Sci 10(4):193–202
6. Tellegen BDH (2014) The gyrator, a new electric network element. Philips Res. Rep. 3: 81–101.
April 1948, archived from the original on 2014–04–23. Retrieved 20 Mar 2010
7. Halkhams I, et al (2017) Design of a dual-band bandpass filter using active inductor principle.
Int J Commun Antenna Propagation (IRECAP) 7(4) ISSN 2039–5086
8. Ren S, Benedik C (2013) RF CMOS active inductor band pass filter with post fabrication
calibration. AEU - Int J Electr Commun 67(12):1058–1067
9. Nawaz AA, et al (2019) A Ka/V band-switchable LNA with 2.8/3.4 dB noise figure. IEEE
Microw Wirel Compon Lett 29(10)
10. Lee W, Hong S (2020) 28 GHz RF front-end structure using CG LNA as a Switch. IEEE
Microw Wirel Compon Lett 30(1)
11. Mudavath M, et al (2020) Design and analysis of CMOS RF receiver front-end of LNA for
wireless applications. Microprocess Microsyst 75
SM Versus PI Tunning for PWM Rectifier
for SEIG in Wind Energy Application

Ouafia Fadi , Soufiane Gaizen, and Ahmed Abbou

Abstract This work presents the sliding mode control applied on the self-excited
induction generator, typically used in electrical energy production and particularly
in wind turbines. The machine is associated with an active load through a bank
of capacitors and operated as a generator. Its stator is associated with the AC/ DC
converter. The SM control is presented by a cascade structure that includes tow loops:
The internal loop controls the machine currents “id and iq” and the external loop
controls the Vdc voltage. The whole model (SEIG/PWM rectifier/load) is simulated
using the Matlab/Simulink software, then the results of this simulation validated
the SM superiority over the typical PI regulator. The proposed SM method hold
various advantages: it is recommended to control the systems that contain multy-
variables, for both linear and non-linear systems, it provides great transient and
steady performances, it’s robust against external disturbances, robust against uncer-
tainties/parameters variations, and offers lower THD harmonics versus the typical
PI.

Keywords SEIG · PWM rectifier · SM strategy · PI controller · THD harmonics

1 Introduction

The wind generation system has gained recently a great importance as a sustainable
and secure source of renewable energy. A three-phase asynchronous machine could
be operated as an asynchronous generator, either connected to the electrical network
or operated in isolated power generation mode with a three-phase capacitance coupled
to the stator windings [1]. This type of generator operated in excitation autonomous
mode, called self-excited induction generator is the most useful generator for remote
windy aeras because it doesn’t require an external power supplies to produce the
excitation magnetic field, the excitation is delivered by supplemental stator terminal
excitation capacitor bank. These references [2–4] have presented several algorithms

O. Fadi (B) · S. Gaizen · A. Abbou

Mohammadia School of Engineers, Mohammed 5 University, P.B765 Rabat, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1675
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_152
1676 O. Fadi et al.

and approaches to determine the minimum value of the excitation capacitor. SEIG
has a several advantages such as simple and robust construction, minimal mainte-
nance, absence of DC power supply for excitation and inherent overload protection
[1]. However, the principal drawback of such generators is the voltage control that
can be highly complex depending on the desired performance, this complexity is
more precisely located at the power converter that is unable to produce a cleaned
and continued DC voltage waveform, this issue is mainly caused by the analytical
nonlinear model of the machine that hold numerous variables and that is very respon-
sive to external disturbances (wind speed and load changes) [1]. The control of these
power converters to maintain the output dc voltage regulated to a desired value and
to generate the stator currents with a lowest possible harmonic distortion has been
widely studied by engineers and researchers. The Conventional techniques of PI
control, cover a wide range in industrial applications. These are linear control tech-
niques and have the advantage of simplicity of implementation and ease of synthesis.
Over time, the PI control applications will be inefficient, especially if the processes
to be controlled have complex and non-linear structures [4]. Likewise, the PI tunning
is used in a cascaded configuration [5]. In such a configuration, the internal loop is
often accomplished by current controller, as well as the controller in the external
loop regulates dc-link output voltage. There are typically two classical families of
these control configurations [6, 7]: the VOC (Voltage oriented control) similarly as
in VFOC (Virtual Flux oriented control) of induction generator, they are based on
the coordinate transformation between stationary α-β and synchronous rotating d-q
reference system. Both strategies guarantee rapid transient response and high static
results via the internal current control loop. however, these configurations exhibit
some disadvantages such as [6]: A coupled control between the active and reac-
tive components of the current; A coordinate transformation and PI controllers are
required. Another proposed method based on instantaneous direct active and reac-
tive power control loops, called Direct Power Control (DPC) is developed in [8].
This method constitutes a viable alternative to the typical vector control methods
(VOC-FVOC). However, the DPC has several disadvantages [9]: Variable switching
frequency, High values of the line inductance and sampling frequency are needed to
obtain an acceptable current waveform, also power and voltage estimation should be
avoided at the moment of switching (it yields high errors).
All these techniques mentioned above are employed when the controlled part
of the machine is minimally disturbed, they may be sufficient if the requirements
on the accuracy and performance of the system are not too stringent. However, in
the opposite case and particularly when the controlled part is subjected to strong
nonlinearities and to temporal variables, it is necessary to design control algorithms
ensuring the robustness of the generator with respect to the uncertainties on the
external parameters and their variations [10].
The sliding mode tunning is a robust control method [10] that has indisputable
advantages for the poorly identified system or with variable parameters. This paper
is divided into 4 important sections: in Sect. 1, the modeling of the asynchronous
generator, powered by a tow-level inverter/rectifier system is presented. In Sect. 2, the
basic principle of sliding mode is treated and detailed by mathematical equations. In
SM Versus PI Tunning for PWM Rectifier for SEIG ... 1677

Sect. 3, the internal and external regulation loops used the conventional PI controllers
are replaced by the regulators based on sliding mode, the simulation results will be
exposed to observe the validity of the PI compared with SM regulators. finally, in
Sect. 4, a recap of the main points of this study are discussed.

2 System Units Modeling

2.1 D-Q Axes SEIG Modeling

The system constitution based on a squirrel cage induction generator, prime mover,
excitation capacitance and a PWM rectifier connected to an active load [8].

disq 1  
= −Ls Rs isq − wr Lm
2
isd + Lm Rr iqr − wr Lm Lr idr + Lr Vsq (1)
dt Ls Lr − Lm
2

disd 1  
= −Lr Rs isd − wr Lm
2
isq + Lm Rr idr + wr Lm Lr iqr + Lr Vsd (2)
dt Ls Lr − Lm
2

dirq 1  
= Lm Rs isq + wr L m L s isd − Ls Rs iqr + wr Ls Lr idr + Lm Vsq (3)
dt Ls Lr − Lm
2

dird 1  
= Lm Rs isd − wr L m L s isq − Ls Rr idr − wr Ls Lr iqr + Lm Vsd (4)
dt Ls Lr − Lm
2

The excitation system of the SEIG presented by [8]:

dVsd 1
= (isd − id ) (5)
dt C
dVsq 1 
= isq − iq (6)
dt C

2.2 D-Q PWM Rectifier Modeling

The system dynamic of the three-phase two-level rectifier can be expressed using
the park transformation, in dq synchronous reference frame as [8]:

diq
L = Vqs − Riq − Lwid − Vrq (7)
dt
1678 O. Fadi et al.

did
L = Vds − Rid + Lwiq − Vrd (8)
dt
dVbus   Vbus
Cbus = ibus − il = Fd id + Fq iq − (9)
dt RL

3 Sliding Mode Tunning Theory

The SM system, is a non-linear control technique, it is characterized by the disconti-

nuity of the control at the passages by a switching surface called the sliding surface.
The basic idea of this control consists of bring the system to a stable switching hyper-
plane (sliding surface) and, to converge the system on this sliding area towards the
desired equilibrium point [10].

3.1 Choice of Sliding Surface

Consider the following state model [10]:

 
Ẋ = [A][X] + [B][U] (10)

Where [X] Rn is the state vector, and [U] Rm is the control vector, with n > m.
Generally, the choice of the sliding surfaces number is equal to the dimension of
the control vector (U) [11]. In order to ensure the convergence of a state variable X
towards its reference value X *, several works propose the following general form:
r−1
d
S(x) = ( + γ) ∗ e(x) (11)
dt
γ : Positive gain; e(X) = X - X*: The error of the variable to be regulated; r:
Relative degree, it is the smallest positive integer representing the number of times
of derivation to obtain the control (d S)/dt=0 ensuring the controllability; S(x): is
an autonomous linear differential equation whose response e (x) tends to zero for a
correct choice of gain γ and this is the main objective of this control.

3.2 Convergence Condition of the SM

The condition of the convergence is the criteria which allow the dynamic of the system
converges towards the sliding surface and remains constant against the disturbances
SM Versus PI Tunning for PWM Rectifier for SEIG ... 1679

[10]. For this reason, A candidate Lyapunov function V (x) > 0 is chosen for the system
state variables and a control law is determined to make this function decreases V(x)
< 0. The appropriate Lyapunov function for the system is expressed as follow [10]:

1
V(x) = S(x)S(x)2 (12)
2
By deriving this one, we obtain:

˙ = S(x)S(x)˙ < 0
V(x) (13)

The Lyapunov candidate function decreases if (13) is negative.

3.3 Control Law Determination

The law control is necessary to attract the trajectory of the desired state towards the
surface and then towards its equilibrium point while ensuring the existence conditions
of the SM [10]. The law control structure has two parts, the first concerns the exact
linearization Ueq and the second Un is a stabilizer part, this one is used to eliminate
the model uncertainties and to dismiss the external disturbances [10].

U(t) = Ueq + Un (14)

The two parts expressions of the control U is determined by the set of equations:

dS ∂S ∂x
Ṡ = = . (15)
dt ∂x ∂t
By replacing (10) and (13) in (14), we find:

∂S   ∂S
Ṡ = [A][X] + [B]Ueq + [B]Un (16)
∂x ∂x

The equivalent control is deduced by considering that: d S(t)

dt
= 0:
 −1 
∂S ∂S
Ueq = − ∗ [B] ∗ ∗ [A] ∗ [X] (17)
∂x ∂x

Replacing the equivalent control by its expression in (15), we obtain the new
expression of the derivative surface [10]:

˙ = ∂S ∗ [B] < 0
S(x) (18)
∂x
1680 O. Fadi et al.

In order to satisfy the condition, the sign of Un must be opposite of S(x) = ∂S∂x
∗[B].
The simplest form that the discrete control can take, is the ‘sign’ function presented
as:

Un = Kx ∗ sing(S(x)) (19)

However, this function generates on the sliding surface, a phenomenon called

chattering which is generally undesirable because it adds high frequency components
to the control spectrum. The chattering issue can be decreased by replacing the “sign”
function with an adequate saturation function which filters out the great frequencies
[10].
⎧
⎪
⎨ −1 if S < −η
S
Sat(S) =
⎪ η if |S| < −η
⎩ 1 if S > −η

η: A small and positive parameter, defines the bandwidth of ‘sat’ function.

4 SM Application on the SEIG

In this study, the SM technique is applied to the asynchronous generator and the
control expressions are established based on the following layout model:
Figure 1 shows the cascade configuration of sliding mode regulator employed
to the SEIG. The external loop output variable presents the current references of
the internal loops [10]. The references output are the direct and quadrature control
voltages to be applied (Vrd ,Vrq ) to the PWM rectifier. The cascade configuration
imposes a choice of three aeras along the two axes (d,q). The internal loops regulate
the currents “id and iq” and the external loop regulates the Vdc voltage.

Fig. 1 Outline schematic of the SM controller

SM Versus PI Tunning for PWM Rectifier for SEIG ... 1681

4.1 Direct Current Regulator Id:

The sliding area related to this regulator is defined as follow [11]:

S(id ) = id∗ − id (20)

The sliding aera derivative is:

dS(id ) di∗ did

= d − (21)
dt dt dt
Substituting the Eq. (8) in Eq. (21) we deduced:
 
dS(id ) di∗ 1 R
= d − Vds − i d + wi q (22)
dt dt L L

The control low Vr∗d is defined as follow:

∗ eq
Vrd = Vrd + Vrd
n
(23)

The equivalent component of the voltage is:

eq di ∗d
Vrd = Vds − Ri d + wLi q − L (24)
dt
The discontinuous component of the voltage is given by:
∗
Vrd = −K d ∗ sat(Sd ) (25)

4.2 Quadrature Current Regulator Iq:

By following the same approach detailed above, the equivalent component of the
voltage Vrq∗ is given by:

di∗q
eq
Vrq = Vqs − Riq − wLid − L (26)
dt
The discontinuous component of the voltage is:
∗
Vrq = −K q ∗ sat(Sq ) (27)
1682 O. Fadi et al.

4.3 DC-Bus Voltage Regulator Vbus :

The sliding area related to regulator Vbus is expressed as follow [11]:

S(Vbus ) = Vbus ∗ −Vbus (28)

 
dS(Vbus ) dV∗bus dVbus dV∗bus ibus Vbus
= − = − − (29)
dt dt dt dt Cbus Cbus ∗ RL
∗
The control low i bus is defined as follow:
∗ eq
ibus = ibus + ibus
n
(30)

With, the discontinuous and equivalent components are as follow:

n
ibus = Kbus ∗ sing(S(Vbus ) (31)

eq Vbus
ibus = (32)
RL

By neglecting the inverter/rectifier losses and assuming that i d = 0, we obtain:

Vbus ∗
Pbus = Vbus *ibus = Vq *iq ⇒ i∗q = *i (33)
Vq bus

Therefore, the control low of dc-bus is expressed as follow:

 
Vbus   Vbus
iq∗
eq
= ibus + n
ibus = ∗ Kbus ∗ sing(S(Vbus + ) (34)
Vq RL

5 Simulation Results and Interpretations

Figure 2 presents the dc bus voltage across the rectifier terminal; the regulation of
this bus is examined by both SM and PI controllers. In spite of the presence of load
step changing and, as illustrate in the figure below, the SM performances of the DC
voltage is better than PI performances, in termers of, the robustness against the load
disturbances, also, the dynamic property is better: there is no overshoot and the SM
response time is smaller than the classical PI time response.
Figure 3 demonstrates the performances of active and reactive powers in the
presence of load variation. During the steady regime, the P power is maintained
constant, and the Q power is maintained at 0 KVA which lead to obtain a good
SM Versus PI Tunning for PWM Rectifier for SEIG ... 1683

Fig. 2 DC voltage in the terminal of the PWM rectifier

Fig. 3 P power, Q power and PF of the SEIG

1684 O. Fadi et al.

power factor (PF = 1), furthermore, The SM ensures a good performance than the
conventional PI regulator (smaller overshoot, smaller response time and constant
steady state).
Figure 4 shows the steady state for the SEIG currents, using both SM and PI control
strategies. As can be observed, the SEIG currents, in SM regulator technique, are
close to sine wave better more than the PI regulator technique. Furthermore, the
Fig. 5 shows that the SM exhibits the best power quality (THD = 1.96%) and, the
PI leads to a dispersed harmonic spectrum with a higher THD of around 6.24%.

Fig. 4 PI versus SM approaches for SEIG currents

Fig. 5 THD of SEIG currents (PI versus SM)

SM Versus PI Tunning for PWM Rectifier for SEIG ... 1685

6 Conclusion

This work proposes a nonlinear control of the self-excited induction generator using
SM approach. In the suggested control technique, stable and independent controls
of the direct and quadrature currents are achieved. This method can be easily used
in the isolated wind power systems. Furthermore, the simulation results demonstrate
that the dc voltage obtained with the SM tunning is faster compared to the clas-
sical PI. Furthermore, the SM controller brings remarkable improvements compared
to the classic PI regulator since the SM regulators offer good static and dynamic
performances (stability, rapidity and precision), that is to say a shorter response time
without overshoot, also they grant better tracking and rejection of load disturbances,
and reduce THD of SEIG currents.

References

1. Kalaivani C, Rajambal K (2020) Modeling of an efficient high power wind energy conversion
system using self-excited multi-phase machines. Microprocess Microsyst 74:103020. https://
doi.org/10.1016/j.micpro.2020.103020
2. Ouafia F, Ahmed A (2018) Elaboration of the minimum capacitor for an isolated self excited
induction generator driven by a wind turbine. In Conference: International Conference of
Computer Science and Renewable Energies, Ouarzazat
3. Chaturvedi Y, Kumar S, Gupta V (2020) Capacitance requirement for rated current and rated
voltage operation of SEIG using whale optimization algorithm. Procedia Comput Sci 167:2581–
2589. https://doi.org/10.1016/j.procs.2020.03.315
4. Silva EO, Vanço WE, Guimarães GC (2020) Capacitor bank sizing for squirrel cage induction
generators operating in distributed systems. IEEE 8:27507–27515
5. Wang F, Zhang Z, Mei X, Rodríguez J, Kennel R (2018) Advanced control strategies of
induction machine: Field oriented control, direct torque control and model predictive control.
Energies 11(1):120
6. Toufik A, Taibi D, Oualid A (2018) Voltage oriented control of self-excited induction generator
for wind energy system with MPPT. In AIP Conference Proceeding
7. Abdelrahem M, Hackl C, Kennel R, Rodriguez J (2019) Sensorless predictive speed control
of permanent-magnet synchronous generators in wind turbine applications. In PCIM Europe,
International Exhibition and Conference for Power Electronics, Intelligent Motion, Renewable
Energy and Energy Management (1–8) VDE
8. Ouafia F, Ahmed A (2020) A direct power control of the PWM rectifier for SEIG feeding
resistive load in wind energy systems. In 5th International Conference on Renewable Energies
for Developing Countries, pp 1–6. IEEE
9. Mohan H, Pathak MK, Dwivedi SK (2019) Direct power control of induction motor drives.
In IEEE 13th International Conference on Compatibility, Power Electronics and Power
Engineering pp 1–5
10. Boudries ZOUBIR, Aberbour A, Idjdarene KASSA (2016) Study on sliding mode virtual
flux-oriented control for three-phase PWM rectifiers. Rev Roum Sci Techn–Électrotechn
Énerg 61(2):153–158
11. Xiong L, Li P, Wang J (2020) High-order sliding mode control of DFIG under unbalanced grid
voltage conditions. Int J Electr Power Energy Syst 117:105608
Improved DTC of the PID Controller
by Using Genetic Algorithm of a Doubly
Fed Induction Motor
Said Mahfoud, Aziz Derouich, Najib El Ouanjli, Mohammed Taoussi,
and Mohammed El Mahfoud

Abstract This paper presents a new strategy for the Direct Torque Control (DTC) of
the Doubly Fed Induction Motor (DFIM) powered by two voltage inverters. The speed
regulation of this control is done by a PID (proportional integral derivative) controller
optimized by a Genetic Algorithm (GA). The classic DTC has several advan-
tages such as (dynamic, robustness, ease of implementation, high performance…).
However, it has ripples of the electromagnetic torque and variable switching frequen-
cies that generate vibrations of the machine causing rapid aging. To overcome the
disadvantages of classic DTC control, the new GA-DTC control strategy based on the
PID controller optimized by the genetic algorithm, have the tendency to operate the
system throughout its operating range. The entire system is implemented and vali-
dated on the Matlab/Simulink environment to improve machine performance and
behavior.

Keywords DFIM · DTC · GA-DTC · PID · GA

1 Introduction

During the mid-80s, the control technique of the induction motor, known as Direct
Torque Control or DTC, has appeared to compete with traditional techniques. This
strategy has been introduced by TAKAHASHI [1] and DEPENBROCK [2]. Its prin-
ciple for a DFIM is based on a direct identification of the control pulses applied to
the voltage inverters switches, in order to preserve the stator and rotor flux and elec-
tromagnetic torque within two predefined hysteresis bands. Such an application of
this technique makes it possible to ensure decoupling between the flux and the torque

S. Mahfoud (B) · A. Derouich · N. El Ouanjli · M. Taoussi

Industrial Technologies and Services Laboratory, Higher School of Technology, Sidi Mohamed
Ben Abdellah University, Fez, Morocco
e-mail: [email protected]
M. E. Mahfoud
Laboratory of Systems Integration and Advanced, Faculty of Sciences Dhar El Mahraz, Sidi
Mohamed Ben Abdellah University, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1687
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_153
1688 S. Mahfoud et al.

control without using the Pulse Width Modulation (PWM) or coordinate transforma-
tion. The important features of DTC are its fast dynamic torque response, robustness,
ease of implementation, and high performance [3, 4]. However, this strategy has two
important disadvantages: First of all, the frequency of switching is highly variable.
Second of all, the torque and flux ripples remains poorly handled in the entire oper-
ating range. It is worthwhile to note that the torque ripples are making additional
noises and vibrations. Therefore, they are causing fatigue in the machine shaft. In
order to reduce the impact of these phenomena on machine life, it is estimated that
intelligent techniques can provide improvements.
Currently, many researchers have proposed solutions to improve the performance
of classic DTC control, which are based on artificial intelligence. These approaches
are very robust to measure noise and parametric variations.
In [5–7], the authors have proposed many techniques to enhance the efficiency
of direct torque control using intelligent techniques. They are called Direct Torque
Fuzzy Control (DTFC), Direct Torque Neural Control (DTNC), and Direct Torque
Neural-Fuzzy Control (DTNFC). They use fuzzy logic controllers or neural networks
or an Adaptive Neuro-Fuzzy Inference System (ANFIS) combining artificial neural
networks and fuzzy logic, which replace truth tables and hysteresis comparators, to
generate the voltage vector in order to direct flux and torque to their references.
In [9] the authors have proposed a new DTC strategy of the DFIM, The PI speed
controller is optimized by using the genetic algorithm of DTC control at the stator, in
order to minimize torque ripples. But this method does not allow us to take advantage
of the double speed band of DFIM [10].
In order to solve this problem, this work is articulated on the application of the
DTC control on both sides of the DFIM and the optimization of PID controller
parameters based on genetic algorithm to resolve the precedent problems. In this
article, the contributions are focused on the following objectives:
• Minimization of the flux and torque ripples
• Reduction of response time and speed rejection time
• Minimization of the speed and torque overshoot
This article is organized according to the following form: Sect. 2 Presents the
mathematical model of the DFIM, Sect. 3 Describe the operation principal and math-
ematical model of the DTC, Sect. 4 is focused to study the genetic algorithm, Sect. 5
is articulated for interpreted the results of simulation and finally the conclusion of
control results and suggestion of future works.

2 Doubly Fed Induction Motor Mathematical Model

The mathematical model of doubly fed induction motor appropriate for studying
the behavior of direct torque control is the model of two-phase represented by the
coordinates (α, β), it is represented by the following equations [11]:
Improved DTC of the PID Controller by Using Genetic Algorithm ... 1689

• Electrical equations:
⎧
⎪
⎪
dψsα
⎪ vsα = Rs .i sα +
⎪
⎪
⎪ dt
⎪
⎪
⎪
⎪ dψ sβ
⎨ vsβ = Rs .i sβ +
dt (1)
⎪
⎪ dψr α
⎪ vr α = Rr .ir α +
⎪ + ωm .ψrβ
⎪
⎪ dt
⎪
⎪
⎪
⎪ dψrβ
⎩ vrβ = Rr .irβ + − ωm .ψr α
dt
• Magnetic equations:
⎧
⎪
⎪ ψ = L s i sα + L m .ir α
⎨ sα
ψsβ = L s i sβ + L m .irβ
(2)
⎪
⎪ ψ = L r ir α + L m .i sα
⎩ rα
ψrβ = L r irβ + L m .i sβ

• Mechanical equations:
⎧
⎨ Tem = p.(ψsα i sβ − ψsβ i sα )
(3)
⎩ J. d + f. = Tem − Tr
dt

3 Direct Torque Control Strategy

The DTC is less sensible in front of parametric variations of the motor and it allows
us to obtain precise and fast torque response. The objective of this strategy is to
control directly the machine torque and flux. In this context, a hysteresis comparator
is used to compare the estimated value and the reference one, and then the state of
the inverters is controlled by a predetermined selection table. Figure 1 illustrates the
structure of the DTC control strategy applied to the doubly fed induction motor.

3.1 Control of the Vectors of Stator and Rotor Flux

In the fixed (α, β) plane associated with the stator and rotor. The rotor and stator flux
are estimated according to the equations below [11].
 t
ψ s (t) = 0 (V s + Rs .I s ).dt
t (4)
ψ r (t) = 0 (V r + Rr .I r ).dt
1690 S. Mahfoud et al.

Fig. 1 Structure of the direct torque control strategy optimized by the Genetic Algorithm

The flux are defined by their modules and their positions, which are given by the
following relationship:

ψ̂ = ψ̂α2 +ψ̂β2
ψ̂
(5)
θ = ar ctg( ψ̂β )
α

By using the flux in the reference (α, β), the electromagnetic torque can be
calculated using a mathematical equation.

T̂em = p.(ψ̂sα .i sβ − ψ̂sβ .i sα ) (6)

Improved DTC of the PID Controller by Using Genetic Algorithm ... 1691

Fig. 2 Flux trajectory

Fig. 3 Hysteresis comparators (a) of the three-levels torque (b) of the two-levels flux

3.2 Hysteresis Comparators

Two hysteresis comparators with two-levels, which are easy to use. Their roles are to
maintain the end of each vector of the flux in a circular crown as shown in Fig. 2, and a
three-levels hysteresis comparator Fig. (3, a), allows controlling the electromagnetic
torque developed by the motor in both directions of rotation, they are generating
either a positive torque or a negative one.

3.3 Elaboration of the Switching Table

Depending on the sector and the evolution of the torque and flux, you should select
the voltage vector to be applied in order to respect the references of the torque and
flux. The truth table for selecting the appropriate vector is shown in Table 1.
1692 S. Mahfoud et al.

Table 1 Inverter sequences

Sector
 Ccpl S1 S2 S3 S4 S5 S6
1 1 ν2 ν3 ν4 ν5 ν6 ν1
0 ν7 ν0 ν7 ν0 ν7 ν0
−1 ν6 ν1 ν2 ν3 ν4 ν5
0 1 ν3 ν4 ν5 ν6 ν1 ν2
0 ν0 ν7 ν0 ν7 ν0 ν7
−1 ν5 ν6 ν1 ν2 ν3 ν4

Fig. 4 Structure of the

speed PID controller
optimization by GA

4 Optimization of the Speed Controller by the Genetic

Algorithm

Genetic algorithm is a research strategy to provide the best solution for non-linear
systems. Classic DTC controller and machine are affected by variable parameters,
which makes the system non-linear, for this situation, a genetic algorithm generates
the optimal values for KP , KI and KD of the PID controller at each occurrence of a
change, in the system in order to master the situation. Figure 4 illustrates the reduced
structure of the system.
The genetic algorithm belongs to the category of evolutionary algorithm, which
uses techniques based on evolutionary biology, such as selection, crossover, and
mutation [12], The sequences of operations involved in GA are described in Fig. 5,
which presents a flowchart that respects the evolutionary rules of a genetic algorithm.

5 Simulation and Interpretation

The implementation of the GA-DTC control and the DFIM is realized in the
Matlab/Simulink platform. The configuration of the global system is done according
to the parameters Tables 5, 6 and 7. To find out the behavior of the system, the latter
is subjected to variable references of speed and torque to put it in situations similar
to reality. The results of the simulation for both controls (GA-DTC and classic DTC)
were tested using a 1.5 kW machine.
Improved DTC of the PID Controller by Using Genetic Algorithm ... 1693

Fig. 5 Flowchart of the speed PID controller optimization by GA

Figures (6, a), (6, b), (6, c) and (6, d), respectively show the variation of the speed,
and the variation of the electromagnetic torque and the stator and rotor flux, which
are obtained without load as a condition of starting and at nominal load from 0.5 s
as normal operating condition.
Figures (6, a), and (6, b) illustrate, the speed and torque characteristics of a doubly
fed induction motor controlled by the GA-DTC, which quickly reach their references
without overshoot compared with a classic DTC during no-load starting and during
nominal load starting.
For the speed characteristic the response time is minimized by a rate greater than
37% (48.5 ms for the classic DTC and 30.4 ms for the proposed GA-DTC), and the
overshoot is reduced by approximately 100% (9 rad/s for classic DTC and 0 rad/s for
proposed GA-DTC control), and the rejection time is reduced by 16.73% (23.3 rad/s
for the classic DTC and 19.4 rad/s for GA-DTC control).
For the characteristic of the electromagnetic torque, the ripples are reduced by
more than 22% (4,264 Nm for classic DTC and 3,297 Nm for the proposed GA-DTC).
1694 S. Mahfoud et al.

Fig. 6 System responses (a)

Speed (b) Torque (c) Stator
flux (d) Rotor flux

For the stator and rotor flux characteristics, the ripples are reduced to 36.05% and
23.74% respectively (0.1079 Wb and 0.069 Wb for the classic DTC and 0.0514 Wb
and 0.0392 Wb for the proposed GA-DTC).
Improved DTC of the PID Controller by Using Genetic Algorithm ... 1695

Table 2 Performance measures DTC classic and GA-DTC

Characteristics Optimized DTC by Classic DTC Improvement (%)
GA
 Response time (ms) 30.4 48.5 37.32
Overshoot (rad/s) 0 9 100
Rejection time (ms) 19.4 23.3 16.73
T Response time (ms) 15.4 16.1 4.35
Overshoot (rad/s) 4.27 5.49 22.22
Ripples (Wb) 3.297 4.264 22.68
ψs Ripples (Wb) 0.069 0.1079 36.05
ψr Ripples (Wb) 0.0392 0.0514 23.74

The results presented in Table 2 showed that the optimization of the DTC
control PID controller by the genetic algorithm improved the classic DTC control
performances (dynamics, robustness…).

6 Conclusion

In this article, the optimization of the PID controller of the classic DTC control is
achieved by a genetic algorithm, which has been applied to a doubly fed induction
motor powered through two inverters. The modeling and details of the system are
introduced and implemented in the Matlab/Simulink platform. The results of the
simulation improve the performance of the doubly fed induction motor, such as
minimizing the response time of 37.32% and the rejection time is improved by
16.73%, and the torque ripples which are minimized by up to 22.68% . The proposed
control has increased the performance of the classic DTC control at the transient and
dynamic regime. In order to make a progress of the research the next work will be
on the implementation of the proposed control on the dSPACE DS1104 board to test
the validity of this technique.

Appendix

(See Tables 3, 4, 5, 6 and 7).

1696 S. Mahfoud et al.

Table 3 Nomenclature
Parameters Description
Vsα , Vsβ , Vrα and Vrβ Stator and rotor voltages in (α, β) plan
Udcs and Udcr Stator and rotor directs voltages
Isα , Isβ , Irα and Irβ Stator and rotor currents in (α, β) plan
 sα ,  sβ ,  rα and  rβ Stator and rotor flux in (α, β) plan
Rs , Rr Stator and rotor resistors
Ls , Lr Stator and rotor inductors
Lm Mutual Inductance
P Number of pairs of poles
ωr Rotor angular speed
ωs Stator angular speed
 Rotation speed
Tem Electromagnetic torque
Tr Resistant torque
f Viscous friction coefficient

Table 4 Abbreviations
Abbreviation Wording
DFIM Doubly Fed Induction Motor
DTC Direct Torque Control
GA Genetic Algorithm
GA-DTC Genetic Algorithm-Direct Torque Control
PID Proportional Integrator Derivator
DTFC Direct Torque Fuzzy Control
DTNC Direct Torque Neural Control
DTNFC Direct Torque Neural-Fuzzy Control
ANFIS Adaptive Neuro-Fuzzy Inference System

Table 5 DFIM parameters

Symbols Values (unit)
Pn 1.5 Kw
Vs 230 V
Vr 130 V
P 2
f 50 Hz
Rs 1.75 
(continued)
Improved DTC of the PID Controller by Using Genetic Algorithm ... 1697

Table 5 (continued)
Symbols Values (unit)
Rr 1.68 
Ls 0.295 H
Lr 0.104 H
Lm 0.165 H
f 0.0027 kg.m2 /s
J 0.001 kg.m2

Table 6 Parameters of the

Parameters/Control GA-DTC Classic DTC
PID of GA-DTC and PID of
Classic DTC KP 96.4889 50
KI 0.1576 4
KD 0.9706 0

Table 7 Parameters of the

Description Type/Value
genetic algorithm
Population size 20
Maximum iteration 50
Selection Uniform
Crossover Roulette wheel selection
Mutation Uniform

References

1. Takahashi I, Ohmori Y (1989) High-performance direct torque control of an induction motor.

IEEE Trans Ind Appl 25:257–264
2. Depenbrock M (1988) Direct Self Control of Inverter-Fed Induction Machines. IEEE Trans
Power Electron PE-3(4):420–429
3. Taoussi M, Karim M, Bossoufi B, Hammoumi D, Lagrioui A (2015) Speed backstepping control
of the double-fed induction machine DRIVE. J Theor Appl Inf Technol 74(02):189–199
4. Mahfoud S, Derouich A, El Ouanjli N, Mohammed T, Hanafi A (2021) Field oriented control
of doubly fed induction motor using speed sliding mode controller. In: E3S web of conferences,
vol 229. EDP Sciences, pp 01061
5. Abbou A, Mahmoudi H (2009) Performance of a sensorless speed control for induction motor
using DTFC strategy and intelligent techniques. J Electr Syst 5(3):64–81
6. Ouanjli NE, Derouich A, Ghzizal AE, Chebabhi A, Taoussi M, Bossoufi B (2018) Direct
torque control strategy based on fuzzy logic controller for a doubly fed induction motor. In:
IOP conference series: earth and environmental science
7. Cirrincione G, Cirrincione M, Chuan L, Pucci M (2016) Direct torque control of induction
motors by use of the GMR . In: Neural networks, proceedings of the international joint
conference, 20–24 July 2003, vol 3, pp 2106–2111
8. Jang J-SR, Sun C-T (1995) Neuro-fuzzy modelling and control. Proc IEEE 83:378–406
9. Piotr F (2001) Neuro-fuzzy control of inverter-fed induction motor drives. Thesis. Lublin
University of Technology, Poland
1698 S. Mahfoud et al.

10. Zemmit A, Messalti S, Harrag A (2017) A new improved DTC of doubly fed induction machine
using GA-based PI controller. Ain Shams Eng J 9:1877–1885
11. El Ouanjli N, Derouich A, El Ghzizal A et al (2019) Modern improvement techniques of direct
torque control for induction motor drives - a review. Prot Control Mod Power Syst 4:11. https://
doi.org/10.1186/s41601-019-0125-5
12. Goldberg DE (1989) Sizing populations for serial and parallel genetic algorithms. In:
Proceeding of the third international conference on genetic algorithms, CL, USA
State Feedback Control of DC-DC
Converter Using LQR Integral
Controller and Kalman Filter Observer
Djamel Taibi , Toufik Amieur , Mohcene Bechouat , Sami Kahla ,
and Moussa Sedraoui

Abstract In this paper, the linear state feedback control using LQR controller for
a DC/DC converter in the case of negative voltages topology is presented in order
to achieve a particular desired behavior. To guarantee a zero steady-state error, we
introduce an integral action, which will work out this problem by assuring that
the steady-state error will end up to zero. For filtering and state estimation with
a low cost and less complexity a state observer is obtained based a Kalman Filter
observer. Detailed simulation study is presented to demonstrating the robustness and
effectiveness of the proposed control scheme.

Keywords Linear quadratic regulator · DC/DC Buck-Boost Converter · Kalman

filter · Static error

1 Introduction

Power supply technology enables technologies that allow us to operate circuits and
electronic systems. All digital and analog circuits require a power source to operate.
Many of these circuits require multiple DC supply voltages. DC power supplies

D. Taibi (B)
Département de Génie Électrique, Université de Kasdi Merbah, Ouargla, Algeria
T. Amieur
Département de Génie Électrique, Université de Larbi Tebessi, Tebessa, Algeria
e-mail: [email protected]
T. Amieur · M. Bechouat · M. Sedraoui
Laboratoires des Télécommunications LT, Université 8 mai 1945, Guelma, Algeria
e-mail: [email protected]
M. Bechouat
Faculté des Sciences et Technologie, Université de Ghardaïa, Noumirat BP 455,
Route Ouargla Ghardaia, 47000 Ghardaia, Algeria
S. Kahla
Centre de Recherche en Technologies Industrielles CRTI, 16014 Cheraga, Alger, Algeria

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1699
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_154
1700 D. Taibi et al.

are widely used in telecommunications, household appliances, defense, and medical

electronics [1]. The DC voltage is generally obtained from a battery or by converting
AC current into DC current using a transformer, rectifier, and filter [1, 2].
In some power supplies, a negative output voltage are required. There are many
different methods for generating a negative voltage in the output from a positive
voltage in the input. A very simple and inexpensive method for obtaining a negative
voltage is the Buck-Boost converter. The main advantage of this converter is the
simplicity of design. The topology requires very few components, which reduces the
cost and complexity of the development.
In [3] a proportional-integral voltage regulator (PI) is extended by a sensor-
less predictive control of Buck-Boost converter using a self-correction differential
current observer. [4] Presents three adaptive optimal PI controllers that can be used
for switching power converters with the unknown load resistance. Robust control
used for the control of DC-DC converters in [5, 6]. Presents LQR control of power
converters. Similar to this paper, [7] introduces an integral action for DC/DC Boost
converter. In [8] a multi-loop based PI control and an LQG method are used to control
a DC/DC Buck converter. The nonlinear control is used in [9] for Boost converter.
This paper describes the LQR controller for a Buck-Boost with a negative output
voltage. Finally, the Kalman filter is a special type of observer that enables optimal
filtering of the different types of noise in the measurement and the system if the
covariance of these noises is known [10].
In Sect. 2, the Buck-Boost converter model is used directly for simulation purpose
using the LQR controller. The LQR is detailed in Sect. 3. The Kalman filter observer
is examined in Sect. 4. The results of the numerical simulation are presented in
Sect. 5. Finally, conclusion are presented in Sect. 6.

2 The Buck-Boost Model

The Buck-Boost converter with negative output voltages topology is as shown in

Fig. 1.
Mosfet Q is ON (Fig. 2), the charging current in the capacitor C is partially
discharged. During the second interval in which the Mosfet Q is switched off (Fig. 3),
the polarity of the voltage on the inductor is reversed and the diode is conductive
[11, 12].

Fig. 1 Buck-boost converter

State Feedback Control of DC-DC Converter ... 1701

Fig. 2 Switched ON of the Buck-Boost converter

Fig. 3 Switched OFF of Buck-Boost converter

2.1 Model of Switched Buck-Boost Converter

By applying Kirchhoff’s laws in the two previous circuits, we get the following
dynamics:
In the switching ON, the following dynamic is obtained [12]:
⎧ di
⎪
⎨L =E
dt (1)
⎩ C dv = − v
⎪
dt R
In the switching OFF, the following dynamic is obtained [12]:
⎧ di
⎪
⎨L =v
dt (2)
⎩ C dv = −i − v
⎪
dt R
From the two Eqs. (1) and (2), we can obtain a single unified model, which is
u ∈ {0, 1}.The dynamics of the converter are therefore as follows [13]:

L di
dt
= (1 − u)v + u E
(3)
C dv
dt
= −(1 − u)i − Rv
1702 D. Taibi et al.

3 LQR Control

3.1 LQR Controller

The Eq. (4), describe the state space model of a system: [14]

ẋ(t) = Ax(t) + Bu(t)
(4)
y(t) = C x(t) + Du(t)

The vector of optimal control is as follows:

u(t) = −K x(t) (5)

To define the optimal control inputs and optimize the state variables, the following
cost function J must be minimized [15]:
 ∞ 
J= x T Qx + u T Ru dt (6)
0

Q and R two positive definite matrix, Q and R are chosen, the LQR control
problem reduces to finding K that minimizes (6).
Solution P of Riccati equation [15]:

P A + A P + Q − P B R −1 B  P = 0 (7)

The gain matrix of the optimal control vector can be calculated using the following
equation [15]:

K = R −1 B  P (8)

Therefore, optimal control equation becomes [16]:

u(t) = −K x(t) = −R −1 B T P x(t) (9)

3.2 LQR Control with Integral Action

The schematic diagram LQR with integral is shown in Fig. 4 [10, 16].
The LQR with integral action require the following extended matrix configuration:



A 0 B  
 = , B̂ = , Ĉ = 0 0 , K̂ = k −ki
−C 0 0
State Feedback Control of DC-DC Converter ... 1703

Fig. 4 Schematic of LQR with integral action

With
k: State-feedback vector gain.
ki : Integral action.
The cost function is known by the following relationship [15]:
 ∞  T
J= e (t)Qe(t) + u T (t)Ru(t) dt
0

The vector e(t) is defined by:


x̂(t) − x(∞)
e(t) = (10)
ξ (t) − ξ (∞)

Where ξ (t) represents the integral action.

The gain K in (12) is obtained by solving the:

 T P + P  − P B̂ R −1 B̂ T P + Q = 0 (11)

K̂ = R −1 B̂ T P (12)

4 Kalmen Filter

4.1 Kalman Observer

We consider that only the output voltage is sensed. Then, to estimate the inductor
current and reduce the noisy environment effects, a Kalman filter is introduced [10]:

ˆ = A x̂(t) + Bu(t) + ω(t)
ẋ(t)
(13)
ŷ(t) = C x̂(t) + ν(t)
1704 D. Taibi et al.

where ω(t) is a process noises, and ν(t) is the voltage sensor noise.
The state equations of the Kalman filter can be carried out as follows [17]:

ẋˆ = (A − K K C)x̂ + Bu + K K y (14)

Where K is the Kalman gain matrix.

The Kalman filter is usually designed for linear time-varying systems is as follows
[17]:

d ẋˆ = A(t)ẋ(t)dt
ˆ ˆ
+ B(t)u(t)dt + K K (t)(y(t)dt − C(t)ẋ(t)dt) (15)

K K (t) = PK (t)C T (t)R −1

K (t) (16)

d
PK (t) = A(t)PK (t) + PK (t)A T (t) − PK (t)C(t)T R −1
K (t)C(t)PK (t) + Q K (t)
dt
(17)

x̂(t0 ) = x0 , P(t0 ) = P0

5 Simulation Results

In the MATLAB/SIMULINK using SimPowerSystems toolbox. In simulation,

the Kalman filter observer algorithm was implemented using the continuous-time
Kalman filter block.
The buck-boost converter with negative output voltages topology was simulated
and switched by the PWM technique. The parameters used during the analysis of
simulation results are shown in Table 1.

Table 1 The Buck-Boost

E Input voltage 20 [V]
parameters
L Inductance 15.91 [mH]
C Capacitance 470 [µF]
R Loading resistance 20 []
Vref The reference output voltage −40 [V]
Iref The reference inductor current 6 [A]
Uref Initial duty cycle 66.67%
VD Diode voltage drop 0.7 [V]
RD Diode on-resistance 0.05 []
RSW Mosfet on-resistance 0.1 []
FSW Commutation frequency 20 [kHz]
State Feedback Control of DC-DC Converter ... 1705

Fig. 5 Open and closed loop control

Open-Loop Open-Loop
Closed- Closed-

Fig. 6 Simulation results of open and closed-loop with LQR controller

5.1 Analysis for Open and Closed-Loop Control

In this case, the simulation of the buck boost converter is carried out for two cases
for the control behavior of the converter in the Open and Closed-Loop control using
LQR. The SimPowerSystems model is shown in Fig. 5.
According to results of the simulation, the waveform of output voltage and the
inductance current have the same waveform characteristic for the open and closed
loop control (Fig. 6).
The value of the output voltage in the closed-loop is approximately −40 V and
about −37 V in the open-loop. Then the average value of the inductance current is
about 6 A and 5.6 A in closed and open-loop respectively. A comparison between
the simulation results shows that the control with closed-loop using LQR has better
dynamics response compared with open-loop as well as in the maximum peak
overshoot/ undershoots, settling time.
1706 D. Taibi et al.

Fig. 7 SimPowerSystems model of buck-boost converter with LQR and Kalman filter observer

Estimated
Measured

Fig. 8 Simulation results for estimated and measured output voltage

Estimated
Measured

Fig. 9 Simulation results for estimated and measured inductor current

State Feedback Control of DC-DC Converter ... 1707

5.2 Analysis for Buck-Boost Converter with LQR

and Kalman Filter Obsever

This section presents a simulation of the dynamic performance of an LQR controller

with a Kalman filter observer. The algorithm of the observer was implemented
using the Continuous-Time Kalman filter block. The Simulink model with LQR
and Kalman filter observer is shown in Fig. 7.
In the Kalman filter observer, the matrices QK and RK are difficult to know exactly,
since the noises w and v are not known. The only possible method is to adjust the QK
and RK values using simulations. In this simulation, we use the values as follows:


109 0 
QK = R K = 103
0 104

The estimated and measured output voltage and inductance current at an output
voltage reference of −40 V are shown in Fig. 8 and Fig. 9. In both static and dynamic
cases, it can be seen that the signals estimated by the KF observer track the real signals
very well.

5.3 Effect of a Measurement Noise

In many applications, the measurement noise caused by the hardware or the environ-
ment has a significant impact on the system. The estimation accuracy of KF is tested
in this paper under noisy output voltage measurement. Figure 10 shows the noise
injected at the output voltage in the range of [0–2 A]. The noise is zero mean, white
Gaussian. The aim of the voltage injection is to observe the low pass filter characteris-
tics of KF. Figure 11 shows the output voltage, the inductor current shown in Fig. 12.
The accuracy of the state estimate can be increased by increasing the covariance of

Fig. 10 Injected noise to output voltage

1708 D. Taibi et al.

Noisey
Estimated
Measured

Fig. 11 Measured and estimated, output voltage with measured noisy

Estimated
Measured

Fig. 12 Measured and estimated inductor current

the measurement noise under noisy conditions, which makes the system model will
have more importance.

6 Conclusion

This paper has demonstrated the process of applying the state feedback control of a
buck-boost converter using LQR with integral action controller. The clear advantages
of state feedback, that has a positive effect on response settling time, reducing the
undesirable peak overshoots and serve having a less oscillated performance, referring
that this approach doesn’t provide a zero static error, this latter has been solved by
adding an integral action to state feedback control. The simulation results found that
adding a Kalman filter reduces the cost and impact of the sensor on LQR tracking
performance.
State Feedback Control of DC-DC Converter ... 1709

References

1. Kazimierczuk MK (2015) Pulse-width modulated DC-DC power converters. Wiley, Hoboken

2. Boyette A (2006) Contrôle-commande d’un générateur asynchrone à double alimentation avec
système de stockage pour la production éolienne. Thèse de doctorat. Université Henri Poincaré-
Nancy 1
3. Zhang Q, Min R, Tong Q, Zou X, Liu Z, Shen A (2014) Sensorless predictive current controlled
DC–DC converter with a self-correction differential current observer. IEEE Trans Industr
Electron 61(12):6747–6757
4. Michel L (2012) Pilotage optimal des IGBT et commande sans-modèle des convertisseurs de
puissance. Thèse de doctorat. Université du Québec à Trois-Rivières
5. Zhuo S, Gaillard A, Xu L, Bai H, Paire D, Gao F (2020) Enhanced robust control of a DC–DC
converter for fuel cell application based on high-order extended state observer. IEEE Trans
Transp Electrif 6(1):278–287
6. Agrawal N, Samanta S, Ghosh S (2020) Modified LQR technique for fuel-cell-integrated boost
converter. IEEE Trans Ind Electron 68(7):5887–5896
7. Battiston A (2014) Modélisation, commande, stabilité et mise en oeuvre des onduleurs à source
impédante: application aux systèmes embarqués. Thèse de doctorat. Université de Lorraine
8. Mojallizadeh MR, Badamchizadeh MA (2018) Hybrid control of single-inductor multiple-
output converters. IEEE Trans Ind Electron 66(1):451–458
9. Gateau G, Maussion P, Meynard T (1997) De la modélisation a la commande non linéaire des
convertisseurs multicellulaires série Application à la fonction hacheur. J Phys III 7(6):1277–
1305
10. Taibi D, Titaouine A, Benchabane F, Bennis O (2015) Stability analysis of the extended Kalman
filter for permanent magnet synchronous motor. J Appl Eng Sci Technol 1(2):51–60
11. Fang S, Wang X (2020) Modeling and analysis method of fractional order buck boost converter.
Int J Circuit Theory Appl 48(9):1493–1510
12. Ahmed NA (2005) Modeling and simulation of AC–DC buck-boost converter fed DC motor
with uniform PWM technique. Electr Power Syst Res 73(3):363–372
13. Hernandez L (2017) Etude et réalisation d’un convertisseur AC/DC Buck Boost réversible à
haut rendement pour alimentation de secours. Thèse de doctorat. Université de toulouse
14. Park M, Kang Y (2021) Experimental verification of a drift controller for autonomous vehicle
tracking: a circular trajectory using LQR method. Int J Control Autom Syst 19(1):404–416
15. Abbas NH, Algamluoli AF (2020) Designing an integral LQR controller for DC-DC X-
converter based on enhanced shuffled frog-leaping optimization algorithm. J Electr Syst
16(2):152–170
16. Dabin V (2018) Commande d’un quadricoptère par rejet actif de perturbations. Thèse de
doctorat. École Polytechnique de Montréal
17. Bai W, Xue W, Huang Y, Fang H (2018) On extended state based Kalman filter design for a
class of nonlinear time-varying uncertain systems. Sci China Inf Sci 61(4):1–16
Backstepping and Indirect Vector
Control for Rotor Side Converter
of Doubly Fed-Induction Generator
with Maximum Power Point Tracking

Elmostafa Chetouani, Youssef Errami, Abdellatif Obbadi,

and Smail Sahnoun

Abstract This manuscript proposes a comparative analysis between the Backstep-

ping and Indirect Field-Oriented Control (IFOC) based on conventional Proportional-
Integral regulator strategies of a wind energy conversion system based on a doubly-
fed induction generator. The stator of the generator is directly coupled to the elec-
trical network. However, the rotor is branched to the grid through two pulse width
modulation converters. This article aims to control the active and reactive power inde-
pendently by the two proposed methods. Furthermore, the maximum power tracking
with speed regulation is established and analyzed. The simulation results, taking into
consideration the wind speed variation of the wind energy chain using doubly-fed
induction of 5 MW, are realized by the Matlab/Simulink environment.

Keywords Backstepping (BS) · Indirect Field-Oriented Control (IFOC) ·

Maximum Power Tracking (MPT) · Proportional-Integral controller (PI)

1 Introduction

In the world as it is today and without electricity, daily life would be difficult to
imagine. It is, therefore, necessary to produce electric energy efficiently and contin-
uously. Following the terrible damage caused by the production of electricity on
the environment employing fossils combustible such as coal, oil, and natural gas.
Renewable energy-based on inexhaustible energy sources remains the best solution
to replace the existing energy sources [1]. Recently, wind energy, which is consid-
ered clean and sustainable energy, has known a great development and exploitation,
thanks to the strong growth in the research domain. The wind energy system has
several configurations. Figure 1 represents the studied system configuration, which
is widely used and dedicated to wind energy applications. The wind turbine shaft is
connected to the doubly fed-induction generator through the gearbox to conform the
slow-speed to the high one. DFIG stator is directly coupled to the electrical network.

E. Chetouani (B) · Y. Errami · A. Obbadi · S. Sahnoun

Laboratory: Electronics, Instrumentation and Energy, Faculty of Science,
University Chouaib Doukkali, Eljadida, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1711
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_155
1712 E. Chetouani et al.

Fig. 1 Whole chain of wind energy conversion system

However, the rotor is branched to the power grid via bidirectional converters. Using
the generator integrated with a wind energy system has many advantages such as the
ability to use a partial sized converter in the rotor to control the power, minimizing
power losses and cost, making fewer efforts on mechanical parts, noise reduction, the
control of active and reactive power, and a controllable power factor [2]. The gener-
ator is a non-linear system and has complex and coupling equations. The vector
control based on the stator or rotor field orientation combined with the Park trans-
formation makes the control of the generator a simple and easier task. Besides, this
vector control permits the command of the DFIG as a separately excited DC machine,
which makes it very popular in the industry. The direct and indirect methods are two
variants of the vector control. These types have been compared by Bouderbala et al.
[3]. The authors have shown that indirect control presents a satisfactory perfor-
mance than direct vector controls. The vector command based on the classical PI
controller has demonstrated performance limitations, especially when the parame-
ters of the generator change. Besides, the non-linear Backstepping approach, based
on the stability of the Lyapunov function, has been proposed by several researchers
to improve the performance of the system. S. Mensou et al. [4] have established
a Backstepping strategy to control the whole wind energy conversion system. The
produced mechanical power from wind energy depends on the characteristics of each
turbine, and eventually the wind variable speed. Consequently, tracking the maximal
power generated is required when the profile airspeed changes. To perform the MPT,
several techniques have been developed, and they can be divided into two categories
[5, 6]. The first one has required the characteristic aerodynamic curve of wind speed
information, and the second one does not need any information about wind speed
to generate excellent speed reference. The objective of this paper is to establish
a comparative study between the Backstepping and the Open Loop Indirect Field
Backstepping and Indirect Vector Control for Rotor Side Converter ... 1713

Oriented strategy for piloting the RSC converter of the induction generator. Also, the
maximum power following with speed control is established to generate the power
reference and the optimal rotation speed driving the generator. The article is organized
in this way: Sect. 2 introduces the modeling of the WECS. Section 3 discusses the
MPT with a speed control strategy computed by the classical PI controller. Section 4
proposes the control of the stator powers (Active and Reactive) by using Backstep-
ping and indirect vector control strategies. Section 5 examines and compares the
simulation results. Finally, Sect. 6 highlights the conclusion.

2 Modeling of the Wind Power System

2.1 Model of the Turbine

The turbine mechanical power is expressed as [7, 8]:

1
PTu = · Cp (λ, β) · ρ.π · R2 · V3W (1)
2
The mechanical torque TTu is written as below:

PTu
TTu = (2)
Tu

Where, ρ is the air density (kg/m3 ), R is the blade ray (m), Tu is the angular
velocity of the turbine, and CP represents the performance factor of the turbine. CP
can be formulated as [8]:
 
π(λ + 0.1)
CP (β, λ) =[0.5 − 0.0167.(β − 2)].sin
18.5 − 0.3.(β − 2)
− 0.00184.(λ − 3).(β − 2) (3)

This coefficient is calculated by the tip speed λ and the angle β of the blade pitch.
The latter is fixed to β = 2◦ for having CPmax . Equation (4) defines the expression of
λ [7]:

R.Tu
λ= (4)
V
1714 E. Chetouani et al.

2.2 Model of the Gearbox

Equation (5) presents the mechanical equation of the system, taking into consider-
ation that the overall mechanical dynamics are brought back to the turbine shaft,
[5–7]:

dmec
Jtot . + f.mec = Tg − Tem (5)
dt
Where
TTu mec
Tg = and GB = (6)
GB Tu

Where Jtot is the overall inertia of wind energy system, TTu is the turbine torque,
TTem is the electro-magnetic torque of the generator, f is the overall viscous coefficient
of friction, mec is the rotational speed at the rotor shaft of the gearbox (rad/s) and
GB is the gearbox multiplier.

2.3 Model of The Doubly Fed-Induction Generator

The Park transformation allows simplifying the general electrical model, which is
determined by the following equations. Stator and rotor voltages equations are [2,
8]:
⎧
⎪
⎪ V = Rs isd + dϕdtsd − ωs ϕsq
⎪ sd
⎨ dϕ
Vsq = Rs isq + dtsq + ωs ϕsd
(7)
⎪ dϕrd
⎪ Vrd = Rr ird + dt − ωr ϕrq
⎪
⎩ V = R i + dϕrq + ω ϕ
rq r rq dt r rd

Where, Vr and Vs are the voltages; is and ir are the currents; ϕs and ϕr are the
flux; Rs and Rr are the resistances; ωs and ωr are the angular frequencies; Ls and Lr
are the inductances; M is the mutual inductance. The r and s denote the rotor and
stator, respectively. The electromagnetic torque can be expressed as follows [8]:

M
Tem = −p. (irq ϕsd − ird ϕsq (8)
Ls

Where p is the number of generator pole pairs.

The stator and rotor powers are expressed as follows [2, 8]:

Ps = Vsd .isd + Vsq .isq (9)

Backstepping and Indirect Vector Control for Rotor Side Converter ... 1715

Qs = Vsd .isd − Vsd .isq (10)

3 Maximum Power Point Tracking

During the normal functioning of the wind turbine, the maximum power control
method is developed to exploit the energy available in the wind as much as possible.
The MPT method with mechanical speed control is established. This technique
consists of maintaining the generator speed at its reference, which is maximized
when the Cp is optimal. The electro-magnetic torque developed by DFIG is equal to
its reference value imposed by the control defined as:

Tem = Tem−opt (11)

The optimal electro-magnetic torque Tem-opt for obtaining a rotation speed equal
to the optimal speed is given as follows:

1
Tem−opt = Kpmppt + Kimppt . .[mec−opt − mec ] (12)
S

Where, KPmppt and Kimppt are the PI controller gains. The optimal speed (mec-opt )
is:
Vw .λopt
mec−opt = GB .Tu−opt ; With Tu−opt = (13)
R
Determination of the PI Gains for MPPT. The PI controller parameters are determined
by the pole compensation method. The time constant of the system (Tsys ) is:

Jtot
Tsys = (14)
f
The gains of the controller are expressed as:

1 −K imppt .Jtot Tsys

Kimppt = And Kpmppt = ; with τ = (15)
τ.f f 1000
1716 E. Chetouani et al.

4 Rotor Side Converter (RSC) Control Strategies

4.1 Powers Control by Using Indirect Field-Oriented Method

To command independently the powers of generator and to deal with the coupling
problem of the system, the indirect vector control is applied. In this paper, the powers
are controlled in the open loop and the rotor currents are controller in the closed-
loop. Stator flux is considered constant and is oriented according to d-axis. The stator
resistance is neglected and the stator voltage equation can be simplified as [2, 8]:

ϕsd = ϕs , ϕsq = 0; Vsd = 0 Vsq = Vs = ωs .ϕs (16)

Equations of rotor voltages are [2, 8]:

    
M2 M2
Vrd = Rr + Lr − .s ird − gωs Lr − irq (17)
Ls Ls
    
M2 M2 Vs M
Vrq = Rr + Lr − .s irq + gωs Lr − ird + g (18)
Ls Ls Ls

From Eqs. (17)–(18), the rotor currents expressions are deduced as follows:
     
M2 M2
ird = Vrd + gωs Lr − irq / Rr + Lr − .s (19)
Ls Ls
    
M2 Vs M M2
irq = [V rq − gωs Lr − ird − g ]/ Rr + Lr − .s (20)
Ls Ls Ls

The current references can be expressed as follows [2, 8]:

∗ Ls
irq =− .P∗ (21)
M.Vs s

∗ Ls Vs2
ird =− .(Qs∗ − ) (22)
M.Vs Ws Ls

The voltage references are expressed easily from Eqs. (17)–(18) as follows:

 1 Vs M
∗ ∗
Vrq = irq − irq .[Kp−rsc1 + Ki−rsc1 . ] + erd + g (23)
S Ls

∗
∗ 1
Vrd = ird − ird .[Kp−rsc2 + Ki−rsc2 . ] + erq (24)
S
Backstepping and Indirect Vector Control for Rotor Side Converter ... 1717

Where:
   
M2 M2
erd = gωs Lr − .ird And erq = gωs Lr − .irq (25)
Ls Ls

Determination of the Gains of PI Controller. To set the PI parameters (Kprsc ,Kirsc ),

the pole compensation method is utilized. The time constant of the system is:

M2
Ts = (Lr − )/Rr (26)
Ls

The equations of PI parameters (Kprsc ,Kirsc ) are given as follows:

1 M2 Kprsc .Rr Ts
Kprsc = .(Lr − ) And Kirsc = 2 With Trsc = (27)
Trsc Ls (Lr − Ls )
M 100

4.2 Powers Control by Backstepping Technique (BS)

The RSC is designed by Backstepping procedure to pilot the powers. For this task,
virtual control should be designed based on the rotor currents. The system stability
and performance are attained by utilizing a Lyapunov theorem, which is exploited to
establish the control of the system. Considering, the Eqs. (19)–(20), the derivation
of the rotor currents equations can be deduced as follows:
 
dird 1 M2
=   (Vrd − Rr ird − g.ωs Lr 1 − .irq ) (28)
dt Lr . 1 − M2 Lr .Ls
Lr .Ls
 
dirq 1 M2 M.Vs .g
=   (Vrq − Rr irq − g.ωs Lr 1 − .irq − ) (29)
dt Lr . 1 − M2 Lr .Ls Ls
Lr .Ls

Calculating the errors between desired rotor currents and the actual one to stabilize
it by Backstepping in the first step as:
∗ ∗
εrsc1 = irq − irq And εrsc2 = ird − ird (30)

The derivative of the errors is given by:

∗
dεrsc1 dirq dirq dεrsc2 di∗ dird
= − And = rd − (31)
dt dt dt dt dt dt
1718 E. Chetouani et al.

So, by replacing the time derivative of currents in Eq. (31), we obtain:

∗  
dεrsc1 dirq 1 M2 M.Vs .g
= −   (Vrq − Rr .irq − g.ωs Lr 1 − .irq − )
dt dt Lr . 1 − M2 Lr .Ls Ls
Lr .Ls
(32)
 
dεrsc2 di∗ 1 M2
= rd −   (Vrd − Rr .ird + g.ωs Lr 1 − .irq ) (33)
dt dt Lr . 1 − M2 Lr .Ls
Lr .Ls

The Lyapunov candidate function is defined, in a second step, as follows [4, 9]:

ε2rsc1 ε2
Lf−rsc = + rsc2 (34)
2 2
The derivation of Lyapunov function is expressed as:

L̇f−rsc = εrsc1 .ε̇rsc1 + εrsc2 .ε̇rsc2 (35)

So, replacing the errors derivation in Eq. (35), the Lyapunov derivative rewrites
as:
⎡   ⎤
∗
⎣
dirq 1 M2 M.Vs .g ⎦
L̇f−rsc =εrsc1 . −   (Vrq − Rr irq − g.ωs Lr 1 − .irq − )
dt 2
Lr . 1 − LM.L Lr .Ls Ls
r s
⎡ ⎤
 
∗ 2
⎢ di 1 M ⎥
+ εrsc2 .⎣ rd −   (Vrd − R ird + g.ωs Lr 1 − .irq )⎦ (36)
dt L . 1− M
2 Lr .Ls
r Lr .Ls r

To ensure the system stability, according to Lyapunov function, the derivation of

Lf−rsc must be negative and is given as follows [7, 9]:

L̇f−rsc = −K1 .ε2rsc1 − K2 .ε2rsc2 (37)

Where, K1 and K2 are positive constants. By equalizing the Eq. (36) and the
Eq. (37), the virtual commands Vrd and Vrq are deduced directly as follows:
⎡  2
 ⎤
∗
dirq Rr g.ωs Lr 1 − LM.L
⎢ +   irq +   .ird ⎥
r s
 ⎢ 2 2 ⎥
⎢ M2
dt Lr . 1 − LM.L Lr . 1 − LM.L ⎥
∗ =L . 1− ⎢ r s r s ⎥
Vrq r .⎢ ⎥ (38)
Lr .Ls ⎢ M.Vs .g ⎥
⎢ Ls ⎥
⎣ +   + K .ε
1 rsc1 ⎦
2
Lr . 1 − LM.L
r s
Backstepping and Indirect Vector Control for Rotor Side Converter ... 1719

 ⎡  2
 ⎤
M 2 di ∗
R g.ωs Lr 1 − LM.L
∗ =L . 1−
Vrd .⎣ rd + 
r
 ird − 
r
 .irq + K2 .εrsc2 ⎦
s
r
Lr .Ls dt 2
Lr . 1 − LM.L
2
Lr . 1 − LM.L
r s r s
(39)

5 Simulation Results and Discussion

With help of the before-mentioned mathematical models, the wind chain is modeled
and simulated employing the parameters presented in Tables 1 and 2 under
Matlab/Simulink software. The wind profile used in this work is presented in Fig. 2.
However, Fig. 3 illustrates the produced mechanical power, which is pursuing the
evolution of the wind speed. From 0.5 (s) to 0.8 (s) the power is equal to 5 MW, which
corresponds to the rated value of the wind velocity that is 12.5 (m/s). However, Fig. 4
shows the mechanical speed computed by the MPT strategy based on PI controller.
This power will be utilized as a reference of power for driving the system. Figures 5
and 6 show the efficiency of the MPT strategy for making constant the tip speed ratio
at its best value, which is λopt = 9.19 and the coefficient of power is preserved at
its maximum (Cpmax = 0.5). So, the maximal power is achieved. As you can see in
Fig. 7, the stator active power is computed by the Backstepping (BS), in which the
positive constants are chosen by trial and error method as follows: K1 = K2 = 9000,
to satisfy convergence condition, and the Indirect Field-Oriented control based on the

Fig. 2 Wind speed profile (m/s)

Fig. 3 Extracted mechanical power by MPT strategy

1720 E. Chetouani et al.

Fig. 4 Mechanical rotation speed of the DFIG

Fig. 5 Tip Speed Ratio (TSR)

Fig. 6 Coefficient of power (Cp)

Fig. 7 Stator active power (W)

Backstepping and Indirect Vector Control for Rotor Side Converter ... 1721

PI controller parameters. It is notable, in Fig. 8, that the time response is ameliorated

from 10 (ms) to 1.4 (ms). Besides, the static error is improved. The desired reactive
power is zero (Qs * = 0) to ensure a power factor equal to 1, as shown in Fig. 9. It
is observable that the response time is also improved, and the stability is guaran-
teed when especially wind changes. Figure 10 shows the unit power factor, which
is approved by the proposed method. The robustness test is established by varying
the stator leakage inductance of 100% (Ls = 2.Ls), as shown in Fig. 11. However,
Fig. 12 shows a zoom of stator active power when the turbine operates at its rated
power. As you can see, the stability is guaranteed by the Backstepping and is lost by
using the PI controller. So, the Backstepping control gives better performance than
the vector control.

Fig. 8 Stator active power (W) -Zoom

Fig. 9 Stator reactive power (VAR)

Fig. 10 Unit power factor (UPF)

1722 E. Chetouani et al.

Fig. 11 Stator active power- robustness test (Ls = 2.Ls)

Fig. 12 Stator active power- robustness test (Ls = 2.Ls) - zoom

6 Conclusion

In this paper, the wind energy conversion system is modeled and simulated based
on variable wind speed. The Maximum Power Tracking with speed regulation is
established. Then, the control of the rotor side converter is designed by employing
the IFOC based on the classical PI controller, in which the gains are calculated
by the pole compensation method, and the Backstepping based on the Lyapunov
stability which the positive constants are chosen. The results of the two strategies
are compared and analyzed. It is observed that the Backstepping brings satisfactory
and attractive results than the indirect vector control under wind speed variation. The
robustness test shows that the proposed method ensures good stability of the system
when the inductance of the DFIG changes. The Backstepping technique improves
the stability, the time response, and ameliorates the static error than indirect vector
control based on the classical PI controller.

Appendix

(See Tables 1 and 2)

Backstepping and Indirect Vector Control for Rotor Side Converter ... 1723

Table 1 Proportional and

Kpmppt Kimppt Kprsc1, 2 Kirsc1, 2
Integral (PI) gains for IFOC
and MPPT control −7.2e + 6 51.87 0.1446 0.2376

Table 2 Set of parameters

Turbine Ray of blade R 51,583 m
used in the simulation
Coefficient of multiplier GB 47,23
Total moment of inertia Jtot 1000 kg.m2
DFIG DFIG rated power Ps 5 MW
Stator inductance Ls 1,2721 mH
Rotor resistance Rr 1,446 m
Rotor inductance Lr 1,1194 mH
Mutual inductance M 0,55182 mH

References

1. Noussi K, Abouloifa A, Katir H, Lachkar I (2019) Modeling and control of a wind turbine
based on a doubly fed induction generator. In: 2019 4th world conference on complex systems
(WCCS), Ouarzazate, Morocco, pp 1–5. https://doi.org/10.1109/ICoCS.2019.8930738
2. Lamnadi M, Trihi M, Bossoufi B, Boulezhar A (2016) Modeling and control of a doubly-
fed induction generator for wind turbine- generator systems. Int J Power Electron Drive Syst
(IJPEDS). 7(3):973–985. https://doi.org/10.11591/ijpeds.v7.i3.pp982-994
3. Bouderbala M, Bossoufi B, Lagrioui A, Taoussi M, Aroussi H, Ihedrane Y (2018) Direct and
indirect vector control of a doubly fed induction generator based in a wind energy conversion
system. Int J Electr Comput Eng (IJECE). 9(3):1531–1540. https://doi.org/10.11591/ijece.v9i3.
pp1531-1540
4. Mensou S, Essadki A, Minka I, Nasser T, Bououlid Idrissi B (2018) Backstepping controller
for a variable wind speed energy conversion system based on a DFIG. Int J Electr Comput Eng
12(9)
5. Omar Baba A, Liu G, Chen X (2020) Classification and evaluation review of maximum
power point tracking methods. Sustainable Futures 2:100020. https://doi.org/10.1016/j.sftr.2020.
100020
6. Errami Y, Ouassaid M, Cherkaoui M, Maaroufi M (2015) Maximum power point tracking control
based on a nonlinear backstepping approach for a permanent magnet synchronous generator wind
energy conversion system connected to a utility grid. Energy Technol 3:743–757. https://doi.
org/10.1002/ente.201500026
7. Errami Y, Obbadi A, Sahnoun S, Benhmida M, Ouassaid M, Maaroufi M (2016) Design of a
nonlinear backstepping control strategy of grid interconnected wind power system based PMSG.
In: AIP conference proceedings, vol 1758, p 030053. https://doi.org/10.1063/1.4959449
8. Bouderbala M, Bossoufi B, Lagrioui A, Taoussi M, Alami Aroussi H, Ihedrane Y (2018) Direct
and indirect vector control of a doubly-fed induction generator based in a wind energy conversion
system. Int J Electr Comput Eng (IJECE) 9(3):1531–1540. https://doi.org/10.11591/ijece.v9i3.
pp1531-1540
9. Nadour M, Essadki A, Nasser T (2017) Comparative analysis between PI & backstepping control
strategies of DFIG driven by wind turbine. Int J Renew Energy Res 7(3):1307–1316
Integral Backstepping Power Control
of DFIG Based Nonlinear Modeling
Using Voltage Oriented Control

Mourad Loucif, Abdelkader Mechernene, and Badre Bossoufi

Abstract In this paper, the control of the wind system, the mechanical part consisting
of the wind turbine with maximum power point tracking control, and the electrical
part using a double fed induction generator is studied, taking into consideration the
full nonlinear model of the DFIG with voltage oriented control. The control strategy
is applied to the rotor side converter with an integral backstepping controller of the
active and reactive stator powers operating at sub-synchronous speed. The simulation
results show a good capability to monitor the approach proposed at different variable
wind speeds.

Keywords Wind turbine · MPPT · DFIG · Voltage oriented control · Integral

backstepping

1 Introduction

The wind has many advantages and has been used by man for many thousands of
years for both sailing and milling. Today, it is one of the main uses of wind energy,
because of its free availability, inexhaustible and its clean energy [1, 2].
DFIG is broadly applied in the Wind Energy Conversion System (WECS) due to
its mature technology and high power production efficiency. With the advantages of
long distance and large transmission capacity, series compensated transmission line
is often used to transfer large capacity wind energy to remote load center.
Backstepping control uses the Lyapunov function to ensure convergence of the
error of position tracking from all available initial conditions [3, 4]. An integral added
is integrated in the reverse design. The error integral is added to the first function
of stabilization in the backstepping approach design to remove the steady state error
and increase the rejection of disturbances [5].

M. Loucif (B) · A. Mechernene

Laboratoire d’Automatique de Tlemcen (LAT), Université de Tlemcen, Chetouane, Algeria
B. Bossoufi
Laboratory of Engineering Modeling and Systems Analysis, SMBA University, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1725
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_156
1726 M. Loucif et al.

The structure of the paper is as follows. Section 1 discusses the modeling and
control of the wind turbine to get the maximum power point tracking with linear
controller PI. Section 2 describes a complete model of the double fed induction
generator within the framework of synchronous rotation dq and a new modeling
approach that uses active and reactive power as states. We adopt the voltage oriented
approach for DFIG to propose a simplified active and reactive stator power control.
The integral backstepping design is discussed in Sect. 3. Section 4 and 5 provide the
numerical results and concluding remarks, respectively.

2 MPPT Control of Wind Turbine

The power and the torque delivered by the wind turbine are defined by [6, 7]:
⎧
⎪
⎪ 1
⎨ Paer = C p (λ, β).ρ.π.R .V
2 3
2
(1)
⎪
⎪ ρ 1
⎩ Taer = C p (λ, β). .π.R 2 .V 3
2 tur b

With ρ is the air density, V is the wind speed; R is the rotor radius.
The expression of power coefficient Cp(λ, β) who is a function of pitch angle β
and tip speed ratio λ, is written as shown below [8]:
 
π.(λ + 0.1)
C p = 0, 5176 − 116.(β − 0, 4). sin − 5.(λ − 21).(β − 0, 4)
14.8 − 0.3.(β − 2)
(2)
tur b .R
λ= (3)
V
The expression of the gearbox G is given by:

Ttur b 
G= = (4)
Tg tur b

The mechanical equation is given by [18]:

d
J. = Tg − Tem − B. (5)
dt
The power coefficient expression (2) has been shown to have a unique maximum,
corresponding to the optimal choice of the tip speed ratio λopt = 9,15 and the power
coefficient take a maximum value of C pmax = 0,5 for a pitch angle is fixed β = 2°
[9].
Integral Backstepping Power Control of DFIG Based Nonlinear Modeling ... 1727

To get the maximum power, we use the MPPT control with wind speed measure-
ment developed in paper [6]. For the optimal tip speed ratio λopt , (3) and (4) provide
the optimal rotational speed generator as

λopt .V
∗ = (6)
R

3 Voltage Oriented Control of DFIG

The mathematical model of DFIG has been obtained after the according to the
voltage equations and equations of flux, expressed in a dq reference frame rotating
at synchronous speed, the differential equations of the stator and rotor circuits of the
generator with stator and rotor current as state variables can be obtained by [10, 13].
⎧
⎪
⎪
di sd
= −a1 .i sd + (a.ω + ωs ).i sq + a3 .ir d + a5 .ω.irq + b1 .Vsd − b3 .Vr d
⎪
⎪
⎪
⎪ dt
⎪
⎪
⎪
⎪ di sq
⎨ = −(a.ω + ωs ).i sd − a1 .i sq − a5 .ω.ir d + a3 .irq + b1 .Vsq − b3 .Vrq
dt
⎪  ω
(7)
⎪ dir d
⎪ = a4 .i sd − a5 .ω.i sq − a2 .ir d + ωs − .irq − b3 .Vsd + b2 .Vr d
⎪
⎪ σ
⎪
⎪ dt
⎪
⎪  ω
⎪ di
⎩ rq = a6 .ω.i sd + a4 .i sq − ωs − .ir d − a2 .irq − b3 .Vsq + b2 .Vrq
dt σ
M2 Rr .Msr Rs .Msr
Where, σ = 1 − L s .Lsr r , a = 1−σ σ
, a1 = σ.LRs
s
, a2 = σ.L Rr
r
, a3 = σ.L s .L r
, a4 = σ.L s .L r
,
a5 = σ.L s , a6 = σ.L r , b1 = σ.L s , b2 = σ.L r , b3 = σ.L s .L r .
Msr Msr 1 1 Msr

Stator and rotor resistances (Rs and Rr ). ω = p. electrical speed and p is
the number of poles pair. Stator and rotor inductances (L s L r ) and the magnetizing
inductance (M sr ). The direct and quadrature stator and rotor currents and voltage are
given as follow (isd , isq , ird , irq and V sd , V sq , V rd , V rq ).
The electromagnetic torque can be described by fundamental torque Eq. (8):
⎧
⎪
⎪ Msr

⎨ Te = p ϕsq .ir d − ϕsd .irq
Ls
(8)
⎪
⎪ d 1
⎩ = (Te − Tl − f.)
dt J

The active and reactive powers of the stator are given as described below [18]:

Ps = Vsd .i sd + Vsq .i sq
(9)
Q s = Vsq .i sd − Vsd .i sq
1728 M. Loucif et al.

Derivatives active and reactive powers of the stator:

⎧ di sd di sq
⎪
⎨ Ṗs = Vsd . + Vsq .
dt dt (10)
⎪
⎩ Q̇ = V . di sd di sq
s sq − Vsd .
dt dt
Substituting (7) in (10), the derivatives stator powers can then be expressed as:
⎧


⎪
⎪ Ṗs = −a1 .Ps − (ωs + a.ω)Q s + a5 .ω.Vsd + a3 .Vsq .irq + b1 . Vsd2 + Vsq2
⎪
⎨ +
a .V − a .ω.V i − b
V .V + V .V 
⎪
3 sd 5 sq r d 3 sd r d sq rq



⎪
⎪ s
Q̇ = (ω + a.ω)P − a .Q + b V .V − Vsq .Vr d + a5 .ω.Vsd + a3 .Vsq ir d
⎪
⎪
s s 1 s 3 sd rq
⎩

+ a5 .ω.Vsq − a3 .Vsd irq
(11)

The idea of the proposed control is a voltage oriented control (VOC), in which
the control reference is chosen to align with the q-axis voltage grid component [11,
12]:

Vsd = 0, Vsq = Vs (cst) (12)

The expressions of the derivatives stator powers (10) are easily simplified as:

Ṗs = −a1 .Ps − (ωs + a.ω)Q s + a3 .Vs .irq + b1 .Vs2 − a5 .ω.Vs .ir d − b3 .Vs .Vrq
Q̇ s = (ωs + a.ω)Ps − a1 .Q s − b3 .Vs .Vr d + a3 .Vs .ir d + a5 .ω.Vs .irq
(13)

4 Integral Backstepping Control of Rotor Side Converter

Backstepping is a nonlinear control scheme which is based on Lyapunov’s stability

theorem. The technique may be applied to synthesize a command function in order
to force a system to follow a desired trajectory [16, 17]. The control law is designed
by iteratively selecting various state variables appropriate as virtual inputs for the
smaller subsystems of the overall system. Lyapunov functions are then designed
for each stable virtual controller. As a result, the final real control law guarantees
the asymptotic balance of the total control system [14]. The additional integrator
improves the robustness of the system against model uncertainties and outside distur-
bances. A more detailed discussion of the backstepping control approach can be seen
in [3, 15].
The active and reactive power tracking errors e1 and e2 are defined respectively
by:
Integral Backstepping Power Control of DFIG Based Nonlinear Modeling ... 1729

e1 = Psr e f − Ps
(14)
e2 = Q rs e f − Q s

ref ref
Where Ps and Q s are the expected active and reactive powers benchmarks.
The derivative of the equation error variables are derived as:

ė1 = Ṗsr e f − Ṗs
(15)
ė2 = Q̇ rs e f − Q̇ s
⎧
⎨ ė = Ṗ r e f + a .Ps + (ωs + a.ω)Q s − b .V 2 − a .Vs .irq + b Vs .Vrq + a .ω.Vs .i
1 s 1 1 s 3 3 5 rd
⎩ ė = Q̇ r e f − (ω + a.ω)P + a .Q − a .V .i + b .V .V − a .ω.V .i
2 s s s 1 s 5 s rd 3 s rd 5 s rq
(16)

Let us consider:
⎧ ⎧
⎨ ė01 = e1 ⎨ ė02 = e2
t t
, (17)
⎩ e01 = λ e1 (τ )dτ ⎩ e02 = λ e2 (τ )dτ
0 0

Select a Lyapunov function candidate:

⎧ 1 1 2
⎪
⎨ V (e1 ) = .e12 + .e
2 2 01 (18)
⎪
⎩ V (e ) = 1 .e2 + 1 2
2 .e
2 2 2 02
The derivative of Lyapunov function are variables gives:
 
V̇ (e1 ) = e1 .ė1 + λe01 .e1 V (e1 ) = e1 .(ė1 + λe01 )
(19)
V̇ (e2 ) = e2 .ė2 + λe02 .e2 V̇ (e2 ) = e2 .(ė2 + λe02 )
⎧ 
⎪ re f
⎨ V̇ (e1 ) = e1 . Ṗs + a1 .Ps + (ωs + a.ω)Q s − b1 .Vs2 + b3 .Vs .Vrq + a5 .ω.Vs .ird − a3 .Vs .irq + λe01

⎪
⎩ V̇ (e2 ) = e2 . Q̇ re f
s − (ωs + a.ω)Ps + a1 .Q s + b3 .Vs .Vrd − a5 .Vs .ird − a5 .ω.Vs irq + λe02
(20)

The control algorithm proposed which is based on integral backstepping has been
introduced:
⎧
⎪ 1  re f

⎪
⎨ Vrq = b .V − Ṗs − a1 .Ps − (ωs + a.ω)Q s + b1 .Vs − a5 .ω.Vs .ird + a3 .Vs .irq − k1 .e1 − λ.e01
2
3 s
⎪
⎪ 1 
⎩ Vrd = re f
− Q̇ s + (ωs + a.ω)Ps − a1 .Q s + a5 .Vs .ird + a5 .ω.Vs .irq − k2 .e2 − λ.e02
b3 .Vs
(21)
1730 M. Loucif et al.

Fig. 1 Diagramme of simulation integral backstepping associate with mechanical part

With λ, k1 and k2 are positives constants.


V̇ (e1 ) = −k1 .e12
(22)
V̇ (e2 ) = −k2 .e22

The diagram of simulation nonlinear integral backstepping powers control of

DFIG associate with MPPT control of wind turbine is shown in the Fig. 1.

5 Interpretation and Results

The purpose of validating the comportment of the association of the mechanical part
with wind turbine, to extract the MPPT we use linear PI controller and electrical part
with DFIG, nonlinear integral backstepping controller is being applied to control the
active and reactive stator power. The wind turbine and DFIG parameters are provided
in the appendix.
The wind speed profile will be simulated in deterministic representation by a sum
of many harmonics is defined by the relation:

V = 7 + 0, 2 sin(0, 1047t) + 2 sin(0, 2665t) + sin(1, 293t) + 0, 2 sin(3, 6645t)

(23)

Figure 2 show the wind speed model used in this validation simulation chosen for
an average speed of 7 m/s, operate into sub-synchronous mode.
It is observed from Fig. 3, that the controller proportional and integral for
maximum power point closely tracks the generator speed as well the profile of the
wind proposed in steady state, and a slight overshoot in transient state. The operation
is always sub-synchronous mode; the slip is always positive in all profile of wind
Integral Backstepping Power Control of DFIG Based Nonlinear Modeling ... 1731

speed propose. Figure 4 the active power of the stator has a negative sign, signi-
fying that the generator produces energy and transmits it to the grid, are correctly
follows almost perfectly its reference generated by the turbine. In order to maintain
a unitary power factor on the grid side, the stator reactive power reference value will
be fixed at zero in order to ensure that the quality of the power delivered to the grid
is optimized, the power produced will track almost their reference after the end of its
transient regime. We can observe that the integral backstepping controller ensures
a exactly decoupling between them. Figure 5 the stator currents in the three phases
network generated by DFIG are sinusoidal shape and which has a excellent power
quality for the electrical grid system of frequency 50 Hz, it is very clear in the zoom
of stator currents. The rotor currents in the three phases are also sinusoidal, present
a good tracking and the frequency of the current varies with the slip shown in Fig. 6.
The electromagnetic torque variation gives the identical profile of the stator active
power.

9
Wind speed (m/s)

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s)

Fig. 2 Variable wind speed profile

170 0.5
157
140 0.4
Generator speed (rad/s)

120
100 0.3
slip

0.2
50
*
0.1
W
W
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s) Time (s)

Fig. 3 Generator speed with MPPT and the slip

1732 M. Loucif et al.

4
* x 10
4000 Ps 1
*
Qs

Reactive power (Var)

2000 Ps
Active power (W)
Q
0.5 s
0

-2000 0

-4000
-0.5
-6000

-8000 -1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s) Time (s)

Fig. 4 Generator active and reactive powers

20 10
Stator current zoom (A)
Stator current (A)

10 5

0 0

-10 -5

-20 -10
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 3 3.02 3.04 3.06 3.08 3.1 3.12 3
Time (s) Time (s)

Fig. 5 Stator currents with zoom

40 20
Torque Tem (Nm)
Rotor current (A)

20 0

0 -20

-20 -40

-40 -60
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s) Time (s)

Fig. 6 Rotor currents and electromagnetic torque

Integral Backstepping Power Control of DFIG Based Nonlinear Modeling ... 1733

6 Conclusion

In this paper, a modeling, control and simulation of the mechanical part of a wind
turbine using a PI controller to extract the MPPT for the variable speed operating
in region two is presented. It also presents the stator active and reactive powers
control of a double-fed induction generator by considering the complete non-linear
model in the dq reference frame, using a stator voltage oriented control to simplify
the derivatives of the stator powers. Backstepping controller with integral action
applied to the rotor side converter. The slip is always positive; DFIG operates in
sub synchronous speed mode. The results of simulation are presented and discussed
confirming its feasibility of the proposed control scheme had a good follow speed
and stator powers response, regardless of variation wind speed profile.

Appendix

(See Tables 1 and 2).

Table 1 Wind parameters

Parameters Value Unit
Nominal power Pn 10 Kw
Air density ρ 1,22 Kg/m3
Rotor radius R 3 m
Turbine inertia Jt 0,02 Kg.m2
Turbine friction coefficient Bt 0,0016 Nm/rad/s
Number of blade 3 –
Gearbox ratio G 5,4 –

Table 2 DFIG parameters

Parameters Value Unit
Nominal power Pn 7, 5 Kw
Nominal speed Nn 1440 Rpm
Number of poles pair p 2 –
Stator resistance Rs 0.455 
Rotor resistance Rr 0.62 
Stator inductance Ls 0,084 mH
Rotor inductance Lr 0,081 mH
Stator-rotor mutual inductance Msr 2,691.105 mH
Moment of inertia J 0,3125 N.m.s2
Coefficient of friction B 6.73e−3 kg.m/s
1734 M. Loucif et al.

References

1. Abolhassani MT, Enjeti P, Toliyat M (2008) Integrated doubly fed electric alternator/active
filter (IDEA), a viable power quality solution for wind energy conversion systems. IEEE Trans
Energy Convers 23(2):642–650
2. Loucif M, Boumediene A, Mechernene A (2014) Robust nonlinear combined backstepping
sliding mode control of DFIG under wind speed variation. In: The second international
conference on electrical engineering and control applications
3. Krstic M, Kanellakopoulos I, Kokotovic PV (1995) Nonlinear and adaptive control design.
Wiley, New York
4. Krstic M, Kanellakopoulos I, Kokotovic PV (1994) Nonlinear design of adaptive controllers
for linear systems. IEEE Trans Autom Control 39:738–752
5. Yaolong T, Jie C, Hualin T, Jun H (2000) Integral backstepping control and experimental imple-
mentation for motion system. In: Proceedings of the IEEE international conference control
applications
6. Loucif M, Boumediene A, Mechernene A (2014) Maximum power point tracking based
on backstepping control of wind turbine. J Electrotehnica Electronica Automatica (EEA).
62(3):103–109
7. Ayrir W, Ourahou M, Haddi A (2018) DFIG stator reactive and active power control based
fuzzy logic. Int J Circ Syst Signal Process 12:262–267
8. El Aimani S (2004) Modélisation de différentes technologies d’éoliennes intégrées dans un
réseau de moyenne tension. Doctorat thesis, Lille, France
9. Loucif M, Boumediene A, Mechernene A (2015) Nonlinear sliding mode power control of
DFIG under wind speed variation and grid connexion. J Electrotehnica Electronica Automatica
(EEA) 63(3):103–110
10. Lihui Y, Xikui M, Dong D (2009) Hopf bifurcation in doubly fed induction generator under
vector control. Chaos Solitons Fractals 41:2741–2749
11. Bourdoulis M, Alexandridis A (2013) A new controller design and analysis of DFIG wind
turbine systems for mpp operation. In: EuroCon. Zagreb, Croatia
12. Bourdoulis M, Alexandridis A (2013) Nonlinear stability analysis of DFIG wind genera-
tors in voltage oriented control operation. In: European Control Conference (ECC), Zürich,
Switzerland
13. Loucif M (2016) Synthèse de lois de commande non-linéaires pour le contrôle d’une machine
asynchrone à double alimentation dédiée à un système aérogénérateur. Doctorat thesis,
Tlemcen, Algerie
14. Laoufi A, Hazzab A, Bousserhane IK, Rahli M (2006) Direct field oriented control using
backstepping technique for induction motor speed control. Int J Appl Eng Res 1(1):37–50
15. Haruna A, Mohamed Z, Basri MAM, Ramli L, Alhassan A (2017) 2-step integral backstepping
control of the two-rotor aero-dynamical system (TRAS). J Fund Appl Sci 9:395–407
16. El Mourabit Y, Derouich A, El Ghzizal A, El Ouanjli N, Zamzoum O (2020) Nonlinear back-
stepping control for PMSG wind turbine used on the real wind profile of the Dakhla-Morocco
city. Int Trans Electr Energ Syst 30:e12297
17. Loucif M, Boumediene A, Mechernene A (2013) Backstepping control of double fed induction
generator driven by wind turbine. In: 3rd international conference on systems and control
18. Loucif M, Boumediene A, Mechernene A (2015) Modeling and direct power control for a
DFIG under wind speed variation. In: 3rd international conference on control, engineering and
information technology
Smart Monitoring PID of Solenoid
Response Based on Fiber Squeezer

Said Amrane, Abdallah Zahidi, Nawfel Azami, Mostafa Abouricha,

Naoual Nasser, and Mohamed Errai

Abstract Connecting of integrated fiber systems depends on the polarization of the

light at the output of polarization controllers based on solenoid, as squeezer fiber
is suitable for coherent fiber optic systems because of its low insertion loss as well
as low power penalty. However, the stability in open loop of the solenoid caused
by their magnetic circuit affects the performance of these polarization controllers
and limits their applications. To improve stability performance, first; a feedback PID
corrector is proposed to correct the dynamic response. Then, the authors propose
to study the effect of electrical parameters variations of the PID gains of solenoid.
Finally, the artificial neural network is using to predict the PID gains that will correct
the solenoid dynamic response. The simulation results show the need to adjust the
PID corrector in the event of system parameter changes to maintain the optimum
dynamic response.

Keywords Solenoid · Fiber squeezer · PID corrector · ANN

S. Amrane (B) · M. Errai

Polydisciplinary Faculty of Taroudant, University Ibn Zohr, Agadir, Morocco
e-mail: [email protected]
M. Errai
e-mail: [email protected]
A. Zahidi · N. Azami
National Institute of Posts and Telecommunications, Rabat, Morocco
e-mail: [email protected]
M. Abouricha
Faculty of Science, University Ibn Zohr, Agadir, Morocco
e-mail: [email protected]
N. Nasser
Faculty of Sciences and Techniques, University Hassan 1st, Settat, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1735
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_157
1736 S. Amrane et al.

1 Introduction

Solenoids are used in several technical applications such as robotics [1], valves
[2, 3], positioners [4], anti-braking systems [5]. They are also used in polarization
controllers to adjust the light power at the output of an optical fiber by applying
mechanical stress to it [6]. This type of actuator is characterized by a stable amplitude
of movement, which does not extend over its entire range [7]. Several studies are
devoted to extending the stability interval of this type of devices such as the nonlinear
magnetic field mapping [8], adaptive control [9], dual solenoid configuration [10],
sliding mode control [11], pulse with modulation control [12] and PID corrector that
is easy to implement [13]. However, it is useful to readjust the PID corrector in case
of the physical parameters variation of the solenoid, which affects the performance
of its response. This variation is considered to be a major problem for this type of
actuator [14–18]. The effect of fluctuations in mechanical parameters is studied in
[19].
To control the solenoid response intelligently, we propose to use an optical fiber as
a force sensor between the frame and the plunger of the solenoid. This force produces
a birefringence on the fiber, which changes the polarization state of light through the
fiber [6]. The change in the polarization state of the light results in the variation of
the light power measured by a photodiode.
First, we propose to correct the light power dynamic response at the output of the
fiber by adjusting the gains of the PID corrector. In the second step, we studied the
effect of electrical parameters variation of the solenoid the gains of the PID corrector.
The electrical parameters including resistance and inductance of solenoid. Finally,
the ANN is used to predict the values of the gains of the PID corrector to allow
maintaining the stability of the system.

2 The Methodology of the Study

2.1 Structure of Solenoid with Fiber Squeezer

The solenoid is the electromagnetic actuator that exerts pressure on the fiber. Its
structure is shown in Fig. 1

2.2 Solenoid Mathematical Modeling

The mathematical model of the solenoid is described in detail in [20]. The mechanical
part of the system can be modeled by Eq. 1:
Smart Monitoring PID of Solenoid Response Based on Fiber Squeezer 1737

Fig. 1 Scheme of the

solenoid with fiber squeezer

 
d2 x 1 dx
= Fem − k(x − x0 ) − B (1)
dt 2 m dt

Where x is the displacement of the armature in (m), m is the mass of the armature
in (Kg), Fem is the force produced by the magnetic field in (N), K is the stiffness of
the spring in (N/m), x0 is the initial air gap between the armature and the backside of
the frame in (m), and B is the system damping coefficient in (N.s/m). The electrical
part is modeled by Eq. 2:
 
di 1 dL(x) dx
= u − Ri − i (2)
dt L(x) dx dt

Where i(t) is the coil current in (A), L is the coil inductance in (H) that depends
on the air gap [7], and R is the coil resistance in (). Equation 3 and 4 models the
optical part.

F
δ = 6.10−5 (3)
λd
Pi
Po = (1 + cos δ) (4)
2
Where δ in (rad) is the magnitude of the phase difference between the polarized
light along the squeezing axis and it is orthogonal [21], F is squeezing force in (N),
λ is the light wavelength used in (m), d is the fiber diameter in (m), Po is the output
light intensity, and Pi is t the input light intensity in (W).
1738 S. Amrane et al.

Fig. 2 Simulink model of

the solenoid with fiber
squeezer

2.3 System Modeling in Simulink Environment

Taking into account Eqs. 1 and 2 that respectively models the mechanical block and
the electrical block; Eqs. 3 and 4 that models the optical block, the model Simulink
is generated to simulate the solenoid as fiber squeezer (Fig. 2).

3 Simulation and Result

3.1 The Proposed Method of Simulation

First, we use the Simulink model to generate the open-loop step response of the
simulated system. After identification from the step response, the system is looped
and corrected using a PID controller. Finally, we study the effect of fluctuation of coil
resistance and coil inductance on PID gains (Kp, Ki and Kd). The parameter variation
ranges are the resistance R for 5 to 50  and inductance L for 10 to 35 mH. The
identification of the system is done using the “tfest” function, of the Matlab/Simulink
environment, to estimate the system transfer function [22]. This transfer function Tf
has the coefficients (num, den2, den3), and can be expressed by the following form:
num
Tf = (5)
s2 + den2 s + den3
Smart Monitoring PID of Solenoid Response Based on Fiber Squeezer 1739

After estimating the transfer function of the system, the gains of the PID corrector
are determined using the “pidtune” function [23].

3.2 Open-Loop System Step Response of Solenoid with Fiber

Squeezer

The Simulink model is running using the values of the solenoid with fiber squeezer
represented in the following Table 1:
Figure 3 represents the curve of variation of the light power at the output of the
fiber as a function of time when a voltage step is applied to the input of the open-loop
system (open-loop step response). The curve shows a stable response with damped
oscillations. The overshoot is relatively large (around 80%).

Table 1 The values of solenoid parameters with fiber squeezer

R () K (N/m) B (Ns/m) M (g) L (mH) Pi (mW)
10 2000 2 200 20 1

Fig. 3 Open-loop step response of solenoid with fiber squeezer

1740 S. Amrane et al.

Table 2 The value of the PID gains used to correct the system response
Kp Ki Kd
2.9 757.5 39193

Fig. 4 Close-loop system step response of solenoid with fiber squeezer

3.3 Close-Loop System Step Response of Solenoid with Fiber

Squeezer

Using the pidtune function, the values of the PID gains obtained are shown in the
following Table 2:
Figure 4 represents the curve of variation of the light power at the output of the
fiber as a function of time when a voltage step is applied to the input of the closed-loop
system (close-loop step response). The curve shows a stable response. The overshoot
improved has decreased significantly (from 80% to 13%.)

3.4 Influence of Coil Electrical Resistance and Coil

Inductance on PID Gains

The effect of the coil resistance and inductance variation the PID gains is shown in
the following maps:
Smart Monitoring PID of Solenoid Response Based on Fiber Squeezer 1741

Fig. 5 Map of Kp = f(R, L)

Fig. 6 Map of Ki = f(R, L)

1742 S. Amrane et al.

Fig. 7 Map of Kd = f(R, L)

3.5 The Monitoring Process of PID Gains

The results of the simulation are stored in a database. Each record consists of:
– Physical parameters of solenoid: the mass m, the resistance of coil R, inductance
L, spring stiffness K, and the damping coefficient B.
– The coefficients of the transfer function: num, den2, and den3.
– The PID gains: Kp, Ki, and Kd.
The generated database is used to train the artificial neural network. The coefficients
of the transfer function represent the inputs of the neural network model while the
gains of the PID corrector represent these outputs. To predict the gains of the PID
corrector, we estimate the transfer function coefficients of the open-loop system by
using the tfest function; the parameters of this function (num, den2, and den3) are
introduced into the ANN to generate the PID gains.

4 Interpretation of Results

Figures 5, 6, and 7 represent the maps of variations of the PID corrector gains as a
function of the resistance R of the coil and the inductance L of the solenoid. These
figures show that the three gains have the same monotony when R and L vary. These
Smart Monitoring PID of Solenoid Response Based on Fiber Squeezer 1743

gains increase with R and decrease with L. We also notice that varying the coil
resistance and the coil inductance affects the solenoid response and therefore the
intensity of the output power. Consequently, it is necessary to readjust the PID gains
in the case of the resistance and inductance variation in order to maintain the response
with sufficient stability.
Using the ANN is useful to predict the gains of the PID corrector in the event of
any fluctuations in the coil resistance or the coil inductance of the solenoid. However,
it is still necessary to study and improve the performances of the ANN and its training
to minimize the error on the prediction of the PID gains.

5 Conclusion

The force applied on the optical fiber by the armature of the solenoid varies the
light power at its output. Mathematical modeling makes it possible to describe the
dynamic response of this system. PID controllers provide powerful control tools for
the transient response. The overshoot is greatly reduced from 80% to 13%. This
study also highlighted the influence of the variations of the coil resistance and coil
inductance on the solenoid response with fiber squeezer. Hence, is necessary to
readjust the PID gains in the case of the coil resistance and inductance variations.
Therefore, to keep the solenoid dynamic response in an optimal state, it suffices
to analyze its response to determine the changes that the solenoid parameters have
undergone in order to readjust the PID gains again. The use of an ANN can be used
for monitoring the PID gains to keep the solenoid dynamic response with sufficient
stability.

References

1. Takai A, et al (2013) Prototyping the flexible solenoid-coil artificial muscle, for exoskeletal
robots. In: ICCAS 2013, pp 1046–1051
2. Kajima T, Kawamura Y (1995) Development of a high-speed solenoid valve: investigation of
solenoids. IEEE Trans Ind Electron 42(1):1–8. https://doi.org/10.1109/41.345838
3. Mutschler K et al (2014) Multi physics network simulation of a solenoid dispensing valve.
Mechatronics 24(3):209–221
4. Chen M-Y et al (2010) A new design of a submicropositioner utilizing electromagnetic actuators
and flexure mechanism. IEEE Trans Ind Electron 57(1):96–106
5. Chu L, et al (2007) Study on the dynamic characteristics of pneumatic ABS solenoid valve for
commercial vehicle. In: 2007 IEEE vehicle power and propulsion conference, Arlington, TX,
USA, pp 641–644
6. Rumbaugh SH et al (1990) Polarization control for coherent fiber-optic systems using nematic
liquid crystals. J. Lightwave Technol 8(3):459–465
7. Li C et al (2016) Enhanced-performance control of an electromagnetic solenoid system using
a digital controller. IEEE Trans Contr Syst Technol 24(5):1805–1811
1744 S. Amrane et al.

8. Lim KW, et al (1994) Proportional control of a solenoid actuator. In: Proceedings of IECON
1994 - 20th annual conference of IEEE industrial electronics, Bologna, Italy, vol 3, pp 2045–
2050
9. Nataraj C, Lee D (2010) Model-based adaptive control for a solenoid-valve system. In: Volume
8: dynamic systems and control, parts A and B, Vancouver, British Columbia, Canada, pp
1009–1015
10. Yu L, Chang TN (2010) Zero vibration on–off position control of dual solenoid actuator. IEEE
Trans Ind Electron 57(7):2519–2526
11. Nguyen T et al (2007) Accurate sliding-mode control of pneumatic systems using low-cost
solenoid valves. IEEE/ASME Trans Mechatron 12(2):216–219
12. Taghizadeh M et al (2009) Modeling and identification of a solenoid valve for PWM control
applications. Compt Rendus Mécan 337(3):131–140
13. Lunge SP, Kurode SR (2013) Proportional actuator from on off solenoid valve using sliding
modes. In: Proceedings of the 1st international and 16th national conference on machines and
mechanisms, IIT Roorkee, India
14. Steiner G et al (2004) Managing uncertainties in electromagnetic design problems with robust
optimization. IEEE Trans Magn 40(2):1094–1099
15. Kabib M et al (2016) Modelling and simulation analysis of solenoid valve for spring constant
influence to dynamic response. JEAS 11:2790
16. Dülk I, Kovácsházy T (2014) Resistance estimation in solenoid actuators by considering
different resistances in the PWM paths. Period Polytech Elec Eng Comp Sci 58(3):109–120
17. Zahidi A, Said A, Azami N, Nasser N (2020) Effect of fiber and solenoid variation parameters on
the elements of a corrector PID for electromagnetic fiber squeezer based polarization controller.
IJECE 10(3):2441
18. Abedallah Z, Said A, Nawfel A, Naoual N (2019) Study and simulation of fiber press control
for solenoid parameters monitoring for Smart preventive maintenance. In: 2019 1st ICSSD,
Rabat, Morocco, pp 1–6
19. Zahidi A, Amrane S, Azami N, Nasser N (2020) Self-tuning PID of the solenoid response
based on fiber squeezer. In: 13th SITA, Rabat, Morocco, pp 1–6
20. Ikeda K, et al (2003) Endless tracking polarization controller
21. Smith AM (1980) Single-mode fibre pressure sensitivity. Electron. Lett. 16(20):773
22. Najdek K (2019) Identification of dual-active-bridge converter transfer function. Electrotech
Rev 1(3):153–156
23. Bansal H (2009) Tuning of PID controllers using simulink. Int J Math Model Simul Appl
2(3):337–344
Fuzzy Logic Based MPPT Control
of an Isolated Hybrid PV-Wind-Battery
Energy System

Hadjer Abadlia and Riad Toufouti

Abstract Recently the generations by the traditional sources are insufficient to meet
the increasing power demands due to continuous depletion of fossil energy sources,
therefore distributed generators (DGs) based renewable energy sources (RES) to
system is necessary to overcome these issues. In this study we present a stand-alone
hybrid power generation system composed by a photovoltaic and wind power genera-
tion sources with energy system storage to achieve high performance while supplying
unbalance AC load, PV system and Win turbine is operated with a DC/DC boost
converter to extract the Maximum Power Point Tracking (MPPT) using the fuzzy
logic technique. The total power generation supply the load via an inverter controlled
by pulse width modulation technique. A Simulink model of the proposed hybrid
system with the MPPT controlled Boost converter of the PV generator and Voltage
regulated voltage source inverter for standalone system is simulated in MATLAB.

Keywords Photovoltaic generator · Wind turbine · Permanent magnet

synchronous generator (PMSG) · Maximum power point tracking · Fuzzy logic
controller (FLC) · Hybrid renewable energy system (HRES)

1 Introduction

Decentralized electricity generation is currently undergoing a very significant

development, much of the energy design is introduced by fossil energy sources.
But these conventional energies drain quickly and intensify greenhouse gas emis-
sions results the air pollution [1]. Faced with the consequences it is important to take
into account the development of alternative energy which is known as renewable
energy sources.
Several sources of renewable energy are increasingly used and developed to be
the future for power systems like photovoltaic (PV), wind, geothermal, biomass,
hydro power are most commonly utilized to meet the needs growing energy [2].

H. Abadlia (B) · R. Toufouti

Department of Electrical Engineering, Laboratory of Electrical and Renewable Energies, LEER,
Souk Ahras University, Souk Ahras, Algeria

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1745
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_158
1746 H. Abadlia and R. Toufouti

They give promising solutions in building sustainable and ecofriendly electric power
[2]. Moreover, they decrease pollution in limiting the emission of carbon dioxide
coming from the conventional methods for energy generation based on fossil fuels.
Wind and photovoltaic power generation systems can be used for power gener-
ation as isolated system, but their benefits are considerably improved when they
are integrated into electrical grid. But the intermittent nature of these two sources
according the time of day, season and year, the random nature of these sources has
led researchers to turn to the combination of them to form the hybrid energy system
(HES) which is very interest by their reliability, their incomparable flexibility and
their attractive cost [3, 4].
HES are a good solution for alimenting rural areas located in isolated regions and
rural remote areas, where the extension of the conventional power network is more
costly and online losses become significant. But the main problem of is that (RES)
must work in a complementary way to manage the response for demand of power in
any moment.
For improvement of the complexity of the RES involved in a hybrid power system
(HPS), several researches have been devoted by several researchers in worldwide.
In order to manage the energy consumption without any problems Many studies
discussed control and energy management sources of HES presented in [5, 8, 9].
Various papers has been presented for the use sizing and optimization of hybrid
renewable systems [4, 6, 7]. For optimization HRES using intelligent techniques,
many advanced optimization techniques have been presented for the optimal design
of hybrid RES especially: Particle Swarm Optimization (PSO), Genetic algorithm
(GA) [12, 16, 23].
To extract the maximum power from RES, many MPPT algorithms are Developed
in literature such as Perturb and Observe technique (P&O), [24], incr-conductance
[25], fuzzy logic [23], sliding mode [26], also the combination of fractional order
control and incremental conductance [27]. These control strategies differ among
them in: simplicity, reject disturbances, sensors required who needed the exact
parameterization, cost and efficiency [28].
In this paper has the objective of modeling and control, inside the Matlab/Simulink
environment of isolated hybrid PV/Wind/Battery, distributed power system under
changing weather conditions to supply a variable AC load and the MPPT controller
is implemented using the fuzzy logic technique to obtained the maximum power
from the PV panels and Wind turbine to ensure the availability of power generation.
The robustness of the proposed control strategy is verified by numerical simulation.

2 Description of the Proposed HES

The proposed hybrid renewables energy resources combine PV and wind energy
sources as fellows in Fig. 1:
wind generator which is used with a diode rectifier and DC/DC converter and a
PV generator includes PV panels connected to high boost converter to obtain the
Fuzzy Logic Based MPPT Control ... 1747

Fig. 1 Configuration of
proposed hybrid system

(MPPT) using fuzzy logic controller. The total energy produced from the wind and
PV sources is used to supply the AC loads through three phase inverter controlled
by pulse width modulation, the output voltage of DC/AC converter is filtered by LC
filter to given pure sine wave in the load. The IGBT invertetr is controlled by sinus
triangular modulation (SPWM).

2.1 Modeling of the HRES

The conception of the hybrid system is mainly dependent on the performance of each
source of energy, which must be first modeled individually and then combined to
evaluate the performance of the system.
• Modeling of PV generator. Solar energy is the most encouraging source of energy
among renewable energy, it is an attractive option due to zero CO2 emission, noise
less in operation and with lower maintenance cost [10]. The PV solar energy is
the direct transformation of light into electric power through solar cells which is
the essential element building the PV panel, PV cell is Generally a semiconductor
device composed of PN Junction almost identical to a diode [11], these cells are
connected in series to form a PV module, a number of PV modules are connected
in series to generate higher voltage which formed a string, and a number of strings
are connected in parallel to given the higher current and formed a PV array for
obtaining the required power [12].
1748 H. Abadlia and R. Toufouti

Fig. 2 PV model of single

diode

Like all the physical systems, the PV cell model is done with different levels
of exactitude depending on the users intended; this operation is realized through an
equivalent circuit with précising parameters and variables, that is why the field of PV
cell equivalent circuit model always knows advancement [13]. The one diode model
is the most used models [10, 13–15], also the model of three diode is developed
[16, 17].
We present in this paper the model of one diode due to their Simplicity [10, 18,
19], and the least number of parameters [20],this classic model consists of a current
source antiparallel with a parallel and series resistor connected in the terminal of the
diode the equivalent circuit is shown in Fig. 2 [21, 22].
Using a loop the voltage and node current laws, the generated current in the PV
generator is given:
    
V + I Rs V + I Rs
I = I ph − I0 ex p q −1 − (1)
nK T Rsh

Where:
Iph : The current generated by PV panel,
Io: reverse saturation diode current off the grid solar, n: the diode ideality constant.
Rs and Rsh series and shunt resistance, K: Boltzmann’s constant q: the electron
charge.
• Wind generator. Wnd turbine convert’s the mechanical energy in the wind into
electrical energy, it’s composed of turbine coupled to a generator wich transforms
the mechanical power into electrical power [29]. The generator can be a (PMSG)
or induction generator, Which are the most common generators that are used in
wind turbines [30]. Power produced by wind turbine is given as fellow: [31–33]

1
Pw = C p (λ, β)ρ AV 3 (2)
2
Where:
ρ: air density (kg/m), Cp : power coefficient, A = π × R2 : intercepting area of the
rotor blades (m), V: average wind speed (m/s), λ: tip speed ratio, β: the blade pitch
angle (in degrees).
Fuzzy Logic Based MPPT Control ... 1749

The power coefficient Cp of the turbine (also known as performance coefficient),

the power coefficient presents a fraction of power in the wind captured by the turbine
and has a theorical maximum of 0.59 [29, 31]. The tip speed ratio (TSR) for turbine
is given as fellow, [33]

Rω
λ= (3)
V
Where:
R: radius of turbine (m), ω: angular speed (rad/s).
• Battery storage and controller by bidirectional converter. The model of battery
used in this study is been introduced in this section. The mathematical equations
based on the actual SOC of the battery as follows [4, 7, 36]:

Vbt = E batt − R I (4)

During charge

Q Q
E batt = E 0 − K It − K I ∗ + A.exp(−bit) (5)
Q − it Q − it

During discharge

Q Q
E batt = E 0 − K It − K I ∗ + A.exp(−bit) (6)
Q − it It + 0.1Q

For charging and discharging we used a DC/DC bidirectional buck/boost

converter.
• MPPT technique. Changing of environment conditions with the increasing
demands of PV system cause to modification voltage, current and also the MPPT
of PV panels this needs optimizing the power generation to minimize the number
of PV panels needed for required load demand and increase the efficiency of the
system with reducing the total cost, It’s provided by the MPPT algorithm [23].

In this work Fuzzy logic controller is used for the control of tracking the maximum
power of the Photovoltaic generator (PVG), because of their advantages such as
robustness, simplicity to design, it require the thorough knowledge about the system
and doesn’t needs the model of the system. The proposed FLC, shown in Fig. 3 [34,
35]. In the simulation and for extract the maximum power from PV panels and Wind
turbine, A Boost Converter is widely used [36].
The inputs variables of FLC are the error E and the change of error CE at sampled
times k defined by

Pph (k) − Pph (k − 1)

E(k) = (7)
V ph (k) − V ph (k − 1)
1750 H. Abadlia and R. Toufouti

Fig. 3 Diagram of a fuzzy logic controller

Table 1. Fuzzy rules table

C E(k) = E(k) − E(k − 1) (8)

Where:
Pph (k) power of the PVG at sampled times k.
E(k) obvious if the load operation point is situated on the left or on the right of the
MPPT on the PV characteristic, and CE(k) give the change direction of this point.
Using Madani’s method the fuzzy inference are given in (Table 1), and the defuzzi-
fication uses the centre of gravity to compute the duty cycle D which is the output
of this FLC: [35]
n  
j=1 μ D j − D j
D= n   (9)
j=1 μ D j

The inputs and E(k), Ce(k) and the output D for the tracking of the maximum
power point are simulated in matlab and illustrated in Fig. 4.
Fuzzy Logic Based MPPT Control ... 1751

2.2 Simulation Results

The proposed hybrid system has been modeled and simulated in this section, The PV
panels are modeled for 5.5 KW generated power, 6 panels are connected in series
to form string and 3 strings are connected in parallel, in total 18 panels are used for
building the solar PV system which based on the fuzzy logic MPPT algorithm. A
4 KW PMSG is driven by wind turbine, the system supplies an AC load of 8 KW
maximum considered as 6 KW principal load and additional load of 2 KW. The
simulation model permit to study different parameters of the system like voltage,
current, power also confirm the efficiency of control strategy used.
With a different metrological conditions, the simulation tests of the HRES are
implemented and interpreted. The air temperature and solar radiation have been
used as the input parameters of the PV generator. The results are obtained to evaluate
the control performance to confirm the efficiency of FLC techniques.

(a) Error e (b) Change of error ce (c) Duty ratio

Fig. 4 I/O membership functions for DC/DC boost converter

Fig. 5 a. Irradiances variation. b. Solar power

1752 H. Abadlia and R. Toufouti

Figure 5 presents the variation of power generated from the PV system in different
values of solar irradiance (1000,400,800 w/m2 ) respectively, The output PV power
takes the same shape of the solar irradiance due to the direct proportional relationship
between them, the MPPT controller give maximum power output at low values of
solar irradiances.
The wind turbine generated power is depended of the wind speed variations the
output wind power depict in Fig. 6. While the wind speed is over 11 m/s, the output
power is limited to 3.5 kW. When the wind speed is less than the wind speed (6 m/s),
the output power is 1 kW.
Figure 7(a-b) the DC voltage and and load powers, we can see the Vdc voltage
I the input inverter is not affected by the power generated by the renewable sources
it’s maintained stable through MPPT controller.
The results simulation shown in Fig. 8 demonstrate that the power of hybrid system
is almost equal to the total power generated by the PV and wind generators that’s
obvious the efficiency of the fuzzy logic technique for the MPPT of PV generator.
When PHRES < Pload, the sum of power generated by the wind and the PV is
insufficient to supply the load demand. For this reason the battery start supplying

Fig 6 a. Wind speed variation. b. Wind power

Fig. 7 a. DC bus voltage VDC b. Active and reactive load powers

Fuzzy Logic Based MPPT Control ... 1753

Fig. 8 Output power for PV generator, Wind generator and hybrid system

load, the state of charge start decreasing and also during discharging current become
positive.
When PHRES > Pload, there is excess power obtainable for generation. The excess
power of HRES is used for battery charging.
From these simulation results we conclude that the HRES based Fuzzy logic
MPPT controller has a good performances in the presence of irradiance and wind
speed variations, is good solution for alimenting rural areas located in isolated site
or hard-to-access locations and do not allow connection to an power network since
the investment cost.

3 Conclusion

In this paper, an AC-linked stand-alone Photovoltaic-Wind-bettery hybrid system

coupled to an alternative AC load is proposed. First, in the work of this paper we
have present a modeling of different power sources that consist of PV- Wind-battery
storage and the MPPT controller to extract the maximum. Secondly, with the metro-
logical conditions (radiation and wind speed), we have applied numerical simula-
tion to validate the performances of a standalone hybrid system by supporting of a
charging and discharging of Battery storage with terminal DC bus. The output PV
and wind energy are modeled and a MPPT controller based of FLC technique, the
simulation result depicts than the combination of the two sources assures the relia-
bilty of the system, better than systems used only one source, with a complementary
way to manage the response for demand of power in any moment.
1754 H. Abadlia and R. Toufouti

References

1. Lopes LAC, Lienhardt AM (2003) A simplified nonlinear power source for simulating PV
panels. In: IEEE 34th annual conference on power electronics specialist, PESC 2003, vol 4.
IEEE, pp 1729–1734
2. Javed MS, Ma T, Jurasz J, Amin M (2020) Y: Solar and wind power generation systems with
pumped hydro storage: review and future perspectives. Renew Energy 148:176–192
3. Bagen, Billinton R (2005) Evaluation of different operating strategies in small stand-alone
power systems. IEEE Trans. Energy Convers 20(3):654–660
4. Ogunjuyigbe ASO, Ayodele TR, Akinola OA (2016) Optimal allocation and sizing of
PV/Wind/Splitdiesel/Battery hybrid energy system for minimizing life cycle cost, carbon emis-
sion and dump energy of remoteresidential building. Appl Energy 171:153–171. https://doi.
org/10.1016/j.apenergy.2016.03.051
5. Bakir H, Kulaksiz A (2020) A: Modelling and voltage control of the solar-wind hybrid micro-
grid with optimized STATCOM using GA and BFA. Eng Sci Technol Int J 23(3):576–584
6. Al Ghussain L, Taylan O (2019) Sizing methodology of a PV/wind hybrid system: case study
in cyprus. Environ Progr Sustain Energy 38(3):e13052
7. Diab AAZ, Sultan HM, Mohamed IS, Kuznetsov ON, Do TD (2019) Application of different
optimization algorithms for optimal sizing of PV/wind/diesel/battery storage stand-alone
hybrid microgrid. IEEE Access 7:119223–119245
8. Boussetta M, Motahhir S, El Bachtiri R, Allouhi A, Khanfara M, Chaibi Y (2019) Design and
embedded implementation of a power management controller for wind-PV-diesel microgrid
system. Int J Photoenergy. https://doi.org/10.1155/2019/8974370
9. Rezkallah M, Chandra A, Saad M, Tremblay M, Singh B, Singh S, Ibrahim H (2018) Composite
control strategy for a PV-wind-diesel based off-grid power generation system supplying unbal-
anced non-linear loads. In 2018 IEEE Industry applications society annual meeting (IAS).
IEEE, pp 1–6
10. Xiong G, Zhang J, Shi D, He Y (2018) Parameter extraction of solar photovoltaic models using
an improved whale optimization algorithm. Energy Convers Manage 174:388–405
11. Sakib S, Siddique MAB (2019) Modeling and simulation of solar photovoltaic cell for the
generation of electricity in UAE. In: 5th international conference on advances in electrical
engineering (ICAEE). IEEE, pp 66–71
12. Boukenoui R, Salhi H, Bradai R, Mellit A (2016) A new intelligent MPPT method for stand-
alone photovoltaic systems operating under fast transient variations of shading patterns. Sol
Energy 124:124–142
13. Chaibi Y, Allouhi A, Malvoni M, Salhi M, Saadani R (2019) Solar irradiance and temperature
influence on the photovoltaic cell equivalent-circuit models. Sol Energy 188:1102–1110
14. Ayang A, Wamkeue R, Ouhrouche M, Djongyang N, Salomé NE, Pombe JK, Ekemb G (2019)
Maximum likelihood parameters estimation of single-diode model of photovoltaic generator.
Renew Energy 130:111–121
15. Yahya-Khotbehsara A, Shahhoseini A (2018) A fast modeling of the double-diode model for
PV modules using combined analytical and numerical approach. Sol Energy 162:403–409
16. Khanna V, Das BK, Bisht D, Singh PK (2015) A three diode model for industrial solar cells
and estimation of solar cell parameters using PSO algorithm. Renew Energy 78:105–113
17. Allam D, Yousri DA, Eteiba MB (2016) Parameters extraction of the three diode model for
the multi-crystalline solar cell/module using Moth-Flame Optimization Algorithm. Energy
Convers Manage 123:535–548
18. Peter AG, Saha AK (2019) Electrical Characteristics improvement of photovoltaic modules
using two-diode model and its application under mismatch conditions. In: Southern African
Universities Power Engineering Conference/Robotics and Mechatronics/Pattern Recognition
Association of South Africa (SAUPEC/RobMech/PRASA). IEEE, pp 328–333
19. Bader S, Ma X, Oelmann B (2019) One-diode photovoltaic model parameters at indoor
illumination levels–a comparison. Sol Energy 180:707–716
Fuzzy Logic Based MPPT Control ... 1755

20. Gnetchejo PJ, Essiane SN, Ele P, Wamkeue R, Wapet DM, Ngoffe SP (2019) Important notes
on parameter estimation of solar photovoltaic cell. Energy Convers Manage 197:111870
21. Praiselin WJ, Edward JB (2017) Improvement of power quality with integration of solar PV
and battery storage system based micro grid operation. In: Innovations in power and advanced
computing technologies (i-PACT). IEEE, pp 1–5
22. Ayodele TR, Ogunjuyigbe ASO, Ekoh EE (2016) Evaluation of numerical algorithms used in
extracting the parameters of a single-diode photovoltaic model. Sustain Energy Technol Assess
13:51–59
23. Yilmaz U, Kircay A, Borekci S (2018) PV system fuzzy logic MPPT method and PI control
as a charge controller. Renew Sustain Energy Rev 81:994–1001
24. John R, Mohammed SS, Zachariah R (2017) Variable step size Perturb and observe MPPT
algorithm for standalone solar photovoltaic system. In: 2017 IEEE international conference on
intelligent techniques in control, optimization and signal processing, pp 1–6
25. Selvan DS (2013) Modeling and simulation of incremental conductance MPPT algorithm for
photovoltaic applications. Int J Sci Eng Technol 2(7):681–685
26. Kchaou A, Naamane A, Koubaa Y, M’sirdi N (2017) Second order sliding mode-based MPPT
control for photovoltaic applications. Sol Energy 155:758–769
27. Al-Dhaifallah M, Nassef AM, Rezk H, Nisar KS (2018) Optimal parameter design of fractional
order control based INC-MPPT for PV system. Sol Energy 159:650–664
28. Faranda R, Leva S (2008) Energy comparison of MPPT techniques for PV Systems. WSEAS
Trans Power Syst 3(6):446–455
29. Chow JH, Sanchez-Gasca JJ (2020) Power System Modeling, Computation, and Control. Wiley,
Hoboken
30. Arturo Soriano L, Yu W, Rubio JDJ (2013) Modeling and control of wind turbine. Math Probl
Eng. https://doi.org/10.1155/2013/982597.
31. Patel MR (2005) Wind and solar power systems: design, analysis, and operation. CRC Press
32. Salehi J, Badr AA (2019) Dynamic modeling and optimal control of a wind turbine with doubly
fed induction generator using imperialist competitive and artificial bee colony algorithms. J
Circ Syst Comput 28(04):1950070
33. Natsheh EM, Albarbar A, Yazdani J (2011) Modeling and control for smart grid integration
of solar/wind energy conversion system. In: 2011 2nd IEEE PES International conference and
exhibition on innovative smart grid technologies, pp 1–8
34. Takun P, Kaitwanidvilai S, Jettanasen C (2012) Maximum power point tracking using fuzzy
logic control for photovoltaic systems. In: World congress on engineering 2012, 4–6 July 2012,
London, UK, vol 2189. International Association of Engineers, pp 986–990
35. Cheikh MA, Larbes C, Kebir GT, Zerguerras A (2007) Maximum power point tracking using
a fuzzy logic control scheme. Rev Energies Renouvelables 10(3):387–395
36. Feddaoui O, Toufouti R, Jamel L, Meziane S (2020) Fuzzy logic control of hybrid systems
including renewable energy in microgrids. Int J Electr Comput Eng (IJECE), 10(6). (2088-8708)
Implementation of a Digital Control
for PV Power Systems

Badreddine Lahfaoui

Abstract Our laboratory LGEM is equipped with a Monocrystalline solar module.

We want to control the PV system Monocrystalline solar module 30W - ET-
M53930. To this end, we have implemented a digital control in the DSPACE 1104
card. Our experiments are carried out at the Oujda Higher School of Technology
(ESTO) at the LGEM laboratory with an ambient temperature of the room around
16 °C. Several tests are performed in order to validate our proposed controller. We
mention that our experimental work is done by implementing a digital control for a
PV system studied in the absence of disturbance: tests with a fixed radiation and a
fixed charge.

Keywords System identification and control · Power electronics · Renewable

energies · Wind turbine · PV systems · Hybrid systems · Control card

1 Introduction

Solar panels are the main sources of any photovoltaic system. They are connected in
parallel or in series to remove their operating voltage or current. The energy supplied
by the system can be used to charge batteries which will provide electricity when
needed.
It can also be used by directly connecting the modules to the load without batteries
(i.e. for a solar pump, water is used as storage), or by plugging them into an electrical
network. It is also possible to combine the output of the PV field with other energy
sources such as a generator or a wind turbine which will serve as back-up, if there is
not enough sunlight.
Photovoltaic systems can be either (Autonomous system without battery,
Autonomous system with battery, Hybrid PV/generator system) the stand-alone
systems, not connected to a power grid. Or the other three types (PV system on

B. Lahfaoui (B)
Mohammed First University Oujda, EST Oujda, LGEM, 60000, Oujda, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1757
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_159
1758 B. Lahfaoui

diesel grid, PV system on decentralized grid, Centralized PV system) the PV systems

connected differently to the electricity grid [1, 2].
To control the power of the PV system to the maximum, it is essential to operate the
photovoltaic installations near this point (voltage VPVpmax and current IPVpmax)
that is to say to have this power even if there are climatic or physical variations.
Several analog and digital controls [3–8] are developed to solve this problem.
The objective of this paper is to present all the practice works of a PV system
connected to the Boost converter using a digital control. For this reason electrical
circuit within the laboratory were implemented in order to test this numerical control.
This paper is organized as follows: we start first in Sect. 2 by presenting the
technical specifications of the photovoltaic source (manufacturer characterization
and also experimental characterization performed in the laboratory).
In Sect. 3 a detailed description of the system architecture is made by presenting
the numerical control algorithm as well as the specifications of the Boost converter
used during the experiments. The Sect. 4 shows the experimental results obtained,
and at the end we finish with a conclusion.

2 The Specifications of the Photovoltaic System

2.1 The Solar Module - ET-M53930

This subset converts solar energy into electrical energy. It is made up of the following
elements:
– A 30 W “Monocrystalline” type photovoltaic solar module.
– An aluminum structure tilting from 15 to 75°.
– 3 safety connectors (4mm) (+, −, PE).
– A non-return diode located in the module connection box.
The ETM53930 solar panel consists of 39 monocrystalline silicon cells connected
in series. Table 1 presents the electrical specifications, of this module, which are
obtained under standard test conditions (STC):
– Solar irradiation of 1000 W/m2 .
– 1.5 air masses.
– And temperature of 25° C.
We also present the electrical performance of this solar module illustrated by me
in the labs which represents on the one hand the electrical power as a function of
voltage and on the other hand the electric current as a function of voltage.
Implementation of a Digital Control for PV Power Systems 1759

Table 1. PV system
Parameters Values
characteristics
Weight 3.5 kg
Dimensions 633 × 427 × 35 mm
Maximum power voltage (Vmp) 19.4V
Maximum power current (Imp) 1.55A
Open circuit voltage (Voc) 23.8V
Short circuit current (Isc) 1.74A
Temp.coeff. of Isc (Tk Isc) 0.06%/°C
Temp.coeff. of Voc (Tk Voc) −0.397%/°C
Temp.coeff. of Pmax (Tk Pmax) −0.549%/°C
Normal operating cell temperature 44.4+− 2 °C

2.2 Experimental Characterization

The electrical specifications of our PV system are determined by making circuits in

the laboratory using an artificial source of solar energy that generates a desired solar
irradiation to attack the photovoltaic system.
For this, several measurements sampling works are taken in our research
laboratory by linking our panel to electronic converter boost.
The voltage and the current at the terminals of the solar panel (Vpv , Ipv ) are
measured at each solar irradiance while varying the value of the resistive load. The
Fig. 1 illustrates all these data taken.

Fig. 1 The experimental electrical characteristics of the PV system

1760 B. Lahfaoui

Fig. 2 The equipment used in the test.

3 Numerical Control of the Solar System

3.1 The Numerical Experimental Control PV System

Wanting to exploit the optimum power of the solar module ET-M53930.For this
we have implemented the numerical control algorithm in the DSPACE 1104 card.
Figure 2 represents the experimental hardware formed by:
– PV source 30W monocrystalline solar module - ET-M53930
– Artificial solar source (W/m2 )
– One Boost
– Load
– Current sensor
– Voltage sensor
– DSPACE 1104 card
– Computer with ControlDesk software

3.2 The Numerical Method Implemented in the Control Card

The Perturb and Observe Numerical control has been implemented on a DSPACE
1104 card to perform power regulation.
Implementation of a Digital Control for PV Power Systems 1761

Fig. 3 Digital control implemented in DSPACE 1104.

The algorithm of this method as illustrated in the Fig. 3 begins with the measure-
ment of the two parameters current and voltage Vpv (k), Ipv (k) to calculate the instan-
taneous power value at the instant k and the previous power at the instant k-1. Also
the algorithm calculates the difference in voltages Vpv = Vpv (k) − Vpv (k−1).
Then the command checks the sign of the value of Ppv and following the sign
verification of the value of Vpv the algorithm increments or decrements the duty
cycle Dpv with a value of Dpv .
This algorithm is experimentally translated into a programming language using
the Simulink tool from MATLAB software and is then sent to the DSPACE control
card handled via the Control Desk software.
1762 B. Lahfaoui

Table 2 Components of the

Elements Value of component
Step-Up converter
Inductances L’1 = 100 Uh
Internal resistance r’1 = 0.2 
Input capacities C1 = 1000 uF/250V
Output capacities C2 = 1000 uF/250V
R(Load) 500 
Switch IRFp450
Diode BYW98-200

Table 3 Measuring
Equipment Parameters
equipment
Current sensor 10A to 1V and 100A to 1V
Voltage sensor ST1000

3.3 Specification of Step-Up Converter

So as to obtain experimental results, we had to go through the realization phase, of

which we carried out Step-Up converter for our solar panel system with the values
and parameters calculated mentioned in Table 2.
The Table 3 shows the set of equipment used during the test.

4 Experimental Results of Numerical Control System

We present above the experimental results of the various study tests in the absence
of disturbance in Figs. 4 which include the power at the terminals of the solar panel
Ppv as function as the value of the duty cycle Dpv .
The optimum power is reached for all of these tests through the use and
implementation of the MPPT control in the DSPACE 1104 card.
– The maximum power at the out of the solar panel oscillates approximately around
18W.
– The optimal duty cycle oscillates approximately around 0.7.
The experimental results illustrated in Fig. 4 are captured by Controdesk software
over a time interval of 500 s and a solar irradiance equal to approximately 1000W/m.
Implementation of a Digital Control for PV Power Systems 1763

Fig. 4 Tracking the maximum power using DSPACE 1104 Card (In Controldesk).
1764 B. Lahfaoui

5 Conclusions

The maximum optimization of renewable energy sources has become essential for
socio-economic and environmental reasons. For this, various research works are
initiated on several sources such as photovoltaic sources, wind sources, and other
sources.
Our research work described in this paper presents the optimization of a photo-
voltaic system driven by a numerical control to satisfy a resistive load under solar
irradiance and constant temperature.
The experience was carried out in the laboratory, LGEM, using several control
measuring equipment and artificial sources.
A validation of the experimental results shown in this paper proves the success of
our proposed architecture.

References

1. Labouret A, Villoz M (2009) Energie solaire photovoltaïque. DUNOD, Paris

2. Royer J, Djiako T, Schiller E (1998) BocarSadaSy “le pompage photovoltaïque” (1998)
3. Lahfaoui B, Zouggar S, Elhafyani ML, Rabhi AH (2019) Implementation of a real-time MPPT
of hybrid renewable energy system composed of wind turbine and solar PV cells. In: Hajji B,
Tina G, Ghoumid K, Rabhi A, Mellit A (eds.) proceedings of the 1st international conference
on electronic engineering and renewable energy. ICEERE 2018. Lecture Notes in Electrical
Engineering, vol 519. Springer, Singapore. https://doi.org/10.1007/978-981-13-1405-6_58
4. Lahfaoui B, Mohammed SZB, Elhafyani ML (2017) Real time study of P&O MPPT control for
small wind PMSG turbine systems using arduino microcontroller. Energy Procedia 111:1000–
1009
5. Lahfaoui B, Zouggar S, Elhafyani ML, Seddik M (2015) Experimental study of P&O MPPT
control for wind PMSG turbine. In: IEEE international, renewable and sustainable energy
conference (IRSEC 2015), Marrakech, Morocco, 10–13 Dec, pp 1–6
6. Lahfaoui B, Zouggar S, Elhafyani ML, Benslimane A (2015) Modeling validation and MPPT
technique of small wind PMSG turbines using DSPACE hardware. In: IEEE international, renew-
able and sustainable energy conference (IRSEC 2015), Marrakech, Morocco, 10–13 Dec, pp
1–6
7. Lahfaoui B, Zouggar S, Elhafyani ML, Seddik M (2016) An Experimental Study of P&O MPPT
Control for Photovoltaic Systems. Int J Power Electron Drive Syst (IJPEDS) 7:954
8. Lahfaoui B, Zouggar S, Elhafyani ML, Kadda FZ (2014) Experimental modeling and control of a
small wind PMSG turbine. In: IEEE international, renewable and sustainable energy conference
(IRSEC 2014), Ouarzazat, Morocco, pp 802–807
Towards a Digital Modeling
of the Optimal Mechanical Properties
of a Green Eco Composite Based
on Renewable Resources
Aziz Moumen, Abdelghani Lakhdar, Mustapha Jammoukh,
and Khalifa Mansouri

Abstract In order to respond to the industrialists, environmental and sustainable

development needs, researchers are currently studying a new category of materials
called eco-composites based on natural biofillers. The goal is to use animal and
vegetable fibers to load polymers. Several sectors integrate these eco-composites to
manufacture their products by taking advantage of the important mechanical proper-
ties obtained by this bio-loading compared to the polymers in pure state and also to
attract consumers to the bio-sourced products. In this research, the optimal proper-
ties of Nylon 66 bio filled by Argan Nut Shell (ANS) Particles with the Mori Tanaka
and the finite element methods were studied. The mechanical properties of the eco
composites studied are optimal at 20 Wt% of particles and by increasing their aspect
ratios. The rigidity decreases slightly by moving away from 0°. The use of these types
of natural bio fillers can replace the use of conventional fillers and move towards eco
composites that respect the environment with low cost and density.

Keywords Eco composite · Sustainable development · Environment · Argan Nut

Shell · Nylon 66 · Digital modeling

1 Introduction

The use of classical loads as reinforcement of polyamides is more and more inad-
visable thanks to their environmental, economic and sustainable development draw-
backs [1, 2]. Scientific research is more and more interested in a new category of
materials called “eco composites” or “bio composites” meeting the requirements
of sustainable development, manufacturers and environmental constraints. They
are based on natural bio fillers also offering a lower cost and density with a high
biodegradability.
In this work, the Argan Nut Shell (ANS) particles will be used in order to give them
the good value in the socioeconomic and environmental context. The Nylon PA66

A. Moumen (B) · A. Lakhdar · M. Jammoukh · K. Mansouri

SSDIA Laboratory, Hassan II University of Casablanca, ENSET Mohammedia, Casablanca,
Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1765
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_160
1766 A. Moumen et al.

Table 1 Characteristics of
ANS Values
ANS
Content 20 Wt%
Orientation 0,45 and 90°
Aspect ratios 0.5,2 and 20

used from the family of thermoplastics has a light weight and good thermomechanical
properties [3–5]. Several Nylons can be quoted: 6,11,12,46 and 66 etc. [6–11]. They
are more used in transportation, automotive, textile and packaging industries. The
PA66 is one of the most used Nylons. Its synthesis is made by the hexamethylene
diamine and adipic poly condensation.
Filled Nylon has better characteristics in terms of strength and stiffness. We will
study in this work the optimal properties of the Nylon PA66 bio filled by ANS
Particles with the Mori Tanaka and finite element methods by varying the content,
orientation and aspect ratios of particles.

2 Materials and Methods

2.1 Materials

The ANS particles are rich in cellulose, hemi cellulose, lignin and Ashes by (25.7%,
34.3%, 34.5% and 5.4% respectively) as determined by Essabir et al. [2]. The Table 1
presents the properties of the particles used.
The Nylon 66 used is from the family of thermoplastics having a light weight and
good thermo-mechanical properties [5, 12, 13].

2.2 Methods

2.2.1 Mori-Tanaka Method

We will use this model to predict the mechanical behavior of Nylon 66 bio filled by
ANS particles. It is advantageous compared to other homogenization models such
as Voigt, Tsai Pagano and Reuss etc. [14–16] it gives more efficiency and reliability
of predicted behaviors. The homogenization parameters are cited in the Table 2.
In order to have better convergence, we are going to use a minimal, maximal and
final times respectively of 0.01, 0.1 and 1 s. The values chosen in Table 2 are the
result of several numerical simulations leading to good convergence and reliability
of the results.
Towards a Digital Modeling of the Optimal Mechanical Properties … 1767

Table 2 Homogenization, loading and orientation control

Parameters Values
Stopping criterion 1E-006
Maximal stage of error 1E-005
Iterations prior to control 4
Highest iterations 20
Increments for angle 12
Trace-tolerance 0.1

Fig. 1 Geometry of Nylon PA66 bio filled by ANS particles a Cylinder particles, b Ellipsoid
particles, c Sphero cylinder particles

2.3 Finite Element Method

A. Geometry of the eco composite

The PA66 bio filled by ANS particles is numerically presented as a RVE in order
to perform the finite element modeling which is powerful in predicting the complex
non linear behaviors [17–21]. The Fig. 1 presents the eco composite geometry.
B. Mechanical loading
We are going to use the periodic boundary conditions that are periodic to all the faces
of the volume element. This is enforced through a large set of equations relating the
degrees of freedom of the nodes lying on one face with those of the corresponding
nodes and lying on the opposite face, on a 2-by-2 basis. Node duplication is used to
prevent issues coming from non periodic meshes. Figure 2 illustrates the definition of
periodic boundary conditions in the case of a mechanical loading with a macroscopic
uniaxial peak strain 11 .
Periodic boundary conditions usually lead to the best predictions when compared
to the Dirichlet and Mixed boundary condition type. It also shows a faster convergence
rate as the size of the volume element increases, but to the expense of increased CPU
time and memory requirements to solve the finite element problem. This decrease of
1768 A. Moumen et al.

Fig. 2 Periodic boundary

conditions

Fig. 3 Uniaxial_1 loading

of the eco composite

performances is due to the potentially large set of constraint equations that has to be
imposed.
A macroscopic uniaxial strain state in the 1-direction is imposed (macroscopic
uniaxial stress state) as presented in the Fig. 3.
C. Incrementation and Mesh
We are going to use a max of 20 increments with initial, minimum, maximum and
final times respectively of 0.1, 0.01, 0.1 and 1. The mesh used is the voxel non-
conforming, having several advantages in terms of convergence, flexibility of use
and the best accuracy [20].
Towards a Digital Modeling of the Optimal Mechanical Properties … 1769

3 Results and Discussion

3.1 Eco Composite Mesh

Figure 4 presents the mesh performed by the finite element method.

3.2 Density

The density of ANS particles is around 1.29 g/cm3 which is lower compared to
mineral or inorganic fillers such as 2.6g/cm3 clay [2] and talc 2.8 g/cm3 [2] showing
a lightness of our bio load which can benefit the polyamides studied.
The Table 3 and the Fig. 5 show the density of the eco-composite based on ANS
particles compared to other composites loaded with mineral and inorganic particles
(clay and talc).

Fig. 4 Mesh of Nylon PA66 bio filled by ANS particles a Cylinder particles, b Ellipsoid particles,
c Sphero cylinder particles

Table 3 Density of different

Density Values
loads and bio composites
using Mori-Tanaka method ANS 1.29
Clay 2.6
Talc 2.8
Nylon 66 1.14
Nylon 66- ANS 1.1671
Nylon 66-Clay 1.2842
Nylon 66-Talc 1.2934
1770 A. Moumen et al.

Fig. 5 Density of different bio composites using Mori-Tanaka method (M-T)

3.3 Effect of the Orientation

3.3.1 Random Orientation

In order to study the effect of ANS particles orientation on the behavior of the polymer
studied, we modeled different orientations ranging from 0° to 90°.The orientation
tensor is defined by:
⎡ ⎤
a11 a12 a13
ai j = ⎣ a12 a22 a23 ⎦ (1)
a13 a23 a33

The trace of the orientation tensor must be equal to1.

a11 + a22 + a33 = 1 (2)

In order to obtain a trace equal to 1, we define a trace tolerance of 0.1. If the result
is far from [1-tol; 1 + tol], the numerical modeling corrects the orientation tensor by
this way (Table 4):
⎡ ⎤ ⎡ a ⎤
a11 a12 a13
11
trace(a)
a12 a13
⎢ ⎥
ai j = ⎣ a12 a22 a23 ⎦is corrected to a = ⎣ a12
 a22
trace(a)
a23 ⎦ (3)
a33
a13 a23 a33 a13 a23 trace(a)

The reference orientation tensor for the three morphologies is as follows:

⎡ ⎤
0.333 0 0
ai j = ⎣ 0 0.333 0 ⎦ (4)
0 0 0.333
Towards a Digital Modeling of the Optimal Mechanical Properties … 1771

Table 4 Global error on

ANS Values
orientation tensor
Cylindrical 0.0779
Ellipsoid 0.0709
Sphero cylindrical 0.0472

⎡ ⎤
0.3016 −0.0157 −0.0218
ai j (cylindrical) = ⎣ 0 0.3919 −0.0138 ⎦ (5)
0 0 0.3065
⎡ ⎤
0.3016 −0.0113 −0.0477
ai j (ellipsoid) = ⎣ 0 0.3263 0.00797 ⎦ (6)
0 0 0.3721
⎡ ⎤
0.3472 0.00351 0.00244
ai j (sphero cyl) = ⎣ 0 0.3517 0.02548 ⎦ (7)
0 0 0.3012

3.4 Fixed Orientation

If we want to express the aligned Representative Volume element, it’s better to choose
the fixed orientation tensor to save computing time. We define the angles θ and φ.
The first one is closed to 90° and the second one varies from 0° to 180° depending
on the orientation in the plane (1–2).
The following figures obtained by the finite element method present the elastic
behaviors of composites for 0°, 45 and 90° orientations (Fig. 6).
The analysis of the curves in the figure below shows a slight stiffness in terms
of Young modulus and the stress at break for cylindrical, sphero cylindrical and
ellipsoid particles at 0°. This rigidity decreases slightly away from 0°.The less rigid
behavior is the same for cylindrical and sphero-cylindrical particles at 45° and 90°
whereas for ellipsoid particles, the behavior at 45° is slightly higher than 90°.
In order to validate the results obtained for the orientation effect, a comparison
between the results of theMori-Tanakaand finite element methods is carried out as
presented by the following figures:
The analysis of the curves of the Figs. 7, 8 and 9 obtained numerically shows that
the results obtained by the finite element method are similar to those obtained by
the Mori-Tanaka model. The rigidity decreases slightly by moving away from 0° in
terms of Young modulus and the stress at break for cylindrical, sphero-cylindrical
and ellipsoid particles.
Compared to the orientation, it is slightly advised to use cylindrical and sphero
cylindrical morphologies especially when using particles in orientations close to 90°.
1772 A. Moumen et al.

Fig. 6 Particle orientation effects for the three morphologies using the finite element method

Fig. 7 Elastic curves for cylindrical and sphero cylindrical particles at 0° by the FE method and
M-T model

3.5 Effect of Aspect Ratio

In order to study the effect of ANS particles aspect ratio on the behavior of the
polymer studied, we modeled several aspect ratios (0.5, 2 and 20). The Figs. 10
and 11 present respectively the tensile curves obtained by theMori-Tanaka and finite
element methods for these different aspect ratios.
Towards a Digital Modeling of the Optimal Mechanical Properties … 1773

Fig. 8 Elastic curves for cylindrical and sphero cylindrical particles at 45° and 90° by the FE
method and M-T model

Fig. 9 Elastic curves for ellipsoid particles at 0°, 450 and 90° by the FE method and M-T model

The results illustrated in Figs. 10 and 11 shows that the Mori-Tanaka and finite
element methods give the same results for aspect ratios of 0.5,2 and 20.The two
numerical methods clearly show that the rigidity of the bio composite improves by
increasing the aspect ratio of Argan Nut Shell. At an aspect ratio of 20, the rigidity is
much better. The stress at break is greater with these high aspect ratios. These results
are valid for all the morphologies studied (cylinder, sphero cylinder and ellipsoid).
1774 A. Moumen et al.

Fig. 10 Tensile curves obtained by the finite element method

Fig. 11 Tensile curves obtained by the Mori-Tanaka model

4 Conclusion

In the present work, the importance of bio loading was studied in order to have green
eco composites meeting the expectations of researchers in terms of respecting the
environment, regulations and sustainable development requirements.
Towards a Digital Modeling of the Optimal Mechanical Properties … 1775

The Mori Tanaka and finite element methods were used to characterize the
mechanical behavior of Nylon 66 bio filled by ANS particles. These two digital
methods have proved a great ability to predict the behavior of the eco composite
studied. They are reliable, simple in use and have more speed in computing.
At 20 Wt% of particles and by increasing their aspect ratios, the mechanical
properties of the eco composites studied are optimum. By moving away from 0°, the
rigidity slightly reduces. The use of these bio fillers will replace the use of traditional
fillers and move towards environment friendly eco composites.

Future Scope
In this study, we determined the mechanical behavior of Nylon PA66 bio loaded by
ANS particles. It will be interesting to study the thermal behavior and apply other
tests on this eco composite such as fatigue and aging tests. Other bio loads of animal
or plant origin can be studied for the reinforcement of this Nylon using experimental
tests and digital modeling.

References

1. Russo P, Simeoli G, Vitiello L, Filippone G (2019) Bio-polyamide 11 hybrid composites

reinforced with basalt/flax interwoven fibers: a tough green composite for semi-structural
applications. Fibers 7:41
2. Essabir H (2015) Bio-composites à base de coque de noix d’arganier: Mise en œuvre,
caractérisation et modélisation du comportement mécanique, p 184
3. Rizkalla S, Dawood M, Schnerch D (2008) Development of a carbon fiber reinforced polymer
system for strengthening steel structures. Compos Part A Appl Sci Manuf 39:388–397
4. Choi EY, Kim MH, Kim CK (2019) Fabrication of carbon fiber grafted with acyl chloride func-
tionalized multi-walled carbon nanotubes for mechanical reinforcement of nylon 6,6. Compos
Sci Technol 178:33–40. https://doi.org/10.1016/j.compscitech.2019.05.012
5. Lyu WY, Chen XM, Li YB, Cao S, Han YM (2019) Thermal stability and heat release effect of
flame retarded PA66 prepared by end-pieces capping technology. Compos Part B Eng 167:34–
43. https://doi.org/10.1016/j.compositesb.2018.12.016
6. Tang Q, Wang K, Ren X, Zhang Q, Lei W, Jiang T, Shi D (2020) Preparation of porous
antibacterial polyamide 6 (PA6) membrane with zinc oxide (ZnO) nanoparticles selectively
localized at the pore walls via reactive extrusion. Sci Total Environ 715:137018
7. Nagasawa N, Tago T, Kudo H, Taguchi M (2017) Radiation-induced crosslinking of
polyamide11 in the presence of triallylisocyanurate. Polym Degrad Stab 136:98–102. https://
doi.org/10.1016/j.polymdegradstab.2016.12.014
8. Li M, Chen AN, Lin X, Wu JM, Chen S, Cheng LJ, Chen Y, Wen SF, Li CH, Shi YS (2019)
Lightweight mullite ceramics with controlled porosity and enhanced properties prepared by
SLS using mechanical mixed FAHSs/polyamide12 composites. Ceram Int 45:20803–20809.
https://doi.org/10.1016/j.ceramint.2019.07.067
9. Umesh GL, Prasad NJK, Rudresh BM, Devegowda M (2020) Influence of nano graphene on
mechanical behavior of PA66/PA6 blend based hybrid nano composites: Effect of micro fillers.
Mater Today Proc 20:228–235
10. Lee CS, Kim HJ, Amanov A, Choo JH, Kim YK, Cho IS (2019) Investigation on very high
cycle fatigue of PA66-GF30 GFRP based on fiber orientation. Compos Sci Technol 180:94–100.
https://doi.org/10.1016/j.compscitech.2019.05.021
11. Spina R, Cavalcante B (2020) Thermal analysis of PA66 grinding. Procedia Manuf 47:910–914
1776 A. Moumen et al.

12. Leconte N, Bourel B, Lauro F, Badulescu C, Markiewicz E (2020) Strength and failure of an
aluminum/PA66 self-piercing riveted assembly at low and moderate loading rates: experiments
and modeling. Int J Impact Eng 142:103587
13. Moumen A, Lakhdar A, Mansouri K (2020) Numerical study of the mechanical behavior of
polyamide 66 reinforced by argan nut shell particles with the finite element method and the
mori-tanaka model. Int J Adv Trends Comput Sci Eng 9:7723–7730. https://doi.org/10.30534/
ijatcse/2020/115952020
14. Moumen A, Jammoukh M, Zahiri L, Mansouri K (2020) Study of the optimal micromechanical
behavior of a polymer reinforced by snail shell particles using the mori-tanaka numerical model.
In: 2020 IEEE international conference of Moroccan geomatics (Morgeo). IEEE, pp 1–6
15. Lakhdar A, Moumen A, Zahiri L, Jammoukh M, Mansouri K (2020) Experimental and numer-
ical study of the mechanical behavior of bio-loaded PVC subjected to aging. Adv Sci Technol
Eng Syst J 5:607–612
16. Chihaoui B, Serra-Parareda F, Tarrés Q, Espinach FX, Boufi S, Delgado-Aguilar M (2020)
Effect of the fiber treatment on the stiffness of date palm fiber reinforced PP composites:
macro and micromechanical evaluation of the young’s modulus. Polymers (Basel) 12:1–21.
https://doi.org/10.3390/POLYM12081693
17. Moumen A, Jammoukh M, Zahiri L, Mansouri K (2020) Numerical modeling of the thermo
mechanical behavior of a polymer reinforced by horn fibers. Int J Adv Trends Comput Sci Eng
9:6541–6548. https://doi.org/10.30534/ijatcse/2020/342942020
18. Lakhdar A, Jammoukh M, Zahiri L, Mansouri K, Moumen A, Salhi B (2020) Numerical and
experimental study of the behavior of PVC material subjected to aging. In: 2020 1st international
conference on innovative research in applied science, engineering and technology (IRASET).
IEEE, pp 1–6
19. Yu G, Shi B, Shen Y, Gu J (2020) A novel finite element method for predicting the tensile
properties of 3D braided composites. Mater Res Exp. 6:125626
20. Rao MVP, Raj D, Harursampath D, Renji K (2020) Estimation of material properties of metal
matrix composites using finite element method in the presence of micromechanics damages.
Mater Today Proc 21:1135–1143
21. Sokolov A, Schetinin V, Kozlov M (2020) Surface finite element for imperfect interface
modeling in elastic properties homogenization. Key Eng Mater 833:101–106
Experimental Evaluation of MEMS
Accelerometers Integrated
into Smartphones: A Case Study
of Bearing Condition Monitoring

Abdelbaset Ait Ben Ahmed , Abdelhamid Touache, Abdelhadi El Hakimi,

and Abderrahim Chamat

Abstract Over the last decade, smartphones have become an essential part of
numerous applications in daily life activities. Contrary to professional equipment,
smartphones are available to everyone and are not expensive, which is why it is a
valuable opportunity to exploit smartphone capabilities for professional applications.
This paper presents an experimental evaluation of integrated MEMS accelerometers
into smartphones to assess their performance for vibration monitoring, with the
expectation to use them for educational and modest industrial purposes. This study
presents an evaluation of two different devices having different specifications, which
evaluates the extracted features of the vibration data collected from both devices.
A shaking table test was carried out to compare the two MEMS accelerometers
results with professional equipment. Finally, a case study was conducted to monitor
the condition of two bearings running at low rotational speed (600 rpm) using the
vibration data acquired through both smartphones. Experimental results indicate that
smartphones can be a useful instrument for monitoring the condition of bearings.

Keywords Vibration measurement · Experimental testing · MEMS

accelerometers · Condition monitoring · Smartphone’s sensors

A. Ait Ben Ahmed (B) · A. Touache · A. El Hakimi

Laboratory of Mechanical Engineering, University of Sidi Mohamed Ben Abdellah, Fez, Morocco
A. Touache
e-mail: [email protected]
A. El Hakimi
e-mail: [email protected]
A. Chamat
Laboratory of Industrial Techniques, University of Sidi Mohamed Ben Abdellah, Fez, Morocco
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1777
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_161
1778 A. Ait Ben Ahmed et al.

1 Introduction

Vibration analysis is very important to identify and to monitor the health of mechan-
ical systems [1], but many people consider that the cost of vibration analysis equip-
ment is very expensive and not at hand for all. Contrary to professional equipment
(Pro. Equip.), the smartphones are characterized by low cost compared with profes-
sional equipment, and by their several functionalities. So, it’s beneficial to look
for replacing professional equipment with smartphones to use them for educational
purposes such as mechanical vibration practical works in the university curriculum,
or modest industrial purposes such as monitoring the rotating machinery operating
under low speed, and correcting the unbalance defect in rotors [2, 3].
Each day, smartphones have more and more features to add to existing devices.
The vast majority of smartphones contain a wide variety of sensors, it has become a
favored tool in the world and even an important part of daily life activities [4]. The
sensor system of smartphones contains motion sensors such as accelerometers and
gyroscopes, magnetic field measurements, light detection, and proximity detection.
Besides, sound can be detected using the microphone when the images are captured
using the camera. Sensors have been used in smartphones since their introduction,
the new smartphones are powerful devices that are always on hand and have access
to a wireless connection at anytime and anywhere [5].
Nowadays, most smartphones are programmable and equipped with autonomous,
inexpensive, small, and low energy sensors, this gives smartphones more powerful
benefits. This paper aims to presents an experimental evaluation of the accelerometer
integrated into the smartphones, to assess its performance in vibration analysis, and
compare them with professional equipment. This study shows the evaluation of two
accelerometers with different specifications and compares their performance with
professional equipment, the comparison will be based on the time and the frequency
features extracted from the vibration data recorded by each accelerometer.

2 Smartphone’s Data Acquisition

2.1 Smartphone’s Accelerometer

Smartphone’s accelerometer measures the acceleration force applied to the device,

which includes the gravitational force. In general, the sensor structure uses a standard
3-axis coordinate system (X, Y, and Z) that can be represented in the Cartesian
coordinates [6], and the accelerometer provides a measurement for each of the X, Y,
and Z directions, as shown in Fig. 1.
In this study, two different smartphones were evaluated, the first labeled old smart-
phone and the second labeled new smartphone. The specifications of each of them
are listed in Table 1.
Experimental Evaluation of MEMS Accelerometers … 1779

Fig. 1 The coordinate

system of smartphone
sensors

Table 1 Specifications of the

Specifications New Smartphone Old Smartphone (Old
smartphone’s accelerometers
(New Sp.) Sp.)
used in this study
Name MTK Inc BOSCH BMC150
Resolution 0.0012 m/s2 0.0095 m/s2
Range ±39 m/s2 (±4 g) ±159 m/s2 (±16 g)
Sampling Up to 100 Hz Up to 100 Hz
frequency

2.2 Data Pre-processing

Pre-processing the accelerometer data is very important to ensure the accuracy and
the quality of the measured data. To measure the real movement of the device it is
recommended to remove the gravity component of the acceleration measurements.
This can be achieved by applying a high-pass filter to eliminate the effect of gravity
from three directions [7]. For noise filtering, a Butterworth Infinite Impulse Response
(IIR) bandpass filter has been applied to remove noise from the vibration signal, with
appropriate cut-off frequencies according to the range of phenomena to be observed.

3 Evaluation of Smartphone’s Accelerometer

3.1 Shake Table Test

In this section, we will present the shake table to evaluate the smartphone’s
accelerometer and compare its results with those of professional equipment.
1780 A. Ait Ben Ahmed et al.

Figure 2 illustrates the experimental setup which contains the rotor system that
produces a uniform vibration (≈16 Hz) using an unbalanced mass. The vibration
of the rotor causes the vibration of the table. The smartphone and professional
equipment capture the vibrations of the table along the Z-axis direction of the
smartphone.
Figure 3 shows the vibration signal obtained using the smartphone and using the
professional equipment during the evaluation test in the Z-axis direction.

Fig. 2 Experimental setup of the evaluation test

Smartphone
1 Pro. Equi.

0.8

0.6
Acceleration (m/s²)

0.4

0.2

-0.2

-0.4

-0.6

-0.8

-1
0 0.05 0.1 0.15 0.2 0.25 0.3
Time (s)

Fig. 3 Smartphone vs. professional equipment vibration signal during the evaluation test
Experimental Evaluation of MEMS Accelerometers … 1781

Table 2 The sampling rate of

Mode Sampling rate
the smartphone’s
accelerometer for each mode Fastest ≈100 Hz
Game ≈50 Hz
Normal ≈2.5 Hz
User interface (UI) ≈1.5 Hz

Table 3 Time-domain
Features New Sp. Old Sp. Pro. Equip.
statistical features of
vibration data (Fastest mode, Mean −0.0004 0.0027 −0.0258
Z-axis direction) RMS 0.5585 0.5437 0.4937
Kurtosis 1.5260 1.6093 1.5134
Skewness 0.0024 −0.0082 0.0167
Variance 0.3125 0.2962 0.2433
Crest factor 1.6090 1.8119 1.4382

3.2 Sampling Frequency

Concerning the sampling frequency, the smartphone’s accelerometer was sampled

regularly due to the implementation of the sampling mechanism by the Android
structure. The Android API provides four different sampling rates for its accelerom-
eter sensor: The Fastest, Game, Normal, and User Interfaces (UI) [8], listed from the
fastest to the slowest.

3.3 Features Extraction

Statistical features of the time-domain, such as mean, root mean square (RMS),
Crest factor, and variance were generally used to identify distinctions between the
vibration signals. More advanced statistical features, such as skewness, and kurtosis
are also applied as features [9, 10], all these features are used in this study to evaluate
the smartphone vibration data. Also, the frequency spectrum of the vibration data
generated by the Fast Fourier transform (FFT) algorithm is exploited for extracting
and evaluating other features of the vibration data [11].

4 Results and Discussion

Tables 3 and 4 show the time-domain features of the vibration signal recorded by
the old and new smartphone compared with the professional equipment results. The
1782 A. Ait Ben Ahmed et al.

Table 4 Time-domain
Features New Sp. Old Sp. Pro. Equip.
statistical features of
vibration data (Game mode, Mean 0.0007 −0.0003 −0.0258
Z-axis direction) RMS 0.5765 0.5144 0.4937
Kurtosis 1.6153 1.7024 1.5134
Skewness −0.0551 −0.1139 0.0167
Variance 0.3330 0.2649 0.2433
Crest factor 1.7442 1.8042 1.4382

measurement concerned just the two first modes (Fastest and Game, see Table 2)
because the sampling rate of each mode is higher than the frequency of vibration
phenomena (Shake table frequency ≈16 Hz).
As shown in Tables 3 and 4, the statistical features of the vibration data,
RMS, Kurtosis, Variance, and Crest factor are more stable for both devices and in
both sampling frequency modes, compared to professional equipment results. And
concerning the frequency-domain, the new smartphone has a clear spectrum than the
spectrum obtained by the old smartphone, which is very noisy. So, this difference
was produced owing to the resolution and the accuracy of the two sensors.
The resolution and the sampling rate of the smartphone’s accelerometer play an
important role in the accuracy of the measurement and the amount of noise in the
vibration signal. Also, the type of sensor mounting on the structure affect the obtained
results, it’s recommended to be adhesive mounting.
Figure 4 shows the frequency-domain of the vibration signal recorded by the old
and new smartphone for the Fastest and Game modes in the Z-axis direction.
Based on Fig. 4, the new smartphone has a clearer spectrum than the spectrum
obtained by the old smartphone which is very noisy. It can therefore be said that the
resolution is responsible for this difference. Besides, the sampling frequency plays
an important role in the spectrum clarity, so that the spectrum of the Fastest mode
(100 Hz) is better than the spectrum of the Game mode (50 Hz).
Generally, the smartphone presents acceptable results, but the low sampling rate
of the smartphone’s accelerometer (Max 100 Hz) still the main disadvantage which
limits the interval of using smartphones in the professional areas. So, in this case, it
is recommended to evaluate the vibration phenomena in which the main frequency
is located below half of the Smartphone sampling rate (i.e. 50 Hz).

5 Case Study: Bearing Condition Monitoring

In this section, two bearings were examined to identify the degree of damage, the
first was a healthy bearing and the second was defective with a localized fault in
one ball of the bearing, as shown in Fig. 5. Bearing fault signals are considered as
amplitude modulated signals. Where the faults excite the resonance frequencies of
the bearing structure. For this reason, it is difficult to identify fault frequencies in the
Experimental Evaluation of MEMS Accelerometers … 1783

0.9 0.9
(a) (b)
0.8 0.8

0.7 0.7
Amplitude (m/s²)

0.6 0.6

Amplitude (m/s²)
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25
Frequency (Hz) Frequency (Hz)

0.6 0.6
(c) (d)
0.5 0.5

Amplitude (m/s²)
0.4 0.4
Amplitude (m/s²)

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25
Frequency (Hz) Frequency (Hz)

Fig. 4 Frequency-domain of vibration signal recorded by both smartphones, a New Sp. Fastest
Mode, b New Sp. Game Mode, c Old Sp. Fastest Mode, (d). Old Sp. Game Mode

Fig. 5 The rotor system with the smartphone mounted tightly to the bearing

signal spectrum. Also, envelope analysis is not useful in our case due to the lower
sampling frequency (Fs) of the smartphone accelerometer. Therefore, we confine to
perform a statistical evaluation.
The smartphone was mounted near the bearing to acquire vibration data in the
three directions (X, Y, and Z axes). Time-domain features, selected in Sect. 4, will
be extracted from the vibration data to examine the health of the two bearings where
the rotor is operating at 600 rpm (10 Hz).
1784 A. Ait Ben Ahmed et al.

0.1 0.8
(a) (b)
0.08
0.6

0.06
0.4
0.04

Acceleration (m/s 2)
Acceleration (m/s 2)

0.2
0.02

0 0

-0.02
-0.2
-0.04
-0.4
-0.06

-0.6
-0.08

-0.1 -0.8
0 5 10 15 0 5 10 15
Time (s) Time (s)

Fig. 6 Time-domain of smartphone vibration data measured in the vertical direction (Y-axis),
a Healthy bearing, b Faulty bearing

Table 5 Time-domain statistical features of vibration data (Old Sp., X-, Y-, and Z-axis)
Features Direction Healthy bearing Faulty bearing Ratio (Healthy/Faulty)
RMS X-axis 0.0479 0.0980 48.8%
Y-axis 0.0120 0.0956 12.5%
Z-axis 0.020 0.0447 44.7%
Kurtosis X-axis 3.8118 5.5676 68.4%
Y-axis 3.1493 15.299 20.5%
Z-axis 3.4848 5.6071 62.1%
Variance X-axis 0.0023 0.0096 23.9%
Y-axis 0.0001 0.0091 1.09%
Z-axis 0.0004 0.002 20.0%
Crest factor X-axis 3.5314 4.7288 74.6%
Y-axis 3.2845 7.7377 42.4%
Z-axis 2.6394 5.6562 46.6%

Figure 6 shows the Time-domain of vibration data recorded by the smartphone

measured in the vertical direction corresponding to the Y-axis coordinate of the
smartphone. As indicated in Fig. 6, the amplitude of vibration of the healthy bearing
is more attenuated compared to the vibration amplitude of the faulty bearing.
Table 5 summarizes the results of the statistical evaluation of the vibration data
for healthy and defective bearings, which are performed according to the directions
of the smartphone’s coordinate system (X, Y, and Z axes).
According to the ratio (healthy/defective), which represents the percentage of the
amount of the statistical feature of the healthy bearing compared to the defective
bearing, the RMS level, Kurtosis value, variance, and crest factor of the defective
bearing were increased in the three directions (X, Y, Z axes).
The Kurtosis value of a normal rolling element bearing is well recognized at 3.
Where, for a defective bearing, the probability density function becomes more peak,
Experimental Evaluation of MEMS Accelerometers … 1785

and the dispersion has also increased around the reference mean value, and the signal
becomes more impulsive.
In sum, for efficient monitoring, it is recommended to make a schedule to systemat-
ically acquire vibration data at predefined points and to establish graphs of the evolu-
tion as a function of time for each feature in order to monitor the bearing degradation
and to create a smartphone-based system for bearing condition monitoring.

6 Conclusion

The evaluation of the two different smartphone’s accelerometers leads us to indicate

that:
• Time-domain statistical features show good stability for most of the features, these
features can be used in condition monitoring.
• The frequency-domain show that the new smartphone has a clean and clear
spectrum than the old smartphone, these features can be exploited to extract
the necessary information from the vibration signal, and identify the vibration
phenomena.
• The difference between the smartphone’s results and the results of the professional
equipment can be justified by the low sampling frequency, and the poor resolution
of the sensor, as well as by the mounting of the sensor to the structure, as in this
study, the mounting is not adhesive, neither for both smartphones nor for the
professional equipment.
In summary, this study has evaluated the smartphone’s accelerometer performance
in vibration analysis to use it for educational purposes to overcome the lack of
measuring equipment and to help students to apply their theoretical knowledge in
vibration analysis at home. Also, it can be used for modest industrial purposes as
indicated in the case study presented in Sect. 5.

Acknowledgment Experimental equipment was provided by the Mechanical Engineering Labo-

ratory, Faculty of Science and Technology, University of Sidi Mohamed Ben Abdellah, Fez,
Morocco.

References

1. Looney M (2014) An introduction to MEMS vibration monitoring. Analog Dialogue 48(06):1–

3
2. Ait Ben Ahmed A, Touache A, El Hakimi A, Chamat A (2019) The possibility of using a
smartphone in a single plane rotor balancing. In: MATEC web of conferences. EDP Sciences,
p 03001. https://doi.org/10.1051/matecconf/201928603001
1786 A. Ait Ben Ahmed et al.

3. Ait Ben Ahmed A, Touache A, El Hakimi A, Chamat A (2020) A new hybrid method for rigid
and flexible rotor balancing without phase response measurements. Aust J Mech Eng. https://
doi.org/10.1080/14484846.2020.1842616
4. Feng M, Fukuda Y, Mizuta M, Ozer E (2015) Citizen sensors for SHM: Use of accelerometer
data from smartphones. Sensors 15(2):2980–2998. https://doi.org/10.3390/s150202980
5. Cerutti S, Magenes G, Bonato P (2010) Special section on smart wearable devices for human
health and protection. Institute of Electrical and Electronics Engineers. https://doi.org/10.1109/
TITB.2010.2048937
6. Google: Sensor—Android Developers. https://developer.android.com/guide/topics/sensors/
sensors_overview. Accessed 29 Nov 2020
7. Google: Sensor—Android Developers. https://developer.android.com/guide/topics/sensors/
sensors_motion. Accessed 11 Sept 2020
8. SensorDelay Enum (Android.Hardware)—Microsoft Docs. https://docs.microsoft.com/en-
us/dotnet/api/android.hardware.sensordelay?view=xamarin-android-sdk-9. Accessed 09 Dec
2020
9. Soong TT (2004) Fundamentals of probability and statistics for engineers. Wiley, Hoboken
10. Caesarendra W, Tjahjowidodo T (2017) A review of feature extraction methods in vibration-
based condition monitoring and its application for degradation trend estimation of low-speed
slew bearing. Machines 5(4):21
11. Chu E, George A (1999) Inside the FFT black box: serial and parallel fast fourier transform
algorithms. CRC Press, Boco Raton
MPPT Using Adaptive Genetic-Fuzzy
Logic Control for Wind Power System

Chakib Alaoui, Hajar Saikouk, and Anass Bakouri

Abstract This study proposes a fuzzy-based MPPT controller used to extract the
maximum power point (MPPT) from the WECS using doubly fed induction generator
(DFIG). The MPPT was accomplished by applying control to the generator speed
and blade pitch angle. The parameters of the FLC are optimized by using a genetic
algorithm (GA). Initial simulation results prove that the proposed MPPT algorithm
possesses good dynamic and steady state performances under different operating
conditions.

Keywords Wind turbine · Doubly fed induction generator (DFIG) · Fuzzy Logic
Controller (FLC) Genetic algorithm optimization

1 Introduction

Several challenges are still confronting the large-scale adoption of WECS (wind
energy conversion system). The most urgent challenge concerns the wind speed
variations at the various spots over the blades covered area. Thus, making direct
sensing of wind speed a challenging task [2, 9]. In addition, deploying mechanical
sensors rises the overall costs of WECS, while reducing its reliability and roughness
[2–4]. In order to meet these challenges, the author of [1] have proposed an estimation
of the actual wind speed indirectly by using electric indicators, such as output electric
power or generator shaft speed using a tachometer which are reasonably easy to
measure according to [4]. The authors of [5–7] combined these results with a dynamic
model of wind turbine. However, these dynamic equations are highly nonlinear and

C. Alaoui (B) · H. Saikouk · A. Bakouri

INSA EuroMediterannée, Euromed University, Fez, Morocco
e-mail: [email protected]
H. Saikouk
e-mail: [email protected]
A. Bakouri
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1787
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_162
1788 C. Alaoui et al.

Grid

Fig. 1 Wind turbine, doubly fed induction generator and power electronic converters block diagram

having hard to find parameters. For that reason, the wind speed is usually estimated
by using some kind system identification, data mining, state observer and/or some
advanced fuzzy techniques [1, 3].
Modern wind turbine systems include embedded controllers to predict the best
pitch angle for the wind turbine blades to operate under various wind speed scenarios.
This will serve to adjust the pitch angle. The objective of the presented project is
to implement a controller that will drive the pitch angle of the blades of the wind
turbine system to the preset pitch angle in order to regulate the input wind power.
Hence extracting the maximum allowable power without damaging the wind turbine
nor the electric generator.
In this project, we consider a wind power system comprising of a wind turbine,
a doubly fed induction generator (DFIG) connected to a three-phase grid through
appropriate AC-DC and DC-AC converters [8]. The proposed system is illustrated
in Fig. 1.
The wind turbine transforms the kinetic energy of the air particles, captured by the
blades, into mechanical power that rotates the electric generator. The static converters
transmit the electric power to the grid. This is accomplished by the drive train system,
rotor/stator of the induction machine and an AC-DC converter coupled with a DC-
AC converter through a capacitor. The control system reads the output power and
adjusts the pitch angle of the blade system of the wind turbine in order to regulate
the output power.
MPPT Using Adaptive Genetic-Fuzzy Logic Control … 1789

2 System Modeling and Implementation

2.1 Wind Turbine

The power accessible from the wind in a wind turbine system is expressed by (1).

P = 0.5ρ Aυ 3 (1)

where ρ is air density (1.225 kg/m3 ), A is the area swept by the blades (m2 ), and v
is wind speed (m/s). Figure 2 shows the wind power as a function of wind speed for
an uncontrolled system.
The power available in the air molecules cannot be completely transferred to a
mechanical power in the shaft of the turbine; it is rather limited by the Betz threshold
of 59%. This limit is expressed by the power coefficient C p of the wind turbine.
It is a function of the blade pitch angle (β) and the blade tip speed ratio (λ). The
mechanical power available at the shaft of the wind turbine becomes (2).

P t = 0.5ρ Aυ 3 C p (β, λ) (2)

where β is the blade pitch angle, C p is the power coefficient of the turbine, λ is the
blade tip to speed ratio (TSR). C p is a nonlinear function of β and λ and given in (3)
and (4).
 
116 −21
C p = 0.5176 − 0.4β − 5 e λi + 0.0068λ (3)
λi
1 1 0.035
= − (4)
λi λ + 0.08β 3β + 1

Fig. 2 Wind power as

function of wind speed
1790 C. Alaoui et al.

Cp

β
λ
Fig. 3 The power coefficient Cp -TSR curve for different pitch angle values

Fig. 4 Output generator power vs Wind speed

Finally, blade tip to speed ratio-TSR is defined in (5), where ω is the blade angular
velocity (rad/s) and R is the wind turbine’s blade radius (m).

ω t R
λ= (5)
υ
Figure 3 illustrates the nonlinear relationship between the blade tip speed ratio
(TSR) and the power coefficient (C p ), with the pitch angle (β).

2.2 Wind Turbine Pith Angle Regulator

Blade pitch regulation means to adjust the angle of attack of the blades of a wind
turbine rotor in order to control the electric power produced by the DFIG. The
main objective of the blade pitch control system is to keep the rotor speed within
MPPT Using Adaptive Genetic-Fuzzy Logic Control … 1791

the acceptable operation boundaries, as shown in Fig. 4. Figure 5 illustrates the

relationship of the mechanical shaft power, transmitted to the DFIG, to the wind
speed and the blades pith angle.
It can be seen from Fig. 5 that the maximum output is determinated when the
pitch angle is zero and thence as the pitch angle increases, as the value of the output
power decreases. Hereafter, it is necessary to control the pitch angle in order to attain
the MPP of the wind turbine.
By applying the TSR control, it is necessary to maintain the TSR to an optimum
value at which extracted power is maximized by controlling the rotational speed of
the DFIG [9–12].
This technique needs a precise measurement of the shaft rotational speed
combined with an estimation of the wind speed in order to get the optimum TSR

Fig. 5 Output power vs. turbine rotational speed for various pith angle values

Fig. 6 Tip to speed ratio based MPPT for wind turbine block diagram
1792 C. Alaoui et al.

Fig. 7 Fuzzy-based MPPT block diagram for a wind power system

(λopt ) of the turbine. The system would then be able to extract the maximum attain-
able power as shown in Fig. 6. The optimum rotational speed is determined as ωopt =
λopt V w /R.

2.3 Fuzzy Based MPPT Controller

Several control approaches [13, 14] have been used that use the Fuzzy Logic Control
(FLC) for WECS MPPT applications, as illustrated in Fig. 7. The foremost advantage
of such controllers is that their respective parameters can be modified rapidly in
response to sudden variations in the system dynamics. When the weather conditions
changes, the performance of a FLC based MPPT is robust. However, their robustness
depends heavily on the knowledge base (KB) and the experience of the operator
in choosing the rule base (RB), the membership functions and tolerable levels of
output errors. Additionally, the memory requisite poses additional limitations in
FLC application. This is the main reason to include a proper optimization algorithm,
such as genetic algorithm (GA), in order to optimize the fuzzy engine of the FLC,
and hence the adaptive FLC.

2.4 Adaptive Fuzzy Controller

The adaptive fuzzy controller system has many advantages when used for non-linear
system. Some of these advantages are:
• Automatically adapts it parameters to new values as the environmental settings
are highly dynamic.
• The parameters are altered to get the desired and fixed output values.
MPPT Using Adaptive Genetic-Fuzzy Logic Control … 1793

• Very recommended for non-linear systems, such as wind power systems.

• The values of the controller K p , K i and K d are constantly tuned by a proposed
adaptive fuzzy knowledge based controller (FKBC).
In addition to the FKBC, two more block are added; (1) process monitor and (2)
an adaptive mechanism.
The process monitor continuously monitors the system and detects any variations
in the process characteristics. It evaluates the performance of the controller in order
to estimate new values for the controller parameters K p , K d and K i . Some of the
parameters that are monitored are: the overshoot, rise-time, settling time, delay rate,
frequency of oscillation, error, absolute error, weighted absolute error, gain, phase
margin among others.
The adaptive mechanism uses the information passed to it by the performance
monitoring unit and updates the controller parameters in order to adapt it to the
targeted results.
The fuzzy parameters that are updated online are typically:
• Scaling factor
• Rules
• Shape of the membership function
• Number of cross points and ratios
• Number of term sets, and definition of rules.
Moreover, there are two types of adaptive FKBC, such as self-tuning controller,
in which the membership function and/or the scaling factor are altered, as illustrated
in Fig. 8 (most common) and self-organized controller, in which the rules are altered,
as illustrated in Fig. 9 (rarely used).
Scaling functions normalize the universes of discourse in which the fuzzy member-
ship functions are described. In a conventional fuzzy system, the scaling functions
are parameterized by a single scaling factor or a lower and/or upper limit for linear
scaling and a contraction/dilatation factor for non-linear scaling. However, in an
optimized fuzzy system, these parameters are adjusted in order to have the scaled
universe of discourse better matches the underlying variable range. This is illustrated
in Figs. 8. and 9.

3 Simulation Results and Discussion

The system shown in Fig. 10 was implemented in Matlab/Simulink environment and

simulated. The first block named ‘controlled system’ receives pulse-width-modulated
signal (PWM) from the fuzzy controller in order to control the power transferred to
the grid through the static converters AC-DC and DC-AC as was illustrated in Fig. 1.
The process parameters, also referred to as the state parameters, are I g , V g , ωm and Pm ,
which are the current and the voltage at the terminal of the generator, the rotational
speed and the mechanical power of the generator respectively. The performance of
1794 C. Alaoui et al.

Fig. 8 Altering the scaling function. Having the scaling function of 200, the fuzzy subset positive
P = [0 0.5 0.7], however, with the scaling function of 400, P = [0 0.25 0.35]

Fig. 9 Adapting the fuzzy set definition. The fuzzy subset P = [0 0.5 0.7] becomes P = [0 0.19
0.4]

the wind power system, represented by the ‘controlled system’ block, is evaluated,
and its output power is used as fitness value for the genetic algorithm (GA). The
GA block produces the optimized parameters for the fuzzy controller block that, in
combination with the state parameters, produces the PWM signal to the control the
converters. Fuzzy modeling of type Takagi–Sugeno was selected in this study since
accuracy was a preferred criterion for the MPPT system.
MPPT Using Adaptive Genetic-Fuzzy Logic Control … 1795

Fig. 10 System’s block diagram

The simulation was executed on a wind power system with specifications listed
in Table 1. The proposed genetic-fuzzy control system for MPPT was done by
Matlab/Simulink. The simulation results, shown in Fig. 11, compare the theoret-
ical values of the generator power and the shaft speed, with the actual output power
and rotational speed as obtained by the proposed method. The wind speed was set
to change from 6 to 8 m/s at t = 2.5 s and from 8 to 10 m/s at t = 4.5 s during the
Matlab simulation.
The Simulation results, as illustrated in Fig. 11, reveal that the proposed method
has good dynamic response, with steady state performances during the course of the
MPPT process.

Table 1 Wind power system:

Symbol Description Numerical value
turbine and generator
specifications R Turbine blade radius 1m
Cp(max) Maximum power coefficient 0.475
λopt Optimal TSR 8.12
Vw Rated wind speed, 10 m/s
ωm Rated rotor speed 1400 rpm
Ig Rated current 23.3 A
P Number of poles 8
Pg Rated power 3 kW
1796 C. Alaoui et al.

Fig. 11 Theoretical vs. simulated output power de the wind power system

4 Conclusion

In this project, a fuzzy logic controller (PLC) was proposed to wind turbine pitch
control in order obtain the MPPT throughout the wind speed values. A genetic-fuzzy
control-based for DFIG wind power system has been discussed and modeled. The
efficacy of the projected MPPT strategy has been proved by initial Matlab/Simulink
simulations.

References

1. Zhiqiang X, Qinghua H, Ehsani M (2012) Estimation of effective wind speed for fixed speed
wind turbines based on frequency domain data fusion. Sustain Energy IEEE Trans 3(1):57e64
2. Tan K, Islam S (2004) Optimum control strategies in energy conversion of PMSG wind turbine
system without mechanical sensors. Energy Convers IEEE Trans 19(2):392e9
3. Wei Q, Wei Z, Aller JM, Harley RG (2008) Wind speed estimation based sensorless output
maximization control for a wind turbine driving a DFIG. Power Electron IEEE Trans
23(3):1156e69
4. Meyer DG, Srinivasan S, Semrau G (2013) Dynamic wind estimation based control for small
wind turbines. Renew Energy 50:259e67
5. Kodama N, Matsuzaka T, Tuchiya K, Arinaga S (1999) Power variation control of a wind
generator by using feed forward control. Renew Energy 16(1):847e50
6. Van der Hooft E, Van Engelen T. Estimated wind speed feed forward control for wind turbine
operation optimisation. In: Conference estimated wind speed feed forward control for wind
turbine operation optimisation
7. Boukhezzar B, Siguerdidjane H (2009) Nonlinear control with wind estimation of a DFIG vari-
able speed wind turbine for power capture optimization. Energy Convers Manag 50(4):885e92
8. SimPowerSystems for Use with Simulink (2004) User’s Guide Version 4, MathWorks, Inc.
Natick, MA
MPPT Using Adaptive Genetic-Fuzzy Logic Control … 1797

9. Buehring IK, Freris LL (1981) Control policies for wind-energy conversion systems. IEE Proc
C Gener Transm Distrib 128:253
10. Abdullah MA, Yatim AHM, Tan CW, Saidur R (2012) A review of maximum power point
tracking algorithms for wind energy systems. Renew Sustain Energy Rev 16(5):3220–3227
11. Knight AM, Peters GE (2005) Simple wind energy controller for an expanded operating range.
IEEE Trans Energy Convers 20:459–466
12. De Kooning JD, Gevaert L, Vyver JVD, Vandoorn TL, Vandevelde L (2013) Online estimation
of the power coefficient versus tip-speed ratio curve of wind turbines. In: 39th annual conference
IEEE industrial electronics society, pp 1792–1797
13. Simoes MG, Bose BK, Spiegel RJ (1997) Fuzzy logic based intelligent control of a variable
speed cage machine wind generation system. IEEE Trans Power Electron 12:87–95
14. Hilloowala RM, Sharaf AM (1996) A rule-based fuzzy logic controller for a PWM inverter in
a stand-alone wind energy conversion scheme. IEEE Trans Ind Appl 32:57–65
Development of an Automated
Measurement Platform for an Electrical
Traction System Test Bench

Amine El Houre, Driss Yousfi, Zakaria Bourzouk, and Mohammed Chaker

Abstract In this paper, a test bench for electrical traction systems is presented
and discussed. A simulation model in Matlab/Simulink is proposed to bypass the
limitations imposed by the mentioned test bench. The developed Simulink model
allows the emulation of the electric vehicle performance for urban driving conditions
using ECE 15 drive cycle. An automated measurement platform is made, with the
aim of improving the test bench, by making it possible to reproduce the simulation
results with fidelity and ease.

Keywords Electric vehicle · Electrical drive · Test bench · Matlab/Simulink ·

DAQ LabJack · Graphical user interface · LabVIEW · Drive cycle

1 Introduction

Today’s global society is very concerned with reducing the negative effects of road
transportation on the environment due to toxic emissions and greenhouse gases.
Therefore, these types of vehicle emissions are legally regulated in multiple countries
across the world, in order to comply with the stricter regulations that are anticipated
to be adopted in the future, vehicle manufacturers are obliged to invest in various
fuel-saving technologies. This has resulted in an increased interest in the electrifi-
cation of vehicles, primarily hybrid electric vehicles (HEVs) which can reduce fuel
consumption compared to conventional vehicles, but also battery electric vehicles
(BEVs) [1, 2]. The latter offer high driveline efficiency and do not emit exhaust
gases, which is why they have been considered CO2 free in regulations to date [3, 4].

A. El Houre (B) · D. Yousfi · M. Chaker

ESETI Laboratory, National School of Applied Sciences, Mohamed First University,
Oujda, Morocco
Z. Bourzouk
ALTRAN, Casablanca, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1799
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_163
1800 A. El Houre et al.

Fig. 1 Experimental test bench

Fig. 2 Test bench structure

Electric vehicles are expected to replace regular vehicles in the near future [5]. It is
therefore very important to conduct research and train a new generation of engineers
in this field.
There are different methods for developing and studying electric vehicles, each
with its own advantages and disadvantages. One of the best-known methods is the
test bench based study. This approach allows researchers to evaluate the perfor-
mance of electric propulsion and develop new control strategies in closed scientific
research environments [6, 7]. In order to maximize the benefit of the test bench,
automation of test measurement and data’s graphical representation is necessary
since it reduces time and effort compared to manual testing, in addition to providing
increased productivity and reliability and increased test predictability [8].
In this paper an experimental test bench of an electric traction motor is presented,
then a Matlab/Simulink test bench simulation model is made and discussed in order
to emulate the performance of the electric vehicle’s traction motor for on-road driving
conditions. Afterwards, a platform for the automation of measurement and graphical
display of data generated from the electric traction motor test bench is built.

2 Experimental Test Bench

Figure 1 illustrates the experimental test bench concerned in this study, and Fig. 2
shows the structure of the experimental test bench. The test bench consists of a
Development of an Automated Measurement Platform … 1801

Fig. 3 Synoptic diagram of the drive cycle based test

10 kW Permanent-magnet (PM) motor, which is controlled efficiently using a vector

controller, this drive motor is connected to a 13 kW PM machine, which acts as a
dynamometer that feeds electrical power to a resistive load through a three phase
rectifier bridge.
The main purpose of this test bench is to evaluate the performance of the motor
that will be integrated into a lightweight electric vehicle. One major advantage of this
test bench is the possibility to estimate the torque applied on the drive motor’s shaft
without the need for a load cell. The estimation of this resistant torque is achieved by
knowing the current absorbed by the resistive load in addition to the torque constant
of the dynamometer. By varying the resistive load, it is possible to control the resistive
torque applied on the drive motor’s shaft, allowing us to emulate in a discrete manner
certain driving conditions on road.
The most recognized method of representing road driving conditions is by
employing standard drive cycles [9–11]. To emulate the electric drive vehicle perfor-
mance on these test cycles using a test bench, we would have to control the speed
of the drive motor and the torque of the dynamometer continuously and precisely in
order to follow the drive cycle and also to mimic the resistive forces acting on the
actual vehicle [6], which is not possible with the current version of the test bench
since the resistive load is varied manually meaning that the generated resistive torque
will not match the requirements of the vehicle on the drive cycle correctly.

3 Simulation Model

Figure 3 shows the synoptic diagram of the drive cycle based test that allows the
emulation of the vehicle performance on city driving conditions, and Fig. 4 demon-
strates the Simulink model created to reproduce the mentioned test. The simulation
idea of the test bench is based on Fajri’s work [6] with certain modifications to suit
1802 A. El Houre et al.

Fig. 4 Simulation model of the drive cycle based test

our case. The motor and dynamometer model parameters used in the simulation were
extracted experimentally in order to get closer to reality.
The drive cycle model generates speed and acceleration references corresponding
to the ECE 15 cycle that represents city driving conditions [12]. These references are
sent to the vehicle’s dynamic model, which describes all the forces exerted on the
vehicle, and calculates the speed and the torque references for the speed controlled
motor and the torque controlled dynamometer, respectively. A mechanical shaft
model is used to simulate the mechanical coupling of the two machines. The test
bench model generates the torque and speed data of the motor; this data will be used
to plot the torque/speed curve. Figure 5 illustrates the simulation results.
Figure 5(a) shows that the dynamometer motor follows almost perfectly the torque
reference, meaning that the torque regulation is working as intended. Figure 5(b)
indicates that the motor torque opposes the dynamometer motor torque with the
same amplitude. However, the torque peaks were caused by the sudden change of
the resistive torque, which affects the speed profile as well. Figure 5(c) confirms the
speed regulation of the motor since it’s almost the same as the reference.
In order to verify the validity of the current simulation model results, another
simulation method was used, which is in our case the Advanced Vehicle Simulator
(ADVISOR) [13].
A vehicle with almost the same specifications as our vehicle is designed and
simulated using ADVISOR. This simulated vehicle follows perfectly the speed profile
of the ECE 15 drive cycle.
Figure 6 indicates that the drive motor torque resulted from the simulation model
and ADVISOR on the ECE 15 drive cycle are nearly similar most of the time.
However, since the vehicle’s specifications for both simulations are not exactly the
same because of the limitations imposed by ADVISOR and the fact that our test bench
model doesn’t include regenerative braking as opposed to ADVISOR, the results of
the two simulations were therefore slightly distorted, particularly with regard to the
negative torques.
Development of an Automated Measurement Platform … 1803

Fig. 5 Simulation results for the drive cycle based test. a Dynamometer motor reference and
generated torque. b Drive motor torque. c Drive motor reference and generated rotational speed

Fig. 6 Drive motor torque comparison between our Simulink model and ADVISOR
1804 A. El Houre et al.

Fig. 7 Simulation result of the drive motor torque/speed data points corresponding to the ECE 15
drive cycle

These results show that our simulation model emulates the electric-drive vehicle
performance on ECE15 drive cycle correctly.
After verifying our simulation results using ADVISOR, the torque/speed data
points of the drive motor according to the ECE 15 cycle were plotted as shown in
Fig. 7.
This simulation model proved that the simulated drive motor based on the actual
motor parameters is able to achieve the whole ECE 15 operating points accurately.
So as to verify this result experimentally, the resistive load and the speed of the drive
motor should be manipulated in a way to reproduce the same operating points as
Fig. 7.
Another way for doing so would be to reproduce the experimental drive motor
acceleration performance and projecting it on the ECE 15 operating points, and if
these points fall within the acceleration torque/speed envelope, then we could say
that these ECE 15 operating points are achievable by the drive motor experimentally.
In both cases, the automation of measurement and graphical display of data gener-
ated from the test bench, would be a priority since we want to be able to follow the
performance of the test bench faithfully and with ease.

4 Automated Measurement Platform

Automation of measurement and data’s graphical representation reduces time and

effort compared to manual testing. It improves productivity, reliability, and increases
test predictability. Figure 8 demonstrates the structure of the made automated
measurement platform.
During the testing of the drive motor, the measurement circuit acquires the
measurand through the sensors. These sensors provide analog electrical signals, that
will be conditioned, in order to be suited for use by the DAQ device, which is the
LabJack U3. This device then processes these signals and sends them in a digital form
through the USB data bus, in order to be displayed and monitored in the LabVIEW
interface.
Development of an Automated Measurement Platform … 1805

Fig. 8 The structure of the automated measurement platform

4.1 Torque and Speed Measurement

The main purpose of the measurement circuit is measuring the required electrical
signals and conditioning them to be usable by the DAQ device.
The most important quantities in our test bench are: current, voltage, speed, torque,
power and efficiency. For the torque T, it’s estimated using the measured load current
and dynamometer torque constant, as described by Eq. (1) which was validated
experimentally. Concerning the speed N, it’s calculated based on the frequency of
the hall effect signal of either the motor or the dynamometer [14, 15], and the pole
number of one of the two machines, as determined by the relation (2).

T = kt × I (1)

Where kt is the torque constant in N.m/A of the dynamometer motor, and I is the
measured load current.
120
N= × f (2)
p

f is the frequency of the hall effect signals, and p is the pole number of the
machine.
As for the power and efficiency, they are deducted from the voltages, currents,
speed and torque.
The calculations of the quantities stated above are implemented in the LabVIEW
interface.

4.2 The LabVIEW Interface

Figure 9 illustrates the user-friendly LabVIEW interface that was created.

1806 A. El Houre et al.

The control and configuration section of the interface Fig. 9, allows the user to
control the execution of the program, to insert the necessary quantities to calculate
the torque and the speed.
It is possible, via this sheet, to save the test results in an Excel file and to display the
elapsed time of the test and errors that might occur during its execution. Figure 9(b)
indicates the XY representation of speed versus torque, source current vs torque,
mechanical power versus torque, efficiency versus torque, and finally the efficiency
map of the electrical drive. The dashboard section showed in Fig. 9(b) visualizes the
important quantities extracted from the test bench in real-time.

Fig. 9 The LabVIEW interface. a The control and configuration section of the interface. b The XY
data representation section. c The dashboard section
Development of an Automated Measurement Platform … 1807

5 Conclusion

In this paper, an experimental test bench was introduced and discussed. This test
bench presents a drawback in terms of the emulation of the electric vehicle perfor-
mance for on-road driving conditions. To overcome this, a simulation model was
proposed in the Matlab/Simulink environment and validated using ADVISOR. This
simulation contains a model of the test bench based on real experimental parameters
of the drive motor, dynamometer, and vehicle dynamics, which allows the repro-
duction of the drive cycle based test. The next objective would be reproducing the
data operating points resulted from this simulation using the actual test bench to
validate that the electrical drive is suitable for our application. In order to do so, an
automated measurement platform was developed to monitor and to follow faithfully
and effectively the performance of the test bench.

Acknowledgement This work was funded by ALTRAN Maroc, a Global Engineering and R&D
Services Company, and was supported by the Moroccan Research Institute for Solar Energy and
New Energies (IRESEN).

References

1. de Santiago J, Bernhoff H, Ekergård B, Eriksson S, Ferhatovic S, Waters R, Leijon M (2012)

Electrical motor drivelines in commercial all-electric vehicles: a review. IEEE Trans Veh
Technol 61:475–484
2. Ehsani M, Gao Y, Miller JM (2007) Hybrid electric vehicles: architecture and motor drives.
Proc IEEE 95:719–728
3. Regulation No 101 of the Economic Commission for Europe of the United Nations
(UN/ECE) (2012/5). https://eur-lex.europa.eu/LexUriServ/LexUriS-erv.do?uri=OJ:L:2012:
138:0001:0077:en:PDF. Accessed 29 Oct 2020
4. Grunditz E (2014) BEV powertrain component sizing with respect to performance, energy
consumption and driving patterns (2014)
5. Ehsani M, Gao Y, Gay S, Emadi A (2004) Modern electric, hybrid electric, and fuel cell
vehicles: fundamentals, theory, and design (2004)
6. Fajri P, Prabhala VAK, Ferdowsi M (2016) Emulating on-road operating conditions for electric-
drive propulsion systems. IEEE Trans Energy Convers 31:1–11
7. Marignetti F, D’Aguanno D, Volpe G (2017) Design and experiments of a test equipment for
hybrid and electric vehicle drivetrains. In: 2017 twelfth international conference on ecological
vehicles and renewable energies (EVER), pp 1–6
8. Washington C, Delgado S (2008) Improve design efficiency and test capabilities with HIL
simulation. In: 2008 IEEE AUTOTESTCON, pp 593–594
9. Lazari P, Wang J, Chen L (2014) A computationally efficient design technique for electric-
vehicle traction machines. IEEE Trans Ind Appl 50:3203–3213
10. Günther S, Ulbrich S, Hofmann W (2014) Driving cycle-based design optimization of interior
permanent magnet synchronous motor drives for electric vehicle application. Presented at the
2014 international symposium on power electronics, electrical drives, automation and motion,
SPEEDAM June 1 2014
11. Watson H (1978) Vehicle driving patterns and measurement methods for energy and emissions
assessment. Bureau of Transport Economics, Canberra
1808 A. El Houre et al.

12. A reference book of driving cycles for use in the measurement of road vehicle emissions:
version 3. IHS (2009)
13. Wipke K, Cuddy M, Burch S (1999) ADVISOR 2.1: a user-friendly advanced powertrain
simulation using a combined backward/forward approach. IEEE Trans Veh Technol 48:1751–
1761
14. Naveen V, Isha TB (2017) A low cost speed estimation technique for closed loop control
of BLDC motor drive. In: 2017 international conference on circuit power and computing
technologies (ICCPCT), pp 1–5
15. Dumitru D. Module 8: Speed Control - NXP Community, https://community.nxp.com/t5/NXP-
Model-Based-Design-Tools/Module-8-Speed-Control/m-p/727476. Accessed 29 Oct 2020
Simulation of a Guidance Law
on a ROS-Based Nonholonomic Mobile
Robot

Nabil Rezoug and Mokhtar Zerikat

Abstract In view of computational challenges that may span autonomous navigation

of mobile robots, mainly due to its complex structure, a low cost, simple structure
scheme is proposed and verified under Gazebo in a ROS-based nonholonomic mobile
robot “Turtlebot Kobuki”. Relying on LQI control strategy, the algorithm ensures the
task of steering the robot through a set of waypoints allowing high performance and
accurate path tracking. Simulating a virtual robot under Gazebo, experimental results
demonstrate the effectiveness of the proposed scheme.

Keywords ROS · Gazebo world · Turtlebot Kobuki · Waypoint guidance · LQI

control · Trajectory tracking

1 Introduction

Wheeled mobile robots play a major role in both industrial and service robotics,
particularly when autonomous motion capabilities are required, however developing
a fully autonomous robot that can navigate in unknown environments is problem-
atic due to difficulties that span dynamics modeling, on-board perception, trajectory
generation, and optimal control, not to mention the heavy computational bill that may
results from such methods. Although several studies denounce limitations that are
imposed by the structural obstruction of the kinematic model, for instance designing
feedback laws, little attention has been drawn to address these obstacles. This paper
develops a waypoint trajectory tracking method that demonstrates the possibility to
overcome the limitation imposed by the kinematic model. One of its most attractive
features is computational simplicity, and the ability to control the accuracy of gener-
ated paths by changing the number of waypoints. in the last decade, Mobile robot
path planning has drawn lot of interest from researchers in deep learning, however,
as reported by [1] real-time processing require a large neural network with multiple
parameters, which represent a computational burden, even for modern CPU, in [2]

N. Rezoug (B) · M. Zerikat

Ecole Nationale Polytechnique Maurice Audin, Oran, Algeria

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1809
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_164
1810 N. Rezoug and M. Zerikat

they used fuzzy PID controller for autonomous motion of a differential drive robot,
in order to solve the path and time optimization problem they developed in [3] a four-
layer Neural network as a controller, in [4] they used as an experimental platform a
car like mobile robot where they focused on designing an effective neural network
controller. By using linearized model predictive control (LMPC) in [5] they dealt
with the system physical constraints, therefore, overcoming the limitations caused
by nonholonomic constraints of mobile robot. In [6] time varying state feedback
control law based on back-stepping technic was implemented on the Turtlebot to test
the performance of the studied strategy in [7] they consider the problem of robust
adaptive trajectory tracking control for a mobile robot dealing with system uncer-
tainties. In order to achieve robust tracking two control strategies were used in [8] a
proportional integral control and sliding mode control, in order to simplify the struc-
ture of the control scheme, a new kinematic control approach was used for position
tracking. Path following is basically dealing with the problem of designing a guidance
law that steers the robot to follow a desired path, based on the approach presented
in [9], and with the sole purpose of avoiding any latency related problem that may
occur from computational complexity, we aim in this paper to design guidance law,
in order to simplify the structure of the control system.
The reminder of this paper is organized as follows: in Sect. 2 a thorough description
of the experimental test bed is presented, Sect. 3 presents the model of the robot and
outline the problem that motivated this work, Sect. 4 describe the strategy used
to control the robot, Sect. 5 shows the simulation results and comparison, Sect. 6
contains the conclusion and future work.

2 Experimental Testbed

To assess the framework reliability, we used for practical experimentation a ROS-

enabled nonholonomic mobile robot, Turtlebot Kobuki 2. It is the new version of
the Turtlebot robot series, powered by Willow Garage and ClearPath Robotics, more
details are provided in [10].
As illustrated in Fig. 1, the robot is equipped with a Kinect sensor, a notebook and
a Kobuki base. It is a suitable mobile platform for developing robot applications. The
robot’s base has two identical no deformable rear wheels each rear wheel is powered
by a motor and equipped with an encoder, the robot also contains proximity sensors
and gyro meter for each axis [10].
To conduct our experiments, Robot Operating System (ROS) is chosen as the main
computing and management unit of the wheeled robot Turtlebot Kobuki. Figure 2
is an overview of the software structure; it provides a simplified illustration on how
different structure interact with each other. Robot Operating System (ROS) handles
sensors and actuators data flow in the system. In the presence of a running ROS master
in our platform, we initialize a MATLAB ROS subsystem in another computer with
the IP address and port number of the ROS master, allowing real time monitoring of
the robot position and velocities [11].
Simulation of a Guidance Law on a ROS-Based Nonholonomic Mobile … 1811

Fig. 1 Experimental platform Turtlebot Kobuki

Fig. 2 Experimental control scheme

3 Robot Model and Path Following Problem Formulation

Typically, the guidance law allows an accurate path tracking in two-dimension frame.
In this section, we briefly review the kinematic model of a differential drive mobile
robot, and present the concept of waypoint guidance method which provides a
rigorous formulation of the problem of steering the differential drive mobile robot
along a desired path.

3.1 Kinematic Model

The kinematic model for nonholonomic ground robot is given by:

1812 N. Rezoug and M. Zerikat

ẋr = Va cos(ψ),
ẏr = Va sin(ψ), (1)
ψ̇r = ω;

Where xr andyr denote the robot position in the fixed frame, and ψ it’s orientation.
Va is the linear velocity, and ω is the angular velocity.

3.2 Waypoint Guidance Strategy

Because of its ability to move in challenging environments, wheeled mobile robots

present a great medium for bringing sensors or tools to specified locations, it is
important to mention that the exact path leading to the target location has little
interest as long as the robot reaches its destination within reasonable amount of time.
In this section we propose a guidance law that drives the robot through reference
locations.
Waypoint guidance strategy drives the robot through each waypoint in a consec-
utive order by taking the shortest possible path from one waypoint to the next, it is
assumed that the waypoint has been reached once the robot is within an acceptance
radius enclosing the current waypoint, for simplifying purposes, we consider the
following assumptions.
• The robot’s path is built by joining a set of reference locations called waypoints.
in this paper we consider a set of waypoints Pk , where k{1, . . . , n}
• Basically we are steering the robot into each of the waypoints P1 , P2 , . . . , Pn in
consecutive order.
• Orientation of the robot around the z axis is defined relative to the North axis
• The proximity distance is defined as an acceptance circle around the waypoint
with radius R = 0.01 m.
• Linear speed is set to 0.3 m/s.
The desired path is defined by the straight-line from the precedent waypoint
Pk (xk ; yk ) to the destination waypoint Pk+1 (xk+1 ; yk+1 ), where (xk ; yk ) and
(xk+1 ; yk+1 ) are respectively the coordinates of the waypoints Pk and Pk+1 in the
inertial frame [9], as shown in Fig. 3.
d is the shortest distance from the position of the robot to the reference path, given
by
   
d = −(xr − xk ) sin ψ p + (yr − yk )cos ψ p (2)

where ψ p is the orientation of the path relative x axe direction, defined by

 
yk+1 − yk
ψ p = tan−1 (3)
xk+1 − xk
Simulation of a Guidance Law on a ROS-Based Nonholonomic Mobile … 1813

Fig. 3 Waypoint following (N: North, E: East)

ψe is the orientation of the robot relative to the desired path

ψe = ψ p − ψ (4)

By differentiating (2) with respect to time and using (1), it follows that

ḋ = −Va sin(ψe ) (5)

ψ̇e = ω (6)

The considered state vector x = [d, ψe ]T

4 Control System Design

In the following we illustrate the functional structure of the control system design
(Fig. 4).
The control law is given by

u(t) = −K x x(t) − K y ∫ y(t)dt (7)

 
With a gain matrix K 1×3 = K x , K y chosen so that the closed loop system is
stable, any classical linear control technique can be applied to ensure convergence
of the controlled robot toward the planned path. To this end, we first derive the linear
model.
1814 N. Rezoug and M. Zerikat

Fig. 4 Schematic diagram of the proposed controller

The model used for linear path is as follow:

ẋ = f (x(t), u(t))

The input u is the angular speed ω, our objective is to design a control law u(t)
that stabilizes the system and consequently steers the Turtlebot to the reference path.
 T
For the equilibrium point xeq = deq , 0 , u eq = 0, where deq is arbitrary, we
obtain a linearized state space model of the system as follows

x = A p x + bpu (8)

Where
 
0 −Va 0
Ap = , bp = (9)
0 0 −1

The regulated output y is the cross-track error d, i.e., the output vector is
 
c= 10 (10)

To provide better robustness against disturbances, the integral of the regulated

output error, denoted e I , is considered for the linear system.
Where

e I = yg − y (11)

yg denotes the reference output. Since the system output is the distance from the
robot to the waypoint reference position, naturally we set yg = 0 in order for the
robot heading direction to converge toward the target location.

e I = 0 − cx (12)
Simulation of a Guidance Law on a ROS-Based Nonholonomic Mobile … 1815

The augmented system is written as follows

Ẋ = AX + Bu

   
ẋ A p 02×1 x b
= + u (13)
ė I −c 0 eI 0

Where A(3 × 3), B(3 × 1) and X 3×1 = [d, ψe , e I ]T

In our paper we relied on Linear Quadratic Regulator, to provide a practical
feedback gain in order to minimize the quadratic cost:

+∞ 
J = ∫ X T (t)Q X (t) + Ru 2 (t) dt (14)
0

Q(3 × 3) and R(1 × 1) are the state weighting matrix and the control weighting
matrix, respectively. As shown in [12], using the state feedback control law (15) will
result in the minimization of the cost function (14)

u(t) = −K X (t) = − R −1 B T P X (t) (15)

P(3 × 3) is the symmetric positive definite solution of the continuous algebraic

Riccati equation.

A T P + P A − P B R −1 B T P + Q = 0 (16)

5 Experimental Results

To show the effectiveness and accurate tracking capabilities of the Turtlebot using
the guidance law presented in this paper, we used two kind of reference trajectories
generated in a 2D frame, square trajectory and circular trajectory. The simulations
were carried out on the Turtlebot Kobuki under Gazebo with parameters as follows.
Diameter: 351.5 mm; Height: 124.8 mm; Weight: 2.35 kg.
The simulation was conducted under GAZEBO to assess the effectiveness of the
trajectory tracking controller, results of the simulation are shown in Figs. 5 and 8.
The environment GAZEBO was used for simulating a virtual “Turtlebot” seeking
to track a squared and circular trajectory. Test results are discussed in the following.
1816 N. Rezoug and M. Zerikat

Fig. 5 Square trajectory diagram of the Turtlebot

The robot initially positioned at the origin (0, 0), is first required to reach the
closest point on the desired trajectory and then follow the given waypoints.
The square trajectory is located at (2, 2); (4, 2); (4, 4); (2, 4). Figure 5 shows
that the robot tracks the given waypoints with a negligible error around the edges of
the square trajectory, the error is mainly due to the imposed range tolerance, which
basically define the exact reach the robot has to get within in order to move to the
next waypoint.
Next, we show the results as presented in [13], where the robot seems to give a
satisfactory result under MATLAB environment, as shown in Fig. 6, however the
simulation result under gazebo world does not follow the desired trajectory Fig. 7.
The results output in [13] can be explained by the fact that under gazebo world
various dynamic properties are taken into account, while under MATLAB simulation
many of these dynamics are not considered. Which justifies the efficiency of the
adopted control law in MATLAB simulation, and the poor performance under Gazebo
world.
Simulation of a Guidance Law on a ROS-Based Nonholonomic Mobile … 1817

Fig. 6 Circular trajectory diagram of the Turtlebot under MATLAB as shown in [13]

Fig. 7 Circular trajectory diagram of the Turtlebot under Gazebo as shown in [13]

The circle is centered in (2, 2), we can see in Fig. 8 that the robot started tracking
with an initial error, which is natural due to the large initial heading error, so basically
there was a time of transition in which the tracking error of heading direction was
stabilized.
1818 N. Rezoug and M. Zerikat

Fig. 8 Circular trajectory diagram of the Turtlebot

6 Conclusion

In this paper, a waypoint guidance law was developed for differential-drive wheeled
mobile robots. An accurate tracking is obtained from various reference trajectories,
3D simulation was developed in Gazebo world using MATLAB-Simulink for moni-
toring and data processing. In future work, we plan to test the proposed scheme in
a real robot, adding more challenging tasks, for instance the difficulty to identify
external dynamics, which will result in a limited dynamic compensation, hence the
necessity for an additional control scheme to compensate the dynamic properties
in order to have more accurate tracking performance. We will aim to achieve the
goal of composite nonlinear control and guarantee that the output tracking error will
ultimately converge to zero. In order to further reduce the oscillation amplitude and
frequency, friction compensation schemes can be incorporated, an extra control level
will focus on improving the robot’s behavior such as obstacle avoidance. To sum up,
this paper lay the ground work for future projects with the sole purpose of improving
the performance of ground robots.
Simulation of a Guidance Law on a ROS-Based Nonholonomic Mobile … 1819

References

1. Vanhoucke V, Senior A, Mao MZ (2011) Improving the speed of neural networks on CPUs.
In: Deep learning and unsupervised feature learning, NIPS
2. Heikkinen J, Minav T, Stotckaia AD (2017) Self-tuning parameter fuzzy PID controller for
autonomous differential drive mobile robot. In: 2017 XX IEEE international conference on
soft computing and measurements (SCM), pp 382–385. https://doi.org/10.1109/SCM.2017.
7970592
3. Parhi DR, Singh MK (2009) Real-time navigational control of mobile robots using an artificial
neural network. Proc Inst Mech Eng Part C: J Mech Eng Sci 223(7):1713–1725
4. Dewi T, Risma P, Oktarina Y, Roseno MT (2017) Neural network controller design for a mobile
robot navigation; a case study. In: 2017 4th international conference on electrical engineering,
computer science and informatics (EECSI)
5. Fetter Lages, W, Augusto Vasconcelos Alves J (2006) Real-time control of a mobile robot using
linearized model predictive control. In: Proceedings Volumes, vol 39, no 16, pp 968–973. ISSN
1474-6670, ISBN 9783902661173
6. Besseghieur, K.L., Tr˛ebiński, R., Kaczmarek, W., Panasiuk, J.: Trajectory tracking control for
a nonholonomic mobile robot under ROS. In: Journal of physics: conference series, volume
1016, 6th international conference on mechatronics and control engineering (ICMCE 2017)
7. Xin L, Wang Q, She J, Li Y (2016) Robust adaptive tracking control of wheeled mobile robot.
Robot Auton Syst 78:36–48 ISSN 0921-8890
8. Becerra HM, Colunga J, Romero J (2018) Simultaneous convergence of position and orientation
of wheeled mobile robots using trajectory planning and robust controllers. Int J Adv Rob Syst
15:172988141875457. https://doi.org/10.1177/1729881418754574
9. Souanef T (2019) Adaptive guidance and control of small unmanned aerial vehicles. https://
doi.org/10.13140/RG.2.2.21153.56166
10. Kobuki Documentation: relase 2.0, Dabit industries
11. Corke P (2015) Integrating ROS and MATLAB [ROS Topics]. IEEE Robot Autom Mag 22:18–
20
12. Zak, SH (2003) Systems and control. vol 198. Oxford University Press, Oxford
13. Bensaci C, Zenir Y, Pomorski D (2017) Control of mobile robot navigation under the
virtual world matlab-gazebo. In: A special issue of the international conference on advanced
Engineering in Petrochemical Industry (ICAEPI’17), Skikda Algeria, vol 2, no 4
Speed Sensorless Direct Torque Control
of Doubly Fed Induction Motor Using
Model Reference Adaptive System

Mohammed El Mahfoud, Badre Bossoufi, Najib El Ouanjli, Said Mahfoud,

Mourad Yessef, and Mohammed Taoussi

Abstract This study emphasis on a direct torque control (DTC) strategy with a speed
estimator utilizing the Model Reference Adaptive System (MRAS) for a doubly fed
induction machine (DFIM). The suggested adaptive reference model uses the values
obtained as the disparity between two voltage-to-current ratios along the d and q axes
for a hypothetical resistance quantity. The resulting sensorless formulation is totally
independent of any expression of stator/rotor resistance. And thereby, DTC command
brings numerous benefits over traditional control strategies like field-oriented control
(FOC) due to its clear architecture, quick dynamic response and less system param-
eter dependency. The suggested control algorithm is simulated and evaluated in
MATLAB/Simulink for a small speed range program.

Keywords DTC · MRAS · DFIM · Speed sensorless

1 Introduction

The DFIM is one of the most common high-power devices, particularly for the fields
of wind and hydropower generation [1]. It is also used in systems of attraction, marine
drives and electric and hybrid cars [2]. But the difficulty of these calculations and
the combination of torque and flux cause many difficulty.
In order to solve these issues, some researchers are focused on designing rigorous
control strategies. Among the first methods, the field orientation control technique
offers a decoupling of flux and torque. This covers a diverse range of modula-
tion velocity, quick torque reactions and high efficiency for several variations in
load [3]. Despite the appealing features of the FOC control, its execution involves
several coordinate transformations and a reliable configuration of a variety of PI

M. El Mahfoud (B) · B. Bossoufi · M. Yessef

Laboratory of Engineering Modeling and Systems Analysis, SMBA University, Fez, Morocco
e-mail: [email protected]
N. El Ouanjli · S. Mahfoud · M. Taoussi
Technologies and Industrial Services Laboratory, SMBA University, Fez, Morocco

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1821
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2_165
1822 M. E. Mahfoud et al.

controllers, and requires an external pulse width modulator (PWM) [1]. To solve
the problems of FOC control, TAKAHASHI implemented Direct Torque Control
in the early 1980s [4]. In comparison to the FOC, the DTC approach focused on
hysteresis comparators and turn tables provides many advantages such as easy execu-
tion, limited reliance on parameters, without any need for current controllers and
several coordinate transformations [5].
While in most industrial uses, speed and/or position detection is needed for robust
and significant command. As a consequence, DFIM speed regulation using the DTC
uses a speed sensor for its application, this speed sensor [6]:
• Diminishes machine performance;
• Raises the cost of the control device and is challenging to install
• Diminished reliability;
• Extra area for assembly,
• The wire.
That is why speed estimation methods are an amazing option. Literature has
proposed various methods such as Luenberger observer, Kalman filter, slipping mode
control, Model Reference Adaptive Systems (MRAS) and artificial intelligence tech-
niques [6, 7]. Of all the techniques, the MRAS Observer for Sensorless DFIM drives
is well known and most desirable due to its flexibility, simple physical perception
and tolerance to variations in system parameters [8].
The MRAS methodology therefore involves many forms of methods based on the
parameter selected for the estimate. Among these methods, the R-MRAS estimator
uses a resistance value corresponding to the voltage/current ratio that provides high
independence and high efficiency [9]. The goal of this paper is to combine the benefits
of DTC control with the efficiency of the MRAS approach for speed control and
estimation.
This paper is organised as follows: Sect. 2 introduces a DFIM modelling. Section 3
addresses traditional DTC approach based on switching tables and hysteresis
controllers. Section 4 addresses speed and flux measurement requirements. Section 5
discusses and interprets simulation curves using MATLAB/SIMULINK environ-
ment. Finally, Sect. 6 summarises the conclusions and recommendations for future
work.

2 Modeling of the DFIM

The DFIM model for DTC is the two-phase system represented in (α, β) coordinates.
The complexity of (a, b, c) representation of the motor is decreased by the Concordia
transformation [5]. Figure 1 indicates the spatial structure of these various frames of
reference.
The angle θ r represents the position of the rotor in relation to the reference axis
of the stator, while the angles θ and θ s show the relative positions of the direct axis
respectively with respect to the rotor and stator axes.
Speed Sensorless Direct Torque Control … 1823

β d (Direct axis)
(Quadrature axis) ω
q a (Rotor)
ω

θ
θ
θ α 0 A
O (Stator)

Fig. 1 Spatial position of the different reference frames

With:

θs = θr + θ (1)

Hence the autopilot equation is expressed by:

dθs dθr dθ
= + ⇒ ωs = ωr + ω (2)
dt dt dt
The DFIM stator and rotor voltage can be represented in (α, β) reference frame
[10]:
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤ ⎡ ⎤
vsα Rs 0 0 0 i sα ψsα 0
⎢v ⎥ ⎢ ⎥ ⎢i ⎥ d ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ Rs 0 0
sβ 0 ⎥ ⎢ sβ ⎥ ⎢ ψsβ ⎥ dθ ⎢ 0 ⎥
⎢ ⎥=⎢ ⎥.⎢ ⎥ + .⎢ ⎥+ .⎢ ⎥ (3)
⎣ vr α ⎦ ⎣ 0 0 Rr 0 ⎦ ⎣ ir α ⎦ dt ⎣ ψr α ⎦ dt ⎣ −ψrβ ⎦
vrβ 0 0 0 Rr irβ ψrβ ψr α

The expression of stator and rotor flux in diphase reference frame (α, β):
⎡ ⎤ ⎡ ⎤ ⎡ i sα ⎤
ψsα Ls 0 M. cos θ −M. sin θ
⎢ψ ⎥ ⎢ ⎢i ⎥
⎢ sβ ⎥ ⎢ 0 Ls M. sin θ M. cos θ ⎥⎥.⎢ sβ ⎥
⎢ ⎥= ⎢ ⎥ (4)
⎣ ψr α ⎦ ⎣ M. cos θ M. sin θ Lr 0 ⎦ ⎣ ir α ⎦
ψrβ −M. sin θ M. cos θ 0 Lr irβ

The electromagnetic torque can be determined from:

Tem = p.(ψsα .i sβ − ψsβ .i sα ) (5)

d
With Tem = TL + J. + f. (6)
dt
1824 M. E. Mahfoud et al.

3 Direct Torque Control Strategy Applied to DFIM

Takahashi proposes the DTC technique for electric motors to address field-oriented
control problems and particularly the sensitivity to variation in machine parameters.
DTC has a lot of advantages, such as reduced dependence on outside disruptions and
machine parameters, simple operation and a relatively quicker response of dynamic
torque to other controls [5]. The power of flux and torque in the stationary structure (α,
β) is decoupled. The switching table is used to choose a suitable voltage vector [10].
The choice of switching states is specifically linked to flux and torque variations.
The electromagnetic torque and flux amplitudes in hysteresis regulator bands are
thus reduced. These regulators control these two dimensions separately. Hysteresis
controller inputs are flux and torque errors and their performance defines the neces-
sary voltage vector to strike the inverter at each switching cycle. The inverter supplies
eight voltage vectors, two of which are null and void. These vectors are chosen from
a switching table that was calculated with the flux, torque and flux vector positions
[11]. The distribution of these vectors is seen in Fig. 2 (α, β).
The flux error εψ is introduced to a two-levels hysteresis comparator, which
produces the binary output (Hψ = 0,1) describing the ideal flux progression. The
torque error εTem is also inserted into a three-levels hysteresis controller, which
produces the variable (HTem = −1, 0, 1) at its output, which corresponds to the
path of the torque creation required.
The switching table provided in Table 1 is drawn up by the variables HTem and
Hψ division as well as the sector presenting details on flux vector locations.

Fig. 2 Voltage vectors

delivered by the inverters
Speed Sensorless Direct Torque Control … 1825

Table 1 Switching table

4 Design of Speed and Flux Estimators

4.1 Speed Estimation by Model Reference Adaptive System

The MRAS Theory is the most popular method used to observe machine parameters
and states using only the measurements of stator and rotor voltage and current [7].
The MRAS observer system consists of two systems with separate mechanisms for
estimating the same state variable depending on the separate inputs [12]:
• The first technique is focused on the equations in the stationary frame (α, β), being
independent of the speed to be estimated; It’s the reference model.
• the adaptive model is the second structure, which relies on the value to be
estimated, is described in the (d,q) reference (Fig. 3).
The values extracted from the models are compared and the variance is used to
operate an adjustment process generating the value to be computed. Centered on
this error, the adjustment mechanism generates the speed value calculated by the PI
controllerIf the algorithm is constructed right, the steady state would go to 0. [6].
Figure 4 reflects.
MRAS observer design.
• Reference model:
vαr vβr
R∗ = − (7)
i αr i βr

• Adaptive model:

vdr vqr
R= ∗ − ∗ (8)
i dr i qr
1826 M. E. Mahfoud et al.

Reference
Model

Comparator PI

α,β
d,q Adaptive R
Model

Integrator

Fig. 3 Structure of the MRAS observer

4.2 Flux and Torque Estimation

In DTC, depending on the errors in the control variables, the switching tables indi-
cate the positions of the inverters switches to generate the three-phase voltages that
attack the DFIM. The discrepancy between approximate variables and their refer-
ences measures these errors. Therefore, flux and torque accuracy is also very critical
to ensure satisfactory performance. Many variables must be defined, the stator and
the rotor current calculated, while the rotor and stator voltages may depend on the
states (Sa , Sb and Sc ).
The electromagnetic torque, rotor flux and stator flux are estimated by following
equations [5]:
  

T em = p.(ψ sα .i sβ − ψ sβ .i sα ) (9)

ψ̂r = ψ̂r2α + ψ̂rβ
2
(10)

ψ̂s = 2 + ψ̂ 2
ψ̂sα sβ (11)

The positions of the rotor and stator flux is calculated from:

ψ̂rβ
θr = ar ctg( ) (12)
ψ̂r α
ψ̂sβ
θs = ar ctg( ) (13)
ψ̂sα

The block diagram for the calculation of flux and torque is illustrated in Fig. 4.
Speed Sensorless Direct Torque Control … 1827

Rotor flux
estimation

Stator flux
estimation

Torque T
estimation

Fig. 4 Block flux and torque calculation diagram

5 Simulation Results and Discussion

Several numerical simulation series are implemented and realized on the

MATLAB/Simulink to evaluate the DFIM performance, robustness and stability.
The main characteristics of this simulation are:
• DC bus voltages: E s = 300 V et E r = 100 V.
• Application of a nominal load (T L = 10 N.m) at t = 1.8 s.
• The sampling frequency f s = 10 kHz.
• The DFIM parameters are listed in Table 2.
The following figure show the simulation results for Step speed of 157 rad/s
(equivalent 1500 rpm) (Fig. 5).

Table 2 Parameters of the DFIM

1828 M. E. Mahfoud et al.

Fig. 5 Simulation results

According to these results, the rotational speed follows the reference value
perfectly and quickly, It attains the reference value after tr = 0.28 s without over-
shooting and static errors at the start. On the other hand, the electromagnetic torque
leads with very rapid dynamics the set reference. At runtime, the torque is maximum
and maintained in constant state at nearly value 0. Also, Where a load torque appli-
cation at t = 1.8 s, a speed change of 1.8 rad/s occurs the time to reject this disruption
is quite quickly equivalent to 30 ms, which illustrates the effectiveness of this regu-
lation technique against changes in load torque. But, the electromagnetic torque has
ripples of the range of Tem = 2.7 N.m. Furthermore, the evolution of each flux
represents a circular course with a constant radius identical to the reference.
Speed Sensorless Direct Torque Control … 1829

The stator and rotor current aspects are sinusoidal in shape and react well to
the changes caused by the load torque. It can also be remembered that the currents
frequencies are equivalent to the autopilot equation’s rotational speed. Finally, the
positions flux are perfectly respected.

6 Conclusion

This paper has presented a speed sensorless DTC scheme for DFIM powered by two
voltage inverters. The goal is to boost traditional DTC efficiency. The hierarchical
DFIM model and the DTC strategy theory were introduced. The control technique
and the speed and flux calculation are comprehensive. After device controls have
been checked with MATLAB/SIMULINK the major results are as below.:
• The MRAS approach can be applied successfully in DFIM drives.
• The fast response, torque dynamics and robustness of the classic DTC are
maintained.
The future studies will concentrate on experimental DTC-MRAS control valida-
tion utilising the dSPACE DS1104 platform.

References

1. Elmahfoud M, Bossoufi B, Taoussi M, El Ouanjli N, Derouich A (2020) Comparative study

between backstepping adaptive and field oriented controls for doubly fed induction motor. Eur
J Electr Eng 22(3):209–221
2. Taoussi M, Karim M, Bossoufi B, Hammoumi D, Lagrioui A, Derouich A (2016) Speed variable
adaptive backstepping control of the doubly-fed induction machine drive. Int J Autom Control
10(1):12–33
3. Elmahfoud M, Bossoufi B, Taoussi M, El Ouanjli N, Derouich A (2019) Rotor field oriented
control of doubly fed induction motor. In: 2019 international conference on optimization and
applications, ICOA 2019
4. Abu-Rub H, Stando D, Kazmierkowski MP (2013) Simple speed sensorless DTC-SVM scheme
for induction motor drives. Bull Polish Acad Sci: Tech Sci 61(2):301–307
5. El Ouanjli N, Motahhir S, Derouich A, El Ghzizal A, Chebabhi A, Taoussi M (2019)
Improved DTC strategy of doubly fed induction motor using fuzzy logic controller. Energy
Rep 5(February):271–279
6. Korzonek M, Tarchala G, Orlowska-Kowalska T (2019) A review on MRAS-type speed
estimators for reliable and efficient induction motor drives. ISA Trans 93:1–13
7. Horch M, Boumediene A, Baghli L (2017) MRAS-based sensorless speed integral backstepping
control for induction machine, using a flux backstepping observer. Int J Power Electron Drive
Syst 8(4):1650–1662
8. El Daoudi S, Lazrak L, El Ouanjli N, Ait Lafkih M (2020) Improved DTC-SPWM strategy
of induction motor by using five-level POD-PWM inverter and MRASSF estimator. Int J Dyn
Control (0123456789)
9. Kumar R, Das S (2017) MRAS-based speed estimation of grid-connected doubly fed induction
machine drive. IET Power Electron 10(7):726–737
1830 M. E. Mahfoud et al.

10. El Ouanjli N, Derouich A, El Ghzizal A, Motahhir S, Chebabhi A, El Mourabit Y, Taoussi M

(2019) Modern improvement techniques of direct torque control for induction motor drives-a
review. Prot Control Mod Power Syst 4(1)
11. Ammar A, Bourek A, Benakcha A (2017) Nonlinear SVM-DTC for induction motor drive using
input-output feedback linearization and high order sliding mode control. ISA Trans 67:428–442
12. Giribabu D, Srivastava SP, Pathak MK (2019) Modified reference model for rotor flux-based
MRAS speed observer using neural network controller. IETE J Res 65(1):80–95
Author Index

A Alaoui, Chakib, 1327, 1585, 1787

A.Monem, O., 1433 Alhazmi, Omar H., 355
Aarab, Abdellah, 123 Alhejaili, Anwar D., 355
Ababssi, Najib, 1223 Alidrissi, Youssef, 279
Abadlia, Hadjer, 1745 Amane, Meryem, 663
Abbou, Ahmed, 155, 1387, 1675 Ambrosetti, Damien, 13
Abd Elsadek, Ahmed, 1433 Amgoune, Hafida, 457
Abdelhadi, Elmoudden, 1245 Amieur, Toufik, 1477, 1699
Abdoun, Farah, 947 Amine, Aouatif, 47, 1101
Abdoun, Otman, 585, 947, 967, 1073 Amine, El Hachimi Mohamed, 741
Abenna, Said, 189 Amrane, Said, 1735
Abou El Hassan, Adil, 259 Amraqui, Samir, 1205, 1489
Aboubakr, Herouane, 1191 Anand, Sampurn, 813
Abouchabaka, Jâafar, 1073 Annaki, Ihababdelbasset, 233
Abounada, Abdelouahed, 1213 Aouragh, Si Lhoussain, 199
Abouricha, Mostafa, 1735 Arsalane, Nabil, 1039
Abouzahir, Mohamed, 399 Arsalane, Sara, 1039
Achchab, Sanaa, 991 Asaidi, Hakima, 543
Achouyab, El Hassan, 1467 Attoui, Issam, 1147, 1171
Adithya, V., 731 Ayad, Hassan, 1643
Aditya, S., 707 Azami, Nawfel, 1735
Agoumi, Ali, 609 Azdoud, Y., 47
Ahmed, Taqi, 741 Azizi, Mostafa, 515
Aicha, Wahabi, 1245 Azizi, Yassine, 515
Aissaoui, Karima, 663 Azzouazi, Mohamed, 781
Ait Belaid, Khaoula, 1643
Ait Ben Ahmed, Abdelbaset, 1777 B
Ait Lahcen, Ayoub, 527 Badri, Abdelmajid, 749, 1365
Ait Nouh, Fatima, 1497 Bahij, Mouad, 165
Aitelkadi, Kenza, 25 Bajit, Abderrahim, 189
Ajaamoum, Mohamed, 1181 Bakhouya, Mohamed, 279
Ajallouda, Lahbib, 335 Bakouri, Anass, 1787
Ajjaj, Souad, 991, 1021 Bakouri, Saber, 25
Akpoviroro, Eric Obar, 959 Balanuta, Daniel Ciprian, 1353
Alami Aroussi, Hala, 1445, 1455 Balboul, Younes, 1081

© The Editor(s) (if applicable) and The Author(s), under exclusive license 1831
to Springer Nature Switzerland AG 2021
S. Motahhir and B. Bossoufi (eds.), Digital Technologies and Applications,
Lecture Notes in Networks and Systems 211,
https://doi.org/10.1007/978-3-030-73882-2
1832 Author Index

Bannari, Rachid, 3 Chadli, Sara, 311, 323

Barakat, M. R., 1409 Chafi, Mohammed, 447
Barara, Mohamed, 1409 Chafiq, Tarik, 111
Bawany, Narmeen Zakaria, 469 Chahid, Yassine, 291
Bechouat, Mohcene, 1477, 1699 Chaker, Mohammed, 1799
Bekkali, Moulhime El, 1081 Chalh, Zakaria, 825
Belahrach, Hassan, 1091, 1643 Chama, Jaride, 741
Belarhzal, Samya, 37 Chamat, Abderrahim, 1777
Belattar, Sara, 585 Chaouni, Samia Benabdellah, 907
Belebrik, Mohamed, 25 Charroud, Mohammed, 1111
Belkacem, Ghania, 1409 Chatri, Chakib, 165
Bellagi, Ahmed, 1505 Chebak, Ahmed, 1377
Bellouki, Mohamed, 543 Cheddadi, Fatima, 301
Benabdellah, Mohammed, 291 Cheddadi, Hafsa, 301
Benazzi, Abdelhamid, 847 Cheddadi, Youssef, 301
Benbrahim, Mohammed, 345, 759 Chennoufi, Khalid, 1159
Bendali, Wadie, 1305 Cherkaoui, Mohamed, 165, 1549, 1561
Benhaim, Mickael, 1497 Chetouani, Elmostafa, 1711
Benhlima, Laila, 1061 Chillali, Sara, 835
Bennouna, Fatima, 825 Chinta, Mukesh, 939
Benomar, Mohammed Lamine, 13 Chojaa, Hamid, 1235, 1279
Bensegueni, Skander, 245 Chrita, Hanae, 597
Benslimane, Mohamed, 871 Chtaina, Noureddine, 25
Bentaleb, Asmae, 527
Berdai, Abdelmajid, 1421 D
Berrada, Mohammed, 99, 663 Dabou, Rachid, 1147
Berrajaa, Achraf, 77 Dahmane, Kaoutar, 1181
Berrich, Jamal, 233 Dahmani, Soufiane, 1003
Berthoz, Alain, 233 Davis, Peter, 719
Bossoufi, Badre, 387, 1255, 1279, 1291, 1445, Dbaghi, Yahya, 1399
1455, 1725, 1821 Deepak, Gerard, 145, 271, 555, 565, 707, 731,
Bouachrine, Brahim, 1181 791, 801, 813
Bouakkaz, Mohammed Salah, 1147, 1171 Derouich, Aziz, 421, 1235, 1687
Bouchaala, Abdel Djabar, 1171 Descombes, Xavier, 13
Boudebbouz, Omar, 1171 Didi, Salah-Eddine, 1081
Bouderbala, Manale, 1445, 1455 Didi, Zaidan, 365
Boudi, El Mostapha, 37 Djekidel, R., 1529
Bouju, Alain, 99 Dusa, Alexandru, 1353
Boukadoum, Ahcene, 1171
Boulaiche, Mehdi, 1135 E
Bouraiou, Ahmed, 1147, 1171 Echchelh, Adil, 1341
Bourhaleb, Mohammed, 233 Eid, Ahmad, 65
Bourzouk, Zakaria, 1799 El Aamery, Salma, 1467
Bousalem, Zakaria, 209, 221 El Abkari, El Hassan, 377, 1051
Boussetta, Mohammed, 1305 El Abkari, Safae, 377, 1051
Boutaghlaline, Anas, 1619 El Akkad, Nabil, 495, 871, 893
Boutasseta, Nadir, 1147, 1171 El Alami, Rachid, 433
Bouzi, Mostafa, 1377 El Alaoui, Mustapha, 433
Bouzoubaa, Karim, 199 El Allali, Naoufal, 543
El Asri, Hayat, 1061
C El Azami, Ikram, 365
Chaatouf, Dounia, 1205, 1489 El Azzab, Driss, 1111
Chabani, Zakariya, 505 El Bakkali, Chakib, 1291
Author Index 1833

El Bekkali, Chakib, 1255 Essakhi, Hassan, 1399

El Bekkali, Moulhime, 1665 Ettifouri, El Hassane, 77
El Bouzekri El Idrissi, Younès, 527 Ez-Zaki, Fatima, 1091, 1643
El Faddouli, Nour-eddine, 621
El Filali, Sanaa, 335
F
EL Garouani, Said, 177
Fadi, Ouafia, 155, 1675
El Gettafi, Mohammed, 1111
Fagroud, Fatima Zahra, 335
El Guabassi, Inssaf, 209, 221
Faham, Hamza, 781
El Habib Daho, Mostafa, 981
Fahmani, Lamyae, 859
El Hajjami, Salma, 99
Faquir, Sanaa, 57
El Hakimi, Abdelhadi, 1777
Farhat, Sadik, 1399
El Hamdani, Wafae, 1665
Fariss, Mourad, 543
El Hmamsy, Youssef, 1517
Faska, Zahra, 893
El Houre, Amine, 1799
Fatima, Cheddadi, 1317
El Houssaini, Mohammed-Alamine, 991, 1021
Fatima-Zohra, Hibbi, 643
El Houssaini, Souad, 991, 1021
Fattah, Mohammed, 1081, 1665
El idrissi, Rafika, 1387
Ferfra, Mohammed, 1159
El Idrissi El-Bouzaidi, Youssra, 967
Ferro, Alexandre Castilla, 233
El Kasmi Alaoui, My Seddiq, 781
Ferroud, Chorouk, 675
El Khadiri, Karim, 433, 1607, 1619, 1631,
1655
El Mahfoud, Mohammed, 1687, 1821 G
EL Majdoubi, Omayma, 947 Gabli, Mohammed, 1003
El Makoussi, Nadia, 1033 Gaizen, Soufiane, 155, 1675
El Mehdi, Abdelmalek, 259 Garfaf, Jamila, 859
El Merrassi, Weam, 1213 Garouani, Moncef, 597
El Mhamdi, Jamal, 377, 1051 Ghammaz, Abdelilah, 1091
El Ouanjli, Najib, 1687, 1821 Gouiouez, Mounir, 575
El Ouardi, Abdelhafid, 411 Gourma, Anass, 1421
El-Attar, A., 47 Guerouate, Fatima, 685
Elbakkali, Outhman, 1341 Gurguiatu, Gelu, 1353
Elboukhari, Mohamed, 515
Elhaj, Nabil, 1409 H
ELkafazi, Ismail, 3 Habib, Mubbra, 925
Elkamoun, Najib, 279 Hachimi, Hayat, 111
Elkhatir, Haimoudi, 643 Haddouch, Khalid, 893
Elkhatiri, Ali, 1549, 1561 Hafian, Asmae, 345
El-Kishky, Hassan, 65 Hain, Mustapha, 135, 1021
Elmouatamid, Abdellatif, 279 Hajjaji, Abdelowahed, 1517
Elmoudden, Aicha, 1399 Hajji, B., 1595
Elouardighi, Abdeljalil, 675, 695 Halkhams, Imane, 1081, 1665
Eloudrhiri Hassani, Abdelhadi, 749 Hamdi, Hassan, 1497
Elsayed, Yasser, 1433 Hamouche, Salima, 505
Emharraf, Mohamed, 311 Hanafi, Ahmed, 421
Ennawaoui, Chouaib, 1517 Hashmi, Mehmood Ahmed Husnain, 925
Ennouni, Assia, 123 Hassani, Moha M’Rabet, 1013
Errai, Mohamed, 1735 Himi, Mahjoub, 1111
Errahimi, Fatima, 301 Houam, Lotfi, 769
Errami, Youssef, 1711 Houicher, S., 1529
Ershadi, Mahmoud, 719 Hussain, Saddam, 925
Essaid, Mohamed, 847 Hussien, Mokhtar, 1433
1834 Author Index

I Lagouir, Marouane, 1365

Ibnyaich, Saida, 1013 Lagriou, Ahmed, 1455
Ikram, Saber, 1317 Lagrioui, Ahmed, 387, 1279
Iqbal, Sana, 469 Lahfaoui, Badreddine, 1757
Lahmar, El Habib Ben, 335
J Lakhdar, Abdelghani, 1765
Jabrane, Oussama, 1111 Lakhliai, Zakia, 1655
Jabrane, Younes, 919, 1033 Lakhssassi, Ahmed, 1607, 1631
Jadli, Aissam, 135 Lakrit, Soufian, 165
Jai Andaloussi, Said, 907 Lamghari, Said, 1497
Jaize, Abderrahman, 135 Lasfar, Abdelali, 655
Jalti, F., 1595 Latif, Rachid, 411
Jammoukh, Mustapha, 1765 Lazouni, Mohammed El Amine, 981
Janan, Mourad Taha, 1191 Lemmassi, Assiya, 421
Jarjar, Abdeltif, 847 Lhayani, Mouna, 1561
Jebbor, Fatine, 1061 Livinti, Petru, 1353
Jefferies, Marcus, 719 Loubna, Atarsia, 1267
Jeghal, Adil, 177 Loucif, Mourad, 1725
Jennane, Rachid, 769 Loulijat, Azeddine, 1223
Jihani, Nassima, 759
Jilbab, Abdelilah, 377, 1051 M
M’dioud, Meriem, 3
K Maaroufi, Mohammed Mouhcine, 609
Kabbaj, Mohammed Nabil, 345, 759 Mabrouki, Mustapha, 1573
Kahla, Sami, 1477, 1699 Machhour, Naoufal, 881
Kaissari, Soufiane, 1051 Maghfour, Mohcine, 695
Kannouf, Nabil, 291 Mahfoud, Said, 1687, 1821
Kaoutar, Senhaji Rhazi, 1245 Mahmoudi, Sidi Ahmed, 981
Karim, Mohammed, 421, 1255 Majid, Muhammad Sakandar, 925
Karima, El hammoumi, 1317 Majout, Btissam, 1255
Karra, Rachid, 655 Makhad, Mohamed, 1223
Karrouchi, Mohammed, 87, 1539 Malik, Saba, 925
Kassmi, Kamal, 87, 1539 Malki, Jamal, 99
Kerrakchou, Imane, 311, 323 Mamri, Ayoub, 399
Khabba, Asma, 1013 Manaswini, S., 145
Khaidar, Mohammed, 279 Mansouri, Khalifa, 1765
Khaldoun, Mehdi, 1497 Mansouri, Merieme, 907
Khalil, Sara, 1549 Marah, Rim, 209, 221
Khankhour, Hala, 1073 Matar, Zaki, 1433
Kharbach, Amina, 323 Mazer, Said, 1081, 1665
Kharroubi, Jamal, 597 Mazri, Tomader, 457, 633
Khatir, Haimoudi El, 585 Mazwa, Khawla, 633
Khemaissia, Seddik, 769 Mbarki, A., 1595
Khouani, Amin, 981 Mebarkia, Meriem, 769
Khrissi, Lahbib, 893 Mechernene, Abdelkader, 1725
Koumina, Abdelaziz, 919, 1033 Mediouni, Mohamed, 1399
Kourchi, Mustapha, 1181 Medromi, Hicham, 859
Mehdaoui, Youness, 433
L Menssouri, Aicha, 1631
Laababid, Younes, 1607 Meraoumia, Abdallah, 769
Laabidine, Nada Zine, 1291 Mermri, El Bekkaye, 1003
Laaroussi, Houria, 685 Mermri, El Bekkaye, 1003
Laayati, Oussama, 1377 Messaoudi, Abdelhafid, 87, 1539
Labbadi, Moussa, 165, 1549, 1561 Mezrhab, Ahmed, 1205, 1489
Author Index 1835

Mohammed, Boussetta, 1317 Redjimi, Kenza, 1135

Mohammed, Karim, 1291 Redjimi, Mohammed, 1135
Mojtahedi, Mohammad, 719 Reha, Abdelati, 919, 1033
Mokhlis, Mohcine, 1387 Rezoug, Nabil, 1809
Morchid, Abdennabi, 433 Riad, Toufouti, 1267
Mostafa, Harti, 99 Rifai, Nada, 1585
Motahhir, Saad, 1147, 1171 Rifi, Mounir, 1039
Moumen, Aziz, 1765 Rithish, Harish, 791
Mounich, Kamilia, 447 Rohini, Valaparla, 939
Mourad, Youssef, 1305
Moussaoui, Hanae, 871 S
Moutaouakil, Abdelhakim, 919, 1033 Saâfi, Ikram, 1505
Muhil Aditya, P., 707 Saber, Ikram, 301, 1305
Saber, Mohammed, 259, 311, 323
N Sabir, Zakaria, 1101
Nahid, Mohammed, 189 Sabor, Jalal, 1585
Najihi, Ikrame, 1517 Sabri, My Abdelouahed, 123
Nasri, Ismail, 87, 1539 Saddik, Amine, 411
Nasri, M’barek, 881 Saeed, Muhammad Tariq, 925
Nasser, Naoual, 1735 Sahel, Aicha, 749
Nassih, B., 47 Sahnoun, Smail, 1711
Neçaibia, Ammar, 1147 Said, Raed, 505
Ngadi, M., 47 Saikouk, Hajar, 1327, 1787
Nisar, Muhammad, 925 Sajid, Muhammad, 925
Nouaiti, Ayoub, 1421 Salhi, Mourad, 1205, 1489
Nouh, Saïd, 781 Salima, Meziane, 1267
Santhanavijayan, A., 145, 271, 555, 565, 707,
O 731, 791, 801, 813
Obbadi, Abdellatif, 1711 Satori, Khalid, 495
Otman, Abdoun, 643 Sayouti, Yassine, 1365
Ouacha, Ali, 1121 Sbihi, Mohamed, 399, 685
Ouajji, Hassan, 1467 Sebbaq, Hanane, 621
Oubella, Mhand, 1181 Sedraoui, Moussa, 1477, 1699
Oughdir, Lahcen, 835 Senhaji, Saloua, 57
Ouladsine, Radouane, 279 Serag, Saif, 1341
Oumidou, Naima, 1549, 1561 Serghini, Abdelhafid, 1003
Outzourhit, Abdelkader, 1497 Settouti, Nesma, 13
Shahzad, Shahbaz Ahmad, 925
Q Sihamman, Noura Ouled, 123
Qazdar, Aimad, 209, 221 Smitha Chowdary, Ch., 939
Qjidaa, Hassan, 433, 1607, 1619, 1631, 1655 Snoussi, Hajar, 87, 1539
Qobbi, Younes, 847 Sobhana, M., 939
Stour, Laila, 609
R Surya, Deepak, 555, 801
Rabbah, Nabila, 959
Rachid, El Bachtiri, 1317 T
Rahali, Ahmed, 1655 Tahiri, Ahmed, 1619, 1631
Rahmoune, Mohammed, 233 Taibi, Djamel, 1607, 1477, 1699, 1655
Raillani, Benyounes, 1205, 1489 Taieb, Ahmed, 1505
Raji, Mohammed, 111 Taleb, Thami Ait, 1191
Ramzi, Mohamed, 1213 Taouni, Abderrahim, 1387
Ramzi, Mustapha, 399 Taoussi, Mohammed, 1235, 1279, 1455, 1687,
Razouki, Hassan, 483 1821
Reddak, Moussa, 1421 Tilioua, Narjiss, 825
1836 Author Index

Timmi, Mohamed, 177 X

Tmiri, Amal, 991 Xiao, Rudan, 13
Tnaji, Khalid, 199
Touache, Abdelhamid, 1777
Touairi, Souad, 1573 Y
Toufouti, Riad, 1745 Yahyaouy, Ali, 57, 177
Touil, Hamza, 495 Yessef, Mourad, 1235, 1279, 1821
Toumi, Hicham, 335 Yousfi, Driss, 1181, 1799
Touzani, Halima Drissi, 57

U Z
ul Huda, Anwaar, 925 Zahidi, Abdallah, 1735
Zahra, Syeda Anam, 925
V Zamzoum, Othmane, 1235
Varghese, Levin, 271 Zaoui, Mohamed, 233
Venna, Deepa, 939 Zellou, Ahmed, 335
Vijaykumar, Chaya, 813 Zerikat, Mokhtar, 1809
Zerraf, Soufiane, 111
W Ziane, Abderrezzaq, 1147
Wadie, Bendali, 1317 Ziani, El Mostafa, 1445
Wahabi, Aicha, 447 Zine-Dine, Khalid, 279
Wakrim, Layla, 1013 Zulfiqar, Ayesha, 469