Advances in Intelligent Systems and Computing 1193

Tawfik Masrour
Anass Cherrafi
Ibtissam El Hassani Editors

Artificial
Intelligence
and Industrial
Applications
Smart Operation Management
Advances in Intelligent Systems and Computing

Volume 1193

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

Advisory Editors
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
Rafael Bello Perez, Faculty of Mathematics, Physics and Computing,
Universidad Central de Las Villas, Santa Clara, Cuba
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
Hani Hagras, School of Computer Science and Electronic Engineering,
University of Essex, Colchester, UK
László T. Kóczy, Department of Automation, Széchenyi István University,
Gyor, Hungary
Vladik Kreinovich, Department of Computer Science, University of Texas
at El Paso, El Paso, TX, USA
Chin-Teng Lin, Department of Electrical Engineering, National Chiao
Tung University, Hsinchu, Taiwan
Jie Lu, Faculty of Engineering and Information Technology,
University of Technology Sydney, Sydney, NSW, Australia
Patricia Melin, Graduate Program of Computer Science, Tijuana Institute
of Technology, Tijuana, Mexico
Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro,
Rio de Janeiro, Brazil
Ngoc Thanh Nguyen , Faculty of Computer Science and Management,
Wrocław University of Technology, Wrocław, Poland
Jun Wang, Department of Mechanical and Automation Engineering,
The Chinese University of Hong Kong, Shatin, Hong Kong
The series “Advances in Intelligent Systems and Computing” contains publications
on theory, applications, and design methods of Intelligent Systems and Intelligent
Computing. Virtually all disciplines such as engineering, natural sciences, computer
and information science, ICT, economics, business, e-commerce, environment,
healthcare, life science are covered. The list of topics spans all the areas of modern
intelligent systems and computing such as: computational intelligence, soft comput-
ing including neural networks, fuzzy systems, evolutionary computing and the fusion
of these paradigms, social intelligence, ambient intelligence, computational neuro-
science, artificial life, virtual worlds and society, cognitive science and systems,
Perception and Vision, DNA and immune based systems, self-organizing and
adaptive systems, e-Learning and teaching, human-centered and human-centric
computing, recommender systems, intelligent control, robotics and mechatronics
including human-machine teaming, knowledge-based paradigms, learning para-
digms, machine ethics, intelligent data analysis, knowledge management, intelligent
agents, intelligent decision making and support, intelligent network security, trust
management, interactive entertainment, Web intelligence and multimedia.
The publications within “Advances in Intelligent Systems and Computing” are
primarily proceedings of important conferences, symposia and congresses. They
cover significant recent developments in the field, both of a foundational and
applicable character. An important characteristic feature of the series is the short
publication time and world-wide distribution. This permits a rapid and broad
dissemination of research results.
** Indexing: The books of this series are submitted to ISI Proceedings,
EI-Compendex, DBLP, SCOPUS, Google Scholar and Springerlink **

More information about this series at http://www.springer.com/series/11156

Tawfik Masrour Anass Cherrafi
• •

Ibtissam El Hassani
Editors

Artificial Intelligence
and Industrial Applications
Smart Operation Management

123
Editors
Tawfik Masrour Anass Cherrafi
Laboratory of Mathematical Modeling, Laboratory of Mathematical Modeling,
Simulation and Smart Systems (L2M3S), Simulation and Smart Systems (L2M3S),
Department of Mathematics and Computer Department of Industrial and Manufacturing
Science, Artificial Intelligence for Engineering, Modeling and Optimization of
Engineering Sciences Team (IASI) Industrial Systems Team (MOSI)
National Graduate School for Arts National Graduate School for Arts
and Crafts, My Ismail University and Crafts, My Ismail University
Meknes, Morocco Meknes, Morocco

Ibtissam El Hassani
Laboratory of Mathematical Modeling,
Simulation and Smart Systems (L2M3S),
Department of Industrial and Manufacturing
Engineering, Artificial Intelligence for
Engineering Sciences Team (IASI)
National Graduate School for Arts
and Crafts, My Ismail University
Meknes, Morocco

ISSN 2194-5357 ISSN 2194-5365 (electronic)

Advances in Intelligent Systems and Computing
ISBN 978-3-030-51185-2 ISBN 978-3-030-51186-9 (eBook)
https://doi.org/10.1007/978-3-030-51186-9
© Springer Nature Switzerland AG 2021
This work is subject to copyright. All rights are reserved 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
Organization

General Chair

Tawfik Masrour, Department of Mathematics and Computer Science, Artificial

Intelligence for Engineering Sciences Team (IASI), Laboratory of Mathematical
Modeling, Simulation and Smart Systems (L2M3S), ENSAM, My Ismail University,
50500 Meknes, Morocco. [email protected] [email protected]

Co-chair

Vincenzo Piuri, Department of Computer Science, University of Milan via Celoria

18, 20133 Milano (MI), Italy

Keynotes Speakers

Amal El Fallah Seghrouchni, Sorbonne University, Paris, France

Mawa Chafii, ENSEA, CY Paris University, Paris, France
Abdellatif Benabdellah, University of Le Havre, France
Ali Siadat, Arts et Métiers Paris Tech Metz, France
Jiju Antony, Heriot-Watt University, Edinburgh, Scotland
Andrew Kusiak, University of Iowa, USA

v
vi Organization

TPC Chairs

Alexandre Dolgui, France

Jose Arturo Garza-Reyes, UK
Abby Ghobadian, UK
Kannan Govindan, Denmark
Janusz Kacprzyk, Poland
Vikas Kumar, UK
Ali Siadat, France

International Scientific Committee

Aalaoui Zinab, Morocco

Abawajy Jemal H., Australia
Aboulaich Rajae, Morocco
Aghezzaf El-Houssaine, Belgium
Ahmadi Abdeslam, Morocco
Ait Moussa Abdelaziz, Morocco
Akhrif Iatimad, Morocco
Aksasse Brahim, Morocco
Al-Mubaid Hisham, USA
Alali Abdelhakim, Morocco
Alami Hassani Aicha, Morocco
Ali Ahad, USA
Ali Siadat, France
Allahverdi Ali, Kuwait
Aly Ayman A., Saudi Arabia
Arbaoui abdelaziz, Morocco
Azizi Abdelmalek, Morocco
Azzeddine MAZROUI, Morocco
Babai Mohamed Zied, France
Badie Kambiz, Iran
Balas Valentina Emilia, Romania
Bansal Jagdish Chand, India
Batouche Mohamed Chawki, Saudi Arabia
Behja hicham, Morocco
Belhadi Amine, Morocco
Ben Abdllah Mohammed, Morocco
Benabbou Rajaa, Morocco
Benaissa Mounir, Tunisia
Benghabrit Asmaa, Morocco
Benoussa Rachid, Morocco
Organization vii

Berrada Mohamed, Morocco

Bouhaddou Imane, Morocco
Brouri Adil, Morocco
Buccafurri Francesco, Italy
Carrabs Francesco, Italy
Castillo Oscar, Mexico
Cerulli Raffaele, Italy
Chaouni Benabdellah Abla, Morocco
Chaouni Benabdellah Naoual, Morocco
Charkaoui Abdelkabir, Morocco
Chbihi Louhdi Mohammed Reda, Morocco
Cherrafi Anass, Morocco
Ciaramella Angelo, Italy
Ciasullo Maria Vincenza, Italy
D’Ambrosio Ciriaco, Italy
Daoudi El-Mostafa, Morocco
De Mello Rodrigo Fernandes, Brazil
Deep Kusum, India
Dolgui Alexandre, France
Ducange Pietro, Italy
El Akili Charaf, Morocco
El Haddadi Anass, Morocco
El Hammoumi Mohammed, Morocco
Ibtissam El Hassani, Morocco
El Jai Mustapha, Morocco
El Jasouli Sidi Yasser, Belgium
El Mazzroui Abb Elaziz, Morocco
El Mghouchi Youness, Morocco
El Ossmani Mustapha, Morocco
Elbaz Jamal, Morocco
Elfezazi Said, Morocco
Elmazroui Azz Addin, Morocco
ES-SBAI Najia, Morocco
Ettifouri El Hassane, Morocco
Ezziyyani Mostafa, Morocco
Faquir Sanaa, Morocco
Fassi Fihri Abdelkader
Fiore Ugo, Italy
Fouad Mohammed Amine, Morocco
Gabli mohamed, Morocco
Gaga Ahmed, Morocco
Gao Xiao-Zhi, Finland
Garza-Reyes Jose Arturo, UK
Ghobadian Abby, UK
Giuseppe Stecca, Italy
viii Organization

Govindan Kannan, Denmark

Grabot Bernard, France
Hajji Tarik, Morocco
Hamzane Ibrahim, Morocco
Harchli Fidaa, Morocco
Hasnaoui Moulay Lahcen, Morocco
Herrera-Viedma Enrique, Spain
Itahriouan Zakaria, Morocco
Jaara El Miloud, Morocco
Jaouad Kharbach, Morocco
Jawab Fouad, Morocco
Kacprzyk Janusz, Poland
Kaya Sid Ali Kamel, Morocco
Khadija Bouzaachane, Morocco
Khireddine Mohamed Salah, Morocco
Khrbach Jawad, Morocco
Kodad Mohssin, Morocco
Krause Paul, UK
Kumar Vikas, UK
Laaroussi Ahmed, Morocco
Lagrioui Ahmed, Morocco
Lasri Larbi, Morocco
Lazraq Aziz, Morocco
Lebbar Maria, Morocco
Leung Henry, Canada
Manssouri Imad, Morocco
Marcelloni Francesco, Italy
Massoud Hassania, Morocco
Medarhri Ibtissam, Morocco
Mkhida Abdelhak, Morocco
Mohiuddin Muhammad, Canada
Moumen Aniss, Morocco
Moussi Mohamed, Morocco
Najib khalid, Morocco
Nee Andrew Y. C., Morocco
Nfaoui Elhabib, Morocco
Nguyen Ngoc Thanh, Poland
Nouari Mohammed, France
Noureddine Boutammachte, Morocco
Novák Vilém, Czech
Ouazzani Jamil, Morocco
Ouerdi Noura, Morocco
Oztemel Ercan, Turkey
Palmieri Francesco, Italy
Pesch Erwin, Germany
Organization ix

Pincemin Sandrine, France

Rachidi Youssef, Morocco
Rahmani Amir Masoud, Iran
Raiconi Andrea, Italy
Rocha-Lona Luis, Mexico
Saadi Adil, Morocco
Sabor Jalal, Morocco
Sachenko Anatoliy, Ukraine
Sael Nawal, Morocco
Saidou Noureddine, Morocco
Sekkat Souhail, Morocco
Senhaji Salwa, Morocco
Serrhini Simohammed, Morocco
Sheta Alaa, USA
Siarry Patrick, France
Soulhi Aziz, Morocco
Staiano Antonino, Italy
Tahiri Ahmed, Morocco
Tarnowska Katarzyna, USA
Tyshchenko Oleksii K., Czech
Tzung-Pei Hong, Taiwan
Zemmouri Elmoukhtar, Morocco
Zéraï Mourad, Tunisia

Local Organizing Committee

Abou El Majd Badr, Med V University

Ahmadi Abdessalam, ENSAM
Benabbou Rajaa, ENSEM
Benghabrit Asmaa, ENSMR
Benghabrit Youssef, ENSAM
Bouayad Aboubakr, ENSAM
Bouhaddou Imane, ENSAM
Chaira Abdellatif, Fac Sci, Meknes
Cherrafi Anass, ENSAM
Ibtissam El Hassani, ENSAM
Hajji Tarij, ENSAM
Masrour Tawfik, ENSAM
Najib Khalid, ENSMR
Saadi Adil, ENSAM
Sekkat Souhail, ENSAM
Zemmouri El Moukhtar, ENSAM
x Organization

Publication Chairs

Badr Abou El Majd, Morocco

Ibtissam El Hassani, Morocco
Ercan Oztemel, Turkey

Poster Chairs

Benabbou Rajaa, Morocco

Souhail Sekkat, Morocco

Registration Chairs

Abdessalam Ahmadi, Morocco

Adil Saadi, Morocco
Jalal Sabor, Morocco

Web Chairs

Ibtissam El Hassani, Morocco

Zemmouri El Moukhtar, Morocco

Public Relations Chairs

Abla Benabdellah Chaouni, Morocco

Asmaa Benghabrit, Morocco
Youssef Benghabrit, Morocco
Imane Bouhaddou, Morocco
Said Ettaqi, Morocco

Industrial Session Chairs

Anass Cherrafi, Morocco

Souhail Sekkat, Morocco
Organization xi

Ph.D. Organizing Committee

Amhraoui ElMehdi, Morocco

Benabdellah Chaouni Abla, Morocco
Eddamiri Siham, Morocco
El Mazgualdi Choumicha, Morocco
El Mekaoui Fatima Zahra, Morocco
Fattah Zakaria, Morocco
Hadi Hajar, Morocco
Jounaidi Ilyass, Morocco
Khdoudi Abdelmoula, Morocco
Moufaddal Meryam, Morocco
Raaidi Safaa, Morocco
Rhazzaf Mohamed, Morocco
Zekhnini Kamar, Morocco
Preface

It is fairly obvious that our world is uncertain, complex, and ambiguous. It is true
today more than ever, and tomorrow will be more challenging. Artificial
Intelligence (AI) can improve responses to major challenges to our communities
from economic and industrial development to health care and disease containment.
Nevertheless, Artificial Intelligence is still growing and improving technology, and
there is a necessity for more creative studies and researches led by both academics
and practitioners.
Artificial Intelligence and Industrial Applications—A2IA’2020, which is the first
edition of an annual international conference organized by the ENSAM—Meknes
at Moulay Ismail University, intends to contribute to this common great goal. It
aims to offer a platform for experts, researchers, academics, and industrial practi-
tioners working in artificial intelligence and its different applications to discuss
problems and solutions, concepts, and theories and map out the directions for future
research. The connections between institutions and individuals working in this field
have to keep growing on and on, and this must have a positive impact on pro-
ductivity and effectiveness of researches.
The main topics of the conference were as follows:
• Smart Operation Management
• Artificial Intelligence: Algorithms and Techniques
• Artificial Intelligence for Information and System Security in Industry
• Artificial Intelligence for Energy
• Artificial Intelligence for Agriculture
• Artificial Intelligence for Healthcare
• Other Applications of Artificial Intelligence
In A2IA’2020 conference proceeding, about 141 papers were received from
around the world. A total of 58 papers were selected for presentation and publi-
cation. In order to maintain a high level of quality, a blind peer review process was
performed by a large international panel of qualified experts in the conference topic
areas. Each submission received at least two reviews, and several received up to

xiii
xiv Preface

five. The papers were evaluated on their relevance to A2IA’2020 tracks and topics,
scientific correctness, and clarity of presentation.
The papers were organized in two parts:
• Artificial Intelligence and Industrial Applications:
Smart Operation Management (Volume 1)
In : Advances in Intelligent Systems and Computing.
• Artificial Intelligence and Industrial Applications:
Artificial Intelligence Techniques for Cyber-Physical, Digital Twin Systems and
Engineering Applications (Volume 2)
In : Lecture Notes in Networks and Systems.
We hope that our readers will discover valuable new ideas and insights.
Lastly, we would like to express our thanks to all contributors in this book
including those whose papers were not included. We would also like to extend our
thanks to all members of the Program Committee and reviewers, who helped us
with their expertise and valuable time. We are tremendously grateful for the pro-
fessional and organizational support from Moulay Ismail University. Finally, our
heartfelt thanks go especially to Springer-Nature.

Meknes, Morocco Tawfik Masrour

Anass Cherrafi
Ibtissam El Hassani
Contents

A Fuzzy Ontology Based Approach to Support Product

Eco-Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Chaimae Abadi, Imad Manssour, and Asmae Abadi
A Genetic-Based SVM Approach for Quality Data
Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Wahb Zouhri, Hamideh Rostami, Lazhar Homri, and Jean-Yves Dantan
Towards a Platform to Implement an Intelligent and Predictive
Maintenance in the Context of Industry 4.0 . . . . . . . . . . . . . . . . . . . . . . 33
El Mehdi Bourezza and Ahmed Mousrij
Towards a Prediction Analysis in an Industrial Context . . . . . . . . . . . . 45
Ilham Battas, Ridouane Oulhiq, Hicham Behja, and Laurent Deshayes
Methodology for Implementation of Industry 4.0 Technologies
in Supply Chain for SMEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Hafsa El-kaime and Saad Lissane Elhaq
A Deep Reinforcement Learning (DRL) Decision Model for Heating
Process Parameters Identification in Automotive Glass
Manufacturing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Choumicha El Mazgualdi, Tawfik Masrour, Ibtissam El Hassani,
and Abdelmoula Khdoudi
Analytic Hierarchy Process (AHP) for Supply Chain 4.0 Risks
Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Kamar Zekhnini, Anass Cherrafi, Imane Bouhaddou,
and Youssef Benghabrit
SmartDFRelevance: A Holonic Agent Based System for Engineering
Industrial Projects in Concurrent Engineering Context . . . . . . . . . . . . . 103
Abla Chaouni Benabdellah, Imane Bouhaddou, and Asmaa Benghabrit

xv
xvi Contents

A Cyber-Physical Warehouse Management System Architecture

in an Industry 4.0 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Mariam Moufaddal, Asmaa Benghabrit, and Imane Bouhaddou
PLM and Smart Technologies for Product and Supply
Chain Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Oulfa Labbi and Abdeslam Ahmadi
Production Systems Simulation Considering Non-productive Times
and Human Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Ismail Taleb, Alain Etienne, and Ali Siadat
Distributed and Embedded System to Control Traffic Collision
Based on Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Tarik Hajji, Tawfik Masrour, Mohammed Ouazzani Jamil,
Zakaria Iathriouan, Sanaa Faquir, and Elmiloud Jaara
The Emergence and Decision Support in Complex System with Fuzzy
Logic Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Mohammed Chennoufi and Fatima Bendella
Markov Decision Processes with Discounted Costs over a Finite
Horizon: Action Elimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Abdellatif Semmouri and Mostafa Jourhmane
Robust Adaptive Fuzzy Path Tracking Control of Differential
Wheeled Mobile Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Brahim Moudoud, Hicham Aissaoui, and Mohammed Diany
Deep Learning Approach for Automated Guided Vehicle System . . . . . 227
Mohamed Rhazzaf and Tawfik Masrour
Path Planning Using Particle Swarm Optimization and Fuzzy Logic . . . 239
Ahmed Oultiligh, Hassan Ayad, Abdeljalil Elkari, and Mostafa Mjahed
Prediction of Robot Localization States Using Hidden
Markov Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
Jaouad Boudnaya, Amine Haytoumi, Omar Eddayer,
and Abdelhak Mkhida
A New Approach for Multi-agent Reinforcement Learning . . . . . . . . . . 263
Elmehdi Amhraoui and Tawfik Masrour
Recommender System for Most Relevant K Pick-Up Points . . . . . . . . . . 277
Ayoub Berdeddouch, Ali Yahyaouy, Younès Bennani, and Rosanna Verde
Feature Detection and Tracking for Visual Effects: Augmented Reality
and Video Stabilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
Houssam Halmaoui and Abdelkrim Haqiq
Contents xvii

Spectral Image Recognition Using Artificial Dynamic Neural Network

in Information Resonance Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
Ivan Peleshchak, Roman Peleshchak, Vasyl Lytvyn, Jan Kopka,
Mariusz Wrzesien, Janusz Korniak, Janusz Kolbusz, and Pawel Rozycki
U-Net Based Model for Obfuscated Human Faces Reconstruction . . . . . 323
Elmoukhtar Zemmouri, Youssef Yakoubi, Mohammed Douimi,
and Adil Saadi
A Machine Learning Assistant for Choosing Operators and Tuning
Their Parameters in Image Processing Tasks . . . . . . . . . . . . . . . . . . . . . 339
Issam Qaffou
Convergence and Parameters Setting of Continuous Hopfield Neural
Networks Applied to Image Restoration Problem . . . . . . . . . . . . . . . . . 351
Joudar Nour-eddine, Ramchoun Hassan, Zakariae En-Naimani,
and Ettaouil Mohamed
The IBN BATTOUTA Air Traffic Control Corpus with Real Life
ADS-B and METAR Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
Kasttet Mohammed Saïd and Lyhyaoui Abdelouahid
Weed Recognition System for Low-Land Rice Precision Farming
Using Deep Learning Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
Olayemi Mikail Olaniyi, Emmanuel Daniya,
Ibrahim Mohammed Abdullahi, Jibril Abdullahi Bala,
and Esther Ayobami Olanrewaju
A Comparative Study Between Mixture and Stepwise Regression
to Model the Parameters of the Composting Process . . . . . . . . . . . . . . . 403
Echarrafi Khadija, Ben Abbou Mohamed, Taleb Mustapha, Rais Zakia,
Ibtissam El Hassani, and El Haji Mounia
Deep Learning Based Sponge Gourd Diseases Recognition
for Commercial Cultivation in Bangladesh . . . . . . . . . . . . . . . . . . . . . . . 415
Tahmina Tashrif Mim, Md. Helal Sheikh, Sadia Chowdhury,
Roksana Akter, Md. Abbas Ali Khan, and Md. Tarek Habib
Exploitation of Vegetation Indices and Random Forest
for Cartography of Rosemary Cover: Application to Gourrama
Region, Morocco . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
Hassan Chafik and Mohamed Berrada

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441

About the Editors

Dr. Tawfik Masrour is a Professor of Applied

Mathematics and Artificial Intelligence in National
High School for the Arts and Crafts (ENSAM-Meknes),
My Ismaïl University UMI, and member of Research
Team Artificial Intelligence for Engineering Sciences
(AIES) and Laboratory of Mathematical Modeling,
Simulation and Smart Systems. (L2M3S). He graduated
from Mohammed V University—Rabat with MSc
degree in Applied Mathematics and Numerical
Analysis and from Jacques-Louis Lions Laboratory,
Pierre and Marie Curie University, Paris with M.A.S
(DEA) in Applied Mathematics, Numerical Analysis
and Computer Science. He obtained his Ph.D., in
Mathematics and informatics, from École des Ponts
ParisTech (ENPC), Paris, France. His research interests
include Control Theory and Artificial Intelligence.
Email: [email protected].

Dr. Anass Cherrafi is an Assistant Professor at the

Department of Industrial Engineering, ENSAM-
Meknes, Moulay Ismail University, Morocco. He holds
a Ph.D. in Industrial Engineering. He has about 7 years of
industry and teaching experience. He has published a
number of articles in leading international journals and
conferences. He has participated as Guest Editor for
special issues in various international journals. His
research interests include industry 4.0, green manufac-
turing, Lean Six Sigma, integrated management systems,
and supply chain management.

xix
xx About the Editors

Dr. Ibtissam El Hassani is a Professor of Industrial

and Logistic Engineering in National High School for
the Arts and Crafts (ENSAM-Meknès), My Ismaïl
University UMI, and Head of Research Team Artificial
Intelligence for Engineering Sciences (AIES). She grad-
uated as an industrial engineer and then got her Ph.D. in
Computer Engineering and Systems Supervision at the
Center for Doctoral Studies at My Ismail University. Her
fields of interest are applications of artificial intelligence in
industrial engineering specially in Lean Manufacturing,
Six Sigma and Continuous Improvement, Production
systems, Quality, Management System, Health and secu-
rity at work, Supply Chain Management, Industrial Risk
Management, and Work Measure. Email: i.elhassani@umi.
ac.ma.
A Fuzzy Ontology Based Approach
to Support Product Eco-Design

Chaimae Abadi , Imad Manssour, and Asmae Abadi

Abstract Nowadays, eco-products have become a real need in the world due to the
awareness of the importance of preserving the environment. Consequently, compa-
nies have become faced to new requirements that should be taken into consideration
from the design process. It is in this context that a new structured approach has been
developed in this paper. It is based on fuzzy ontologies which make the comparison
between many eco-design alternatives and the selection of the best one automatic
and simple. The use of this artificial intelligent tool allows also the handling and the
consideration of fuzzy and uncertain information in the design process which makes
the studies in eco-design more efficient and precise. A case study is presented in
order to demonstrate the efficacy of the proposed methodology.

Keywords Artificial intelligence · Ontology · Fuzzy ontologies · Eco-design ·

Structured approach

1 Introduction

Producing eco-products has become one of the main objectives of the majority of
companies because of the need of protecting the environment and natural resources.
Thus, this requirement should be taken into consideration from the earlier stages
of product life cycle. In this context, the eco-design has appeared and different
definitions and approaches of this concept have been developed. Actually, eco-design
aims to lead the world towards a sustainable future [1].
Platcheck defines eco-design as “a holistic view in that, starting from the moment
we know the environmental problems and its causes, we begin to influence the

C. Abadi (B) · I. Manssour

Laboratory of Mechanics, Mechatronics and Command. Team of Electrical Energy, Maintenance
and Innovation, Ecole Nationale Supérieure d’Arts et Métiers, Moulay Ismail University, B.P.
15290 El Mansour, Meknes, Morocco
e-mail: [email protected]
A. Abadi
INSA, Euro-Mediterranean University of Fez, Fes, Morocco

© Springer Nature Switzerland AG 2021 1

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_1
2 C. Abadi et al.

conception, the materials selection, the production, the use, the reuse, the recy-
cling and final disposition of industrial products” [2]. In addition to that, according
to Ihobe, the Eco-design concept means that the environment is taken into account
when making decisions during the product development process as an additional
factor that traditionally has been taken into account [3].
There are a variety of eco-design tools that have been proposed in the literature
which are classified generally into four categories: guidelines/standards, checklists,
comparative tools, and analytical methods. The most used method of eco-design
is the LCA (life cycle assessment) method. This method consists on supplying a
quantitative data on the environmental impact of the studied product along its life
cycle, since extraction and production of materials to the end of its life, using different
environmental indicators [4].
In order to rank the impacts caused by a product, the ABC analysis is used.
Electronic product environmental assessment tool (EPEAT) is used also for rating and
ranking. In fact, it allows rating the life cycle issues to suggest the design changes for
eco-labeling. In addition to that, there are other eco-design tools like: The eco-design
indicator tool that Builds the life cycle profile of a product for the measurement of its
impact, MET-matrix (material, energy and toxicity) which permits the identification
of the environmental load of the materials, energy and toxicity of products and Tool
for environmentally sound product innovation (TESPI) that allows the determination
of the features of greener products [5].
Contrary to the majority of the previous eco-design approaches which focus on the
analytic side of the problem, in this paper, the proposed approach aims to automate
this phase using the characteristics and the capacities of fuzzy ontologies as a new
emerging artificial intelligence tool.
To do so, reviews about fuzzy ontologies researches will be presented in section
one. Then, the proposed approach and its contributions will be developed. Finally,
a case of study will be submitted in order to show the efficacy of the developed
approach.

2 Fuzzy Ontologies as an Emerging Artificial Intelligence

Tool

An ontology is defined as an explicit and formal specification of the concepts, indi-

viduals and relationships which exist in some area of interest. It is created by defining
axioms that describe the properties of these entities [6]. Many times, ontologies are
used for knowledge representation. But, according to [7], the role of ontologies differs
from a context to another.
In fact, ontologies present different advantages. However, they have a few limita-
tions too. Actually, the description logics were the base of ontologies, at first, which
makes them unable to deal with imprecise information [8, 9]. As a solution to this
problem, Fuzzy Ontologies were appeared.
A Fuzzy Ontology Based Approach to Support Product Eco-Design 3

By referring to [10, 11], a fuzzy Ontology is defined as a quintuple OF = {I, C, R,

F, A} where I is a set of individuals, C is a set of concepts, R is a set of relations, F is
a set of fuzzy relations and A is the set of axioms. Consequently, it can be concluded
that a fuzzy ontology is a generalization of a crisp ontology since it considers fuzzy
membership degrees of relations and it presents intuitively more expressiveness [12].
Through the concept of fuzzy DLs and specially the fuzzy DL reasoner, fuzzy
ontologies have become able to express fuzzy relations and to manage fuzzy and
uncertain information [13]. Hence, they are used in many applications, for examples:
robotics [14], image interpretation [15], ambient intelligence [16], and information
retrieval [17] and of course decision making [18].

3 The Proposed Fuzzy Ontology-Based Approach

In order to support designers in the eco-design process of a new industrial eco-

product, a new approach is presented in this paper. The proposed approach is based on
the fuzzy ontologies concept. Actually, this artificial intelligent tool allows mainly,
as it is noted in the previous section, the automation of decision-making. So, the
selection of the most appropriate alternative in an eco-design problem will be done
in an automated way.

3.1 Description of the Proposed Approach

The developed approach concerns in particular the eco-design of a new industrial

product. As it is shown in Fig. 1 that presents its flowchart, the proposed approach
is compound of the following steps:

Step 1: Defining decision eco-criteria and their weights according to each member
of the design team
Step 2: Collecting design team members’ performances for each alternative with
respect to each criterion using linguistic variables. This step is done
following four stages.
• Step 2.1: The first stage of the second step is to collect design team
members performances in the pre-manufacturing phase of the studied
product. In this stage, each member should take into consideration three
criteria which are: The material extraction of the studied product, design
of environment and the material processing.
• Step 2.2: In this second stage, the purpose is to collect design team
members performances in the manufacturing phase of the studied product
according to three criteria: The production energy used, the hazardous
waste produced and the renewable energy used.
4 C. Abadi et al.

Fig. 1 The proposed fuzzy ontology based approach to support the product eco-design process
A Fuzzy Ontology Based Approach to Support Product Eco-Design 5

• Step 2.3: Collecting design team members’ performances in the use phase
of the studied product represents the third stage of the second step of the
proposed approach. In this stage, three criteria are taken into account:
Emissions, functionality and hazardous waste generated.
• Step 2.4: The last stage in this step is to collect design team members
performances in the post-use phase of the studied product. Each member
should consider in this stage three different criteria which are: The
recyclability, the remanufacturability and the redesign of our eco-
product.
Step 3: Encoding the different membership functions of linguistic values, which
designers have used to weight alternatives, in the fuzzy ontology. This oper-
ation is based on the data types noted in the case study presented in this
paper.
Step 4: Encoding the different design team members’ performances in the fuzzy
ontology.
Step 5: Computing the relative rank value for each eco-design alternative for each
member of the design team on the basis of weighted sum concepts and by
using the fuzzy DL reasoned.
Step 6: Computing the final rank value of each eco-design alternative using the
same method of the previous step.
Step 7: Selecting the best eco-design alternative by interrogating the fuzzy ontology.
It is the alternative which maximizes satiability degree of the fuzzy concept
(FinalValoration-i).

3.2 The Role of the Fuzzy Ontology in the Proposed

Approach

In the developed approach, the fuzzy ontology plays a key role in the automating of the
eco-design process. Actually, it rounds up all the collaborators in a common model.
Thus, it allows the exchange of information between members and applications
without difficulties. In addition to that, it facilitates the understanding and provides
the semantic interoperability within the multidisciplinary teams. Those two functions
are ensured through the formalization of semantics.
Moreover, in addition to its ability of knowledge representation, the fuzzy
ontology permits the computation of the final ranking of the different eco-design
alternatives.
So that we can conclude that the proposed approach uses the fuzzy ontology as
an evaluation tool of the different eco-design alternatives.
Table 1 represents the different characteristics of the fuzzy ontologies that allow
the insurance of its different mentioned functions.
6 C. Abadi et al.

Table 1 The role of the Fuzzy Ontology in the proposed approach

The fuzzy ontology characteristic Its roles
Expressiveness Through this feature, a unified knowledge model is
developed. It contains crisp information, imprecise data
using fuzzy relations and membership functions. Thus, the
majority of semantic interoperability problems and of
designers’ conflicts are resolved
Inference Through the main role that plays the fuzzy DL reasonner
FuzzyDL [13], the inference capacities of the fuzzy
ontology will be used in the different steps of the developed
approach as following:
– Knowledge recuperation: by modeling eco-design criteria
in OWL2, it will be possible to query the fuzzy ontology
and to benefit from the ancient similar projects
– Deduction of new eco-design alternatives by combining
several technical solutions
– Time gain and facilitation of the MCDM process by
clustering the different evaluated alternatives in a
pertinent consistent group
– Automation of the computation of the relative and final
ranks of the eco-design alternatives
Storage This characteristic enables the capitalization and the storage
of all the data and knowledge manipulated and generated
within the project in particular during the eco-design phase.
Thus, the design information and results can be reused in
other eco-design projects

4 Implementation of the Proposed Approach: Case Study

In this section, an industrial case of study is developed to illustrate the implementation

of the proposed eco-design ontology-based approach. In fact, an industrial company
specialized in the manufacturing of washing machines is desired to choose the most
relevant eco-concept design solution from a set of five pre-evaluated eco-friendly
alternatives.
To do so, a design team composed of three experts specialized in environ-
mental impact product studies (D1 , D2 and D3 ) has been formed to evaluate the
five eco-product conceptual alternatives. In their environmental impact evaluations,
we consider that the experts will use four common eco-criteria. Each criterion repre-
sents the expected environmental impact of the eco-product in a specific phase of its
product lifecycle. We denote the used eco-criteria as follows (Fig. 2):
• C 1 : Pre-manufacturing
• C 2 : Manufacturing
• C 3 : Use
• C 4 : Post-use.
A Fuzzy Ontology Based Approach to Support Product Eco-Design 7

Fig. 2 The considered eco-design MCDM problem

Table 2 The importance

Designer D1 D2 D3
degree of each designer in the
team and the weight affected Importance degree 0.32 0.36 0.32
by each designer to each C 1 weight 0.14 0.18 0.35
criterion
C 2 weight 0.24 0.41 0.39
C 3 weight 0.31 0.19 0.12
C 4 weight 0.31 0.22 0.14

The first inputs of our design making problem are the importance
degree of each designer in the team and the weight affected by each
designer to each criterion. Table 2 summarizes these data.
Then, the design team members have expressed their performance on each eco-
design alternative with respect to each criterion using the six different linguistic
variables described in Fig. 3. The next step in our ontology-based approach was

Fig. 3 The linguistic variables used for alternatives rating and their corresponding membership
functions
8 C. Abadi et al.

Table 3 The corresponding

Linguistic variables Corresponding fuzzy numbers
fuzzy numbers for each
linguistic variable VP Left-shoulder (1, 2)
P Triangular (1, 2, 3)
F Triangular (2, 3, 4)
G Triangular (3, 4, 5)
VG Triangular (4, 5, 6)
E Right-shoulder (5, 6)

to exploit the fuzzy ontology expressivity in order to directly convert the linguistic
variables into fuzzy numbers as in Table 3.
To do so, we encoded six data types in our fuzzy ontology and we associated to
each one of them a specific annotation property [13, 18]. The role of the last ones
was to define, for each linguistic variable, the parameters of its corresponding fuzzy
membership function.
An example of the annotation properties that we have defined in the fuzzy ontology
in order to define the linguistic variables “Excellent” and “Fair” is given in Fig. 4.
The ontology editor used for the implementation is Protégé 5.0.
Then, the decision making matrix has been constructed as represented in Table 4.
In fact, each designer Dk has expressed his performance for each eco-design
alternative EDi with respect to each eco-criterion C j .

Fig. 4 The linguistic variables used for alternatives rating and their corresponding membership
functions
A Fuzzy Ontology Based Approach to Support Product Eco-Design 9

Table 4 The eco-design alternatives evaluation according to the design team members
Eco-criterion Eco-design alternative Decision makers
D1 D2 D3
C1 ED1 G E VG
ED2 P F F
ED3 F G VP
ED4 VG VG P
ED5 E P E
C2 ED1 G P VG
ED2 VG F F
ED3 F G VP
ED4 VP VG P
ED5 P E E
C3 ED1 VG VP G
ED2 F P VG
ED3 VP F F
ED4 P G VP
ED5 E VG P
C4 ED1 P G VP
ED2 F VG P
ED3 G F F
ED4 VG VP G
ED5 E P VG

In order to encode these information in the fuzzy ontology, a data property ‘hasS-
core’ has been defined. In our case study, for example, Designer 3 has judged the
eco-design alternative 5 according to criterion 2 as “Excellent”.
To encode this performance, we have defined a new concept entitled Perfor-
mance352 as follows: Performance352 = ∃ hasScore. Excellent, where Excellent
is the datatype that we have previously defined to encode the linguistic variable
“Excellent” as the right-shoulder fuzzy number (5, 6). The same methodology has
been used to encode all the other performances.
The next step is to proceed to calculating the relative and final rates of the different
eco-design alternatives. To do so, we defined a concept denoted RelativeRank. Then,
for each eco-design alternative EDi and for each designer Dk , a sub concept Rela-
tiveRank ki has been defined and annotated as a weighted sum concept [13] as in
Fig. 5.
The last stage of our proposed methodology was to compute the final rank values of
each eco-design alternative. To do so, the same encoding steps have been performed.
10 C. Abadi et al.

Fig. 5 The annotation property defining the concept RelativeRank13

In fact, we defined for each eco-design alternative EDi , a specific weighted sum sub-
concept denoted FinalRank i which takes into account the relative rates calculated
before and the importance degree of each design team member.
An example of the fuzzy annotation that we have defined to characterize the final
rank of the eco-design 1 is given in Fig. 6.
Finally, using the fuzzy DL reasoner FuzzyDL [13], we deduce the best eco-design
alternative solution. It is the concept design solution which maximizes the stability
degree of the fuzzy concept FinalRank i. Table 5 summarizes the final rank value of
the five alternatives for our case study.
We conclude then that the ranking of the five eco-design alternatives is: ED5 >
ED1 > ED2 > ED4 > ED3 and the optimal eco-design alternative that the design team
needs to extend in the detailed design phase is ED* = ED5 .
A Fuzzy Ontology Based Approach to Support Product Eco-Design 11

Fig. 6 Example of the annotation property defining the concepts FinalRank

Table 5 The final rank value

Eco-design alternative Final score
of each eco-design alternative
ED1 3,6324
ED2 3,2308
ED3 2,6396
ED4 3,1012
ED5 4,5812

5 Conclusion

In this paper, we propose a new fuzzy ontology-based approach to support designers

in the eco-design process. The use of this new artificial intelligence tool allows us, at
the same time, to encode the subjective performances of designers and to calculate
in an automated way the final rank of each eco-design alternative.
The proposed approach retains the advantages of classical ontologies and allows
in addition the modeling and reasoning on the fuzzy and imprecise information
12 C. Abadi et al.

manipulated in the eco-design process. To demonstrate the feasibility of the proposed

approach, a case study was proposed using the ontology editor Protégé 5.0 to choose
the best eco-product design from several alternatives based on different eco-design
criteria.
This work demonstrates that fuzzy ontologies can be used as an efficient tool to
support the decision making process in the different stages of product eco-design.
Thus, the use of fuzzy ontologies potentialities to support other stages of the product
development process will be investigated in our future works such as: Manufacturing
phase, control phase, recycling phase.

References

1. Meadows, D.H., Meadows, D., Randers, J., et al.: Los límites del crecimiento: informe al Club
de Roma sobre el predicamento de la humanidad. fondo de cultura económica (1972)
2. Platcheck, E.R., Schaeffer, L., Kindlein, W., Jr., Candido, L.: Methodology of ecodesign for
the development of more sustainable electro-electronic equipment. J. Clean. Prod. 16, 75–86
(2008)
3. IHOBE. Manual práctico de ecodiseño: operativa de implantación en siete pasos (2000)
4. Germani, M., Dufrene, M., Mandolini, M., Marconi, M., Zwolinski, P.: Integrated software plat-
form for green engineering design and product sustainability. In: Re-engineering Manufacturing
for Sustainability, pp. 87–92. Springer, Singapore (2013)
5. Singh, P.K., Sarkar, P.: Eco-design approaches for developing eco-friendly products: a review.
In: Advances in Industrial and Production Engineering, pp. 185–192. Springer, Singapore
(2019).
6. Mädche, A., Staab, S., Studer, R.: Handbook on ontologies. In: Handbook on Ontologies,
International Handbooks on Information Systems, pp. 173–190. Springer (2004).
7. Mezei, J., Wikström, R., Carlsson, C.: Aggregating linguistic expert knowledge in type-2 fuzzy
ontologies. Appl. Soft Comput. 35, 911–920 (2015)
8. Morente-Molinera, J.A., Pérez, I.J., Chiclana, F., Herrera-Viedma, E.: A novel group decision
making method to overcome the Web 2.0 challenges. In: 2015 IEEE International Conference
on Systems, Man, and Cybernetics, pp. 2233–2238. IEEE (2015)
9. Lukasiewicz, T., Straccia, U.: Managing uncertainty and vagueness in description logics for
the semantic web. Web Semant. Sci. Serv. Agents World Wide Web 6, 291–308 (2008)
10. Baader, F., Calvanese, D., McGuinness, D., Patel-Schneider, P., Nardi, D.: The Description
Logic Handbook: Theory, Implementation and Applications. Cambridge University press
(2003)
11. Calegari, S., Ciucci, D.: Fuzzy ontology, fuzzy description logics and fuzzy-owl. In: Interna-
tional Workshop on Fuzzy Logic and Applications, pp. 118–126. Springer, Berlin, Heidelberg
(2007)
12. Carlsson, C., Brunelli, M., Mezei, J.: Decision making with a fuzzy ontology. Soft. Comput.
16, 1143–1152 (2012)
13. Bobillo, F., Straccia, U.: The fuzzy ontology reasoner fuzzyDL. Knowl.-Based Syst. 95, 12–34
(2016)
14. Eich, M., Hartanto, R., Kasperski, S., Natarajan, S., Wollenberg, J.: Towards coordinated multi-
robot missions for lunar sample collection in an unknown environment. J. Field Robot. 31,
35–74 (2014)
15. Dasiopoulou, S., Kompatsiaris, I., Strintzis, M.G.: Investigating fuzzy DLs-based reasoning in
semantic image analysis. Multimed. Tools Appl. 49, 167–194 (2010)
16. Rodríguez, N.D., Cuéllar, M.P., Lilius, J., Calvo-Flores, M.D.: A fuzzy ontology for semantic
modelling and recognition of human behaviour. Knowl.-Based Syst. 66, 46–60 (2014)
A Fuzzy Ontology Based Approach to Support Product Eco-Design 13

17. Calegari, S., Sanchez, E.: Object-fuzzy concept network: an enrichment of ontologies in
semantic information retrieval. J. Am. Soc. Inform. Sci. Technol. 59, 2171–2185 (2008)
18. Abadi, A., Ben-Azza, H., Sekkat, S., Zemmouri, E.M.: A Fuzzy Ontology-based approach
for multi-criteria decision-making: a use case in supplier selection. Congrès International Du
Génie Industriel Et Du Management Des Systèmes, Meknes (2017)
A Genetic-Based SVM Approach
for Quality Data Classification

Wahb Zouhri, Hamideh Rostami, Lazhar Homri, and Jean-Yves Dantan

Abstract With the emergence of the Industry 4.0 and the big data era, many organi-
zations had recourse to data-based approaches for Quality management. One of the
main aims of the data-based approaches in manufacturing industries is quality classi-
fication. Classification methods provide many solutions related to quality problems
such a defect detection and conformity prediction. In that context, this paper iden-
tifies a suitable technique (Support Vector Machine) for quality data classification,
as it proposes an appropriate approach to optimize its performances. The proposed
approach is tested on a chemical manufacturing dataset and a rolling process dataset,
in order to evaluate its efficiency.

Keywords Quality management · Quality 4.0 · GA-SVM · Robustness

1 Introduction

Industry 4.0, also known as the fourth industrial revolution, is characterized by the
application of information and communication technologies to industry, and that
by making a network connection that allows the communication between produc-
tion systems, components, and people. This network connection aims to meet the
requirement for unique and customized products while maintaining equivalent costs,

W. Zouhri (B) · L. Homri · J.-Y. Dantan

Arts et Metiers Institute of Technology, Université de Lorraine, LCFC, HESAM University, 57070
Metz, France
e-mail: [email protected]
L. Homri
e-mail: [email protected]
J.-Y. Dantan
e-mail: [email protected]
H. Rostami
Data science and Engineering group, ASML, De Run 6501, 5504 DR Veldhoven, The Netherlands
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 15

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_2
16 W. Zouhri et al.

despite the low production volumes generated. These forms of connections make it
possible to gather new masses of data on the network and thus, new knowledge [1].
To extract valuable knowledge, industries working with large amounts of data aim
to discover hidden patterns and insights, and to make decision using methods such
as statistics, predictive analytics, and machine learning, in order to help companies
in their decision making.
One of the main industries that are taking benefit of data analytics capabili-
ties is manufacturing. Manufacturing companies are having recourse to data-based
approaches to improve their standard in product, process design and management
in order to remain competitive. The extracted knowledge is used within manufac-
turing systems for many purposes such as: maintenance, supply-chain management,
cost estimations, demand forecasting, material requirements, enterprise resource
planning, process optimization, scheduling, and quality control [2].
Quality 4.0 represents an opportunity to utilize those data-based approaches to
improve the quality within manufacturing systems, and that by predicting failures
and then by informing what action to take to change the outcome. The purposes
of data-based approaches for quality management can be summed up in three main
categories [3]:
• To describe the production process in order to identify the different factors that
directly influence the quality.
• To predict the quality using predictive models, which allows finding the optimal
settings that guarantee the best quality.
• To identify defective products from non-defective products using classification
tools.
Various manufacturing industries exploited classification tools to improve the
quality. This choice is justified by the ability of these tools to guarantee a set of
solutions related to quality problems, such as defect detection, extraction of optimal
adjustment ranges, and conformity prediction [2].
El Attar and Hamery [4] designed a learning system using the ID3 classifier to
automatically integrate knowledge into an expert system to guide helicopter blade
repair. The C4.5 method was used in a hard disk drive production company to identify
the best settings, which increased production quality by 12% [5].
Decision trees have been tested within a chip manufacturing industry to help iden-
tify the causes of defects, making decisions immediately and, eventually, reducing
the cycle time required to solve quality problems [6]. Decision trees were also used
to define the number of kanbans in a dynamic JIT (just-in-time) factory. The number
of kanbans is important for the efficient operation of a JIT production system. The
results showed that decision trees classifiers represent a practical approach with
special capabilities to dynamically adjust the number of Kanbans [7, 8].
Hou et al. [9] have integrated neural networks for monitoring the manufacturing
process and diagnosing failures. A neural network was used to classify quality
defects such as uneven thicknesses, and to determine the cause-and-effect relation-
ship between parameters such as process temperature and output quality measure-
ments. Cho and Leu [10] presented neural network applications for monitoring and
A Genetic-Based SVM Approach for Quality Data Classification 17

controlling manufacturing processes such as machining processes, semiconductors

manufacturing processes, and injection molding processes, while shedding light on
the influence of this tool in improving the quality of these processes. Also, neural
networks have been used for monitoring and controlling the induction hardening
process to improve the hardness of metal products [11, 12].
A quality approach based on the exploitation of data using neural networks has
been proposed by Thomas et al. [13] to classify quality results, in order to avoid
the occurrence of defects. The approach consists first in designing different neural
networks and then choosing the one that guarantees the best results. This approach
has been tested in a paint industry, and it did improve the product quality by sending
warnings if there is a risk of defects, or by stopping production if one of the param-
eters is set on an outlier value. Bahlmann et al. [14], have implemented a five-step
approach for classifying the different textile seams, the approach consists in locating
the seams in images acquired in grayscale, then using neural networks, and based on
the undulations around the seam, the classification of seam quality is performed on
a scale from 1 to 5. This approach has made it possible to classify the quality of a
seam within one second with an accuracy of 80%, which is much less than the time
required for an expert, which is estimated at 30 s.
Gryllias and Antoniadis [15], proposed a two-stage hybrid approach based for the
automated diagnosis of defective rolling element bearings. A support vector machine
(SVM) model is first trained using simulation data, then, vibration measurements
resulting from the machine under condition monitoring are imported and processed
directly by the already trained SVM. Fernández-Francos et al. [16] presented an
automatic method for bearing fault detection and diagnosis. The method uses the
information contained in the vibration signals and a one-class v-SVM to discriminate
between normal and faulty conditions. A weighted support vector machines (WSVM)
was proposed by Xanthopoulos and Razzaghi [17] for automated process monitoring
and early fault diagnosis.
Diao et al. [18] propose a quality control approach by improving the most influen-
tial factors. The approach begins by identifying the factors that lead to quality prob-
lems, known as dominant factors (DFs), using a principal component analysis (PCA).
Then, a quality prediction model to improve DFs is proposed based on the SVM.
An additional weight is introduced into the SVM model to improve and increase its
quality prediction accuracy. Thus, the quality of the product can be guaranteed by
controlling the dominant factors. This approach was tested on a coating process, and
the result showed the feasibility of this methodology which allowed a better control
of the parameters, and consequently an improvement of the quality of this process.
Baccarini et al. [19] used vibration as a basis for diagnosing induction motor faults.
Four SVM-based models have been developed to classify the three recurring defects,
using the ‘’one against all” approach, which identifies whether the vibrations emitted
by the engine indicate one of the three defects or not. Similarly, Jegadeeshwaran and
Sugumaran [20], proposed an approach for diagnosing hydraulic brake system failure
using vibration signals and classification methods. The C4.5 decision tree was used
to select the relevant parameters, then classification was performed using two types
18 W. Zouhri et al.

of SVMs with different kernels, whose basic radial function (RBF) ensured better
accuracy.
The short review above, shows the great potential of the different classifiers in
improving the quality within manufacturing industries. Although, the support vector
machine was selected as the technique to study among the different classifiers, due to
its easy use, and its ability to handle large dimensional dataset with different features.
In addition, due to the issue that the quality data are not often linearly separable, SVM
use different tricks, called Kernel functions to classify non-linear data [2]. However,
the selection of the optimal kernel function is challenging. Some results may be bad
not because the data is noisy or the used learning algorithm is weak, but due to the bad
selection of the kernel space and the parameters that define it. Hence, an approach
should be followed in order to find the best one (Kernel function) that guarantees the
best predictive performances. To this end, an approach based on genetic algorithm
(GA-SVM) is proposed to optimize the performances of SVM.
The genetic algorithm based approaches rely on bio-inspired operators such
as mutation, crossover and selection to evolve toward high-quality solutions. This
evolution usually starts from a population of randomly generated individuals that
undergo recombination in order to produce better generations, and thus, better solu-
tions [21]. In the SVM case, the genetic algorithm aims to find the best kernel space
with the optimal predictive performances. Once defined, the kernel space robust-
ness is assessed, and that by evaluating how measurement uncertainties affect its
performances.
Accordingly, and after introducing the different applications of classification
methods within manufacturing industries in Sect. 1, a detailed description of the
developed approach is addressed in Sect. 2, as well as the results gotten by applying
it on the studied datasets. Finally, Sect. 3 discusses the different results, as it presents
some concluding remarks.

2 GA-SVM: Genetic-Based SVM Approach for Quality

Data Classification

It has been shown that evolutionary algorithms such as genetic algorithms or evolu-
tion strategies are efficient and robust approaches to solve a wide range of optimiza-
tion problems [22]. Accordingly, to optimize the SVM’s predictive performances,
different evolutionary algorithms have been utilized such as genetic algorithm [23],
particle swarm optimization [24], artificial bee colony algorithm [25], and firefly
algorithm [26]. Among these, genetic algorithm (GA) has been widely used to solve
optimization problems, and consequently in this paper, a genetic-based SVM model
(GA-SVM) is developed to find the best kernel function as well as its parameters, in
order to achieve SVM’s highest accuracy.
This section describes the design of the proposed algorithm to classify the quality
data.
A Genetic-Based SVM Approach for Quality Data Classification 19

Fig. 1 GA-SVM approach for quality data classification—a graphical description

In the following, the different parameters and steps of the proposed genetic-based
SVM approach are presented. The genetic representation and the initialization of
the first population are discussed at first. After that, the fitness function is described,
then, the selection, the crossover, and the mutation techniques used in the approach
are presented in detail.
The approach is evaluated through an application on two datasets: a chem-
ical manufacturing data, and a rolling process data. Then, the robustness of the
classification results regarding the measurement uncertainties is assessed (Fig. 1).

2.1 Genetic Representation and Initialization

One of the most critical decisions that significantly affects the performance of a
genetic algorithm while implementing it, is deciding the representation to use in
order to represent the solutions. It has been observed that unsuitable genetic repre-
sentation can lead to poor performance of a GA, hence, choosing a proper genetic
representation is necessary for the success of a GA. The representation of the solution
must be complete and must contain the information needed to represent a solution
to the problem, otherwise, the search will be either larger or poorer than necessary.
In this paper and in order to define a proper genetic representation, the different
parameters to optimize need to be defined. Accordingly, Table 1 represents the
different kernel functions and their different parameters that need to be optimized.
Since different kernel functions have different parameters, the representation of
the solution varies for each kernel. In this paper, an integer representation is used,
where each parameter is represented by a set of genes depending on its range of
20 W. Zouhri et al.

Table 1 Kernel functions and their parameters

Kernel Expression Parameters to tune Range of variation
 T d
Polynomial αx y + r Penalty C [1, 1000]
Constant r [−49, 50]
Slope α [1, 50]
Degree d [1, 10]
 
Sigmoid tanh αx T y + r Penalty C [1, 1000]
Constant r [−49, 50]
Slope α [1, 50]
 
RBF − ||x−y||
2 Penalty C [1, 1000]
2σ 2
e Parameter σ [0.001, 10]

variation. Consequently, the representation of each kernel functions, and the links
that map the genotypes to the phenotypes, are shown in Fig. 2.
Once the three representation are defined, a first population needs to be initialized.
The first population should be diverse to avoid premature convergence, at the same
time, its size should not be too large as it can slow down the algorithm [27]. Many
methods can be found in the literature to initialize the population in a GA, including
random generation, structured generation, and combination of random and structured
generation. The random generation was applied in this paper, as it has been noticed
that a randomly generated population leads to optimality due to its diversity [27].

Polynomial Kernel A0 A1 A2 A3 A4 A5 A6 A7
Variation ranges [0-9] [0-9] [0-9] [0-9] [0-9] [0-4] [0-9] [0-9]

Mapping
A0.100+A1.10+A2+1 A3.10+A4-49 A5.10+A6+1 A7+1
Solution C r α d

Sigmoid Kernel A0 A1 A2 A3 A4 A5 A6
Variation ranges [0-9] [0-9] [0-9] [0-9] [0-9] [0-4] [0-9]

Mapping
A0.100+A1.10+A2+1 A3.10+A4-49 A5.10+A6+1
Solution C r α

RBF Kernel A0 A1 A2 A3 A4 A5 A6
Variation ranges [0-9] [0-9] [0-9] [0-9] [0-9] [0-9] [0-9]

Mapping
A0.100+A1.10+A2+1 A3+A4.0,1+A5.0,01+(A6+1).0,001
Solution C σ

Fig. 2 Genetic representations of Kernels functions

A Genetic-Based SVM Approach for Quality Data Classification 21

2.2 GA-SVM Parameters

• Fitness evaluation

Through any GA, it is necessary to be able to evaluate how good a potential solution
is, compared to other solutions. To do so, a fitness function is defined. The fitness
values are then used for the selection of individuals (parents), on which crossover
and mutation operations will be applied.
For the proposed approach, to get the fitness value of a potential solution, the
following steps should be followed:
1. Split the dataset into training (train), validation (dev), and test (test) sets.
2. Train the SVM model using one potential solution (chromosome).
3. Get the SVM accuracy on the validation set, which represents the fitness value
of the potential solution.
• Selection operation

Selection operation in a genetic algorithm consists of choosing individuals (parents)

for reproduction. Parent selection is very important to the convergence rate of the
genetic algorithm (GA), as good parents generate and yield better and fitter solutions
[28].
In this paper, a 3-way tournament selection is used. This selection method consists
on selecting randomly 3 individuals from the population, and then the one that guar-
antees the best accuracy on the validation set is selected to become a parent. The
choice of this selection method is due to its simplicity, efficiency in parallel and
non-parallel architecture, and also to the non-necessity to sort the population [29].
• Crossover operator

By this operator, individuals (parents) share information by crossing their geno-

types to create better individuals (children). Three different types of crossover are
mainly used in the literature: one-point crossover, two-point crossover, and uniform
crossover [30].
The uniform crossover was considered in this work. In a uniform crossover, each
gene is treated separately. In this, a coin is flipped for each gene to decide whether
or not it’ll be included in the off-spring. This operator doesn’t require the definition
of the crossover rate parameter, as it avoids the genes interchanging of two different
parameters (e.g. C genes with σ genes). Figure 3 depicts the concept of the uniform
crossover.
• Mutation operator

Mutation may be defined as a small random modification in the chromosome to get a

new solution. It is used to introduce diversity in the genetic population and to release
from local minima [31]. To that end, a random resetting mutation is used, where
22 W. Zouhri et al.

Fig. 3 Uniform crossover

Parent 1 0 4 6 0 9 1 5
example
Parent 2 3 7 9 6 8 3 9

Coin 0 0 1 1 0 0 1

Child 1 0 4 9 6 9 1 9
Child 2 3 7 6 0 8 3 5

a random value from the set of permissible values is assigned to randomly chosen
genes.
At this stage, the different techniques and steps of the proposed approach are
described. The Algorithm 1 gives a summary of the approach.
Algorithm 1: Genetic-based SVM approach for Quality data classification
Require Quality dataset for classification
Choose a kernel function of SVM to be optimized
Define the genetic representation of the kernel’s parameters
Define the maximum number of generations: Max_Gen
Generate an initial random population: Pop
Calculate the fitness of each individual of Pop
for each i Max_Gen do:
Select the individuals according to their fitness scores: Pop
Perform uniform crossover on Pop’s individuals: Pop
Perform random resetting mutation on Pop’s individuals: Pop
Calculate the fitness of each individual of Pop
Pick the individual with the highest fitness score: Best_Param[i]
end for
Pick the best individual out of all generations: BEST_KERNEL
return (BEST_KERNEL)
By applying the GA based SVM approach on the two quality datasets, the
following results were found (Table 2). It should be mentioned that having three
types of kernels, resulted in running the approach three times in order to optimize
each one of these types. Once the three runs are performed, the best out of the three
results is picked as the optimal kernel function.
The results show that the RBF kernels give better results compared to the polyno-
mial and the sigmoid kernels for the two datasets, see Tables 2, and 4 in Appendix.
With only two parameters, RBF kernels were easier to tune. Moreover, the optimal
kernel functions did show great performances when tested on new unseen data (test
sets), with an 83.24% accuracy on the chemical dataset, and an 87.67% accuracy on
the rolling process data.
A last test was performed to evaluate the robustness of the optimal kernel functions.
In this case, the Monte-Carlo simulation is used to quantify: “how perturbing the
Table 2 GA-SVM results
Dataset Sets dimensions Optimal kernel Accuracy
Train Dev Test Kernel Parameters Train (%) Dev (%) Test (%)
Chemical data 1168 × 11 585 × 11 584 × 11 RBF C=8 σ = 0.444 83.98 87.15 83.24
Rolling process data 437 × 95 219 × 95 219 × 95 RBF C = 274 σ = 0.051 100 99.08 87.67
Genetic algorithm parameters:
– Population’s size: 100
– Number of generations: 50
A Genetic-Based SVM Approach for Quality Data Classification
23
24 W. Zouhri et al.

Table 3 Robustness assessment regarding measurement uncertainties

Dataset Accuracy of noiseless Average accuracy of Accuracy drop
set (optimal kernel 1000 noisy sets
functions)
Dev (%) Test (%) Dev (%) Test (%) Dev (%) Test (%)
Chemical 87.15 83.24 82.15 80.59 5.00 2.65
data
Rolling 99.08 87.67 98.09 86.81 0.99 0.85
process data

validation set and the test set by Gaussian measurement uncertainties will impact
the SVM model prediction accuracy”. The evaluation is performed by measuring
the average accuracy of 1000 datasets perturbed with a “2.5%”-Gaussian noise in
the first chemical data example and a “15%”-Gaussian noise in the rolling process
data example. These noise rates were estimated by taking into account the precision
and the accuracy of the sensors needed within these two manufacturing systems, and
the environment impact. Table 3 shows the different accuracies drops, due to the
measurement uncertainties.
After performing the Monte-Carlo simulations on both datasets, drops in the
accuracies of the validation and test sets were noticed. For that reason, an awareness
should be raised about the ubiquitous nature of measurement uncertainties in quality
data, and their impact on the predictive performances of the classification methods.

3 Discussion and Conclusion

The average fitness of each generation is calculated in order to see if the results are
evolving and converging to an optimal solution. Also, a short comparison between
the different kernel types is performed in order to select the best one to recommend
for quality data classification.
To show how the GA-SVM progress through generations, Figs. 4 and 5, show the
evolution of the fitness’s average over generations, while the Figs. 6 and 7 show the
evolution of the fittest individual. Linear trend-lines are also added to the graphs to
observe if the fitness is getting better or not over the generations. That is necessary
to prove the efficiency of the proposed approach.
The results showed that the RBF kernel functions have a higher fitness’s average
compared the polynomial and the sigmoid kernel functions, as it has been noticed
that the convergence of the RBF kernel is also faster. For those reasons, and for
being the optimal kernel in both datasets, the RBF kernel is recommended for the
classification of the quality data.
It has been also noticed that the evolution of the fitness’s average doesn’t mean that
the fittest individual over all the generations will appear in the last one. Finally, the
robustness assessment performed using Monte Carlo simulations showed that having
A Genetic-Based SVM Approach for Quality Data Classification 25

Average fitness of every generaƟon RBF

- Chemical Dataset -
0.85
0.84
0.83
0.82
0.81
0.8
0.79
0 10 20 30 40 50

Fig. 4 Average fitness evolution (RBF kernel; Chemical Data)

Average fitness of every generaƟon Sigmoid

- Chemical Dataset -
0.66
0.65
0.64
0.63
0.62
0.61
0.6
0 10 20 30 40 50

Fig. 5 Average fitness evolution (Sigmoid kernel; Chemical Data)

The fiƩest individual of every generaƟon RBF

- Chemical Dataset -
0.872
0.871
0.87
0.869
0.868
0.867
0.866
0.865
0.864
0 10 20 30 40 50

Fig. 6 Fittest individual evolution (RBF kernel; Chemical Data)

26 W. Zouhri et al.

The fiest individual of every generaon

Sigmoid
- Chemical Dataset -
0.9

0.85

0.8

0.75

0.7
0 10 20 30 40 50

Fig. 7 Fittest individual evolution (Sigmoid kernel; Chemical Data)

high prediction accuracies doesn’t mean that the results are robust. Accordingly, a
robustness evaluation should also be performed in order to see how the determined
optimal kernel functions using the GA-SVM approach are robust to measurement
uncertainties.
To conclude, the aim of this study was to analyze the performances of the SVM
technique when dealing with a quality data classification issue. As any other machine
learning tool, the SVM performances are related to the determination of the hyper-
parameters’ values. To do so, a GA-based SVM approach was proposed for the
identification of best hyperparameters allowing the highest accuracy on the valida-
tion set. The approach was applied to two quality datasets. The approach identified
the RBF kernel as the most suitable one for the classification of quality data. At
the end, a robustness evaluation was performed in order to emphasize the impact of
measurement uncertainties on the classifiers’ predictive performances. As a future
work, the proposed GA approach will be applied on Multi-layer perceptron (GA-
MLP) in order to evaluate its efficiency, and to compare it to GA-SVM, also, the key
parameters’ uncertainties will be identified when dealing with SVM approach. That
is necessary to validate the efficiency and reliability of the technique when noisy
data is considered in the context of quality management task.

Appendix

See Tables 4, 5 and 6.

Table 4 Optimal sigmoid and polynomial kernel functions
Optimal sigmoid kernel Accuracy sigmoid Optimal polynomial kernel Accuracy polynomial
Dataset C r α Train (%) Dev (%) Test (%) C r α d Train (%) Dev(%) Test (%)
Chemical data 903 −5 1 83.56 86.13 83.93 12 −8 31 1 80.22 84.93 81.03
Rolling process 419 −27 1 89.47 93.12 59.82 422 26 41 1 100.00 99.08 80.82
data
A Genetic-Based SVM Approach for Quality Data Classification
27
28 W. Zouhri et al.

Table 5 Average fitness of every generation—Chemical Data

Kernels Generations
1 2 3 4 5 6 7 8 9 10 11 12 13
RBF 72.5 87.9 90.4 90.7 91.8 94.0 93.2 91.4 95.7 94.8 95.0 93.1 94.2
(%)
Sigmoid 50.4 51.1 51.1 50.8 50.2 50.0 50.0 50.0 50.0 50.2 50.4 50.7 50.3
(%)
Poly 97.9 98.3 98.5 98.3 98.5 98.4 98.5 98.4 98.4 98.4 98.3 98.2 98.4
(%)
14 15 16 17 18 19 20 21 22 23 24 25 26
RBF 92.9 93.8 92.6 90.5 90.0 92.4 93.0 94.9 95.5 93.8 94.0 93.2 92.2
(%)
Sigmoid 50.0 50.2 50.1 50.1 50.1 50.3 50.0 50.2 50.4 50.1 50.5 50.8 50.8
(%)
Poly 98.7 98.7 98.5 97.5 97.6 98.4 98.4 98.7 97.7 98.6 98.5 98.4 97.5
(%)
27 28 29 30 31 32 33 34 35 36 37 38 39
RBF 91.2 95.2 95.8 91.5 90.5 94.7 94.0 94.8 92.4 93.1 93.8 93.1 95.4
(%)
Sigmoid 51.0 51.4 50.8 50.8 51.3 50.1 51.0 50.6 50.7 50.2 50.5 50.1 50.1
(%)
Poly 98.3 98.5 98.7 98.4 98.5 98.5 98.6 98.5 98.2 98.5 98.5 98.3 98.4
(%)
40 41 42 43 44 45 46 47 48 49 50
RBF 93.1 93.6 92.6 93.4 94.2 95.0 91.9 92.4 93.3 95.4 92.6
(%)
Sigmoid 50.1 50.3 50.4 50.2 50.5 51.4 50.3 50.4 50.6 50.2 50.9
(%)
Poly 98.5 98.4 97.5 98.4 98.6 98.4 98.6 98.2 97.6 97.5 98.7
(%)

Table 6 Average fitness of every generation—Rolling Data

Kernels Generations
1 2 3 4 5 6 7 8 9 10 11 12 13
RBF 79.6 80.8 82.5 83.6 83.9 84.0 84.6 84.3 84.1 84.1 84.2 84.1 84.3
(%)
Sigmoid 61.2 61.4 61.5 60.4 62.4 60.9 62.4 61.3 62.5 62.1 62.3 61.5 63.0
(%)
Poly 67.3 67.0 67.4 66.0 67.1 68.0 68.5 68.3 67.7 67.4 69.0 71.1 70.7
14 15 16 17 18 19 20 21 22 23 24 25 26
RBF 84.3 84.1 84.5 84.1 84.2 83.9 84.2 84.0 84.0 84.5 83.8 84.3 84.0
(%)
(continued)
A Genetic-Based SVM Approach for Quality Data Classification 29

Table 6 (continued)
Kernels Generations
1 2 3 4 5 6 7 8 9 10 11 12 13
Sigmoid 64.2 63.9 63.3 63.4 63.9 63.5 63.5 63.3 63.5 64.1 64.5 62.7 63.4
(%)
Poly 69.1 69.6 68.1 70.7 70.0 73.1 72.1 66.8 70.8 70.3 70.4 72.3 70.8
(%)
27 28 29 30 31 32 33 34 35 36 37 38 39
RBF 84.0 84.2 84.1 84.0 83.3 83.8 84.0 83.7 84.0 84.2 83.6 84.5 84.6
(%)
Sigmoid 63.1 63.1 61.7 62.5 62.8 62.4 63.1 62.6 63.2 63.6 63.1 62.9 62.3
(%)
Poly 71.5 70.8 71.6 699. 71.5 68.6 69.2 66.9 69.7 68.8 71.4 71.2 71.9
(%)
40 41 42 43 44 45 46 47 48 49 50
RBF 84.2 83.7 83.9 84.0 84.6 84.7 84.2 84.3 84.2 84.3 84.5
(%)
Sigmoid 62.3 63.0 63.8 645. 64.4 64.2 63.8 63.7 64.3 64.5 65.4
(%)
Poly 70.7 71.9 72.2 70.0 71.3 72.1 72.4 72.8 72.6 71.7 74.0
(%)

References

1. Alcácer, V., Cruz-Machado, V.: Scanning the Industry 4.0: a literature review on technologies
for manufacturing systems. Eng. Sci. Technol. Int. J. 22(3), 899-919 (2019). https://doi.org/10.
1016/j.jestch.2019.01.006
2. Rostami, H., Dantan, J.-Y., Homri, L.: Review of data mining applications for quality assess-
ment in manufacturing industry: support vector machines. Int. J. Metrol. Qual. Eng. 6(4), 401
(2015). https://doi.org/10.1051/ijmqe/2015023
3. Köksal, G., Batmaz, İ., Testik, M.C.: A review of data mining applications for quality improve-
ment in manufacturing industry. Expert. Syst. Appl. 38(10), 13448–13467 (2011). https://doi.
org/10.1016/j.eswa.2011.04.063
4. Attar, M.E., Hamery, X.: Industrial expert system acquired by machine learning. Appl. Artif.
Intell. 8(4), 497–542 (1994). https://doi.org/10.1051/ijmqe/201502310.1080/088395194089
45457
5. Siltepavet, A., Sinthupinyo, S., Chongstitvatana, P.: Improving quality of products in hard drive
manufacturing by decision tree technique. J. Comput. Sci. Issues 9(3), 7 (2012)
6. Adidela, D.R.: Construction of fuzzy decision tree using expectation maximization algorithm.
J. Comput. Sci. Manag. Res. 1(3), 9 (2012)
7. Wray, B.A., Rakes, T.R., Rees, L.P.: Neural network identification of critical factors in a
dynamic just-in-time kanban environment. J. Intell. Manufact. 8(2), 83–96 (1997). https://doi.
org/10.1023/A:1018548519287
8. Markham, I.S., Mathieu, R.G., Wray, B.A.: A rule induction approach for determining the
number of kanbans in a just-in-time production system. Comput. Ind. Eng. 34(4), 717–727
(1998). 10.1016/S0360–8352(98)00099-0
30 W. Zouhri et al.

9. (Tony) Hou, T.-H., Liu, W.-L., Lin, L.: Intelligent remote monitoring and diagnosis of manu-
facturing processes using an integrated approach of neural networks and rough sets. J. Intell.
Manufact. 14(2), 239–253 (2003).https://doi.org/10.1023/A:1022911715996
10. Cho, H. S., Leu, M.C.: Artificial neural networks in manufacturing processes: monitoring and
control. IFAC Proc. Vol. 31(15), 529–537 (1998). https://doi.org/10.1016/S1474-6670(17)406
07-0
11. Zahran, B.M.: Using neural networks to predict the hardness of aluminum alloys. Eng. Technol.
Appl. Sci. Res. 5(1), 4 (2015)
12. Pouraliakbar, H., Khalaj, M., Nazerfakhari, M., Khalaj, G.: Artificial neural networks for hard-
ness prediction of HAZ with chemical composition and tensile test of X70 pipeline steels. J.
Iron Steel Res. Int. 22(5), 446–450 (2015). 10.1016/S1006-706X(15)30025-X
13. Noyel, M., Thomas, P., Charpentier, P., Thomas, A., Beauprêtre, B.: Improving production
process performance thanks to neuronal analysis. IFAC Proc. Vol. 46(7), 432–437 (2013).
https://doi.org/10.3182/20130522-3-BR-4036.00055
14. Bahlmann, C., Heidemann, G., Ritter, H.: Artificial neural networks for automated quality
control of textile seams. Pattern Recognit. 32(6), 1049–1060 (1999). https://doi.org/10.1016/
S0031-3203(98)00128-9
15. Gryllias, K.C., Antoniadis, I.A.: A Support Vector Machine approach based on physical model
training for rolling element bearing fault detection in industrial environments. Eng. Appl. Artif.
Intell. 25(2), 326–344 (2012). https://doi.org/10.1016/j.engappai.2011.09.010
16. Fernández-Francos, D., Martínez-Rego, D., Fontenla-Romero, O., Alonso-Betanzos, A.: Auto-
matic bearing fault diagnosis based on one-class ν-SVM. Comput. Ind. Eng. 64(1), 357–365
(2013). https://doi.org/10.1016/j.cie.2012.10.013
17. Xanthopoulos, P., Razzaghi, T.: A weighted support vector machine method for control chart
pattern recognition. Comput. Ind. Eng. 70, 134–149 (2014). https://doi.org/10.1016/j.cie.2014.
01.014
18. Diao, G., Zhao, L., Yao, Y.: A dynamic quality control approach by improving dominant factors
based on improved principal component analysis. Int. J. Prod. Res. 53(14), 4287–4303 (2015).
https://doi.org/10.1080/00207543.2014.997400
19. Baccarini, L.M.R., Rocha, V.V., Silva, de Menezes, B.R., Caminhas, W.M.: SVM practical
industrial application for mechanical faults diagnostic. Expert Syst. Appl. 38(6), 6980–6984
(2011). https://doi.org/10.1016/j.eswa.2010.12.017
20. Jegadeeshwaran, R., Sugumaran, V.: Fault diagnosis of automobile hydraulic brake system
using statistical features and support vector machines. Mech. Syst. Signal Process. 52–53,
436–446 (2015). https://doi.org/10.1016/j.ymssp.2014.08.007
21. Lessmann, S., Stahlbock, R., Crone, S.F.: Optimizing Hyperparameters of Support Vector
Machines by Genetic Algorithms, p. 7
22. Bäck, T., Schwefel, H.-P.: An overview of evolutionary algorithms for parameter optimization.
Evol. Comput. 1(1), 1–23 (1993). https://doi.org/10.1162/evco.1993.1.1.1
23. Chou, J.-S., Cheng, M.-Y., Wu, Y.-W., Pham, A.-D.: Optimizing parameters of support vector
machine using fast messy genetic algorithm for dispute classification. Expert Syst. Appl. 41(8),
3955–3964 (2014). https://doi.org/10.1016/j.eswa.2013.12.035
24. Lin, S.-W., Ying, K.-C., Chen, S.-C., Lee, Z.-J.: Particle swarm optimization for parameter
determination and feature selection of support vector machines. Expert Syst. Appl. 35(4),
1817–1824 (2008). https://doi.org/10.1016/j.eswa.2007.08.088
25. Yang, D., Liu, Y., Li, S., Li, X., Ma, L.: Gear fault diagnosis based on support vector machine
optimized by artificial bee colony algorithm. Mech. Mach. Theory 90, 219–229 (2015). https://
doi.org/10.1016/j.mechmachtheory.2015.03.013
26. Olatomiwa, L., Mekhilef, S., Shamshirband, S., Mohammadi, K., Petković, D., Sudheer, C.: A
support vector machine–firefly algorithm-based model for global solar radiation prediction.
Solar Energy 115, 632–644 (2015). https://doi.org/10.1016/j.solener.2015.03.015
27. Baker, B.M., Ayechew, M.A.: A genetic algorithm for the vehicle routing problem. Comput.
Oper. Res. 30(5), 787–800 (2003). 10.1016/S0305–0548(02)00051–5
28. Jebari, K., Madiafi, M.: Selection methods for genetic algorithms. Int. J. Emer. Sci. 3(4) (2013)
A Genetic-Based SVM Approach for Quality Data Classification 31

29. Fang, Y., Li, J.: A review of tournament selection in genetic programming. In: Cai, Z., Hu, C.,
Kang, Z., Liu, Y. (eds.) Advances in Computation and Intelligence, vol. 6382, pp. 181–192.
Springer, Berlin, Heidelberg (2010)
30. Walchand College of Engineering, U. A.J., S. P.D., Government College of Engineering, Karad.:
Crossover operators in genetic algorithms: a review. IJSC 06(01), 1083–1092 (2015). https://
doi.org/10.21917/ijsc.2015.0150
31. Angelova, M., Pencheva, T.: Tuning genetic algorithm parameters to improve convergence
time. Int. J. Chem. Eng. 2011, 1–7 (2011). https://doi.org/10.1155/2011/646917
Towards a Platform to Implement
an Intelligent and Predictive
Maintenance in the Context
of Industry 4.0

El Mehdi Bourezza and Ahmed Mousrij

Abstract In the world of the 4th industrial revolution and in order to master
the production tools, the company must have a relevant maintenance manage-
ment system. Therefore, it is necessary to research and develop new maintenance
approaches in the context of Industry 4.0, in order to digitize the manufacturing
process and generate information to detect failures and act in real time. This paper
aims to present the draft of an implementation approach for an intelligent platform
of industrial maintenance, aligned with the principles of Industry 4.0. This platform
consists in acquiring and conditioning the data to analyze them in order to detect
failures and to estimate the time of the good functioning of a device. Then the choice
of the appropriate procedure is provided by the decision support, which sends it in
turn for it to be planned and executed. Finally, an evaluation module to check the
smooth execution.

Keywords Industry 4.0 · Predictive maintenance · Maintenance architecture

platform · Intelligent approach

1 Introduction

In the world of the 4th industrial revolution and the changing consumer behavior,
more and more demanding, in a very competitive market. The company must control
their production tools while having a relevant maintenance management system.
Therefore, it is necessary to research and develop new maintenance approaches that
are part of Industry 4.0, in order to digitize the manufacturing process and generate
information to detect failures and act in real time. In other words, the intercon-
nection of the physical environment allows the acquisition and processing of data

E. M. Bourezza (B) · A. Mousrij

Laboratory of Engineering of Industrial Management and Innovation, Hassan First University,
Settat, BP, Morocco
e-mail: [email protected]
A. Mousrij
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 33

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_3
34 E. M. Bourezza and A. Mousrij

to extract knowledge and optimize the production process in terms of downtime,

planning interventions, the cost of maintenance, etc. The use of intelligent mainte-
nance approaches requires the application of advanced analysis techniques and data
processing (IoT, Big Data, IA, Cloud Computing …) which represent the basis of the
concepts of Industry 4.0. These approaches are decision supports and handle large
amounts of data. However, processing a mass of real-time data is a major challenge
for decision-making systems [1]. In addition, despite the advantages of implementing
the 4.0 maintenance system, the maintenance architectures proposed as part of the
industry of the future are not able to work in an integrated way and are not yet mature
enough [2]. This paper aims to develop an intelligent platform for implementation
of industrial maintenance, aligned with the principles of Industry 4.0, which allows
the acquisition of data collected in production lines. After analyzing the informa-
tion collected, we arrive at the Monitoring and Failure Prediction phase in order
to implement maintenance procedures. These will be planned for being performed
using augmented reality. Finally, the purpose of the evaluation phase is to monitor
the execution of procedures in an efficient manner. To understand the scope of our
project, we will first go through a literature review of published works in order to
build a comparative study between the proposed architectures. Secondly, we will
identify the conditions for success as well as the limitations that can hinder the
implementation of intelligent maintenance. These studies will then allow presenting
in Sect. 3, the draft of an intelligent platform of industrial maintenance in context
4.0 to apply to Moroccan SMEs for a global and efficient maintenance management.
The last section is devoted to conclusions and future work.

2 Literature Review and Comparison

Since the arrival of industry 4.0 concepts, many architectures have been proposed by
academics to implement intelligent maintenance in production lines [2] and [3] have
proposed two architectures of an intelligent and predictive maintenance system that
meet Industry 4.0 principles. The architectures focus on the in-depth analysis of the
information collected for the early detection of possible failures and provide intelli-
gent decision support guided during maintenance interventions. However, these two
proposals do not deal with the scheduling module of the procedures to be executed
as well as the evaluation module, which makes it possible to control the execu-
tion. In addition, no method used for the analysis and design of architectures. The
approaches proposed by Bousdekis and Mentzas [4] and Al-Najjar et al. [5] have
the same characteristics and constitute the same stages as the two previous proposi-
tions. Nevertheless, they do not follow a frame of reference like the OSA-CBM of
MIMOSA used by other architectures and which presents an application of the ISO
13,374 standard for the monitoring of the state and diagnosis of the machines. More-
over, we notice the absence of the visualization module, which makes it possible
to follow the instantaneous changes [6]. presented a smart predictive maintenance
system that consists of the modules needed to achieve the goal of Industry 4.0. On
Towards a Platform to Implement an Intelligent and Predictive Maintenance … 35

the other hand, it does not talk about execution and evaluation modules, as well as the
lack of a methodology in the design. Other architectural proposals for the implemen-
tation of intelligent and predictive maintenance as well as the approaches described
previously are summarized in the comparative Table 1. By analysing this table, we
find that all the architectures do not deal with the planning modules of the mainte-
nance and evaluation procedures of the interventions. We also note the absence of a
method of analysis and design of proposed approaches. Some architectures also do
not meet the OSA-CBM standard and others do not consider the operational aspect
of maintenance.

3 Key Factors to Develop Successfully the Maintenance

System 4.0

The implementation of a 4.0 maintenance system looks like a typical change project
with all its challenges. In addition, to develop a system of intelligent and predictive
maintenance, it is important to consider all the key factors of success. According to
[1], it is necessary to take into account all the interdependencies between the modules
constituting the maintenance system 4.0 [8]. have added other important criteria for
the development of a maintenance system aligned with the principles of Industry 4.0.
We quote in this sense: The integration of physical resources, the interoperability of
heterogeneous modules thanks to the exchange of data, the flexibility of the proposed
architecture to adapt to the changes as well as to have a generic and evolutive aspect
in order to follow the conditions of use. In addition, in the era of Industry 4.0, there
is a variety of technology groups (Fig. 1). Therefore, to have a given 4.0 process,
one must choose the appropriate technology group, as well as their methods and
techniques to adopt.
On the other hand, it is important to follow a standard repository that provides the
necessary steps and best practices to develop an intelligent and predictive system of
maintenance. The OSA-CBM standard from MIMOSA [13] is used to detail the ISO
13347 standard for equipment monitoring and diagnostics. It comprises six steps
as shown in Fig. 2 explained at the level of the bibliographic reference. We will
build on this standard to develop our intelligent industrial maintenance platform.
However, a 4.0 maintenance system faces a number of challenges that we need to
consider including: the large amount of data from different sources can influence the
accessibility, quality and fusion of information [14]. In addition, the 4.0 maintenance
system must be able to process data with industrial big data and have the flexibility
to acquire and store information [15]. Finally, it is possible to fall in front of useless
maintenance works because of the inaccurate predictors, whereas the precision of
the forecasts is a critical point that can generate waste in term of time, production,
cost etc.…
 36

Table 1 Comparison between proposed architectures

Data Data analysis Visualization of Failure Decision Maintenance Maintenance Evaluation of Standard
Collection data detection making planning performance the procedure OSA_CBM
[2] ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✕ ✓
[4] ✓ ✓ ✕ ✓ ✓ ✓ ✓ ✕ ✕
[7] ✓ ✓ ✓ ✓ ✓ ✕ ✕ ✕ ✓
[8] ✓ ✓ ✓ ✓ ✓ ✕ ✕ ✕ ✕
[9] ✓ ✓ ✓ ✓ ✓ ✕ ✓ ✕ ✕
[10] ✓ ✓ ✕ ✓ ✓ ✕ ✓ ✕ ✓
[5] ✓ ✓ ✕ ✓ ✓ ✕ ✓ ✕ ✕
[4] ✓ ✓ ✓ ✓ ✓ ✕ ✓ ✕ ✓
[6] ✓ ✓ ✓ ✓ ✓ ✓ ✕ ✕ ✓
[11] ✓ ✓ ✕ ✓ ✓ ✕ ✕ ✕ ✓
[12] ✓ ✓ ✕ ✓ ✓ ✕ ✕ ✕ ✕
E. M. Bourezza and A. Mousrij
Towards a Platform to Implement an Intelligent and Predictive Maintenance … 37

Fig. 1 Industry 4.0

technology groups

Big
Data

Cloud IA

CPS
Process IOT
4.0

Simul-
Robot
ation

AR

Fig. 2 Steps of the

OSA-CBM standard Advisory Generetion
Prognostics Assessment
Health Assessment
State Detection
Data Manipulation
Data Acquisition

Sensors / Transducer / Manual Entry

4 Platform of Industrial Maintenance in a Context 4.0

The literature review discussed in the second section and the comparative study
between the different proposals, as well as the consideration of key success factors
and compliance with the MIMOSA OSA-CBM guidelines, allowed us to develop
the architecture of our maintenance platform 4.0 articulated in six complementary
phases, Fig. 3.
Figure 5 presents the synthesis of our scientifically designed platform using the
SADT method Fig. 4, (Structured Analysis and Design Technics). It is a systemic
approach to modeling a complex system or an operating process. Each function of
the modeled system (process) is presented as a subsystem as follows.
38 E. M. Bourezza and A. Mousrij

Fig. 3 The different phases

of the 4.0 industrial Evaluation
maintenance platform
Execution

Decision Support

Diagnosis & Pronosis

Data Analysis

Data Acquisition & Conditioning

Fig. 4 Method SADT Energy

Configuration

Setting

Input Exploitation
Global function
Output

System

The actigram is characterized by an input/output flow, its global function and

all the constraints that must be taken into account.The proposed architecture of our
platform consists of several modules to design a functional system, which allows
implementing an intelligent process of industrial maintenance based on the principles
of Industry 4.0. In the following paper, we will detail the modules of our platform.

4.1 Data Acquisition and Conditioning Module

This module is divided into two components, the acquisition and conditioning of
data. The acquisition component is used to collect automatically or manually (HMI)
information integrating the maintenance process from several different sources using
the Internet of Things. On the other hand, the conditioning component consists in
filtering the data by the elimination of disturbances and the smoothing of the signals
that carry noise. In addition, it allows merging the data to avoid unnecessary and
Towards a Platform to Implement an Intelligent and Predictive Maintenance … 39

repetitive information. It is also used to transform the data into an appropriate format
so that it can be used in a way that other modules can understand.
Finally, the last operation deals with reducing the size of the data by applying
methods of extracting signal characteristics in this direction.

4.2 Data Analysis Module

It contains a database for temporary storage, it allows analyzing the data by applying
new emerging technologies such as advanced analysis and machine learning to detect
and identify failures in order to generate knowledge allowing to trigger an automatic
correction without human intervention or initiate a maintenance procedure to return
the system to its normal state. When analyzing new data, this module takes historical
data into account, which allows new insights from these correlations. Neural networks
(RNN) such as LSTM (Long Short Term Memory) and GRU (Gated Recurrent Units)
applied in maintenance are used at this stage to control deep learning (Fig. 5).

IOT, Sensors, Ta blette, Mobile .. .

Internet
Availibility, Sec urity
Energy

Devices, Acquire &

Production,
condition data
Exploit able dat a

Process ...
Sta tistica l Methods
Machine le arning, Dee p Lea rning
Availibility, Sec urity
SoŌware
Energy
Acquisition System Level
Updates
Store & Proceedings
Analyze Data facts
Rules
Dashboa rd
Internet
KPI
Datawarehouse Energy

&
Analysis system
Visualize, detect Maintenance
failures + Prognosis Warning

Hardware Augme nted R ea lity

C loud C omputing
Maintenance expert

Level Dynamic Monitoring

Energy

Order Intervention
Choose maintenance Maintenance + Procedure
procedure ApplicaƟon
Heuristic algorithm s
Exec ution time

Human R esourc es, Materia l .. .

Level
Availability, state of produc tion a nd logistics
Decision Making

Schedule
Scheduled Procedure

maintenance tasks
AR Proc edure Visualiza tion
HMI
HMD
Schedule
Executed procedure

maintenance tasks

Perform
M odels
Sta nda rds
the maintenance
Sec urity

Maintenance
Evaluate and technician
standardize the
Executed Procedure procedure performed

Control system

Fig. 5 Block diagram of the intelligent maintenance platform

40 E. M. Bourezza and A. Mousrij

Sensors/Tranducer/Manual Entry

Fig. 6 Platform architecture

4.3 Diagnosis Module and Prognosis

This module provides real-time monitoring by generating a maintenance warning

whenever a failure occurs. It consists of three components: visualization, diagnosis
and prognosis.
The visualization represents a dashboard that allows you to follow the evolution of
physical quantities in real time such as pressure, temperature, vibration, flow, etc.…
It also makes it possible to visualize and report the critical state of the production
process when an indicator leaves its tolerance range.
The diagnostics component uses deep learning to detect probable failures,
applying the instructions and rules generated by the analysis module, to evaluate
the existing facts that are in the acquisition module database.
Towards a Platform to Implement an Intelligent and Predictive Maintenance … 41

The third component of this module is the prognosis for estimating the operating
time of a system component prior to failure using the RUL (Remaining Useful Life)
concept.

4.4 Decision Support Module

After receiving the maintenance warning, the decision support module looks in its
database for the correct procedure to remedy a given problem. In the event that the
requested procedure does not exist, a process of creation or adaptation that uses cloud
computing will be launched to feed and enrich the database with new procedures to
ensure effective interventions.

4.5 Planning Module

It makes it possible to plan the maintenance intervention sent by the decision support
module, by the application of the heuristic algorithms, the taking into account of the
necessary resources (human, material …) and the execution time determined by the
operator according to the maintenance procedure. It also considers the availability
of the system as well as the state of production.

5 Platform Architecture

The functional decomposition of our platform allows us to propose a hierarchical

architecture Fig. 6 to be established at the level of the industrial sector that supports
the set of challenges discussed in Sect. 3 by ensuring an intelligent and efficient
management. This architecture makes it possible to manage the heterogeneous data
collected from different sources in an intelligent way, in order to guarantee the imple-
mentation of intelligent maintenance within the industrial entities. This architecture
consists of three levels, namely a hardware level, a software level and an applica-
tion level. The hardware level is the chain of acquisition of data collected from a
distributed sensor network that represents the physical environment by taking into
account the constraints of synchronization, correlation and data quality. It also allows
data to be transmitted to higher levels of our architecture using different communi-
cation protocols. As a third function of this level is to ensure the digital conversion
of data in order to have a data model exploitable in terms of analysis and processing.
The software level consists of storing, analyzing, processing data and making deci-
sions using artificial intelligence algorithms. It allows dynamic data management to
42 E. M. Bourezza and A. Mousrij

be viewed in real time for the purpose of monitoring equipment status. The applica-
tion level is the top level of this architecture, it allows to deploy the results of this
platform through interaction with stakeholders.

6 Predictive Maintenance Platform for Vibration Analysis

of Rotating Machines

Our predictive maintenance platform constitutes a decision-making system whose

objective is to ensure real-time monitoring of the equipment of a given production
line (Fig. 7). There are several techniques for this: vibration analysis, thermographic
analysis, oil analysis, etc. In our case, we chose the vibrational analysis of rotating
machines as an application case because it is more used to detect the appearance and
evolution of most mechanical defects. The advantage of our platform is to guarantee
real-time analysis and decision-making. The figure below illustrates the functioning
of our platform, it makes it possible to acquire signals delivered by the vibration
sensors installed on precise measurement points (bearings, fixing points, etc.). These
signals will be subjected to conditioning in order to improve their qualities and to
meet application constraints. Arriving at the analysis block, this information will be
compared against the reference models of the predetermined vibration phenomena.
Each reference model will be developed empirically, it has characteristics of judgment
which will specify the admissible vibration levels. The tolerance intervals for each
vibration phenomenon as well as the instantaneous variation will be displayed on a

VibraƟon Phenomenon
Unbalance Misalignment Bearings Others

MathemaƟcal OR Empirical model

Jugement Features

Reference model

Data
SC1 Analysis

SC2
MulƟplexer

A/D
SC3 Converter
SC4

SCn Database

Fig. 7 Predictive maintenance platform for vibration analysis of rotating machines

Towards a Platform to Implement an Intelligent and Predictive Maintenance … 43

dashboard which allows monitoring the state of the system in order to make decisions
and act at the right time. The maintenance intervention can be in real time as it can
be planned in a systematic maintenance program, it depends on the severity of the
vibration phenomenon.

7 Conclusions

The architecture proposed in this paper presents a decision-making platform whose

objective is to ensure real-time monitoring of the equipment of a given production
line.
Our platform consists in acquiring and conditioning the data to be analyzed in
order to detect breakdowns and to estimate the time of good functioning of an equip-
ment. Then, the choice of the appropriate procedure is ensured by the decision aid,
which in turn sends it so that it can be planned and executed. We chose the vibrational
analysis of rotating machines as an application case because it is more used to detect
the appearance and evolution of most mechanical faults, but it should be noted that
it is possible to implant the architecture using thermographic analysis.
The proposed version is an initial version. Throughout our research, this architec-
ture will be subject to several modifications until total satisfaction. Our future work
will be devoted to the development of reference models of vibrational phenomena
from empirical models developed following an experimental study. In addition, we
will also devote ourselves to the design of the hardware part of our platform.

References

1. Thoben, K.-D., Ait-Alla, A., Franke, M., Hribernik, K., Lütjen, M., Freitag, M.: Real-time
predictive maintenance based on complex event processing. In: Enterprise Interoperability,
pp. 291–296 (2018). 10.1002/9781119564034.ch36
2. Cachada, A., Barbosa, J., Leitno, P., Gcraldcs, C. A. S., Deusdado, L., Costa, J., Romero,
L.: Maintenance 4.0: intelligent and predictive maintenance system architecture. In: 2018 IEEE
23rd International Conference on Emerging Technologies and Factory Automation (ETFA)
(2018). https://doi.org/10.1109/etfa.2018.8502489
3. Bousdekis, A., Lepenioti, K., Ntalaperas, D., Vergeti, D., Apostolou, D., Boursinos, V.: A RAMI
4.0 view of predictive maintenance: software architecture, platform and case study in steel
industry. In: Proper, H., Stirna, J. (eds.) Advanced Information Systems Engineering Work-
shops. CAiSE 2019. Lecture Notes in Business Information Processing, vol. 349. Springer,
Cham (2019)
4. Bousdekis, A., Mentzas, G.: Condition-based predictive maintenance in the frame of industry
4.0. In: Lödding, H., Riedel, R., Thoben, K.D., von Cieminski, G., Kiritsis, D. (eds.) Advances
in Production Management Systems. The Path to Intelligent, Collaborative and Sustainable
Manufacturing. APMS 2017. IFIP Advances in Information and Communication Technology,
vol. 513. Springer, Cham (2017)
44 E. M. Bourezza and A. Mousrij

5. Al-Najjar, B., Algabroun, H., Jonsson, H.: Smart maintenance model using cyber physical
system. In: International Conference on “Role of Industrial Engineering in Industry 4.0
Paradigm” (ICIEIND), Bhubaneswar, India, September 27–30, pp. 1–6 (2018)
6. Wang,K.: Intelligent Predictive Maintenance (IPdM) system Industry 4.0 scenario. WIT Trans.
Eng. Sci. 113 (2016). https://doi.org/10.2495/IWAMA150301
7. Canito A. et al.: An architecture for proactive maintenance in the machinery industry. In: De
Paz, J., Julián, V., Villarrubia, G., Marreiros, G., Novais, P. (eds.) Ambient Intelligence—
Software and Applications—8th International Symposium on Ambient Intelligence (ISAmI
2017). Advances in Intelligent Systems and Computing, vol. 615. Springer, Cham (2017)
8. Peres, R.S., Dionisio, A., Leitao, P., Barata, J.: IDARTS—Towards intelligent data analysis
and real-time supervision for industry 4.0. Comput. Ind. 101, 138–146 (2018). https://doi.org/
10.1016/j.compind.2018.07.004
9. Li, Z.: A Framework of Intelligent Fault Diagnosis and Prognosis in the Industry 4.0 Era,
Doctoral theses at Norwegian University of Science and Technology (2018)
10. Algabroun, H., Iftikhar, M.U., Al-Najjar, B., Weyns, D.: Maintenance 4.0 framework using
self-adaptive software architecture. In: Proceedings of 2nd International Conference on
Maintenance Engineering, IncoME-II 2017. The University of Manchester, UK (2017)
11. Ferreira, L.L., Albano, M., Silva, J., Martinho, D., Marreiros, G., di Orio, G., Ferreira, H.,
et al.: A pilot for proactive maintenance in industry 4.0. In: 2017 IEEE 13th International
Workshop on Factory Communication Systems (WFCS) (2017). 10.1109/wfcs.2017.7991952
12. Galar, D., Thaduri, A., Catelani, M., Ciani, L.: Context awareness for maintenance decision
making: a diagnosis and prognosis approach. Measurement 67, 137–150 (2015). https://doi.
org/10.1016/j.measurement.2015.01.015
13. Mimosa—An Operations and Maintenance Information Open System Alliance, “MIMOSA
OSA-CBM,” 2010. [Online]. https://www.mimosa.org/mimosa-osa-cbm.
14. Aljumaili, M., Wandt, K., Karim, R., Tretten, P.: eMaintenance ontologies for data quality
support. J. Qual. Maint. Eng. 21(3), 358–374 (2015). https://doi.org/10.1108/JQME-09-2014-
0048
15. Liu, J., Dietz, T., Carpenter, S.R., Alberti, M., Folke, C., Moran, E., Taylor, W.W., et al.:
Complexity of coupled human and natural systems. Science 317(5844), 1513–1516 (2007).
https://doi.org/10.1126/science.1144004
Towards a Prediction Analysis
in an Industrial Context

Ilham Battas, Ridouane Oulhiq, Hicham Behja, and Laurent Deshayes

Abstract The efficiency of a mine’s fixed facilities depends on several parameters,

which makes its analysis and improvement very complex. The purpose of this article
is to propose an approach to predict the efficiency of the fixed facilities based on data
mining algorithms. The input of this approach is the data concerning the parameters
influencing the efficiency. The output is the predicted efficiency of the facilities. This
approach consists of four steps: domain comprehension, data preprocessing, devel-
opment of prediction models and finally validation and verification of the proposed
models. It has been applied for a Moroccan mining company and the results of this
case study are also presented.

Keywords Data mining · Prediction · Industrial efficiency

1 Introduction

In the recent years, data mining has been used by a large number of companies within
different industrial sectors around the world. Data mining combines statistical anal-
ysis, artificial intelligence and advanced technologies to extract relationships and

I. Battas (B)
Research Foundation for Development and Innovation in Science and Engineering, 16469
Casablanca, Morocco
e-mail: [email protected]
I. Battas · H. Behja
Engineering Research Laboratory (LRI), Modeling System Architecture and Modeling Team
(EASM), National and High School of Electricity and Mechanic (ENSEM) Hassan II University,
8118 Casablanca, Morocco
e-mail: [email protected]; [email protected]
I. Battas · R. Oulhiq · L. Deshayes
Innovation Lab for Operations, Mohammed VI Polytechnic University, 43150 Benguerir, Morocco
e-mail: [email protected]
L. Deshayes
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 45

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_4
46 I. Battas et al.

patterns from huge databases [1]. To carry out a data mining project, various tools
can be used such as segmentation, where the goal is to discover groups and structures
within the studied databases, association tools to identify rules, and finally predic-
tion tools [2]. Data mining is a part of the general process of knowledge discovery
in databases (KDD), defined as “the nontrivial process of identifying valid, novel,
potentially useful, and ultimately understandable patterns in data”. It summarily
involves preparing data, searching for patterns, and evaluating and refining the
extracted knowledge, all repeated in several iterations [3]. Manufacturing has become
a potentially promising area of application for data mining to extract knowledge for
performance improvement purposes. Indeed, there exists several indicators to eval-
uate the performance of a company; a crucial one is the efficiency of facilities, which
combines several factors such as the efficiency of materials and human resources,
effectiveness of different resources management strategies and other external and
internal factors. The analysis of efficiency in the industrial context is often limited
to standard methods witch are strongly related to the companies core business and
mainly inspired by lean manufacturing and lean management methods. Indeed, the
current manufacturing industry methods define the predicted efficiency as the result
of the ratio between the predicted quantity to be produced by production lines and the
production cycle time. While the data mining tools uses both the history data of the
predicted efficiency and the history data of all parameters influencing the efficiency,
such as field data and meteorological data, to determine predicted efficiency. This
work is being carried out within the framework of industry 4.0. It proposes a new
approach to the development of a system based on intelligent algorithms to predict
the efficiency of industrial facilities. As powerful tools, these algorithms will then
be adapted to support managers decision-making.
In this paper, we present a data mining approach for predicting facilities efficiency
based on extensive data acquisition over all facilities. Its main objective is to predict
efficiency based on influencing factors within the process chain. Thus, these critical
factors could be identified as a basis for efficiency improvement by identifying the
modifications to be made in order to achieve a desired efficiency.
This approach consists in four main steps. The first consists in having in-depth
information and knowledge about the application domain, the mining sector’s fixed
installations, in order to determine the set of parameters influencing the screening
unit’s efficiency. The second concerns the data preprocessing of these parameters,
in order to improve their quality before being used by the third step, which aims to
develop predictive models that will be validated and verified with different evaluation
criteria during the last step. In the case study section, the results will also be detailed.

2 Domain Comprehension

The domain comprehension is a very important step, it allows to understand the

application domain and to have a detailed view of its functioning, which allows to
Towards a Prediction Analysis in an Industrial Context 47

better understand the problem and to determine all the parameters that can have an
impact on the problem.

3 Data Preprocessing

Data pre-processing is a technique used to improve the quality of the database and
real-world data warehouses. It includes five steps detailed bellow [4].
• Data comprehension: analyze the existing data in order to use them correctly. The
target variable is the one to be predicted and the explanatory variables are the
independent variables that will help us predict the target [2].
• Data integration: fuses data from multiple sources into a consistent database
[2, 4, 5].
• Data cleaning: reduce and weaken the effect of missing and outlier values in the
database [2, 5, 6].
• Data standardization: consist of applying the scaling of functionalities. Indeed
models are based on Euclidean distance, if, the values of one column (x) are
much higher than those of another column (y), the square difference between the
lower and upper value is greater, the lower value will almost be treated as if it
did not exist. For this reason, it is necessary to transform all our variables into a
single scale [2, 4].
• Data reduction: Reducing the number of variables consists in selecting the
variables that are highly correlated with each other and keeping only one [2, 4].

3.1 Model Development

Data mining offers algorithms and tools for the discovery of comprehensible models
from a large amount of data. These algorithms fall into three categories: prediction,
segmentation and frequent pattern search.
Predictive models are designed to determine a function (or model) that combines
inputs and outputs. Depending on the nature of the output variable, two subcategories
of tools exist:
• If the output variable is of a discrete type, the classification will be used to build
the model that will allow the records to be correctly classified.
• If the output variable is of a continuous type, the usual statistical regression tools
are most commonly used.

Segmentation is unsupervised learning (which does not define “inputs” or “out-

puts”). It aims to identify sets of elements that share certain similarities. Indeed,
segmentation algorithms maximize homogeneity within each set and maximize
48 I. Battas et al.

heterogeneity between sets. Different methods are used to define these groups:
k-means, hierarchical algorithms, neural networks…
Description (or identification of frequent patterns) consists in explaining the existing
relationships in the data. Link analysis and visualization techniques are commonly
used for this purpose. In fact, visualization techniques are used to simplify the
comprehension of data using adapted graphical representations.
In this study we will focus on predictive regression models. In the litera-
ture, there are several predictive regression models; we have chosen to adopt the
most commonly used models, namely: Multiple linear regression, Support vector
regression, Regression tree, and K nearest neighboring [7].

3.2 Model Evaluation

Among the evaluation criteria most commonly used for regression models are the
score R2 and the root mean square error RMSE [8, 9].

4 Case Study

The different parts that have been presented along the previous sections have been
applied to the database of mine’s fixed facilities. They are critical and important
parts, thus developing a system for the prediction of their efficiency can represents
several benefits for performance enhancement.

4.1 Domain Comprehension

The fixed facilities process is divided into three main steps described below (Fig. 1).

Destoning
The destoning facility is located 2 km from the screening facility. The connection
between these two facilities is provided by the T1 conveyor. The destoning facility
consists of:
• Feeding hoppers: The two hoppers are designed to receive the phosphate
transported by the trucks.
• Destoning: destoning is done by screens. The rejection of the screens (class >
90 mm) feeds a channel which discharges into the crusher.
• Crushing: the jaw crusher is designed to crush the screen reject (class > 90 mm)
and reduce the phosphate blocks to a size of 0 to 300 mm, to be transported by
the conveyors.
Towards a Prediction Analysis in an Industrial Context 49

Unscreened Steril
phosphate

Steril storage
Destoning Intermediate Screening
storage

Storage of screened
phosphate
Screened
phosphate

Fig. 1 Fixed facilities process

• Conveyors: Phosphate conveyor belts, which connect fixed facilities and storage
parks.

Intermediate Storage
• The destoned phosphate is transported by the conveyors to be stored in rafters
using the storage machines (Stackers).
• Resumption of the destoned phosphate: This resumption is ensured by the shovel
wheel, which feeds the B7 conveyor to discharge into the T1 conveyor. This
conveyor transports the phosphate to the screening hopper with a maximum
throughput of 1200 t/h.

Screening
• The destoned phosphate is conveyed by the T1 conveyor and discharged into
a distribution chute to feed the five hoppers. The destoned phosphate is then
screened by five screens at the exit of the hoppers.
• The screened phosphate is conveyed by the T13 conveyor to the storage area. The
sterile material is conveyed by the T6 conveyor for a final screening with the CR7
screen before being sent to the sterile storage area.

Tarin Loading
• Homogenization: Mineral layers of different contents are extracted, a homoge-
nization is necessary to achieve a mixture of the desired quality. This is achieved
by storing the destoned and screened phosphate in alternate passes.
• Shovel wheel: It allows the phosphate to be resumed and the T14a and T14b
conveyors to be loaded (depending on availability) to convey the phosphate to the
loading hopper, via the T15 and T17 conveyors.
50 I. Battas et al.

The product is removed from the storage area by a take-back shovel wheel and
then transported by conveyors to a hopper that supplies the wagons. A train generally
consists of 60 wagons with an average capacity of 61 tonne.

4.2 Data Preprocessing

Pre-processing is necessary to generate a quality data and to improve the efficiency

of data mining models. In order to achieve this, we started by the comprehension
of the data, indeed the data are on 2 databases, one for the year 2017 and the other
for the year 2018. Each of these databases contains 2 csv files. The first file contains
the variables related to the T1 conveyor stops and the second file contains the rest
of the data such as the shift, the shift head, the control room operator, the screening
efficiency and many other additional information (see Table 1). Table 2 shows the
number of observations included in each of the two databases 2017 and 2018.
After the data understanding step, we moved on to the integration of the different
files, but when setting up the database we encountered some problems such as the
naming problem. The same data may have different names or the same name may
have different formats due to spelling mistakes, lower case instead of upper case, etc.).
Therefore we decided to unify the names of the different variables and then encode
them as integers because it is complicated for machines to understand and work
with texts rather than numbers, as the models are based on mathematical equations
and calculations. Afterwards we started the cleaning data step where we deleted all
lines containing more than 80% missing values, we deleted lines with incorrect and
insignificant values, and we handled missing values and outliers. Then we normalized
the data and finally tried to obtain a reduced representation of the dataset, smaller in
volume, but which produces the same (or almost) analytical results.
After this pre-processing the number of observations mentioned in Table 3 is no
longer valid. Initially there were around 4400 observations, and in the end there are
2700 observations (Table 3). Approximately 1700 observations have been deleted
for various reasons mentioned above.

4.3 Model Development

Databases used: During our study of the different prediction algorithms, we found
that they are highly dependent on the types of variables existing in the database. For
this reason we have subtracted from our raw database the following databases:
• Complete database: On this database we have treated all variables containing
missing, outliers, erroneous, and insignificant values.
• Numerical variable database: We have kept only the numerical variables in this
database.
Towards a Prediction Analysis in an Industrial Context 51

Table 1 Available variables

Data Heading level Signification Type
Explanatory variables 1 Date Date of a day of the year Date
2 Week Week number in the year Integer
3 Shift N° shift of phosphate of Integer
the extraction
4 Shift head Name of the head of the String
phosphate extraction shift
5 Control room operator Name of control room String
operator
6 Driver of the shovel Name of the driver the String
Shovel
7 Quantity started in tonne Quantity of intermediate Float
stock between destoning
and screening
8 Layer Nature of the phosphate String
layer
9 ORIGIN Park of destoning stock String
10 START Start of the layer Time
11 END End of the layer Time
12 THC (tonne) Screened Wet Tonne Float
13 THE (tonne) THE = THC + sterile Float
wet destoned
14 HMT1 (hour) Conveyor running hours Float
T1
15 HMT13 (hour) Conveyor running hours Float
T13
16 Rate of the steril % Quantity of the Float
sterile/Quantity of
phosphate screened
17 Efficiency T1 (t/h) THE/HMT1 Float
18 Efficiency T13 (t/h) THC/HMT13 Float
19 Number of stops Number of conveyor Float
stops T1
20 Duration of stops (min) Duration of conveyor Float
stops T1
24 TEMPERATURE Daily temperature of Ben Float
Guerir
25 PRECIPITATION Daily precipitation of Ben Float
Guerir
26 PREVIOUS LAYER Nature of the previous String
phosphate layer
27 PREVIOUS ORIGIN Previous destoning stock String
park
(continued)
52 I. Battas et al.

Table 1 (continued)
Data Heading level Signification Type
Target variable 21 Screening unit efficiency Screened Wet Float
(t/h) Tonne/number of hours

Table 2 Number of
Database Number of observations
observations included in
databases 2017 2225
2018 2236

Table 3 Number of
Number of observations
observations before and after
preprocessing Before preprocessing 4400
After preprocessing 2700

• Correlated variables database: on this database we are limited to variables that

strongly influence the performance of fixed facilities.

General parameters for the experiment: All results were obtained using the Python
language, with the sklearn library containing all models.
The models we used are Multiple Linear Regression, Support Vector Machine-
Regression, Regression Tree, k-Nearest Neighbors.
Concerning the values of the experimental parameters, we kept those defined by
default in the software except for the “random state” parameter which is used to
choose the same sample when we test the model. Indeed, if we set this parameter
equal to an integer 10 for example, the model will always take the same sample. We
will then be able to compare the models with each other because the same sample
will have been selected. However, if we do not set a random state, the model will
randomly take a different sample for each test than before.

4.4 Model Evaluation

Results. We present below the results obtained from the R2 score and the Root Mean
Square Error RMSE of the different models application on each of the databases:
complete database, database with numerical variables, and database with correlated
variables.
Towards a Prediction Analysis in an Industrial Context 53

Table 4 The maximum

Models Score R2 RMSE
percentage of error of
overfitting for the correlated Multiple Linear Regression 0.97 40.81
variables database Support Vector Machine-Regression 0.95 49.66
(SVR)
Regression Tree 0.83 91.83
K Nearest Neighboring (KNN) 0.73 115.21

Table 5 Results of the

Models Score R2 RMSE
evaluation criteria of the
models applied on the Multiple Linear Regression 0.97 40.58
numerical variables database Support Vector Machine-Regression (SVR) 0.95 49.52
Regression Tree 0.83 90.98
K Nearest Neighboring (KNN) 0.89 71.57

Complete database
We present in Table 4, the application results of the models: Multiple Linear
Regression, Support Vector Machine-Regression, Regression Tree, and K Nearest
Neighboring on the complete database.
For the complete database the best model to predict screening efficiency, using
regression algorithms is multiple linear regression. The score R2 for this model is
more than 97% and its root mean square error is the lowest which is equal to 40.81.
Numerical variable database
Table 5 bellow summarizes the application results of the prediction models on
numerical variable database.
The best model, even for this database, is the multiple linear regression with a
score R2 equal to 97% and a root mean square error equal to 40.58.
Correlated variables database
In Table 6 we show the application results of the prediction models on the correlated
variables database.
Again, we note that the best result found for the correlated variables database is
the multiple linear regression with a score R2 equal to 87% and a root mean square
error equal to 80.81.

Table 6 Results of the

Models Score R2 RMSE
evaluation criteria of the
models applied on the Multiple Linear Regression 0.87 80.81
correlated variables database Support Vector Machine-Regression (SVR) 0.82 92.49
Regression Tree 0.83 97.07
K Nearest Neighboring (KNN) 0.85 86.66
54 I. Battas et al.

Verification of the presence of overfitting: It is important to check if there is no

overfitting. Several techniques are available, such as the use of two separate files and
cross-validation. In this study, we opted for the “Use of two separate files” method, its
principle is to separate all observations into two files, the first file serves as learning
data and the second as test data. The observation file is therefore used to build the
learning model and the test file to validate the model. The percentage of division
of all observations differs from one author to another, usually between 60–70% of
observations are in the learning file and 30–40% in the test file and then it is necessary
to determine a maximum acceptable deviation that should not be exceeded between
the two files, this percentage depends on several factors. One of the important factors
in determining this percentage is the sample size. In our case we set a percentage
of 30% for test data and a percentage of 70% for learning data and then we tested
all our models with this technique and we accepted a maximum difference of 5%
between the learning file and the test file. A difference greater than this percentage
would mean that our model is too adapted. We tolerated a maximum error of 5%
between the R2 of the learning file and the R2 of the test file. So, with a difference of
less than 5%, we assumed that there was no overfitting [10].
Complete database
Table 7 shows the margin of error percentages of the different models applied on the
complete database.
Numerical variable database
In Table 8 we give the margin of error percentages of the different models applied
on the numerical variable database.

Table 7 The maximum error percentage of overfitting for the complete database
Models Score R2 of learning R2 of test Margin of error %
Multiple Linear Regression 0.96 0.97 1
Support Vector Machine-Regression 0.96 0.95 1
(SVR)
Regression Tree 0.87 0.83 4
K Nearest Neighboring (KNN) 0.77 0.73 4

Table 8 The maximum percentage of error of overfitting for the numerical variable database
Models Score R2 of learning R2 of test Margin of error (%)
Multiple Linear Regression 0.96 0.97 1
Support Vector Machine-Regression 0.96 0.95 1
(SVR)
Regression Tree 0.87 0.83 4
K Nearest Neighboring (KNN) 0.84 0.89 5
Towards a Prediction Analysis in an Industrial Context 55

Table 9 The maximum percentage of error of overfitting for the correlated variables database
Models Score R2 of learning R2 of test Margin of error %
Multiple Linear Regression 0.88 0.87 1
Support Vector Regression (SVR) 0.84 0.82 2
Regression Tree 0.87 0.83 4
K Nearest Neighboring (KNN) 0.82 0.85 3

Table 10 Best results

Type of database Best model Score R2 RMSE
obtained for each database
Complete database Multiple linear 0.97 40.81
regression
Numerical variable Multiple linear 0.97 40.58
database regression
Correlated variables Multiple linear 0.87 80.81
database regression

Correlated variables database

Table 9 illustrates the margin of error percentages of the different models applied on
the numerical variable database.
From this study we notice that we have not exceeded the maximum error of
overfitting for each model applied to each database.
Summary of the Results Achieved: Table 10 summarizes the best results obtained
for each database.
According to this table, the best model for predicting the performance of the
screening unit is the linear regression applied on the numerical variable database.

5 Conclusion and Perspectives

Through this work, we presented in detail the methodology followed for data mining
in four steps, namely: domain comprehension, data preprocessing, model develop-
ment and finally model evaluation. Subsequently, we applied this methodology to
the screening unit database in order to predict the efficiency of the fixed facilities
screening unit based on many explanatory variables for three types of databases
(complete database, numerical variables database, correlated variables database).
To obtain the best combination of model and database, we tested and compared
these models against the evaluation criteria while ensuring that there was no over-
fitting. These tests showed that the best model for predicting the efficiency of the
screening unit is the linear regression applied on the basis of data from the numerical
variables.
56 I. Battas et al.

This study illustrates the importance of regression models in predicting the effi-
ciency of the fixed facilities. However, some limitations must be taken into account.
First, other databases and case studies should be envisaged. Second, other data mining
methods such as classification algorithms and other methods for handling missing
and outliers data need to be tested with a larger database in order to extract new knowl-
edge and build models that can help managers make more accurate decisions. Also,
professionals must be further involved in this study to provide additional knowledge
and information in this domain. Finally, methodological and technical assistance to
users (novices or experts) of the data mining process must be available. Certainly a
blind application of data mining methods on the data in hand can lead to the discovery
of incomprehensible or even useless knowledge for the end user. Thus, the effective
implementation of a data mining project requires advanced knowledge and appro-
priate decisions on a number of specialized techniques (data preparation, attribute
transformation, choice of algorithms and parameters, methods for evaluating results,
etc.) [11, 12].

Acknowledgments We would like to thank our colleagues from Mohammed VI Polytechnic

University and OCP Group for giving us access to the required resources and their unconditional
support.

References

1. Benbelkacem, S., Kadri, F., Chaabane, S., Atmani, B.: A data mining-based approach to predict
strain situations in hospital emergency department systems. In: International Conference on
Modeling, Optimization and Simulation. Nancy, France, p. 8 (2014)
2. Agarwal, V.: Research on data preprocessing and categorization technique for smartphone
review analysis. Int. J. Comput. Appl. 131, 30–36 (2015). https://doi.org/10.5120/ijca20159
07309
3. Fayyad, U.M., Piatetsky-Shapiro, G., Smyth, P.: From data mining to knowledge discovery:
an overview. In: Advances in Knowledge Discovery and Data Mining, pp. 1–34. American
Association for Artificial Intelligence, Menlo Park, CA, USA (1996)
4. Scott, H.: Data Mining: Data Preprocessing (2016). https://slideplayer.com/slide/6466501/.
Accessed 10 June 2019
5. Wei, J.: Research on data preprocessing in supermarket customers data mining. In: 2nd Interna-
tional Conference on Information Engineering and Computer Science, pp. 1–4. IEEE, Wuhan,
China (2010)
6. García, S., Luengo, J., Herrera, F.: Tutorial on practical tips of the most influential data prepro-
cessing algorithms in data mining. Knowl.-Based Syst. 98, 1–29 (2016). https://doi.org/10.
1016/j.knosys.2015.12.006
7. Agard, B., Kusiak, A.: Exploration des Bases de Données Industrielles à l’Aide du Data Mining
– Perspectives. In: 9th National AIP PRIMECA Colloquium 9 (2005)
8. Nagelkerke, N.J.D.: Miscellanea A note on a general definition of the coefficient of
determination. Biometrika Trust 4 (2008)
9. Berry, M.J.A., Linoff, G.S.: Data Mining Techniques: for Marketing, Sales, and Customer
Relationship Management, 2nd edn. Wiley, Indianapolis, Ind (2004)
10. Bellavance, F.: Préparation de données pour le data mining. HEC Montréal University Course
(2017)
Towards a Prediction Analysis in an Industrial Context 57

11. Zemmouri, E.M., Behja, H., Marzak, A., Trousse, B.: Ontology-based knowledge model for
multi-view KDD process. Int. J. Mobile Comput. Multimed. Commun. 4, 21–33 (2012)
12. Zemmouri, E.L.M., Behja, H., Marzak, A.: Towards a knowledge model for multi-view KDD
process. In: 3rd International Conference on Next Generation Networks and Services (NGNS),
pp. 18–22. IEEE, Hammamet, Tunisia (2011)
Methodology for Implementation
of Industry 4.0 Technologies in Supply
Chain for SMEs

Hafsa El-kaime and Saad Lissane Elhaq

Abstract Over the past years, several initiatives have been emerged: industry 4.0 in
Germany, smart manufacturing in the United States of America, internet + in China,
future industry in France. Different nominations with an only objective, which is to
make the company smarter: transformation in the organization and management of
production as well as the entire logistics chain. Even the importance of this indus-
trial revolution, there is a lack of research in the work of implementing Industry 4.0
concepts for small and medium-sized enterprises (SMEs). In order to facilitate the
transformation, this work treats an approach through the presentation of a method-
ology for implementing Industry 4.0 technologies in supply chains. The objective
of this article is to define an overview of Industry 4.0 and its impact on the opti-
mization of the supply chain for SMEs, based on the literature review study, and
pilot guide for a 4.0 transition strategy based on the method called DMAIC. In this
study a framework of a methodology for implementing industry 4.0 is proposed and
the elaboration of model is given in order to establish the technological means of
industry 4.0 according to the objectives performance desired by the company.

Keywords Industry 4.0 · Implementation methodology · Supply chain · SMES ·

DMAIC

1 Introduction

The supply chain has existed since the start of industrialization, the overall objective
is to transform and transport materials or products, to add value, and to satisfy a
request at each stage of the process [1].

H. El-kaime (B) · S. L. Elhaq

Laboratory of Engineering Research, National Higher School of Electricity and Mechanics
(ENSEM), Hassan II University of Casablanca, Casablanca, Morocco
e-mail: [email protected]
S. L. Elhaq
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 59

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_5
60 H. El-kaime and S. L. Elhaq

The appearance of the 4th industrial revolution impacted the supply chain, it
was turned upside down; it became smarter, more transparent and more flexible at
all levels. Today we are talking about Supply Chain 4.0: it is the establishment of
Industry 4.0 technologies in supply chains.
The arrival of the 4th industrial revolution introduced the concept of the internet
and intelligence into global supply chains through connected machines and tools.
Thus, the production environment is directly linked to the distribution area. This
constant connection between the digital world and the physical system is one of the
main challenges of the industry of the future [2].
Increased digital interconnectivity between humans, machines and human-
machine systems, where hierarchies will be increasingly dismantled. Will be trans-
lated by a vertical and horizontal integration along the value chain [3]. This link has
great potential for optimizing manufacturing processes, reducing costs, deadlines,
improving productivity and quality. Therefore, companies must embrace this digital
transformation and adapt their systems to changes in order to improve their supply
chains, for greater flexibility to react to the individual needs of each client and to
remain competitive in the market [4].
Thanks to these advantages, the majority of companies will wish adhere to this
paradigm in order to have a smart and flexible factory, but they don’t know how to
carry out this transformation towards the digital 4.0 world. Also, it turns out that it
is very difficult for SMEs, which often haven’t the financial and human resources
necessary to invest in recent technologies [5]. The weight of SMEs in the country’s
economy has a particular importance. In Germany, for example 99.6% of companies
are SMEs [6]. Also in Morocco, all studies agree today, on the fact that almost 95%
of the national economic fabric is made up of SMEs [7]. Therefore, this group of
companies is an essential economic engine, it is necessary to support these medium-
sized enterprises in this transition 4.0, identify an approach to follow and define a
methodology for implementing industry 4.0 technologies in the supply chain.
In the existing literature, hundreds of varied research studies on the application
of Industry 4.0 technologies in different fields of industry [8]. However, there is a
lack of research into the implementing Industry 4.0 concepts strategy for small and
medium-sized businesses [6]. The “how” is not well detailed, we have the means
and the pillars of implementation, all the data to join this digital transformation,
but the organizational transformation methodology to help companies to position
themselves and define a strategy allowing them to prioritize the transformation of
their supply chain to 4.0 doesn’t exist [9].
Industry 4.0: a vision, a concept which according to the industrial enterprise is
now SMART. This is also the aim of the supply chain: reactivity, flexibility and
security. So what is Industry 4.0? What is supply chain 4.0? How can SMEs join
this transition 4.0 in the logistics chains to benefit from all the advantages of this 4th
industrial revolution? The answer to these research questions is based on an analysis
of an in-depth literature review, which the final objective is to provide a guide of
piloting an industry 4.0 project in the supply chain for SMEs.
The rest of the article is organized as follows: the following section is a synthesis
of the general concepts of industry 4.0, Sect. 3 is a literature review on supply chain
Methodology for Implementation of Industry 4.0 Technologies … 61

4.0 in SMEs, the operational performance objectives of industry 4.0 and the method-
ologies of transition 4.0 and Sect. 4 is a detailed description of our methodology
for implementing industry 4.0 in SMEs based on the results of the case studies in
the literature to formalize the steps in the form of an industry 4.0 guide for SMEs
in Sect. 5. We conclude with a conclusion and we provide suggestions for further
research.

2 Industry 4.0: Theoretical Background

The term industry 4.0 was proposed for the first time in 2010 by the German govern-
ment as part of the “action plan for a high-tech strategy 2020” [10]. This term was
invented to identify the complete transformation of all industrial production through
the fusion of internet technologies and industrial manufacturing processes. It is a
global transformation using digital integration and intelligent engineering, it is cited
as the next level of manufacturing, where machines will redefine themselves in the
way they communicate and perform individual functions [8]. The goal of the estab-
lishment of Industry 4.0 is to develop a new generation of intelligent factories based
on the digitalization of manufacturing processes; these factories are characterized by
production flexibility through the use of real-time configurable machines that allow
customized production [11].
The field of industry 4.0 is very young since the first article was only published
in 2012, moreover the number of scientific productions has increased exponentially
compared to the number of publications in the two indexing platforms: Web Of
Science and Scopus [8].

2.1 Definition of SME

Small and medium-sized enterprises (SMEs) are independent companies, which

employ fewer than a certain number of employees and have a specific turnover or
balance sheet. Consequently, most of the variables defining SMEs vary from country
to country. For example, SMES in Morocco are defined on the following bases:
Turnover is between 10 MDHS (1.000.000 e) and 175 MDHS (17.500.000 e), the
permanent workforce is between 10 and 250.

2.2 Evolution of the Industrial Environment

The industrial environment has evolved in a discontinuous manner over the centuries.
These discontinuities, which are presented in Fig. 1, lead to four industrial revolutions
62 H. El-kaime and S. L. Elhaq

Fig. 1 Industrial revolution

The first industrial revolution used steam to mechanize production, the second
used electrical energy to produce a large volume of identical products at low unit
cost, this period was marked by mass production and the third used electronics and
information technology to automate production. Today, a fourth industrial revolution
is underway, characterized by the ubiquitous use of sensors and actuators commu-
nicating via a network, allows a real-time connection between systems, machines,
tools, employees, products and other objects defining what is called IOT [12].

2.3 Main Industry 4.0 Technologies

The essential components of industry 4.0 included the following technologies: CPS,
additive manufacturing, virtual reality, cloud computing, big data and data science
[8]. According to the Boston consulting group, 9 technology groups of industry 4.0,
represented in Table 1, transform the production of isolated and optimized cells into
a fully integrated, automated and optimized workflow [13].
Table 1 presents a concise summary of the technological means of industry 4.0
that we have synthesized from the literature review, these pillars that we call industry
4.0 implementation means are available and accessible, sometimes at different costs
and levels of development [10]. The objective is to exploit these technology groups
in order to establish industry 4.0 in the company and provide the supply chain with
new operational performance objectives.
Methodology for Implementation of Industry 4.0 Technologies … 63

Table 1 Industry 4.0 technologies

Technology 4.0 Description
Cloud computing Def Storage and analysis huge amounts of data [13]
Obj Computing and storage capacity is shared by partitioning physical
resources using virtualization technologies
Fct Share information across multiple systems and networks in real
time, promote planning and utilization of shared resources, process
control and performance evaluation [10]
IOT Def Global network of interconnected objects based on standard
communication protocols [14]
Obj Communication in real time of physical objects makes it possible to
control the state of products or systems and to facilitate the
decentralization of decision-making in the face of a hazard [10]
Fct Advanced connectivity of products, systems and services by
providing ubiquitous information
Big data Def Heterogeneous interconnected objects generate a large amount of
data of different types
Obj Generating decisions in real time of great importance
Fct Improved decision making and process optimization
Simulation Def Modeling to reflect the physical world in a virtual world
Obj Optimization of all industrial processes [15]
Fct Analysis of product behavior, performance of production lines and
coordination of multi-site networks, thereby reducing machine
configuration time and increasing quality [11]
CPS Def Systems in which natural systems created by man (physical space)
are tightly integrated with computing, communication and control
systems (cyberspace)
Obj Management of the interconnection between physical active
systems and their computing capacity [13]
Fct Real-time data acquisition requires an appropriate architecture,
called CPS
Cyber Security Def Emergence of data analysis and sharing technologies have forced
users and manufacturers to secure their information [16]
Obj Guarantee of the data which are stored and which pass on the
network so that they are not vulnerable
Fct Communication of objects with their environment and active
participation in the reconfiguration of the system in real time
Virtual Reality Def Improvement of the real environment of a human being thanks to a
virtual environment
Obj Superposition in the user’s vision of real objects from his classic
vision with virtual objects added by computer [17]
Fct Interaction between the user, real and virtual objects [18]
(continued)
64 H. El-kaime and S. L. Elhaq

Table 1 (continued)
Technology 4.0 Description
Printing 3D Def Manufacture of objects from a 3D model through a process in which
the layers of materials are laid under precise computer control
Obj Favoring the personalized demand of customers who request
products of small quantities
Fct Manufacture of complex shape products in record time
Cobotif Def Active cooperation with human operators during all industrial
activities
Obj Increase productivity at lower cost
Fct Assistance for operators in their most difficult tasks, by
supplementing their effort in their movements [19]

3 Literature Review

The literature search was conducted using the Springer, Science direct, IEE, Google
Scholar and Wikipedia databases. The literature review included research in gener-
alities on industry 4.0, supply chain 4.0, SMEs in the age of industry 4.0, operational
performance objectives and transition 4.0 methodology.

3.1 Supply Chain 4.0

Supply chain 4.0 has come to solve various problems: on-time delivery with missed
deadlines, shortage of stocks, administrative burdens, inability to predict order book,
operating burdens, quality problems, inability to organize to produce... By using data,
digital, artificial intelligence and digital to optimize supplies. Integrated information
flows are essential in the supply chain 4.0 (Fig. 2). The interconnection of the entire
supply chain is necessary to succeed in achieving the potentials foreseen for Industry
4.0. Having a look at different research streams that concern the role of Industry 4.0
for supply chain management, we identify several publications that regard the role
of Industry 4.0 [20]. Further investigations consider the role of Industry 4.0 respec-
tively the Internet of Things in supply chain quality management, business process
management, logistics trajectory and cloud logistics. In general, the potentials and

Information flow

Procurement Production Distribution Sales

Physical flow

Fig. 2 Main functions of SCM

Methodology for Implementation of Industry 4.0 Technologies … 65

challenges of Industry 4.0 discussed in these publications show that Industry 4.0 seeks
to achieve benefits in all areas of application. The areas of the supply chain that will
be most positively affected by the introduction of Industry 4.0 are order fulfillment
53.84%, transport logistics 61.54%. With regard to the warehouse 66.6%, within
the purchasing function, Industry 4.0 shows 71.43% opportunities [21]. The most
relevant benefits from the implementation of Industry 4.0 are increased flexibility,
quality standards, efficiency and productivity. This will enable mass customization,
allowing companies to meet customers’ demands, creating value through constantly
introducing new products and services to the market. Finally, industry 4.0’s techno-
logical capabilities could be opportunities or threats, depending on the context of
implementation.

3.2 SMEs in the Age of Industry 4.0

The European Commission defines an SME as a company with less than 250
employees and a turnover of less than e50 million. The importance of SMEs in
the country’s economy is particularly important. In Morocco, all studies now agree
that almost 95% of the national economic fabric is made up of SMEs [7]. This
group of companies is therefore an essential economic driver. So it is important for
SMEs to follow industrial development and use recent technologies to improve their
industrial performance, to have significant productivity and efficiency gains for orga-
nizations. However, this technological transition remains an evolving process that
manufacturing SMEs have very unevenly integrated for the time being [22].
Table 2 summarizes the literature review on the presence of industry 4.0 means
of implementation in SMEs, only a few technology groups are presented in SMEs:
Cloud Computing and IOT are the most established and used in SMEs [10].
In the following paragraph, we present a detailed analysis of the operation of the
various technology groups in industry 4.0 in SMEs.

Table 2 Presence of industry

Technology 4.0 Presence in SMEs Obstacle
4.0 technologies in SME
IOT/RFID ✓ Absence of reliable
data [23]
Cloud computing ✓ Digital maturity [10]
Big Data ✓ Human competence
[24]
CPS, Cyber Human competence
Security [10]
Simulation, 3D ✓ Financial resource
Printing [10]
Cobotif, Virtual Financial resource
Reality [10]
66 H. El-kaime and S. L. Elhaq

IOT: Several researchers use IOT in SMEs with RFID technology to obtain informa-
tion from production flows in real time [23] They show that RFID and IOT improve
collaboration between SMEs. The majority of SMEs don’t have reliable data. So the
use of RFID to map production flows during a Lean Manufacturing implementation
process is the right solution. This system allows reliable data on production flows
to be obtained quickly and thus to target continuous improvement initiatives to be
launched as a priority.
Cloud Computing: The most commonly used means of implementation in empirical
cases, more than 65% of articles report the use of cloud computing [10], especially
with the aim of building virtual collaborative ventures between SMEs.
However, on the basis of the observations, SMEs do not all have the skills and
capabilities to meet the complex needs of customers, cloud computing promotes the
development of industrial collaboration between several partners by pooling their
skills and different businesses to meet customer demand.
Big data, CPS, Cyber Security, Cobotif… : Some authors argue that SMEs do not
consider the potential value within their data [24]. So data analysis tools, such as
Big data and data analytics, are not much used in their decision-making. The lack
of research in the field of Big data in SMEs confirms the observation of [24], they
showed the weakness of SMEs in development research, they stressed that cloud
computing is a viable solution for SMEs for a potential use of massive data analysis.
Also, few articles deal with cyber security and CPS in SMEs, they are not present
in SMEs [10], the same for cobotives and 3D printing are not exploited by SMEs,
and these means of implementation are very expensive with a long-term return on
investment.

3.3 Operational Performance Objectives of Industry 4.0

Through the introduction of industry 4.0 techniques in the supply chain, the opera-
tional performance objectives most achieved in SMEs are: flexibility and productivity
improvement [10].
According to Table 3, which we have summarized in the literature, we concluded
that the adoption of fundamental industry 4.0 technologies such as cloud computing,
IOT, Big Data in supply chains have a positive impact on several operational
performance factors especially flexibility, improved productivity and reduced lead
times.
Supply chain 4.0 is therefore a necessity to be adopted in companies, especially
SMEs, but it remains to be seen how to proceed, the following section presents some
proposals for a methodology for the implementation of industrial 4.0 technologies
in the supply chain.
Methodology for Implementation of Industry 4.0 Technologies … 67

Table 3 Operational performance objectives of Industry 4.0

Objective Description Technology 4.0
Flexibility – Synchronization of the flow and Decentralized – Cloud
decision making Computing
– Quick response to any market development
– Collaboration between partner companies
– Ability to react quickly to market fluctuations
– Business expertise within a corporate network
with tools for collaboration [23]
Cost reduction – Order products from the internet from – Cloud
anywhere on the earth Computing
– Synchronization of flows [25] allows a – Big Data
reduction in stocks – IOT
– Strong increase in business activity and
optimization of resources, thanks to the best
sharing between customer and supplier
Improved productivity – Optimization of production based on data – IOT
generated by connected objects [23] – Cloud computing
– Optimization of the productivity of all partners – Big data
(customer–supplier) using a Cloud computing
platform
– Optimization of production planning through
analysis of customer data
Quality improvement – Improvement of the quality control of the – RFID
production process through the use of RFID – Big data
technology [26]
– Improvement of product quality using historical
data from sensors
Delay reduction – Synchronization of the flows between all the – Cloud computing
partners allows a reduction of the deadlines [23] – IOT
– Realization of a cartography of the production
flows using the IOT in complementarity of
Lean Manufacturing
– Accurately measures the waiting times of each
station
– Identification of bottlenecks to optimize and
target workstations in priorities

3.4 Transition 4.0 in the Supply Chain

In the existing literature varied research on the applications of industry 4.0 tech-
nologies in different types of industry. However, little works that addresses the
methodology for implementing industry 4.0 concepts for small and medium-sized
enterprises. Among these works, we quote:
• A method that allows companies participating in a given supply chain to success-
fully transition to industry 4.0, through a business performance assessment model
that is part of a supply chain to support them in the transformation 4.0. This model
68 H. El-kaime and S. L. Elhaq

makes it possible both to assess the current level of maturity of each company and
to have a global vision of the entire supply chain of which they are part. The gap
analysis defines the transformation 4.0 strategy to be implemented [9], structured
according to the steps of the Deming wheel (Plan-Do-Check-Act)
The proposed methodology, which is qualitative, focuses on the Do and Check
steps, but the ACT step, which focuses on the implementation of actions based on
the application of industry resources 4.0, has not been detailed.
• A method based on a qualitative study that combines three areas of intervention:
Industry 4.0, customer experience and supply chain management [6]. The authors
identified eight factors (degree of customization, stakeholder integration, manu-
facturing flexibility, process transparency, automation, inventory level, customer
data analysis and customer independence) to be used to optimize supply chain
processes in accordance with Industry 4.0 standards and to increase customer satis-
faction. Based on the results of the application of these factors in a German SME,
the authors presented a five-step framework for the implementation of industry
4.0 in the SME supply chain: 1/Establish a team of experts, 2/Prepare the data
collection process, 3/Define the optimal realization of industry 4.0, 4/Analyze the
real status of the industry 4.0 realization, 5/Develop an action plan.
The definition of the methodology is based on a case study based on the realization
of Industry 4.0 factors in an SME, these presented factors do not define an optimal
implementation applicable in Industry 4.0 in general.
This study must be validated by other companies and industries to ensure general
practical applicability. The framework only presents a set of areas that help to discover
the potential for supply chain improvement. This leads to a very individual definition
of the optimal use case for Industry 4.0 and also requires specific know-how to
develop a real implementation plan.

4 Methodology of Industry 4.0 Implementation in Supply

Chain for SMEs

4.1 Principe

The methodology that we propose is based on a synthesis of literature review of the

industry 4.0 implementation in SMEs, structured according to the DMAIC approach,
in order to suggest an implementing model for the industry 4.0 technologies in SMEs.
The literature review shows that industry 4.0 has an impact on supply chain
management in the context of SMEs, via the pillars of industry 4.0: Cloud computing,
Big data, IOT…In opposite side, there is a lack of research on how to proceed:
How SMEs can join this transition 4.0 in their supply chains to benefit from all the
Methodology for Implementation of Industry 4.0 Technologies … 69

advantages of this 4th revolution? How SMEs with a limited means and financial
resources can establish the technological tools to have a smart supply chain?

4.2 Research Method: DMAIC

Our methodology is structured according to the DMAIC approach, and consists of

five steps: Define, Measure, Analyze, Innovate and control.
• Define: Determination of the general framework of the project.
• Measure: Development of the current VSM.
• Analyze: Analysis of the results obtained.
• Innovate: Implementation of improvement actions.
• Check: Measure the relevance of improvement actions.

4.3 Digital Maturity

To propose an improvement in any field, it is first necessary to have a very good

understanding of the envisaged place, the same thing for the transition 4.0, it is
necessary to study the degree of digital maturity of the company to offer support
adapted to the situation. The digital maturity of industrial organization is made up
of several elements that make it possible to identify areas for improvement for the
company.
We have synthesized the definition of the maturity level of the industrial company
in the form of a diagram containing five components (Fig. 3).
This study of the level of maturity makes it possible to choose the technological
solutions of industry 4.0 and the necessary equipment to acquire, according to the
intellectual level of human capital, in order to guarantee that all the company’s actors
adhere to the changes.
We have identified in the literature two models in the Table 4 for assessing a
company’s performance in relation to Industry 4.0.
Based on the maturity models developed in the literature, we propose three levels
of digital maturity, each level representing a category of SMEs:
• level = 1, for SMEs who are experts, they have established transformation 4.0
• level = 2, for SMEs that are intermediate, they have started the transformation
4.0
• level = 3, for SMEs who are beginners, they question the general concepts of
industry 4.0.
70 H. El-kaime and S. L. Elhaq

• Size, number of these locations, seniority, type of management, equipment....

Company
identity card

Human
• Average age, versatility of employees, skills assessment....
Capital
Identity
Sheet

Financial • Cost price, sales price, percentage of average turnover...

analysis of
the company

• Planning tool, flow type, flow mapping by VSM, identification of bottlenecks,

Industrial deployment of Lean Manufacturing tools, finite or infinite capacity planning...
organization

infrastructur
• Systems used (ERP, MES), automated machines, robots...
e used

Fig. 3 Five components to measure digital maturity

Table 4 Maturity models 4.0

Model Maturity level Type of company References
4.0 Readiness Model outsider, beginner Newcomers [27]
intermediate, experienced learners
expert, top performer Leaders
Maturity Index for SMEs standard, Big data, smart data, SME [28]
dark factory, industrial
ecosystem

4.4 Supply Chain 4.0 Implementation Framework

The proposed supply chain 4.0 framework is based on the deployment of industry 4.0
tools in supply chain departments, structured by the DMAIC approach, to achieve
operational performance objectives (Fig. 4).

4.5 Methodology for Deploying a Transformation 4.0

Strategy

We propose a structured methodology according to the DMAIC steps (Table 5).

Methodology for Implementation of Industry 4.0 Technologies … 71

Core functions of Operational performance

Supply chain objectives
Technology 4.0

-Procurement -Flexibility
IOT, Big Data, Cloud
-Production Computing, CPS, CS, -Cost Reduction
Simulation RV, 3D Printing,
-Distribution -Time Reduction
Cobotif
-Sales
- Productivity Improvement

DMAIC
- Quality improvement

Fig. 4 Supply chain 4.0 implementation framework

Table 5 Transformation 4.0 DMAIC

Step Objective Tools
Define – Working group QQOQCP
– Performance objective SIPOC
– Supply chain area Diagram GANT
– Process of the area
– Specifications
– Project planning
Measure – List all ONVA in the area Brainstorming
– List internal and external requirements VSM
– Measure the digital maturity level PARETO
Analyze – Analyze the degree of maturity Ishikawa
– Analyze the VSM Brainstorming
Innovate – Define the action plan Big data
– Introduce industry 4.0 technologies to solve each problem Cloud computing
– Trace the future VSM IOT
CPS…
Check – Monitor improvements in real time Standards of control
– Compare the state before and after the introduction of
industry 4.0
– Capitalize the solutions in the form of a guide for managing
an industrial site 4.0

Define: The first step of the DMAIC approach focuses on defining the objectives
and limits of the project. During this phase, it is strongly recommended to start by
establishing precise specifications of needs and describing the company’s situation.
This includes defining at first with precision:
• Define the project group: It is important to work with a multidisciplinary group
composed of a project manager, method support, production leader, scheduling
and operational agent… The project group must be supervised by a steering
72 H. El-kaime and S. L. Elhaq

committee composed of different managers: Production, logistics, sales and site

manager.
• Define the provisional planning of the project in order to organize the steps of
transformation 4.0; the GANT diagram can be used as a tool to plan the tasks.
• Form the General Concept of Industry 4.0 Project Group, including the basic tools
of Industry 4.0 to facilitate the involvement of all people in the project.
• Define the SIPOC diagram (Suppliers, Input, Process, Output, Customer) of the
supply chain; SIPOC is a technique to model a process. This tool consists of a
mapping of the process that we want to improve. It present the entire flow from
suppliers to customers.

Measure: In a second step, it is necessary to map the process to establish a VSM of

the flow. Then it is necessary to study the degree of digital maturity of the company to
propose a support adapted to the situation. This step consists of gathering information,
collecting data, identifying malfunctions and identifying sources of waste. The work
consists of:
• List the waste terms of each area of the supply chain
• List the problems experienced in each area through brainstorming
• Perform a VSM mapping to target the 4.0 transition
• Measure the maturity level of the area in order to evaluate its digital maturity
position.

Analyze: After a global study through the two phases “Define and Measure” the
“Analyze” phase makes it possible to identify the extent of the problems, the level
of maturity of the area in relation to industry 4.0, the analysis of the causes and then
to formulate an action plan. Analyze the most achieved operational performance
objectives of industry 4.0: flexibility and productivity improvement... in the supply
chain.
Innovate: Following the Analysis phase, and through brainstorming and a detailed
understanding of the maturity level of the study area, an action plan can be developed
to take full advantage of the benefits of industry 4.0, where the pillars of industry 4.0
will be introduced into the supply chain to make it smart.
The integration of industry 4.0 techniques into supply chain management is very
effective. In the innovation phase, two categories of means are considered: hardware
and digital, for the first type, the fundamental action is that companies must adopt
these technologies; in fact, no industry 4.0 without investment and adoption of new
intelligent machines is necessary for the implementation of industry 4.0 in an SME
at the beginning. The implementation of all pillars of industry 4.0 requires significant
investment for an industry, such as 3D printing, cobotives and virtual reality.
According to the literature review, SMEs do not have the financial resources and
human skills to carry out all nine pillars of industry 4.0. However, we will only focus
on the means that are very common in the literature and that mark the 4.0 transition
in SMEs such as: 3D printing, cobotives, Big data, Cloud computing, IOT.
Methodology for Implementation of Industry 4.0 Technologies … 73

Control: This step consists of observing the improvement actions implemented and
following up, in order to show the gain achieved by the transition 4.0. And to get
feedback, we have proposed a guide to managing an industrial 4.0 site in the form
of a flowchart, with the objective of making it a transition 4.0 model for SMEs.

5 Guide to Managing an Industrial 4.0 Site in SME

Based on the bibliographic research, a five-step approach structured according to the

DMAIC method is advised: As a first step, It is strongly recommended to start by
establishing precise specifications of requirements and to describing the situation of
the company or the factory, this includes properly framing the 4.0 transition project. In
a second step, it is necessary to draw up a VSM mapping of the organization, then it is
necessary to study the degree of digital maturity of the company to propose a support
adapted to the situation. This step consists gathering information and collecting data.
In the third step, it is relevant to analyze the most achieved operational performance
objectives: flexibility, productivity improvement, cost and delay reduction in the
supply chain in order to assess the real status of the industry 4.0 achievement and
internal weaknesses. Finally, in the fourth step, it is recommended to develop an
action plan to take full advantage of the benefits of Industry 4.0 in the company
(Fig. 5).

6 Conclusion and Notes for Future Work

This work aimed to present a methodology for supply chain transformation 4.0 in
SMEs, with an emphasis on industry 4.0 technologies and the most cited operational
performance objectives, through a detailed review of the literature review The main
conclusions of this study are: firstly, definition of digital maturity levels for SMEs,
secondly development of a methodology for deploying Industry 4.0 tools in SMEs,
finally capitalization of information and feedback in the form of a flowchart of the use
of industry 4.0 resources in the supply chain according to the performance objectives
targeted by SMEs.
This article proposes a five-step approach structured according to the DMAIC
approach method. First, establish a precise specification of needs and a descrip-
tion of the current situation of the company. Second, measure the level of digital
maturity of the business and map organization cartography VSM. Third, analyze the
most achieved operational performance objectives: flexibility, improved productivity,
reduction of costs and delays in the supply chain in order to assess the real status
of the achievement of Industry 4.0. Finally, it is recommended to develop an action
plan for the use of the nine pillars of Industry 4.0, in order to take full advantage of
the benefits of Industry 4.0 in the enterprise.
74 H. El-kaime and S. L. Elhaq

Yes

Cloud computing IOT RFID BD 3D Printing RV Cobotif

Fig. 5 Industry 4.0 deployment in the supply chain for SMEs

Methodology for Implementation of Industry 4.0 Technologies … 75

Future research is recommended in several directions: The methodology for

achieving Industry 4.0 and its objectives must be validated by an SME, through
a case study, to guarantee practical applicability of the proposed solution. In addi-
tion the costs are not taken into account, a detailed cost-benefit analysis is strongly
recommended before implementing Industry 4.0 digital technologies. Also this work
does not treat the risks associated to the new concepts adaptation related to Industry
4.0, which could be added as research in future work.

References

1. Janvier-James, A.M.: A new introduction to supply chains and supply chain management:
definitions and theories perspective. IBR 5, p194 (2011). https://doi.org/10.5539/ibr.v5n1p194
2. Lee, J., Bagheri, B., Kao, H.A.: A cyber-physical systems architecture for Industry 4.0-based
manufacturing systems. Manuf. Lett. 3, 18–23 (2015). https://doi.org/10.1016/j.mfglet.2014.
12.001
3. Kagermann, H.: Change through digitization—value creation in the age of Industry 4.0. In:
Albach, H., Meffert, H., Pinkwart, A., Reichwald, R. (eds.) Management of Permanent Change,
pp. 23–45. Springer Fachmedien Wiesbaden, Wiesbaden (2015)
4. Colotla, I., Fæste, A., Heidmann, A., Winther, A., Høngaard Andersen, P., Duvold, T.: Winning
the industry 4.0 RACE.pdf (2016)
5. Tangible Industry 4.0: une approche basée sur des scénarios pour l’apprentissage de l’avenir de
la production - ScienceDirect. https://www.sciencedirect.com/science/article/pii/S22128271
16301500. Accessed 9 Dec 2019
6. Bär, K., Herbert-Hansen, Z.N.L., Khalid, W.: Considering Industry 4.0 aspects in the supply
chain for an SME. Prod Eng Res Devel 12, 747–758 (2018). https://doi.org/10.1007/s11740-
018-0851-y
7. Landaburu E La PME, moteur de l‘économie marocaine. 10
8. Muhuri, P.K., Shukla, A.K., Abraham, A.: Industry 4.0: A bibliometric analysis and detailed
overview. Eng. Appl. Artif. Intell. 78, 218–235 (2019). https://doi.org/10.1016/j.engappai.
2018.11.007
9. Ioana, D., François, M., Jean-Louis, M. : Méthodologie et outil de définition de la stratégie de
transition 4.0 pour la chaine logistique (2019). https://ideas.repec.org/p/ulp/sbbeta/2019-14.
html. https://ideas.repec.org/p/ulp/sbbeta/2019-14.html. Accessed 16 Dec 2019
10. Moeuf, A.: Identification des risques, opportunités et facteurs critiques de succès de l’industrie
4.0 pour la performance industrielle des PME (2018)
11. Vaidya, S., Ambad, P., Bhosle, S.: Industry 4.0—A Glimpse. Procedia Manuf. 20, 233–238
(2018). https://doi.org/10.1016/j.promfg.2018.02.034
12. Stankovic, J.A.: Research directions for the Internet of Things. IEEE Internet Things J. 1, 3–9
(2014). https://doi.org/10.1109/JIOT.2014.2312291
13. Bortolini, M., Ferrari, E., Gamberi, M., Pilati, F., Faccio, M.: Assembly system design in the
Industry 4.0 era: a general framework. IFAC-PapersOnLine 50, 5700–5705 (2017). https://doi.
org/10.1016/j.ifacol.2017.08.1121
14. Atzori, L., Iera, A., Morabito, G.: The Internet of Things: a survey. Comput. Netw. 54, 2787–
2805 (2010). https://doi.org/10.1016/j.comnet.2010.05.010
15. Azevedo, A., Almeida, A.: Factory templates for digital factories framework. Robot. Comput.-
Integr. Manuf. 27, 755–771 (2011). https://doi.org/10.1016/j.rcim.2011.02.004
16. Holtewert, P., Wutzke, R., Seidelmann, J., Bauernhansl, T.: Virtual fort knox federative, secure
and cloud-based platform for manufacturing. Procedia CIRP 7, 527–532 (2013). https://doi.
org/10.1016/j.procir.2013.06.027
76 H. El-kaime and S. L. Elhaq

17. Lee, J., Han, S., Yang, J.: Construction of a computer-simulated mixed reality environment for
virtual factory layout planning. Comput. Ind. 62, 86–98 (2011). https://doi.org/10.1016/j.com
pind.2010.07.001
18. Azuma, R., Baillot, Y., Behringer, R., Feiner, S., Julier, S., MacIntyre, B.: Recent Advances in
augmented reality. IEEE Comput. Graph. Appl. 14 (2001)
19. Khalid, A., Kirisci, P., Ghrairi, Z., Thoben, K.D., Pannek, J.: A methodology to develop collab-
orative robotic cyber physical systems for production environments. Logist. Res. 9, 23 (2016).
https://doi.org/10.1007/s12159-016-0151-x
20. Müller, J.M., Voigt, K.I.: The impact of Industry 4.0 on supply chains in engineer-to-order
industries—an exploratory case study. IFAC-PapersOnLine 51, 122–127 (2018). https://doi.
org/10.1016/j.ifacol.2018.08.245
21. Tjahjono, B., Esplugues, C., Ares, E., Pelaez, G.: What does Industry 4.0 mean to supply chain?
Procedia Manufact. 13, 1175–1182 (2017). https://doi.org/10.1016/j.promfg.2017.09.191
22. Hennebert, M.A., Cayrat, C., Morissette, L.: Transformation des PME manufacturières : entre
promesses et réalités de l’industrie 4.0. Gestion 44, 86–89 (2019)
23. Ren, L., Zhang, L., Tao, F., Zhao, C., Chai, X., Zhao, X.: Cloud manufacturing: from concept to
practice. Enterp. Inf. Syst. 9, 186–209 (2015). https://doi.org/10.1080/17517575.2013.839055
24. Bi, Z., Cochran, D.: Big data analytics with applications. J. Manag. Anal. 1, 249–265 (2014).
https://doi.org/10.1080/23270012.2014.992985
25. Chalal, M., Boucher, X., Marques, G.: Decision support system for servitization of industrial
SMEs: a modelling and simulation approach. J. Decis. Syst. 24, 355–382 (2015). https://doi.
org/10.1080/12460125.2015.1074836
26. Segura Velandia, D.M., Kaur, N., Whittow, W.G., Conway, P.P., West, A.A.: Towards indus-
trial internet of things: crankshaft monitoring, traceability and tracking using RFID. Robot.
Comput.-Integr. Manuf.s 41, 66–77 (2016). https://doi.org/10.1016/j.rcim.2016.02.004
27. Akdil, K.Y., Ustundag, A., Cevikcan, E.: Maturity and readiness model for Industry 4.0 strategy.
In: Industry 4.0: Managing the Digital Transformation, pp. 61–94. Springer International
Publishing, Cham (2018)
28. Häberer, S., Fabian, B., Katrin Lau, L.: Development of an Industrie 4.0 Maturity Index for
Small and Medium-Sized Enterprises (2017). https://www.researchgate.net/publication/320
415942.
A Deep Reinforcement Learning (DRL)
Decision Model for Heating Process
Parameters Identification in Automotive
Glass Manufacturing

Choumicha El Mazgualdi, Tawfik Masrour, Ibtissam El Hassani,

and Abdelmoula Khdoudi

Abstract This research investigates the applicability of Deep Reinforcement

Learning (DRL) to control the heating process parameters of tempered glass in indus-
trial electric furnace. In most cases, these heating process parameters, also called
recipe, are given by a trial and error procedure according to the expert process expe-
rience. In order to optimize the time and the cost associated to this recipe choice, we
developed an offline decision system which consists of a deep reinforcement learning
framework, using Deep Q-Network (DQN) algorithm, and a self-prediction artificial
neural network model. This decision system is used to define the main heating param-
eters (the glass transfer speed and the zone temperature) based on the desired outlet
temperature of the glass, and it has the capacity to improve its performance without
further human assistance. The results show that our DQN algorithm converges to
the optimal policy, and our decision system provides good recipe for the heating
process with deviation not exceeding process limits. To our knowledge, it is the first
demonstrated usage of deep reinforcement learning for heating process of tempered
glass specifically and tempering process in general. This work also provides the basis
for dealing with the problem of energy consumption during the tempering process
in electric furnace.

Keywords Deep Reinforcement Learning · Q-Learning · Deep Q-Network ·

Decision system · Process parameters identification · Heating process · Tempered
glass

C. E. Mazgualdi (B) · T. Masrour · I. E. Hassani · A. Khdoudi

Artificial Intelligence for Engineering Science Team, Laboratory of Mathematical Modeling,
Simulation and Smart Systems (L2M3S), ENSAM, Moulay Ismail University, Meknes, Morocco
e-mail: [email protected]
T. Masrour
e-mail: [email protected]
I. E. Hassani
e-mail: [email protected]
A. Khdoudi
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 77

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_6
78 C. E. Mazgualdi et al.

1 Introduction

Tempered glass is a type of safety glass manufactured by a chemical or thermal

treatment that consist of extreme heating and rapid cooling. In thermal tempering
process, the glass is heated to a certain temperature in tempering furnace and is then
cooled rapidly. This is one of the most complex manufacturing process that toughens
the glass both physically and thermally [1, 2].
In thermal tempering glass manufacturing process, especially automotive glass
production process, temperature control of tempering furnace is the main control
parameter that has an important influence on the quality of the products. Another
parameter that can be took into account is the glass speed through the different zones
of the furnace. Thus, a good choice of those two parameters at the launch of a new
project allows the process engineers to avoid many trials and errors to fix the adequate
recipe. In general, this parameters choice, also called recipe, is given according to
the expert process experience but, in most cases, it must be adjusted by the process
engineers which is time and resources consuming.
The aim of the present work is to develop a decision system that consists of
solving the problem of process parameters identification using Deep Q-Learning
(DQL) algorithm. Using ANSYS software, a series of simulation experiments are
conducted and background data are extracted. Then, a self-prediction model and a
decision agent are trained successively based on these data. The developed DQL
algorithm is then utilized for training the decision model to be able to decide, based
on the glass characteristics and the desired temperature of the glass at the output of
the furnace, which heating process parameters to choose in order to meet production
requirements.
The paper is structured as follows: in the next Sect. 2 the background on RL
needed to understand the proposed decision system schema is introduced, including
basic notions from RL and DRL field. Section 3 provides a detailed description
of the heating process control formulation into the MDP and the adoption of the
DQN to solve this decision-making problem. In Sect. 4, we present and discuss the
experiments results. Finally, Sect. 5 offers some final remarks and suggestions for
further work.

2 Background on Reinforcement Learning

2.1 Reinforcement Learning

Reinforcement learning (RL) is a subclass of machine learning which differs from

classical supervised and unsupervised learning by using a trial-and-error paradigm
to learn a strategy. It employs an agent that iteratively interacts with the environment
and modifies its actions according to a policy π. The environment responds to every
A Deep Reinforcement Learning (DRL) Decision Model … 79

action by changing its state and provides a reward as a response. The agent tries to
maximize the reward received from the environment by via a learned value function.
In order to formalize a reinforcement learning problem, the Markov Decision
Process (MDP) can be considered as an idealized form that provides precise theo-
retical statements. Hence, a reinforcement learning problem can be completely
described by a tuple of the Markov decision Process framework <S, A, P, R>, where
S denotes the state-space which consists of all the possible states the agent can be
in. A denotes the action-space which contains all the possible actions to take in the
different states. P is the transition function or probability, it contains all the proba-
bility that, given a specific action in a certain state, the process moves to a particular
state [3].
Finally, the reward function R represents the value of transitioning from one state
to another.
MDP problems can be solved in different ways, one of the most traditional methods
are Dynamic Programming (DP), which is a collection of algorithms such as policy
iteration, value iteration, and linear programming [4]. These methods are known by
their assumption of a perfect model of the environment and their huge computational
time. Hence, if the process is not finite or the transition probabilities are not available,
DP becomes more difficult to apply. One famous alternative to DP is model-free
algorithms which approximate the value function based on the experience and then,
make decisions by accessing information that has been stored in a policy or a value
function [5].
Q-Learning represents one famous model-free algorithm that systematically
updates action value function q(s, a) without the need of a model of the environ-
ment. The basic idea of Q-learning is to learn the state-action pair value, which
represents the long-term expected utility Q(s, a) for each pair of state and action.
The learned Q-value function, using this method, directly approximates the optimal
action-value function q∗ .
The Q-learning is much more used in cases where action-state spaces are discrete
and finite. This is due to the fact that the q-values are represented in a lookup table.
However, for most applications, especially industrial ones, the state space is very
large or even continuous and thus requires the q-function to be represented with a
function approximation. Artificial Neural Networks (ANN) are powerful function
approximators widely used especially for nonlinear functions.

2.2 Deep Reinforcement Learning

Deep Reinforcement Learning (DRL) is an approach that efficiently combines the

decision-making ability of the conventional reinforcement learning with the percep-
tion of deep learning [6]. It uses deep neural networks as function approximation
of the action-value function q(s, a) and it is called Deep Q-Network (DQN) [7].
The input layer of the Deep Neural Networks (DNN) is the state of the RL algo-
rithm and the output layer constitutes the predicted q-values of the state-action pairs.
80 C. E. Mazgualdi et al.

The parameters of the DNN are updated iteratively, by minimizing the loss func-
tion (Eq. 1), until the difference between the target value and the predicted value
converges.
   2 
L i (θi ) = E r + γ max Q s  , a  , θi−1 − Q(s, a, θi ) (1)

3 DQN Problem Formulation for the Heating Process

The proposed system is a DQN agent that optimizes the process parameters for the
heating process of tempered glass.
In this work, we developed an offline decision model in combination with a
quality self-prediction model [8]. A simulation software analysis was used to collect
background data, which consist of the quality index and its corresponding process
parameters. Then, a self-prediction model is trained by these data to predict the
quality index based on the process condition. This model is used as our simulated
industrial furnace and provides a training environment for the decision model by
applying a DQN algorithm.

3.1 Background Data Acquisition

ANSYS is a simulation platform that provides high-fidelity virtual prototypes to

simulate the behavior of complete products in their actual working environments.
ANSYS is used in our work to simulate the behavior of our industrial furnace
under its working condition. Then, we used this software to acquire a certain amount
of background data that consists of the outlet temperature of the glass under the
selected process parameters.

3.2 Self-prediction Model Development

Using the background data collected from our simulation software, we established
different prediction mapping models to predict the outlet temperature of glass based
on the process condition as inputs to the models.
In our study, we choose Random Forest algorithm because it shows a good perfor-
mance compared to the other algorithms, and it is performing a better mapping of the
complex relationship between the process parameters and quality indexes. Random
Forest is a powerful ensemble learning method based on decision tree algorithm, and
it is recommended for its ability to deal with small and large sample sizes [9].
A Deep Reinforcement Learning (DRL) Decision Model … 81

Fig. 1 Industrial electric furnace illustration

Our study case for the experiments consist of heating similar pieces of glass inside
an industrial electric furnace composed of 9 zones (Fig. 1). Each zone has a specific
temperature, and for all the 9 zones, one speed is defined. For every type of glass a
recipe is defined according to the experience of the expert process, then some trials
are mandatory to achieve the desired temperature of the glass at the output of the
furnace. This recipe consist of defining the temperature at each zone plus the speed
with which the glass will pass through the 9 zones. The outlet temperature is the key
dimension for production requirements, and it is specific to the type and dimensions
of the glass in production.

3.3 DQN Decision Model

We used our trained self-prediction model as a simulated environment of the real

industrial furnace. This helped us to avoid extra cost caused by the huge number of
trials we will need to acquire background data. Then, our DQN agent will directly
interact with the simulated environment, instead of the real one, to learn the optimal
strategy for process parameters identification. Figure 2 illustrates the framework for
the proposed decision model to address our problem.

Fig. 2 Proposed structure of DQN decision model

82 C. E. Mazgualdi et al.

State space S
The state representation is a combination of the current temperature values of the 9
zones, the speed of the glass and the outlet temperature of the glass. It is described
as follows:

st = (T1 , T2 , . . . , T9 , v, Texit ) ∈ S (2)

where Ti is the current temperature value of the ith zone of the furnace, v is the
current speed of the glass, and Texit is the temperature of the glass at the output of the
furnace. When the agent takes an action under the current state st , it receives a reward
from the environment, and a new state st+1 can be observed under the new process
parameter conditions. By repeating the procedure and interacting with the simulated
furnace environment, an optimal recipe that meets the quality requirements for the
outlet temperature of the glass can be acquired.
Action space A
In our study, we consider an environment in which the temperature of each zone Ti ,
and the glass speed across the furnace can be adjusted via a discrete set of values
consisting of 20 discrete actions to either increase or decrease glass speed or zone
temperature, denoted as follows:

possible actions = (a1 , a2 , . . . , a20 ) (3)

where ai ∈ {a1 , . . . , a9 } refers to the action increase the temperature of the ith zone
by 10 °C, and a j ∈ {a11 , . . . , a19 } refers to the action decrease the temperature of the
(j – 10)th zone by 10 °C. Finally, actions a10 and a20 refer, respectively, to increase
and decrease glass speed by 10 mm. In order to avoid unnecessary trials, we defined
low and high boundaries for either the zone temperature and glass speed. The zone
temperature are between 500 and 700 °C, and the glass speed is between 150 and
200 mm.
For example, if in time step t, the DQN agent takes the action a2 , which means
that the temperature of zone 2 will be increased by 10 °C, then, at can be denoted as
follows:

at = (0, a2 , 0, . . . , 0) (4)

Thus, the agent will find itself in a new state denoted:

st+1 = (T1 , T2 + 10, . . . , T9 , v, N ewT exit ) (5)

where N ewT exit is the predicted value of the outlet temperature of glass, based on the
new set of process parameters generated by the agent action, using our self-prediction
model.
A Deep Reinforcement Learning (DRL) Decision Model … 83

Reward function R
When the DQN agent takes an action, a new set of process parameters is gener-
ated. Then, the temperature of glass at the output of the furnace is predicted using
our self-prediction model, as a consequence, the agent receives a numerical reward
that quantifies the goodness or badness of its action. For our purposes, the agent
should obtain an optimal process parameter setting that allows it to reach the target
temperature at the output of the furnace. Thus, the reward function is given by:
   
 Pr eviousTexit − Ttarget  − Curr ent Texit − Ttarget 
rt = (6)
Ttarget

In our case study,

 we fixed a tolerance interval
 for the outlet temperature of the
glass, which is Ttarget − 5 ◦ C, Ttarget + 5 ◦ C . Then, at the end of every episode, an
additional reward of + 1 is added to the reward function if the outlet temperature is
within the tolerance interval.

4 Experiments Validation and Results

In this section, the experiment settings are explained and simulations are performed
to acquire background data using ANSYS software for training both self-prediction
model and DQN agent. The effectiveness and robustness of the proposed decision
system are evaluated by validation experiments on an industrial case study.

4.1 Offline Simulation and Prediction Model Training

The experimental case utilized in this study was presented in Sect. 3 (3.2), as a remind
it consist of heating similar pieces of glass (sidelites of cars) inside an industrial
electric furnace composed of 9 zones. Each zone has a specific temperature, and
for all the 9 zones, one speed is defined. Thus, the process parameters are the 9
temperatures plus the speed of the glass and the quality index is the temperature of
the glass at the output of the furnace. This temperature is associated to the glass model
to produce and it has a tolerance interval of ±5 °C. Our purpose is to automate the
choice of the process parameters to meet the production requirement for a specific
type of glass, which is the temperature of the glass at the output of the furnace. The
first step consists of collecting background data to train our self-prediction model.
This self-prediction model will constitute our simulated environment.
We, first, create a simulated furnace adopting the necessary characteristics of the
real furnace and importing the needed functionalities such as physical and thermal
laws that describe the real behavior of the electric furnace. After validating our
84 C. E. Mazgualdi et al.

Fig. 3 Plot of the predicted values versus real values (Target values)

simulated furnace, hundred simulations were performed, and then the background
data were extracted to feed our self-prediction model.
The developed self-prediction model which is a Random Forest algorithm using
120 decision trees, is then trained using the background data. It is expected to approx-
imate the relationship between the process parameters, which are the 9 temperatures
plus the glass speed, and the outlet temperature of the glass.
Figure 3 shows the plot of the predicted values versus the target values over 60
observations. We can conclude that our model perform a good fit of the real data
with a mean absolute error MAE = 0.807.

4.2 Decision System

The DQN agent is trained by interactions with the self-prediction model as a simu-
lated environment. We evaluate our DQN algorithm using the implementation in the
OpenAI Baselines project [10]. The benefit of interfacing with OpenAI Gym [11] is
that it is an actively developed interface which provides various environments and
features useful for training, it offers a reference implementation of RL algorithms.
The DQN algorithm is run with default hyper-parameters except for the number of
simulations steps and the learning rate, which we set respectively to 1000 steps and
0.00001. The controller policy is the default two hidden layer, 64 nodes multi-layer
perceptron (MLP) policy. Training and evaluations were run on Windows 10 with an
eight-core i7-7700 CPU and an NVIDIA GeForce GT 1060 graphics card.
A Deep Reinforcement Learning (DRL) Decision Model … 85

Fig. 4 Cumulative reward

performance evaluation of
DQN algorithm

The convergence of the reward function during the training process is shown in
Fig. 4.
The cumulative reward curve represents the total rewards in an episode obtained
by the agent. The curve is smoothed by the moving average. It can be observed that
the rewards value oscillates considerably in most time. This is because of the ε-greedy
policy adopted for the Q-learning algorithm, which means that the agent is unfamiliar
with the environment, and thus it need to explore the entire state space in order to
learn proper strategies. However, with more iterations, the reward kept increasing
and eventually, after 450,000 time steps it converge to a policy good enough for
identifying the good process parameters to make in use for a given product.
The proposed decision system is tested for different configurations of process
parameters proposed by the process engineers to evaluate the model robustness.
Figure 6 provides the results proposed by our decision model for a given product
in which the temperature of the glass at the output of the furnace was 685 °C. The
proposed set of process parameters allows, after 44 iterations, an outlet temperature
of 682 °C, which respect the tolerance imposed by the client.

Fig. 6 DQN Testing results for a given glass model starting from a randomized state
86 C. E. Mazgualdi et al.

5 Conclusion and Future Work

In this paper, RL with the DQN agent was successfully applied to process parameters
identification for the heating process of tempered glass. The proposed decision system
consist of a reinforcement learning based model integrated with a Random Forest self-
prediction algorithm as a simulated environment. The system automatically identify
the process parameters to make in use for a given product based on production
and quality requirements. Validation experiments were performed and demonstrates
that, based on the background data provided by ANSYS simulations, the trained
self-prediction model accurately predicts the temperature of the glass at the output
of the furnace with an MAE of 0.807. Then, the DQN agent is trained using the
prediction model as the simulated furnace environment to learn a good strategy for
process parameters identification without human intervention or any prior expert
knowledge. The system presents a rapid and stable convergence and it is able to
replace the traditional method for process parameters identification based on physical
trials which is costly and time consuming. In future work, we would like to further
investigate other toolkits in DRL regime such as policy gradient based RL especially
Proximal Policy Optimization (PPO) [12] and Trust Region Policy Optimization
(TRPO) [13]. The reward function could also be improved, one option to do this
is by introducing the time so that the agent is forced to obtain the optimal process
parameter settings within the shortest time step. Finally, further researches will focus
on integrating the optimization of the energy consumption without compromising
the product quality.

References

1. Ford, M.: How is tempered glass made? In: Scientific American (2001)
2. Jonathan Barr: Understanding the Glass Tempering Process, Germany (2003)
3. Otterlo, M., Wiering, M.: Reinforcement learning and Markov decision processes. In: Wiering,
M., Otterlo, M. (eds.) Reinforcement Learning: State-Of-The-Art, pp. 3–42. Springer, Berlin
(2012)
4. Bellman, R. E.: Dynamic programming. Ser. A Rand Corporation research study. Princeton,
N.J. Princeton University Press (1957)
5. Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. MIT Press (2018)
6. Nielsen, M. A.: Neural Networks and Deep Learning (2015)
7. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves,
A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., et al.: Human-level control through deep
reinforcement learning. Nature 518(7540), 529–533 (2015)
8. Guo, F., Zhou, X., Liu, J., Zhang, Y., Li, D. and Zhou, H.: A reinforcement learning decision
model for online process parameters optimization from offline data in injection molding. Appl.
Soft Comput. 85 (2019)
9. Ali, J.: Random forests and decision trees. Int. J. Comput. Sci. Issues (IJCSI) 9, 272–272 (2012)
10. Prafulla, D., Christopher, H., Oleg, K., Alex N., Matthias P., Alec R., et al.: OpenAI Baselines.
GitHub (2017). From https://github.com/openai/baselines
11. Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J. and Zaremba,
W.: Openai gym (2016). arXiv preprint arXiv:1606.01540
A Deep Reinforcement Learning (DRL) Decision Model … 87

12. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization
algorithms (2017). arXiv preprint arXiv:1707.06347
13. Schulman, J., Levine, S., Abbeel, P., Jordan, M. and Moritz, P.: Trust region policy optimization.
In: International conference on machine learning, pp. 1889–1897 (2015)
Analytic Hierarchy Process (AHP)
for Supply Chain 4.0 Risks Management

Kamar Zekhnini , Anass Cherrafi, Imane Bouhaddou,

and Youssef Benghabrit

Abstract Currently, supply chain management incorporates technologies from

Industry 4.0, Internet of Things (IoT) and Smart Manufacturing to improve perfor-
mance and customer satisfaction. Supply chain digitalization helps organizations
meet new customer needs, supply issues and existing efficiency improvement stan-
dards. In other words, digitalization leads to a faster, flexible and efficient supply
chain. However, this digitalization has created new supply chain risks such as infor-
mation and cybersecurity risks. The purpose of this paper is to discuss key supply
chains 4.0 risks and examine the relationship between them. This study also presents
a literature review related to the subject to well understand the Supply chain 4.0
risk management. The authors use the approach of the analytical hierarchy model
(AHP) to analyze the supply chain 4.0 risks. The suggested study is useful for both
academics and practitioners, as it illustrated the relatedness of supply chains 4.0
risks. This would help organizations confront new risks facing interconnected supply
chains. The findings would help secure supply chains and strengthen organizations’
risk management.

Keywords Supply chain management 4.0 · Supply chain 4.0 risks · Risk
management · AHP

K. Zekhnini (B) · A. Cherrafi · I. Bouhaddou · Y. Benghabrit

LM2I Laboratory, ENSAM, Moulay Ismail University, 50500 Meknes, Morocco
e-mail: [email protected]
A. Cherrafi
e-mail: [email protected]
I. Bouhaddou
e-mail: [email protected]
Y. Benghabrit
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 89

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_7
90 K. Zekhnini et al.

1 Introduction

Digital technologies have stimulated huge advances in manufacturing development.

The role of new technologies in increasing industry competitiveness has become more
crucial in the digitalization era. In other words, the way the information is shared
has changed dramatically with digital technologies. All industries are impacted by
these new technologies [1]. Consequently, they altered the exchange of information
between supply chains (SCs).
There’s a considerable amount of risk sources in SCs. This encompasses unpre-
dictable exchange rates, supply disruptions, work accidents, competitive markets,
political instability, unstable demand, and natural disasters [2]. Furthermore, many
other sources have been identified through the digitization of SCs.
Thus, adopting digital supply chain systems, make companies face significant
risks [3]. In this sense, this study is of great importance to organizations. This provides
an outline of risk management in the digital age. Because risk management is impor-
tant in the presence of multitude of uncertainties in order to maintain the supply
chain efficiently [4].
Over the years, many researchers have focused on supply chain risk management
(SCRM) by contributing in the fields of risk definition, operationalization and miti-
gation [4]. To date, little attention has been given to risk management in supply chain
4.0 [5, 6]. The aim of this study is therefore to fill the literature gap with the prioriti-
zation of supply chain risks in the digital era. It addresses the new category of risks
born with the supply chains 4.0. In addition, this paper aims to provide academicians
with another supply chain 4.0 risk management perspectives.
This article is organized as follows: in the following section, we present the
literature review to well understand the supply chain 4.0 risk management. Section 3
describes the research methodology used for the study. Section 4 illustrates the appli-
cation of AHP for Supply chain 4.0 risks. And finally, Sect. 5 discusses findings
results and presents some guidelines to overcome most crucial risks facing supply
chains 4.0.

2 Literature Review

2.1 Industry 4.0

The advanced industrial factory digitization is called Industry 4.0 [7]. Industry 4.0
is referred to the fourth industrial revolution [8]. It is also known as ‘digital manu-
facturing’, ‘industrial internet’, or ‘integrated Industry’. It is expected to have the
potential to affect entire industries by changing the way products are produced,
manufactured, distributed and charged [9].
In other words, industry 4.0 is a view of smart factories designed with digital
cyber-physical systems [8]. It also refers to recent advances in technology where the
Analytic Hierarchy Process (AHP) for Supply Chain 4.0 Risks Management 91

Fig. 1 Industrial revolution

internet serves as an infrastructure for connecting physical objects, human partici-

pants, smart machines and processes across institutional borders to create a new kind
of networked, smart and agile value chain [10].
The Industry 4.0 relies on many advanced technologies such as cloud computing,
cyber-physical systems, 3D printing, internet of things, artificial intelligence, big data
analytics, advanced sensor-driven manufacturing technologies, augmented reality
and advanced robotics [11, 12].
We may identify four phases (Fig. 1) in the ongoing process called the Industrial
Revolution from the first mechanical loom, dating back to 1784, precisely 235 years
ago [13]. In fact, the first industrial revolution was symbolized by the implementation
of mechanical manufacturing systems using water and steam power. It started in the
1800s [14]. By the beginning of the twentieth century, the second industrial revolution
was described by the use of electricity in mass production. In the mid-twentieth
century, the third industrial revolution has known the introduction of automation and
microelectronic engineering into manufacturing [14]. The fourth industrial revolution
named ‘industry 4.0’ [15] ties the Internet of Things (IOT) to production [16].
The Industry 4.0 can make a great revolution in the management of the global
supply chain by allowing reaching unprecedented operative efficiency levels [17].
Moreover, by using new technologies, in the course of Industry 4.0, time and effort
can be reduced within entire supply chains [18]. Also, the supply chain becomes
more flexible and more transparent [19]. Furthermore, factories and businesses that
have adopted Industry 4.0 will know improvements in different areas [20]. In other
words, industry 4.0 offers the supply chain direct cost savings, Enhanced speed,
increased employee productivity, competitiveness in the global market and increased
profitability [21].
92 K. Zekhnini et al.

2.2 Supply Chain Management 4.0

Industry 4.0, the Internet of Things (IoT) and a host of new technologies allow
supply chain partners to more efficiently track, predict and respond to consumer
demand [22]. Hence, Supply chain 4.0 or digital supply chain (DSC) is a supply
chain, based on Web-enabled capabilities. A true DSC goes far beyond the hybrid
use of paper-based and IT-enabled processes, to rely entirely on synchronization,
system integration, and “smart” information-producing capabilities [23].
A digital supply chain is made up of hardware, software and communication
networks systems that enable interaction among globally distributed organizations.
It orchestrates the supply chain activities such as buying, making, storing, moving and
selling a product [24]. In other words, DSC is a process that uses new technologies
to identify processes from market to market (from distributor to supplier networks)
to understand, react and orchestrate bi-directionally [25].
The notion of “Supply chain management 4.0” will bring a variety of advantages
to the development of organizations. In fact, the digitalization enables the evolution
of the next generation of supply chains offering both flexibility and efficiency.

2.3 Supply Chain 4.0 Risks

When companies adopt digital supply chain systems to operate and coordinate with
their partners, they face significant risks. In other words, implementation and use
of IT systems in inter-organizational contexts, like digital supply chain networks, is
considered very vulnerable [3]. The risky aspect of digital supply chain systems is
due to the fact that successful implementation of these systems requires focal firms
to adapt to the external environment and multiple external parties, frequently beyond
their control [26].
Therefore, organizations need a risk management system to deal with risks. Risk
management is a structured approach that allows companies to know what the risk
is, who is at risk and what safeguards are appropriate for such risks [27]. Many risks
could threaten companies such as IT failures, piracy, theft, and cyber-attacks. Such
risks are one of the leading causes of reputation crises and loss of reputation value.
They can damage the supply chain performance [28]. Identified risks found in the
literature are presented in (Table 1).

3 Methodology

Various researchers in different sectors, such as education, industry, manufacturing,

engineering, etc., have widely used AHP methodology to solve various multi-criteria
decision problems [33–36].
Analytic Hierarchy Process (AHP) for Supply Chain 4.0 Risks Management 93

Table 1 Supply chain 4.0

Risks Type of risk References
risks
Macroeconomic Financial risks [29]
fluctuation
Supply chain overall
coordination
Product arrival variability Supply risks [29, 30]
(delays)
Loss of suppliers
Fluctuating Demand risks [29]
Demand
Lack of skilled workers Operational risks [29]
Breakdown of fractioning,
Natural/man-made Environmental risks [29]
disasters
Cyber-attack Industry 4.0 risks [3, 31, 32]
Information security risks
Risks to
computer-security

The Analytic Hierarchy Process (AHP) is a multi-criteria decision-making

approach. It was introduced by Saaty (1977 and 1994). In other words, the AHP
is a tool for decision support that can be used to solve complex decision issues.
This uses a hierarchical multi-level hierarchy of priorities, criteria sub criteria, and
alternatives in order to rank the importance of each variable [37].
Basically, AHP uses algebra metrics-based mathematical approach. It was used
as a method to determine the relevance of factors for decision-making or problem-
solving in order to achieve an objective. AHP incorporates in analysis the qualitative
and quantitative approach and integrates it as a single analytical problem into the
context. AHP uses the qualitative approach to converting issues into hierarchy. On the
other hand, based on a quantitative approach, it uses more of the pair-wise comparison
strategy to achieve more reliable responses and reliability [38].
AHP has been a favorite research decision method in many fields, such as engi-
neering, food, industry, ecology, health, and government, due to its mathematical
simplicity and flexibility. In addition, The AHP not only helps analysts make the
best decision, but also gives them a good reason for the choices they make. So,
AHP methodology is used in this study to rank and prioritize supply chain risks in
the digital era. This provides a framework for dealing with logical, quantitative and
qualitative dimensions of multiple criteria situations.
94 K. Zekhnini et al.

Fig. 2 Hierarchy of risks factors

4 AHP to Supply Chain 4.0 Risk Management

4.1 Developing the AHP Model

We first established a risk hierarchy as shown in Fig. 2 and then used pair-wise
comparisons [39] to determine the impact of risks on the supply chain 4.0. This
phase requires the formulation of an AHP model hierarchy Containing objective,
factors, sub-factors, and results. Note that the first level of the hierarchy is our goal
which is prioritizing supply chain 4.0 risks. The second level in the hierarchy is
constituted by the criteria we will use to decide. Hence, to accomplish our goal, five
major risks are defined namely financial risks, supply risks, demand risks, operational
risks and industry 4.0 risks. The third level of hierarchy consists of 11 risks in the five
types of risks. The fourth level of hierarchy consists of the results which is effective
risk prioritization in supply chain 4.0.

4.2 Weights Evaluation

Not all the criteria will have the same importance. Therefore, the second step in the
AHP process is to derive the relative priorities (weights) for the criteria. The pair-
wise comparison decision matrices are developed by expert guidance to assess the
relative importance of the supply chain 4.0 risks.
Analytic Hierarchy Process (AHP) for Supply Chain 4.0 Risks Management 95

4.2.1 Establishment of Pair-Wise Comparison Matrices

The precise assessment of the relevant data is one of the most important steps in many
decision-making processes. So, in term of each criterion, comparisons are used in
pairs to determine the relative importance of each alternative. In this method, the
decision-makers express their preferences on the importance of one pair comparison
at a time.
Comparisons are made on a nine-point scale that transforms individual prefer-
ences into available alternatives. The values of the AHP pairwise comparisons are
determined by Saaty’s (1980) scale. The available values for the pairwise compar-
isons are set according to this scale: {9, 8, 7, 6, 5, 4, 3, 2, 1, 1/2, 1/3, 1/4, 1/5, 1/6, 1/7,
1/8, 1/9} (Table 2). Overall, the matrix of comparison is defined as = (ai j ) where A
= reciprocal matrix with 1 in the diagonal and the lower triangular matrix is filled
using formula ai j = a1ji .
Given a judgment matrix with pairwise comparisons, the corresponding maximum
left eigenvector is approximated by using the geometric mean of each row. That is,
the elements in each row are multiplied with each other and then the n-th root is taken
(where n is the number of elements in the row). Tables 3, 4, 5, 6, 7 and 8 presents
the pair-wise comparison matrix for different level of hierarchy.

Table 2 The comparison scale of pair wise Saaty [40]

Intensity of importance Definition Explanation
1 Equal importance Two elements contribute equally to the
objective
3 Moderate importance Experience and judgment slightly favor
one element over another
5 Strong importance Experience and judgment strongly favor
one element over another
7 Very strong importance One element is favored very strongly over
another, its dominance is demonstrated in
practice
9 Extreme importance The evidence favoring one element over
another is of the highest possible order of
affirmation

Table 3 Pair-wise comparisons of risks types

Criterea I S D O F Weights
I 1 3 5 4 4 0.449
S 1/3 1 3 5 4 0.276
D 1/5 1/3 1 2 2 0.116
O ¼ 1/5 1/2 1 2 0.090
F ¼ ¼ 1/2 ½ 1 0.068
Sum 2.033 4.783 10.000 12.500 13 CR = 0.09
96 K. Zekhnini et al.

Table 4 Pair-wise
Criterea F1 F2 Weights
comparison judgement: of
financial risks (F) F1 1 5 0.833
F2 1/5 1 0.167
Sum 1.200 6.000 CR = 0

Table 5 Pair-wise
Criterea S1 S2 Weights
comparison judgement: of
supply risks (S) S1 1 2 0.667
S2 ½ 1 0.333
Sum 1.500 3.000 CR = 0

Table 6 Pair-wise
Criterea D1 D2 Weights
comparison judgement: of
demand risks (D) D1 1 2 0.667
D2 ½ 1 0.333
Sum 1.500 3.000 CR = 0

Table 7 Pair-wise
Criterea O1 O2 Weights
comparison judgement: of
operational risks (O) O1 1 7 0.875
O2 1/7 1 0.125
Sum 1.143 8.000 CR = 0

4.2.2 Calculation of the Degree of Consistency

One of AHP’s important tasks is to measure the estimated vector consistency level.
Thus, it is necessary to check the consistency once decisions have been entered.
AHP calculates a consistency ratio (CR) comparing the consistency index (CI) of the
matrix in question versus the consistency index of a random-like matrix (RI). Saaty
(2012) provides the calculated RI value for matrices of different sizes as shown in
Table 9.

Table 8 Pair-wise comparison judgement: of industry 4.0 risks (F)

Criterea I1 I2 I3 Weights
I1 1 6 2 0.705
I2 1/6 1 3 0.295
I3 ½ 1/3 1 0.225
Sum 1.167 7.000 5.000 CR = 0.87
Analytic Hierarchy Process (AHP) for Supply Chain 4.0 Risks Management 97

Table 9 Average random index values

N 1 2 3 4 5 4 7 8 9
RCI 0 0 0.58 0.90 1.12 1.24 1.32 1.41 1.45

If all the comparisons are perfectly consistent, then the following relation should
always be true for any combination of comparisons taken from the judgment matrix:
a i j = a i k ∗ a k j . The following steps are taken to calculate the CR coefficient. First,
the consistency index (CI) needs to be estimated. This is done by adding the columns
in the judgment matrix and multiply the resulting vector by the vector of priorities
(i.e., the approximated eigenvector) obtained earlier. This yields an approximation
of the maximum eigenvalue, denoted by λmax . So we have:
• CI = (λmax – n)/(n – 1)
• and CR = CI/RCI
Saaty (2012) has shown that it is acceptable to continue the AHP analysis with
a consistency ratio (CR) of 0.10 or less. If the consistency ratio reaches more than
0.10, the judgements need to be updated and correct the source of the inconsistency.

5 Discussion and Managerial Implication

The final rankings of various supply chain 4.0 risks are presented in Table 10.
According to Risk rankings, we can determine crucial risks that need managers’
attention. I1 (Cyber-attack) is the most important risk factor because it has a highest
priority weight (0.32). The second most critical risk factor is S1 (Product arrival
variability(delays)) with priority weight of 0.18. The third most important risk factor

Table 10 Supply chain 4.0 risk rank

Criterea Local Weights Global Weights Overall weights Risk rank
D1 0.667 0.116 0.0773 7
D2 0.333 0.116 0.0387 9
F1 0.833 0.068 0.0567 8
F2 0.167 0.068 0.0113 10
I1 0.705 0.449 0.3164 1
I2 0.295 0.449 0.1326 3
I3 0.225 0.449 0.1012 4
O1 0.875 0.09 0.07875 6
O2 0.125 0.09 0.01125 11
S1 0.667 0.276 0.184092 2
S2 0.333 0.276 0.091908 5
98 K. Zekhnini et al.

is I2 (Information security risks) with the risk exposure of 0.13. The fourth important
risk factor is I3 (Risks to computer-security) with the risk exposure of 0.10. The fifth
important risk factor is S2 (Loss of suppliers) with the risk exposure of 0.09. Then,
we have O1 (Lack of skilled workers), D1 (Fluctuating), F1 (Macroeconomic fluc-
tuation), D2 (Demand), F2 (Supply chain overall coordination), and O2 (Breakdown
of fractioning). The risk exposure for first two risks I1 and S1 is considerably high
than the other risks. In fact, the first two important risks add up to 50% of global
weight.
The graph (Fig. 3) represents the drops in risk weights of various risks. It shows that
I1 has the highest risk weight. Then S1 risk weight decreases by 13%. Consequently,
these two risks should take managers high attention to been mitigated in real time.
According to the figure, we can divide the risks into clusters on the basis of the
weights range. In other words, risks cluster where managers need to put same amount
of attention are I2, I3. Next cluster that managers should not neglect is S2, O1, D1,
F1. Interestingly, we can observe that the industry 4.0 and supply risks are the most
crucial risks.
The supply chain cyber-attacks are higher now than ever. Industries must under-
stand how to secure their supply chains from these expanding attacks. Having a
secure supply chain means that the organizations softwares are up to date. Training
employees on cyber security would reduce the risk of cyber-attacks. In addition,
managers have to think completely about supplier risk in other ways. They need to
establish management strategies, with both proactive and reactive mechanisms. This
should include identifying risk-fraught materials, educating risk management staff,
creating methods for breaching their complex supply chains, and assuring a clear
visibility of the whole supply chain in real time. To sum up, increasing visibility
in the supply chain, establishing a trusting relationship with suppliers and having
strategic plans to face threats will help organizations reduce supply chain risks.
Table 11 presents some actions to mitigate the most crucial SC 4.0 risks.

Fig. 3 Final weights of each 0.35

supply chain 4.0 risk
0.30
0.25
0.20
0.15
0.10
0.05
0.00
I1 S1 I2 I3 S2 O1 D1 F1 D2 F2 O2
Analytic Hierarchy Process (AHP) for Supply Chain 4.0 Risks Management 99

Table 11 Supply chain 4.0 mitigation actions

Risk Actions
Cyber-attack (I1) – Identify and document asset vulnerabilities
– Continuous employee education
– Monitor network traffic for suspicious activity
– Upgrade software immediately
– Harden external facing web applications
Product arrival variability (delays) (S1) – Evaluate your supplier’s processes with an ISO 9001
quality audit
– Regular analysis of all available supplier information
– Having a well-defined response strategy and crisis
management plan
– Using Supply Chain Predictive Analytics to reduce
risks
Information security risks (I2) – Upgrade Authentication inside and out
– Know where sensitive data resides
– Develop data protection strategy
– Include encryption monitoring
Risks to computer-security (I3) – Implement rigorous application development testing
– Perform periodic penetration assessments and
vulnerability assessments
– Use reputable anti-virus
Loss of suppliers (S2) – Diversify Suppliers
– Maintaining a close relationship with a supplier
– Invoking certain policies, procedures, processes or
systems
– Switching suppliers

6 Conclusion

Supply chain risk management has shown excellent interest in manufacturing studies
and this study is an effort to outline the current risk management for the digital supply
chains. The identification of key risks and future study paths could be acquired from
these analyzes.
Due to the supply chain’s interconnectivity and complexity, managers are
becoming more conscious of the supply chain risk. Notwithstanding this, the supply
chains 4.0 risks have not been widely studied so far. In other words, managers often
handle issues with the supply chain without discerning risks. Because of this liter-
ature gap, the present paper examined the supply chain 4.0 risk factors. We listed
11 risks under 5 major risk factors. AHP method has been used to prioritize Supply
chain 4.0 Risks.
Our paper presents an overview about supply chain 4.0 risks and how risk priori-
tization can be implemented by AHP. The research has implications for both theory
and practice. It can be a roadmap for research beginners in supply chain 4.0 risk
management to explore and work on the future issues of the subject. It, also, presents
100 K. Zekhnini et al.

some actions to mitigate the most crucial supply chain 4.0 risks. Moreover, the find-
ings of this study can enable the practitioners in the entire supply chain to achieve
successful risk management. In the other hand, this study also presents some limi-
tations due to the small number of papers used for the analysis of the subject. Our
results remain to be discussed and reinforced in future studies. In addition, due to the
various new risks that could arise in the supply chain 4.0, it is desirable to review the
AHP calculus so far and consider introducing new risks or new level in the hierarchy
model.

References

1. Ivanov, D., Dolgui, A., Sokolov, B.: The impact of digital technology and Industry 4.0 on the
ripple effect and supply chain risk analytics. Int. J. Prod. Res. 57, 829–846 (2019). https://doi.
org/10.1080/00207543.2018.1488086
2. Er Kara, M, Oktay Fırat, S., Ghadge, A.: A data mining-based framework for supply chain risk
management. Comput. Ind. Eng. 105570 (2018). https://doi.org/10.1016/j.cie.2018.12.017
3. Xue, L., Zhang, C., Ling, H., Zhao, X.: Risk mitigation in supply chain digitization: system
modularity and information technology governance. J. Manag. Inf. Syst. 30, 325–352 (2013).
https://doi.org/10.2753/mis0742-1222300110
4. Ho, W., Zheng, T., Yildiz, H., Talluri, S.: Supply chain risk management: a literature review.
Int. J. Prod. Res. 53, 5031–5069 (2015). https://doi.org/10.1080/00207543.2015.1030467
5. Gottlieb, S., Ivanov, D., Das, A.: Case studies of the digital technology impacts on supply chain
disruption risk management. In: Schröder, M., Wegner, K. (eds.) Logistik im Wandel der Zeit
– Von der Produktionssteuerung zu vernetzten Supply Chains, pp. 23–52. Springer Fachmedien
Wiesbaden, Wiesbaden (2019)
6. Radanliev, P, Treall, L., Santos, O., Montalvo, R.M.: Cyber risk from IoT technologies in the
supply chain-discussion on supply chains decision support system for the digital economy
(2019). https://doi.org/10.13140/rg.2.2.17286.22080
7. Glas, A.H., Kleemann, F.C.: The Impact of Industry 4.0 on procurement and supply
management: a conceptual and qualitative analysis. Int. J. Bus. Manag. Inven. 5, 55–66 (2016a)
8. Thames, L., Schaefer, D.: Software-defined cloud manufacturing for Industry 4.0. Procedia
CIRP 52, 12–17 (2016). https://doi.org/10.1016/j.procir.2016.07.041
9. Hofmann, E., Rüsch, M.: Industry 4.0 and the current status as well as future prospects on
logistics. Comput. Ind. 89, 23–34 (2017). https://doi.org/10.1016/j.compind.2017.04.002
10. Schumacher, A., Erol, S., Sihn, W.: A maturity model for assessing Industry 4.0 readiness and
maturity of manufacturing enterprises. Procedia CIRP 52, 161–166 (2016). https://doi.org/10.
1016/j.procir.2016.07.040
11. Rajnai, Z., Kocsis, I.: Labor market risks of industry 4.0, digitization, robots and AI. In:
SISY 2017—IEEE 15th International Symposium on Intelligent Systems and Informatics,
Proceedings, pp. 343–346 (2017). https://doi.org/10.1109/SISY.2017.8080580
12. Ivanov, D., Dolgui, A., Sokolov, B.: The impact of digital technology and Industry 4.0 on the
ripple effect and supply chain risk analytics. Int. J. Prod. Res. 0, 1–18 (2018). https://doi.org/
10.1080/00207543.2018.1488086
13. Bloem, J., van Doorn, M., Duivestein, S., Excoffier, D., Maas, R., van Ommeren, E.: The Fourth
Industrial Revolution, p. 40
14. Da, Xu.L., Xu, E.L., Li, L.: Industry 4.0: state of the art and future trends. Int. J. Prod. Res. 56,
2941–2962 (2018). https://doi.org/10.1080/00207543.2018.1444806
15. Hahn, G.J.: Industry 4.0: a supply chain innovation perspective. Int. J. Prod. Res. 0, 1–17
(2019). https://doi.org/10.1080/00207543.2019.1641642
Analytic Hierarchy Process (AHP) for Supply Chain 4.0 Risks Management 101

16. da Silva, V.L., Kovaleski, J.L., Pagani, R.N.: Technology transfer in the supply chain oriented
to industry 4.0: a literature review. Technol. Anal. Strat. Manag. 0, 1–17 (2018). https://doi.
org/10.1080/09537325.2018.1524135
17. Rodriguez Molano, J.I., Contreras Bravo, L.E., Trujillo, E.R.: Supply chain architecture model
based in the industry 4.0, validated through a mobile application. CES 10, 1581–1594 (2017).
https://doi.org/10.12988/ces.2017.711186
18. Müller, J.M., Voigt, K.I.: The impact of Industry 4.0 on supply chains in engineer-to-order
industries—an exploratory case study. IFAC-PapersOnLine 51, 122–127 (2018). https://doi.
org/10.1016/j.ifacol.2018.08.245
19. Schröder, M., Indorf, M., Kersten, W.: Industry 4.0 and its impact on supply chain risk
management. Transp. Syst. 11, (2014)
20. Glas, A.H., Kleemann, F.C.: The impact of industry 4.0 on procurement and supply
management: a conceptual and qualitative analysis. Int. J. Bus. Manag. Inven. 5, 55–66 (2016)
21. Recknagel, R.: The Supply Chain Manager’s Guide to Industry 4.0. Supply Chain Excell. 7
(2017)
22. Brinch, M., Stentoft, J.: Digital supply chains: Still more “wannabe” than practise. Dilf
Orient. 54(2), 22–28 (2017)
23. Rouse, M., Cobb, M.: Transport Layer Security (TLS) (2016). http://searchsecurity.techtarget.
com/definition.Transp ort-Layer-Security-TLS. Accessed 19 Nov 2016
24. Bhargava, B., Ranchal, R., Othmane, L.B.: Secure information sharing in digital supply chains.
In: 2013 3rd IEEE international advance computing conference (IACC). IEEE, pp 1636–1640
(2013)
25. Cecere, L.: Embracing the digital supply chain. In: Supply Chain Shaman (2016). https://
www.supplychainshaman.com/demand/demanddriven/embracing-the-digital-supply-chain/.
Accessed 2 Jan 2020
26. Riggins, F.J., Mukhopadhyay, T.: Interdependent benefits from interorganizational systems:
opportunities for business partner reengineering. J. Manag. Inf. Syst. 11, 37–57 (1994)
27. Tupa, J., Simota, J., Steiner, F.: Aspects of risk management implementation for Industry 4.0.
Procedia Manuf. 11, 1223–1230 (2017). https://doi.org/10.1016/j.promfg.2017.07.248
28. Gaudenzi, B., Siciliano, G.: Managing IT and cyber risks in supply chains. In: Khojasteh, Y.
(ed.) Supply Chain Risk Management, pp. 85–96. Springer Singapore, Singapore (2018)
29. Faizal, K., Palaniappan, P.L.K.: Risk assessment and management in supply chain. Glob. J.
Res. Eng. 14, 19–30 (2014)
30. Yang, Q., Wang, Y., Ren, Y.: Research on financial risk management model of internet supply
chain based on data science. Cogn. Syst. Res. 56, 50–55 (2019). https://doi.org/10.1016/j.cog
sys.2019.02.001
31. Colicchia, C., Creazza, A., Menachof, D.A.: Managing cyber and information risks in supply
chains: insights from an exploratory analysis. Supply Chain. Manag. 24, 215–240 (2019).
https://doi.org/10.1108/SCM-09-2017-0289
32. Schröder M, Indorf M, Kersten W (2014) Industry 4.0 and Its Impact on Supply Chain Risk
Management, pp. 15–18
33. Al-Barqawi, H., Zayed, T.: Infrastructure management: integrated AHP/ANN model to evaluate
municipal water mains’ performance. J. Infrastruct. Syst. 14, 305–318 (2008)
34. Barbarosoglu, G., Yazgac, T.: An application of the analytic hierarchy process to the supplier
selection problem. Prod. Inven. Manag. J. 38, 14 (1997)
35. Mathiyazhagan, K., Diabat, A., Al-Refaie, A., Xu, L.: Application of analytical hierarchy
process to evaluate pressures to implement green supply chain management. J. Clean. Prod.
107, 229–236 (2015)
36. Vaidya, O.S., Kumar, S.: Analytic hierarchy process: an overview of applications. Eur. J. Oper.
Res. 169, 1–29 (2006)
37. Triantaphyllou, E., Mann, S.H.: Using the analytic hierarchy process for decision making in
engineering applications: some challenges, p. 11 (1995)
38. Safian, E.E.M., Nawawi, A.H.: The evolution of Analytical Hierarchy Process (AHP) as a
decision making tool in property sectors. 4 (2011)
102 K. Zekhnini et al.

39. Wind, Y., Saaty, T.L.: Marketing applications of the analytic hierarchy process. Manage. Sci.
26, 641–658 (1980). https://doi.org/10.1287/mnsc.26.7.641
40. Saaty, T.L.: Fundamentals of Decision Making and PRIORITY THeory with the Analytic
Hierarchy Process. RWS Publications (2000)
SmartDFRelevance: A Holonic Agent
Based System for Engineering Industrial
Projects in Concurrent Engineering
Context

Abla Chaouni Benabdellah, Imane Bouhaddou, and Asmaa Benghabrit

Abstract In the era of industry 4.0, the design of complex engineered systems
is a challenging mission, and most researchers would argue that it is linked to
intentional action and it cannot emerge out of complexity. In fact, to achieve the
common ‘cost/quality/delay’ targets, engineers must work together to create the
best overall solution from their individual components, and exploit and regenerate
the required intellectual capital (Knowledge). Therefore, to be effective, we need a
knowledge based system that takes into consideration all the users’ needed methods
to create share reuse and evaluate knowledge. Thus, using complex adaptive system
theory, mainly Distributed Artificial Intelligence (holonic agent paradigm), a new
knowledge-based system can be designed to address this research issue. In this regard,
the purpose of this paper is first to provide a comparative analysis of the appropriate
method that considers the global challenges for designing a product in a concur-
rent engineering context. Second, a holonic multi-agent system, called SmartD-
FRelevance is proposed based on Agent-oriented Software Process for Complex
Engineering Systems methodology.

Keywords Complex adaptive system theory · Knowledge based system ·

Distributed artificial intelligence · Concurrent engineering · Holonic multi-agents
system

A. C. Benabdellah (B) · I. Bouhaddou

LM2I Laboratory ENSAM, Moulay Ismaïl University, Meknès, Morocco
e-mail: [email protected]
I. Bouhaddou
e-mail: [email protected]
A. Benghabrit
LMAID Laboratory, ENSMR, Mohammed V University, Rabat, Morocco
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 103

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_8
104 A. C. Benabdellah et al.

1 Introduction

Industries are engaging in constant performance enhancement activities in the current

economic and industrial environment to remain competitive in their core business.
In general, performance enhancement levers can concern (1) maximizing product
efficiency and organizational design processes or (2) improving internal process
efficiency or even (3) improving human performance by valorizing knowledge and
competences.
The first axis is well recognized and investigated by enterprises. It embeds methods
and tools such as functional analysis, dependability, statistical process control or
modelling and simulation. The second axis integrates, for instance, all the method-
ologies and tools of project management, agile methodologies, engineering system,
or quality management systems. While the third one remain complex due to the
intrinsic nature of the knowledge and its volatile dimension. In fact, the design of a
new product in terms of cost, flexibility, assembly, quality, safety, serviceability and
environmental issues has specific characteristics that directly affect how knowledge
is performed. In addition, giving an outline of a project or recalling the collabora-
tive work of past projects is still a challenging mission for practitioners. As a result,
designers must take into account the experience of older systems, work with draft
texts, with professional understanding and coordination, and validate the veracity
of information without knowing it. Thus, in order to design product in concurrent
engineering context, experts from different disciplines must work together to create
from their individual components the best overall structure. This scenario produces
a heterogeneous and hierarchical ecosystem similar to a multi-agent system (MAS)
where agents (actors, experts, designers, etc.) are organized to solve individual parts
of a complex issue in a coordinated way.
Notwithstanding, the development of a knowledge management tool, piloted by
agents, based on role monitoring and organizational perspective is of crucial impor-
tance [1]. Indeed, the interest in using a MAS is first to be able to benefit from the
capacities of autonomy, reactivity, proactivity, and social skills. Second, MAS is
working in a decentralized manner is capable of using distributed and incomplete
information and knowledge sources [2]. In addition to that, the advantages of using an
organizational approach derive from its potential to illustrate the processes of knowl-
edge sharing and development, rather than relying on knowledge formalization and
modeling [3]. It is also capable of offering many advantages such as heterogeneity
of languages, modularity, multiple possible architectures and security of the system.
However, for successful application and deployment of MAS, methodologies are
essential. Several methods have been proposed and most of them recognize that the
building process is a traditional software system. In particular, they all recognize the
idea that a MAS can be considered as an organization of individuals in which it play a
role. Therefore, the current methodologies are not knowledge-based driven, they do
not take into account all the processes of user-made knowledge creation, sharing, and
evaluation [4]. For this purpose, we propose in this paper a new holonic knowledge-
based system, called SmartDFRelevance, following the Agent-oriented Software
SmartDFRelevance: A Holonic Agent Based System … 105

Process for Complex Engineering Systems methodology (ASPECS) dedicated to

the analysis, design, and deployment of complex engineered systems. More clearly,
the proposed holonic multi-agent knowledge-based system aims at creating project
memories for indexing and organizing the knowledge to be capitalized during engi-
neering industrial projects in concurrent engineering context (MemoDFRelevance).
It provides also a domain ontology (OntoDFRelevance) to perceive the vocabulary
and semantics of the knowledge used during the engineering project while consid-
ering different virtues such as assembly, manufacture, safety, quality, supply chain,
environment (DesignForRelevance) [5].
This paper is structured as follows; Sect. 2 presents the basic terminologies.
Section 3 presents an extensive review of the existing knowledge management
systems found in the literature. Section 4 presents a comparative analysis of the
various organizational models for the deployment of multi-agent systems. The
interest and the particularity of the ASPECS methodology in bringing together the
advantages of organizational approaches as well as those of the holonic vision in
modeling complex engineered systems is also presented in this section. Section 5
describes the overall architecture of the proposed SmartDFRelevance (Smart Design
for Relevance) knowledge management system. This new knowledge-based system
ensures the whole knowledge lifecycle identification, creation and acquisition,
passing through validation and capitalization up to sharing and reuse. The conclusions
and future works are drawn in the final section.

2 Basic Terminologies

2.1 Multi-Agents Systems (MAS)

Due to its ability to solve complex computing issues [6, 7], Distributed Artificial
Intelligence (DAI) has gained considerable attention from academia in recent years.
Its algorithms are generally divided into three groups, depending on the specific
approaches used to solve the tasks: parallel AI, Distributed Problem Solving (DPS)
and Multi-Agent Systems (MAS). In fact, Parallel AI includes the development
of concurrent algorithms, languages, and architectures to increase the efficiency of
traditional AI algorithms by using sequential tasks [7]. DPS involves dividing a task
into subtasks. Each subtask is assigned to a node between a set of cooperative nodes,
known as computing entities while MAS consists of independent entities known as
agents.
However, there are multitudes of definitions for agents that derive from different
application-specific agent features. Etymologically, the word agent is inspired by
the Latin “agere” which means, an agent is an entity that acts. Drogoul defined
an “agent as an entity that can exercise control over local processes of perception,
communication, learning, reasoning, decision-making or execution” [8]. The authors
in [6] defined “an agent as a computer system, situated in an environment, and which
106 A. C. Benabdellah et al.

Fig. 1 Multi-Agent Simulation principle according to Drogoul [8]

autonomously and flexibly to achieve the objectives for which it was designed”.
Ferber described “an agent as an autonomous physical or abstract entity which is
capable of acting on itself and on its environment, which can communicate with
other agents in a common environment and whose behavior is the result of his obser-
vations, his knowledge and his interactions with other agents which tends to satisfy
his objectives” [9]. Authors in [10] defined an “agent as an entity which is placed in an
environment and senses different parameters that are used to make a decision based
on the goal of the entity. The entity performs the necessary action on the environ-
ment based on this decision”. Therefore, a literature analysis shows that there is no
exact definition of agent, but generally, there is a consensus on common properties
(autonomy, rationality, responsiveness, adaptability, flexibility, proactivity, socia-
bility, interactivity, etc.) and keywords (entity, environment, parameters, action, etc.)
that characterize the concept of agent [11]. However, while an agent working on
its own is capable of taking action (based on autonomy), the agents’ real benefit
can only be gained by collaborating with other agents in partnership. In this sense,
multiple agents working together to simulate and understand complex systems [12]
are referred to as Multi-Agents System (MAS).
According to Wooldridge [6] a “MAS is a set of heterogeneous agents with own
and active motivations to achieve their goals. Each agent can share and organize
their roles and activities with other agents. Besides, agents can deal with complex
situations within these systems by sharing their knowledge of the environment”.
More clearly, the phenomena are categorized into a series of acting or communicating
elements. Each of these elements is modeled by an agent, and the overall model is
the result of agent interactions [9, 13]. The multi-agent simulation principle can be
illustrated in Fig. 1.

2.2 Holonic Multi-Agents Systems (HMAS)

Despite the benefits of MAS, this approach presents different challenges and some
open issues need to be answered. One such concern is how an organization can be
viewed as agents [1]. In other words, how to interpret the fact that a group of agents
SmartDFRelevance: A Holonic Agent Based System … 107

displays a specific behavior in interaction and that they behave as if they were a
single entity at a certain level of abstraction [14]. Many researchers have studied this
issue and have suggested several models influenced by their experience in various
fields [15]. Thus, due to the usefulness of holons where goals can be recursively
decomposed into subtasks that can be assigned to individual holons [16], Holonic
system is choosen in this paper. Questions arise, what is a Holonic agent? What are
the main differences between holons and agents?
In 1967, the Hungarian writer and philosopher Arthur Koestler [18], who had
studied social organisms and organizations, pointed out that an entity that can be a
whole or a part cannot be considered an absolute in real life. He then introduced the
concept of Holon’s “holos”, meaning the whole and the “on” suffix that indicates
a component like “proton.” Therefore, a holon is defined by Koestler as “part of a
whole or a larger organization that fulfills three conditions: stability, autonomy, and
cooperation” [18]. Besides, the holon can sometimes be seen as an atomic entity,
sometimes like a collection of interacting holons, depending on the level of observa-
tion. Likewise, a group of separate holons can be viewed as a collection of interacting
entities, or as parts of a holon of a higher level [16]. At a higher level, the composed
holon is called a super-holon and the holons that make up a super-holon are called
sub-holons or member holons (Fig. 2).
Thus, holons are by definition composed of other holons while agents are not
necessarily composed of other agents. This does not mean agents cannot be composed
of agents, but the general assumption is that “agents are atomic entities.” Nonetheless,
as stated by Hilaire et al. [14], almost all the proposed agent architectures have not
discussed the general issue of how to handle “agent” sets as higher-order entities,
such as how to treat organizations as agents. In this respect, Holonic Multi-agents
system (HMAS) attempt to address this problem.

Fig. 2 Vertical and horizontal decomposition of a Holon [17]

108 A. C. Benabdellah et al.

3 The Agent Paradigm in Knowledge Engineering

Artificial Intelligence (AI) research has centered in earlier days on establishing

formalisms, mechanisms of inference, and strategies for knowledge-based systems
(KBS) operationalization. Generally, the development efforts were limited to small
KBSs to assess the feasibility of different approaches. Despite the very promising
results provided, the transition of this technology to industrial use to build large
KBSs in many cases failed. Therefore, there is a need for alternative methodological
approaches to deal with these unsatisfactory situations. For this purpose, the field of
Knowledge Engineering (KE) was born and a variety of concepts, approaches, and
techniques that have greatly improved the information creation, use, and deployment
process was created [19].
As an engineering discipline that requires a high level of human expertise and
involves integrating knowledge into computer systems, KE means to gather, study,
organize and represent knowledge. In other words, it is defined as a modeling process
that represents knowledge to store it, to communicate it or to externally manipulate
[20]. Nonetheless, since knowledge cannot be reduced to a static resource-based
perspective due to its inherently dynamic nature, KE solutions need to focus on
knowledge models with richer semantics and improved traceability [21]. According
to this statement, and compared to other KE strategies, a multi-agent organiza-
tional structure allows information exchange and development processes to be
explained instead of focusing on information formalization and simulation [3]. In
fact, according to Kusch [22], the services that the multi-agent systems can offer are:
• Finding, collecting, evaluating and classifying knowledge from different data
sources.
• Addressing the system’s integration or omission of knowledge.
• Explaining the quality and reliability of the standardized knowledge of the system.
• Learning gradually throughout the knowledge management process.
Multi-agent systems are therefore well suited for managing heterogeneous and
distributed knowledge by their ability to adapt to dynamic environments. In addi-
tion, coordination between agents gives them the collective ability to manage large
amounts of data. To do so, several published works have implemented such services
upon agent cooperation to address complicated problems related to the knowledge
types and assistant management agents depending on the actor’s needs.
In the domain of complex design [23], Monceyron [24] has created an architec-
ture of blackboard construction, EXPORT, applied to harbor design. More recently,
after Scalabrin et al. [25] work on cognitive agent systems, Shen and Barthès [26]
developed a generalized multi-agent framework, DIDE, for mechanical engineering
design. Barthès and de Azevedo [27] suggested a framework for the study and imple-
mentation of multi-agent schemes. With the same perspective, Barthès et Tacla [28]
creates a multi-agent system to support the “ba” externalization approach used by
the R&D department. A context for how knowledge is produced and shared [29] is
given by the “ba” theory.
SmartDFRelevance: A Holonic Agent Based System … 109

Several other published works focus on the efficiency and information sharing
effects of organizational structures within a group [30, 31]. In [32], Dignum and
Meyer illustrate the social aspect of the organization by describing the agency orga-
nizations as collections of individuals and their interactions that are governed by
social order processes and created for a common goal by autonomous actors. More-
over, Boissier and Demazeau [33] argue that an agent organization can be interpreted
easily as a collection of restrictions followed by a collective of agents to facilitate the
accomplishment of their objectives. From these two claims, it is apparent that agents
must take into account the group’s priorities within an organization so that they can
accomplish their own goals. In addition to that, Guizzardi [34] emphasizes the effec-
tiveness of knowledge sharing among agent organizations. He suggests including the
idea of the information within organizations: an agent association is a knowledge-
sharing environment in which agents work together and share expertise to carry out
their work. Carley [35] states that human organizations ‘research and review helps
to develop computational models that can be used to improve three organizational
perspectives: structure, sharing of information and social cooperation. Together with
teamwork, stigmergy, adaptivity, and evolution, the agent model allows the creation
of software engineering applications to improve human organizations’ performance.
Further research has discussed the potential of the agent to run distributed data
and solve problems of collaboration in a system of action. In [32], Van Elts proposes
taking into account a domain’s relational aspect along with the needs and goals of
the participants in the same domain. This approach is referred to as Agent Mediated
Knowledge Management (AMKM). For dynamic environments, these organizations
make information acquisition simpler. With the same perspective, the creation of
frameworks such as FRODO [3], CoMMA [36], Edamok [37] and KRAFT [38]
was developed as complementary methods for information management (workflows,
ontologies, information analysis structures, and so on). All these studies have led
to the Multi-Agent Information system (MAIS) that handles and uses centralized
information to view, update, and compile heterogeneous information [36, 39].
Nevertheless, even if agents’organization allow to consider social and collabo-
ration aspects of project teams, there is a need to provide the ability to handle the
information to these agents. To do so, several authors use ontology combined with the
benefits of a multi-agents system to understand the world of truth. Buccafurri et al.
[40] and Wooldridge [6] use ontology to provide agents with an abstract description
of their related individual users desires and behaviors. Other researchers use ontology
to assist agents in selecting the most appropriate agents to be contacted for knowl-
edge sharing purposes [41, 42]. Such structures are typically designed to prevent
other agents from accessing the ontologies of other agents; they maintain an indi-
vidualistic conception of agent communities. Another researcher design agents that
create their ontologies automatically by tracking the behavior of users [43]. In this
approach, the agents can extract logical rules that characterize the actions of the user
from the user interests defined in the ontology automatically. Further Researchers
have developed their multi-agent structures by following a shared view of the agent
societies where the ontology of each agent is believed to be available, even partially
to the other agents [44].
110 A. C. Benabdellah et al.

However and based on this exhaustive overview, we conclude the need to leverage
and combine the benefits of the organizational perspective, the ontology and the
multi-agent approach. For this purpose, the proposed MAS called SmartDFRele-
vance is composed of a common ontology used by the agents to perceive the whole
knowledge of the user’s domain. In addition to that, we consider using a social
and collaborative approach based on role modeling and coordination throughout the
project. However, in order to propose the new holonic knowledge-based system, we
need first to choose the appropriate design approach for the deployment of HMAS.
To do so, the next section compares and analyzes the main existing organizational
MAS design approaches to address their key elements and the possible shortcoming
when implemented as a KE approach.

4 Design of a Knowledge-Based System: ASPECS

Methodology

Many agent-oriented methodologies with a specific organizational vision have

been developed in recent years to support modeling or designing complex engi-
neered systems. Several authors evaluated the characteristics of these methodologies
taking into consideration various technical parameters (criteria) for evaluating and
comparing these methodologies [1, 15, 17]. In fact, they have criteria related to the
methodology process such as development life-cycle (waterfall or iterative), coverage
life-cycle (analysis, design, implementation, etc.) or the application domain (inde-
pendent or dependent). Others are related to the methodology itself such as system
structure (one or different abstraction level), agent-centered or organization-centered.
Furthers are related to the supportive feature of the methodology such as dynamic
structure, open system, standard integration, and ontology consideration. But in
general, methodologies can be categorized into two main categories: Rules-based
organizational Methodologies and Non-rules-based organizational methodologies.
Table 1 gives an overview and a classification of the existing methodologies. A
description of each category is also presented.
Notwithstanding, based on Table 1and the different comparative studies analyzed
in the literature [1, 5, 15, 51], we can state that all of these methodologies were
suggested to address issues resulting from the current semantic gap between the oper-
ating environment and the software system. Each of these organizational methods
enables a multi-agent structure to be defined by the positions, relationships, and
organizations of the agents. They consider the multi-agent as a virtual structure inde-
pendent of human organizations [4]. However, Knowledge is a semantic description
of a collection of human-generated facts in a given context. Therefore, although the
information sent to and between agents is well defined, the whole knowledge lifecycle
is not taken into consideration most of the time. In addition, there is no methodology
capable of designing a system with different granularities, which is the particularity
SmartDFRelevance: A Holonic Agent Based System … 111

Table 1 Classification of the existing methodologies

Categories Description Methodologies References

Rules-based organizational – The design of the system is based on Engineering for Software Agents [45]
methodologies social structures of agents and eliciting (INGENIAS)
behavioral rules between agents [35]
Model of Organization for multi-agent [46]
– Organizations of agents are known to
Systems (MOİSE+)
arise from the definition of these
requirements ANEMONA [47]

OPERA [48]

Non-rules-based organizational – For each group of agents, this concept Agent, Group and Role (AGR) [9]
methodologies first defines an organizational structure
Multiagent Systems Engineering [49]
and then suggests common goals
(MaSE)
– Such methods move further into the
classification of relationships of agents TROPOS [50]
after their initial stage

of the HMAS. Thus, to deal with such issues, Cossentino et al. proposed an Agent-
oriented Software Process for Complex Engineering Systems (ASPECS) method-
ology with the intrinsic ability to consider the combination of holonic and classical
agent-oriented approaches with knowledge-based approaches (ontology) [17]. More
specifically, ASPECS is a systematic framework that specifies the key principles for
the proposed study, design and development of MAS and HMAS [52]. The ASPECS
lifecycle includes three phases that are: (1) system requirements, (2) agent society
design and (3) implementation and deployment [17]. In the next section, the archi-
tecture of the proposed knowledge-based system following ASPECS methodology
is presented.

5 Architecture of the SmartDFRelevance

Knowledge-Based System

In this section, we detail the design process of the proposed knowledge-based system
SmartDFRelevance. The description begins with a section detailing the activities of
System Requirements Analysis process and then continues with another section
explaining the final design of the proposed SmartDFRelevance Holonic multi-agent
(Agent Society Design phase). We notice that the purpose of the system requirements
analysis is to provide a complete description of the problem based on the principles
outlined in the meta-model problem domain. While the design phase of the Agent
Society aims to design a society of agents whose global behavior can provide an
effective solution to the issue addressed in this paper.
112 A. C. Benabdellah et al.

5.1 System Requirements Analysis Phase

of SmartDFRelevance

The system Requirements analysis phase begins with a Domain Requirements

Description (DRD) activity where requirements are described using traditional
methods such as use cases. In our context, SmartDFRelevance is a knowledge
management system whose main purpose is to handle multi-source information
during the product development process while considering different virtues repre-
sented by Design for X paradigm such as Design for Assembly and Manufacture,
Design for Service, Design for safety, etc. [51]. Five primary processes are illustrated
in this sense (Fig. 3):
• Knowledge Identification and extraction
• Knowledge Validation
• Knowledge Capitalization
• Knowledge Diffusion
• Knowledge Reuse.
Thus, in order to analyze these goals, we chose the i* framework [53] to represent
the goal-oriented analysis of our knowledge management system. In this sense and
according to Fig. 3, our knowledge management system (SmartDFRelevance) is
realized by the contribution of 5 Soft Goals. Besides, these objectives are linked

Fig. 3 Domain requirements description of the SmartDFRelevance knowledge management system

SmartDFRelevance: A Holonic Agent Based System … 113

by the “And” decomposition because it is by their combination that the knowledge

management system can tend towards the overall objective.
However, to provide a conceptual overview of the SmartDFRelevance knowledge
management system, we will describe the problem ontology (Problem Ontology
Description (POS)), which is the second main activity of the first phase of ASPECS
needs analysis. Indeed, this activity deepens the understanding of the problem with
a description of the concepts that make up the problem domain. To do this, the
ontology domain is represented using a class diagram where the concepts, attributes,
and actions are identified by specific stereotypes [54].
The UML diagram in Fig. 4 shows our proposed ontology related to the area of
interest. More clearly, the system describes systematically the activities to follow for
the development of an industrial product, from its beginning of life (feasibility) to
its end of life (recycling or rejection). The product development process is made of
different phases. A phase consists of a set of different activities. Each activity can
itself be composed of sub-activities and carried out by professional actors. At the end
of an activity, a deliverable is created. As for the phase, it delivers a state of the product
such as a concept or a prototype. Each phase is characterized by its requirements,
control parameters, resources, current state, and progress. The phases and activities
are then modeled by organizations, which are composed of different roles. Each
organization is then broken down into several sub-organizations, which follow a
process carried out by several or one activities. Each role represents the behavior of

Fig. 4 Problem Ontology Description (POD) for SmartDFRelevance system

114 A. C. Benabdellah et al.

actors and makes it possible to determine their behavior during the activity through
his interactions, skills, and knowledge. Thus, a role has one or more skills and these
require a series of knowledge to characterize them. In addition, an actor can play
several roles in depends on his competences and capacities. Therefore, the state of the
product is contained in the knowledge associated with the interactions’ results shared
during the interaction between roles. This knowledge is then structured according
to MemoDFRelevance project memory and stored and conceptualized according to
OntoDFRelevance ontology.
The identification of organizations corresponds to the third activity in the System
Requirements analysis phase, which is “Organization Identification (OID)” in
ASPECS. Each goal/plan is assigned to an organization that is responsible for
accomplishing it. Each organization is then defined by a set of roles, their inter-
actions and a common context. According to Fig. 5, the overall organization, called
SmartDFRelevance is broken down into five sub-organizations.
In this figure, the “Boundary Roles” represent either humans interacting with
the system, such as professional actors or experts, or the OntoDFRelevance and
MemoDFRelevance components used by the HMAS for knowledge management.
The professional actor is monitored by the detector role. This role monitors the
activities of the user by detecting the name of the project he is working on, his role
in this project, the current activity, and possibly the term in the search field. The
skills and knowledge used during the product development process are presented
in the MemoDFRelevance project memory that constitutes a “Boundary role”. The
expert “Boundary role” represents the human experts in interaction with the system
who are responsible for validating the knowledge capitalized in the project memory
MemoDFRelevance and conceptualized from the ontology OntoDFRelevance. All

Fig. 5 Organization Identification (OID) of the SmartDFRelevance system

SmartDFRelevance: A Holonic Agent Based System … 115

these components interact with the five organizations “Knowledge identification”,

“Knowledge validation”, “Knowledge capitalization”, “Knowledge dissemination”
and “Knowledge reuse”.
The first organization, Identification, aims to meet the goal of defining the knowl-
edge that will be extracted for reuse. It mainly includes 4 sub-organizations, namely:
identification, extraction and annotation, classification and sharing. The second orga-
nization, Validation, aims to satisfy the aim of validating the knowledge resulting
from the “Identification” organization or added by the project agents. It includes 3
sub-organizations, namely: sharing, validation and evaluation. The third organiza-
tion, Capitalization, aims to annotate, exploit and formalize the knowledge validated
by the Validation organization. It includes 3 sub-organizations, namely: Storage,
operation and annotation. The fourth organization, Diffusion, aims to share and
disseminate knowledge among systems’ agents. It also includes 3 sub-organizations,
namely: organization, updating and dissemination. Finally, the fifth organization,
Reutilization, aims to reuse the knowledge stored in project memory or Metier
repository.
However, after the identification and the organizational decomposition of the
knowledge management system “SmartDFRelevance”, we decompose each global
behavior into several sub-roles to better understand the operation. The identification
of interactions and roles corresponds to the fourth activity “Interactions and Role
Identification (IRI)” in ASPECS. The finest behavior is represented by a Role. Role
interaction must be defined in the same organization that provides the context for
interaction. The goal of each Role must then contribute to the accomplishment (of a
part) of the goal of the organization to which it belongs. Thus, to support this decom-
position we use the strategic reasoning diagrams of the i *framework to decompose
each non-elementary goal into sub-goals or plans [55].
For example for the Reutilization organization, Fig. 6 shows its strategic reasoning
diagram. For this soft goal to be achieved, at least one of the four plans, namely,
carrying out the PUSH/Pull mechanism, carrying out the automatic assistance,
carrying out the alert system or carrying out the transfer must be executed. These
four plans differ in their usefulness for agents according to the different stages of
their integration process. It should be noted that PUSH activity in our context refers
to the fact that the system tracks user activities, detects their role, the name of the
product as well as its activity and automatically proposes the stored knowledge corre-
sponding to the role of the user [56]. On the other hand, the Pull activity offers a
personalized search for knowledge, the actor types a term in the search field and he
obtains the knowledge related to this term filtered according to his role and activity
in the OntoDFRelevance model [56]. For a beginner, for example, it is interesting
to automatically offer him the knowledge that allows him to learn the basics of his
work and extract knowledge from the project memory (MemoDFRelevance). For a
more experienced actor, an alert system, helping to avoid errors, allows him to carry
out projects with less effort and in a shorter time. Thus for the achievement of the
objective of the Reutilization organization, four main roles are involved (represented
by circles in Fig. 6):
116 A. C. Benabdellah et al.

Fig. 6 Interactions and Role Identification (IRI) for reutilization organization of the SmartDFRel-
evance system

• Context Agent detects the organizational context linked to the actor, namely his
role, his activity, the name of the project and, possibly, the term sought.
• Reasoning Project and Metier Agents responsible for extracting knowledge
from project memory and Metier repository.
• Reuse Agents responsible for formulating requests and proposing results and
suggestions.
• Assistant Agent in charge of identifying and transferring reused knowledge to
agents and actors.
As a summary, at the end of the system requirements phase, the scope of the
knowledge-based system to be developed, as well as the organizational hierarchy that
composes it, are now identified and partially specified. At each level of this hierarchy,
the organizations and the roles and interactions that compose it are described. The
first specification of the behavior of the roles has already been made. The goal now
is to design a society of agents whose overall behavior is capable of providing an
effective solution to the complex problem described in the system requirements
analysis phase.

5.2 Agent Society Design Phase of SmartDFRelevance

At this stage, the problem has been formed in terms of organizations, roles, capa-
bilities and interactions. The goal of the second phase in ASPECS methodology
is to develop a system of agents whose global activity is capable of providing an
SmartDFRelevance: A Holonic Agent Based System … 117

appropriate solution to the problem identified in the previous phase and satisfying
the related specifications. In other words, this agent society design phase results in
a description of the agent community involved in the solution in terms of social
relations and interdependencies between individuals (Holons and/or Agents).
The first activity in this phase is Solution Ontology Description (SOD) in which
the ontology of the problem described during the first phase is refined by adding
new concepts relevant to the agent-based solution and refining existing ones. This
ontology’s definitions, predicates and actions are now also meant to be used to define
knowledge shared between positions in communication. This means specifying all
the predicates used to communicate information in correspondence as well as the
acts that Holons/Agents will take and influence the position of the environment in
which they exist (as a logical reflection of ontology). Therefore, with respect to
the proposed knowledge management system (SmartDFRelevance), ontology (see
Fig. 4) has been filled with function principles to fully support the process of knowl-
edge management by adding actor, holon, sub-holons, super-holons, group, etc. At
the end of this activity, the ASPECS development process is divided into two devel-
opment sub-branches. The first and foremost being dedicated to the system’s orga-
nizational design, the second being dedicated to the identification and design of the
system’s component agents.
The second step is to classify agents (Agent Identification (AI)) that will make
up the program hierarchy’s lowest level and their responsibilities in terms of roles
implemented. Their goals must, therefore, correspond to the goals of these roles. We
take the roles described in the previous models and define the types of agents (cogni-
tive or reactive) that interpret these roles. Moreover, as in the process of knowledge
management, two types of knowledge are distinguished (Project Knowledge and
Metier Knowledge), the HMAS must take into account the actors who will consider
both project and Metier knowledge. To do this, the actors in our SmartDFRelevance
system form three communities (Fig. 7) [44]:
• Metier Agents (AM) are working on the positions of Metier agents. By explaining
the role of their character and their place in the process of product development,
they will recreate the company. Therefore, a Metier agent wants the informa-
tion conveyed from the position and expertise of his Metier actor. This cate-
gory includes cognitive agents (Process Engineers Agents, Designers Engineers
Agents, Project Manager Agent, etc.) as well as the context agent and the assistant
agent.
• The Project Knowledge Management Agents (AMCP) aim to store knowledge
using project memory MemoDFRelevance. In order to ensure their reliability, they
must also validate knowledge. These agents’ third objective is to manipulate,
disseminate and re-use this knowledge in order to propose solutions that will
be communicated to Metier agents. This category is composed of the project
Identification agent, project Capitalization agent, project Validation agent, project
Sharing agent, project Reutilization agent and project Reasoning agent.
• The Metier Knowledge Management Agents (AMCM) plan to store knowledge
from all projects in a knowledge base called Metier Repository. Unlike AMCPs,
118 A. C. Benabdellah et al.

Fig. 7 Agent ıdentification for SmartDFRelevance system

they also need to ensure durability by encouraging Metier actors to test them. This
category is composed of the Metier Identification agent, Metier Capitalization
agent, Metier Validation agent, Metier Sharing agent, Metier Reutilization agent
and Metier Reasoning agent.
After having described the distribution of roles between the agents, it is advisable
to end the analysis by describing the interactions between the agents and their actions.
Thus, the functionality of the Metier Agents is to detect the role, activity and project
linked to the user as well as the term sought via the role of the context agent. The
AM sends a message containing this information to the identification agent so that he
can deduce the associated knowledge. After having identified candidate knowledge
among the six types presented in the OntoDFRelevance, the latter transmits a message
to the capitalization agent who annotates the knowledge. The latter then communicate
the knowledge to the AMCP agents. The AMCPs build the project memory via
the role of the capitalization agent after having obtained the knowledge validation
agreement from the validation agent. During the validation activity, the Metier agents
offer assistance from their Metier agents by providing them with the validation criteria
as well as the knowledge already capitalized in MemoDFRelevance. In this context,
the agents responsible for project reasoning consult the OntoDFRelevance ontology
to formulate the requests that they communicate to the sharing agents. These respond
to requests to disseminate the knowledge capitalized during the project.
Concerning the communication between the AMCPs and AMCMs, the AMCPs
archive knowledge after their validation with Metier agents. When the project is
finished, they transmit this knowledge at AMCM. These build the Metier repository
containing the knowledge capitalized during all the projects. The Metier repository
evolves as knowledge is capitalized during projects.
SmartDFRelevance: A Holonic Agent Based System … 119

Finally, concerning the communication between AMs and AMCMs, AMs formu-
late requests based on the OntoDFRelevance ontology. Knowing the specialties of
each AMCM, they can draw up the right request to the right Metier Agent who
in this case represents a professional actor. AMCM responds to requests using the
knowledge capitalized by the Metier capitalization agent and validated by the Metier
validation agent in the Metier repository. In most cases, the Metier validation agent
requests assistance from AMCP.
After explaining the interaction between systems’ agents, we are now engaged
in the “Holarchy Design (HD)” activity to finish and complete the design phase
of the agent society. This seeks to provide an overarching overview that integrates
and summarizes all the previous activities and concepts in a single graph explaining
the overall structure of the system and the rules governing its complex use. More
clearly, the holonic architecture of the SmartDFRelevance knowledge-based system
is illustrated in Fig. 8.
Notwithstanding, the Implementation and Deployment phase that aims to incorpo-
rate and deploy the holonic agent-oriented approach developed in the previous phase
by applying Janus facilities enable the definition of the system architecture and the
development of related source code and evaluation. Therefore, SmartDFRelevance
agents support the knowledge management process that is a complex process by
using an ontology that allows them to annotate and infer knowledge in order to find
the appropriate knowledge that human actors need. SmartDFRelevance agents also
share and propose the proactive reuse of knowledge not only to one actor, but also
to all the actors who are likely to be concerned about knowledge. Consequently, this
feature provides the intelligence for these agents and makes the proposed knowledge-
based system (SmartFDFRelevance) as a complete smart holonic system, flexible and
autonomous.

6 Conclusion

In this paper, we have proposed a new holonic knowledge-based system, called

SmartDFRelevance, following the ASPECS methodology dedicated to the analysis,
design, and deployment of complex engineered systems. More clearly, the obtained
results are:
• A definition of the needs of a knowledge management system called SmartD-
FRelevance in the form of a goal tree. The goal-oriented approach is chosen in
order to model the objectives of the system as well as the actors involved and their
dependencies for the achievement of each goal;
• A domain ontology which conceptualizes the knowledge capitalized in our project
memory MemoDFRelevance;
• A set of organizations composed of interacting roles aimed at satisfying the goals
of the system;
120 A. C. Benabdellah et al.

Fig. 8 Holarchy design for SmartDFRelevance knowledge-based system

• A set of plan roles for the identification of roles and interactions which aim to break
down a global behavior embodied by an organization into sub-organizations;
• A set of abilities to define the generic behavior of roles by identifying the know-
how that a role requires of the individual who will perform it;
• A list of agents and their responsibilities in terms of roles implemented.
• A holarchy of the SmartDFRelevance system where the results of all the previous
work are combined and summarized in a single graph.
SmartDFRelevance: A Holonic Agent Based System … 121

As a summary, the proposed SmartDFRelevance knowledge-based system assists

the business actor throughout the product development process. It offers assistance
to the business actor by identifying, validating, capitalizing, sharing and reusing
knowledge while considering an organizational approach. We notice that our design
was mainly based on the work of Monticolo [57] and BenMiled [56]. Monticolo had
proposed a knowledge capitalization approach following an organizational approach
[57]. In contrast, BenMiled extended Monticolo’s work by adding the knowledge
reuse system [56]. Our contribution to this work is twofold. On the one hand, the
analysis process makes it possible to highlight the different components and their
links with a view to achieving the objectives of the system. On the other hand and on
the basis of this analysis, we have proposed mechanisms to assist in the identification
and acquisition of knowledge, the validation of knowledge; the capitalization of
knowledge, the sharing of knowledge as well as the re-use of knowledge.
As a perspective, we recommend first enriching our multi-agent system’s oper-
ational framework and adding new ideas to the paradigm used. The application of
logic processes, rules or methods of learning may be important for the purpose of
product development circumstances. Second, we plan to focus on the application
of formal notations and procedures that have been already used for the analysis of
function actions. In future works, these issues will be discussed to provide a complete
knowledge-based system that is appropriate and can be used by the largest number
of designers in different industries.

References

1. Chaouni Benabdellah, A., Bouhaddou, I., Benghabrit, A.: Holonic multi-agent system
for modeling complexity structures of product development process. In: 2019 4th World
Conference on Complex Systems (WCCS), pp. 1–6. IEEE (2019)
2. Monostori, L., Váncza, J., Annals, S.K.-C.: Agent-based systems for manufacturing. Elsevier
(2006)
3. Abecker, A., Bernardi, A., International, L. van E.-P.: Agent technology for distributed
organizational memories. dfki.uni-kl.de (2003)
4. Girodon, J., Monticolo, D., Bonjour, E., Engineering, M.P.-A.: An organizational approach to
designing an intelligent knowledge-based system: application to the decision-making process
in design projects. Elsevier (2015)
5. Benabdellah, A.C., Bouhaddou, I., Benghabrit, A., Benghabrit, O.: A systematic review of
design for X techniques from 1980 to 2018: concepts, applications, and perspectives. Int. J.
Adv. Manuf. Technol., 1–30 (2019). https://doi.org/10.1007/s00170-019-03418-6
6. Wooldridge, M., Jennings, N.R., Kinny, D.: The Gaia methodology for agent-oriented analysis
and design. Springer (2000)
7. Smith, P.B., Bond, M.H.: Social psychology: Across cultures. Allyn & Bacon (1999)
8. Drogoul, A., Ferber, J.: Multi-agent simulation as a tool for modeling societies: application to
social differentiation in ant colonies. Presented at the (1994)
9. Ferber, J., Weiss, G.: Multi-agent systems: an introduction to distributed artificial intelligence
(1999)
10. Dorri, A., Kanhere, S., Access, R.J.-I.: Multi-agent systems: a survey. ieeexplore.ieee.org
(2018)
122 A. C. Benabdellah et al.

11. García-Magariño, I., Gutiérrez, C.: Agent-oriented modeling and development of a system for
crisis management. Elsevier (2013)
12. Chaouni Benabdellah, A., Bouhaddou, I., Benghabrit, A.: Supply chain challenges with
complex adaptive system perspective (2018)
13. Durand, D.: La systémique (10e éd.). Paris Press. Universitaires de France (2006)
14. Hilaire, V., Koukam, A., Gruer, P., Müller, J.P.: Formal specification and prototyping of multi-
agent systems. Springer (2000)
15. Isern, D., Sánchez, D., Moreno, A.: Organizational structures supported by agent-oriented
methodologies. J. Syst. Softw. 84, 169–184 (2011)
16. Giret, A., Botti, V.: Holons and agents. J. Intell. Manuf. 15, 645–659 (2004). https://doi.org/
10.1023/B:JIMS.0000037714.56201.a3
17. Cossentino, M., Gaud, N., Hilaire, V., Galland, S., Koukam, A.: ASPECS: an agent-oriented
software process for engineering complex systems. Auton. Agent. Multi. Agent. Syst. 20,
260–304 (2010). https://doi.org/10.1007/s10458-009-9099-4
18. Koestler, A., Zeppelins, P.D.-R.: Arthur Koestler. books.google.com (1905)
19. Studer, R., Benjamins, V., Fensel, D.: Knowledge engineering: principles and methods. Elsevier
(1998)
20. Gandon, F.: Distributed artificial intelligence and knowledge management: ontologies and
multi-agent systems for a corporate semantic web (2002)
21. Verhagen, W.J.C., Bermell-Garcia, P., van Dijk, R.E.C., Curran, R.: A critical review of
knowledge-based engineering: an identification of research challenges. Adv. Eng. Informat.
26, 5–15 (2012). https://doi.org/10.1016/j.aei.2011.06.004
22. Argente, E., Julian, V., Botti, V.: Multi-agent system development based on organizations.
Elsevier (2006)
23. Benabdellah, A., Benghabrit, A. and Bouhaddou, I: Complexity drivers in engineering design:
toward a decision support system based on an organizational perspective. J. Eng. Design
Technol. (2020), Vol. ahead-of-print No. ahead-of-print. https://doi.org/10.1108/JEDT-11-
2019-0299
24. Monceyron, E., Barthès, J.-P.A.: Rev. Sci. Tech. la Concept, 49–68 (1992)
25. Scalabrin, E.E., Vandenberghe, L., Azevedo, H., Barthès, J.-P.A.: A generic model of cognitive
agent to develop open systems. Presented at the October 23 (1996)
26. Shen, W., Barthes, J.P.A.: Asn experimental multi-agent environment for engineering design.
Int. J. Coop. Inf. Syst. 05, 131–151 (1996). https://doi.org/10.1142/S0218843096000063
27. Sato, G.Y., de Azevedo, H.J.S., Barthès, J.P.A.: Agent and multi-agent applications to support
distributed communities of practice: a short review. Auton. Agent. Multi. Agent. Syst. 25,
87–129 (2012). https://doi.org/10.1007/s10458-011-9170-9
28. Barthès, J.P.A., Tacla, C.A.: Agent-supported portals and knowledge management in complex
R&D projects. Elsevier (2002)
29. Nonaka, I., Takeuchi, H.: The knowledge-creating company: how Japanese companies create
the dynamics of innovation (1995)
30. Liao, C., Chuang, S.H., To, P.L.: How knowledge management mediates the relationship
between environment and organizational structure. Elsevier (2011)
31. Huang, T.C., Lin, B.H., Yang, T.H.: Herd behavior and idiosyncratic volatility. Elsevier (2015)
32. van Elst, L., Dignum, V., Abecker, A.: Towards agent-mediated knowledge management.
Presented at the (2004)
33. Boissier, O., Demazeau, Y.: ASIC: An architechture for social and individual control and its
application to computer vision. Presented at the (1996)
34. Guizzardi, R.S.S., Aroyo, L., Wagner, G.: Agent-oriented knowledge management in learning
environments: a peer-to-peer helpdesk case study. Presented at the (2004)
35. Carley, K.M.: Computational organization science: anew frontier. Natl Acad Sci (2002)
36. Gandon, F., Berthelot, L., Dieng-Kuntz, R.: International, 6th, 2002, undefined: a multi-agent
platform for a corporate semantic web. hal.inria.fr (2002)
37. Bonifacio, M., Bouquet, P., Traverso, P.: Enabling distributed knowledge management:
MANAGERIAL and technological implications (2002)
SmartDFRelevance: A Holonic Agent Based System … 123

38. Preece, A., Hui, K., Gray, A., Marti, P., Bench-Capon, T., Jones, D., Cui, Z.: The KRAFT archi-
tecture for knowledge fusion and transformation. In: Research and Development in Intelligent
Systems XVI. pp. 23–38. Springer, London (2000)
39. Champin, P., Prié, Y., ICCBR, A.M.: Musette: Modelling uses and tasks for tracing experience.
yannickprie.net (2003)
40. Buccafurri, F., Rosaci, D., Sarnè, G.M.L., Ursino, D.: An agent-based hierarchical clustering
approach for E-commerce environments. Presented at the (2002)
41. Braga, R.M., Werner, C.M., Mattoso, M.: Odyssey-Search: A multi-agent system for compo-
nent information search and retrieval. Elsevier (2006)
42. Bravo, M., Perez, J., Sosa, V., et al.: Ontology support for communicating agents in negotiation
processes. ieeexplore.ieee.org (2005)
43. Amailef, K., Lu, J.: Ontology-supported case-based reasoning approach for intelligent m-
Government emergency response services. Elsevier (2013)
44. Monticolo, D., Mihaita, S., Darwich, H., Hilaire, V.: An agent-based system to build project
memories during engineering projects. Elsevier (2014)
45. Bernon, C., Cossentino, Pavón, J.: Agent-oriented software engineering. cambridge.org (2005)
46. Bordini, R., Hübner, J., Wooldridge, M.: Programming multi-agent systems in AgentSpeak
using Jason (2007)
47. Giret, A., Garcia, E., Botti, V.: An engineering framework for service-oriented intelligent
manufacturing systems. Comput. Ind. 81, 116–127 (2016)
48. Aldewereld, H., Dignum, V.: OperettA: Organization-oriented development environment.
Presented at the (2011)
49. DeLoach, T.L.: Soft-sided beverage cooler (2001)
50. Bresciani, P., Perini, A., Giorgini, P., Giunchiglia, F., Mylopoulos, J.: Tropos: an agent-oriented
software development methodology. Auton. Agent. Multi. Agent. Syst. 8, 203–236 (2004).
https://doi.org/10.1023/B:AGNT.0000018806.20944.ef
51. Benabdellah, A.C., Bouhaddou, I., Benghabrit, A., Benghabrit, O.: Design for relevance
concurrent engineering approach: integration of IATF 16949 requirements and design for X
techniques. Res. Eng. Design (2020). https://doi.org/10.1007/s00163-020-00339-4
52. Mazigh, B., Garoui, M., Koukam, A.: Use of formal languages to consolidate a Holonic MAS
methodology: a specification approach for analysing problem and agency domains. J. Simul.
7, 159–169 (2013). https://doi.org/10.1057/jos.2012.24
53. Conklin, J., Begeman, M. L.: gIBIS: A hypertext tool for exploratory policy discussion. dl.
acm.org (1988)
54. Bellifemine, F., Poggi, A., Rimassa, G.: Developing multi-agent systems with a FIPA-compliant
agent framework. Softw. Pract. Exp. 31, 103–128 (2001). https://doi.org/10.1002/1097-024
X(200102)31:2%3c103::AID-SPE358%3e3.0.CO;2-O
55. Yu, E.S.: Towards modelling and reasoning support for early-phase requirements engineering.
ieeexplore.ieee.orgieeexplore.ieee.org (1997)
56. Miled, A.B.: Vers un système de réutilisation des connaissances en ingénierie de conception
(2011)
57. Monticolo, D., Gabriel, A., Barrios, P.C.: Une approche de conception de systèmes multi-agents
dédiés à la gestion des connaissances. search.proquest.com (2018)
A Cyber-Physical Warehouse
Management System Architecture
in an Industry 4.0 Context

Mariam Moufaddal , Asmaa Benghabrit , and Imane Bouhaddou

Abstract Cyber Physical Systems are used whenever complex physical systems
need to communicate with the digital world to optimize performance and increase
efficiency. They play an increasingly important role in industrial processes and in
the control of production (smart factory), particularly in the context of the Internet
of Things. Cyber-physical systems consist of objects, equipped with electronics and
software. They are connected to each other via the Internet in order to form a simple
network system. This concerns the sensors and components that allow the movement
or control of a mechanism or a system and also actuators linking Cyber-Physical
Systems to the outside world. Sensors allow systems to acquire and process data.
This data is then made available to the various services connected to the network,
which use these actuators to directly impact the measures taken in the real world.
This leads to the fusion of the physical world and virtual space in the Internet of
Things. In this paper, we propose a Cyber-Physical Warehouse Management System
environment and architecture where warehouse operations are linked as a single
platform assuring consumers satisfaction.

Keywords Cyber-physical system · Warehouse management system · Internet of

things

M. Moufaddal (B) · A. Benghabrit · I. Bouhaddou

LM2I Laboratory, ENSAM, Moulay Ismaïl University Meknes, Meknes, Morocco
e-mail: [email protected]
A. Benghabrit
e-mail: [email protected]
I. Bouhaddou
e-mail: [email protected]
A. Benghabrit
LMAID Laboratory, ENSMR, Mohamed V University Rabat, Rabat, Morocco

© Springer Nature Switzerland AG 2021 125

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_9
126 M. Moufaddal et al.

1 Introduction

The current industrial revolution, called Industry 4.0, is based on a strong intercon-
nection of objects and/or actors and also of the real world and the virtual world
via innovations in ICT (Information and Communication Technologies) [1]. This
affects the industrial ecosystem more broadly and will impact the world of logistics,
already described as “Logistics 4.0”. Indeed, according to Kohler and Weisz [2],
the digital revolution is gradually erasing the boundaries between B2B (Business to
Business) and B2C (Business to Customer), by directly interconnecting supply and
demand. The interactive and transparent collaboration offered by the evolutions of the
Internet world eliminates the traditional organization in sectors, branches and trades,
and makes penetrate the customer within the development of the added value, within
a relation of customer-supplier hybrid co-design. In this context, the advantage will
go to those who know how, by associating their customers and offering them new
possibilities, to create unexpected possibilities from complementary businesses.
The technological leaps observed in recent years, notably linked to the develop-
ment of digital communication tools, have led to paradigm shifts which must integrate
this new situation. Industry 4.0 is based on various technological advances, among
which we can mainly cite CPS (Cyber-Physical System), IoT (Internet of Things),
and IoS (Internet of Services) [3, 4]. The IoT represents an intelligent ICT infras-
tructure that allows communication and cooperation in real time between objects
(machines and devices) as well as the connection between the physical and virtual
worlds [5] in a dynamic connected environment, giving access to remote services
(IoS). A CPS is an autonomous on-board system, equipped with sensors to perceive
its environment, capable of acting on physical processes by means of actuators. The
CPSs are connected to each other via digital networks and are able to use remote
services to assist them. These various concepts can globally be considered as exten-
sions of robotics and automation, going in the direction of the digitalization of the
world [6].
CPSs therefore allow instant and continuous interaction between physical and
virtual elements and also with external actors. Objects are becoming more and
more autonomous and systems more and more reconfigurable as access to informa-
tion becomes possible anytime and from anywhere. Added to this, the intelligence
capacity given to objects and systems, or the ability to go and solicit via the commu-
nication capacity of remote services able to provide this intelligence. In this context,
Industry 4.0 sees the factory, and even the entire production chain, as a gigantic
cyber-physical system largely self-regulated by machine-object interactions. Bene-
fits in terms of flexibility and agility, savings in time, quality and costs are expected
[2]. This results in the concept of cyber-physical production systems (CPPS) [7]. As a
corollary to the ongoing transformation of the manufacturing world, the organization
of the logistics sector must also evolve.
In the scope of this paper, we consider a supply chain as a network constituted
of different nodes responsible for Procurement, Manufacturing, Warehousing and
Distribution activities. For each activity, Industry 4.0, through its technologies, has
A Cyber-Physical Warehouse Management System … 127

brought profound changes in the way it is handled and the way it interacts with other
processes and actors. Indeed, warehousing is no exception.
A warehouse is where goods are stored and handled. Receiving and preparation
to delivery are two of the activities that take place at a warehouse. But to be able
to receive goods, warehouse capacity must be checked. This information should be
communicated to the transporter before to know when to deliver the goods to this
warehouse. Similarly, information about the availability of a transporter should be
communicated to the warehouse responsible before preparing any orders.
To tackle this matter, the paper is structured as follows: after introductory state-
ments, Cyber-Physical systems are being defined, the overall architecture will be
presented and their contribution on the remarkable changes in the supply chain
processes will be illustrated. Next, warehouse functions and operations will be high-
lighted as well as what opportunities CPSs can bring. Based on the result of our
analysis, a CPS environment and architecture will be proposed, highlighting how the
digital and physical worlds will be combined for better handling of goods. Concluding
remarks and future research directions will be given at the end of the paper.

2 The Role of CPS and Its Usefulness

Many technologies related to sensors, communication, embedded systems, connected

objects or decision support have been able to achieve a high degree of innovation
and maturity of use, relative to their sector namely the industry 4.0 far technology
Cyber-Physical system (CPS) (Fig. 1). The strong connection of the physical, the
service and the digital world can improve the quality of information required for
planning, optimization and operation of industrial systems [8].
The Internet has transformed how humans interact and communicate with one
another, revolutionized how and where information is accessed, and even changed
how people buy and sell products. Similarly, CPSs will transform how humans
interact with and control the physical world around us [10]. Cyber means computa-
tion, communication, and control that are discrete, switched, and logical; the term

Fig. 1 CPS basics [9]

128 M. Moufaddal et al.

Fig. 2 Human interaction

with systems [12]

physical refers to natural and human-made systems governed by the laws of physics
and operated in a continuous time.
CPSs are embedded systems that use sensing devices to retrieve data and act
on physical processes using actuators. They are connected to each other via digital
networks, they use all the data and services available and benefit from multimodal
human–machine interfaces [2]. Within the factory of the future, also considered as a
smart factory, CPS will enable the communication between humans, machines and
products alike [11]. As they are able to acquire and process data, they can self-control
certain tasks and interact with humans via interfaces [12] as shown in Fig. 2.
CPSs are also autonomous on-board systems, equipped with sensors to perceive
their environment, capable of acting on physical processes by means of actuators
[13]. The CPSs are connected to each other via digital networks and are able to use
remote services to assist them.
The explosion of interest in CPS is first of all correlated with the ability to embark
a real-time processing capacity which allows either a strong interaction with the
user of the CPS, or a total consideration in real time of the environment allowing
intelligent automation of the CPS.
Kohler and Weisz [2] pointed out that as a prerequisite for the implementation of
CPS are considered, the compatibility and interoperability of systems which raises
the question of communication standards between machines. Once connected, the
systems communicate with each other and are able to self-regulate without central
control. The factory is configurable as needed with modules that can be added or
removed using plug and work functions. Finally, a virtual factory model is used to
test the different module configurations, but also to simulate and steer the entire
production process. Similarly, using these CPS functionalities, this approach can be
applied to any industrial system as it will allow gains as follow:
• Interoperability will promote Machine-to-Machine and Human-to-Machine inter-
faces;
• Modularity will enhance atomicity and flexibility;
• Real-time processing capacity will provide instant inventory information, impor-
tant deadlines and expected delays;
A Cyber-Physical Warehouse Management System … 129

• Virtualization will enable an overview of the to-be set up model/framework

allowing any possible optimization requirements;

A CPS generally consists of two main functional components: (a) The advanced
connectivity that ensures real-time data acquisition from the physical world and
information feedback from the cyber space; (b) intelligent data management, compu-
tational and analytics capability that constructs the cyber space [14]. Some technolo-
gies that are closely connected to the CPS are IoT, wireless sensor networks, and
cloud computing. Wireless sensor networks are regarded to be a vital component
of CPS [15]. The Internet technology provides essential approaches to enhancing
the performance of cyber physical systems [16]. A 5C level CPS structure has been
proposed [17]. It defines how people construct a CPS from the initial data acquisition,
then analytics, and to the final value creation:
• C1—At the Connection level, the CPS operates on a Plug & Play network and
uses data sent by a network of sensors;
• C2—At the Conversion level, the CPS knows how to process information and
transcribe it into higher level information;
• C3—At Cyber level, the CPS has knowledge of other environmental CPS and can
interact with them to enrich its own information processing;
• C4—At the Cognition level, the CPS is capable of establishing a diagnosis based
on simulations of its own behavior and a differential analysis of sensor data;
• C5—At the Configuration level, the CPS can adapt itself in the event of a failure,
reconfigure or adjust its parameters autonomously in order to return to nominal
behavior.

However, the full integration of the 5 levels within a CPS is currently only very rarely
achieved and is not always justified depending on the type of application. Indeed,
among the 5C levels, the cognition and configuration levels are the most difficult to
achieve [18].

3 CPS is Important to Supply Chains

Industrial process control systems are widely used to provide autonomous control
over business processes through control loops. As CPS is usually defined as integra-
tion of computation with physical processes, the most direct application of CPS in
Industry 4.0 scenarios is enhanced process control [19]. As supply chains are a set of
business processes, they are ideally placed to integrate CPS as a way of monitoring
and for instant decision-making. CPS is a system that deals with the physical as well
as the informational aspects of processes. Indeed, it can provide broad controls over
complex and large industrial processes through a heterogeneous network architecture
of sensors, actuators, and processors, because CPS can integrate all the mechanisms
to reach and maintain a synchronized state [20].
130 M. Moufaddal et al.

Supply chains are embedded in dynamic environments. Their management is

challenged by the occurrence of perturbations, as well as by complex production and
transport networks [21]. Many research topics related to CPS in Supply Chains
were reviewed [19, 21–29]. Soon after, researchers started to point out specific
processes to take the best out of these embedded systems. In this context, Oks et al.
[30] proposed an application map with the aim of distinguishing various oppor-
tunity areas for applying industrial cyber physical systems of Industry 4.0 within
which integrated supply chain, e-procurements and logistics were identified as one
of the improvements categories. Lee et al. [17] also proposed a Multi-Agent System
(MAS) approach for addressing sustainable supplier evaluation and selection process
to provide a proper communication channel, structured information exchange and
visibility among suppliers and manufacturers.
After procurement related issues, researchers focused on manufacturing opera-
tions by proposing approaches and design of CPSs to tackle the diverse problems they
face. Lee et al. [14] proposed a unified 5-level architecture as a guideline for imple-
mentation of CPS. Liu and Jiang [31] continued in the same perspective and provided
a Cyber-physical System (CPS) architecture adapted to shop floor for achieving the
goals of intelligent manufacturing. The proposed architecture provides a guideline
to construct a CPS system from the hardware interconnection, to the data acquisi-
tion, processing, and visualization, and the final knowledge acquisition and learning.
However, in smart factories, maintenance is still an important aspect to safeguard the
performance of their production. Especially in case of failures of machine compo-
nents diagnosis is a time-consuming task. For these reasons, Schneider et al. [32]
presented an approach for a cyber-physical failure management system, which uses
information from machines such as programmable logic controller or sensor data and
Information Technology (IT) systems to support the diagnosis and repairing process.
Following this, several challenges have been raised up on how to handle flexibility,
optimization, and interoperability in production lifecycle. In this context, Jeon et al.
[33] proposed a model-driven approach by ETRI CPS Modeling Language (ECML)
to solve the challenges. Many other research subjects were studied just to name a
few [7, 16, 33–42].
However, warehousing and distribution processes receive less interest from
researchers. According to the PwC’s Global Industry 4.0 Survey [43], which is the
biggest worldwide survey of its kind with over 2,000 participants from nine major
industrial sectors and 26 countries, 21% of the survey respondents belonged to indus-
trial manufacturing whereas only 9% of the respondents belonged to transportation
and logistics fields as illustrated in Fig. 3.
Indeed, for warehousing and distribution processes, industry 4.0 is no longer a
‘future trend’. It is now at the heart of the industry and research agenda. Companies
should start combining advanced connectivity and advanced automation, Cyber-
Physical systems, cloud computing, sensors, connected capability, computer powered
processes, intelligent algorithms and Internet of Things services to transform their
businesses.
In the rest of the paper, we focus the study on building smart warehouses. To do
so, we need to focus on the technologies towards building smart warehouses [44].
A Cyber-Physical Warehouse Management System … 131

Fig. 3 Industry split of survey respondents [43]

The insight to make traditional warehouses smart lies in using Cyber-Physical

System (CPS) [14]. A CPS can monitor and create a virtual copy of the real-world
industry processes, and thus we can know the status of each industry process, control
and make proper decisions in a real time manner [45]. In other words, CPS can bring
the virtual and physical worlds together to construct an entirely networked world,
where smart objects communicate and interact with each other.

4 Warehouse Operations

Warehouses have always served as a vital hub in the flow of goods within a supply
chain. But in today’s economic climate, they also serve as a key source of competitive
advantage for logistics providers who can deliver fast, cost-efficient, and increasingly
flexible warehousing operations for their customers [46]. A warehouse is a facility
in the supply chain to consolidate products to reduce transportation cost, achieve
economies of scale in manufacturing or in purchasing [47] or provide value added
processes and shorten response time [48].
Indeed, warehousing has been recognized as one of the main operations where
companies can provide tailored services for their customers and gain competitive
advantage [49]. There are various types of warehouses: they can be classified into
production warehouses and distribution centers [50] and by their roles in the supply
chain they can be classified as raw materials warehouses, work-in-process ware-
houses, finished good warehouses, distribution warehouses, fulfillment warehouses,
local warehouses direct to customer demand, and value-added service warehouses
[51].
132 M. Moufaddal et al.

Fig. 4 A typical warehouse processes [47]

Warehouses have been going through various challenges such as—supply chains
are becoming more integrated and shorter, globalized operation, customers are
more demanding and technology changes are occurring rapidly. In order to cope
up with these challenges organizations are adopting innovative approaches such as
Warehouse Management System (WMS) [49].
A WMS primarily aims to control the movement and storage of materials within a
warehouse and provide assistance through the execution of warehousing operations.
A warehouse management system (WMS) is a database driven computer application,
to improve the efficiency of the warehouse by directing cutaways and to maintain
accurate inventory by recording warehouse transactions. The systems also direct and
optimize stock based on real-time information about the status of utilization. It often
utilize Auto ID Data Capture (AIDC) technology, such as barcode scanners, mobile
computers, wireless LANs (Local Area Network) and Radio-frequency identification
(RFID) to efficiently monitor the flow of products [49].
Even though warehouses can serve quite different ends, most share the same
general pattern of material flow. Essentially, typical warehouse processes include
inbound processes and outbound processes [47] (Fig. 4).
• Inbound processes concern the receiving and put-away operations [49].

Receiving may begin with advance notification of the arrival of goods. This allows
the warehouse to schedule receipt and unloading to coordinate efficiently with other
activities within the warehouse. Once the product has arrived, it is unloaded and
possibly staged for put away. It is likely to register its arrival so that ownership
is assumed, payments dispatched, and so that it is known to be available to fulfill
customer demand. Product will be inspected, and any exceptions noted, such as
damage, incorrect counts, wrong descriptions, and so on.
In put away operations, an appropriate storage location must be determined before
considering of moving the product. This is very important because information about
where the product is stored defines how quickly and at what cost it will later be
retrieved for a customer. This requires knowing at all times what storage locations
are available, how large they are, how much weight they can bear, and so on. This
information will subsequently be used to construct efficient pick lists to guide the
order-picking operators in retrieving the product for customers.
A Cyber-Physical Warehouse Management System … 133

Receiving operations and put away operations account respectively for about 10%
and 15% of operating costs in a typical distribution center [51] and IoT technologies
are expected to further reduce this.
• Outbound processes include Order-picking, packing and shipping [48, 52–56].

The outbound processes of the warehouse are initiated by receipt of a customer order,
which may be thought of as a shopping list. Each entry on the list is referred to as an
order-line and typically consists of the item and quantity requested. The warehouse
management system (WMS) then checks the order against available inventory and
identifies any shortages. In addition, the WMS may reorganize the list to match
the layout and operations of the warehouse for greater efficiency. Pick-lines are
instructions to the order-pickers, telling them where and what to pick and in what
quantity and units of measure. Each pick-line represents a location to be visited,
which get organized into pick lists to achieve still more efficiencies.
Packing can be labor-intensive because each piece of a customer order must be
handled; but there is little walking. And because each piece will be handled, this is
a convenient time to check that the customer order is complete and accurate. Order
accuracy is a key measure of service to the customer, which is, in turn, that on which
most businesses compete.
Shipping generally handles larger units than picking, because packing has consol-
idated the items into fewer containers (cases, pallets). Consequently, there is still less
labor here. There may be some walking if product is staged before being loaded into
freight carriers. Product is likely to be staged if it must be loaded in reverse order
of delivery or if shipping long distances, when one must work hard to completely
fill each trailer. Staging freight creates more work because staged freight must be
double handled.

5 The Shift to CPS for Autonomous and Smart Warehouses

In a traditional warehouse, the operations of pickup, delivery, and accounting

are accomplished by storekeepers. There are several drawbacks of conventional
warehouses. First, it is time-consuming to store/fetch inventories into/from them.
Second, the usage of storekeepers is a waste of human resources. Finally, it is not
environmentally friendly to record the stocks using account books.
The adoption of the Industry 4.0 paradigm will introduce remarkable changes in
the way warehouse works these days. Especially, the introduction of ‘smart’ manage-
ment throughout the proper adoption and implementation of Warehouse Manage-
ment Systems (WMS) which will transform the warehouse activities into the future
requirements of the inbound logistics according to the Industry 4.0 paradigm [57].
Therefore, it is profoundly essential to make traditional warehouse smart.
However, one of the key paradigms of the Industry 4.0 is the use of modern
information technology (IT). The insight to make traditional warehouses smart lies
134 M. Moufaddal et al.

Fig. 5 The shift from hierarchy to interconnection [59]

in using Industry 4.0 far technology Cyber-Physical System (CPS) [14]. A smart
warehouse is an automated, unmanned, and paperless warehouse when conducting
the operations of pickup, delivery, and bookkeeping [45].
Automated warehouse systems play a key role in manufacturing systems and
are usually controlled using hierarchical and centralized control architectures and
conventional automation programming techniques [58]. However, in an entirely inter-
connected world, each field device in a warehouse is able to initiate an exchange of
data to different levels. This interconnectedness will lead to dissolution of classical,
rigid hierarchies of warehouse processes [59] as shown in Fig. 5. In this context,
Basile et al. [58] present preliminary results in developing a flexible, modular and
distributed control architecture for automated warehouse systems using Function
Blocks and a CPS perspective.
Although CPS has the potential to bring revolution to traditional warehouses, it
is not easy to integrate CPS techniques into smart warehouses for Industry 4.0. In
this context, Liu et al. [45] discussed how the state-of-the-art techniques in cyber-
physical systems has facilitated building smart warehouses to achieve the promising
vision of industry 4.0. They stated that a CPS-based smart warehouse contains four
main components: CPS devices, inventories, robots, and human beings.
For CPS devices, there can be thousands of CPS devices working together. Thus,
the communication among the CPS devices needs to be scheduled efficiently. For
inventories, it is imperative to timely know their status and location context according
to the data reported by the attached devices. For robots, they are required to collab-
oratively accomplish some tasks that are repetitive and harmful to human beings.
For human beings, their gestures or other activities to perform warehouse operations
need to be recognized. The gestures and other activities may be extracted using the
CPS devices deployed in the environment.
Therefore, they focused on four significant issues when applying CPS techniques
in smart warehouses. First, efficient CPS data collection: when limited communi-
cation bandwidth meets numerous CPS devices, more efforts need to be engaged
to study efficient wireless communication scheduling strategies. Second, accurate
A Cyber-Physical Warehouse Management System … 135

and robust localization: localization is the basis for many fundamental operations in
smart warehouses, but still needs to be improved from various aspects like accuracy
and robustness. Third, multi-robot collaboration: smart robots will take the place
of humans to accomplish most tasks particularly in a harsh environment, and smart
and fully distributed robot collaborating algorithms should be investigated. Fourth,
human activity recognition: it can be applied in human–computer interaction for
remote machine operations.
To be more specific, Grzeszick et al. [53] stated that methods like human activity
recognition (HAR) became of increasing interest in industrial settings. In this context,
they introduced a novel deep neural network architecture for HAR. A convolutional
neural network (CNN), which employs temporal convolutions, was applied to the
sequential data of multiple intertial measurement units (IMUs). The network was
designed to separately handle different sensor values and IMUs, joining the infor-
mation step-by-step within the architecture. For illustration purposes, an evaluation
was performed using data from the order picking process recorded in two different
warehouses. The influence of different design choices in the network architecture, as
well as pre- and post-processing, was evaluated. Finally, crucial steps for learning a
good classification network for the task of HAR in a complex industrial setting were
shown.
Following the same CPS perspective, Basile et al. [58] adopted a service oriented
multi-agent approach to obtain, from a cyber-physical system perspective, a formal
model of a complex and real automated warehouse system. The proposed model is
based on Colored Modified Hybrid Petri Net and aims at representing both cyber and
physical aspects with high fidelity [58]. The objective here is modeling a general real
warehouse, from a cyber physical perspective, as a system of cooperating agents.
With this purpose, a Colored Modified Hybrid Petri Net (CMHPN) model, made up
of two interacting sub-nets, one for the cyber and one for the physical part of each
agent, was proposed.
The need for more flexible, adaptable and customer-oriented warehouse opera-
tions has been increasingly identified as an important issue by today’s warehouse
companies. This is due to the rapidly changing preferences of the customers that
use their services. Furthermore, logistics within manufacturing sites like warehouse
and shop floors are rationalized by RFID so that materials’ movements could be
real-time visualized and tracked [60]. This gives rise to product intelligence concept.
Motivated by manufacturing and other logistics operations, Giannikas et al. [61]
argued on the potential application of product intelligence in warehouse operations.
It was illustrated as an approach that can help warehouse companies, that manage
a high variety of products and a big number of individual customers, address the
related issues. This is particularly true in third-party logistics warehouses. Here, the
operations are required to become more customer-oriented and more responsive to
requests with different characteristics and needs in an efficient manner. Finally, they
discussed the opportunities of such an approach using a real example of a third-
party logistics warehouse company and presented the benefits it can bring in their
warehouse management systems.
136 M. Moufaddal et al.

6 A Proposed CPS Framework for Warehouse Operations

Handling

In a supply chain, warehousing function is very critical as it acts as a node in linking

the material flows between the supplier and customer. In today’s competitive market
environment, companies are continuously forced to improve their warehousing oper-
ations. Many companies have also customized their value proposition to increase their
customer service levels, which has led to changes in the role of warehouses.
The use of information systems for warehouse management is studied extensively
in literature. Complexity of warehouse management is indicated among others by
amount and heterogeneity of handled products, the extent of overlap between them,
amount and type of technology as well as characteristics of associated processes.
As the complexity increases it becomes necessary to use Warehouse Management
Systems (WMS) for handling warehouse resources and to monitor warehouse oper-
ations. The warehouses with high amount of processed order lines and amount of
stock keeping units will be best supported by customized software. It is laborious
to update day by day operations of stock status, areas of forklifts and stock keeping
units (SKUs) in real-time by utilizing the bar-code-based or manual-based warehouse
management systems [62].

6.1 The Cyber-Physical Warehouse Management System

(CPWMS) Environment

Innovations that occur within industry 4.0 create new technologies that facilitate
system management. The Cyber-Physical System (CPS) can track and create a virtual
copy of the actual process that can be used to monitor process performance [63]. In
addition, the CPS allows all components within the system to communicate with one
another i.e. that the physical components are virtually connected, allowing for cost
savings in parallel with increasing efficiency [39]. Warehouses based on this tech-
nology include RFID sensors, Bluetooth technology, Wi-Fi access points, cameras
and robots that are coordinated in the system to perform a defined task. The role of a
man in such a warehouse is primarily related to monitoring and reprogramming the
system if necessary [45].
However, the CPS technology also enables inter-machine co-operation between
robotic systems, thus reducing the need for human work. Such designed robotic
systems also enable human movement to recognize what enables employees to use
robotic help in carrying out activities [64]. By replacing human labor with auto-
mated and robotized systems, or through the implementation of such systems as
humanitarian aid, the efficiency of the warehouse system increases. Activities that
are dangerous for a man can be robotized, thus reducing the risk of injury and unlucky
cases [63].
A Cyber-Physical Warehouse Management System … 137

The combination of human and artificial resources has not been part of mainstream
automation practice where (1) robots and humans are generally kept away from each
other, and (2) humans must adhere to work procedures as rigid as the rest of the
automated warehousing environment [65].
Indeed, on one hand, robots exhibit high precision and repeatability, can handle
heavy loads and operate without performance deterioration even in difficult or
dangerous environments. However, robot control systems quickly reach their limits
in recognizing and handling unexpected situations, as reflected by the relatively rigid
plans and robot programs widespread in today’s automated systems [65]. Humans
on the other hand, tackle unexpected situations better, are aware of a much larger
part of the environment than formally declared and show more dexterity in complex
or sensitive tasks. Humans, however, are more prone to error, stress or fatigue [66],
and their employment underlies strict health and safety regulations. For this reason,
human recognition activity is an important issue among others when building a CPS
for warehouse operation [45, 53, 67, 68].
To cope with all these challenges and issues, we are proposing a Cyber-Physical
Warehouse Management System (CPWMS) where a smart autonomous warehouse
sees the day. It is indeed a WMS but added to it a layer of intelligence. The system
will be controlling the movement and storage of materials within a warehouse. It will
be able to collect, provide and analyze in real-time information about each product
in terms of who the supplier is, where it is located in the warehouse, what is its
destination and when it should be executed thanks to cloud computing and big data
technologies. The CPWMS will be able to plan resources and activities to synchronize
the flow of goods in the warehouse. It will be able to track each resource (operators,
forklifts, etc.) in the warehouse thanks to connected devices (wearables, etc.) and
the internet of things. The system also will be offering additional functionality like
transportation, dock door, and value-added logistics planning which help to optimize
the warehouse operations as a whole. A typical smart warehouse [46] is shown in
Fig. 6.
In addition, imagine a future where IT systems are not created by computer
analysts speaking the languages of Java and C but instead by business managers
speaking the languages of supply chain, customer service or product development. It
is a future made possible by Service Oriented Architecture (SOA)—an evolution in
the way enterprise IT systems can be built [69]. With an SOA, business applications
are constructed of independent, reusable, interoperable services that can be reconfig-
ured without vast amounts of technical labor. The fundamental building blocks of an
SOA are web services. An SOA is a collection of web services brought together to
accomplish business tasks (checking a product availability or generating an invoice)
and it is a way of designing and building a set of Information Technology applica-
tions where application components and Web Services make their functions available
on the same access channel for mutual use [70]. Because the services can interact
with systems outside a single organization, they provide the ability for companies to
collaborate with customers and suppliers [69].
In the light of these concepts, our CPWMS will be supported by the following:
138 M. Moufaddal et al.

Fig. 6 A typical Smart warehouse [46]

• The Internet of Things (IoT) is the concept that refers to the connection of any
device with an intelligent network through the use of the Internet. IoT has affected
the way in which cyber-physical systems (CPS) interact, are controlled, managed
and monitored. It comprises communicating smart systems using IP addresses.
This enables each physical object being equipped with a unique IP address [16].
• The Internet of Services (IoS) comprises new service paradigms such as being
provided by the service-oriented architecture (SOA) or the REST-technology. It
highlights the fact of reusing and composing existing resources and services [70].
• The Internet of Data (IoD) enables to store and transfer mass data appropriately,
and to provide new and innovative analysis methods for interpreting mass data.

In this new industrial model, the work environments will be fully automated thanks
to the appropriate interaction between IoT and CPS, while the data will now be
processed online thanks to the use of Cloud Computing. This concept and its inno-
vative level of organization and control of the entire value-added system, which
considers the complete cycle of products with the ultimate goal of guaranteeing the
supply of the needs of consumers, is made up of many areas called smart and some
industry 4.0 technologies as illustrated in Fig. 7.
A Cyber-Physical Warehouse Management System … 139

Fig. 7 The CPWMS

environment
Big Data
Cloud technologies
compuƟng /analyƟcs

Smart warehouse Smart

CPS products
Smart
vehicles
/robots
Augmented
reality

6.2 The Cyber-Physical Warehouse Management System

Architecture

Before talking about any CPS warehouse architecture, it is necessary to evaluate and
measure the degree of digitalization of the concerned warehouse. In this context,
Tao and Zhang [5] defined four stages to illustrate the process as shown in Fig. 8.
At the first one, due to the lack of effective information means, warehouses depend
on physical space completely, leading to low efficiency, accuracy and transparency.
Then with the developments of information technologies, computer aided systems
begin to be applied in warehousing activities, but as the interaction methods are
weak, virtual space is out of step of the physical one. At the third stage, benefited

Fig. 8 Stages of physical and virtual worlds convergence [5]

140 M. Moufaddal et al.

Cyber layer Physical layer

Client tier
Human-machine interface

Send order/service request

Action and execute tier

Actuator unit

Locate IT

Security assessment tier

Receive commands Perform commands
vulnerabilities &
risks
Service and decision-making tier
API Task analysis Task schedule

Decision-making Services interaction Data security

Access security
Data processing tier Data storage Send data to action tier

Heterogeneous Generate control Device security

data processing commands

Sensor and capture tier Sensor unit Receive/send data

Preliminary data Periodicity/continuous

pre-processing data retrieval

Fig. 9 A CPWMS architecture

from communication technologies, sensors, IoT, etc., interaction between the two
spaces exists. In the future, with the continuous developments of new information
technologies, virtual space will gradually play the equally important role with the
physical one and the two-way connection will also be enhanced, which supports the
further convergence.
In an era where virtual and physical spaces begin to further interact and converge
and after the investigation of available Cyber-Physical Systems for warehouses, we
are proposing a six tiers CPS architecture as follows and as shown in Fig. 9.
Sensor and capture tier: Sensor tier is the data source for the above tiers. At
this level, no enhanced processing and analysis capabilities are evoked. Instead,
their main task is to control the sensors which are connected to the tier, ask the
value of one or more attributes periodically and/or continuously as well as to send
and receive messages from and to the data processing tier and perform preliminary
pre-processing. The functions of this tier are environment awareness which mainly
achieved by sensors WSN (Wireless Sensor Network) is one of the basic techniques
of this sensor tier.
Data processing tier: The huge amount of data is not enough, what really makes
it useful is the way it can improve decision-making. This is when smart connections
become relevant, the information must be significant, and displayed at the right time
in the right place. Internet of things objects generate data. Cloud allows to have
access to this data/information from anywhere in a simplified way while Big data
analytics get value out of it. In this perspective, this tier provides resources for the
execution of control algorithms. It consists of the computational devices and storage
A Cyber-Physical Warehouse Management System … 141

devices, providing the heterogeneous data processing such as normalization, noise

reduction, data storage and other similar functions. It takes requests from the sensors
in the sensor/capture tier and generate new control commands from these inputs,
which they send back to the execute tier.
Service and decision-making tier: This tier provides the typical functions of the
whole system, including the APIs to clients, task scheduling, decision-making, task
analysis and so on. In this tier, a number of services are deployed and interact with
each other.
Action and execute tier: At this level, a physical action is asked to be executed
by means actuators. The actuator may be any kind of physical device. Orders and
commands are received and executed.
Client tier: The client tiers serve as a human-machine interface. It can be a tablet,
a computer, a smartphone or a web browser.
Security assessment tier: Security assessment tier is an explicit study to locate IT
security vulnerabilities and risks along the rest of the tiers. It covers devices security,
data security, and access security.
The overall architecture is able to be divided into a number of small indepen-
dent services which respond to specific functionalities. Each tier operates in its own
process and communicates with the other tiers via light mechanisms (service tier or
any other middleware system). As warehouses are considered as high load systems,
this architecture is better suited. The architecture has the advantage of fully preserving
the nominal performance of a warehouse management system because the nominal
handling and control loops are kept within the system.
This architecture is considered as the information system (IS) of a warehouse
which is most often made up of a multi-level architecture, in which each level has
a well-defined role. The highest level in the hierarchy is Warehouse Management
System. This system manages more commercial aspects such as the processing of
customer orders, the allocation of stock, etc. It allows editions such as load slips,
delivery notes and invoices, based on order preparation and shipping information. It
is not necessarily connected in real time to handling equipment, see not connected
at all.
However, here comes the role of the cyber-physical interaction. It is able to coor-
dinate the different systems of material handling equipment. It would direct real-
time data exchanges with handling equipment and also provides a unified human–
machine interface (HMI) for monitoring, control and diagnostics. As a keystone in
the operational management of handling equipment, the Cyber-Physical Warehouse
Management System is the essential link between a simple WMS and the equipment
control systems. It receives information from the sensor tier and coordinates the
various control systems in real time (conveyors, label printers, etc.), by means data
processing tier, to carry out the daily workload. At each decision point, the overall
system determines the most efficient routing and transmits the instructions to the
action and execute tier to perform the expected action.
At the lowest level, we find the sensor tier. It is constituted of Inputs/Outputs (I/O)
devices such as RFID tags or QR codes embedded into products (smart products) and
pallets and allow the physical action of handling equipment and total traceability of
142 M. Moufaddal et al.

packages, according to the given directives. Concretely, a sensor unit only takes care
of one machine (robot palletizer, sorter …) or part of the system (conveyor section,
sorting stations …).
Cyber-physical systems consist of objects, equipped with electronics and soft-
ware. They are connected to each other or via the Internet in order to form a simple
network system. This concerns the sensors and components that allow the movement
or control of a mechanism or system also called an actuator linking CPS to the outside
world. Sensors allow systems to acquire data. They are embedded in products, pallets
or any equipment that needs to be tracked. Then, by choosing Cloud Computing, the
necessary IT infrastructure will be at disposal quickly and efficiently, without making
large investments. It allows also to put the necessary IT resources at disposal, and
easily adapt the capacity, up or down, according to the activity peaks. Cloud helps to
integrate data from numerous sources and uses Big Data Analytics to process data.
This data is then made available to the various services connected to the network,
which use these actuators to directly impact the measures taken in the real world.
This leads to the fusion of the physical world (warehouse with all the operations)
and virtual space (with all the data capture, data processing).
Ultimately, our proposed architecture for a warehouse is similar to/reflects the
organizational structure of human activities. Managers determine the workload to be
accomplished daily, then supervisors manage operators’ activities in real-time. Each
operator performs a specific task based on their expertise (preparation, Pick to Light,
conveying, …). As soon as an operator has completed a task, the supervisor gives a
new task based on the current workload. Once the orders are ready, the supervisor
reports to managers on the progress of the preparations as well as all other relevant
information.
IoT would be of great use in a warehouse managed using our proposed archi-
tecture. Indeed, wireless readers capture information transmitted from each pallet
because it arrives through inbound doors. This information might incorporate data
on the item such as volume and measurements, which could be gathered and sent
to the WMS for handling. This capability disposes of the time-consuming assign-
ment of manual checking and volume checking of pallets. Cameras connected to the
portals seem moreover to be utilized for damage detection, by checking pallets for
defects.
Apart from products put away in a warehouse, IoT can drive ideal resource utiliza-
tion. By interfacing equipments and vehicles, IoT empowers warehouse supervi-
sors to monitor all resources in real time. Supervisors can be cautioned when a
resource is being overused or when an unoccupied resource should be sent to do other
assignments. For illustration, sensors may well be conveyed to track how frequently
resources, such as conveyer belts, are used or unused, and at what frequency. Applying
the different Big data Analysis (BDA) techniques and processing these captured data
might recognize ideal capacity rates and assignments for the resources.
In order to respond to the growing complexity of computing systems, due in partic-
ular to the rapid and permanent progress of information technologies, new paradigms
and architectural solutions based on self-adaptive, self-organized structures are to be
developed. These must allow, on one hand, the provision of sufficient computing
A Cyber-Physical Warehouse Management System … 143

power responding to severe time constraints (real-time processing). On the other

hand, to have great flexibility and adaptability to respond to changes in processing
or unforeseen failures characterizing a context of changing operating environment
of the system.

7 Conclusion and Future Research Directions

As a complex, real-time, embedded and integrated system, CPS integrates various

devices equipped with sensing, identification, processing, communication, and
networking capabilities. It has provided an outstanding foundation to build advanced
industrial systems and applications by integrating innovative functionalities through
Internet of Things (IoT) to enable connection of the physical reality operations with
computing and communication infrastructures [71].
The ongoing transformation of the supply chain area requires an evolution in
the organization of the logistics sector which should allow better pooling of logistics
resources, by interconnecting several logistics actors. This allows better performance,
in terms of effectiveness, efficiency and quality of service.
In this paper, we introduced the concept of Cyber-Physical system as the industry
4.0 revolutionary technology along with its contribution on the remarkable changes
in supply chain processes. Next, we focused the study on warehouse functions and
operations and the opportunities CPS can bring. After investigation of available CPS
architectures in literature, we proposed a Cyber-Physical Warehouse Management
System environment and architecture. Building a smart warehouse requires integra-
tion of different technologies each one is responsible for a part of the work. This
is why we provided a system environment where this CPWMS could evolve and
accomplish the tasks it is supposed to. The warehouse should be equipped with
smart products, smart vehicles and robots, cloud computing to store data gener-
ated from connected devices and big data analytics to process and provide decision
making tools. Augmented reality could also be of great use as it assists operators in
accomplishing their daily tasks efficiently.
Our CPWMS was a 6-tier architecture. The sensing tier where data is perceived
from sensors, the data tier where data is being transferred for processing purposes,
the service tier, the execute tier where actuators are asked to perform commands,
the client tier as the human machine interface and finally the security tier where IT
vulnerabilities and risks are identified. Through this architecture, we tried to present
a smart warehouse architecture where data flows are illustrated.
As future research direction, we propose to develop further this architecture and
build a framework where activities of a warehouse are illustrated, and both physical
and data flows will be highlighted from the order received from the client to his
satisfaction.
Furthermore, although the progress in CPS research, companies are still unable
to fully adopt it as an underlying technology. We propose to study the main reasons
144 M. Moufaddal et al.

faced by companies while adopting CPS and formulate a paradigm shift from central-
ized control systems to decentralized event-triggered control systems that operate at
multiple scales and have the capability to reorganize and reconfigure.
Moreover, different CPS applications might need to collaborate to achieve a
specific mission. For instance, the mobile-health application might need to collabo-
rate with the transportation system to get an ambulance as fast as possible in the case of
patient’s biomedical sensors revealing an emergency. In such scenarios, data analytics
becomes a challenging task as it requires to be able to put the different analysis frag-
ments from different CPS applications together to provide broader conclusions and
decisions making.
Industrial process control systems are widely used to provide autonomous control
over business processes through control loops. As CPS is usually defined as integra-
tion of computation with physical processes, the most direct application of CPS in
Industry 4.0 scenarios is enhanced process control [19]. As supply chains are a set of
business processes, they are ideally placed to integrate CPS as a way of monitoring
and for instant decision-making. CPS can provide broad controls over complex and
large industrial processes through a heterogeneous network architecture of sensors,
actuators, and processors, because CPS can integrate all the mechanisms to reach
and maintain a synchronized state [20]. A typical supply chain contains four business
processes: Procurement, Manufacturing/Quality control and packaging, warehousing
and distribution/return. So as future research direction, we propose to study how in
the era of industry 4.0 and using IoT technologies, Big data is transforming CPS data
into useful knowledge to fulfill warehouse processes in an optimal and a flexible
way.

References

1. El Kadiri, S., Grabot, B., Thoben, K.D., Hribernik, K., Emmanouilidis, C., von Cieminski, G.,
Kiritsis, D.: Current trends on ICT technologies for enterprise information systems. Comput.
Ind. 79, 14–33 (2016). https://doi.org/10.1016/j.compind.2015.06.008
2. Kohler, D., Weisz, J.-D.: Industrie 4.0: quelles stratégies numériques? La numérisation de
l’industrie dans les entreprises du Mittelstand allemand. Kohler Consulting & Coaching (2015)
3. Hermann, M., Pentek, T., Otto, B.: Design principles for Industrie 4.0 scenarios. In: 2016 49th
Hawaii International Conference on System Sciences (HICSS), pp. 3928–3937 (2016)
4. Kagermann, H.: Recommendations for Implementing the Strategic Initiative INDUSTRIE 4.0:
Securing the Future of German Manufacturing Industry; Final Report of the Industrie 4.0
Working Group. Forschungsunion (2013)
5. Tao, F., Zhang, M.: Digital twin shop-floor: a new shop-floor paradigm towards smart manu-
facturing. IEEE Access 5, 20418–20427 (2017). https://doi.org/10.1109/ACCESS.2017.275
6069
6. Noyer, J.-M.: L’Internet des Objets, l’Internet of “Everything”: quelques remarques sur
l’intensification du plissement numérique du monde. IdO 17 (2017). https://doi.org/10.21494/
ISTE.OP.2017.0134
7. Monostori, L.: Cyber-physical production systems: roots, expectations and R&D challenges.
Procedia CIRP 17, 9–13 (2014). https://doi.org/10.1016/j.procir.2014.03.115
A Cyber-Physical Warehouse Management System … 145

8. Landherr, M., Schneider, U., Bauernhansl, T.: The application center Industrie 4.0—industry-
driven manufacturing, research and development. Procedia CIRP 57, 26–31 (2016). https://doi.
org/10.1016/j.procir.2016.11.006
9. Cardin, O.: Contribution à la conception, l’évaluation et l’implémentation de systèmes de
production cyber-physiques. Université de Nantes (2016)
10. Rajkumar, R.: A cyber-physical future. Proc. IEEE 100, 1309–1312 (2012). https://doi.org/10.
1109/JPROC.2012.2189915
11. Damm, W., Achatz, R., Beetz, K., Broy, M., Daembkes, H., Grimm, K., Liggesmeyer, P.:
Nationale roadmap embedded systems. In: Broy, M. (ed.) Cyber-Physical Systems, pp. 67–136.
Springer, Berlin, Heidelberg (2010)
12. Broy, M.: Cyber-physical systems—Wissenschaftliche Herausforderungen Bei Der Entwick-
lung. In: Broy, M. (ed.) Cyber-Physical Systems, pp. 17–31. Springer, Berlin, Heidelberg
(2010)
13. Wittenberg, C.: Human-CPS interaction—requirements and human-machine interaction
methods for the Industry 4.0. IFAC-PapersOnLine 49, 420–425 (2016). https://doi.org/10.1016/
j.ifacol.2016.10.602
14. Lee, J., Bagheri, B., Kao, H.A.: A cyber-physical systems architecture for Industry 4.0-based
manufacturing systems. Manuf. Lett. 3, 18–23 (2015). https://doi.org/10.1016/j.mfglet.2014.
12.001
15. Kos, A., Tomažič, S., Salom, J., Trifunovic, N., Valero, M., Milutinovic, V.: New benchmarking
methodology and programming model for big data processing. Int. J. Distrib. Sens. Netw. 11,
271752 (2015). https://doi.org/10.1155/2015/271752
16. Wang, L., Wang, G.: Big data in cyber-physical systems, digital manufacturing and Industry
4.0. Int. J. Eng. Manuf. 6, 1–8 (2016). https://doi.org/10.5815/ijem.2016.04.01
17. Lee, J., Ardakani, H.D., Yang, S., Bagheri, B.: Industrial big data analytics and cyber-physical
systems for future maintenance & service innovation. Procedia CIRP 38, 3–7 (2015). https://
doi.org/10.1016/j.procir.2015.08.026
18. Kao, H.A., Jin, W., Siegel, D., Lee, J.: A cyber physical interface for automation systems—
methodology and examples. Machines 3, 93–106 (2015). https://doi.org/10.3390/machines3
020093
19. Sanchez, B.B., Alcarria, R., Sanchez-de-Rivera, D.: Enhancing process control in Industry 4.0
scenarios using cyber-physical systems. J. Wirel. Mob. Netw. Ubiquitous Comput. Depend.
Appl. 7, 41–64 (2016)
20. Chen, H.: Applications of cyber-physical system: a literature review. J. Ind. Integr. Manag. 02,
1750012 (2017). https://doi.org/10.1142/S2424862217500129
21. Frazzon, E.M., Silva, L.S., Hurtado, P.A.: Synchronizing and improving supply chains through
the application of cyber-physical systems. IFAC-PapersOnLine 48, 2059–2064 (2015). https://
doi.org/10.1016/j.ifacol.2015.06.392
22. Klötzer, C., Pflaum, A.: Cyber-Physical Systems (CPS) in Supply Chain Management : A
definitional approach (2016)
23. Frazzon, E.M., Hartmann, J., Makuschewitz, T., Scholz-Reiter, B.: Towards socio-cyber-
physical systems in production networks. Procedia CIRP 7, 49–54 (2013). https://doi.org/
10.1016/j.procir.2013.05.009
24. Klötzer, C., Pflaum, A.: Cyber-physical systems as the technical foundation for problem solu-
tions in manufacturing, logistics and supply chain management. In: 2015 5th International
Conference on the Internet of Things (IOT), pp. 12–19 (2015)
25. Tu, M., Lim, M.K., Yang, M.F.: IoT-based production logistics and supply chain system—part
2. Ind. Manag. Data Syst. (2018). https://doi.org/10.1108/IMDS-11-2016-0504
26. Cardin, O., Leitão, P., Thomas, A.: Cyber-physical systems for future industrial systems 3
(2016)
27. Rajkumar, R., Lee, I., Sha, L., Stankovic, J.: Cyber-physical systems: the next computing
revolution. In: Design Automation Conference, pp. 731–736 (2010)
28. Mosterman, P.J., Zander, J.: Industry 4.0 as a cyber-physical system study. Softw. Syst. Model.
15, 17–29 (2016). https://doi.org/10.1007/s10270-015-0493-x
146 M. Moufaddal et al.

29. Nagy, J., Oláh, J., Erdei, E., Máté, D., Popp, J.: The role and impact of Industry 4.0 and
the internet of things on the business strategy of the value chain—the case of Hungary.
Sustainability 10, 1–25 (2018)
30. Oks, S.J., Fritzsche, A., Möslein, K.M.: An application map for industrial cyber-physical
systems. In: Jeschke, S., Brecher, C., Song, H., Rawat, D.B. (eds.) Industrial internet of things:
cybermanufacturing systems, pp. 21–46. Springer International Publishing, Cham (2017)
31. Liu, C., Jiang, P.: A cyber-physical system architecture in shop floor for intelligent manufac-
turing. Procedia CIRP 56, 372–377 (2016). https://doi.org/10.1016/j.procir.2016.10.059
32. Schneider, M., Lucke, D., Adolf, T.: A cyber-physical failure management system for smart
factories. Procedia CIRP 81, 300–305 (2019). https://doi.org/10.1016/j.procir.2019.03.052
33. Jeon, J., Kang, S., Chun, I.: CPS-based Model-Driven Approach to Smart Manufacturing
Systems. IARIA, pp. 133–135 (2016)
34. Zhang, Y., Guo, Z., Lv, J., Liu, Y.: A framework for smart production-logistics systems based
on CPS and industrial IoT. IEEE Trans. Ind. Inf. 14, 4019–4032 (2018). https://doi.org/10.
1109/TII.2018.2845683
35. Yan, J., Zhang, M., Fu, Z.: An intralogistics-oriented cyber-physical system for workshop in
the context of Industry 4.0. Procedia Manuf. 35, 1178–1183 (2019). https://doi.org/10.1016/j.
promfg.2019.06.074
36. Dworschak, B., Zaiser, H.: Competences for cyber-physical systems in manufacturing—first
findings and scenarios. Procedia CIRP 25, 345–350 (2014). https://doi.org/10.1016/j.procir.
2014.10.048
37. Wang, L., Ji, W.: Cloud enabled CPS and big data in manufacturing. In: Ni, J., Majstorovic,
V.D., Djurdjanovic, D. (eds.) Proceedings of 3rd International Conference on the Industry 4.0
Model for Advanced Manufacturing, pp. 265–292. Springer International Publishing (2018)
38. Wright, P.: Cyber-physical product manufacturing. Manuf. Lett. 2, 49–53 (2014). https://doi.
org/10.1016/j.mfglet.2013.10.001
39. Kim, S., Park, S.: CPS (Cyber Physical System) based manufacturing system optimization.
Procedia Comput. Sci. 122, 518–524 (2017). https://doi.org/10.1016/j.procs.2017.11.401
40. Monostori, L., Kádár, B., Bauernhansl, T., Kondoh, S., Kumara, S., Reinhart, G., Sauer, O.,
Schuh, G., Sihn, W., Ueda, K.: Cyber-physical systems in manufacturing. CIRP Ann. 65,
621–641 (2016). https://doi.org/10.1016/j.cirp.2016.06.005
41. Taisch, M.: Role of CPS in Manufacturing 18 (2015)
42. Lugaresi, G.: Simulation Modeling and Production Control in CPS-Based Manufacturing
Systems. ACM, New York, NY (2017)
43. Geissbauer, R., Vedso, J., Schrauf, S.: Industry 4.0: building the digital enterprise. PwC (2016)
44. Jabbar, S., Khan, M., Silva, B.N., Han, K.: A REST-based industrial web of things’ framework
for smart warehousing. J. Supercomput. 74, 4419–4433 (2018). https://doi.org/10.1007/s11
227-016-1937-y
45. Liu, X., Cao, J., Yang, Y., Jiang, S.: CPS-based smart warehouse for Industry 4.0: a survey of
the underlying technologies. Computers 7(13) (2018). https://doi.org/10.3390/computers701
0013
46. Macaulay, J., Chung, G., Buckalew, L.: Internet of Things in Logistics. DHL Customer
Solutions & Innovation (2015)
47. Bartholdi, J.J., Hackman, S.T.: Warehouse & Distribution Science. The Supply Chain and
Logistics Institute, School of Industrial and Systems Engineering, Georgia Institute of
Technology, Atlanta, GA 30332-0205 USA (2014)
48. Gong, Y., Koster, R.D.: A polling-based dynamic order picking system for online retailers. IIE
Trans. 40, 1070–1082 (2008). https://doi.org/10.1080/07408170802167670
49. Ramaa, A., Subramanya, K.N., Rangaswamy, T.M.: Impact of warehouse management system
in a supply chain. Int. J. Comput. Appl. 54 (2012)
50. Ghiani, G., Laporte, G., Musmanno, R.: Introduction to Logistics Systems Management. Wiley
(2013)
51. Frazelle, E.H.: World-Class Warehousing and Material Handling. McGraw Hill Professional
(2001)
A Cyber-Physical Warehouse Management System … 147

52. Poon, T.C., Choy, K.L., Chow, H.K.H., Lau, H.C.W., Chan, F.T.S., Ho, K.C.: A RFID case-based
logistics resource management system for managing order-picking operations in warehouses.
Expert Syst. Appl. 36, 8277–8301 (2009). https://doi.org/10.1016/j.eswa.2008.10.011
53. Grzeszick, R., Lenk, J.M., Rueda, F.M., Fink, G.A., Feldhorst, S., ten Hompel, M.: Deep neural
network based human activity recognition for the order picking process. In: Proceedings of the
4th international Workshop on Sensor-based Activity Recognition and Interaction—iWOAR
’17, pp. 1–6. ACM Press, Rostock, Germany (2017)
54. Xie, L., Thieme, N., Krenzler, R.K., Li, H.: Efficient order picking methods in robotic mobile
fulfillment systems (2019). arXiv:1902.03092. https://doi.org/10.13140/RG.2.2.137748.30082
55. Merkuryev, Y., Burinskiene, A., Merkuryeva, G.: Warehouse order picking process. In:
Merkuryev, Y., Merkuryeva, G., Piera, M.À., Guasch, A. (eds.) Simulation-Based Case Studies
in Logistics, pp. 147–165. Springer London, London (2009)
56. Ottjes, J.A., Hoogenes, E.: Order picking and traffic simulation in distribution centres. Int. J.
Phys. Distrib. Mater. Manag. 18, 14–21 (1988). https://doi.org/10.1108/eb014696
57. Berttram, P., Schrauf, S.: Industry 4.0: how digitization makes the supply chain more efficient,
agile, and customer-focused. https://www.strategyand.pwc.com/reports/digitization-more-eff
icient (2018). Accessed 10 May 2018
58. Basile, F., Chiacchio, P., Coppola, J.: A cyber-physical view of automated warehouse systems.
In: 2016 IEEE International Conference on Automation Science and Engineering (CASE)
(2016), pp. 407–412
59. Bartodziej, C.J.: The concept Industry 4.0. In: Bartodziej, C.J. (ed.) The Concept Industry 4.0
: An Empirical Analysis of Technologies and Applications in Production Logistics, pp. 27–50.
Springer Fachmedien Wiesbaden, Wiesbaden (2017)
60. Dai, Q., Zhong, R., Huang, G.Q., Qu, T., Zhang, T., Luo, T.Y.: Radio frequency identification-
enabled real-time manufacturing execution system: a case study in an automotive part manu-
facturer. Int. J. Comput. Integr. Manuf. 25, 51–65 (2012). https://doi.org/10.1080/0951192X.
2011.562546
61. Giannikas, V., Lu, W., McFarlane, D., Hyde, J.: Product intelligence in warehouse management:
a case study. In: Mařík, V., Lastra, J.L.M., Skobelev, P. (eds.) Industrial Applications of Holonic
and Multi-Agent Systems, pp. 224–235. Springer, Berlin Heidelberg (2013)
62. Faber, N., de Koster, R. (Marinus) B.M., van de Velde, S.L.: Linking warehouse complexity
to warehouse planning and control structure: an exploratory study of the use of warehouse
management information systems. Int. J. Phys. Distrib. Logist. Manag. 32, 381–395 (2002).
https://doi.org/10.1108/09600030210434161
63. Buntak, K., Kovačić, M., Mutavdžija, M.: Internet of things and smart warehouses as the
future of logistics. Tehnički glasnik 13, 248–253 (2019). https://doi.org/10.31803/tg-201902
15200430
64. Ding, W.: Study of smart warehouse management system based on the IOT. In: Du, Z. (ed.) Intel-
ligence Computation and Evolutionary Computation, pp. 203–207. Springer, Berlin, Heidelberg
(2013)
65. Wang, X.V., Kemény, Z., Váncza, J., Wang, L.: Human–robot collaborative assembly in cyber-
physical production: classification framework and implementation. CIRP Ann. 66, 5–8 (2017).
https://doi.org/10.1016/j.cirp.2017.04.101
66. Fong, T., Thorpe, C., Baur, C.: Collaboration, dialogue, human-robot interaction. In: Jarvis,
R.A., Zelinsky, A. (eds.) Robotics Research, pp. 255–266. Springer, Berlin, Heidelberg (2003)
67. Zheng, X., Wang, M., Ordieres-Meré, J.: Comparison of data preprocessing approaches for
applying deep learning to human activity recognition in the context of Industry 4.0. Sensors
(Basel) 18 (2018). https://doi.org/10.3390/s18072146
68. Valle, A., Jacobsen, D.P.: Optimización de una red logística de distribución global e impacto
de la Industria 4.0 en centros de almacenamiento 16 (2019)
69. Rippert, D.J.: The building blocks of a simpler future are in place. Financial Times (2006)
70. Reis, J.Z., Gonçalves, R.F.: The role of Internet of Services (IoS) on Industry 4.0 through
the Service Oriented Architecture (SOA). In: Moon, I., Lee, G.M., Park, J., Kiritsis, D., von
Cieminski, G. (eds.) Advances in Production Management Systems. Smart Manufacturing for
Industry 4.0, pp. 20–26. Springer International Publishing, Cham (2018)
148 M. Moufaddal et al.

71. Lanting CJM, Lionetto A (2015) Smart Systems and Cyber Physical Systems paradigms in an
IoT and Industry/ie4.0 context 14
PLM and Smart Technologies
for Product and Supply Chain Design

Oulfa Labbi and Abdeslam Ahmadi

Abstract The emergence of advanced smart technologies (e.g. Internet-of Things,

Big Data Analytics, and Artificial Intelligence techniques) has created a number of
innovations in handling supply chain activities and thus accelerate the transforma-
tion of the traditional supply chain towards a smart supply chain. Many existing
studies focused only on applying smart technologies to facilitate smart manufac-
turing, but few ones addressed investigations covering smart technologies in the
whole product lifecycle, namely the design and recovery stages. This paper interest
focus on how smart technologies combined to product lifecycle management (PLM)
contributes to realize sharing and integration of data in different lifecycle stages
of the product? And how this could help to conduct a smart product and supply
chain design? Moreover, the paper investigates the use of smart technologies espe-
cially during the product design phase where the supply chain configuration will be
conducted. For this purpose, a framework is addressed and an UML model for smart
product and supply chain design is proposed.

Keywords Smart technologies · PLM · UML · Product design phase · Smart

supply chain design

1 Introduction

Nowadays companies are seeking to explore new sources of competitiveness through

the optimization of their supply chains and through strong coordination between their
partners. Such cooperation requires a massive product information exchange between

O. Labbi (B)
Department of Industrial Engineering, Ecole Nationale des Sciences Appliquées, Sidi Mohamed
Ben Abdellah University, Avenue My Abdallah Km 5, Route d’Imouzzer, BP 72, 30000 Fes,
Morocco
e-mail: [email protected]
A. Ahmadi
Department of Mathematics and Informatics, Ecole Nationale Supérieure d’Arts et Métiers,
Moulay Ismail University, B.P. 15290 El Mansour, 50500 Meknes, Morocco
e-mail: [email protected]
© Springer Nature Switzerland AG 2021 149
T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_10
150 O. Labbi and A. Ahmadi

all supply chain stakeholders. For this purpose, many advanced manufacturing tech-
nologies were put forward, such as computer aided design (CAD), manufacturing
execution system (MES), computer aided manufacturing (CAM), enterprise resource
planning (ERP), and recently product lifecycle management (PLM).
PLM enables comparison, evaluation and optimization of the different product
requirements, linking production information (specifications, models, results of
performance, best practices, and reviews) to design thanks to knowledge manage-
ment it provides [1]. It integrates and makes available all data collected during all
phases of the product lifecycle for all the stakeholders across the extended enterprise
[2].
In fact, with traditional PLM systems, massive data can be gathered through the
whole lifecycle and are available with various degrees of complexity but unfortu-
nately there is a lack in its exploitation. However, with the promotion and imple-
mentation of advanced smart technologies (e.g. Internet-of Things (IoT), Big Data
Analytics (BDA), and Artificial Intelligence (AI)), the lifecycle data and knowledge
can be effectively facilitated, more reasonable and better exploited enabling all supply
chain components to be perceptible, predictable, and optimizable [3–5]. Through this
context, the aim of this paper is to show how smart technologies combined to PLM
contribute to good performance and optimization of the supply chain and its products.
Browsing literature researches, many authors addressed the problem of joint
product and supply chain design [6, 7], others investigated the PLM as an approach
and also as a technology to support the simultaneous product and supply chain
configuration [8–10] and recently others addressed the smart enabling technolo-
gies for supply chain management [11–13] but the product and supply chain design
using smart technologies is seldom addressed. Indeed, most researches concerned
the middle of product lifecycle (MOL) including production, inventory and main-
tenance to address issues regarding intelligent manufacturing [14, 15], intelligent
inventory [16] and intelligent maintenance [17], few researchers performed inves-
tigations covering smart enabling technologies in the whole lifecycle, especially
the design and recovery phases. Thus, this paper interests the product design phase
and proposes to combine PLM and smart technologies to conduct a joint and smart
product and supply chain design. Indeed, most benefits of collaboration among supply
chain partners lie in the design phase of the product lifecycle since the cost of design
changes increases as the design phase of the product lifecycle ends and the manu-
facturing phase starts [18]. Therefore, it is important to integrate product and supply
chain design decisions during the beginning of product lifecycle (BOL) [19].
To sum up, this paper proposes to conduct product and supply chain design during
the product design phase using the advantages of implementing smart technologies
in the whole product lifecycle. For this purpose, a conceptual framework and an
UML model is proposed.
The rest of paper will be organized as follows: in Sect. 2, we give an overview of
smart technologies for PLM and a literature review of smart technologies for supply
chain management. In Sect. 3, we present the research approach by proposing a
conceptual framework for smart product and supply chain design. Section 4 presents
PLM and Smart Technologies for Product and Supply Chain Design 151

the UML model proposed. Finally, a conclusion with perspectives is presented at the
end of the paper.

2 Literature Review

2.1 Smart Technologies for PLM

Product Lifecycle Management is a concept with multiple interpretations. The most

common use of PLM is about how to manage product lifecycle gathered informa-
tion for future needs for design and manufacturing, as well as for spare parts and
maintenance management. However, collecting maintenance and service, as well
as product usage information have not been as a fundamental part of PLM. In this
section, we will provide advantages of implementing smart technologies to improve
product lifecycle management through several research works.
For the BOL phase, designers will be able to exploit expertise and know-how of
all actors implemented in the product’s lifecycle and thus improve product design
and quality using smart technologies such as big data analytics techniques (BDA)
and cloud based technologies. For instance, Bohlouli et al. [20], proposed a cloud-
based product design knowledge integration framework to integrate knowledge in the
collaborative product design procedure for sustainable and innovative product design.
BDA for product design was investigated by tao et al in their digital twin-driven
product design, manufacturing and service framework [21].
Using smart technologies in MOL phase including manufacturing, maintenance
and distribution was addressed in many researches. Karkkainen et al. [22] conducted
a solution design experiment for implementing a tracking based inventory manage-
ment system in temporary storage locationsof a project delivery chain. Other works
covering smart maintenance, manufacturing and distribution proposed a complete
and always up-to-date report about the status of the product and real-time assistance
through the usage of artificial intelligence (AI), internet of things (IOT) technologies
and radio frequency identification (RFID) systems [23–25].
End of lifecycle (EOL) phase is few addressed. Jun et al. [26] used RFID device
as a mobile memory to record and update the real-time degeneration status and
lifetime data of components in the EOL stage. Jensen and Remmen [27] analyzed
how IoT technology contributes to stimulate and implement high quality EOL product
management strategies.
Finally, works handling smart technologies for supporting PLM systems are
seldom addressed. In this context, we find the work of Dekhtiar et al. [28] that have
conducted an engineering case study with the objective to design a software product
based on deep learning techniques that could support Product Lifecycle Manage-
ment (PLM) in terms of activity. In their work, deep learning was used for big data
application in PLM to tackle the data heterogeneity issue in terms of type and format,
to resist to noisy data as possible and to conduct a limited system-redesign costs in
152 O. Labbi and A. Ahmadi

case of situation changes. For this purpose, they construct a dataset for the digital
Mock up (DMU) which will act as a knowledge-base for the deep learning model.

2.2 Smart Technologies for Supply Chain Management

The advanced smart technologies have created a number of innovations in handling

supply chain management and activities. Many researchers have deployed smart
enabling technologies to tackle supply chain issues such as supplier selection
problem, supply chain planning and conducting smart supply chain functions (e.g.
smart manufacturing, smart maintenance, smart inventory, …).
Oh et al. [11], have developed in their study a profit-effective and response efficient
tactical supply planning model to find an optimal trade-off between profit and lead
time. In their work, tactical supply planning was conducted in smart manufacturing
supply chain context; it recognized the ability of a pseudo real-time decision-making
constrained by the planning horizon. In other researches, smart technologies were
applied to facilitate the implementation of sustainable production and consumption
strategies [29, 30]. Zheng and Wu [31] have proposed a smart spare parts inventory
management system to establish transparency between manufacturers and suppliers
and to reduce the inventory costs. As far as intelligent maintenance is concerned, the
use of IoT provides real-time data of the whole lifecycle to improve maintenance
and service decisions [32].
In this paper, smart technologies will be used to improve PLM traditional func-
tionalities to conduct better product and its supply chain design. Next section will
present the approach adopted to do so.

3 Research Adopted Approach

This section explains the approach adopted for conducting a smart joint product and
supply chain design. Indeed, the product and its supply chain design will be handled
by combining the following:
• Smart technologies integrated to Product Lifecycle Management (PLM) solutions
to manage all life cycle data incorporated at the product design.
• Integration of supply chain constraints at the product design. For instance,
constraints of suppliers, transportation, storage conditions, and manufacturing
processes are introduced in the digital mock up (DMU) during the design phase.
• Integration of product constraints in the supply chain design allowing supply chain
partners to consider specificities of the designed product in their functions and
duties.
• Considering the product lifecycle phases as a horizon decision for the supply chain
optimization and configuration. In fact, based on the application of advanced smart
PLM and Smart Technologies for Product and Supply Chain Design 153

PLM Data Smart Technologies SC functions

Product Life Cycle Phases

Smart design
Product BDA
Development
Data Smart production

Smart inventory
Production AI Smart distribution
Data and tracking
IOT Smart
maintenance
Operational Smart use
support RFID Smart recovery
Data and recycling

Fig. 1 PLM and smart technologies for product and supply chain design framework

technologies, smart supply chain is designed to be responsive and customized to be

flexible in order to face product lifecycle changes such as growing and decreasing
of the product demand, opening and closing of facilities and so on.
• Focusing on the feedback and sharing of lifecycle data among all stakeholders
in the supply chain (clients, suppliers, manufacturers, distributors, recyclers..) to
take optimal decisions since each decision taken in one phase of the lifecycle
affects other phases of the lifecycle.

Figure 1 illustrates PLM and smart technologies framework for smart product and
supply chain design. Indeed, PLM data, Smart technologies, and smart supply chain
functions are the three key components of our approach. In fact, between this three
elements: knowledge flows, data flows, data sources and main lifecycle stages are
interconnected. Next sections describe each component in details.

3.1 Smart Technologies for PLM Support

PLM data represents data to be managed through PLM systems. Indeed, it refers to
large amounts of multi-source, heterogeneous data generated throughout the product
lifecycle. Generally speaking, we can classify PLM data to many categories:
154 O. Labbi and A. Ahmadi

• Manufacturing data collected from manufacturing information systems (e.g.,

MES, ERP, SCM)
• Equipment data collected by industrial IoT technologies, which includes data
related to real-time performance, operating conditions, and the maintenance
history of production equipment [28].
• User data collected from internet platforms involving user profiles, user prefer-
ences towards products/services, as well as user behavior.

In our work, we considered that data to be managed are classified into three main cate-
gories namely, product development data, production data and operational support
data [9]. Product development data describe how the product is designed, manu-
factured, operated or used, serviced and then retired. Production data focus on all
activities associated with the production and the distribution of the product. Opera-
tional support data deal with the enterprise’s core resources, such as people, finances
and other resources required for supporting the enterprise.
These three categories of data are collected based on the usage of advanced smart
technologies. In this work, we focus especially on BDA, AI, IoT and RFID systems.
For instance, IoT technologies with RFID systems could be used to provide real-time
data collected from equipment’s process, inventory warehouses, internet platforms
for product users, etc. For instance, customers’ data gathered through internet plat-
forms and IoT will be as an input for product development data and will guide
designers to dress new products requirements respecting users’ preferences. BDA
will be used to extract useful and valuable information from big data generated in
the whole product lifecycle. It will help to manage heterogeneity which makes any
data usage difficult. AI techniques’ usage concerns mainly optimal decision making
thanks to their capability to intelligently recognize and learn business models such
as supplier selection decisions, production scheduling, etc…

3.2 PLM and Smart Technologies for Smart Supply Chain

Functions

As shown in Fig. 1, from product lifecycle phase’s perspective, the implementation

of advanced smart technologies will lead to smart design, smart production, smart
inventory, smart distribution and tracking, smart maintenance and smart use and
recovery.
In fact, smart product design is achieved by applying BDA, for instance, to handle
data related to customer demands such as customer preferences and behaviors. This
will help to predict market demands and to develop new product design strategies. In
addition, taking into account the whole product lifecycle data will help designers to
integrate other phases constraints (manufacturing, recycling phases) and conduct an
integrated product design, such as design for manufacturing, design for maintenance
and sustainable eco-design.
PLM and Smart Technologies for Product and Supply Chain Design 155

Smart production, smart inventory and smart distribution and tracking are
conducted thanks to IoT combined to RFID tagging systems. For instance, to have a
real-time production and inventory visibility, all manufacturing items (e.g. products,
trolleys, machines, ranks) are connected to smart sensors and RFID tags and then
connected to the product management system to provide real-time data visibility. In
addition, the IoT platform can indicate the condition of each product, such as the
storage conditions or expiration date, to ensure the quality of the products in the
supply chain.
Smart maintenance concerns mainly the prediction side. In fact, by attaching
sensors to equipment or parts, they can immediately alert the supply chain profes-
sional when it is operating outside acceptable levels. When using for instance,
machine learning, the IoT platform can determine whether the part needs to be
replaced or repaired and then automatically trigger the correct process.
Smart recovery involves data gathered at the recycling phase and how to use it for
optimal recovery decisions making such as remanufacturing, recycle or disposal of
products.
To sum up, the deployment of smart technologies in the whole product lifecycle
improves substantially the supply chain performance. In addition, considering the
lifecycle feedback data leads to a great expertise and provides a better knowledge
management. The idea of this paper is to exploit data gathered during the whole
product lifecycle from the different supply chain stakeholders to conduct an optimal
product and supply chain design. Next section describes the model proposed for
conducting this joint design.

3.3 PLM and Smart Technologies for Product and Supply

Chain Design

The product and its supply chain design are conducted during the design phase. In
fact, the configuration of the supply chain is studied during the product design phase,
which means that the product configuration is not yet finalized. In the product’s DMU,
specifications of the product to be designed and constraints of supply chain actors
are both integrated. This design of the product is supported by data gathered from
smart supply chain functions:
• Market research and marketing is made thanks to BDA and IoT techniques. Indeed,
the demand for the new product to be designed could be estimated based on
knowledge of user or consumer preferences and behaviors and also based on
the feedback of other lifecycle phases regarding demand evolution for previous
products.
• Designing a product means determining BOMs, components, manufacturing
processes and also determining the way it will be maintained, used and recovered.
156 O. Labbi and A. Ahmadi

Considering the whole lifecycle data gathered the different supply chain func-
tions will assist designers to consider production and other logistical constraints
to reduce design errors and iterations.

From the same perspective of product lifecycle, the design of the supply chain is
also improved by considering feedback and sharing data among all supply chain
partners and also by implying then during the design stage. This sharing will assist
designers on one side and will assist also each supply chain partner to operate in good
performing conditions. For instance, involving suppliers while designing the product
will give them an idea about their capacity to supply components of this product.
Manufacturers could also anticipate equipment and facilities to produce this product.
Supporting PLM by smart technologies will contribute also to a sustainable
product and supply chain design. For instance, to meet product demands, production
planning is optimized including expected quantities recovered or recycled. In fact,
through analyses of the lifecycle data and IoT, the remaining lifetime of products
or parts can be predicted in real-time helping to make the right recovery decisions
(reuse or remanufacturing for example).

4 The Model Proposed for Smart Product and Supply

Chain Design

As mentioned beforehand, we will consider the product design phase to conduct the
product and its supply chain design for the high impact of this phase decisions on the
other lifecycle phases. We use Unified Model Language (UML) to model the static
aspect of the product and its supply chain design (Fig. 2).
To remember, the model proposed is supported by PLM functionalities empow-
ered by smart technologies implementation. In other words, since PLM is based on
the principle of experience capitalization and data archiving, product and logistics
data of the UML model proposed are considered as inputs for the product design.
For instance, the demand for designed product could be estimated from behaviors of
the existing products’ demand in each product lifecycle phase from introduction to
decline phases. As explained before, this information comes from the usage of smart
technologies (IoT, BDA,..). Furthermore, all data regarding supply chain partners are
also gathered the same way and thus could be integrated in the DMU of the product.
DMU is the process of building a numerical (digital 3D) representation of a
product to conduct tests that will predict product function and performance in the
real world.
While developing the DMU, we are reducing the need for physical product
prototyping that is the most expensive aspect of the product development. In fact,
before starting the product’s DMU development, the design team should carry out
constraints of old supply chain partners, technical constraints of the manufacturers
and customer specifications to list the new product requirements [9].
PLM and Smart Technologies for Product and Supply Chain Design 157

Fig. 2 UML class diagram modelling the product and supply chain design

In the model proposed, the class “Product_DMU” is the numerical representation

of the product. It provides technical data of the product such as bill of materials
(BOM) or components. For each product lifecycle phase, represented in the model
by the class “Cyclephase” and for each product, there is a specific supply chain
configuration (customers, suppliers, manufacturers, production process, distribution
and recycling centers). This is for controlling the evolution of supply chain structure
in accordance with the different lifecycle stages.
The ESC Node class represents each node or partner belonging to the supply
chain. It could be either a supplier or a manufacturer or any other partner. Each ESC
node could have many production processes (equipment) to realize its product. We
consider the cardinality 0…* between ESC node class and equipment class to take
into account the case where the ESC node is a distribution center (it doesn’t have a
production machine).
158 O. Labbi and A. Ahmadi

In this model, we showed that the product’s DMU comprises not only data
regarding product architecture and features of its components but also incorporates
the logistical constraints of all supply chain partners. This was through association
classes that result from relationship between different classes such as: supplying,
order, production, inventory and recycling classes. In these association classes, we
are informed about logistical data such as costs and quantities.

5 Conclusion

This work belongs to works covering the use of smart technologies to handle supply
chain activities. In fact, many researchers only focused upon one stage of the lifecycle
(production stage of BOL and operation or maintenance stage of MOL) to tackle
intelligent manufacturing and intelligent maintenance issues. There is a lack in works
dealing with the implementation of smart technologies in the whole lifecycle to
exploit data sharing and knowledge interaction among the various lifecycle stages
for a better supply chain performance. This paper considers the product lifecycle
perspective to propose a framework that combines the PLM and smart technologies
(e.g. BDA, IoT, AI, …) for product and supply chain design.
The aim of the paper is to exploit the benefits of implementing smart technologies
to empower the traditional PLM functionalities to conduct an optimal product design.
Obviously, when we talk about the product design, it is mandatory to think about
its product’s supply chain (which suppliers for this product, which manufacturing
process, how the product will be stored, distributed and recovered). In this context,
the paper tackles the design of the product and its supply chain in the same time
during the product design. This is for reporting in the upstream level, the constraints
of other downstream supply chain functions.
To design the product and its supply chain, smart technologies are integrated to
Product Lifecycle Management (PLM) solutions to manage all life cycle data incor-
porated at the product design. In addition, supply chain constraints are incorporated
at the digital mock up during the design phase. For this purpose, an UML model was
proposed to describe our approach for the design of the product and its supply chain.
In this work, we presented how the design of the product and its supply chain
could be handled exploiting the advantages afforded by the implementation of smart
technologies to PLM but we did not focus on how the optimization of the supply
chain is conducted. Future works will cover smart technologies and applications for
supply chain optimization issues.
PLM and Smart Technologies for Product and Supply Chain Design 159

References

1. Trotta, M.G., di Torino, P.: Product lifecycle management: sustainability and knowledge
management as keys in a complex system of product development. J. Ind. Eng. Manag. 3(2),
309–322 (2010)
2. Sudarsan, R., et al.: A product information modeling framework for product lifecycle
management. Comput. Aided Des. 37, 1399–1411 (2005)
3. Büyüközkan, G., Göçer, F.: Digital supply chain: literature review and a proposed framework
for future research. Comput. Ind. 97, 157–177 (2018)
4. Oh, J., Jeong, B.: Tactical supply planning in smart manufacturing supply chain. Robot.
Comput. Integr. Manufact. 55, 217–233 (2019)
5. Ren, S., Zhang, Y., Liu, Y., Sakao, T., Huisingh, D., Almeida, C.M.V.B.: A comprehensive
review of big data analytics throughout product lifecycle to support sustainable smart manu-
facturing: a framework, challenges and future research directions. J. Clean. Prod. (2018). https://
doi.org/10.1016/j.jclepro.2018.11.025
6. Yang, D., Jiao, R.J., Ji, Y., Du, G., Helo, P., Valente, A.: Joint optimization for coordinated
configuration of product families and supply chains by a leader-follower Stackelberg game.
Eur. J. Oper. Res. (2015). https://doi.org/10.1016/j.ejor.2015.04.022
7. Labbi, O., Ouzizi, L., Douimi, M., Ahmadi, A.: Genetic algorithm combined with Taguchi
method for optimisation of supply chain configuration considering new product design. Int. J.
Logist. Syst. Manag. 31(4), 531–561 (2018)
8. Labbi, O., Ouzizi, L., Douimi, M.: A dynamic model for the redesign of a product and
its upstream supply chain integrating PLM (Product Lifecycle Management) approach
In: Proceedings of the International Conference on Industrial Engineering and Systems
Management (IEEE-IESM’2015), pp. 1187–1196 (2015)
9. Labbi, O., Ouzizi, L., Douimi, M., Aoura, Y.: A model to design a product and its extended
supply chain integrating PLM (product lifecycle management) solution. Int. J. Sci. Eng. Res.
7(10), 1190–1220 (2016)
10. Bouhaddou, I., Benabdelhafid, A., Ouzizi, L., Benghabrit, Y.: PLM (product lifecycle manage-
ment) approach to design a product and its optimized supply chain. Int. J. Bus. Perform. Supply
Chain Model. 6(3/4), 255–275 (2014)
11. Jisoo, O., Bongju, J.: Tactical supply planning in smart manufacturing supply chain. Robot.
Comput. Integr. Manuf. 55, 217–233 (2019)
12. Frazzon, E.M., Silva, L.S., Hurtado, P.A.: Synchronizing and improving supply chains through
the application of cyber physical systems. IFAC PapersOnLine 48(3), 2059–2064 (2015).
https://doi.org/10.1016/j.ifacol.2015.06.392
13. Ivanov, D., Dolgui, A., Sokolov, B., Werner, F., Ivanova, M.: A dynamic model and an algorithm
for short-term supply chain scheduling in the smart factory industry 4.0. Int. J. Prod. Res. 54(2),
386–402 (2016). https://doi.org/10.1080/00207543.2014.999958
14. Mittal, S., Khan, M.A., Wuest, T.: Smart manufacturing: characteristics and technologies. In:
IFIP International Conference on Product Lifecycle Management, pp. 539–548 (2016)
15. Renzi, C., Leali, F., Cavazzuti, M., Andrisano, A.O.: A review on artificial intelligence appli-
cations to the optimal design of dedicated and reconfigurable manufacturing systems. Int. J.
Adv. Manuf. Technol. 72, 403–418 (2014). https://doi.org/10.1007/s00170-014-5674-1
16. Boone, C.A., Skipper, J.B., Hazen, B.T.: A framework for investigating the role of big data in
service parts management. J. Clean. Prod. 153, 687–691 (2017). https://doi.org/10.1016/j.jcl
epro.2016.09.201
17. Bahga, A., Madisetti, V.K.: Analyzing massive machine maintenance data in a computing
cloud. IEEE Trans. Parallel Distrib. Syst. 23(10), 1831–1843 (2012)
18. Gokhan, N.M.: Development of a Simultaneous Design for Supply Chain Process for the
Optimization of the Product Design and Supply Chain Configuration Problem. Ph.D. thesis,
University of Pittsburgh (2007).
160 O. Labbi and A. Ahmadi

19. Chiu, M.C., Okudan, G.E. : An investigation on the impact of product modularity level on
supply chain performance metrics: an industrial case study. J. Intell. Manuf. 30(4), 129–145
(2014)
20. Bohlouli, M., Holland, A., Fathi, M.: Knowledge integration of collaborative product
design using cloud computing infrastructure. In: IEEE International Conference on Electro
Information Technology, pp. 1–8 (2011). https://doi.org/10.1109/EIT.2011.5978611
21. Tao, F., Cheng, J., Qi, Q., Zhang, M., Zhang, H., Sui, F.: Digital twin-driven product design,
manufacturing and service with big data. Int. J. Adv. Manuf. Technol. 94, 3563–3576 (2018).
https://doi.org/10.1007/s00170-017-0233-1
22. Karkkainen, M., Ala-Risku, T., Framling, K., Collin, J., Holmstrom, J.: Implementing inventory
transparency to temporary storage locations: a solution design experiment in project business.
Int. J. Manag. Proj. Bus. 3(2), 292–306 (2010)
23. Laalaoui, Y., Bouguila, N.: Pre-run-time scheduling in real-time systems: current researches
and artificial intelligence perspectives. Expert Syst. Appl. 41, 2196–2210 (2014). https://doi.
org/10.1016/j.eswa.2013.09.018
24. Kumar, A., Shankar, R., Thakur, L.S.: A big data driven sustainable manufacturing framework
for condition-based maintenance prediction. J. Comput. Sci. 27, 428–439 (2018). https://doi.
org/10.1016/j.jocs.2017.06.006
25. Abdel-Basset, M., Manogaran, G., Mohamed, M.: Internet of Things (IoT) and its impact on
supply chain: a framework for building smart, secure and efficient systems. Future Gener.
Comput. Syst. 86, 614–628 (2018)
26. Jun, H.B., Shin, J.H., Kim, Y.S., Kiritsis, D., Xirouchakis, P.: A framework for RFID appli-
cations in product lifecycle management. Int. J. Comput. Integr. Manuf. 22, 595–615 (2009).
https://doi.org/10.1080/09511920701501753
27. Jensen, J.P., Remmen, A.: Enabling circular economy through product stewardship. Procedia
Manuf. 8, 377–384 (2017). https://doi.org/10.1016/j.promfg.2017.02.048
28. Dekhtiara, J., Durupta, A., Bricognea, M., Eynarda, B., Rowsonb, H., Kiritsis, D.: Deep learning
for big data applications in CAD and PLM-research review, opportunities and case study.
Comput. Ind. 100, 227–243 (2018)
29. Zhang, Y., Ren, S., Liu, Y., Si, S.: A big data analytics architecture for cleaner manufacturing
and maintenance processes of complex products. J. Clean. Prod. 142, 626–641 (2017)
30. Roy, V., Singh, S.: Mapping the business focus in sustainable production and consumption
literature: review and research framework. J. Clean. Prod. 150, 224–236 (2017). https://doi.
org/10.1016/j.jclepro.2017.03.040
31. Zheng, M., Wu, K.: Smart spare parts management systems in semiconductor manufacturing.
Ind. Manag. Data Syst. 117, 754–763 (2017). https://doi.org/10.1108/IMDS-06-2016-0242
32. Zhang, Y., Zhang, G., Wang, J., Sun, S., Si, S., Yang, T.: Real-time information capturing and
integration framework of the internet of manufacturing things. Int. J. Comput. Integr. Manuf.
28(8), 811–822 (2015)
Production Systems Simulation
Considering Non-productive Times
and Human Factors

Ismail Taleb, Alain Etienne, and Ali Siadat

Abstract For decades and all around the world, the main goal of industries is finan-
cial and economic profit. Lately, new approaches additionally aim to better the oper-
ator’s working conditions. With the help of simulation, these approaches can be tested
in a much cheaper, faster and easier manner before real-life implementation. In this
paper, new strategies have been tested to improve working conditions for the operator
while maintaining high performance and productivity. The implementation of human
factors and margins of maneuver into a production system is explored to improve
working conditions through non-productive times. One main non-productive time
is considered: unnecessary movement. A model describing a real-life scenario is
suggested, as well as strategies to reduce and leverage non-productive times.

Keywords Simulation · Non-productive times · Human factors · Margins of

maneuver · Production systems

1 Introduction

Non-productive times mostly known as times related to non-value-added activities

are times that most industries look to remove, reduce or exploit. By doing so, they
reduce costs, improve lead time and thus increase profit [1], in the other hand, these
times were found to be important to the operator, adding variability which leads
to better working conditions [2]. Thus, the competitiveness in this era is increasing
continuously, demanding lower costs, better quality and high variety which translates
to smaller lot size and shorter lead time, thus flexibility and responsiveness is a
primordial aspect of companies’ concern [3]. Human-centered production systems
are one of the favorite candidates, even though automatic technologies are evolving
rapidly, most of the industries rely on these systems thanks to these advantages they
have [4].

I. Taleb (B) · A. Etienne · A. Siadat

Arts et Métiers Institute of Technology, Université de Lorraine, LCFC, HESAM Université,
57070 Metz, France
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 161

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_11
162 I. Taleb et al.

Most studies have focused more on one of the concepts discussed earlier without
looking at the bigger picture and the interaction of these concepts when they are
together. Abubakar and Wang [3] studied the effects of human factors such as age
on a human-centered production system.
A novel approach studied the interaction between some human factors (such as
fatigue, experience) and time allowances, to grant operators better working conditions
[5]. Digiesi et al. [6] shows that their models can describe the dynamic behavior of
workers regarding tiredness which can have huge impacts on flow time, buffer level
and throughput. They found that severe tiredness, usually towards the end of the
shift, creates accumulating effects on the buffers, slowing down production. They
suggest using buffers to cope with variability caused by fatigue. Margins of maneuver
[7] is a concept taking into consideration all of an operator’s internal and external
resources to enable an increase in control over his working conditions. They found
that increasing the operator’s control over his working conditions has direct benefits
over his physical and mental health, and thus indirect benefits over production and
productivity.
Productivity is linked to reducing non-productive times that are inspired by non-
value-added activities mainly studied in Lean Manufacturing, also known as wastes
or Muda (無駄). The fundamental focus of their studies is to “lean” the produc-
tion system, meaning to systematically remove any non-value-added activity, thus
reducing costs and increasing benefits. In practice, the aim is to reduce or exploit
these activities as they can’t be entirely removed. These activities are rarely tied to
times, hence the need to define non-productive times to be able to mathematically
quantify them.
This paper aims to study the effect of each of these concepts on a production
system, in particular on non-productive times such as unnecessary movement, as
well as their interaction between each other. The paper is organized as follows. First,
these key concepts are defined. Then, their behaviors are studied separately through
a simulation. And finally, their interactions are studied in a production system as well
as their influence on non-productive times.

2 State of the Art

This problem is related to production systems and contains multiple key concepts
acting at the same time: non-productive times, human factors, margins of maneuver.
This first section aims to define each one of these key concepts individually. The
next section aims to introduce the model simulating their collective behavior.
Production Systems Simulation Considering … 163

2.1 Production Systems Management

The production system is the main component on which all these parameters vary.
It has been the central focus of changes throughout history. El Mouayni et al. [8]
regroups the major steps the production system has undergone since the introduction
of standards and interchangeable pieces. The steppingstone of these changes was
the Taylorism, driven by the constant increase in demand, it brought innovative
measures such as the fragmenting of tasks and abilities, rather than multitasking. This
new approach based on small repetitive tasks increased worker performance. Then
comes the Fordism, which combined the previous approach with other technological
advances defining the first model of mass production. The Toyota Production System
(TPS) appeared later, where the main goal was to eliminate waste in order to reduce
cost, consequently, increase profit. Since then, many new approaches have been
developed by different industries with different tools, but one main goal remains:
increase profit [9]. Newer approaches focus now, in addition to increasing profit, on
improving working conditions and Human physical and mental health.
In fact, Industry 4.0 is believed to provide better working conditions via social
sustainability or fully automated machinery, leading to less demanding physical labor
[10].

2.2 Human Factors

In this new era of high competitiveness, increasing profitability is the core goal of
many businesses but it is not the only one. For instance, some of the other goals that
they aim to achieve are better working conditions [7], better physical and mental
(psychosocial) health [4] during and after work.
There are two main houses of study, the first one concentrates on the overall human
factors as a result [11, 12]. The second one studies the underlying reasons behind
these human factors such as physical, cognitive, cardiovascular and psychological
effects [13, 14]. The first one is more interesting to our study since it is related to
engineering whereas the second one is related to biology, studying for example, the
effect of fatigue on each part of the body separately.
Many studies have tried to identify the most important human factors contributing
to performance and working conditions. Givi et al. [11] have identified three mains
contributors: (1) the first one is the pair Fatigue-Recovery factor, (2) the second
pair is the Learning-Forgetting factor, and (3) the last one is the human error factor.
According to them, these 3 factors are the main ones in simulating the performance
as well as the health of the worker. This model has some drawbacks such as its
deterministic behavior, its focus on physical fatigue only, and the interpretation of
the results in a real situation. The authors assumed that fatigue and recovery follow
an exponential form with maximum benefit in the earlier phases and are as follows:
164 I. Taleb et al.

F(t) = 1 − e−λt (1)

R(τi ) = F(t)e−μτ i (2)

The learning-forgetting model is much more complicated considering many

parameters are considered, but the main formulas are shown in (3) and (4)

Tx = T1 x−b (3)

Tni = T1 (ui + ni )−b (4)

These human factors are directly tied to working conditions. And, one of the ways
to improve them is through the use of margins of maneuver, which will be discussed
in the next section.

2.3 Margins of Maneuver

Margins of maneuver are a new concept, very few articles talk about this subject, the
oldest article found dates to 1997 according to [7]. They are mainly used as temporal,
spatial or organizational buffers to allow the worker to have some control over his
workload and working conditions [7]. El Mouayni et al. [15] have used temporal
margins of maneuver to give the operator improved breaks according to their speed
allowing him to get better break allocations without losing performance. Margins
of maneuver are beneficial ergonomically for the worker and economically for the
company [7].
They make the operator more autonomous resulting in more economical profit.
El Mouayni et al. [16] proposed a model to simulate margins of maneuver (called
Time Margins) described as so:

T M i = (Cmax − pi ) × T M E E i (5)

where T M E E i is the Mean Time between Events, and described as such (in this
case, an event being the arrival of a new task):

E i − E i−1 + (i + 1) × T M E E i−1
T M E Ei = ; ∀i ≥ 1 (6)
i
Production Systems Simulation Considering … 165

2.4 Non-productive Times

Non-productive times, in the other hand, is wasted time. The term is based on Lean
Manufacturing which uses the terms waste and non-value-added activities. In this
field, these 2 concepts are well defined, but the concept of time is not always tied to
it. Some examples of waste are tied to movement, inventory, cost, …etc.
Activities are split between value-added and non-value-added activities. So, Low
et al. [17] thought of adding a third group of non-value-added activities but necessary,
meaning that it is impossible to remove waste entirely but there is a mandatory
minimum limit. The next section “Methodology” aims at explaining the logic chosen
for this study, starting by introducing a new, more exhaustive taxonomy for non-
productive times, then defining the model used for the simulation, explaining the
results found. Finally, the last section seeks to discuss these results and thus conclude
on them.

3 Methodology

3.1 New Non-productive Times Approach

Before presenting the model, the new non-productive times taxonomy is introduced.
It is developed from multiple Lean concepts (Muda) and other non-productive activ-
ities [18−23] (Fig. 1) using an Eisenhower matrix to identify different types of
non-productive times depending on the cadence and quality (Fig. 2).
In this article, the use of “non-productive time” implies all cases that are not
productive such as Poor or Bad Productivity (see Fig. 2).

Fig. 1 New taxonomy for non-productive times

166 I. Taleb et al.

Fig. 2 Eisenhower matrix with 2 entries: quality and cadence and 5 outputs

3.2 Model

In this study, the main focus is human-centered production systems, which means
that the center of the model should be the Human. The model describing the oper-
ator is based on the human factors described earlier. Around the operator are the
main components of the production system: machines, products and other governing
concepts like tasks and scheduling. Each one of these components and concepts are
described with the goal of simulation, to better understand the interactions between
each of them.
The model is built for a discrete-event agent-based simulation, where every
component of the model is modeled separately, and their combined behavior defines
different interactions and results. The inputs are generally rigid and are determined
by the situation for example the takt time is governed by demand and other factors.
Implementing the margins of maneuver into the organizational level means using
control parameters such as the number of operator or the maximum distance between
tasks to accept them (Fig. 3).
The robustness of the model has been tested through different inputs, then vali-
dated through comparison of real-life experiment. The model as is, is complex enough
to return realistic results and can create NP-hard problems due to the complexity of
scheduling problems [24]. Thus, some hypotheses have been made to simplify it,
such as:
• The tasks done by operators are manufacturing tasks (assembly): this is
crucial to our study since the models used in this study (fatigue/recovery and
learning/forgetting) are based on manufacturing tasks.
• One operator per machine and one machine is operated by one operator.
Production Systems Simulation Considering … 167

Fig. 3 Proposed model with inputs, control parameters and outputs

• Machine cycle times follow deterministic laws, except for maintenance which
follows stochastic laws.
• A station is modeled by an operator, a machine, an input buffer and related
distances.

3.3 Simulation

• Why discrete-event agent-based simulation?

Before simulating the model, this section aims to identify the reasons behind the
choice of simulation, then the choice of Discrete Event and Agent Based Simulation
(DES/ABS). These simulation approaches are often used when simulating human
behavior in Operational Research and similar fields [25]. They give the ability to
model individual behaviors of different parts of the simulation heterogeneously and
enable the creation of populations from these entities to simulate. Due to the nature
of the model, the criteria necessary to simulate it can be written as such:
• Agent-based simulation: to model individual behaviors of populations.
• Discrete-event simulation (can also be continuous simulation): to simulate the
flow of time and dependencies.
• Taking into consideration multiple control parameters in the form of variables
such as distances, resources and desired output.
• Dynamically change these control parameters for eventual optimization.
168 I. Taleb et al.

• Which software?

Multiple tools, software and languages were identified but only 3 were final candi-
dates: AnyLogic and Simio which are object-oriented, multi-paradigm simulation
software. Or, coding a simulation tool from scratch using the coding language Python.
After studying different aspects of these 3 solutions, the software AnyLogic was
chosen due to the benefits it had for this particular example (ABS and DES).
• Simulation:

In this section, the model is implemented into AnyLogic in the form of different
agents: Operators, Stations and the Production Site (Fig. 4). Each agent is modeled
separately, and a population of agents is then generated from it. In this example, the
production site is an agent with a population of 1, which contains 6 operators and 15
stations.
The current goal is to find the optimal strategy to achieve better working conditions
without losing performance. For this example, it translates to lowering the distance

Fig. 4 Class diagram of the

three types of agent used in
the simulation (Factory,
operator and station)
Production Systems Simulation Considering … 169

traveled during a workday while having the same throughput. Considering that bigger
distances mean more time wasted not producing, this is the main goal of the study:
lowering the non-productive time that is unnecessary movement through a better
allocation of tasks for operators.
The model of the operator and the station are also implemented into separate agents
in AnyLogic using statecharts to describe their logics, variables and parameters to
store and change values and statistics to store final results (Fig. 5). The software allows
the creation of functions which enables the possibility to develop more intricate
and complex logics. An example of the functions developed is the “sendRequest()”
function, it activates once the station is ready to receive an operator, thus sending
a request for an available operator nearby to take control of the station. (Notice
“nearby” is an optimization condition in our model, and it is quantifiable). The
flowchart in Fig. 5 illustrates fragments of logics followed by operators and stations
in the simulation.
The accuracy of the model was tested through comparison of simulation results
with real-life results, such as values of total distance traveled, operator productivity.

Fig. 5 Flowchart: Logics of a station agent (left) and an operator agent (right)
170 I. Taleb et al.

Fig. 6 Evolution of peak

Total distance traveled (km)

total distance traveled (km) 22
20
by strategy 18 17

12

Strategy Strategy Strategy Strategy Strategy

#1 #2 #3 #4 #5

3.4 Results

To test the robustness of the model, we made several trials with different input param-
eters, while analyzing the output for any variance (Fig. 6). One of the parameters we
decided to test for robustness are working hours and maximum distance between the
operator and the station, 5 different strategies were agreed upon:
• Strategy #1: Regular simulation (default parameters)
• Strategy #2: 10-h workday and max distance of 20 m
• Strategy #3: 8-h workday and max distance of 10 m
• Strategy #4: 8-h workday and max distance of 5 m
• Strategy #5: 8-h workday and no constraint on distance

We notice that the 3rd strategy returns the best results, with a peak for 12 km total
distance traveled per workday.

4 Conclusion

Using simulation, we were able to better understand the behavior and interactions of
these different key concepts operating in the same environment. In this case study,
we managed to identify a new non-productive time related to unnecessary move-
ment with a better allocation strategy. We were able to prove that waiting in the
current position can be a form of micro-break (margin of maneuver) that can improve
productivity and working conditions at the same time (see strategy #3 vs. strategy
#5). However, if the constraint over distance is too demanding (say <5 m), the effects
become counterproductive (see strategy #3 vs. strategy #4).
In this paper, we were able to identify the main parameters influencing the perfor-
mance and working conditions of an operator in production systems via simula-
tion. We identified the main quantifiable human factors and their behavior, as well
as margins of maneuver which were found to be a buffer between the operator’s
Production Systems Simulation Considering … 171

human factors and the machine’s parameters. Finally, we identified and developed a
taxonomy of non-productive times related to production systems and implemented
them into the simulated model.
We were able to manually find the optimal strategy for better working conditions
without losing performance. Thus, our aim for future works can be summarized as
such:
• Develop a more complex model with multiple objectives.
• Add variability to the model and study its effects through non-productive times
(through fatigue and machine breakdown – not demonstrated in this paper).
• Better implement margins of maneuver by considering buffer sizes and the
production system as a whole.

To summarize, we understand that non-productive times are more subtle than previ-
ously thought, as they can hide as productivity. Additionally, human factors such as
fatigue can be absorbed, mitigated and regulated by margins of maneuver such as
waiting for a better situation (strategy #3 for example). Through simulation, many
different goals and strategies were discussed and considered, from purely economic
goals to highly human-centered ones, finally, we can achieve a compromise of both
economical and ergonomic goals.

References

1. Rivera, L., Chen, F.F.: Measuring the impact of Lean tools on the cost–time investment of a
product using cost–time profiles 23, 684–689 (2007)
2. El Mouayni, I., Etienne, A., Siadat, A., Dantan, J., Lux, A.: AEN-PRO: Agent-based simulation
tool for performance and working conditions assessment in production systems using workers’
margins of manoeuver. IFAC-PapersOnLine 50(1), 14236–14241 (2017)
3. Abubakar, M.I., Wang, Q.: Key human factors and their effects on human centered assembly
performance. Int. J. Ind. Ergon. 69, 48–57 (2019)
4. Phillips, R.O.: What is fatigue and how does it affect the safety performance of human transport
operators? (2014)
5. Petronijevic, J., Etienne, A., Dantan, J.: Human factors under uncertainty: a manufacturing
systems design using simulation-optimisation approach. Comput. Ind. Eng. 127, 665–676
(2019)
6. Digiesi, S., Kock, A.A.A., Mummolo, G., Rooda, J.E.: The effect of dynamic worker behavior
on flowline performance. IFAC Proc. 39(3), 341–346 (2006)
7. Lanfranchi, J.B., Duveau, A.: Explicative models of musculoskeletal disorders (MSD): from
biomechanical and psychosocial factors to clinical analysis of ergonomics. Rev. Eur. Psychol.
Appl. 58(4), 201–213 (2008)
8. El Mouayni, I., Etienne, A., Dantan, J., Lux, A.: Développement d ’ une méthode outillée
pour l ’ amélioration des aspects santé-sécurité dès les premières phases de la conception des
systèmes de production par la prise en compte des marges de manœuvre temporelles (2018)
9. Ebert, P., Freibichler, W.: Nudge management: applying behavioural science to increase
knowledge worker productivity. J. Organ. Des. 6(1), 4 (2017)
10. Ghobakhloo, M.: Industry 4.0, digitization, and opportunities for sustainability. J. Clean. Prod.
252, 119869 (2020)
172 I. Taleb et al.

11. Givi, Z.S., Jaber, M.Y., Neumann, W.P.: Modelling worker reliability with learning and fatigue.
Appl. Math. Model. 39(17), 5186–5199 (2015)
12. Jaber, M.Y., Givi, Z.S., Neumann, W.P.: Incorporating human fatigue and recovery into the
learning-forgetting process. Appl. Math. Model. 37(12–13), 7287–7299 (2013)
13. Kolus, A., Wells, R., Neumann, P.: Production quality and human factors engineering: a
systematic review and theoretical framework. Appl. Ergon. 73, 55–89 (2018)
14. Neumann, W.P., Kolus, A., Wells, R.W.: Human factors in production system design and quality
performance—a systematic review. IFAC-PapersOnLine 49(12), 1721–1724 (2016)
15. El Mouayni, I., Etienne, A., Siadat, A., Dantan, J., Lux, A.: A simulation-based approach
for enhancing health aspects in production systems by integrating work margins. IFAC-
PapersOnLine 49(12), 1697–1702 (2016)
16. El Mouayni, I., Etienne, A., Lux, A., Siadat, A., Dantan, J.: A simulation-based approach for
time allowances assessment during production system design with consideration of worker’s
fatigue, learning and reliability. Comput. Ind. Eng. 139, 105650 (2020)
17. Low, S.N., Chong, S.H., Sim, H.Y., Razalli, S., Kamaruddin, S.: Measurement of overall
performance effectiveness in setup improvement. J. Ind. Eng. 2014, 1–7 (2014)
18. Stamatis, D.H.: The OEE Primer: Understanding Overall Equipment Effectiveness, Reliability,
and Maintainability (2010)
19. Manotas Duque, D.F., Cadavid, L.R.: Lean manufacturing measurement: the relationship
between lean activities and lean metrics. Estud. Gerenciales 23(105), 69–83 (2007)
20. Garre, P., Nikhil Bharadwaj, V.V.S., Shiva Shashank, P., Harish, M., Sai Dheeraj, M.: Applying
lean in aerospace manufacturing. Mater. Today Proc. 4(8), 8439–8446 (2017)
21. Foulloy, L., Clivillé, V., Berrah, L.: A fuzzy temporal approach to the Overall Equipment
Effectiveness measurement. Comput. Ind. Eng. 127, 103–115 (2019)
22. Singh, R., Shah, D.B., Gohil, A.M., Shah, M.H.: Overall Equipment Effectiveness (OEE)
calculation—automation through hardware & software development. Procedia Eng. 51, 579–
584 (2013)
23. Badiger, A.S., Gandhinathan, R.: A proposal: evaluation of OEE and impact of six big losses
on equipment earning capacity. Int. J. Process Manag. Benchmarking 2(3), 234–248 (2008)
24. Bozejko, W., Hejducki, Z., Wodecki, M.: Flowshop scheduling of construction processes with
uncertain parameters. Arch. Civ. Mech. Eng. 19(1), 194–204 (2019)
25. Majid, M.A., Fakhreldin, M., Zuhairi, K.Z.: Comparing Discrete Event and Agent Based
Simulation in Modelling Human Behaviour at Airport Check-in Counter, vol. 1, pp. 510–522
Distributed and Embedded System
to Control Traffic Collision Based
on Artificial Intelligence

Tarik Hajji , Tawfik Masrour , Mohammed Ouazzani Jamil,

Zakaria Iathriouan, Sanaa Faquir, and Elmiloud Jaara

Abstract Traffic accidents are the leading cause of death and serious injury. Many
accidents occur at controlled intersections, such as those with traffic lights. Artificial
intelligence consists of a wide range of computer techniques intended for problems
such as pairing models, language processing and solving very complex and ill-defined
problems. The artificial intelligence situation analyzer can use model analysis of
previous vehicle data and previous signal violation events to determine the decelera-
tion of the threshold, or otherwise determine the probability of a signal violation. In
this article, we present an on-board system to prevent traffic collisions. This system is
based on the extraction of six functionalities which characterize mobile vehicles and
an unsupervised learning approach based on an artificial neural network to classify
functionalities. Our approach began with the description of a motion vector which
describes the state of a moving object. Then, the learning database was designed.
After various computer simulations, an electronic system was designed to integrate
our solution.

T. Hajji (B) · T. Masrour

Laboratory of Mathematical Modeling, Simulation and Smart Systems (L2M3S), Artificial
Intelligence for Engineering Sciences Team, ENSAM-University My ISMAIL, Meknes, Morocco
e-mail: [email protected]
T. Masrour
e-mail: [email protected]
M. Ouazzani Jamil · Z. Iathriouan · S. Faquir
Faculté des Sciences de l’Ingénieur, Université Privée de Fès, Fes, Morocco
e-mail: [email protected]
Z. Iathriouan
e-mail: [email protected]
S. Faquir
e-mail: [email protected]
E. Jaara
Faculté des Sciences d’Oujda, Université Mohammed Premier, Oujda, Morocco
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 173

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_12
174 T. Hajji et al.

Keywords Artificial intelligence · Distributed computing · Embedded system ·

Neural networks · Images and videos processing · Smart cities

1 Introduction

A traffic collision, also called a motor vehicle collision (MVC) among other terms,
occurs when a vehicle collides with another vehicle, pedestrian, animal, road debris,
or other stationary obstruction, such as a tree, pole or building. Traffic collisions
often result in injury, death, and property damage.
In 2013, 54 million people worldwide sustained injuries from traffic collisions
[1]. This resulted in 1.4 million deaths in 2013, up from 1.1 million deaths in 1990
[2] and about 68,000 of these occurred in children less than five years old [2].
Almost all high-income countries have decreasing death rates, while the majority
of low-income countries have increasing death rates due to traffic collisions. Middle-
income countries have the highest rate with 20 deaths per 100,000 inhabitants,
accounting for 80% of all road fatalities with 52% of all vehicles. While the death
rate in Africa is the highest (24.1 per 100,000 inhabitants), the lowest rate is to be
found in Europe (10.3 per 100,000 inhabitants) [3].

1.1 Distributed System

Distributed computing is a field of computer science that studies distributed systems.

A distributed system is a system whose components are located on different
networked computers that communicate and coordinate their actions by passing
messages to each other [4]. The components interact between them in order to achieve
a common goal. Three significant characteristics of distributed systems are: concur-
rency of components, lack of a global clock, and independent failure of components
[4].

1.2 Embedded System

An embedded system is a computer system with a dedicated function within a larger

mechanical or electrical system, often with real-time computing constraints [5, 6]. It
is embedded as part of a complete device often including hardware and mechanical
parts. Embedded systems control many devices in common use today [7]. Ninety-
eight percent of all microprocessors are manufactured as components of embedded
systems [8].
Distributed and Embedded System to Control Traffic … 175

1.3 Neural Networks (ANN)

ANN (Fig. 1) is a problem-solving tool inspired by the biological neural networks that
constitute human brains, and it’s a network of many very simple processors (“units”),
each possibly having a local memory. It’s based on a collection of connected artificial
neurons organized in layers. Most multilayer perceptron networks are composed of
three layers: the input layer, the hidden layer, and the output layer [9].
An artificial neuron (Fig. 2) is a function able to realize a non-linear sum of its
inputs. Most ANN are based on “training” rules where the weights of connections
are adjusted on the basis of the presented patterns.
The Back-Propagation algorithm [10], considered as one of the most popular
learning algorithms, is a method used in ANN to calculate the error contribution of
each neuron after processing a batch of data.

Fig. 1 Structure of an ANN containing one hidden layer (4 neurons), one output layer (2 neurons),
and three input layers

Fig. 2 Artificial neural model

176 T. Hajji et al.

1.4 Related Work

The control of traffic collision based on Artificial Intelligence started long before.
In the paper [11] briefly reviews the current state of development of the artificial
intelligence field and explores how these techniques might be applied to air traffic
control.
In [12] we have an apparatus for controlling a traffic signal at an intersection under-
stood a vehicle sensor providing vehicle data, such as vehicle speed and vehicle posi-
tion, and, optionally, an ambient condition sensor, providing ambient condition data
for the intersection, and a signal controller controlling the traffic signal. The signal
controller includes an artificial intelligence based situational analyzer receiving the
vehicle data and, optionally, ambient condition data and a time signal.
In the paper [13] the authors describe an alternative method based on the use of
artificial neural networks (ANN) in order to work out a model that relates to the
analysis of vehicular accidents in Milan.
The paper in [14] paper focuses on predicting the severity of freeway traffic
accidents by employing twelve accident-related parameters in a genetic algorithm
(GA), pattern search and artificial neural network (ANN) modelling methods.
An implementation of a collision detection system based on a motion detector
(MD) and Artificial Neural Network (ANN) was presented in [15]. The system was
designed to recognize and classify motor vehicle collisions basing on the traffic
cameras video-streams analysis. To test MD and ANN separately, a small set of
video records from traffic cameras were used and results showed that the chosen
implementation of the pretrained ANN was not performant as It often detected small-
scale details like side mirrors, taxi signs, wheels, etc.
Artificial Neural Networks (ANN) modelling with Multiple Layers Perceptron
(MPL) were also used in [16] to classify the severity of road accidents as this method
is recommended for treating non-linear problems, whose distributions are not normal,
which is the case for road accidents.
An automated system for the identification of motor vehicular accidents that uses
artificial neural network approach was developed in [17] to estimate the probability
of occurrence, based on recorded acoustic signals. Results showed that system could
identify actual accidents with 100%. The system was also able to prioritize the time
domain signal features over those of the frequency domain, in the identification
process.

2 Methodology

Our project aims to minimize the rate of road collision in Morocco by a road accident
prediction system based on artificial neural networks. The project analyzes in real
time the data captured by road surveillance cameras whose purpose is to inform road
users of the likelihood of a road collision or disruption in normal road traffic (Fig. 3).
Distributed and Embedded System to Control Traffic … 177

Fig. 3 General architecture of the project

For this reason, an embedded system to prevent road collision was designed. The
general functional specifications of the system are explained in Fig. 4. The system
is composed of three parts:
• Server side: that captures the video, the data processing, the generation of a score
that represents the probability of having an accident and the broadcaste of the
data.
• Client side: in the form of several applications that make the reception and the
interpretation of the signals.
• Data analysis and big data side: which is responsible for the massive collection
of data generated by the system in order to build a predictive map of accidents?

Fig. 4 Functional specifications of the system

178 T. Hajji et al.

Table 1 Learning data base built

MO Mask Trajectory Speed Acceleration Direction Surface Score
1 625 9 145 5 195 97 3
2 47 6 129 −4 223 24 4
3 58 3 58 −3 136 4 1
4 417 7 35 −8 0 73 4
5 75 6 116 6 67 78 6

2.1 Learning Data Base

We have experimentally built a knowledge database. We created a prototype of a

roundabout after extracting six common features (Mask, Trajectory, Speed, Accel-
eration, Direction, and Surface), we used a camera to monitor and simulate moving
objects that we prepared in our laboratory, and then we programmed valid and invalid
configurations. A score that represents the probability from 1 to 7 of having an
accident was also calculated for each moving object (1: good… 7: catastrophique).
Finally, we computed the motion vectors for each moving object. Results are
summarized in Table 1.

2.2 Neuronal Structures

We did a comparative study to find the most efficient neural architecture as shown
in Table 2.
We used Alyuda-NeuroIntelligence [18] for the automatic search of neural archi-
tecture for our prediction model. Alyuda’s neural network software is successfully
used by thousands of experts to solve complex data mining problems, empower
pattern recognition and predictive modeling, build classifiers and neural net simula-
tors, design trading systems and forecasting solutions. Alyuda offers us a high-end
professional neural network software system to get the maximum predictive power
from artificial neural network technology.
We used ANN composed of an input layer with 6 artificial neurons, a single hidden
layer with 3 artificial neurons and an output layer composed of one single artificial
neuron. We used the retro propagation of the gradient algorithm for learning. The
neural structure used is shown in Fig. 5.

2.3 Embedded Prototype

We designed a prototype to do the actual tests of our system. Table 3 contains all the
hardware configuration that we have exploited:
Distributed and Embedded System to Control Traffic … 179

Table 2 Comparative study results

ANN Quadratic error Number of iterations Recognition rate
Learning database (%) Test database (%)
6:3:1 0.01 106 100 97
109 100 99
1010 100 99.8
0.0001 106 100 98
109 100 99.6
1010 100 100
6:2:1 0.01 106 100 97
109 100 99
1010 100 99.8
0.0001 106 100 98
109 100 99.6
1010 100 100
6:3:2:1 0.01 106 100 97
109 100 99
1010 100 99.8
0.0001 106 100 98
109 100 99.6
1010 100 100

Fig. 5 The Ann structure

used

The server embedded is shown in Fig. 6 and the receiver is shown in Fig. 7. We
have developed two types of computer-based simulation applications that analyze
video and a mobile application.
180 T. Hajji et al.

Table 3 Hardware configuration exploited

Product Specification Price ($)
Micro Processor Original Rapberry Pi 3, Modele B, 1 GB, Quad-Cores 53
Screen LCD Raspberry Pi, Official 800*480 100
Cable HDMI/HDMI Pour Raspberry P1 4
Cable FC-26P, Cable Pour Raspberry P1 4
Adapter T Gpio Adapter Board 4
Food 5v/2a Alimentation Micro Usb Raspberry Pi 5
Wireless Nrf24l01 Wireless Transceiver Module + Antenne *2 25
Wireless NRF24L01 Wireless Transceiver Module 3
Camera Raspberry Pi Caméra 5mp 36
Micro Controller Arduino UNO 11
Micro Controller Arduino Mega 16
Viewer LCD 16 4
Viewer Adabptateur LCD I2C 1
Memory Micro SD Card CLASS 10 10
Breadboard Mini Breadboard 400 Points 2

Fig. 6 The server and the receiver embedded prototype

Distributed and Embedded System to Control Traffic … 181

Fig. 7 Example of analyzed

videos

Table 4 Results of analyzed videos

MO Mask Trajectory Speed Acceleration Direction Surface Score
1 625 9 145 5 195 97 4
2 47 6 129 −4 223 24 4
3 20 3 58 −3 136 4 1
4 417 7 35 −8 0 73 4
5 75 6 116 6 67 78 6

3 Validation Test and Results

To test our system, we analyzed videos on the internet that show accidents. Table
4 shows some of characteristics of moving objects captured and Fig. 7 shows an
example of these videos.
The following diagram shows the difference between the results returned by our
system and the real results (Figs. 8 and 9).
Based on these results, we can conclude that with the ANN characterized by: an
input layer composed of 6 artificial neurons, a hidden layer composed of 3 artificial
neurons, an output layer composed of a single artificial neuron, and a quadratic error
average of the order of 10−4 and with an iteration number equal to 109 , we have a
perfect recognition rate equal to 100% tested on the learning database.

4 Conclusion

In this work, we gave the conception of an embedded system able to make the
prevention of road collision. We dealt with automatic detection of undesirable events
by surveillance cameras in the roundabouts of cities in order to propose intelligent,
rapid and effective intervention.
182 T. Hajji et al.

Fig. 8 Results returned by our system

Fig. 9 The real results

We used neural networks as a prediction tool on a learning database and a vali-

dation database. We used a motion vector of 6 characteristics to represent a moving
object in the monitored area. We achieved a recognition rate of more than 98% for
the extraction of the characteristics used and 90% of accident prevented.

References

1. Global Burden of Disease Study 2013, Collaborators (22 Aug 2015): Global, regional, and
national incidence, prevalence, and years lived with disability for 301 acute and chronic diseases
and injuries in 188 countries, 1990-2013: a systematic analysis for the Global Burden of
Disease Study 2013. Lancet 386(9995), 743–800. https://doi.org/10.1016/s0140-6736(15)606
92-4. PMC 4561509. PMID 26063472
Distributed and Embedded System to Control Traffic … 183

2. GBD 2013 Mortality and Causes of Death, Collaborators (17 Dec 2014): Global, regional, and
national age-sex specific all-cause and cause-specific mortality for 240 causes of death, 1990–
2013: a systematic analysis for the Global Burden of Disease Study 2013. Lancet 385(9963),
117–171. https://doi.org/10.1016/S0140-6736(14)61682-2. PMC 4340604. PMID 25530442
3. Global Status Report on Road Safety 2013: Supporting a Decade of Action (PDF) (in English
and Russian). World Health Organization WHO, Geneva, Switzerland (2013). ISBN 978 92 4
156456 4. Accessed 3 Oct 2014
4. Anthony, R.J.: Chapter 6—Distributed systems. In: Anthony, R.J. (eds.) Systems Programming,
pp. 383–474. Morgan Kaufmann (2016). ISBN 9780128007297. https://doi.org/10.1016/B978-
0-12-800729-7.00006-6
5. Barr, M.: Embedded system glossary. Neutrino Technical University. Accessed 21 Apr 2007
6. Heath, S.: Embedded System Design. EDN Series for Design Engineers (zed) newness, p. 2
(2003). ISBN 978-0-7506-5546-0
7. Barr, M., Massa, A.J.: Introduction, Programming Embedded System, with C and GNU
Development tools@’reilly, pp. 1–2 (2006). ISBN 978-0-596-00983-0
8. Barr, M.: Real mean programme in C. Embedded systems design. Tech in sights (business
media), p. 2. Accessed 23 Dec 2009
9. Hajji, T., Itahriouan, Z., Jamil, M.O.: Securing digital images integrity using artificial neural
networks. IOP Conf. Ser. Mater. Sci. Eng. 353(1), 012–016 (2018). https://stacks.iop.org/1757-
899X/353/i=1/a=012016
10. Yu, X., Efe, M.O., Kaynak, O.: A general backpropagation algorithm for feedforward neural
networks learning. IEEE Trans. Neural Netw. 13(1), 251–254 (2002). https://doi.org/10.1109/
72.977323
11. Wolterman, M.: Infrastructure-based collision warning using artificial intelligence. U.S. Patent
No. 7,317,406, 8 Jan 2008
12. Gosling, G.D.: Identification of artificial intelligence applications in air traffic control. Transp.
Res. Part A: Gen. 21(1), 27–38 (1987)
13. Mussone, L., Ferrari, A., Oneta, M.: An analysis of urban collisions using an artificial
intelligence model. Accid. Anal. Prev. 31(6), 705–718 (1999)
14. Kunt, M.M., Aghayan, I., Noii, N.: Prediction for traffic accident severity: comparing the
artificial neural network, genetic algorithm, combined genetic algorithm and pattern search
methods. Transport 26(4), 353–366 (2011)
15. Atoyan, A., Zolotov, O., Romanovskaya, Y.: Application of the motion detector and the artificial
neural network to detect vehicle collisions: a case study. EPJ Web Conf. 224, 04001 (2019).
https://doi.org/10.1051/epjconf/201922404001
16. Chuerubim, M.L., Valejo, A., Bezerra, B.S., da Silva, I.: Artificial neural networks restriction
for road accidents severity classification in unbalanced database. Sigma J. Eng. Nat Sci. 37(3),
927–940 (2019)
17. Murray, R., D’Arbasie, A.: Automated identification of vehicular accidents from acoustic
signals using artificial neural networks. West Indian J. Eng. 41(2), 55–61 (2019)
18. Tarik, H., Jamil, O.M.: Weather data for the prevention of agricultural production with convolu-
tional neural networks. In: 2019 International Conference on Wireless Technologies, Embedded
and Intelligent Systems (WITS), pp. 1–6. IEEE, Apr 2019
The Emergence and Decision Support
in Complex System with Fuzzy Logic
Control

Mohammed Chennoufi and Fatima Bendella

Abstract This paper proposes a new approach for a crowd navigation in evacuation
which is a good example of a complex system. The mathematical concepts such as
differential equations is used in simple system but when the complexity increases,
these concepts become less effective. The processing of complex systems such as
crowd navigation requires the manipulation of vague, imprecise, uncertain, or both
imprecise and uncertain information. The proposed approach use path finding for
static obstacle and perception of the environment with spatial behaviour for dynamic
obstacle, which affects their travel speeds and emerging appeared phenomena. Our
system is modelled by agent and tested by a processing simulation, several modules
such as A* planning, physical factors of agents have been programmed with java
and successfully inserted into a system. The paper presents many simulations for
evacuation of crowd behaviour and emergence of the crowd. Experimental results
show how our agents can plan the best path with any collision in an unknown envi-
ronment. Our application can be used as a framework to simulate real situation of
complex system.

Keywords Emergence · Decision support · Complex system · Fuzzy controller

1 Introduction

This work concerns the microscopic level of a complex system and more precisely the
interactions between crowd individuals in evacuation. The collision avoidance is a
primordial phase in the interaction individuals-environment, whether the avoidance

M. Chennoufi (B)
Laboratory LISIDD, Institute of Maintenance and Industrial Safety,
IMSI, University of Oran 2, Oran, Algeria
e-mail: [email protected]
F. Bendella
Department of Computer Science, University of Science and Technology of Oran USTO,
Oran, Algeria
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 185

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_13
186 M. Chennoufi and F. Bendella

of static (wall) or dynamic obstacles (neighborhood individuals). We distinguish

several microscopic models of collision avoidance such as social forces [12], rule-
based models [21] and cellular automata models [2]. The difference between them
is in the discretization of space and time. A (virtual) social force is analogous to
real force such as repulsive interaction, friction force, dissipation, solving motion
Newton’s equations for each individual. In the rule-based model, the displacement
of the crowd is governed by behavioral rules of the form “if condition then action”.
Recently there has been a lot of work on crowds and their behavior as models that
check the characteristics of complex systems [21, 25]. We proposed a collision
avoidance algorithm based on Newtons forces and inspired by the Reynolds model
[21], it’s based on simple rules with the addition of several criteria. After seeing to
plan the way of our individuals’ crowd towards the arrival destination by avoiding
the static obstacles by a simple A* algorithm [9, 26] and fuzzy system for dynamic
obstacles. We launch our simulation of crowd movement by activating the module
of collision avoidance, the latter is based on a local perception of the environment
(Angle of vision, comfort distance, density) and a speed of movement (desired speed).
Each individual in the course of his or her path realizes one of these interactions to
follow, flee or avoids a dynamic collision if it exists by decreasing or increasing its
speed and or by a slight deviation to the right or to the left then a return towards
its planning. This one will influence instantly on its initial trajectory, for helping the
decision taken individually by the agents to choose the best path, we use the fuzzy
logic controller with different inputs factors. The case study of crowd navigation is
a good example because it responds to the characteristics of a complex system. For
example, if an initial parameter at the microscopic level between people in the crowd
is not respected “crossing, collision, distance” it will influence at the macroscopic
level, resulting in the emergence of unexpected behavior “panic, queueing …” which
is difficult to predict without recourse to simulation.
The rest of the article is organized as follows: the second section describes a
previous work while emergence and chaotic of complex system is presented in the
section three. The fourth section presents the proposed model. In section five we
present a simulation of our model with discussions of emerging phenomena. Finally,
we present conclusions and future works.

2 Related Works

Helbing et al. [13] proposes a model in which agents move in the calculated direction
on the basis of the equation of generalized Newton force. This technique enables to
realistically simulate the fundamental community of the effects such as the formation
of routes and bottlenecks. Other mathematical models and special cases exist in
the literature inspired by the work of Helbing. Other works on the simulation of
evacuation of crowds based on cellular automata. Schadschneider [24] proposes that
the interactions are based on the concept floor field where individuals follow the
virtual traces field by other individuals. In the same microscopic axis there is the
The Emergence and Decision Support in Complex System with Fuzzy Logic Control 187

agent-based approach where the characteristics of an autonomous agent may help in

well modeling the reality. Shao and Terzopoulos [25] and Ji and Gao [14] propose an
evacuation of a dance hall where the crowd is made up of several groups dynamically
led by chiefs. The goal is to see the effect of the number of agents on the effectiveness
of the evacuation in case there are a lot of leaders, the evacuation may take more time
than if it is done without a guide, and this is what increases the risk of an imbalance
towards chaos. Pillac et al. [20] presents a path generation approach based on conflicts
for the evacuation planning.
Abdelhak et al. [1] added emotions as cognitive characteristics of agents to the
behavior of crowds. Lin et al. [15] recommends an adaptive distributed guiding pro-
tocol which takes several factors such as dangerous areas, the distance to the exits,
their width, and the degree of each location by considering the congestion. The goal
is to help the moving objects in order to avoid a dangerous region. The simulation
results show that the proposed protocol can guide mobile objects (people) for a fast
exit. He et al. [11] presents a complete analysis of logic models of evacuation crowd
behavior as well as the interactive effects of information (objective and subjective)
and the study of the psychology. Olya [18] propose a dynamic program to find the
shortest path in a network having exponential and gamma probability distributions
as arc lengths, its objective is to extend the shortest path problem in dynamic and
stochastic networks where link travel times are defined as gamma or exponential
probability distributions. In the same axis [17] present a general treatment of shortest
path problems by dynamic program in networks having normal probability distri-
butions as arc lengths by extension the general shortest path problem in dynamic
and stochastic networks where link travel times are defined as normal probability
distributions. Lu et al. [16] simulate the crowd evacuation phenomenon through the
simulation of the artificial fish improved speed. The methodology used by Samah
et al. [22] is the modification of Dijkstra algorithm to find the most sure and short
way. The first modification implies the restriction of the direction of the node or
plan, and the second modification blocked the nodes which are touched by the fire.
In El-Hadidy and El-Bagoury [7] a new search model that finds a three-dimensional
randomly located target in a known zone by a single searcher has purposed to obtain
an optimal search strategy that minimises the expected value of the time required
to detect the target. Several models have been developed to simulate the evacuation
either in the microscopic or macroscopic axis but little work exists on the path plan-
ning of a dense panicked crowd. In the same axes [3] presented a collision avoidance
technique for a crowd robust navigation of individuals in evacuation. His algorithm
is inspired from the Reynolds model.

3 Emergence and Chaotic of Complex System

Complexity theory is used for modeling and analyzing complex systems in different
fields from computer science to ecology and engineering. Complexity is a funda-
mental characteristic of our world that does not lend itself to our traditional methods
of modern science.
188 M. Chennoufi and F. Bendella

Fig. 1 Microscopic and

macroscopic level

Complex systems are also seen as non-linear systems whose skills from interac-
tions between components exceed the sum of the individual contributions of each
component [10].
The most important thing about these systems is that they seem to have nothing
in common. But, they often share structural models and means of similar [5].
Two levels are distinguished in complex systems:
– A microscopic level (low level) which represents the local properties between the
components of the system.
– A macroscopic level (high level) which represents the whole system with emer-
gence of the new properties (Fig. 1).
A complex system is characterized by non-linearity, interactions, self-
organization, robustness, emergence and chaos. A system is linear if we add or
multiply factors to equations we get another system, unlike the non-linearity where
the change of the output is not proportional to the change of the input (no explicit
solutions). Studying interactions is difficult because they can be direct, indirect, with
or without feedback. In the case of direct interactions it is difficult to follow these
sequences over time between the entities of the system, imagine if we add the indi-
rect case as well as the dynamics of the environment which becomes more and more
complex and difficult. Different levels of interaction exist: interactions between sys-
tems of the same level and interactions between different levels of systems. For the
robustness and self organization are sets of mechanisms that lead to the emergence
of an emergence property at the global (macroscopic) level of the system through
interactions between the elements of the finer (microscopic) levels. In mathematics,
chaos theory is defined as systems that are incapable of establishing complicated
rules and following simple laws. This theory studies the behaviour of dynamic sys-
tems that are very sensitive to initial conditions. It only occurs in non-linear systems.
The main characteristics of chaotic systems are:
– Chaos results from a deterministic process.
– It occurs only in non-linear systems.
– It occurs in retroactive systems; Where events of the past affect current events and
current events affect events of the future.
The Emergence and Decision Support in Complex System with Fuzzy Logic Control 189

Fig. 2 Example of complex system a emergence of pedestrian, b emergence of fish

– The details of the chaotic behavior are hypersensitive to the change of the initial
conditions.
– It can occur from relatively simple systems:—with discrete time.
– Information on initial conditions is irretrievably lost.
It is not yet possible to determine in advance the particular path that the dynamic
process will follow in order to move towards chaos. Emergence is the process of deriv-
ing new (unexpected) behaviors, structures or patterns. The latter appearing in time
through simple interactions between entities (see Fig. 2a) and it is this phenomenon
that distinguishes a complex system of a complicated system. These emergent prop-
erties are not visible at the microscopic level. They are chaotic systems and occur
only in non-linear systems. It is not yet possible to determine in advance the particular
path that the dynamic process will follow in order to move towards chaos.

4 The Proposed Model

Avoiding collisions is a crucial phase in the paths of individuals, the detection of

collisions will depend on the shape and size of fixed or dynamic obstacles. When
making an attempt to avoid an obstacle, there are many decisions to make. The latter
ones depend on the available paths, finding the best and the shortest path depends on
the algorithms used.
This section describes our approach of autonomous agent and its behaviors with
the group of its neighbors to provide emergence phenomena. We have modeled our
complex system (the crowd) on an agent-based system at the microscopic level, and
study the emergence at the macroscopic level. Each agent perceives its local environ-
ment (the input of fuzzy controller), the action performed by the agent which is the
force in order to avoid any type of collision by separations, cohesion and at the same
time to go towards a path. An instant update for speed, force and acceleration, which
190 M. Chennoufi and F. Bendella

Fig. 3 Proposed model

allows changement of agent position (rules of fuzzy controller). Figure 3 illustrates

the modeling and the simulation of proposed model. In the modeling part we observe
4 steps:
– The local perception of the environment for each agent.
– The action performed by the agent which is the force in order to avoid any type of
collision by separations, cohesion and at the same time to go towards a path.
– An instant update for speed, force, acceleration and deceleration, which allows the
change agent position (rules of fuzzy controller).
– Decision support to control our system to not switch into chaos (output of fuzzy
controller).
The simulation part contains the Java implementation of a set of autonomous agents
and their local interactions as well as the appearance of the emerging phenomena.
The Emergence and Decision Support in Complex System with Fuzzy Logic Control 191

4.1 Mathematical Modeling of the Problem

In our approach we consider that each agent is represented by a circle Fig. 4 with:

– Position (X, Y ): is the spatial location in X ,Y coordinates of the agent in the virtual
space.
– Comfort distance (d): is the minimum distance between agents and obstacle to
avoid collision.
– Distance goal (dg): is the minimum distance between agent and goal.
– Risk zone (R I D): it covers a comfort distance d.
– Comfort zone (C O D): it covers a d/2.
– Radius of vision (α): Custom domain of vision.
– Angle of deviation (θ ): the agent is steering with angle θ to the path, its represents
the angle between position of the agent and the goal.
– Radius of comfort R = (2 ∗ d ∗ tan g(α/2)
– Speed (v): the speed varies with the state of health and mobility, ranging from 0
(when the health or mobility = 0) at 4 m/s (when running in panic) Figs. 4 and 5.

4.2 Fuzzy Logic for Crowd Control

Fuzzy logic can be seen as an extension of Boolean logic; moreover, it makes it

possible to treat linguistic variables whose values are words or expressions of the
natural language. In this section, we present the basic properties of classical sets and
fuzzy set [8].

Fig. 4 Agent planning in part of his path

192 M. Chennoufi and F. Bendella

Fig. 5 Illustration fuzzy of collision avoidance with input distance and angle

In a large scale, Fuzzy Logic [27] is ideal for modeling and controlling a crowd
movement for collision avoidance. This fuzzy is used to represent an imprecise
value between true and false. The input variables of the regulator are transformed into
linguistic variables with the definition of membership functions. Thus, this operation
consists of determining the degree of value to a fuzzy set as the work of [6, 23].

4.2.1 Inputs and Outputs from Fuzzy Controller

Figures 6 and 7 shows The fuzzy set of the environment. The input variables of the
fuzzy rules for collision avoidance are distance d and angle α which are respectively
decomposed into three and five fuzzy sets. In our approach, we observe three distances
for d (ZER: zero distance), (COD: comfort distance), (RID: risk distance) and five
angles from vision α (SMV: small vision), (GRV: Great vision), (ZERO: zero vision),
(NSMV: negatives small vision), (NGRV: negative great vision). The distance has a
value between 0 and ∞.
In addition, considering the relation between the agent and the goal the distance
information between the agent and the goal dg and an angle between the targets θ
are added into state space. We observe two distances for dg (N: near), (F: far) and
three angle for θ (Nega: Negative), (Z: Zero), (Pos: Positive).
The output variables speed and danger degree which are respectively decomposed
into three fuzzy sets: (low, Average and high) from speed with value between 0 and 4
and (low, small and high) from danger deviation which has a value between 0 and 5.
After that the rules have been evaluated, we use an aggregation centre-of-gravity
function for velocity and rotation to merge the outputs Fig. 8.
The Emergence and Decision Support in Complex System with Fuzzy Logic Control 193

Fig. 6 The controller input variable distance

Fig. 7 The controller input variable distance α

4.3 Planning and Decision Support

Decision support is used to control the crowd to not switch into chaos with the out-
put of the fuzzy controller. Each individual is modeled by a physiological cognitive
autonomous agent as personality, goals to be accomplished and existing social force
relations with other agents. When an agent acquires new information (By partial
perceptions of the environment), it automatically updates its basic beliefs. The plans
identify strategies that the agent can follow. We used five factors of personality vector
of OCEAN model [4] (Extrovert, Introvert) (agreeableness, unagreeableness) (Con-
scientiousness, unconscientiousness) (Neuroticism, Emotional stability) (Openness
to Ex[H]perience, Closed to experience), these factors range from [−1, 1]. For emo-
tions, we chose the emotional theory OCC defined by [19]. At every moment the
agent feels a mood that reflects its emotional state over time. Figure 8f illustrates the
194 M. Chennoufi and F. Bendella

Fig. 8 Fuzzy input set membership functions for classifying the distance d (a), angle α (b), distance
dg (c) and angle θ (d) to the nearest obstacle in fuzzy terms, fuzzy output value for obstacle avoidance
and path finding, desired speed (e) and danger deviation (f)

Fig. 9 Surface viewer control

The Emergence and Decision Support in Complex System with Fuzzy Logic Control 195

decision taken by the agent to choose the best path based on its partial perception of
the environment and the spread of panic by the neighboring agents. The agent will
replace the initial destination with a new one (Fig. 9).
In general, the decision process encompasses all activities between the time a
stimulus for action is perceived, and when the commitment to action is performed.
The agent perceives its environment (a panic or a dense crowd danger at its path),
consults its knowledge base, thus the choice of route is taken in a list of exits stored
in an order of the closest distance at each node.

5 Implementation and Results

Java language and Processing are used to implement our model on a machine core
I7 and 8 GB of RAM. We present in this virtual environment:
– Static obstacles: modeled by red triangles that agents must avoid.
– Dynamic obstacles which represent the autonomous agents modeled by grey cir-
cles.
– A best path with an exit.
– Interaction between agents with red lines.
After several experiments, we obtain these results, Adding the module of cohesion,
the emergence begins to produce with separation and cohesion. Figure 10 represents
the activation of the local interaction module modeling with a red line. One notices
well the collision avoidance between the near neighbor and from a certain distance
there is not interaction.

Fig. 10 The local interaction between agents

196 M. Chennoufi and F. Bendella

Fig. 11 Emerging phenomena

The proposed system is chaotic because it is very sensible to the initial parameters
taking the case of the speed = 4 and a collision force of 0.1. Figure 11 illustrates
different representatives with 20, 60 and 100 agents and reproduces emerging phe-
nomena as line and group formation with avoided static and dynamic obstacles.

6 Conclusion

In this paper, we proposed an emerging phenomena model of autonomous agents

and decision making of complex system. The proposed model contains modeling
and simulation of the autonomous agents in evacuations based on Newton’s force
for collision avoidance and best A* algorithm for path planning. This architecture
allowed the appearance of emerging phenomena without our system they switch to
chaos. Our approach is improved and controlled with fuzzy logic allows the pedes-
trian to move in an environment without collision with optimization inputs and out-
puts rules. This model has various potential extensions to be explored in the future
work. It can be integrated with the physiological factors and real environment. This
The Emergence and Decision Support in Complex System with Fuzzy Logic Control 197

extension has the potential to simulate other interesting crowd behaviors such as
panic. We are interested in the spatial part of the environment and thinking a new
model of collision learning.

References

1. Abdelhak, H., Ayesh, A., Olivier, D.: Cognitive emotional based architecture for crowd simu-
lation. J. Intell. Comput. 3(2), 55–66 (2012)
2. Chenney, S.: Flow tiles. In: Proceedings of the 2004 ACM SIGGRAPH/Eurographics Sympo-
sium on Computer Animation, pp. 233–242. Eurographics Association (2004)
3. Chennoufi, M., Bendella, F., Bouzid, M.: Multi-agent simulation collision avoidance of com-
plex system: application to evacuation crowd behavior. Int. J. Ambient Comput. Intell. (IJACI)
9(1), 43–59 (2018)
4. Costa, J., Paul, T.: Toward a new generation of personality theories: theoretical contexts for the
five-factor model. In: The Five-Factor Model of Personality: Theoretical Perspectives, vol. 51
(1996)
5. Couture, M., Valcartier, D.: Complexity and Chaos-State-of-the-Art; Overview of Theoretical
Concepts. Minister of National Defence, Canada (2007)
6. Deepak, D., John, S.J.: Information systems on hesitant fuzzy sets. Int. J. Rough Sets Data
Anal. (IJRSDA) 3(1), 71–97 (2016)
7. El-Hadidy, M.A.A., El-Bagoury, A.A.A.H.: Optimal search strategy for a three-dimensional
randomly located target. Int. J. Oper. Res. 29(1), 115–126 (2017)
8. Glorennec, P.Y.: Algorithmes d’apprentissage pour systèmes d’inférence floue. Hermes (1999)
9. Hart, P.E., Nflsson, N., Raphael, B.: A formal bas# s for heunsuc deterrmnatmn of mmmmm
path cost. IEEE Trans. SSC 4, 100–107 (1968)
10. Hassas, S.: Systèmes Complexes à Base de Multi-agents Situés. University Claude Bernard
Lyon (2003)
11. He, G.G., Yang, Y., Chen, Z.H., Gu, C.H., Pan, Z.G.: A review of behavior mechanisms and
crowd evacuation animation in emergency exercises. J. Zhejiang Univ. Sci. C 14(7), 477–485
(2013)
12. Helbing, D., Farkas, I., Vicsek, T.: Simulating dynamical features of escape panic. Nature
407(6803), 487 (2000)
13. Helbing, D., Farkas, I.J., Molnar, P., Vicsek, T.: Simulation of pedestrian crowds in normal and
evacuation situations. Pedestr. Evacuation Dyn. 21(2), 21–58 (2002)
14. Ji, Q., Gao, C.: Simulating crowd evacuation with a leader-follower model. IJCSES 1(4), 249–
252 (2006)
15. Lin, C.H., Chen, P.Y., Chen, W.T.: An adaptive guiding protocol for crowd evacuation based
on wireless sensor networks. In: 2013 IEEE 77th Vehicular Technology Conference (VTC
Spring), pp. 1–5. IEEE (2013)
16. Lu, D., Zhang, G., Liu, Y., Wang, D., Liu, H.: AFSA based path planning method for crowd
evacuation. J. Inform. Comput. Sci. 11(11), 3815–3823 (2014)
17. Olya, M.H.: Applying Dijkstra’s algorithm for general shortest path problem with normal
probability distribution arc length. Int. J. Oper. Res. 21(2), 143–154 (2014)
18. Olya, M.H.: Finding shortest path in a combined exponential-gamma probability distribution
arc length. Int. J. Oper. Res. 21(1), 25–37 (2014)
19. Ortony, A., Clore, G.L., Collins, A.: The Cognitive Structure of Emotions. Cambridge Univer-
sity Press (1990)
198 M. Chennoufi and F. Bendella

20. Pillac, V., Van Hentenryck, P., Even, C.: A conflict-based path-generation heuristic for evacu-
ation planning. Trans. Res. Part B: Methodol. 83, 136–150 (2016)
21. Reynolds, C.W.: Flocks, herds and schools: a distributed behavioral model. In: ACM SIG-
GRAPH Computer Graphics, vol. 21, pp. 25–34. ACM (1987)
22. Samah, K., Hussin, B., Basari, A.S.H.: Modification of Dijkstra’s algorithm for safest and
shortest path during emergency evacuation. Appl. Math. Sci. 9(31), 1531–1541 (2015)
23. Sarkar, M., Banerjee, S., Hassanien, A.E.: Evaluating the degree of trust under context sensitive
relational database hierarchy using hybrid intelligent approach. Int. J. Rough Sets Data Anal.
(IJRSDA) 2(1), 1–21 (2015)
24. Schadschneider, A.: Cellular automaton approach to pedestrian dynamics-theory (2001).
arXiv:cond-mat/0112117
25. Shao, W., Terzopoulos, D.: Autonomous pedestrians. Graph. Models 69(5–6), 246–274 (2007)
26. Sun, X., Koenig, S.: The fringe-saving A* search algorithm-a feasibility study. IJCAI 7, 2391–
2397 (2007)
27. Zimmermann, H.J.: Fuzzy Set Theory—and Its Applications (1991)
Markov Decision Processes with
Discounted Costs over a Finite Horizon:
Action Elimination

Abdellatif Semmouri and Mostafa Jourhmane

Abstract Markov Decision Processes (MDPs) are a popular decision model for
stochastic systems. Since the computational complexity is an open problem, all
researchers are interesting to find methods and technical tools in order to solve the
proposed problem. In this paper, we develop the backward induction algorithm to
calculate optimal policies and value functions for solving finite horizon discrete-time
MDPs in the discounted case. Firstly, we present the backward induction algorithm
for solving Markov decision problem employing the total discounted expected cost
criterion over a finite planning horizon. Second, we propose a new test to iden-
tify non-optimal decisions in the same context. Combining the preceding presented
results, we give an efficient algorithm by linking the recursive approach and the
action elimination procedures. It is a powerful analytical tool used for sequential
decision making under uncertainty that have been widely used in many industrial
manufacturing, financial fields and artificial intelligence.

Keywords Discounted Markov decision processes · Induction algorithm ·

Suboptimal action · Optimization

1 Introduction

Action elimination procedures have undergone expanded research in recent years

as a considerable interest in Markov decision problems (MDPs) that were applied
in modeling of many situations. Among those areas we recall Finance, Engineer-
ing, Statistics and Artificial Intelligence. In this literature, several tests have been
proposed by determining upper and lower bounds on the discounted criterion over
infinite horizon. The first works in this background, have been began by MacQueen

A. Semmouri (B) · M. Jourhmane

FST, Sultan Moulay Slimane University, Beni Mellal, Morocco
e-mail: [email protected]
M. Jourhmane
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 199

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_14
200 A. Semmouri and M. Jourhmane

(1978, 1979) and Porteus (1971). Afterwards, It followed a wealth of results by

other authors such that Grinold (1973), Hastings and Mello (1967), Puterman (1971),
Sadjadi and Bestwick (1973), White (1973) and Semmouri and Jourhmane (2019).
Despite of these considerable progresses, the results concerning the action elimina-
tion are almost rare on the side the discounted criterion over finite planning horizon.
In this manuscript, we will establish an efficient algorithm which is constructed based
on the combination of the forward algorithm that is a part of dynamic programming
and a new test of sub-optimality. For incorporating our work into the previous con-
tributions, we will show its novelty and then its scientific character. Next, we will
give an illustrated numerical example.
The structure of this paper is as follows. Section 1 gives a brief historical overview
of previous works in action elimination procedures. In this regard, several tests were
provided on discounted reward over infinite horizon. In Sect. 2, we give some relevant
definitions and concepts about Markov decision process model over a finite horizon.
Next, Sect. 3 contains the main results in our work. First, we formulate backward and
forward induction algorithms in finite horizon to compute optimal discounted cost.
Second, we provide a new test to eliminate non-optimal actions and we establish
the corresponding algorithm with numerical examples. Finally, the conclusion is
presented in Sect. 4.

2 Preliminaries

2.1 Sample Space, Policies and Measures

At finite time points t = 1, 2, . . . , T , where T is a given integer, a system is observed

by a decision maker to be in one of the states of a finite state space S under the state
process {X t , t = 1, . . . , T } which takes values in S. At each epoch t, the decision
maker controls the system by choosing an action a from the finite action space A. If
the action a is selected in state i, the following happens independently of the history
of the process at time point t: an immediate cost ct (i, a) is lost and the state of the
t
system at the next  decision time point is state j with transition probability pia j , such
that pia j ≥ 0 and j∈S pia j = 1.
t t

Hence, the sample space is given by Ω = {S × A}T , so that the typical realization
ω can be represented as ω = (x1 , a1 , x2 , a2 , . . . , x T , aT ). The state and action random
variables X t , At for t = 1, 2, . . . , T are then defined as the coordinate mappings
X t (ω) := xt and At (ω) := at . Throughout, The sample space Ω will equipped with
the σ-algebra B generated by the random variables (X t , At , t = 1, 2, . . . , T ).
In order to give a formal definition of a policy, first let Q be the set of all probability
measures on the action space A, i.e.:

Q := {(q1 , q2 , . . . , q|A| ) : q1 + q2 + · · · + q|A| = 1, qi  0, 1  i  |A|

Markov Decision Processes with Discounted Costs … 201

where |A| is the cardinality of A. Then a policy π is defined to be a sequence

π = (π 1 , π 2 , . . . , π T ) where π t is a mapping from {S × A}t−1 × S to Q. We write
πit , 1  i  |A|, for the ith component of π t .
For a fixed policy π and initial state i, we can now construct the probability
measure Piπ for the measurable space (Ω, B). The finite-dimensional distributions
of the probability measure Piu are defined as follows

Piπ (X 1 = x) = 1; (1)

Piπ (At = a/ X 1 = x1 , A1 = a1 , . . . , X t−1 = xt−1 , At−1 = at−1 , X t = xt )

= π t (x1 , a1 , . . . , xt , at , xt ); (2)

Piπ (X t+1 = y/ X 1 = x1 , A1 = a1 , . . . , X t−1 = xt−1 , At−1 = at−1 , X t = x, At = a)

= pxay . (3)

where pxay is the law of motion, which is given and determined from the physical
of the problem.
In order to indicate the dependence of the probabilities on the policy π, the notation
Pπ (B) will denote the probability of an event B occurring when policy π is used.
From a standard application of the Kolmogorov consistency theorem, we know there
exists a unique probability measure Piπ , on (Ω, B) such that (1), (2) and (3) hold for
all possible histories and all t  1 where Piπ (B) = Pπ (B/ X 1 = i). Thus, for each
policy π and initial state i, we have constructed a probability space (Ω, B, Piπ ).
Let Π the set of all policies. In terms of our model, we will be interesting in
a particular class of policies that have very important stochastic properties. The
decision rule that it depends on previous system states and actions only through the
current state of the system, it is said to be Markovian (memoryless). The set of all
Markovian policies are denoted by Π M . Among the policies of Π M , each decision
rule that specifies an action with certainty for each current state is called deterministic
(or pure). Their set is denoted by Π M D .
With our definitions we have

ΠMD ⊂ ΠM ⊂ Π

For t = 1, 2, . . . , T , it can be easily seen that a decision rule π t defines a proba-

bility transition matrix
P(π t ) = [ p( j/i, π t )]i, j∈S

with entries given by 

p( j/i, π t ) = t
pia j πi,a
t

a∈A(i)
202 A. Semmouri and M. Jourhmane

Fig. 1 Markov decision problem model


j = 1 for all i ∈ S and a ∈ A(i), P(π ) is a stochastic matrix (all rows
t t
Since j∈S pia
sum to 1) that uniquely defines a Markov chain.
A discrete-time Markov Control Model (MCM) consists of a tuple
 

S, A = A(i), Q = { pt ( j/i, a)}, c = {ct (i, a)}
i∈S

The MCM is interpreted as representing a controlled stochastic system which is

observed at times t = 1, 2, . . . , T (Fig. 1).
Using the decision rule π t at stage t, we define the immediate expected cost
column vector by
c(π t ) := [c(i, π t )]i∈S

where, for each i ∈ S, 

c(i, π t ) := ct (i, a)πi,a
t

a∈A(i)

2.2 Total Discounted Expected Cost in Finite Horizon

Definition 1 For all policy π ∈ Π and initial state i ∈ S, we define the total γ-
discounted expected cost over a finite horizon after T periods by


T
CγT (i, π) := γ t−1 Eiπ [c(X t , At )] (4)
t=1

where c(., .) is the cost function.

Markov Decision Processes with Discounted Costs … 203

Obviously, we have


T  
CγT (i, π) = γ t−1 Piπ (X t = j, At = a)ct ( j, a)
t=1 j∈S a∈A( j)

Here T is a finite integer that represents the number of horizons and γ is a positive
constant that represents a discounted factor (0 ≤ γ ≺ 1).
By writing γ = 1+r 1
, the nonnegative number r is interpreted as a compound
interest rate in the market.
Interchanging the finite summation and the expectation in (4), the total discounted
expected cost over the planning horizon may also be defined as the expected total
γ-discounted cost  T 

Cγ (i, π) = Eπ
T i
γ c(X t , At )
t−1

t=1

In vector notation, the value vector CγT (π) = (CγT (i, π))i∈S becomes


T
CγT (π) = γ t−1 P(π 1 )P(π 2 ) . . . P(π t−1 )c(π t ), for π = (π 1 , π 2 , . . . , π T ) ∈ Π
t=1

where P(π 0 ) := I and I is the identity matrix of size n.

Definition 2 (i) The optimal value function for the discounted criterion is defined
by
CγT (i):= min CγT (i, π); i ∈ S
π∈Π

(ii) A policy π  is said to be T -optimal if

CγT (i, π  ) = CγT (i) for all i ∈ S

Since the solution of a finite-horizon MDP depends on the optimality equation

⎧ ⎫
⎨  ⎬
C t (i) = min ct (i, a) + γ t
pia jC
t+1
for t = 1, 2, . . . , T and
a∈A(i) ⎩ ⎭
j∈S

C T +1 (i) = 0, ∀i ∈ S,

an optimal policy has the property that whatever the initial state and decision are, the
remaining decisions constitutes an optimal policy with regard to the state resulting
from the first decision (Fig. 2).
204 A. Semmouri and M. Jourhmane

Fig. 2 MDP and artificial

intelligence

3 Results and Computational Examples

3.1 Backward Induction Algorithm

The backward induction algorithm is an iterative procedure. Starting at the end of

the planning horizon one computes the value for the previous periods. Effectively,
this method employs a recursion formula by C t can be expressed in terms of C t+1
for t = T, T − 1, . . . , 1. Further, an optimal policy is obtained after T iterations.
Lemma 1 For any starting state i and any policy π ∈ Π , there exists a Markov
policy π̃ ∈ Π M , such that
(i) Piπ̃ (X t = j, At = a) = Piπ (X t = j, At = a), t ∈ N, j ∈ S, a ∈ A( j)
(ii) Eiπ̃ [ct (X t , At )] = Eiπ [ct (X t , At )], t ∈ N
where N denote the set of all nonnegative integer.

Proof From Corollary 2.1 (pp. 22) in Feinberg and Shwartz (2002), it follows the
assertion (i). Using (i), we obtain
 
Eiπ̃ [ct (X t , At )] = Piπ̃ (X t = j, At = a)ct ( j, a)
j∈S a∈A( j)
 
= Piπ (X t = j, At = a)ct ( j, a) = Eiπ [ct (X t , At )] 
j∈S a∈A( j)

Theorem 1 Let CiT +1 = 0, i ∈ S. Let for t = T, T − 1, . . . , 1 consecutively, respec-

tively a deterministic decision π ∗t and a vector C t ∈ R|S| can be defined as
⎧ ⎫
⎨  ⎬
{C(π ∗t )}i + γ{P(π ∗t )C t+1 }i = min ct (i, a) + γ t
pia t+1
j Cj
a∈A(i) ⎩ ⎭
j∈S
Markov Decision Processes with Discounted Costs … 205

where Vi is the ith component of the vector V ∈ R|S| and

C t = c(π ∗t ) + γ P(π ∗t )C t+1

Then, π ∗ = (π ∗1 , π ∗2 , . . . , π ∗T ) is an optimal policy and C 1 is the corresponding

value vector CγT .

Proof We use induction on T . Let π = (π 1 , π 2 , . . . , π T ) be an arbitrary policy.

T = 1:
  
Cγ1 (i, π) = Piπ (X 1 = j, A1 = a)c1 ( j, a) = c1 (i, a)πi,1 j
j∈S a∈A( j) a∈A( j)


≥ [ min c1 (i, a)] πi,1 j = min c1 (i, a) = Ci1 = Cγ1 (i, π ∗ ), i ∈ S
a∈A(i) a∈A(i)
a∈A( j)

Assume that the result has been shown for T = 1, 2, . . . , t. Take an arbitrary state i ∈
S. From Lemma 1 it follows that there exists a Markov policy π̃ = (π̃ 1 , π̃ 2 , . . . , π̃ t )
such that
Cγt+1 (i, π̃) = Cγt+1 (i, π)

Define the Markov policy ρ = (ρ1 , ρ2 , . . . , ρt ) by ρkj,a = π̃ k+1

j,a for k = 1, 2, . . . , t.
From the induction assumption it follows that Cγ ( j, ρ) ≥ C j , j ∈ S, the fact that for
t 2

a planning horizon of t + 1 periods, C 2 is the same as C 1 for a planning horizon of

t periods. Hence,
⎧ ⎫
 ⎨  ⎬
Cγt+1 (i, π) = Cγt+1 (i, π̃) = π̃i,a
1
c1 (i, a) + γ 1
pia j Cγ ( j, ρ)
t
⎩ ⎭
a∈A(i) j∈S
⎧ ⎫
 ⎨  ⎬
≥ π̃i,a
1
c1 (i, a) + γ 1
pia j Cj
2
⎩ ⎭
a∈A(i) j∈S
⎧ ⎫
⎨  ⎬ 
≥ min c1 (i, a) + γ 1
pia j Cj
2
π̃i,a
1
a∈A(i) ⎩ ⎭
j∈S a∈A(i)
⎧ ⎫
⎨  ⎬
= min c1 (i, a) + γ 1
pia j Cj
2
= Ci1
a∈A(i) ⎩ ⎭
j∈S
206 A. Semmouri and M. Jourhmane

On the other hand

C 1 = c(π ∗1 ) + γ P(π ∗1 )C 2 = c(π ∗1 ) + γ P(π ∗1 )[c(π ∗2 ) + γ P(π ∗2 )C 3 ]

= c(π ∗1 ) + γ P(π ∗1 )c(π ∗2 ) + γ 2 P(π ∗1 )P(π ∗2 )C 3
= c(π ∗1 ) + γ P(π ∗1 )C 2 + γ 2 P(π ∗1 )P(π ∗2 )C 3
= ...............

t+1
= γ k−1 [P(π ∗1 )P(π ∗2 ) . . . P(π ∗(k−1) )c(π ∗k )] = Cγt+1 (π ∗ )
k=1

Next, Cγt+1 (i, π ∗ ) = Ci1 ≤ Cγt+1 (i, π) for all i ∈ S, π ∈ Π . Therefore, π ∗ is an

optimal policy and C 1 is the corresponding value vector. 

By Theorem 1, we provide the following algorithm based on the backward induc-

tion to compute CγT .

Algorithm 1: Backward Induction Algorithm (BIA)

Input: MDP
Output: The optimal value vector and a T-optimal policy
1 Initialization: CiT +1 = 0 for all i ∈ S
2 for t=T to 1 do
3 for i ∈ S do ⎧ ⎫
⎨  ⎬
t+1
4 Ci = min ct (i, a) + γ
t t
pia j C j
a∈A(i) ⎩ ⎭
j∈S
⎧ ⎫
⎨  ⎬
5 π ∗t (i) = arg min ct (i, a) + γ t
pia C jt+1
a∈A(i) ⎩ ⎭
j
j∈S

6 return C 1 and π ∗

Remark 1 A change of previous notations in the Algorithm 1 by

Vit = CiT −t+1 , i ∈ S, t = 0, 1, . . . , T.

we give it a simpler and practical form based on the forward induction.

Markov Decision Processes with Discounted Costs … 207

Algorithm 2: Forward Induction Algorithm (FIA)

Input: MDP
Output: The optimal value vector and a T-optimal policy
1 Initialization : Vi0 = 0 for all i ∈ S
2 for t=1 to T do
3 for i ∈ S do ⎧ ⎫
⎨  ⎬
t−1
4 Vi = min ct (i, a) + γ
t t
pia j Vj
a∈A(i) ⎩ ⎭
j∈S
⎧ ⎫
⎨  ⎬
5 ∗t
π (i) = arg min ct (i, a) + γ t
pia j Vjt−1
a∈A(i) ⎩ ⎭
j∈S

6 return VT and π∗

Fig. 3 Graph
representation 1

Example 1 S = {1, 2}; A(1) = A(2) = {a1 , a2 , a3 }; γ = 0.7; T = 10.

Transition probabilities:
p1a1 1 = 1; p1a2 1 = 0.7; p1a3 1 = 1; p1a2 2 = 0.3; p2a1 2 = 0.4; p2a2 2 = 1; p2a3 2 = 1;
p2a1 1 = 0.6.
Costs:
c(1, a1 ) = −1; c(1, a2 ) = 2; c(1, a3 ) = 4; c(2, a1 ) = 7; c(2, a2 ) = −4; c(2, a3 ) =
3 (Fig. 3).

By applying the Algorithm 2 (FIA), we obtain the following numerical results

(Table 1).
Hence, an optimal policy and the corresponding value vector are given by

π = (π 1 , π 2 , . . . , π 10 ) and Cγ10 = (−3.239174917, −12.95669967).

where

π t (1) = a1 and π = (π t (2) = a2 for t = 1, . . . , 10.

208 A. Semmouri and M. Jourhmane

Table 1 Application of Algorithm 2 (FIA)

Iteration t Cγt (1) Cγt (2) π t (1) π t (2)
1 −1 −4 a1 a2
2 −1.7 −6.8 a1 a2
3 −2.19 −8.76 a1 a2
4 −2.533 −10.132 a1 a2
5 −2.7731 −11.0924 a1 a2
6 −2.94117 −11.76468 a1 a2
7 −3.058819 −12.235276 a1 a2
8 −3.1411733 −12.5646932 a1 a2
9 −3.19882131 −12.79528524 a1 a2
10 −3.239174917 −12.95669967 a1 a2

3.2 New Test of Non-optimal Decisions

The action elimination approach based on specific tests is a powerful tool that is
widely used in solving Markov decision problems. Since it helps to identify non-
optimal decisions, it reduces computational complexity. Hence, we provide a new
test in Theorem 2. For thus we will need to use the following lemma.
Throughout this subsection, we assume that all costs {ct (i, a), i ∈ S, a ∈ A(i), t =
1, . . . , T } are nonnegative and stationary with stationarity of transition probabilities
in time and let
c = max |c(i, a)|
i∈S,a∈A(i)

Next, consider the Bellman operator Φ : R|S| → R|S| given for all vector V =
(Vi )i∈S ∈ R|S| by
⎧ ⎫
⎨  ⎬
(ΦV )i = min c(i, a) + γ pia j Vj
a∈A(i) ⎩ ⎭
j∈S

Lemma 2 Let (V n )n≥0 the sequence of vectors given by Algorithm 2. The mapping Φ
is a monotone increasing contraction mapping on R|S| , with respect to the supremum
norm, with contraction factor γ. Moreover

γn − γT
Vin ≤ CγT (i) ≤ Vin + c, for all i ∈ S, 0 ≤ n ≤ T. (5)
1−γ

Proof The monotonicity property of Φ follows from Proposition 6.3.2 of Puterman

(1994) and Lemma 1.1.1 of Bertsekas (2007). Using Proposition 6.2.4 of Puter-
Markov Decision Processes with Discounted Costs … 209

man (1994) and Proposition 1.4.5 of Bertsekas (2007), the property of contraction
mapping is guaranteed.

Next, the proof of Theorem 6.2.3 of Puterman (1994) gives the following result,

γn − γT γn − γT
Vin − c ≤ CγT (i) ≤ Vin + c (6)
1−γ 1−γ

Since Vi0 = 0, i ∈ S and the fact that Φ is an increasing mapping, it follows

Vin ≤ CγT (i), for 0 ≤ n ≤ T (7)

Therefore, (6) and (7) give the requested result in assertion (5). 
We take inspiration from Lemma 2 to give birth to a new test that will appear in
the following theorem.
Theorem 2 (New Test of nonoptimality) Let (V n )n≥0 the sequence of vectors given
by Algorithm 2 (FIA). If for T ≥ 2 and 1 ≤ n ≤ T − 1

 γn − γT
c(i, a ) + γ pia j Vjn Vin + c (8)
j∈S
1−γ

at some epoch n, then any markovian policy which uses action a in state i is nonop-
timal.

Proof Suppose that the action a is optimal in state i and in stage T . Then, we will
have
 
CγT (i) = ViT = c(i, a ) + γ pia j ViT −1 ≥ c(i, a ) + γ pia j Vin (9)
j∈S j∈S

From (8) and (9), we show

γn − γT
CγT (i) Vin + c (10)
1−γ

Combining (6) and (10), it follows that CγT (i) CγT (i). What is absurd. This ends
the proof. 

3.3 New Iterative Algorithm

By combining Algorithm 2 (FIA) and the preceding test (8) in theorem 2, we give
the following new iterative algorithm.
210 A. Semmouri and M. Jourhmane

Algorithm 3: Improved Forward Induction Algorithm (IFIA)

Input: MDP
Output: the optimal value vector, an optimal policy
1 Initialization : Vi0 = 0 for all i ∈ S
2 for n=1 to N do
3 for i ∈ S do
4 for a ∈ A(i) do 
5 Q n (i, a) = c(i, a) + γ pia j Vjn−1
j∈S

6 Vin = min Q n (i, a)

a∈A(i)
7 π ∗n (i) = arg min Q n (i, a)
a∈A(i)
8 for a ∈ A(i) do
γ n −γ N
9 B n (i, a) = Q n+1 (i, a) − Vin − 1−γ c
10 if B n (i, a) 0 then
11 A(i) ← A(i) − {a}

12 return V N and π ∗

Fig. 4 Graph
representation 2

Example 2 Consider an MDP with the following data: S = {1, 2}; A(1) = A(2) =
{a1 , a2 , a3 }; γ = 0.8; T = 20.
Transition probabilities:
p1a1 1 = 1; p1a2 1 = 0.8; p1a3 1 = 1; p1a2 2 = 0.2; p2a1 2 = 0.6; p2a2 2 = 1; p2a3 2 = 1;
p2a1 1 = 0.4.
Costs:
c(1, a1 ) = 17; c(1, a2 ) = 20; c(1, a3 ) = 12; c(2, a1 ) = 18; c(2, a2 ) = 10;
c(2, a3 ) = 14 (Fig. 4).

Applying Algorithm 3 (IFIA), we obtain the numerical results as approximations

of the function B t (i, a) and the optimal value vector C0.8
20
to three digits after the
decimal point (Table 2).
The optimal value vector CγT at iteration T = 20 for γ = 0.8 is

C0.8
20
= (59.308, 49.423)
Markov Decision Processes with Discounted Costs … 211

Table 2 Values of B t (., .) by IFIA

Iteration t B t (1, a1 ) B t (1, a2 ) B t (1, a3 ) B t (2, a1 ) B t (2, a2 ) B t (2, a3 )
0 −81.847 −78.847 −86.847 −80.847 −88.847 −84.847
1 −64.247 −61.567 −69.247 −61.887 −70.847 −66.847
2 −50.167 −47.743 −55.167 −46.719 −56.447 −52.447
3 −38.903 −36.684 −43.903 −34.585 −44.927 −40.927
4 −29.892 −27.837 −34.892 −24.877 −35.711 −31.711
5 −22.683 −20.759 −27.683 −17.111 −28.338 −24.338
6 −16.916 −15.096 −21.916 −10.898 −22.440 −18.440
7 −12.302 −10.566 −17.302 −5.928 −17.721 −13.721
8 −8.611 −6.943 −13.611 −1.952 −13.947 −9.947
9 −5.658 −4.043 −10.658 1.229 −10.927 −6.927
10 −3.296 −1.724 −8.296 3.774 −8.511 −4.511
11 −1.406 0.131 −6.406 5.810 −6.578 −2.578
12 0.106 1.616 −4.894 7.438 −5.032 −1.032
13 1.315 2.803 −3.685 8.741 −3.795 0.205
14 2.283 3.753 −2.717 9.784 −2.805 1.195
15 3.057 4.513 −1.943 10.617 −2.014 1.986
16 3.676 5.121 −1.324 11.285 −1.380 2.620
17 4.171 5.607 −0.829 11.818 −0.874 3.126
18 4.568 5.996 −0.432 12.245 −0.468 3.532
19 4.885 6.308 −0.115 12.587 −0.144 3.856

Fig. 5 Discount factor and actions eliminated

212 A. Semmouri and M. Jourhmane

In addition, the new action spaces given by Algorithm 3 (IFIA) are:

A(1) = {a3 } after 12th iteration and A(2) = {a2 } after 13th iteration.

The following figure gives the cumulative percentage of actions eliminated at each
iteration based on values of the discount factor γ (Fig. 5).

Remark 2 The cumulative percentage of actions eliminated is less important as

soon as the discount factor γ is approaching from 1.

4 Conclusion

In our manuscript, we have updated both two successive recursion methods, based
on bellman’s equation, named backward induction algorithm (BIA) and forward
induction algorithm (FIA) to compute an optimal policy that minimizes the total
discounted expected cost over a finite horizon. Next, we have provided a new test
of suboptimality. The test already proposed in Theorem 2 and the forward induction
algorithm (FIA) are combined to provide an improved algorithm (IFIA) to solve
Markov decision problem in the finite planning horizon under the discounted criterion
by eliminating nonoptimal decisions.
Since the standard value iteration algorithm is finite in Markov decision processes
with finite state space and finite action spaces, our test can be extendable to infinite
horizon by taking T very large as possible. In the future work, we will focus our efforts
to determine the desired horizon T such that the extension already mentioned is valid
with minimum computational complexity. Thus, we will establish new successful
and effective test to rule out suboptimal actions at each epoch if there is possible.
In addition, we have illustrated some numerical examples to make sense of our
contribution.
Although works on action elimination procedures considering the discounted
reward criterion in Markov decision problems cited in the beginning of Sect. 1, this
scientific wealth is not yet extended and very developed in the discounted cost case
over a finite horizon.
In the future work, we will model the functioning of a mobile drill as a robotic
one on fixed rails by the use of Markov decision processes and results presented in
this paper. This device drills holes at specific positions on a metal plate. Hence, this
study will offer an innovative framework that links decision modeling and artificial
intelligence in practical fields.

Acknowledgments The authors would like to thank the following people. Firstly, Professor Dr. C.
Daoui of Sultan Moulay Slimane University, Beni Mellal, Morocco for his help and encouraging
during the period of research. Secondly, Mr. Lekbir Tansaoui, ELT teacher, co-author and textbook
designer in Mokhtar Essoussi High School, Oued Zem, Morocco for proofreading this paper. We
also wish to express our sincere thanks to all members of the organizing committee of the Conference
A2IA’2020 and referees for careful reading of the manuscript, valuable suggestions and of a number
of helpful remarks.
Markov Decision Processes with Discounted Costs … 213

References

1. Bertsekas, D.P.: Dynamic Programming and Optimal Control, 3rd edn. Athena Scientific, Bel-
mont (2007)
2. Derman, C.: Finte State Markovian Decision Processes. Wiley, New York (1970)
3. Feinberg, E.A., Shwartz, A.: Handbook of Markov Decision Processes: Methods and Applica-
tions. Kluwer, New York (2002)
4. Filar, J., Vrieze, K.: Competitive Markov Decision Processes. Springer, New York (1997)
5. Grinold, R.: Elimination of suboptimal actions in Markov decision problems. Oper. Res. 21,
848–851 (1973)
6. Hastings, N.A.J., Mello, J.M.C.: Tests for suboptimal actions in discounted Markov program-
ming. Manag. Sci. 19, 1019–1022 (1973)
7. Howard, R.A.: Dynamic Programming and Markov Processes. MIT Press. Publisher, Cam-
bridge (1960)
8. Hu, Q., Yue, K.W.: Markov Decision Processes with Their Applications. Springer, US (2008)
9. MacQueen, J.: A modified dynamic programming method for Markovian decision problems.
J. Math. Anal. Appl. 14, 38–43 (1966)
10. MacQueen, J.: A test for suboptimal actions in Markovian decision problems. Oper. Res. 15,
559–561 (1967)
11. Porteus, E.L.: Some bounds for discounted sequential decision processes. Manag. Sci. 18, 7–11
(1971)
12. Puterman, M.L.: Markov Decision Processes: Discrete Stochastic Dynamic Programming, 2nd
edn. Wiley, New Jersey (1994)
13. Puterman, M.L., Shin, M.C.: Modified policy iteration algorithms for discounted Markov deci-
sion problems. Manag. Sci. 24, 1127–1137 (1978)
14. Sadjadi, D., Bestwick, P.F.: A stagewise action elimination algorithm for the discounted semi-
Markov problem. J. Oper. Res. Soc. 30, 633–637 (1979)
15. Semmouri, A., Jourhmane, M.: Markov decision processes with discounted cost: the action
elimination procedures. In: IEEE, ICCSRE 2019, pp. 1–6. IEEE, Agadir (2019). https://doi.
org/10.1109/ICCSRE.2019.8807578
16. White, D.J.: The determination of approximately optimal policies in Markov decision processes
by the use of bounds. J. Oper. Res. Soc. 33, 253–259 (1982)
17. White, D.J.: Markov Decision Processes. Wiley (1993)
Robust Adaptive Fuzzy Path Tracking
Control of Differential Wheeled Mobile
Robot

Brahim Moudoud, Hicham Aissaoui, and Mohammed Diany

Abstract This paper presents robust fuzzy path tracking control of Differential
Wheeled Mobile Robot (DWMR) with bounded disturbances and uncertainties. The
proposed controller is based on dynamic model of robot and actuators. The dynamic
sliding mode control and fuzzy logic methods are used to computing the control
voltage. This later ensures the convergence of trajectory and velocity errors to zero.
The effect of chattering phenomenon is reduced by adjusting the switching control
law adaptively. The Lyapunov theory is used to verify the stability of the proposed
controller. Finally, to demonstrate the effectiveness of the proposed method, we use
the simulations in the Matlab/Simulink environment.

Keywords Path tracking · Differential wheeled mobile robot · Adaptive fuzzy

control · Bounded disturbances and uncertainties

1 Introduction

Wheeled Mobile Robot (WMR) is the typical non linear, complex and nonholonomic
system. Controlling the motion of the WMR, in different environments, requires exact
knowledge of the kinematic and dynamic model as well as the disturbances. In this
context, path tracking has recently become an interesting research problem.
To resolve the path tracking problem of WMR, many researchers have designed
the controllers using different non linear and intelligent methods. These methods
can be classified into several categories, namely, those based on the sliding mode

B. Moudoud (B) · H. Aissaoui · M. Diany

Laboratory of Sustainable Development, Faculty of Sciences and Technologies,
Sultan Moulay Slimane University, Beni Mellal, Morocco
e-mail: [email protected]
H. Aissaoui
e-mail: [email protected]
M. Diany
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 215

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_15
216 B. Moudoud et al.

control strategy (SMC) [1–4], those based on the backstepping technique [5, 6], so-
called adaptive [4] method. This latter is used to make the system adaptive, using
the previous techniques combined with other strategies such as fuzzy logic, neuronal
networks [7], etc. For example, In [8] the adaptive control law is used to estimate the
dynamic parameters such; the mass and inertia of the robot.
In [3] three methods are combined; the backstepping is used to design the kine-
matic control, the sliding mode control is used in dynamic control and fuzzy technic
is employed to adapt the switching gain of the control law.
From a robustness point of view, the sliding mode control method is the most used,
however the chatter phenomenon appears when the switching gain is high. This has
a detrimental effect on the control signal. The challenge is to achieve a compromise
between robustness and degradation of the chattering phenomenon.
The main contributions of this work are:
(i) Developing the dynamic DWMR model, including that of the actuator.
(ii) Synthesis of a dynamic control law capable of rejecting the effects of bounded
disturbances and parameter uncertainties while ensuring precise convergence
to zero of the posture and speed errors.
(iii) The proposed controller decreases the chatter phenomenon.
The present document is organized as follows. The kinematic and dynamic model
of DWMR and actuators are presents in Sect. 2. Section 3 chows the design of kine-
matic and dynamic controller. The stability of the proposed algorithm analysed in
Sect. 4. Some simulation results are shown and discussed in Sect. 5. Finally, the
document is concluded.

2 Modeling of Wheeled Mobile Robot

In this section, the kinematic and dynamic model of Differential driving Wheeled
Mobile Robot is presented. The position and orientation of the DWMR in the different
coordinate systems used are shown in Fig. 1. The posture of the robot is defined by
q = [x, y, ϕ]T in the inertial frame and by q1 = [x1 , y1 , ϕ]T in the robot frame. 2b
is the distance between the right and left wheels, r is the wheel’s radius and d is the
distance between C(x, y) (center of mass) and the middle of the drive wheels axis.

2.1 Kinematic Modeling

This subsection presents the description Kinematic of DWMR. The relationship

between q and q1 is given by: [9]
Robust Adaptive Fuzzy Path Tracking Control of Differential Wheeled Mobile Robot 217

Fig. 1 Wheeled mobile

robot and coordinate systems

⎡ ⎤⎡ ⎤
cos(ϕ) − sin(ϕ) 0 x1
q = R(ϕ)q1 = ⎣ sin(ϕ) cos(ϕ) 0⎦ ⎣ y1 ⎦ (1)
0 0 1 ϕ

where R(ϕ) is the orthogonal matrix rotation. The linear and angular velocities of
the robot in the robot frame are given as follows [9]

v = vr +v
2
l
= r (ωr2+ωl )
(2)
ω = 2b = r (ωr2b−ωl )
vr −vl

where ωr and ωl are the angular velocities of right and left wheels respectively. By
differentiating (1) and using (2), we can write the equations of motion of C as [7]:
⎡ ⎤ ⎡ ⎤
ẋ cos(ϕ) −d sin(ϕ)  
v
q̇ = ⎣ ẏ ⎦ = S(q)η = ⎣ sin(ϕ) d cos(ϕ) ⎦ (3)
ω
ϕ̇ 0 1

where η = [v, ω]T is the velocity vector.

2.2 Dynamic Modeling

In this subsection, the dynamic model of mechanical system and DC motors is pre-
sented. The state equation describing the dynamic model of the DWMR robot is:

M η̇ + A(η)η + D = Bτ (4)
218 B. Moudoud et al.
    1 1 
m0 0 −md ϕ̇
where: M = ,A= , B = br −br and τ = [τr , τl ]T . τr and
0 I md ϕ̇ 0 r r
τl are the right and left torque input of the mobile robot respectively. m represents
the total mass of the robot and I its moment of inertia by rearranging Eq. (4), we can
write:

η̇ = A(η)η + Bτ + D (5)

where, A = −M −1 A, B = M −1 B and D = −M −1 D = [d1 , d2 ]T to complete the

dynamic model, we can include actuator modeling. The DWMR is driven by two
DC motors assumed to be identical. The electromechanical equations of each motor
are defined as follows [9]:
⎧
⎨ u mi = R Imi + L ddtImi + ei
e = ke ωmi i = 1, 2 (6)
⎩ i
τi = kτ Imi

where kτ and ke are the torque constant and the back EMF constant respectively. R
and L denote the resistance and inductance of the armature circuit of each DC motor,
respectively.
By considering the gear ratios N and ignoring the inductance L, the Eq. (6)
becomes:

N kτ N 2 ke kτ
τi = u mi − ωi , i = 1, 2 (7)
R R
With ω1 = ωr , ω2 = ωl , u m1 = u mr and u m2 = u ml . We obtain:

τr = τ1 = k1 u mr − k2 ωr
(8)
τl = τ2 = k1 u ml − k2 ωl
2
where k1 = NRkτ and k2 = N Rke kτ .
Considering (5) and (8), the dynamic model can be described by follows equation:

η̇ = Am (η)η + Bm u m + Dm (9)
   k1 k1 
− r2k2 m2 dω
where, Am = , Bm = bk
rm rm , Dm = [dm1 , dm2 ]T represents
− mdωI
− 2kI 2 ( br )2 rI
1
− bk
rI
1

the disturbances and parameter uncertainties. u m = [u mr , u ml ]T is the voltage control

input.
Robust Adaptive Fuzzy Path Tracking Control of Differential Wheeled Mobile Robot 219

3 Controller Design

3.1 Kinematic Algorithm

The kinematic controller is based on back-stepping method which consists of cal-

culating the auxiliary velocity vector ηc = [vc , ωc ]. We define the deviation of the
DWMR by eq = qr − q = [xr − x, yr − y, ϕr − ϕ], where qr is the desired trajec-
tory and q the actual posture of the mobile robot. In the robot frame, the trajectory
error is defined by: e p = [ex , e y , eϕ ]T = R(ϕ)(qr − q). We can write:
⎡ ⎤ ⎡ ⎤⎡ ⎤
ex cos(ϕ) − sin(ϕ) 0 xr − x
e p = ⎣ e y ⎦ = ⎣ sin(ϕ) cos(ϕ) 0⎦ ⎣ yr − y ⎦
eϕ 0 0 1 ϕr − ϕ

The object of this algorithm is to computed the velocity vector ηc which is the
input of the dynamic controller. Fierro and Lewis [5, 2] by considering the derivative
of e p and the reference velocity vector ηr = [vr , ωr ], We obtain:

vc = vr cos(eϕ ) + k x ex
(10)
ωc = ωr + vr k y e y + vr kϕ sin(eϕ )

k x , k y and kϕ are the positive constants.

3.2 Dynamic Controller Design for DWMR

In this subsection, three dynamic controllers are presented. First, the sliding mode
control (SMC) approach is introduce based on dynamic model of DWMR including
the actuator model. Second, the fuzzy logic sliding mode control (FSMC) is designed
in order to reducing the effect of chattering phenomenon. Finally, in order to guarantee
that the robot follows the desired path with a minimum of position and speed error and
with a reduction of the chattering phenomenon, another dynamic control is proposed.

3.2.1 SMC Design

We assume at first that disturbances and uncertainties are zero. In order to find the
voltage input, we propose the auxiliary velocity tracking error and its derivative
as [5]:
ec = [ec1 , ec2 ]T = ηc − η = [vc − v, ωc − ω]T (11)
220 B. Moudoud et al.

We define the sliding surface by:

S = [S1 , S2 ]T = ec + β ec dt (12)

According to (12), when the system is on the sliding surface S = 0, the ec speed
error will converge to zero. From the concept of SMC, the necessary condition for
the state trajectory stay on sliding surface is Ṡ(t → ∞) → 0 then ec (t → ∞) → 0.
Therefor, the equivalent control law u meq is:

u meq = [u meqr , u meql ]T = Bm−1 (η˙c − Am (η)η + βec ) (13)

This control law can make the system state on the sliding surface if the dynamic
model is known exactly. But, in practical applications, this is not the case. In order
to satisfy the sliding condition, we must employ discontinuous control law u ms in
(13). Then, the control law becomes:

u m = u meq + u ms = [u mr , u ml ]T = Bm−1 (η˙c − Am (η)η + βec + γsgn(S)) (14)

where β = diag(β1 , β1 ), γ = diag(γ1 , γ1 ) and β1 , β2 , γ1 , γ2 ≥ 0.

3.2.2 FSMC Design

In the SMC control, robustness is derived from the switching term. However, for the
important values of γ, the switching of the system around the sliding surface causes
the chattering phenomenon. In order to reducing the effect of chattering phenomenon,
the switching gain is adapted by using the fuzzy logic method. The FSMC control is
defined as:
u m = Bm−1 (η˙c − Am (η)η + βec + ρ̂γsgn(S)) (15)

where, ρ̂ = diag(ρ̂1 , ρ̂2 ) the adaptive switching gain and ρ̂1 , ρ̂2 are variable in the
range [0, 0.5] [3].

3.2.3 Proposed Controller

The proposed control law has two objectives. The first is to benefit from the advan-
tages of FSMC control in terms of reducing the effect of chattering phenomenon.
The second is to guarantee the robustness of the system against to disturbances and
uncertainties. This control law is presented as follows:

u m = Bm−1 (η˙c − Am (η)η + βec + ρ̂γsgn(S) + δS) (16)

With δ = diag(δ1 , δ2 ), δ1 and δ2 are the positive constants.

Robust Adaptive Fuzzy Path Tracking Control of Differential Wheeled Mobile Robot 221

4 Stability Analysis

In this section, The stability of proposed control law is studied by using the Lyapunov
theory. The candidate function is chosen as:

L = L1 + L2 (17)

The stabilizing function of the kinematic controller is defined as [10]:

1 2 1 − cos(eϕ )
L1 = (e + e2y ) + (18)
2 x ky

The derivative of L 1 is:

vr kθ sin2 (eϕ )
L̇ 1 = −k x ex2 − (19)
ky

It’s clear that L 1 ≥ and L̇ 1 ≤ 0. Consequently, the kinematic controller is asymptot-

ically stable
1
L2 = ST S ≥ 0 (20)
2
Differentiating (20), we obtain:

L̇ 2 = S T Ṡ = S T (u˙c − u̇ + βec ) (21)

by replacing η̇ in (9) and u m in (16), we obtain:

L̇ 2 = S T (−ρ̂γsgn(S) − δS − Dm ) (22)

Consider that disturbances and uncertainties are bounded |dmi |  ζ. If the control
parameters γi and δi are chosen properly, then we obtain, L̇ 2 ≤ 0, i = 1, 2. Therefore,
the proposed controller is asymptotically stable.

5 Simulation Results

This section presents the simulation results using Matlab/simulink environment. To

evaluate the performances of the proposed method, ⎧ we consider the circular desired
⎨ ẋr = vr cos(ϕr )
trajectory defined by qr = [xr , yr , ϕr ]T , where: ẏr = vr sin(ϕr ) .
⎩
ϕ̇r = ωr
222 B. Moudoud et al.

Fig. 2 Disturbances and uncertainties

Table 1 Simulation parameters

Kinematic parameters d = 0.15 m, b = 0.15 m, r = 0.07 m
Dynamic parameters m = 2 Kg, I = 1 Kg m2
Actuator parameters N = 34, R = 6 , km = 0.13 N m−1

Fig. 3 Trajectory (x, y)

The desired linear and angular velocities are vr = 0.15 m s−1 and ωr = 0.1 rad s−1
respectively. The initial posture (position and orientation) of DWMR is q(0) =
[1, −0.5, 0]T . The bounded disturbances and uncertainties are presents in Fig. 2.
The simulation parameters are defined in Table 1.
In Fig. 3 the reference and actual trajectories are shown. As shown in Figs. 4, 5 and
6, the proposed method has precise tracking over SMC and FSMC methods. However,
the posture error (position and orientation) converges to zero and the effects of
bounded disturbances are rejected. Others, The linear and angular velocities converge
to the desired input velocities as illustrated in Figs. 9 and 10.
The computed voltages (right and left) are presented in Figs. 7 and 8, respectively.
It can be seen that the proposed controller reduces the oscillations of the control
signal (Figs. 9 and 10).
Robust Adaptive Fuzzy Path Tracking Control of Differential Wheeled Mobile Robot 223

Fig. 4 x-axis error

Fig. 5 y-axis error

Fig. 6 Orientation error

224 B. Moudoud et al.

Fig. 7 Right computed

voltage

Fig. 8 Left computed

voltage

Fig. 9 Linear velocity

Robust Adaptive Fuzzy Path Tracking Control of Differential Wheeled Mobile Robot 225

Fig. 10 Angular velocity

6 Conclusion

This paper discussed an adaptive dynamic controller of differential wheeled mobile

robot based on sliding mode control and fuzzy logic methods. The proposed algo-
rithm, based on robot and actuator dynamics, tracks the desired path with precision
and the rejection of perturbations and uncertainties of the model. A second advantage
of this control law is the reduction of oscillations of the control voltage of the actua-
tors. The efficiency of this controller, compared to the SMC and FSMC controllers,
is illustrated using simulations in the Matlab/Simulink environment.

References

1. Koubaa, Y., Boukattaya, M., Damak, T.: Adaptive sliding mode control for trajectory tracking
of nonholonomic mobile robot with uncertain kinematics and dynamics. Appl. Artif. Intell.
924–938 (2018)
2. Wang, Z.P., Yang, W.R., Ding, G.X.: Sliding mode control for trajectory tracking of non-
holonomic wheeled mobile robots based on neural dynamic model. In: Second WRI Global
Congress on Intelligent Systems, pp. 270–273 (2010)
3. Wu, X., Jin, P., Zou, T., Qi, Z., Xiao, H., Lou, P.: Backstepping trajectory tracking based on
fuzzy sliding mode control for differential mobile robots. J. Intell. Robot. Syst. 109–121 (2019)
4. Zhai, J., Song, Z.: Adaptive sliding mode trajectory tracking control for wheeled mobile robots.
Int. J. Control 1–8 (2018)
5. Fierro, R., Lewis, F.L.: Control of a nonholonomic mobile robot: backstepping kinematics into
dynamics. In: Proceedings of the 34th Conference on Decision and Control, New Orleans, pp.
3805–3810 (1995)
226 B. Moudoud et al.

6. Sen, P.T.H., Minh, N.Q., Anh, D.T.T., Anh, D.T.T., Minh, P.X.: A new tracking control algorithm
for a wheeled mobile robot based on backstepping and hierarchical sliding mode techniques.
In: First International Symposium on Instrumentation, Control, Artificial Intelligence, and
Robotics (ICA-SYMP), Bangkok, Thailand, pp. 25–28 (2019)
7. Fierro, R., Lewis, F.L.: Control of a nonholonomic mobile robot using neural networks. IEEE
Trans. Neural Netw. 12 (1998)
8. Mevo, B.B., Saad, M.R., Fareh, R.: Adaptive sliding mode control of wheeled mobile robot
with nonlinear model and uncertainties. In: 2018 IEEE Canadian Conference on Electrical and
Computer Engineering (CCECE), Canada, p. 5 (2018)
9. Ahmad Abu Hatab R.D.: Dynamic modelling of differential-drive mobile robots using Lagrange
and Newton-Euler methodologies: a unified framework. Adv. Robot. Autom. 02(02) (2013)
10. Kanayama, Y., Kimura, Y., Miyazaki, F., Noguchi, T.: A stable tracking control method for an
autonomous mobile robot, pp. 384–389 (1990)
Deep Learning Approach for Automated
Guided Vehicle System

Mohamed Rhazzaf and Tawfik Masrour

Abstract Automated guided vehicles system (AGVS) is a new logistics problem

area, and the most demanded in terms of performance in view of the exponential
growth of product traffic in the world. The present paper aims to to propose a new
approach, to deal with AGVS problems, based on deep reinforcement learning algo-
rithms, as an alternative to classic methods. Indeed, the classical approaches are based
on a pre-established policy of vehicle movement rules, while our method deduces the
movement rules based on trial/error reinforcement learning approach, and effectively
gives good results in relatively small moving areas and a limited number of agents.

Keywords Artificial intelligence · Automated guided vehicles system · Deep

learning · Reinforcement learning · Multi-agents system

1 Introduction

Today, smart process automation in the industrial sector is one of the main foundation
of Industry 4.0. Indeed, the new challenges demanded by the world trade impose
minimization of human intervention in repeated processes, and increasingly injecting
artificial intelligence into the industrial logic.
Automated guided vehicle system (AGVS) is one of the concrete examples of
these challenges, especially in the area of product transshipment. AGVS is a system
consisting in group of vehicles sharing the same environment and which is responsible
for transporting products from a given departure to a given destination. Such a system

M. Rhazzaf (B)
Artificial Intelligence for Engineering Sciences Team (IASI), Doctoral Studies Center,
ENSAM-Meknes, University My Ismail, Meknes, Morocco
e-mail: [email protected]
T. Masrour
Department of Mathematics and Informatics,Artificial Intelligence for Engineering Sciences
Team (IASI), ENSAM-Meknes, University My Ismail, Meknes, Morocco
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 227

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_16
228 M. Rhazzaf and T. Masrour

requires management of task allocation and resolving access conflicts to the common
areas.
The technical specifications of an AVGS require an advanced algorithmic design
to complete the desired tasks, referring in particular to product allocation system
for vehicles, and also the management of access conflicts. On the other hand, the
use of machine learning methods, and more precisely deep reinforcement learning
in these systems makes it possible to remove the complexity of their formalism, and
subsequently simplify the implementation tasks for architects, and this is what we
are going to put in our model.
Our study will be distributed as follows: Section 2 presents the background of
AGV systems, and their implementations. Section 3 illuminates the concept of the
Markov chain, reinforcement learning, and deep learning. Next we will present in
Section 4 the vehicle control rules, and the reinforcement learning model. Section 5
states the results obtained by the model, and the Section 6 presents a conclusion and
perspectives.

2 Background

An AGV is a driverless robot for transporting products, it is designed to move products

in a two-dimensional space. This technology has been used since the 1950s in many
sectors, namely the automotive industry, warehouses, distribution transshipments,
etc.
An AGVS is a system that brings together all of the AGVs, with the circulation
platform and the transport network, and essentially the control system [1]. It is
considered to be the main tool for the intra-logistics sector according to Gotting’s
2000 statistics [2]. Depending on the context and the application fields, AGVs and the
transport network differ from case to case, but the basic logic of the control system
remains almost the same. The control system of an AGVS takes care of the tasks
of assigning orders to AGVs, indicating directions, and resolving access conflicts to
common areas or the deadlocks constraints.
There are several criteria for AGVS classification, we cite the number of AGVs that
occur in the environment, the dimensions of the environment, and the AGVs moves
mechanisms. On the other hand, the control systems classifications are arranged by
the strategy for shipping products, and the mechanisms for managing deadlocks and
access conflicts [3].
According to Egbela and Tanchoco [4], different heuristics can manage the prob-
lems of shipping products in AGVS, such as the rule of the random vehicle, the
nearest vehicle, the farthest vehicle, of the most chosen vehicle, and that of the least
chosen vehicle.
For the management of deadlocks and access conflicts [5], there are three man-
agement classes. The first is purely hardware, which is linked to the design of the
environment itself, and which structurally prevents conflicts between vehicles. The
second requires the division of the environment into zones, so that the control system
Deep Learning Approach for Automated Guided Vehicle System 229

allows the presence of only one vehicle at a time in a given zone, this is a pseudo-
hardware solution. And the third alternative integrates algorithms assigning the paths
for AGVs in order to avoid access conflicts, we speak of a purely software solution.

3 Deep Reinforcement Learning

3.1 Reinforcement Learning

Reinforcement learning [6] is a set of machine learning techniques, it differs from

supervised learning [7] which requires the presence of labeled data, and unsupervised
learning [8] which feeds on unlabeled data. This paradigm is based on the sequential
interaction between an artificial agent and a given environment. The agent reacts
through the application of an action, which can modify the environment characteris-
tics. Then the environment returns its present state to the agent by adding the reward
for the action taken by the agent leading to this state. The agent thereafter aims to
maximize the cumulative rewards acquired from the environment in the long term,
which is an optimization problem.
A Reinforcement learning problem must be modeled by what is called the Markov
decision process (MDP) [9] which can be presented in simple terms as a sequential
decision process, it is defined by the 4 tuples (S, A, Pa , Ra ) where S represents all
the states of the environment, A indicates all the possible actions, Pa expresses the
transition function between states, and Ra indicates the immediate reward function
received after the transition between two states by the effect of a given action.
Two notions are important in each Reinforcement learning implementation: The
policy π(s|a) which defines the mapping between the states of the environment
and the actions of the agent, and this mapping can be deterministic as it can be
stochastic. The value function of a policy π on the other hand, expresses the expec-
tation of accumulating long-term rewards respecting π . There are two value function
alternatives, the state value function Vπ (s)= E π [ ∞ k=0 γ Rt+k+1 |st = s] and the
k
∞
state-action value function qπ (s, a) = E π [ k=0 γ Rt+k+1 |st = s, at = a] where γ
k

expresses the discount factor.

3.2 Deep Learning in Reinforcement Learning

The concept of artificial neuron was started in the 1950s with the introduction of
so-called perceptron, it’s a computational unit that simulates the behavior of neu-
ronal cells. Neural network is an architecture of neural layers that communicate and
collaborate with each other to approximate complex or nonlinear patterns.
It was not until the early 90s that this concept was used at a strong cadence, and
this due to a computational power provided by the new generations of computers
230 M. Rhazzaf and T. Masrour

(CPU, GPU), as well as the data availability. However the results remain limited to
specific contexts, especially of image processing.
The beginning of deep learning took place in the late 1990s when we realized that
more layers on the neural network we implement, the more it is suitable for high
level features. The concept was first introduced with Yan LuCun [10] in the form of
a convolutional neural network (CNN) that deals with hand writing digits with great
efficiency. Afterwards, several deep neural network architectures appeared, namely
the recurrent neural network (RNN) [11], autos encoder [12].
The first promising results of the use of deep learning in reinforcement learning
problems was by [13], using DQN (Deep Q Network), applied on the ATARI games
as a case study. The idea is to approximate the state-action value function q(s, a)
by the neural networks in the context of the Q-Learning approach minimizing the
quantity:
L i (θi ) = E s,a [(yi − Q(s, a; θi ))2 ]

Q(s  , a  ; θi−1 |s, a)]

dqn
yi = E s  [R + γ max
a

with the addition of Relay Memory concept, to avoid the correlation problems, and the
Target Network to delaying update of the neural network . D2QN (Double Deep Q-
Network) [14] is an improvement on DQN in the sense that we differentiate between
the estimation network and the evaluation network, to avoid overestimation of state-
action value function of the neuron network, by replacing yi with

= E s  [R + γ Q(s  , arg max Q(s  , a, θt ); θt− |s, a)]

d2qn
yi
a

After that, a new method proved its performance that we will use it in our study,
that of the D3QN (Dueling Double Deep Q-Network) [15], which uses a mix of the
value functions V (s) and q(s, a) in the context of a called advantage value function
A(s, a) to have the optimal policy.

4 Deep Reinforcement Learning AGVS Model

In the design of our AGV system, we consider a two dimensions D1 ∗ D2 discreet

plane, containing a finite number of vehicles moving in discrete steps, and we also
consider a finite number of products to deliver, each vehicle has the right to deliver
a single product at a time.
To design our AGVS in Markov Decision Process, it is necessary to define its 4
tuples, namely the states, the actions, the transitions and the reward function (Fig. 1).
Deep Learning Approach for Automated Guided Vehicle System 231

Fig. 1 The AGV system sketch, a gripe system with cells, the agents and the products are dispatched
inside the zone

4.1 States

They gather information from agents and products to be delivered. The agents infor-
mation contains the agents’ positions and for each agent who has a product to deliver
its destination position. The information of a product are simply their starting posi-
tions.

4.2 Actions

Each vehicle has 5 possible actions (up, down, right, left, stay), but for our AGVS the
vehicles interact with the environment at the same time, so an AGVS agent’s action
is the combination of all vehicles actions, that is, it is a vector containing the actions
of each vehicle (Fig. 2).

Fig. 2 AGV actions, up,

down, right, left, and stay
232 M. Rhazzaf and T. Masrour

4.3 Transitions

They are done through the behavior of vehicles according to the physical rules of the
environment. These rules can be summarized as follows:

– An action leading a vehicle out of the grid is prohibited, and the vehicle in question
maintains the position of its previous state.
– An action leading two or more vehicles to try to access a common position at the
same time is prohibited, and the vehicles in question maintain the positions of their
previous state.
– A free vehicle making an action which leads it to an area where there is a product
to be delivered will have the new position and the product to be delivered.
– A vehicle having a product to deliver making an action which leads it to the
destination of its product will have the new position with free situation to have
the possibility of delivering another product, and the product in question becomes
delivered.

4.4 Rewards System

The reward of an AGVS action on a state is the average of the elementary rewards of
each action made by a vehicle in the system, and for each vehicle we can summarize
the rewards collected as shown in Table 1.
To tackle a reinforcement learning problem, many methods exist, we cite for
example dynamic programming [16], and Monte Carlo [17], but the goal is always
to get the optimal value function for each state-action combination possible. Such
an approach requires a relatively small state space.

Table 1 Reward roles

Role Reward
Staying without product −0.08
Staying with product −0.2
Collision with border without product −0.2
Collision with border with product −0.4
Collision with vehicle without product −0.6
Collision with vehicle with product −0.3
Stand by without product −0.1
Stand by with product −0.2
Loading product 0.8
Delivered product 1
Deep Learning Approach for Automated Guided Vehicle System 233

Fig. 3 The AGV system learning procedure, it start with 4 by 4 grid and step by step we use the
transfer learning and increase the grid dimensions until reaching the final one

The MDP design of our AGVS result a huge state space, which prompts us to
use deep reinforcement learning. One such method aims to approximate the value
function by the neural network using the Q-learning approximation [13]. There are
different deep reinforcement learning alternatives, but we will use D3QN [15].

5 Simulation and Results

To simulate our AGVS model, we maintain our goal to reach a stable AGVS for a
7 ∗ 7 grid, but first we begin with a 4 ∗ 4 grid for K vehicles and 5 products, after
that, gradually, we will increase the grid space by adding a unit in one of the two
dimensions, and using the transfer learning that imports the trained model from the
previous grid. So we will apply he procedure for a 4 ∗ 5, 5 ∗ 5, 6 ∗ 5, and so on until
we reach the 7 ∗ 7 grid.
As we mentioned above, we will use the D3QN algorithm. The algorithm trains
a hidden 4-layer neural network to ensure the non-linearity of the approximation
of the state action value function q(s, a).We do not use CNN architecture because
we use height level features based on what we said in Chap. 4 as using AGVs and
products information (positions, states , …). The number of neurons in each layer
will be estimated heuristically by the relation:

h 0 = (8 ∗ agent ∗ pr oduct + 6 ∗ agent + 9 ∗ agent ∗ (agent − 1))

There are many hyper parameters to adjust, including:

– The Bach size: between 600 and 4000
– The Relay memory size: between 20000 and 80000
– The discount factor: 0.99
– The learning rate: 0.001
– The exploration value: varies between 30 and 60% at the start of learning, and it
decreases linearly to 10% when the loss gets to 10−2.2
– Network Target: we use hard update target network at 1000 to 4000 steps (the
soft update is not stable in our case).
The increase in the number of agents exponentially increases the number of
actions, which impacts the choice of the hyper parameters, for example having more
agents, intuitively influences the size of the bash, indeed it must be increased to ensure
234 M. Rhazzaf and T. Masrour

(a) Average Loss/Mean-Q value for 1 agent (b) Average AGV moves for 1 agent with
with h0 in each hidden layer h0 in each hidden layer

(c) Average Loss/Mean-Q value for 1 (d) Average AGV moves for 1 agents with
agents with 2h0 in each hidden layer 2h0 in each hidden layer

(e) Average Loss/Mean-Q value for 1 agent (f) Average AGV moves for 1 agent with
with 3h0 in each hidden layer 3h0 in each hidden layer

Fig. 4 Statistic results for 1 agents and 5 products

the asymptotic presence of all possible actions, so we must raise the exploration, and
the training iteration number, for the same reason (Figs. 3 and 4).
In Fig. 5 we can notice some spikes in the average of AGV moves, actually they
indicate the stages we apply transfer learning at the moment of the map transition.
We should also mention that the elevation of level average in some graphs, espe-
cially in the AGV average moves, is actually due to the growth strategy learning of
dimension environment (Fig. 6).

6 Conclusion

In this paper we apply a deep reinforcement learning approach to a problem

that requires some intuition about collaborative Multi-agents System. The results
obtained in the simulation above show a stable regime for one and two agents, but
the system maintains good results as we don’t reach higher dimension values.
Deep Learning Approach for Automated Guided Vehicle System 235

(a) Average Loss/Mean-Q value for 2 agent (b) Average AGV moves for 2 agent with
with h0 in each hidden layer h0 in each hidden layer

(c) Average Loss/Mean-Q value for 2 (d) Average AGV moves for 2 agents with
agents with 2h0 in each hidden layer 2h0 in each hidden layer

(e) Average Loss/Mean-Q value for 2 agent (f) Average AGV moves for 2 agent with
with 3h0 in each hidden layer 3h0 in each hidden layer

Fig. 5 Statistic results for 2 agents and 5 products

As we begin with three agents, the system becomes unstable, even with multiple
layer sizes, and multiple layer numbers. This chaotic behavior is due to many reasons,
we cite the size of the action space (5m where m is the number of agents),the hyper-
parameter setting sensibility, and the need of more data that engage all the possible
actions, but increasing memory replay does not resolve the problem because we
cannot be sure about the data quality regenerated by the reinforcement learning
agent.
Because what we cited above, as perspectives, we can focus on reducing the
actions space for the large number of agents using an asynchronous action decision
approach, and also we can look for some solutions to the large dimension environment
problem using cooperative approach between agents groups.
236 M. Rhazzaf and T. Masrour

(a) Average Loss/Mean-Q value for 3 agent (b) Average AGV moves for 3 agent with
with h0 in each hidden layer h0 in each hidden layer

(c) Average Loss/Mean-Q value for 3 (d) Average AGV moves for 3 agents with
agents with 2h0 in each hidden layer 2h0 in each hidden layer

(e) Average Loss/Mean-Q value for 3 agent (f) Average AGV moves for 3 agent with
with 3h0 in each hidden layer 3h0 in each hidden layer

Fig. 6 Statistic results for 3 agents and 5 products

References

1. Muller, T.: Automatic Guided Vehicles. IFS (Publ.) Ltd., U.K. (1983)
2. Gotting, H.H.: Automation and steering of vehicles in ports. Port Technol. Int. 10, 101–111
(2000)
3. Co, G.C., Tanchoco, J.M.A.: A review of research on AGVs vehicle management. Eng. Costs
Prod. Econ. 21, 35–42 (1991)
4. Egbelu, P.J., Tanchoco, J.M.A.: Characterization of automatic guided vehicle dispatching rules.
Int. J. Prod. Res. 22(3), 359–374 (1984)
5. Xu, F., Van Brussel, H., Nuttin, M., Moreas, R.: Concepts for dynamic obstacle avoidance and
their extended application in underground navigation. Robot. Auton. Syst. 42, 1–15 (2003)
6. Sutton, R.S., Barto, A.G.: Introduction to Reinforcement Learning, vol. 2, no. 4. MIT press,
Cambridge (1998)
7. Russell, S.J., Norvig, P.: Artificial Intelligence: A Modern Approach. Pearson Education Lim-
ited, Malaysia (2016)
8. Barlow, H.B.: Unsupervised learning. Neural Comput. 1(3), 295–311 (1989)
9. White, C.C.: Markov Decision Processes, pp. 484–486. Springer, US (2001)
10. LeCun, Y., Bengio, Y.: Convolutional Networks for Images, Speech, and Time Series. The
Handbook of Brain Theory and Neural Networks, vol. 3361, no. 10 (1995)
Deep Learning Approach for Automated Guided Vehicle System 237

11. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Trans. Sig. Proc.
45(11), 2673–2681 (1997)
12. Hinton, G.E., Zemel, R.S.: Autoencoders, minimum description length and Helmholtz free
energy. In: Advances in Neural Information Processing Systems, pp. 3–10 (1994)
13. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Petersen, S.:
Human-level control through deep reinforcement learning. Nature 518(7540), 529–533 (2015)
14. Van Hasselt, H., Guez, A., Silver, D.: Deep reinforcement learning with double q-learning. In:
Thirtieth AAAI Conference on Artificial Intelligence (2016)
15. Wang, Z., Schaul, T., Hessel, M., Van Hasselt, H., Lanctot, M., De Freitas, N.: Dueling net-
work architectures for deep reinforcement learning. In: Proceedings of the 33rd International
Conference on Machine Learning, vol. 48, pp. 1995–2003 (2016)
16. Bellman, R.: Dynamic programming. Science 153(3731), 34–37 (1966)
17. Hastings, W.K.: Monte Carlo sampling methods using Markov chains and their applications
(1970)
Path Planning Using Particle Swarm
Optimization and Fuzzy Logic

Ahmed Oultiligh, Hassan Ayad , Abdeljalil Elkari ,

and Mostafa Mjahed

Abstract Path planning is one of the research axes in mobile robotics, it allows us
to plan a path linked, the robot’s starting point and the goal to guarantee autonomous
navigation for the mobile robot. This work presents a quantitative methodology to
compare the Particle Swarm Optimization algorithm and Fuzzy Logic, applied to
solve the path-planning problem, taking into account their basic concepts and results
obtained.

Keywords Robot mobile · PSO · Autonomous navigation · Fuzzy logic · Path

planning · Optimal path

1 Introduction

Robot motion planning is an important constituent of robotics, as it gives a mobile

robot the ability and strength to plan its motion, and moves autonomously between
two points for accomplishing different tasks. The fundamental task of path planning is
to generate the shortest path for a mobile robot towards the goal point while avoiding
any contact with the obstacles.
Generally, robot path planning problems are categorized as global path plan-
ning and local path planning. In global path planning, the algorithm has to prepare
the entire path with coordinates using different types of methods before the robot

A. Oultiligh (B) · H. Ayad · A. Elkari

Faculty of Sciences and Technology, Marrakesh, Morocco
e-mail: [email protected]
H. Ayad
e-mail: [email protected]
A. Elkari
e-mail: [email protected]
M. Mjahed
Royal School of Aeronautics, Marrakesh, Morocco
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 239

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_17
240 A. Oultiligh et al.

starts its motion. Rather, local path planning takes place completely in an unknown
environment when the mobile robot is moving.
A lot of research has been done on path planning. So Far, various path plan-
ning algorithms and methods have been presented by researchers, such as the clas-
sical path planning technique including, Visibility Graphs and Dijkstra [1], Voronoi
Diagrams [2], Rapid Random Tree Exploration [3], Artificial Potential Field [4, 5],
A* algorithm [6], and many others. Each one with its strengths and weaknesses. The
heuristic technique of path planning including, the fuzzy logic technique which is
mainly based on the IF–THEN rules [7–9], the artificial neural network (ANN) which
is very popular among heuristic approach methods because of its parallel learning
and processing capability [10]. In addition to these two approaches, various searches
have used algorithms inspired by nature such as Colony Optimization Ant (CAO)
[11, 12], Genetic Algorithm [13], Particle Swarm Optimization (PSO) [14–16] and
Cuckoo-search [17]. Also, some works use the hybrid techniques in path planning,
for example, the Neuro-Fuzzy technique where the fuzzy logic controller uses ANN
as a pre-processing part, GA-Fuzzy Logic [18, 19], GA-Artificial Bee Colony [20]
and many other hybrid techniques.
Although these algorithms give some search results, there are still many problems.
For example, the artificial potential field method establishes a simple and efficient
virtual physical environment, but still presents some problems such as local minimum
and oscillation. The genetic algorithm is a common method, a large number of studies
show that the genetic algorithm has achieved good results in path planning, but there
are still some weaknesses, such as the use of a random method to generate the
initial population, although the diversity of the population has met the requirements,
it may lead to infeasible trajectories and slow convergence [13]. Fuzzy logic also
has a problem which is the difficulty to construct and adjust the correct rank of the
membership function. Researchers have tried to solve this problem by combining
fuzzy logic with other algorithms, such as the Wind-Driven Optimization algorithm
(WDO). The use of a hybrid Fuzzy-WDO algorithm can achieve the navigation of
mobile robots and collision avoidance in a static and dynamic environment [21].
This article presents the use of Particle Swarm Optimization and Fuzzy Logic,
for planning the optimal path of a mobile robot in an environment with polygonal
obstacles. In literature review, there can be found serval researchers have been worked
on Fuzzy logic technique for mobile robot navigation. Some researchers proposed a
new approach of behavior-based fuzzy control for mobile robot [9], in which they
based on a behavior approach to control the mobile robot. The rest of the document
is structured as follows:
Section 2 discusses Particle Swarm Optimization and Fuzzy Logic as adaptive
methods for path planning. Section 3 demonstrates the effectiveness of the proposed
planners using the simulation results, and Sect. 4 provides the conclusion and future
recommendations.
Path Planning Using Particle Swarm Optimization and Fuzzy Logic 241

2 Adaptative Methods for Path Planning

2.1 Path Planning Using PSO Algorithm

Particle Swarm Optimization is a very popular algorithm in recent years. Also, many
researchers find that PSO is simpler, it can be easily applied in optimization problems,
this algorithm is used in mobile robotics especially for planning an optimal path.
Indeed, a smaller number of evaluations of this algorithm can converge it towards
the solution. In this section, we present the MATLAB implementation of PSO for
optimal path planning.
The PSO simulates the movement of a swarm of particles in a research space
multidimensional. The position of each particle presents a possible solution (e. g. a
path). The particle moves to a better position (e. g. the optimal path). Each particle is
located by its position and velocity, two parameters that must be adjusted according
to the previous position, the best position previously occupied by the particle and the
best position previously occupied by the best particle of the swarm at each iteration
of the simulation. Taking these adjustment rules into account, the swarm of particles
moves towards an optimal solution, whereas it is initially random. The adjustment
of the velocity and position of a particle is done according to the following equations
[16]:

Vi = ω Vi−1 + c1r1 (Pbest − X i−1 ) + c2 r2 (G best − X i−1 ) (1)

X i = X i−1 + Vi (2)

Vi The speed of the particle i.

Pbest The best position of the particle i.
Gbest The best position of the best particle in the swarm.
Xi The position of the particle i.
c1 and c2 constants.
ω is the inertia weight.
r 1 and r 2 random value vectors between 0 and 1.

To meet the objective of planning an optimal path for a mobile robot, we use PSO
in MATLAB to simulate the movement of a swarm of paths in a 2D space to a position
that gives us the optimal path as much as possible by avoiding obstacles. To do this,
we present our swarm (path) by a set of particles (following the coordinate points (x,
y)). At the beginning of the algorithm, the particles are randomly initialized. In our
case, the random generation of these particles is done in such a way as to eliminate
the points that are inside an obstacle, to avoid the choice of random path that will
cross the obstacle. The implementation of PSO in MATLAB is presented in Fig. 1.
242 A. Oultiligh et al.

Initialization of the parameters

Initialization of the position of the particles

Initialization of particle velocity

Evaluation of the cost of particles

Update of the best local position

Update of the best global position

Calculation of the new particle velocity

Calculation of the new particle position

No Yes
Finished?

Smoothing of the best path.

End.

Fig. 1 The PSO algorithm for path planning

2.2 Path Planning Using Fuzzy Logic

In this part, fuzzy logic is used to control the navigation of a unicycle type mobile
robot; Pionner3D-X. Fuzzy behaviors are adopted as illustrating the decomposition
of the global navigation task into a set of elementary behaviors including, Converge
towards the Goal (CG), Avoid Obstacles (AO) and Follow the Wall (FW). The fuzzy
controller used contains, in addition to the three previous behaviors, a supervisor
who allows the robot to choose at any time the behavior to be activated, according
to the sensory data.
The fuzzy controller used has six inputs; “Dist-L, Dist-LF, Dist-F, Dist-RF and
Dist-R”, which are the information captured by ultrasonic sensors of the Pionner3D-
X robot that can record the Distance between the robot and obstacles (Fig. 2a). And
the orientation angle of the robot (Angle-RG) which is calculated according to the
current position of the robot and the goal (Fig. 2b). The navigator outputs are the
speeds of the wheels either left and right. At each moment the system reads the
coordinates (xg, yg), and calculates the Distance separating the robot and the goal
Path Planning Using Particle Swarm Optimization and Fuzzy Logic 243

Fig. 2 a The sensors of the Pioneer 3D-X used, b position of the robot and the goal

and its angle of orientation according to the following equations:

  
yg
Dist-RG = xg 2 + yg 2 , Angle-RG = arctan (3)
xg

Figure 3 shows the structure of the fuzzy controller used. The outputs of each
fuzzy behavior are the speeds of the robot wheels. The supervisor chose between the
different behaviors; according to these inputs, he activates the necessary behavior to
respond to the situation of the robot in the environment. For example, if the sensors
do not detect any obstacles, the supervisor activates the behavior “Converge towards
the Goal” to control the robot.
The Conception of Fuzzy Behaviors. The different behaviors (Converge towards
the Goal, Avoid Obstacles and Follow the Wall) combine the data received at the
inputs, then each one processes a specific command according to a set of fuzzy rules
to be simple to interpret, readable and with command efficiency to control the robot.
Converge Towards the Goal (CG). The behavior “Converge towards the Goal”, allows
the robot to reach the goal when there is no obstacle using the knowledge of its

Supervisor

Angle-RG AcƟvaƟon of a behavior

Converge towards the Goal

Dist-L
LeŌ Speed (LS)
Dist-F LS
Avoid the obstacles
Dist-R RS
Right Speed (RS)
Dist-LF

Dist-RF Follow the wall

Fig. 3 The structure of the fuzzy controller used

244 A. Oultiligh et al.

Negative Zero Positive

1 1
Degree of mf

Degree of mf
0,5 0,5

-200 -35 -10 0 10 35 200 -4 -2 0 1 2 3

Input variable “Angle-RG” Output variables “LS” and “RS”
(a) (b)

Fig. 4 a Membership functions of the Angle-RG variable, b the left and right speeds of the robot

Table 1 Fuzzy rules of the

Angle-RG Left speed (LS) Right speed (RS)
behavior “Converge towards
the Goal” Negative Normal Back
Zero Big Big
Positive Back Normal

orientation angle. The assumed fuzzy controller uses a single variable (Angle-RG)
to generate the two robot speeds (LS and LR). The membership functions used to
fuzzify the input variable (Angle-RG) are trapezoidal (Fig. 4a). Both speeds are
reduced to singletons (Fig. 4b): These speeds configurations are the same for the
other behaviors.
After the definition of the linguistic variables, 3 inference rules are established to
define the link between the input–output variables for this behavior (Table 1).
Avoid Obstacles (AO). “Avoid Obstacles” behavior gives the robot the ability to
avoid an obstacle when one or more objects are detected at a certain Distance from
the mobile robot by its ultrasonic sensors. The membership functions used to fuzzify
his input variables are trapezoidal (Fig. 5). The different fuzzy rules used in the
inference step are shown in Table 2.
Follow the Wall (FW). “Follow the Wall” behavior consists of avoiding collision with
obstacles that are in a particular position in the environment. When the robot moves
towards the goal, its orientation angle becomes almost zero, this angle remains zero
until the robot reacts to avoid such an obstacle. If an obstacle is detected just in front
of the robot, the robot’s reaction should be one of the following: turn left or right
to avoid the obstacle. At that moment the navigator must choose the right reaction
since we want to guarantee better navigation (the shortest possible path). Moreover,
if the robot turns in one of these directions, it must remain in a path parallel to the
obstacle so that it does not move away from the goal, it can continue its convergence
towards the goal as soon as it passes the obstacle. This behavior is called “Follow
Path Planning Using Particle Swarm Optimization and Fuzzy Logic 245

Fig. 5 Membership
functions of variables: Near Far
Dist-L, Dist-R, Dist-F,
1
Dist-RF and Dist-LF

Degree of mf
0,5

0 0.5 0.6 1
Inputs variables “DistL”, “Dist-R”, “Dist-RF” and “Dist-LF”

Table 2 Fuzzy rules of behavior “Avoid Obstacles”

Dist-L Dist-F Dist-R Dist-LF Dist-RF Left Speed (LS) Right Speed (RS)
Far Far Near – – Big Big
Near Far Far – – Big Big
Far Near Far Far Near Back Normal
Far Near Far Near Far Normal Back
Far Near Far Near Near Normal Back
Far Near Far Far Far Normal Back

the Wall". The input variables are the different sensors used. This behavior processes
and generates the two speeds according to 6 fuzzy rules presented in Table 3.
Supervisor. The supervisor is a fuzzy controller with six inputs (all the inputs of
the fuzzy controller) and one output. The controller’s rule base links these inputs
and, depending on their states, generates an output according to which the algorithm
activates one of the three behaviors (Converge towards the Goal, Avoid Obstacles or
Follow the Wall). The input variables are fuzzified by trapezoidal form membership

Table 3 Fuzzy rules of behavior “Follow the Wall”

Dist-L Dist-F Dist-R Dist-LF Dist-RF Left Speed (LS) Right Speed (RS)
Near Near Near – – Normal Back
Near Near Far – – Normal Back
Far Near Near – – Back Normal
Near Far Near – – Big Big
Far Far Far Near Far Normal Back
Far Far Far Far Near Back Normal
246 A. Oultiligh et al.

Table 4 Fuzzy rules of the “Supervisor”

Dist-L Dist-F Dist-R Angle-RG Dist-LF Dist-RF Behavior
Far Far Far – Far Far CG
Far Far Far – Far Near AO
Far Far Far – Near Far AO
Near Near Far – – – AO
Near Near Near – – – AO
Near Far Near – – – AO
Far Near Near – – – AO
Near Far Far Negative – – CG
Near Far Far Zero – – CG
Far Far Near Positive – – CG
Far Far Near Zero – – CG
Far Far Near Negative – – FW
Far Near Far Zero – – FW
Near Far Far Positive – – FW

functions (Figs. 4a, b and 5). The different fuzzy rules of this block are presented in
Table 4. The implementation of the Fuzzy controller is described in Fig. 6.

3 Simulations and Results

In this section, the different methods are simulated using the V-rep software. In
addition to PSO and Fuzzy logic, two other methods Dijkstra and Voronoï have been
applied, to compare the results obtained by (PSO, Fuzzy logic, Dijkstra and Voronoï).

3.1 The PSO Algorithm

Table 5 presents the parameters of PSO and the values where chosen in our
implementation. Figure 7 presents the paths obtained by PSO.

3.2 Fuzzy Logic Controller

The simulation of the fuzzy controller is presented in Fig. 8.

Path Planning Using Particle Swarm Optimization and Fuzzy Logic 247

Read the positions of the robot and the goal

Calculate the distance between the robot and the goal

No Yes
The robot reaches the goal?

Read the sensors and calculate the Stop the robot

orientation angle of the robot

Supervisor End.

Converge towards the Goal

Avoid Obstacles

Follow the Wall

No
Yes
The robot reaches the goal?

Stop the robot

End.

Fig. 6 The robot’s autonomous navigation algorithm

Table 5 Parameters of the PSO algorithm

The parameter Number of iterations Number of population ω C1 C2
Suggested values 300 850 0.98 2.5 1.9
248 A. Oultiligh et al.

Fig. 7 The paths obtained by PSO for two scenarios

Fig. 8 The paths obtained by Fuzzy logic

3.3 Dijkstra Algorithm

Figure 9 presents the paths obtained by Dijkstra.

Fig. 9 The paths obtained by Dijkstra algorithm

Path Planning Using Particle Swarm Optimization and Fuzzy Logic 249

Fig. 10 The paths obtained by Voronoï diagram

3.4 Voronoï Diagram

Figure 10 presents the paths obtained by Voronoï method.

To make a comparison according to the length of the path, we simulate path
planning in the V-rep software using a 2D space (x = 6 m, y = 5.2 m) with 5
obstacles. Table 6 presents a comparative study of the results obtained.
We observed the operation of those methods and found that they guide the robot
through its environment so that it arrives at the final destination without collision
with obstacles by following the planned path.
The PSO algorithm worked efficiently to find the best solution after a defined
number of iterations. The PSO manages to find the shortest path by avoiding obsta-
cles. This algorithm ensures the planning of a short trajectory and it can be adapted
to more complicated situations.
The built fuzzy controller uses the different behaviors according to the situation
where the robot is at each moment, each behavior decides according to the data
acquired by the sensors. This method does not guarantee an optimal path, even if it
can control the robot in an unknown complicated environment.

Table 6 Comparative table

Method PSO Fuzzy logic Dijkstra Voronoï
Scenario Source Length of the path 6.2864 10.3421 6.3421 8.3421
(a) (x = 0.5 m, y = (m)
0.5 m)
Goal
(x = 5 m, y = 4 m)
Scenario Source Length of the path 5.3437 8.3421 5.3737 7.3737
(b) (x = 0.5 m, y = (m)
0.5 m)
Goal
(x = 5 m, y = 3 m)
250 A. Oultiligh et al.

By analyzing the results of the Dijkstra algorithm, we noticed its effectiveness in

optimizing the trajectory, but it has certain limits, its implementation is becoming
more and more complex for more complicated environments.
By analyzing the result of the Voronoï method, we noticed that in the Voronoi
diagram, the path generation is away from the obstacles so that the robot will stay safe
without any collusion, but the path planned is not claimed as shortest. This method
gives maximum clearance to the path but becomes more complex in implementation.

4 Conclusion

In this paper, a quantitative comparison is presented of different path planning

methods with different fundamental characteristics. The different methods have been
successfully tested. The robot always manages to reach the target while avoiding
obstacles. The following conclusions can be drawn:
• With the two algorithms PSO and Dijkstra, the shortest possible path can
be planned, except that the application of Dijkstra becomes more difficult to
implement for complex environments.
• The method based on the Voronoï diagram can guarantee navigation away from
obstacles, but on the other hand, it cannot give the optimal path.
• Comparing these different methods, we can conclude that, to perform the
autonomous navigation of a mobile robot in a known or unknown environment,
PSO and Fuzzy Logic are the best to use. Moreover, the PSO is the most powerful
method since it gives as result the shortest possible path.

The purpose of the future work is to propose new algorithms to contribute to the path
planning of a mobile robot, by a conception of an intelligent control system able to
solve complex driving situations.

References

1. Syed Abdullah, F., Sani Lyal, A., Mokhairi, M., Azrul Amri, J.: Robotic indoor path planning
using Dijkstra algorithm with multi-layer dictionaries. Paper presented at the 2nd International
Conference on Information Science and Security, Seoul, Korea, 2015. https://doi.org/10.1109/
ICISSEC.2015.7371031
2. Ayawli, B.B.K., Mei, X., Shen, M., Appiah, A., Kyeremeh, F.: Mobile robot path planning
in a dynamic environment using Voronoï diagram and computation geometry technique 7,
86026–86040 (2019). https://doi.org/10.1109/access.2019.2925623
3. Seil, R., Oskoei, M.: Mobile robot path planning by RRT* in dynamic environments. Int. J.
Intell. Syst. Appl. 7, 24–30 (2015). https://doi.org/10.5815/ijisa.2015.05.04
4. Kovacs, B., Szayer, G., Tajti, F., Burdelis, M., Korondi, P.: A novel potential field method
for path planning of mobile robots by adapting animal motion attributes. Robot. Auton. Syst.
(2016). https://doi.org/10.1016/j.robot.2016.01.007
Path Planning Using Particle Swarm Optimization and Fuzzy Logic 251

5. Azzabi, A., Nouri, Kh.: Path planning for autonomous mobile robot using the potential field
method. Paper presented at the International Conference on Advanced Systems and Electric
Technologies, Tunisia, 14–17 July 2017
6. Batik Garip, Z., Karayel, D., Ozkan, S., Atali, G.: Path planning for multiple mobile robots
using A* algorithm. Acta Phys. Pol. A 132, 685–688 (2017). https://doi.org/10.12693/APhysP
olA.132.685
7. Chouhan, A.S., Parhi, D.R., Chhotray, A.: Control and balancing of two-wheeled mobile robots
using Sugeno fuzzy logic in the domain of AI techniques. Paper presented at Emerging Trends
in Engineering, Sciences and Manufacturing, Sarang, India (2018)
8. Elkari, B., Ayad, H., Elkari, A., Mjahed, M., Pozna, C.: Design and FPGA implementation of
a new Intelligent behaviors fusion for mobile robot using fuzzy logic. Int. Rev. Autom. Control
12 (2019). https://doi.org/10.15866/ireaco.v12i1.14802
9. Elkari, B., Ayad, H., Elkari, A., Mjahed, M.: A new approach of behavior-based fuzzy control
for mobile robot navigation. Int. Rev. Autom. Control 10 (2017). https://doi.org/10.15866/ire
vaaco.v10il.14390
10. Cacerez Florez, C., Rosarion, J., Amaya, D.: Control structure for a car-like robot using artificial
neural networks and genetic algorithms. Neural Comput. Appl. (2018). https://doi.org/10.1007/
s00521.018.3514.1
11. Chen, G., Liu, J.: Mobile robot path planning using ant colony algorithm and improved potential
field method. Comput. Intell. Neurosci. (2019). https://doi.org/10.1155/2019.1932812
12. Ganapathy, V., Sudhakara, P., Jia lie, T., Parasuraman, S.: Mobile robot navigation using
amended ant colony algorithm. Indian J. Sci. Technol. (2016). https://doi.org/10.17485/ijst.
2016.v9i45.102431
13. Wu, M., Chen, E., Shi, Q., Zhou, L., Chen, Z., Li, M.: Path planning of a mobile robot based on
improved genetic algorithm. Chin. Autom. Congr. (2017). https://doi.org/10.1109/cac.2017.
8243983
14. Kun, Su., Wang, Y.: XinNan Hu. Robot path planning based on random coding particle swarm
optimization, Int. J. Adv. Comput. Sci. Appl. (2015)
15. Shahab Alam, M., Usman Rafique, M., Umer Khan, M.: Mobile robot path planning in static
environments using particle swarm optimization. Int. J. Comput. Sci. Electron. Eng. (2015)
16. Dewang, H.S., Mohanty, P.K., Kundu, S.: A robust path planning for mobile robot using smart
particle swarm optimization 133, 290–297 (2018) (Elsevier Ltd.). https://doi.org/10.1016/j.
procs.2018.07.036
17. Mohanty, P.K., Parhi, D.R.: Optimal path planning for a mobile robot using cuckoo search
algorithm. J. Exp. Theor. Artif. Intell. (2016). https://doi.org/10.1080/0952813X.2014.971442
18. Bakdi, A., et al.: Optimal path planning and execution for mobile robots using genetic algorithm
and adaptative fuzzy-logic control. Robot. Auton. Syst. (2016). https://doi.org/10.1016/j.robot.
2016.10.008
19. Singh, S., Roy, S., Singh, K., Khelchandra, T.: Motion planning of a mobile robot using the
fuzzy-GA method along with three path concepts in a dynamic environment. J. Intell. Fuzzy
Syst. 35, 1445–1457 (2018). https://doi.org/10.3233/jifs.169686
20. Carballo, E.A.S., Morales, L., Trujillo-Romero, F.: Path planning for a mobile robot using a
genetic algorithm and artificial bee colony (2017). https://doi.org/10.1109/icmeae.2017.16
21. Pande, A., Parhi, D.R.: Optimum path planning of mobile robot in unknown static and dynamic
environments using Fuzzy-Wind Driven Optimization algorithm. Def. Technol. (2017). https://
doi.org/10.1016/j.dt.2017.01.001
Prediction of Robot Localization States
Using Hidden Markov Models

Jaouad Boudnaya, Amine Haytoumi, Omar Eddayer, and Abdelhak Mkhida

Abstract The Groundhog robot is a robot built for the first time by the CMU
(Carnegie Mellon University) Mine Mapping Team in 2003, which could explore
and create the map of an abandoned coal mine. The prediction of the robot localiza-
tion with precision still a big problem, for this reason, our study consists of a robot,
which can move within an area of 9 squares. This robot is equipped with the sensing
system, which detects obstacles in four directions: north, south, east and west. The
sensors have an error rate of e = 25%. The objective of this work is the prediction
of the robot localization using Hidden Markov Model.

Keywords Hidden Markov model · Robot localization · Groundhog robot ·

Prediction

1 Introduction

The Groundhog robot is a robot built by the Carnegie Mellon University (CMU) in
2003 to explore unknown places for mine for example or other goals [1]. It could
explore and create the map of an abandoned coal mine. This robot explored tunnels,
which were too toxic for people to enter and where oxygen levels were too low for
humans to remain conscious. The idea of this robot is generalized thereafter and
applied to other fields, but the prediction of the robot localization with precision still

J. Boudnaya (B) · A. Haytoumi · O. Eddayer · A. Mkhida

ENSAM Meknes, Laboratory of Mechanics, Mechatronics and Control (L2MC), Moulay Ismail
University, Marjane 2, Al-Mansour Road Agouray, PO Box 15290, 50500 Meknes, Morocco
e-mail: [email protected]
A. Haytoumi
e-mail: [email protected]
O. Eddayer
e-mail: [email protected]
A. Mkhida
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 253

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_18
254 J. Boudnaya et al.

a big problem. In this context, several works in the literature study the determination
or the prediction of the robot localization using many methods.
In the paper [2], the authors propose a probabilistic framework for modelling the
robot’s state and sensory information based on a Switching State-Space Model. It
generalizes two probabilistic model: the Kalman filter linear models and the Hidden
Markov Models.
The work presented in the paper [3], presents a method to determine the robot’s
position and orientation in dynamic environments. It is based on places recognition,
and a match between places recognized and the sequence of places that the mobile
robot is able to see during a run from an initial place to an ending place.
In the paper [4], the authors propose an approach to view-based localization
and navigation in outdoor environments. It allows to estimate the robot position
comparing image sequences and to cope with changes in a robot’s heading and
speed as well as view changes using wide-angle images and a Markov localization
scheme.
In [5], the authors propose an approach to estimate the robot coordinates using
real-time measurements by analyzing data coming both from the robot on-board
sensors and from environment objects. Bounded-error state estimation can be an
alternative to particle filtering which is sensitive to non-consistent measures, large
measure errors, and drift of evolution model.
The authors of the paper [6] propose a LiDAR-based robot localization method in
a complex oil and gas environment. Robot localization is achieved in six Degrees of
Freedom thanks to a particle filter framework. A time-efficient likelihood function
is introduced and experiments are carried out in real environments.
In the paper [7], the authors use a methodology to estimate the robot pose by
integrating the Internet of Things. It adds a new dimension to this process by enabling
communications with smart objects at anytime and anywhere.
In our study, we focus on a robot, which can move within an area of 9 squares.
The objective of this work is the prediction of the robot localization using Hidden
Markov Model. It is a suitable approach for the problem of prediction. The modeling
takes into account the robot, which is equipped with sensors that can detect obstacles
in four directions: north, south, east and west. The environment of the system is also
took into account; it contains some obstacles between squares. The proposed paper
is structured as follows: Sect. 2 introduces the Hidden Markov Models. Section 3
presents our case study when considering the hidden Markov Model of the robot and
its environment as well as the results of the prediction of localization. Finally, Sect. 4
concludes the paper.
Prediction of Robot Localization States Using Hidden Markov … 255

2 Markov Models

2.1 Generalities

The markovien approach is the most known for methods used in the probability
treatment of the systems with dynamic behaviour. It is an analytical approach by
states based on the identification of the various states of the concerned system then on
the analysis of the evolution of the mentioned system between the mentioned states.
This approach bears the name of its founder Andrei Markov (1856–1922), a Russian
mathematician. It is especially known for its work on the stochastic processes. A main
subject of its research became later known under the name of the Markov chains and
process of Markov. He published his first results in 1906, which were generalized in
a countable infinite space of states by Andrei Kolmogorov in 1936 [8].

2.2 Markovien Process

A stochastic process is a set of random variables (Xt) t ≥ 0 in values in the whole of

the observations.
A stochastic process is markovien if the probability of passage of the present stage
to the following one does not depend on the past. In other term:
For each (t1, t2, …, tn) tq t1 < t2 < ··· <tn − 1 < tn

P[X(tn) = Xn/X(tn − 1) = Xn − 1, . . . , X(t1) = X1] = P[X(tn) = Xn/X(tn − 1) = Xn − 1] (1)

2.3 Hidden Markov Model (HMM)

The Hidden Markov Model (HMM) is a simple way to model sequential data [9]. It is
composed of a Markov process with states that are not directly observable (hidden).
There exists some state X; which we wish to estimate; that changes over time t
and it depends only on previous state in time t − 1 and not on the events that occurred
before.
We cannot directly observe each state (Hidden state). However, we can observe
the evidence E (a piece of information correlated with the state) which helps us to
estimate X (see Fig. 1).
256 J. Boudnaya et al.

Fig. 1 Hidden Markov model [10]

3 Case Study

3.1 Description of the Robot System

In our study, we consider a robot, which can move within an area of 9 squares. In
the mini-warehouse, there is 4 shelfs located between squares S1S2, S2S3, S7S8 and
S8S9. So, the environment of the Robot consists of 9 descrete states (see Fig. 2).
The time is also discrete. At each subsequent time step, the robot is programmed
to change its position with a probability of 90% and moves randomly to a different
neighboring square. The probabilistic graphical model is represented in Fig. 3.
Our robot is equipped with the sensing system composed of a compass and a
proximity sensor, which detects obstacles in four directions: north, south, east and
west. The sensor values are conditionally independent given the position of the robot
when the robot makes a move; we receive four readings from the sensing system.
Moreover, the sensors have an error rate of e = 25%.
Our model consists of hidden states X1, X2, …, X9 (the unknown location of a
robot in time) and known pieces of evidence E1, E2, …, E5 (the subsequent readings
from the sensors).
There are two tools, which we use to localize the robot [11]:
• Filtering- estimation of the state in time Xt, knowing the state Xt-1 and the
evidence Et.

Fig. 2 The environment of

the robot
Prediction of Robot Localization States Using Hidden Markov … 257

Fig. 3 Probabilistic graphical model

f1 : t = a ∗ Ot ∗ T ∗ f1 : t − 1 (2)

• Prediction filtering without evidence. We make a guess about the Xt knowing only
the state Xt − 1.

f1 : t = d ∗ T ∗ f1 : t − 1 (3)

where:
f1:t current probability vector in time t.
f1:t − 1 previous probability vector in time t − 1.
α normalisation constant.
Ot observation matrix for the evidence in time t.
T transition matrix.

3.2 Transition Model

The transition model is the probability of a transition from state i to state j:

Ti, j = P(Xt = j|Xt − 1 = i) (4)

258 J. Boudnaya et al.

The full transition matrix T, which contains all the possible transitions from state i
to state j, will have 9 × 9 = 81 entries.
⎧ ⎫
⎪
⎪ 0.1 0 0 0.3 0 0 0 0 0 ⎪⎪
⎪
⎪ ⎪
⎪
⎪ 0 0.1 0 0 0.225 0 0 0 0 ⎪⎪
⎪
⎪
⎪ ⎪
⎪ 0
⎪ 0 0.1 0 0 0.3 0 0 0 ⎪⎪
⎪
⎪
⎪ ⎪
⎪ 0.9
⎨ 0 0 0.1 0.225 0 0.9 0 0 ⎪⎪
⎬
T= 0 0.9 0 0.3 0.1 0.3 0 0.9 0
⎪
⎪ ⎪
⎪
⎪ 0 0 0.9 0 0.225 0.1 0 0 0.9 ⎪
⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ 0 0 0 0.3 0 0 0.1 0 0 ⎪⎪
⎪
⎪
⎪
⎪ ⎪
⎪
⎪ 0 0 0 0 0.225 0 0 0.1 0 ⎪⎪
⎪
⎩ ⎭
0 0 0 0 0 0.3 0 0 0.1

We assume that we do not know the initial position of the robot. The logical
approach is to assume a uniform probability distribution over all squares. Since our
environment consists of 9 states, so, the probability that the robot starts in square i is:

P(X0 = i) = 1/9

We can express this probability for all 9 states as follow:

⎧ ⎫ ⎧ ⎫
⎪
⎪ P(X 0 = S1 ) ⎪
⎪ ⎪
⎪ 1/9 ⎪
⎪
⎪
⎪ ⎪ ⎪ ⎪
⎪
⎪ P(X 0 = S2 ) ⎪
⎪
⎪
⎪
⎪
⎪ 1/9 ⎪
⎪
⎪
⎪
⎪ ⎪ ⎪ ⎪
⎪
⎪ P(X 0 = S3 ) ⎪
⎪
⎪
⎪
⎪
⎪ 1/9 ⎪
⎪
⎪
⎪
⎪ ⎪ ⎪ ⎪
⎪
⎨ P(X 0 = S4 ) ⎬ ⎨ 1/9 ⎪
⎪
⎪ ⎪
⎪ ⎪
⎬
f 0 = P(X 0 = S5 ) = 1/9
⎪
⎪ ⎪ 1/9 ⎪
⎪ ⎪
⎪
⎪ P(X 0 = S6 ) ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ P(X 0 = S7 ) ⎪
⎪ ⎪
⎪ 1/9 ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ P(X 0 = ⎪
S8 ) ⎪ ⎪
⎪ ⎪ 1/9 ⎪
⎪
⎪
⎩ ⎪
⎭ ⎩ ⎪
⎭
P(X 0 = S9 ) 1/9

3.3 Evidence Model

For our system, observation is made using direction sensors or an electric compass.
The observation reports represent the detection of the various obstacles and walls
in our warehouse.
Since we are in state i, we have to find the probability that the sensor returns the
reading j:

Oi,j = P(E = j|X = i) = (1 − e)4−d ∗ ed (5)

Prediction of Robot Localization States Using Hidden Markov … 259

where:
e an error rate of a sensor (25%.).
d the discrepancy- a number of signals that are different- between the true values
for tile i and the actual reading.

For our case the different possible states of observations are:

SWE as image of the states {S1, S2, S3}.
NWE as image of the states {S7, S8, S9}.
E as image of the state {S6}.
W as image of the state {S4}.
No detection as image of the state {S5}.
For example, Table 1 gives the different possible states of the robot, knowing that
the reading of the sensors detected is SWE, with the detection d and the probability
P.
The observation matrixes are as follow:
⎧ ⎫ ⎧ ⎫
⎪
⎪ 0.32 0 0 0 0 0 0 0 0 ⎪⎪ ⎪
⎪ 0.04 0 0 0 0 0 0 0 0 ⎪ ⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ 0 ⎪⎪ ⎪
⎪ 0 ⎪ ⎪
⎪
⎪
0 0.32 0 0 0 0 0 0 ⎪
⎪ ⎪
⎪
0 0.02 0 0 0 0 0 0 ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ 0 0 0.32 0 0 0 0 0 0 ⎪
⎪ ⎪
⎪ 0 0 0.04 0 0 0 0 0 0 ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪
⎨ 0 0 0 0.04 0 0 0 0 0 ⎪⎪
⎬
⎪
⎪
⎨ 0 0 0 0.04 0 0 0 0 0 ⎪ ⎪
⎬
SWE = 0 0 0 0 0.01 0 0 0 0 NEW = 0 0 0 0 0.0 0 0 0 0
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ 0 0 ⎪⎪ ⎪
⎪ 0 ⎪ ⎪
⎪
⎪
0 0 0 0 0.04 0 0 ⎪
⎪ ⎪
⎪
0 0 0 0 0 0.04 0 0 ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ 0 0 0 0 0 0 0.04 0 0 ⎪
⎪ ⎪
⎪ 0 0 0 0 0 0 0.32 0 0 ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ 0 0.04 0 ⎪ ⎪ ⎪
⎪ 0 0.32 0 ⎪ ⎪
⎪ 0
⎪
0 0 0 0 0 ⎪
⎪ ⎪
⎪
0 0 0 0 0 0 ⎪
⎪
⎩ ⎭ ⎩ ⎭
0 0 0 0 0 0 0 0 0.04 0 0 0 0 0 0 0 0 0.32
⎧ ⎫ ⎧ ⎫
⎪
⎪ 0.04 0 0 0 0 0 0 0 0 ⎪⎪ ⎪
⎪ 0.04 0 0 0 0 0 0 0 0 ⎪ ⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ 0 ⎪⎪ ⎪
⎪ 0 ⎪ ⎪
⎪
⎪
0 0.04 0 0 0 0 0 0 ⎪
⎪ ⎪
⎪
0 0.04 0 0 0 0 0 0 ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪ 0 0 0.04 0 0 0 0 0 0 ⎪⎪ ⎪
⎪ 0 0 0.04 0 0 0 0 0 0 ⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ 0 0 ⎪⎪ ⎪
⎪ 0 ⎪ ⎪
⎨ 0 0 0.04 0 0 0 0 ⎬ ⎨ 0 0 0 0.32 0 0 0 0 ⎬
E= 0 0 0 0 0.11 0 0 0 0 W= 0 0 0 0 0.11 0 0 0 0
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ 0 ⎪⎪ ⎪
⎪ 0 ⎪ ⎪
⎪
⎪
0 0 0 0 0 0.32 0 0 ⎪
⎪ ⎪
⎪
0 0 0 0 0 0.04 0 0 ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪ 0 0 0 0 0 0 0.04 0 0 ⎪⎪ ⎪
⎪ 0 0 0 0 0 0 0.04 0 0 ⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ 0 0 0.04 0 ⎪ ⎪ ⎪
⎪ 0 0.04 0 ⎪ ⎪
⎪
⎪
0 0 0 0 0 ⎪
⎪ ⎪
⎪
0 0 0 0 0 0 ⎪
⎪
⎩ ⎭ ⎩ ⎭
0 0 0 0 0 0 0 0 0.04 0 0 0 0 0 0 0 0 0.04
⎧ ⎫
⎪
⎪ 0.01 0 0 0 0 0 0 0 0 ⎪ ⎪
⎪
⎪ ⎪
⎪
⎪
⎪ 0 ⎪ ⎪
⎪
⎪
0 0.01 0 0 0 0 0 0 ⎪
⎪
⎪
⎪ ⎪
⎪
⎪ 0 0 0.01 0 0 0 0 0 0 ⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ 0 ⎪ ⎪
⎨ 0 0 0 0.11 0 0 0 0 ⎬
No detection = 0 0 0 0 0.32 0 0 0 0
⎪
⎪ ⎪
⎪
⎪
⎪ 0 ⎪ ⎪
⎪
⎪
0 0 0 0 0 0.11 0 0 ⎪
⎪
⎪
⎪ ⎪
⎪ 0
⎪ 0 0 0 0 0 0.01 0 0 ⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ 0 0 0.01 0 ⎪ ⎪
⎪
⎪
0 0 0 0 0 ⎪
⎪
⎩ ⎭
0 0 0 0 0 0 0 0 0.01
260 J. Boudnaya et al.

Table 1 Possible states of the robot for the observation SWE

i Sensor status in position i Current state given by our sensor d P
S1 0 0.32

S2 0 0.32

S3 0 0.32

S4 2 0.04

S5 3 0.01

S6 2 0.04

S7 2 0.04

S8 2 0.04

S9 2 0.04

3.4 Results and Discussions

Our implementation processes each time step in a sequential way. Thus, our HMM
implementation evaluates the probability that the robot is in a state for each time
step.
For example, for the three first steps (see Fig. 4), we have:
• In the state at time t1 (step 1), the robot has a high probability of being in the state
1 and 7 with an equiprobability of 20% for both.
Prediction of Robot Localization States Using Hidden Markov … 261

Fig. 4 Robot Probability of being in each state for the three first steps

• In the state at time t2 (step 2), the robot has a very high probability of being in
the central edge (State 5), with a probability almost equal to 30%.
• In the state at time t3 (step 3), the robot has a very high probability of being in
State 6 with a probability of 50%.

Therefore, we can easily know the different possible scenarios through the HMM
model, that it gives the different probabilities after time steps. Then, we can calculate
the probability of all predicted paths through its observations series (SWE E W…).

Path = any sequence of states

= Pmaxstep1 , Pmaxstep2 , Pmaxstep3 , Pmaxstep.. , Pmaxstep.n
= E step1 , E step2 , E step3 , E step4 , E step.. E step... E stepn

4 Conclusion

In this paper, we have shown how to model a robot, which can move within an area of
9 squares by using Hidden Markov Model. This robot is equipped with sensors, which
detect obstacles in four directions: north, south, east and west. They have an error
rate of e = 25%. The objective of this work is the prediction of the robot localization
by taking into account the robot, the sensors and the environment obstacles.
In our future works, we will try to develop this model to estimate the Remaining
Useful Life (RUL) for the robot by using the history of data collection from sensors
in each localization to feed the survival model by data. We will also strive to improve
the limitations of the study by using imprecise Hidden Markov Models to calculate
the upper and lower bounds of the RUL.

Acknowledgments In the term of this paper, we thank the Laboratory of Mechanic, Mechatronics
and Control (L2MC) of the ENSAM MEKNES. We do not forget both our colleagues and experts
for the information source. In addition, we thank the steering committee of the Artificial Intelligence
& Industrial Applications (A2IA’2020) to allow us to communicate our works and results.
262 J. Boudnaya et al.

References

1. Ferguson, D., Morris, A., Hähnel, D., Baker, C., Omohundro, Z., Reverte, C., Thayer, S.,
Whittaker, C., Whittaker, W., Burgard, W., Thrun, S.: An autonomous robotic system for
mapping abandoned mines. In: Thrun, S., Saul, L., Scholkopf, B. (eds.) Proceedings of NIPS.
MIT Press (2003)
2. Baltzakis, H., Trahanas, P.: Hybrid mobile robot localization using switching state-space
models. In: Proceedings of the 2002 IEEE International Conference on Robotics & Automation,
Washington, pp. 366–373 (2002)
3. Aycard, O., Laroche, P., Charpillet, F.: Mobile robot localization in dynamic environments using
places recognition. In: Proceedings of the 1998 IEEE International Conference on Robotics &
Automation, Belgium, pp. 3135–3140 (1998)
4. Oishi, S., Inoue, Y., Miura, J., Tanaka, S.: SeqSLAM++: view-based robot localization and
navigation. Robot. Auton. Syst. 112, 13–21 (2019)
5. Colle, E., Galerne, S.: A robust set approach for mobile robot localization in ambient environ-
ment. Auton. Robots 43(3), 557–573 (2019) (Springer). https://doi.org/10.1007/s10514-018-
9727-4. hal-01758277
6. Merriaux, P., Dupuis, Y., Boutteau, R., Vasseur, P., Savatier, X.: Robust robot localization in
a complex oil and gas industrial environment. J. Field Robot. 35(2), 213–230 (2018) (Wiley).
hal-01535781
7. Pagetti, C.: Module de sûreté de fonctionnement. Chaine de Markov, pp. 34–41. ENSEEIHT
3ème TR-Option SE (2012)
8. Colle, E., Galerne, S.: A multi hypothesis set approach form mobile robot localization using
heterogeneous measurements provided by the Internet of Things. Robot. Auton. Syst. 96, 102
113 (2017) (Elsevier). https://doi.org/10.1016/j.robot.2017.05.016. hal-01621289
9. Eddy, S.R.: Hidden Markov models. Sequences and topology. Curr. Opin. Struct. Biol. 6,
361–365 (1996)
10. Delmas, A.: Contribution à l’estimation de la durée de vie résiduelle des systèmes en présence
d’incertitudes, Thèse de Doctorat de l’Université de Technologie de Compiègne (2019)
11. Practical tutorial—Robot localization using Hidden Markov Models. https://dtransposed.git
hub.io/blog/Robot-Localization. Accessed 20 Jan 2020
A New Approach for Multi-agent
Reinforcement Learning

Elmehdi Amhraoui and Tawfik Masrour

Abstract The classical approach of reinforcement learning for single agent is based
on the concept of reward that comes only from the environment. By trial-and-error, the
agent has to learn to maximize its total accumulated reward. Several algorithms and
techniques were developed for a single agent reinforcement learning. Our purpose
is to benefit from all done work in reinforcement learning of an agent and extend
it to multi-agent system. we have proposed a new approach that is based on the
following idea: Communication or cooperation of a team can be achieved through
mutual reinforcement of agents, that means agents can give and receive rewards
from other agents and not only from the environment. The goal of each agent is to
maximize the global accumulated reward received from environment and from other
agents. We treat the multi-agent system as a unique entity. We define, state, action of
the system and the value of each state with respect to the global policy of the system.

Keywords Artificial intelligence · Reinforcement learning · Multi-agent system ·

Markov decision process · Markov game · Control theory

1 Introduction

Multi-agent reinforcement learning is a difficult problem in artificial intelligence with

a growing interest in theoretical research and practical applications. In 1993, Ming
Tan demonstrated that agents learn not only by trial and error but also by cooperation
by sharing instant information [1]. In 2000, Y. Nagayuki; S. Ishii and K. Doya

E. Amhraoui (B)
Artificial Intelligence for Engineering Sciences Team (IASI), Doctoral Studies Center,
ENSAM-Meknes, University My Ismail, Meknes, Morocco
e-mail: [email protected]
T. Masrour
Department of Mathematics and Informatics, Artificial Intelligence for Engineering Sciences
Team (IASI), ENSAM-Meknes, University My Ismail, Meknes, Morocco
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 263

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_19
264 E. Amhraoui and T. Masrour

have proposed a multi-agent reinforcement learning method based on predicting the

actions of other agents [2]. However, most research on reinforcement learning has
focused on single agents. In fact, several approaches and algorithms exist for learning
by reinforcement of a single agent and the most famous approach is Q-reinforcement
learning [3, 4].
In this article, we have proposed a consistent model of multi-agents system and
also extend all existing techniques and algorithms in the reinforcement learning of
an agent to the reinforcement learning of multi-agent system.
Our model is in principle based on the idea of distributing reward between envi-
ronment and other agents, that means that the reward of each agent is equal to the
reward coming from the environment plus the reward coming from other agents.
One of the greatest advantages of our model is maintaining the general outlines
of Markov game concept, moreover no external information or oracle is needed for
the system learning.
In order to validate our model, we consider a list of grid world examples with
different size, different goals and different number of agents. In each example, each
agent has to choose a unique goal and to optimize its trajectory.

2 Background

2.1 What Is an Agent?

An agent is anything that can perceive its environment via sensors and acts upon that
environment through actuators. Intuitive examples of agents are humans or robots
[3]. In this section we will study the problem of how an agent can choose the best
possible action at each time step, given what it knows about its environment.
For simplification purposes, we will assume a discrete set of time steps t =
0, 1, 2, . . . ,. In each time step t the agent must choose an action at from a finite
set of actions A.
We will call a state of the environment and denote by st all relevant information
that is contained in the environment at time step t. The set of all states of the environ-
ment will be denoted by S. The environment can be either discrete or continuous.
Environment with a finite number of states will be our main focus.

2.2 Observability

An important property that characterizes a world (environment) from the point of

view of an agent depends on the perception of the agent. We say that the environment
is (fully) observable to an agent if its observation θt completely reveals the state of
the environment, that is, st = θt .
A New Approach for Multi-agent Reinforcement Learning 265

On the other side, in a partially observable environment, the observation θt of the

agent provides only partial information about the state st in the form of a stochastic
observation model. Partial observability implies that the current observation θt does
not completely reveal the true state of the environment, the agent is therefore forced
to assign a probability p(st |θt ) to each state st .
Partial observability generally results from the noise in the agent’s sensors. For
example, due to sensor defect, every time the agent visits the same state it may
perceive something different. For the sake of simplification, we will assume in the
next sections that the world is fully observable to the agent.

2.3 Transition Model

As mentioned previously, at each time step t the agent takes an action at from a
finite set of actions A and the environment state changes as a result of this action.
A transition model (or environment model) specifies how the environment changes
when the agent chooses an action. If the current state is st and the agent takes an
action at , we can distinguish the following two cases:
– deterministic model: In this case, the transition model maps a state–action pair
(st , at ) to a single new state st+1 . In the grid world example, every move changes
the position of the agent in a deterministic manner.
– Stochastic model: In this case, the transition model maps a state–action pair (st , at )
to a probability distribution p(st+1 |st , at ) over states. st+1 is a random variable with
values in S, each value with a probability p(st+1 |st , at ). Most real control problems
involve stochastic transition models.

2.4 Utility-Based Decision Making

An agent holds preferences between different states. For example, in a grid world,
agent will mostly prefer to move forward the goal, and will prefer less to move far
away from the goal, and so on.
To formalize the notion of state preferences, we can assign to each state s a
number U (s) ∈ R that is called the utility of state s. For two states s and s  , we have
U (s) > U (s  ) if and only if the agent prefers state s to state s  , and U (s) = U (s  ) if
and only if the agent does not differentiate between s and s  .
Suppose we know the utility function U , the fundamental question is how to
exploit this function in order to choose the best action at each time step t. Assuming
that the environment has a stochastic transition model p(st+1 |st , at ) and that the
agent is actually in that state st , the Utility-based decision making is based on the
maximization of the expected utility. That means that the optimal action at∗ of the
agent at time step t should maximize the expected utility, that is:
266 E. Amhraoui and T. Masrour

at∗ = argmaxat E(U (s  )) = argmaxat p(st+1 |st , at )U (st+1 ) (1)
st+1

We will call the optimal policy of the agent the function denoted by π ∗ , that maps
each environment state to an optimal action, that is:

π ∗ (st ) = argmaxat p(st+1 |st , at )U (st+1 ) (2)
st+1

2.5 Reinforcement Learning (RL)

RL is learning how to map environment states to actions so as to maximize a numerical

reward signal. The agent is not told which actions to take, but instead has to discover
which actions yield the most reward by trying them.
To formalize the problem of reinforcement learning we will use ideas from optimal
control theory, specifically the concept of Markov Decision Process (MDP).

2.5.1 Finite Markov Decision Process (FMDP)

FMDP is a framework for learning by interaction with the environment to achieve

a goal, the learner is called an agent, and at each time step, the agent has to select
an action and the environment responds by giving a reward (wich is a numerical
value that the agent seeks to maximize over time through its choice of actions) and
presenting a new state Fig. 1. Formally speaking, a Finite Markov Decision Process
is defined by the following elements:

Fig. 1 The agent-environment interaction in a Markov decision process

A New Approach for Multi-agent Reinforcement Learning 267

1. A finite set of states denoted by S,

2. A finite set of actions denoted by A,
a
3. The transition probabilities Ps,s  that satisfies the Markov property.


 = P(st+1 = s |st = s, at = a)
a
Ps,s (3)

4. A reward function R : S × A −→ R, so that the agent receives the reward

R(s, a) while taking the action a at state s,
5. A discount factor μ, which shows the present value of future rewards. In next
section, we will show that for μ ∈ [0, 1], the return of the agent is well defined
for bounded reward functions.

2.5.2 Return and Value Function

Previously we have said that the purpose of the agent is to maximize the total accu-
mulated reward, this idea is based on a fundamental hypothesis wich says: All of
what we mean by goals can be achieved through the maximization of the expected
value of the cumulative received reward [3].
In general we seek to maximize the expected return after time step t, where the
return is denoted by G t . The return is defined as the sum of discounted reward:
∞

Gt = μk Rt+k+1 (4)
k=0

0 ≤ μ ≤ 1 is the discount factor and Rt+k+1 is the received reward at time step
t + k + 1.
Most of reinforcement learning algorithms are based on estimating value functions
of states (or of state-action pairs). Value function of state estimates how good it is
for the agent to be in a given state, and value function of state-action pair estimates
how good it is to perform a given action-state pair.
The value functions are defined in terms of expected return. Keeping in mind that
the return that the agent can expect depends on which actions it will take. As a result,
value functions are defined with respect to policies.
For an MDP, the value function of a state s under a policy π , denoted vπ (s), is
the expected return when starting in s and following π . Formally, we have:

vπ (s) = E π (G t |st = s) (5)

Similarly, we define the value function of an action-state pair (s, a) under a policy
π , denoted qπ (s, a), as the expected return starting from s, taking the action a, and
then following policy π . formally, we can write:

qπ (s, a) = E π (G t |st = s, at = a) (6)

268 E. Amhraoui and T. Masrour

2.5.3 Bellman Equation for Value Functions

A fundamental property of value functions is that they satisfy a recursive relationship

called Bellman equation. For any policy π and for any state s ∈ S, the state value
function satisfies the following equation:
   
 
vπ (s) = π(a|s) a
Ps,s  r (s, a, s ) + μvπ (s ) (7)
a s

The equation above for v p expresses a relationship between the value of a state and
the values of its successor states. For this reason that equation represents the basis
of many methods to compute and learn v p .
The state-action value function, qπ satisfies also a recursive relationship called
Bellman equation for qπ . For any policy π and for any state-action pair (s, a), the
Bellman equation for qπ is:
 
 
    
qπ (s, a) = a
Ps,s  r (s, a, s ) + μ π(a |s )qπ (s , a ) (8)
s a

2.5.4 Optimal Policies and Optimal Value Functions

For a finite MDP, we can define a partial order over all possible policies. We say that
a policy π is similar or better than another policy π  if and only if vπ (s) ≥ vπ  (s) for
all s ∈ S. Formally speaking:

π ≥ π  ⇔ vπ (s) ≥ vπ  (s) ∀s ∈ S (9)

For a finite MDP, there is always at least one policy that is better than or equal to
all other policies, it is called optimal policy and denote by π ∗ . While there may
be more than one optimal policy, they share the same value state function, called
optimal value state function, which is denoted by v ∗ .

v ∗ (s) = maxπ vπ (s) (10)

Optimal policies share also the same optimal state-action value function, denoted q ∗
and defined as follow:
q ∗ (s, a) = maxπ qπ (s, a) (11)

The optimal value function v ∗ satisfies a recursive equation called Bellman opti-
mality equation:
  
v ∗ (s) = maxa a
Ps,s  ∗ 
 r (s, a, s ) + μv (s ) (12)
s
A New Approach for Multi-agent Reinforcement Learning 269

2.5.5 Policy Evaluation

Here we consider the problem of computing the state value function of a given policy
π , this problem is called policy evaluation. We suppose that the environment’s
a
dynamic is completely known, wich means that Ps,s  is known for each element

(s, a, s ) in S × A × S. As we have seen previously, the state value function satisfies
Bellman equation:
   
 
∀s ∈ S vπ (s) = π(a|s) a
Ps,s  r (s, a, s ) + μv(s ) (13)
a s

The system above is a system of |S| linear equations with |S| unknowns. For our
purpose we will use iterative method to approximate the exact value of vπ (s) instead
of direct computational method.
Below we give the pseudo-code of iterative policy evaluation algorithm:
Algorithm 1: Iterative policy evaluation for approximating V vπ
Input: π The policy to be evaluated, precision .
initialization: Arbitrary initialization of V (s) for all s ∈ S and
V ( terminal ) = 0
←− 0
while ≤  do
for Each s ∈ S do
v(s) ←− V (s)  a  
 
V (s) ←− a π(a|s) s  Ps,s  r (s, a, s ) + μV (s )

←− max( , v(s) − V (s))

end
end
Result: V

2.5.6 Policy Improvement

The reason why we compute the value function of a given policy is to help us to find
a better policy by improving the existing one.
Let π be a deterministic policy for which vπ is known, we define a new policy π 
as follows:

π  (s) = argmaxa qπ (s, a) (14)

  
 
= argmaxa a
Ps,s  r (s, a, s ) + μvπ (s ) (15)
s

By the Policy Improvement Theorem [3], π  is better or equal to π . The process

of constructing a new policy from an old policy, by making it greedy with respect to
the value function, is called policy improvement.
270 E. Amhraoui and T. Masrour

2.5.7 Policy Iteration

After viewing the basics of policy improvement method, it is obvious to think on

continuing improvement until we obtain an optimal policy, we call this method Policy
iteration algorithm. A complete algorithm of this method is given below:
Algorithm 2: Policy Iteration algorithm for approximating π ∗
1-Evaluation step:
Input: precision .
initialization: Arbitrary initialization of V (s) and π(s) for all s ∈ S
←− 0
while ≤  do
for each s ∈ S do
v(s) ←− V (s)  a  
 
V (s) ←− a π(a|s) s  Ps,s  r (s, a, s ) + μV (s )

←− max( , v(s) − V (s))

end
end
Result: V
2-Improvement step:
policy-stable←− true
for each s ∈ S do
old-action ←− π(s)  
 
π(s) ←− argmaxa s  Ps,s a
 r (s, a, s ) + μV (s )

if old-action= π(s) then

policy-stable←− false
end
end
if policy-stable= true then
Stop and return π
end
else
Go to 1.
end

3 Overview About Multi-agent System

In this section we will show how to extend the reinforcement learning approach to
multi-agent system (MAS) where several agents are simultaneously taking actions.
Firstly we will describe the generalization of MDP framework called Markov game
or stochastic game. Secondly we will discuss different existing models for rein-
forcement learning of MAS.
A New Approach for Multi-agent Reinforcement Learning 271

3.1 Markov Game

A Markov game [5] can be regarded as the multi-agent extension of MDP. Formally,
a Markov game is defined by the following components:
• A set of n > 1 agents,
• A finite set of environment states denoted by S,
• For each agent i, a finite set of actions denoted by Ai ,
• A stochastic transition model P(s  |s, a) such that a is the joint action at state s,
• For each agent i, a reward function Ri : S × A × S −→ R, that gives the agent i
a reward Ri (s, a, s  ) when the joint action a is taken at state s,
• A discount factor μ.
The main difference between MDP and Markov game is that in Markov game the
transition depends on the joint action of agents, and each agent may receive a different
reward.
Reinforcement learning for MAS is still an active area of research and we will
briefly discuss many approaches that have been proposed in the literature.

3.2 Independent Learning

Independent reinforcement learning in a Markov game means that each agent of the
system learns separately it’s policy without considering other agents. Each agent can
use the policy iteration algorithm that we have treated previously to find its optimal
policy as in MDP. However, experiences show that this approach does not perform
well in the case of Markov game [5].

3.3 Coupled Learning

Better result can be obtained if each agent tries to include interactions with other
agents in its learning phase. In this case their ways of learning are coupled.
A standard approach to model coupled learning is to consider the value functions
of each agent over joint action of the system. A simple update of policy iteration
algorithm is given below:


qi (s, a) = Ri (s, a) + μ a
Ps,s  Vi (s ) ∀s, a (16)
s

Vi (s) = C(q1 (s, a), · · · , qn (s, a)) ∀s (17)

where C is an appropriate function for problem in hand.

272 E. Amhraoui and T. Masrour

3.4 Cooperative Multi-agent Learning

In this section, we consider a fully cooperative multi-agent system and each agent
has a different reward function Ri . A collaborative multi-agent Markov decision
process (CM-MDP) is special case of Markov game model where each agent seeks
to maximize the future global reward R:


n
R(s, a) = Ri (s, a) (18)
i=1

Assuming that each agent i observes tuple (s, a, Ri , s  ), following the general multi-
agent learning approach of coupled learning, we can derive a Q-learning algorithm
for a CM-MDP.

Q i (s, a) = (1 − λ)Q i (s, a) + λ(Ri (s, a) + μVi (s  )) (19)


Vi (s  ) = Q i (s  , a ∗ ) where a ∗ = argmaxa  Q i (s  , a  ) (20)
i

When all agents share the same reward function, the multi-agent system can be
converted to one ‘big agent’ and then one can use standard MDP techniques [6].

4 Basics of Our Model for Fully Cooperative Agents

The model we have proposed is based on the following hypothesis: All what we mean
by collaboration and communication between agents can be achieved by exchanging
rewards between agents, that means that each agent can send and receive reward from
others and not only from the environment, and then each agent seeks to maximize the
global accumulated reward of the system. It is clear that our model is fundamentally
different from existing approaches in the reinforcement learning paradigm. In fact
the reward was a monopoly and a special property of the environment only, and from
now on, it may come from other agents.

5 Tests and Model Validation

In this section we will validate our model by considering a list of grid world examples
with different size, different goals and different number of agents. In each example,
agents have to learn to choose different goals and to optimize their trajectories.
A New Approach for Multi-agent Reinforcement Learning 273

Fig. 2 Grid with size of 4 × 4

Table 1 Reward system for example 1

State ENV to A1 ENV to A2 A1 to A2 A2 to A1
S1t = 12 and S2t = 15 0 0 0 0
S1t = 15 and S2t = 12 0 0 0 0
S1t = S2t 0 0 −1 0
Other cases −1 −1 0 0

5.1 Test 1

We consider a grid with the size of 4 × 4 cells, each cell of the grid is associated
with a number between 0 and 15 as in Fig. 2.
The first example consists of giving two fixed goals G 1 = 12, G 2 = 15 and two
agents A1 , A2 , each agent has to learn to choose and achieve a different goal. We
denote by Sit the state of the agent i at time step t. The reward system that we have
adopted for this example is explained in the Table 1.
After setting the reward system, we have applied our model using the policy
iteration method and we have obtained good results, in fact each agent has learned
to choose a different goal and proceed forward it as shown in Fig. 3.
The Figure below represent the evolution of the norm of value state function with
respect to the improvement steps of policy iteration method.
274 E. Amhraoui and T. Masrour

Fig. 3 Dynamics of two agents toward two goals

Fig. 4 Dynamics of tree agents toward tree goals

5.2 Test 2

In The second example, we use a grid with the size of 6 × 6, 3 fixed goals G 1 =
30, G 2 = 33, G 3 = 35 and tree agents A1 , A2 , A3 . As in the first example, each agent
has to choose a different goal. We denote by Sit the state of the agent i at time step t.
For the reward system, we will use the extension of the one used in the first example.
In figure Fig. 4, we show the dynamics of 3 agents after learning.
In figures below, we show the evolution of accumulated reward and the norm of
value state function with respect to the improvement step of policy iteration method.

As expected, the reward evolution figure shows how the policy iteration algorithm
improves the initial policy in each step. Furthermore, the other figure shows the fact
that the policy iteration algorithm converges to the optimal policy.

6 Conclusion and Perspectives

Until now, our model works well in the previous tests despite that we have used a
simple algorithm for learning, which is the policy iteration algorithm. In the future
work, we will combine our approach with other algorithms in reinforcement learning
A New Approach for Multi-agent Reinforcement Learning 275

such as Q-learning [7] and Deep reinforcement learning [8, 9]. In addition, we will
test the flexibility of our model by considering more complex problem, such as the
Automated Guided Vehicle System [10]. Finally we will compare the flexibility of
our model with other existing approaches in reinforcement learning of multi-agent
system.

References

1. Tan, M.: Cost-sensitive learning of classification knowledge and its applications in robotics.
Mach. Learn. 13(1), 7–33 (1993)
2. Nagayuki, Y., Ishii, S., Doya, K.: Multi-agent reinforcement learning: An approach based on the
other agent’s internal model. In: Proceedings Fourth International Conference on MultiAgent
Systems, pp. 215–221. IEEE (2000)
3. Sutton, R.S., Barto, A.G.: Introduction to Reinforcement Learning, vol. 2, no. 4. MIT Press,
Cambridge (1998)
4. Gosavi, A.: Reinforcement learning: a tutorial survey and recent advances. INFORMS J. Com-
put. 21(2), 178–192 (2009)
5. Vlassis, N.: A concise introduction to multiagent systems and distributed artificial intelligence.
Synthesis Lect. Artif. Intell. Mach. Learn. 1(1), 1–71 (2007)
6. Kok, J.R., Vlassis, N.: Sparse cooperative Q-learning. In: Proceedings of the Twenty-First
International Conference on Machine Learning, p. 61. ACM (2004)
7. Watkins, C.J., Dayan, P.: Q-learning. Mach. Learn. 8(3–4), 279–292 (1992)
8. Van Hasselt, H., Guez, A., Silver, D.: Deep reinforcement learning with double q-learning. In:
Thirtieth AAAI Conference on Artificial Intelligence (2016)
9. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Petersen, S.:
Human-level control through deep reinforcement learning. Nature 518(7540), 529–533 (2015)
10. Erol, R., Sahin, C., Baykasoglu, A., Kaplanoglu, V.: A multi-agent based approach to dynamic
scheduling of machines and automated guided vehicles in manufacturing systems. Appl. Soft
Comput. 12(6), 1720–1732 (2012)
Recommender System for Most Relevant
K Pick-Up Points

Ayoub Berdeddouch , Ali Yahyaouy , Younès Bennani ,

and Rosanna Verde

Abstract According to Microsoft, by 2025, 100% of new cars will be connected

and by 2030, 15% of new cars will be autonomous, and will take care of sending,
receiving and analyzing “large amounts of data”. Automobiles are becoming data
centers on wheels. All these pieces of information can be used by many stakehold-
ers (road authorities, leasing companies, municipalities, car manufacturers, insurers,
workshops, emergency services, etc.) in the mobility space to improve processes.
Taxi’s industry is a concrete example that is evolving rapidly. New competitors and
technologies are changing the way traditional taxi services do business. In spite of
this growth has created new effectualness, it has also created new challenges. Exploit-
ing and understanding of the taxi supply and demand could increase the efficiency
of the city’s taxi fleet management system. This paper describes a recommender
system that recommends the most relevant pick-up points for a driver. It is based on
spatiotemporal features, through data preprocessing and real-time recommendations.
These recommendations could inform the drivers on where to position their taxi’s. We
implemented a recommender system based on KMeans+Regression. Putting in the
test with various features sets, we were able to achieve positive results. The effective-
ness of our approach is demonstrated by the results, achieving approximately 98%
accuracy in most cases.

Keywords Recommender system · Spatiotemporal features ·

Kmeans+regression · Pick-up points · Taxi · Mobility

A. Berdeddouch (B) · A. Yahyaouy

LISAC Laboratory, University Sidi Mohamed Ben Abdellah, Fez, Morocco
e-mail: [email protected]
A. Yahyaouy
e-mail: [email protected]
Y. Bennani
LIPN Laboratory, University Paris 13, Villetaneuse, France
e-mail: [email protected]
R. Verde
Department of Mathemathics and Physics, University of Campania, Caserta, Italy
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 277

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_20
278 A. Berdeddouch et al.

1 Introduction

In recent years, after talking about the Godwin point for a long time, the world has
discovered the Terminator point of saying that strong Artificial Intelligence (AI) will
replace and subjugate humans. In the climax of this trendy subject, the duel at the top
between the champions of both camps: Elon Musk at the head of IA-Fearful on one
side and Mark Zuckerberg on the other. Ultimately, this new Buzz-Word that is AI
echoes others used so far: “Smart Data”, “Machine Learning”, “Computer Science”,
etc. Because if the name changes and gives the illusion of novelty, the reality has
been tangible for many years.
The novelty nevertheless resides in the fact that today, there is no need to be part
of the GAFAM1 to create an AI. Artifact has made it its mission: to bring marketing
into the air of AI and to create, alongside brands, AI that matters. For their client’s
experience and for their organization. It is clear and well accepted that IA will change
our lives, but we do not know when it will happen. Harnessing these technologies
will allow businesses, as individuals, to accomplish more tasks with less effort and
to automate complex processes. AI and machine learning technologies will have a
significant impact on the automotive and mobility industries as they will bring new
products and business models rather than just productivity improvements. One of
the ways to make autonomous cars safer is to train their algorithms in simulated
environments. “We are working a lot on the efficiency of this simulation, which is
linked to the quality of the data. The challenge of the research is to constitute the
most realistic test bases possible” specifies Paul Labrogère.2
How then to go from the intelligent vehicle to the intelligent transport system?
The challenge is crucial with the doubling of the demand for mobility by 2050, and
an urban population which will have grown by 70% to reach 6.6 billion people (out
of 9.8 billion humans in total). With traffic congestion already critical in many cities
and time spent in traffic jams which are growing by 2.6% per year worldwide. The
solution could go through the concept of Mobility as a service (MaaS). Its objective
is to guarantee the user access to the most efficient mobility, at a given time, to go
from point A to point B with a multimodal offer that can combine public transport,
cycling, Mountain biking or just walking. MaaS is the future of transportation, but
it is a future that already exists in many cases. With the combined know-how of
traditional car manufacturers, technical leaders, ambitious startups and governing
bodies, the MaaS movement is well underway.
The ability to recommend taxi ridership could present valuable insights to city
planners and taxi dispatchers in answering questions such as how to position cabs
where they are most required, how many ones to transmit, and how ridership varies
over time. Our project focuses on recommending the most relevant K pick-up points
for a driver given a specific time, a location within the city, a detour, and the number
of pickups. The problem is formulated in terms of the following inputs and outputs:

1 GAFAM: the five big firms, Google, Amazon, Facebook, Apple, Microsoft.
2 Paul Labrogère : DG of IRT SystemX.
Recommender System for Most Relevant K Pick-Up Points 279

Inputs:
• Day of Week: Monday
• Hour: 8 am
• Driver Location (lat, long) : −74.006149, 40.744919
• Deviation Threshold: 43-Metres
• Number of Pickup points: 10.
Output: 10 most relevant Pick-up points ranked by relevance.

2 Related Work

With the recent development of the wireless location, network communication, and
data mining technology, using geolocation to provide insight and commercial ser-
vices for taxi drivers have attracted the attention of academics and industries. Many
methods exist, such as Microsoft Research Asia [1–5], Rutgers University [6–8]
and Wuhan University [9], aim at reducing empty running time, improving energy
efficiency and maximizing economic benefits.
Zhang, et al. [1] implemented a recommender system for pick-up points for Taxi-
drivers based on Spatiotemporal Clustering (applying KMeans and hierarchical clus-
tering algorithms). Firstly, they capture the historical pickup points by analyzing
their intervals. Then, cluster these points at different times and different regions to
create–candidate pick-up points. Secondly, after ranking the candidate pick-up points
around the taxi, the top-5 valuable pick-up points are recommended for taxi drivers.
Yu Zheng, et al. implemented a series of researches based on GPS trajectory: GeoLife
(Geography Life) [2, 3]. It is an application system that uses GPS as data-centered
and shown on the electronic map which shares GPS data and exchange experience.
T-Drive (Taxi Drive) [4, 5] which is an application system based on vehicle GPS
data can provide business services for taxi driver and help for traffic supervision,
city planning and so on. One study [6] about the T-Drive is divided taxi status into
three situations: load driving, no-load driving and parking. Base on the status change,
researchers according to the trajectory of experienced drivers to find the location of
taxi pick-up or unload passengers and count the pick probability for each location,
and then recommend picking-up and waiting points for taxi drivers and passengers.
Yong Ge et al. presented an energy-efficient mobile recommender system [7]
which is a taxi driving, fraud detection system [8] and a taxi business intelligence
system [9] based on the GPS data. The taxi business intelligence systems help taxi
companies to improve their business performances by understanding the behaviors
of both drivers and customers. The taxi driving fraud detection system extract two
aspects of shreds of evidence at first: travel route evidence and driving distance
evidence. Combining the two aspects of pieces of evidence based on DempsterShafer
theory to reveal the taxi driving fraud. The energy-effcient mobile recommender
system extracts energy-effcient transportation patterns from GPS traces and utilizes
these patterns to help traffc department directing traffc. Based on GPS data collected
280 A. Berdeddouch et al.

by taxis, Qingquan Li et al. implemented researched experience and regular pattern

of taxi drivers in routing. They put forward the routing empirical model and path
planning algorithm to offer help for vehicle travel [10].
Brendan Guillouet et al. proposed a new method for predicting the final destination
of vehicle journeys based on their initial partial trajectories. They first examined
how they obtained a grouping of trajectories describing user behavior. Next, they
explained how they modeled the main traffic flows using a mixture of 2D Gaussian
distributions. This gave them a density-based grouping of locations, which produced
a grid-based on data from similar points in each pattern. Thus, they presented how
this model can be used to predict the final destination of a new trajectory according to
their first locations by following a two-step procedure: First assign the new trajectory
to the clusters to which they probably belong. Second, they use the characteristics
of the trajectories within these groups to predict the final destination.
Generally speaking, taxi-drivers expect that the top-k valuable points are rapidly
given nearby the time and within a detour of their location. However, in the existing
methods, nearly all the GPS data of a city are computed for recommendations. Due to
too much spatiotemporal noise, most of the existing methods are not accurate enough.
Therefore, we propose a novel method for recommending the most relevant K pick-
up points for a driver based on spatiotemporal features and a deviation threshold for
the driver.

3 Methodology

3.1 DataSet

Yellow Taxi data: these trips are made by New York City’s iconic yellow taxis, that
have been recorded and provided to the Taxi and Limousine Commission (TLC) since
2009; see their homepage [11]. The raw data includes many fields, such as VendorID,
Pick-up-datetime, Drop-off-datetime,etc. More details about the data dictionary and
trip record user guide [12, 13].
Each month contains more than 10 million trips. Due to our constrained com-
putational assets, we only consider taxi trips that originate somewhere within the
28-square-mile rectangular region (see Fig. 1). For further limitation, we just con-
sider taxi rides in the long stretch of January.

3.2 Evaluation

Train and Test

So as to assess the exhibition of our system, we split the informations into a prepa-
ration set and testing set, 80% and 20% respectively.
Recommender System for Most Relevant K Pick-Up Points 281

Fig. 1 We only consider taxi rides in the portion of New York City between latitudes 40.93 to
40.49 and longitudes −74.25 to −73.65

Evaluation Metrics
Evaluating a model is a core part of building an effective machine learning model,
by getting feedback from the metrics, making improvement and continuing until
achieving a desirable accuracy. In our industry, we think about various types of
measurements to assess our models. The decision of metric totally relies upon the
kind of model and the execution plan of the model. So on our project we used:
• Root Mean Squared Error: is the most popular evaluation metric used in regression
problems. It follows an assumption that error are unbiased and follow a normal
distribution.


1  N 
RMSE =  (yi − ŷi )2 = M S E(baseline) (1)
N i=1

• R-Squared: coefficient of determination, so as to assess how well the model, per-

form relatively to the variance of the data set.

M S E(model)
R − Squar ed = 1 − (2)
M S E(baseline)
282 A. Berdeddouch et al.

3.3 Features Extraction

Below are some feature templates we use to extract features from each data point,
along with a rough intuition as to why these features might be predictive.
• Hour of day ∈ [0, 23]. A daily cycle.
• Day of week ∈ [0, 6].
• Deviation threshold, or detour that should a driver take to get to a pickup point.
• [Latitude, Longitude]. the geolocation of pick-up points.
• Distance . The distance of trips, from pick-up to drop-off points.
The heatmap of correlation matrix (see Fig. 2) shows the dependence between fea-
tures, the more they are close to 1, the more dependent they are.

Fig. 2 Heatmap of correlation matrix

Recommender System for Most Relevant K Pick-Up Points 283

3.4 Framework of the Recommender System of Most

Relevant K Pick-Up Points

The ability to use geolocation historical data to recommend pick-up points for a
driver present valuable insights to city planners and taxi dispatchers in answering
questions such as how to position cabs where they are most needed, how many taxis
to dispatch, how ridership varies over time, and how to increase profits and save
energy. The position of passenger concentrated is always changing over time and the
location of the taxi is dynamical. So, our project focuses on recommending the most
relevant k pick-up point for a driver given a specific time window and a location
within the city. The framework of our recommender system is illustrated (see Fig. 3).
It is divided into two parts:
1. Preprocessing the historical taxi trips.
2. Real-time recommendation.
Firstly, we acquire the historical geolocation trips data set [11], passes them to the
preprocessing phase. Moreover, we extract spatiotemporal features such as pick-up
Latitude, Longitude, day of the week, hour, .etc. Additionally, these preprocessed
data are passed as params to the KMeans algorithm to create clusters within the
city, that will be divided into train and test data (80–20%). Then, these clusters are
passed to a pipeline succession of polynomial features and linear regression in order
to calculate the R-Squared and RMSE and to evaluate our model. Also, we predict
pick-up points, which will be stored to be passed into the second phase. The first part
is performed in an offline basis.
Secondly, the real-time recommendation goes as follow: given the time, geolocation
of the driver, K pick-up points, and within a derivation threshold we recommend
these K most relevant pick-up points with a ranking by relevance to the driver. The
experimental results show that our method can effectively recommend these relevant
pick-up points.

Fig. 3 Framework of the recommender system of most relevant K pick-up points

284 A. Berdeddouch et al.

3.5 Recommender System of Most Relevant K Pick-Up Point

Using KMEANS+REGRESSION

After preprocessing the historical trip data and extracting spatiotemporal features.
We pass this data to the KMeans algorithm in order to create clusters within the city.
The main idea of K-Means algorithm is clustering sample data into k groups through
iteration and making criterion function which evaluation clustering performance to
achieve optimal.
When using the selected distance as sample data similarity measurement to divide
categories if the distance between two samples is shorter, then the two-sample dif-
ference is smaller. We employ Euclidean distance as the algorithm’s similarity mea-
surement and use Prediction Error of Square Sum to evaluate the clustering perfor-
mance. Assuming dataset X contains k categories: X 1 , X 2 , . . . , X k , each category
contains the number of samples is N1 , N2 , . . . , Nk , and the centroid of each category
is C1 , C2 , . . . , Ck , the function of Prediction Error of Square Sum as follow


k 
E= | p − mi|2 (3)
i=1 p∈Xi

The pseudocode of K-Means is provided in Algorithm 1.

—————————————————————–
Algorithm 1. K-MEANS ALGORITHM
—————————————————————–
Input: Number of clusters K , number of samples in database n
Output: K clusters
1. assign initial value for means;
2. for j = 1 to n
3. assign each X j to the closest clusters;
4. for i = 1 to K
5. X i = |C1i | x∈Ci x
6. compute E
7. until E stop changing;
8. return E[K ];
9. end for
10. end for
—————————————————————–

4 Experiments and Results

As mentioned in the Sect. 3.1 we used a dataset detailing more than 10 million taxi
trips. The configurations of our computer are as follow:
• I ntel(R)Cor e(T M)i7 − 46000C PU @2.10G H z2.7Ghz
Recommender System for Most Relevant K Pick-Up Points 285

• 8GB memory,
• 250GB SDD storage,
• Windows 10 64-bits operating system.
Google Colab. configuration is as follow:

• CPU: 1xsingle core hyper threaded i.e. (1 core, 2 threads) Xeon Processors
@2.3Ghz (No Turbo Boost), 45MB Cache
• GPU: 1xTesla K80, having 2496 CUDA cores, compute 3.7, 12GB (11.439GB
Usable) GDDR5 VRAM
• 12 GB memory
• 40 GB Storage.
The algorithm is implemented using Jupyter Notebook, Python libraries such as:
Pandas/Dask, Scikit-learn, matplotlib, geocoder, etc.
Determining the optimal number of cluster in a data set is a fundamental issue
in partitioning clustering, sush as K-Means clustering, which requires the user to
specify the number of clusters K to be generated. In order to do that we performed
direct methods, that consists of optimizing a criterion, such as the within cluster sums

Fig. 4 Silouhette method

286 A. Berdeddouch et al.

Fig. 5 Elbow method

Fig. 6 K-Means K = 100

of squares or the average silhouette. The corresponding methods are named Elbow
and silhouette methods, respectively (see Figs. 4, 5, and 6).
Due to the results of the elbow and silouhette methods, respectively we opted for
k = 50, (see Fig. 7).
Recommender System for Most Relevant K Pick-Up Points 287

Fig. 7 K-Means K = 50

Table 1 The results of KMeans+Regression Model

Model R-Squared (R 2 ) Root Mean Squared Error
(RMSE)
KMeans+Regression 0.9862687 0,0177

In the process to obtain the most relevant K pick-up points based on spatiotem-
poral features we utilize the clustering algorithm K-Means combined with Linear
Regression. We obtained the results shown below Table 1.
Our system, KMeans+Regression performed best, achieving an RMSE value of
0.0177. Using greater than 8,000,000 training examples, the test and training
R-squared (R 2 ) scores are practically identical and well above 0.98, indicating that
the system achieves both low bias and low variance.

5 Conclusion and Future Work

Conclusion
To reduce the time to find a passenger and saving fuel comsuption as well as to
increase the profit of taxi drivers, we proposed a recommender system for the most
K relevant pick-up points. We built the recommender system with a dataset containing
more than 1 million historical trip, and evaluated the system by a vast experiments.
As a result, the system accurately predicts the K most relevant pick-up points. In the
future, we plan to deploy our recommender in the real world so as to further validate
and improve the effectiveness and robustness of this system.
288 A. Berdeddouch et al.

Future work
Deep Recommendation
Our system achieved good results. We may be able to achieve better results using a
clustering, neural network, since neural networks can automatically tune and model
feature interactions. Instead of manually determining which features to combine in
order to capture feature interactions, we could let the learning algorithm perform
this task. Also, we would like to add extra features such as Weather, points of inter-
est (POI)- The types of POI include bus stations, commercial buildings, residential
quarters, etc)-, etc.
Recommending Taxis
The taxi recommender aims to provide the taxi drivers with the best parking places
and the routes to these parking places. But how to define a “good” parking place?
Different drivers may have different definitions. However, it’s intuitively obvious
that a good parking place should bring a high probability (during the routes or at the
parking place) to get a passenger, a short waiting time, a short queue length at the
parking place and a long distance of the next trip.
Recommending Passengers
Different from the taxis, the passengers do not want to walk too long for hailing a taxi.
If a passenger is close to at least one parking place, we provide him with the nearby
parking places with the maximum expected queue length. Otherwise, we suggest the
passenger with the nearby road segments (in a walking distance) considering two
criteria:
1. The possibility to find a vacant taxi
2. The average waiting time.

References

1. Zhang, M., Liu, J., et al.: Recommending pick-up points for taxi-drivers based on spatio-
temporal clustering. In: Second International Conference on Cloud and Green Computing, vol.
67–72 (2012)
2. Zheng, Y., Li, Q., Chen, Y., et al.: Understanding mobility based on GPS data. In: Proceedings
of Conference on Ubiquitous Computing, ACM, pp. 312–321 (2008)
3. Zheng, Y., Zhang, L., Xie, X., et al.: Mining interesting locations and travel sequences from
GPS trajectories. In: Proceedings of International Conference on World Wide Web, ACM, pp.
791–800 (2009)
4. Yuan, J., Zheng, Y., Zhang, C., et al.: T-Drive: driving directions based on taxi trajectories. In:
Proceedings of the 18th SIGSPATIAL International Conference on Advances in Geographic
Information Systems, GIS  10, pp. 99–108 (2010)
5. Yuan, J., Zheng, Y., Xie, X., et al.: Driving with Knowledge from the physical world. In:
Proceedings of 17th SIGKDD Conference on Knowledge Discovery and Data Mining, ACM,
pp. 316–324 (2011)
6. Yuan, J., Zheng, Y., Zhang, L., et al.: Where to find my next passenger? In: Proceedings of
the 13th ACM International Conference on Ubiquitous Computing, UbiComp, pp. 109–118
(2011)
Recommender System for Most Relevant K Pick-Up Points 289

7. Ge, Y., Liu, C., Xiong, H., et al.: A taxi business intelligence system. In: Proceedings of the
17th ACM SIGKDD international conference on Knowledge discovery and data mining, KDD
11, pp. 735–738 (2011)
8. Ge, Y., Xiong, H., Liu, C., et al.: A taxi driving fraud detection system. In: Proceedings of the
ICDM 2011 IEEE International Conference on Data Mining, IEEE, pp. 181–190 (2011)
9. Ge, Y., Xiong, H., Tuzhilin, A., et al.: An energy-efficient mobile recommender system. In:
Proceedings of the 16th ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, ACM, pp. 899–908 (2010)
10. Tang, L., Chang, X., Li. Q.: The knowledge modeling and route planning based on taxi’s
experience. Acta Geodartica et Cartographica Sinica, 404–409 (2010)
11. TLC Trip Records Homepage: https://www1.nyc.gov/site/tlc/about/tlc-trip-record-data.page.
Accessed 12 Jan 2020
12. Data Dictionary: https://www1.nyc.gov/assets/tlc/downloads/pdf/data_dictionary_trip_
records_yellow.pdf. Accessed 12 Jan 2020
13. Trip Record User Guide: https://www1.nyc.gov/assets/tlc/downloads/pdf/trip_record_user_
guide.pdf. Accessed 12 Jan 2020
Feature Detection and Tracking
for Visual Effects: Augmented Reality
and Video Stabilization

Houssam Halmaoui and Abdelkrim Haqiq

Abstract In this article, we propose two visual effects applications of feature track-
ing: augmented reality and video stabilization. The originality of the proposed meth-
ods lies in the combination and the adaptation of computer vision algorithms to
visual effects applications in order to improve the process of visual effects in film-
making. The first step involves feature detection and tracking. Different multi-scale
feature detection algorithms have been tested: Harris-Laplace, Hessian-Laplace and
Laplacian of Gaussian. An experiment was conducted to evaluate the tracking quan-
titatively. Then, we evaluate tracking for visual effects applications. For augmented
reality, the result of tracking combined with perspective transformation are used to
add an image to a planar surface of a scene filmed by a moving camera. We also
evaluate the tracking for 3D objects insertion or matchmoving. For video stabiliza-
tion, the result of tracking is used to estimate the geometrical transformation between
frames.

Keywords Multiscale feature · Homography · Visual effects · Augmented

reality · Video stabilization

1 Introduction

With the development of digital cinema, computer vision and image processing
become omnipresent and indispensable in the entire movie production process
[2, 6, 22] from video acquisition to projection and in every post-production step in

H. Halmaoui (B)
ISMAC - Higher Institute of Audiovisual and Film Professions, Avenue Allal Al Fassi,
Rabat, Morocco
e-mail: [email protected]
H. Halmaoui · A. Haqiq
IR2M Laboratory, Faculty of Sciences and Techniques, Hassan 1st University,
Settat, Morocco
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 291

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_21
292 H. Halmaoui and A. Haqiq

Fig. 1 Proposed method steps

Fig. 2 Example of corner detection and tracking

between: visual effects or VFX, Computer Graphical Imagery or CGI, color grading,
and editing. The classical approaches of image processing, such as segmentation,
stereo matching, optical flow, tracking, alpha matting, compression or inpainting,
were adapted by film industry VFX software providers for artistic usage. On the
one hand, the operating principle of those methods is not known to VFX artists who
spend more time on the creative process and learning new software features. On the
other hand, in the image processing research community, there are far fewer VFX
applications than robotics, medical, video surveillance or road safety, and that is
because the needs of VFX artists are little-known by researchers. The originality of
the proposed methods lies in the combination and the adaptation of computer vision
algorithms to visual effects applications in order to improve the process of visual
effects in filmmaking. In this article, we are interested in the use of feature tracking
in the VFX realm. The steps of the proposed method are illustrated in Fig. 1. We
propose two VFX applications: tracking of a planar surface of the scene in order to
insert an image in it(augmented reality) and stabilization of a video sequence. The
challenge of this study is to propose reliable algorithms that require minimum user
interaction to save the user’s time. We will also address the limitations of the methods,
such as noise and camera speed, of which the user must be aware. Tracking consists
of matching similar regions of consecutive images of a video acquired by a moving
camera. There are two types of tracking: object and feature tracking. Object tracking
consists of detecting and tracking known objects such as people or cars. Feature
tracking (see Fig. 2), which interests us, consists of detecting and tracking patterns
such as corners, lines, or similar connected regions (blobs). Tracking applications
are numerous: stereo matching, image registration, matchmoving, object recognition
and augmented reality.
In the following section, we present the first stage of tracking which consists of
detecting the more reliable regions for matching.
Feature Detection and Tracking for Visual Effects … 293

2 Feature Detection

2.1 Harris Corner

We call feature a rectangle W in the image that is reliable for tracking. The detection
aim is to extract the best features such that each one has a unique match in the other
images. The problem of detection can be considered as a segmentation of the image
in three regions: uniform, line and corner. Figure 2 shows that the features tracked are
not found on line or in a uniform region. Indeed, lines and uniform regions are not
reliable for tracking because in their neighborhoods there are several similar regions.
Corners are locally unique, and therefore they are good candidates for tracking.
Moravec’s algorithm [20] uses a sliding window to calculate the intensity variation
of a block W in several directions. A uniform region corresponds to a quasi-constant
luminance variation. For a line, the variation is very weak. For an edge, the variation
is significant but only in the direction perpendicular to the gradient. For a corner, the
variation is significant in all directions. Therefore, a feature is detected if the variation
of the corresponding block is above a threshold in all directions. The mathematical
formulation of the problem is:

E(u, v) = w(x, y)(I (x + u, y + v) − I (x, y))2 (1)
x,y∈W

The E(u, v) function measures the difference between the pixel intensities in a win-
dow W and W shifted by a vector (u, v). The binary function w(x, y) takes the
value of 1 inside the window and 0 otherwise. The vector (u, v) takes the values
[(1, 1); (1, 0); (0, 1); (−1, 1)]. The detection is performed by calculating the min-
imum value of E(u, v) between the four directions, we retain only local maxima,
and finally a threshold is applied to reduce the noise in order to detect only reliable
features. The disadvantage of this detector is that it is anisotropic because we only
consider four directions to measure the similarity. In the Harris algorithm [13], the
problem is reformulated to take into account all directions. After a Taylor develop-
ment of Eq. 1 around (u, v) = (0, 0) and matrix writing we get:
⎡ ⎤
 ∂ I (x, y) 2  ∂ I (x, y) ∂ I (x, y)
 T ⎢ w(x, y) w(x, y) ⎥ 
u ⎢ x,y ∂x x,y ∂x ∂y ⎥ u
E(u, v) = ⎢ ⎥
v ⎣ ∂ I (x, y) ∂ I (x, y)  ∂ I (x, y) 2 ⎦ v
w(x, y) w(x, y)
x,y ∂x ∂y x,y ∂y
(2)

The matrix in the term on the right, which we denote H , is called the Harris matrix.
Let (α, β) be the eigenvalues of H and (e1 , e2 ) the corresponding eigenvectors.
The eigenvalues are proportional to the principal curvatures of the auto-correlation
function plotted on Fig. 3. The curves show the classification of the different regions
of the image according to the eigenvalues:
294 H. Halmaoui and A. Haqiq

Fig. 3 region classification

according to eigenvalues
(α, β) of the Harris matrix H

– If the eigenvalues are small, the intensity of the window is uniform. Indeed, we
have ∂∂ xI = ∂∂ yI = 0, hence α = β = 0.
– If β is weak and α important, the region corresponds to an edge. The gradient
( ∂∂ xI , ∂∂ yI ) is perpendicular to the contour, e1 normal to the direction of the contour,
and the value of α not negligible. While e2 is parallel to the direction of the contour
and the value of β is weak.
– If both eigenvalues are important the window contains a corner or a blob.
To avoid calculating the eigenvalues, we use a quality measure C (shown on the
right axis of Fig. 3) called the response function:

C = det (H ) − k.trace(H )2 (3)

When the value of C is important, the region corresponds to a corner or a blob. Then,
the detection is performed by thresholding C. The value of the threshold depends on
the application. The parameter k controls the sensitivity of the detector: when k is
low, more features are detected but the reliability of the detector decreases because
of low contrast regions. We generally take k = 0.04. The C value of a feature will
be significant for several blocks in its neighborhood. As a result, several features
are detected around the same region (see Fig. 4). We retain only the block with the
highest measure of quality by calculating the local maxima of C in a neighborhood
of size N × N . Indeed, the further one moves away from the center of the feature,
the more the measure of quality will decrease. In practice, one seeks to obtain a large
number of detections, and for that, we can reduce the value of N .
The detector is covariant by rotation since the detection criterion depends solely
on the eigenvalues and not on the eigenvectors. We talk about covariance and not
invariance since the position of the feature after the rotation will be approximately
the same.
Feature Detection and Tracking for Visual Effects … 295

Fig. 4 Detection of features before and after non local maxima suppression of C

For implementation, the gradient is approximated by:

∂ I (x, y)
= I (x + 1, y) − I (x − 1, y)
∂x
In order to reduce the high-frequency noise, the gradient is estimated by the
convolution of the image with the derivative of a Gaussian (we will discuss the
choice of the Gaussian function in the following section):

∂ I (x, y) ∂G(x, y, σ D )
= I (x, y) ∗
∂x ∂x
where
1 1
G(x, y, σ ) = exp − 2 (x 2 + y 2 )
2π σ 2 2σ

The binary function w in the Harris matrix can be replaced by a symmetrical radial
function to give more weight to the pixels closest to the center (x0 , y0 ) of the feature:

1 1
w(x, y) = G(x − x0 , y − y0 , σ I ) = exp − 2 ((x − x0 )2 + (y − y0)2 )
2π σ I
2
2σ I

where σ I = aσ D and a ∈ [1; 2].

Figure 5 shows some results obtained by our implementation. The disadvantage
of the Harris detector is the fixed size of the W window, which limits the detection
to corners of the same size. We will see in the next section how to modify the
Harris algorithm by using the scale-space theory in order to allow multi-scale feature
detection.
296 H. Halmaoui and A. Haqiq

Fig. 5 Comparison of feature detectors. From left to right: Harris, Harris Laplace, Hessian Laplace
and LoG

2.2 Harris Laplace

The most straightforward method to perform multi-scale feature detection is to sub-

sample the image. The set of reduced images is called pyramid resolution (see Fig. 6).
We perform the detection of the features on each reduced image using the same
window W . This amounts to the same thing as detecting larger features when the
resolution decreases. The problem with this method is that resampling introduces
structures not present in the original image. However, this method is useful in some
cases to speed up processing. An alternative [17] to avoid introducing artifacts is
to convolve the image by a Gaussian in order to reduce the intrinsic resolution of
the image, which means that the number of image pixels does not change, but the
smallest details become unrecognizable (see Fig. 6). Scale-space is the set of images
built by this method. A variation of this method is to use diffusion [21] to reduce
resolution without blurring the edges of the image.

Fig. 6 Pyramid resolution set (first line) and scale-space set (second line)
Feature Detection and Tracking for Visual Effects … 297

We start by building the scale-space set by convolving the image by multiple

Gaussians:
L(x, y, σ D ) = G(x, y, σ D ) ∗ I (x, y)

with σ D taking increasing values in an interval usually formed by a geometric

sequence {σ0 , σ0 .b, σ0 .b2 , . . .}. The scale decreases when σ D increases. We usually
take σ0 = 1.5 and b ∈ [1.2; 1.4]. The new Harris Matrix can be written as:
⎡ ⎤
∂ L(x, y, σ D ) 2 ∂ L(x, y, σ D ) ∂ L(x, y, σ D )
⎢ ∂x ∂x ∂y ⎥
H (x, y, σ D , σ I ) = G(x, y, σ I ) ∗ ⎢
⎣ ∂ L(x, y, σ ) ∂ L(x, y, σ ) 2
⎥
⎦
D D ∂ L(x, y, σ D )
∂x ∂y ∂y
(4)
where ∂ L(x,y,σ ∂x
D)
= ∂G(x,y,σ D )
∂x
∗ I (x, y).
The detection is performed on each scale as previously. However, before per-
forming detection, we need to normalize the matrix H in order to have comparable
matrices between the different scales. Indeed, for tracking, we need to compare fea-
ture detected at different scales because of the objects size variation in the video.
We consider H and H  the Harris matrices, respectively, for high and low resolution
images. If a window in low resolution image is centered on (x  , y  ) on the scale
(σ D , σ I ), the same window in high resolution is centered on (x, y) at scale (σ D , σ I ).
It can be shown [10] that:
1     
H (x, y, σ D , σ I ) = H (x , y , σ D , σ I )
b2
Therefore, in order to normalize the Harris matrix, we multiply it by the compen-
sation term b2 = (σ D /σ D )2 .
A single corner or blob can be detected at different scales. We must choose the
scale that best represents the structure of the feature. For this we retain only local
maxima on the scale dimension of the normalized Laplacian:
 
 ∂ 2 G(x, y, σ D ) ∂ 2 G(x, y, σ D ) 
N L(x, y, σ D ) = σ I2 + ∗ I (x, y) (5)
∂x2 ∂ y2

Figure 5 shows some results obtained by our implementation. We notice that the
features are detected at natural scales and that similar features are detected at the
same scale (scale covariance).
Note also that it is possible to make the detector invariant to affine transformations
[19]. However, this is not interesting in the case of our video tracking study since
the geometric variations between two consecutive images are negligible.
298 H. Halmaoui and A. Haqiq

2.3 Hessian Laplace and LoG

For better detection of blobs, the Harris matrix can be replaced by the scale normal-
ized Hessian matrix:
⎡ 2 ⎤
∂ L(x, y, σ D ) ∂ 2 L(x, y, σ D )
⎢ ∂x2 ∂ x∂ y ⎥
S(x, y, σ D ) = σ D2 ⎢ ⎥
⎣ ∂ 2 L(x, y, σ D ) ∂ 2 L(x, y, σ D ) ⎦ (6)
∂ x∂ y ∂ y2

In the Hessian Laplace method, instead of using the quality measure C, we use the
determinant of S, which corresponds to the Hessian Laplace. For scale selection, we
use the trace of S which corresponds to Laplacian of Gaussians LoG. This method
allows for better rejection of the edges.
In the LoG method, we use the trace of S for both detection and scale selection.
It allows better detection of blobs but also responds to edges.
Figure 5 shows some examples of detection obtained by our implementation.
The LoG can be approximated by the difference of Gaussians DoG to speed
up processing as in SIFT [18] without big loss of accuracy. Note that there are
several fast detectors such as FAST [24], BRIEF [7], ORB [25] and BRISK [16].
The disadvantage of fast detectors is that they are less accurate.

3 Features Descriptor

Once we have the position and scale of the feature, we calculate a descriptor D.
Let f and f  be two features in two different images such that f  = T ( f ) with T
a geometric and photometric transformation (change of point of view, scale, rota-
tion or luminance variation). A good descriptor D must verify D( f ) = D( f  ). The
most simple descriptor is a vector formed by the set of pixel intensity values of the
detected feature. To perform tracking, we compare the intensities of the features.
Many tracking algorithms use this method because of its speed and satisfactory
results when geometric and photometric variations are low, which is often the case
in video sequences. In general, the variations can be significant when the camera
moves fast, and we must use an invariant descriptor.
We start by choosing the relevant pixels that will contribute to the descriptor. We
consider a support region larger than the size of the detected feature, having a size
equal to 3 or 6 times the size of the feature (see Fig. 7).
Then, in order to be able to compare the descriptors of two features with different
orientations, we detect in each support region the orientation of the dominant gradient,
and we change the orientation so that the dominant gradient is horizontal (see Fig. 7).
Note that the same principle is used by SIFT algorithm. To detect the orientation of
Feature Detection and Tracking for Visual Effects … 299

Fig. 7 From left to right: Feature Detected, Support Region, Histogram Orientation, Detection of
the dominant gradient, rotation of the Support Region

the dominant gradient, we start by calculating the magnitude M and the orientation
θ of the gradient for each pixel in the support region D:


M(x, y) = (D(x + 1, y) − D(x − 1, y))2 + (D(x, y + 1) − D(x, y − 1))2

D(x, y + 1) − D(x, y − 1)
θ = arctan
D(x + 1, y) − D(x − 1, y)

The histogram of the orientations is calculated on a scale of 36 bins separated by a

step of 10. Each pixel is incremented by the magnitude M weighted with a Gaussian
M(x, y) · G(x, y, σ ) to give more weight to the pixels of the center. With σ equal
to 1.5 times the scale of the feature.
Orientation is determined by detecting the maximum peak of the histogram, as
shown in Fig. 7. If the histogram contains multiple peaks, one can generate several
features at the same location and retains only the best candidate during matching.
After reorienting the patch, we re-sample the window in size 41 × 41 or 81 × 81
to have descriptors of the same size regardless of the size of the features. Note that
now we have a feature invariant to scale, translation, and orientation. Finally, in
order to compensate for luminance variations, the intensity of the final patch D f is
normalized with its mean μ and its variance s:

(D f − μ)
Df  =
s

4 Tracking

To perform tracking, we compare the ai values of the descriptor A of each feature, and
for each scale, in the first image, to the bi values of the descriptor B of each feature,
and for each scale, in the next image. In order for the matching to be exhaustive, a
candidate feature must be retained only once all the features at all scales are compared
300 H. Halmaoui and A. Haqiq

Fig. 8 Feature tracking between the first (first line) and the last images (second line) of a video
sequence of length 18 s by using Harris Laplace (left), Hessian Laplace (middle) and LoG detectors
(right)

to the reference feature. For comparison, one uses the SSD (Sum Squared Difference)
measurement and retains the candidate feature which minimizes this distance:

ssd(A, B) = (ai − bi )2
i

Other measures such as Euclidean distance or correlation can be used.

We eliminate candidates whose SSD exceeds a threshold that generally varies
between 0.1 and 0.8 depending on the application. We limit the research zone in the
second image to the neighborhood of the reference feature in the first image. The size
of the neighborhood depends on the camera displacement speed. Generally, between
two consecutive images, a feature moves from 1 to 20 pixels. Figure 8 shows some
results obtained by our implementation. Note that only the features visible throughout
the entire sequence are tracked and displayed. Figure 8 shows the features detected in
the first image and the features remaining at the end of the sequence, that is 18 s long.
The longer the sequence and the more significant the movement of the camera, the
fewer features left at the end. We also see that with the Hessian and LoG detectors,
we can track more features because of their better response to blobs.

5 Quantitative Evaluation

We carried out an experimental evaluation to measure the stability of the tracking.

We used a tripod-mounted mobile phone to capture a video of a fixed scene (without
any camera movement), in order to have ground truth. Given that the camera and
the scene are static, a feature should ideally be detected in the same position. We
evaluated three kinds of sequences: no change of lighting, gradual change of lighting,
Feature Detection and Tracking for Visual Effects … 301

Fig. 9 Sample images from the experiment: result of the tracking on the first and last images of
the sequence

Fig. 10 The score of repeatability of the sequence with no change of lighting, using a features
position tolerance of 0 pixels (left) and 2 pixels (right)

and instantaneous change of lighting. To evaluate the tracking algorithm, we compute

a repeatability score [15, 26] between each frame of the sequence and the first frame
of the sequence considered as reference: S = NRf with R the number of repeatable
detections, which corresponds to the number of features whose position has not
changed compared to the reference image. We can introduce a tolerance interval of
the displacement of features. N f is the number of features visible throughout the
entire sequence.
Figure 9 shows some images from a sequence used for the experiment.
For the first experiment of a fixed scene with no change of lighting, the only factor
influencing the position of features between two consecutive images is the camera
sensor noise. The plots in Fig. 10 compares the scores of repeatability of the three
detectors, using a tolerance of 0 and 2 pixels.
Figures 11 and 12 show the results of the experiments with a gradual change of
lighting and with an instantaneous change of lighting.
Hessian-Laplace and LoG detectors show quasi-equivalent scores that outperform
Harris-Laplace detector (similar results were found in [4]). In the first experiment,
the lighting is constant throughout the entire sequence. The score stabilizes after a
certain time (see explanation below). When setting a zero tolerance threshold (we
only consider, for the calculation of the score, the features whose position has not
changed), we get a score of up to about 0.7. The score improves as the tolerance
threshold increases. A score of up to about 0.9 is obtained for a tolerance of ±2
302 H. Halmaoui and A. Haqiq

Fig. 11 The score of repeatability of the sequence with a gradual change of lighting, using a features
position tolerance of 0 pixels (left) and 2 pixels (right)

Fig. 12 From left to right: the score of repeatability of the sequence with instantaneous change of
lighting, using a features position tolerance of 0 (left) and 2 pixels (right)

pixels. In the second experiment, the lighting varies inhomogeneously in the scene.
The plot of the score shows that the detection score decreases over time due to the
change in lighting. Finally, in the last experiment, the lighting varies instantaneously.
The evaluation results show that the score decreases instantaneously at the moment
of the change of lighting.
We also note, in all evaluation curves, a drop in the score at the beginning of the
sequence. This is due to the detector’s sensitivity to low contrast features. Indeed,
some features with low contrast will appear in the first images of the sequence but
not in the next ones. It is possible to reduce the value of this drop in the score by
adjusting the sensitivity parameters of the detector, we will get a better score but also
fewer features, which is not relevant for most tracking applications.
The evaluation results give us an idea of the tracking performance. In the following,
we will evaluate the tracking by combining it with VFX algorithms.

6 VFX Applications

In this section, we propose to use the result of detection and tracking in a video
acquired with a camera in motion for VFX applications (see Fig. 1). An advantage
of the proposed method is its modular aspect. Each step of detection, extraction of
Feature Detection and Tracking for Visual Effects … 303

the descriptors, and tracking, can be replaced by a different algorithm in order to

propose the most appropriate solution for each vfx application.

6.1 Augmented Reality

Augmented reality consists of adding or enhancing elements in a video of a real

scene. Here, we propose to insert a 2D image in a planar surface selected by the
user. The appearance of the inserted image must match the perspectives of the scene
while the camera moves. For this, the result of position of the features belonging to
the planar surface is used to calculate the perspective transformation (homography)
between every two consecutive images of the video.
We start by detecting the features using the multi-scale Hessian Laplace detector
which, as we mentioned previously, gives the best compromise between the detection
of blobs and the rejection of the contours, in addition to its invariance by translation
and scale. The descriptors are extracted as explained previously in order to obtain
features invariant for rotation and luminance change. Then we track features between
consecutive images by matching their descriptors using an SSD measure.
The results of the tracking of the first and last images of the video are displayed to
the user to verify that there are no matching errors, often due to significant geometric
and photometric variations. Tracking errors can be considered outliers and detected
by an algorithm such as Ransac [11] which is not relevant in our case because
we need to have a data model, whereas, we consider random and unknown camera
movements. For a better matching result, VFX scenes are often filmed in a controlled
environment by adding markers. For our study, we also add markers to the planar
surface. Note that there are non-generic detectors [5, 12, 23, 27] that only work with
artificial markers.
As illustrated in the example in Fig. 13 of a video sequence of length 18 s, the
user draws on the first image a rectangle on the planar surface, where he wishes to
insert the image, and then selects the features belonging to the flat surface and which
have been well tracked (without matching errors).
The next step is the homography estimation between consecutive images.

6.1.1 Homography

We consider xi the position of a feature i belonging to the planar surface and xi  the
position of the same feature in the next image. The perspective transformation from
xi to xi corresponds to the 3 × 3 matrix H such that:

xi = H · xi
304 H. Halmaoui and A. Haqiq

Fig. 13 The planar surface selected by the user (left) and the corresponding perspective transfor-
mation (right) after the estimation and application of homography using the result of tracking

where ⎡ ⎤
h1 h2 h3
H = ⎣h 4 h 5 h 6 ⎦
h7 h8 h9
 T  T
and xi = xi yi wi the homogeneous coordinates. For wi = 1, xi yi corre-
sponds to image coordinates. Therefore, the equation is defined up to a factor. That
is, by multiplying H by a factor, the equality remains valid. This reduces the degree
of freedom of H from 9 to 8. We estimate H with Direct Linear Transformation
algorithm [14]. Note that there are methods for estimating homography using deep
learning [9].
We begin by writing the homography equation as a scalar product:

xi × Hxxi = 0
 T  T
We note the lines of H : h 1 = h 1 h 2 h 3 , h 2 = h 4 h 5 h 6 , and h 3 =
 T
h 7 h 8 h 9 . The previous equation becomes:
⎡  3T ⎤
yi h xi − wi h 2T xi
x i × Hxxi = ⎣wi h 1T xi − xih 3T xi ⎦
xih 2T xi − yih 1T xi

As h j T xi = xi T h j , we obtain:
⎡ T ⎤ ⎡ 1⎤
0 −wi xi T yi xi T h
⎣ wi xi T 0 T −xi xi T ⎦ ⎣h 2 ⎦ = 0
−yi xi T xi xi T 0T h3

We note this equation in the form:

Ai h = 0
Feature Detection and Tracking for Visual Effects … 305

where ⎡ 1⎤
h
h= ⎣h 2⎦

h3

and ⎡ ⎤
0 T −wi xi T yi xi T
Ai = ⎣ wi xi T 0 T −xi xi T ⎦
 T
−yi xi xi xi T 0T

It can be seen that the third line of Ai can be obtained, up to a scale, from the sum
of the first line multiplied by xi and the second line multiplied by yi . The matrix Ai
is therefore of rank 2. The equation is thus equivalent to:
⎡ ⎤
 T  h1
0 −wi xi T yi xi T ⎣ 2 ⎦
h =0
wi xi T 0 T −xi xi T
h3

Hence, we consider the new 2 × 9 matrix Ai :

 T 
0 −wi xi T −yi xi T
Ai =
wi xi T 0 T −xi xi T

Each pair of feature xi and xi gives us two equations. Therefore, to solve the
homography problem, we use four pairs of features that provide us with a set of
equations: Ahh = 0. The matrix A is constructed from the Ai matrix lines of each
pair of features.
Since the position of the features is subject to noise, we use more than four features,
and we search for an approximate solution corresponding to the best projective
transformation H , which minimizes a cost function. The function to be minimized
is:

min = ||Ahh ||
h ,||hh ||=1

We use the constraint ||hh || = 1 in order to avoid the trivial solution h = 0.

The solution is obtained using the least squares method by calculating the SVD
of A = U DV T and deducting h as the last column of V .
The image coordinates (xi , yi ) have values that range from 1 to hundreds or
thousands, while wi = 1. Therefore it is necessary to normalize the values of A to
have similar orders of magnitude [8] before applying the DLT algorithm. We start
by calculating for each set of feature xi (and xi ) the transformation T (and T  ),
consisting of translating and scaling xi into xi∗ such that the
√ centroid of the set of
features xi∗ is (0, 0) and the mean distance to the origin is 2:
306 H. Halmaoui and A. Haqiq
⎡ ⎤⎡ ⎤
s −1 0 0 1 0 −m 1
T = ⎣ 0 s −1 0⎦ ⎣0 1 −m 2 ⎦
0 0 1 00 1

where m 1 and m 2 the centroid coordinates of xi :

1 1
n n
m1 = xi and m 2 = yi
n i=1 n i=1

and  0.5
1 
s= (xi − m 1 )2 + (yi − m 2 )2
2n i

The transformation T  from xi to xi∗ is calculated in the same way. Then, we
apply the previous DLT algorithm to find H ∗ . Finally, we de-normalize H ∗ to find
H:

H = T −1 H ∗ T

In our application, we estimate H for every two consecutive images in the video.
Then we apply these homographies to the insertion mask chosen by the user in order
to track the mask throughout the sequence.
Finally, we calculate the homography between the rectangular image to be inserted
and the tracked mask and apply it in order to warp the rectangular image to the mask
position. Figure 14 shows an example of results.
As we pointed out in the tracking section, the longer the video duration is and
the larger the camera movement, the fewer the tracked features will be. This also
affects the stability of the tracking: the position of the tracked features will be less
precise because of the accumulation of errors. Our challenge was to succeed to track
features on a longer duration. The parameters allowing to have more features are the
detection threshold, the threshold SSD of the matching and the area of research. The
values of these parameters depend on each video. Figure 15 shows an example of a
result on a relatively long video with a duration of 18 s with an important change of
perspective. We notice that small errors in position can be introduced. Note also that
in order to obtain a more stable homography result, we used more markers: more
than the minimal number of four points needed for homography. Another example
is shown on Fig. 16.

6.2 Matchmoving

Matchmoving consists of inserting 3D objects into a real-life scene video. For that,
we must determine for each image of the video, the position and the orientation of the
Feature Detection and Tracking for Visual Effects … 307

Fig. 14 Images from the original video (first and third lines) and the corresponding result of image
insertion (second and fourth lines). The lengths of the video are respectively 20 and 13 s

Fig. 15 Example of image insertion on a video of 18 s length with important perspective change

Fig. 16 Example of image insertion on a video of 20 s length

308 H. Halmaoui and A. Haqiq

Fig. 17 Matchmoving: 3D objects insertion on video by using the result of tracking

camera in a referential of the scene. The first stage of matchmoving is the tracking
of features. We used the result of the tracking to supply a matchmoving software
allowing the estimation of the movement of the camera and the insertion of the 3D
objects. Figure 17 shows an example of result. We see that the perspective of 3D
objects corresponds to the movement of the camera. Further study of matchmoving
will be the subject of future research.

6.3 Video Stalbilization

Here we propose another application of feature tracking which consists of stabilizing

video acquired by a hand-held camera to correct image vibration or to add some
artistic effects to the video. We suppose that between two successive images of the
sequence, the image is rotated, scaled, and translated. The geometric transformation
[3] is written as:

xi+1 = xi · s · cos(θ ) − yi · s · sin(θ ) + Tx

yi+1 = xi · s · sin(θ ) + yi · s · cos(θ ) + Ty

with (xi , yi ) the feature position in image i and (xi+1 , yi+1 ) that of its correspondent
in the next image. s, θ and (Tx , Ty ) correspond, respectively, to the coefficients of
scaling, rotation and translation. The set of tracked features are used to estimate the
different parameters of the transformation with the linear least square method.
Finally, in order to stabilize the video, we apply to each image the inverse trans-
formation: scale 1/s, rotation −θ and translation (−Tx , −Ty ).
Figure 18 shows an example of the results obtained. It can be seen that the stabi-
lization algorithm compensates for the movement of the camera.
Feature Detection and Tracking for Visual Effects … 309

Fig. 18 Images from a video before (first line) and after stabilization (second line)

6.4 Note on the Implementation

All the algorithms have been implemented from scratch using Matlab, except for the
algorithm of homography [1]. All videos were acquired with a mobile phone. The
video resolution is equal to 640 × 320. Empirically, we find that the parameters that
give best result for almost all video, and which we use to obtain the results presented
here, is: σ0 = 1.5; b = 1.4; k = 0.04; a = 1.3; Harris threshold of 10−9 ; Hessian
Laplace threshold of 10−3 ; LoG threshold of 10−1 ; Five scales: [σ0 · b, σ0 · b2 , σ0 ·
b3 , σ0 · b4 , σ0 · b5 ]; Neighborhood size for matching of 20 pixels; Descriptors size of
81 × 81; SSD threshold of 2; Support region size of 6 × W ; The Image input values
are normalized between 0 and 1.

7 Conclusion and Perspectives

The objective of our study was to develop visual effects approaches applicable to
video acquired with a mobile camera. The qualitative and quantitative evaluation on
videos of real scenes attests to the effectiveness of the proposed algorithms. However,
the method has limitations such as noise matching errors. User intervention can allow
the elimination of bad matching results. The user interaction can be avoided by using
an outliers elimination algorithm such as Ransac, except that it will be necessary
to adjust the parameters of this algorithm manually according to each sequence.
Feature tracking can be improved by using a Kalman filter to estimate the position
of features not visible throughout the entire sequence and which in our case have
been discarded. One of the limitations of our AR algorithm is that it does not handle
the case when foreground objects occlude the insertion plane. We also propose to
310 H. Halmaoui and A. Haqiq

investigate dedicated VFX markers features detectors. Finally, note that the modular
aspect of our method makes it more flexible. We propose as a perspective to evaluate
the methods proposed on a video database, which will also include moving objects, by
testing and comparing other algorithms in each module, and to consider the real-time
aspect and other VFX applications.

References

1. https://www.robots.ox.ac.uk/~vgg/hzbook/code/
2. Arnold, J., Frater, M., Pickering, M.: Digital Television: Technology and Standards. Wiley
(2007)
3. Battiato, S., Gallo, G., Puglisi, G., Scellato, S.: Sift features tracking for video stabilization. In:
14th International Conference on Image Analysis and Processing (ICIAP 2007), pp. 825–830.
IEEE (2007)
4. Bäuml, M., Stiefelhagen, R.: Evaluation of local features for person re-identification in image
sequences. In: 2011 8th IEEE International Conference on Advanced Video and Signal Based
Surveillance (AVSS), pp. 291–296. IEEE (2011)
5. Bergamasco, F., Albarelli, A., Cosmo, L., Rodola, E., Torsello, A.: An accurate and robust
artificial marker based on cyclic codes. IEEE Trans. Pattern Anal. Mach. Intell. 38(12), 2359–
2373 (2016)
6. Bertalmío, M.: Image Processing for Cinema. CRC Press (2014)
7. Calonder, M., Lepetit, V., Strecha, C., Fua, P.: Brief: binary robust independent elementary
features. In: European Conference on Computer Vision, pp. 778–792. Springer (2010)
8. Chojnacki, W., Brooks, M.J.: Revisiting Hartley’s normalized eight-point algorithm. IEEE
Trans. Pattern Anal. Mach. Intell. 25(9), 1172–1177 (2003)
9. DeTone, D., Malisiewicz, T., Rabinovich, A.: Deep image homography estimation.
arXiv:1606.03798 (2016)
10. Dufournaud, Y., Schmid, C., Horaud, R.: Matching images with different resolutions. In: Pro-
ceedings IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2000 (Cat.
No. PR00662), vol. 1, pp. 612–618. IEEE (2000)
11. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with
applications to image analysis and automated cartography. Commun. ACM 24(6), 381–395
(1981)
12. Garrido-Jurado, S., Muñoz-Salinas, R., Madrid-Cuevas, F.J., Marín-Jiménez, M.J.: Automatic
generation and detection of highly reliable fiducial markers under occlusion. Pattern Recogn.
47(6), 2280–2292 (2014)
13. Harris, C.G., Stephens, M., et al.: A combined corner and edge detector. Citeseer (1988)
14. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cambridge Univer-
sity Press (2003)
15. Hassaballah, M., Abdelmgeid, A.A., Alshazly, H.A.: Image features detection, description and
matching. In: Image Feature Detectors and Descriptors, pp. 11–45. Springer (2016)
16. Leutenegger, S., Chli, M., Siegwart, R.: Brisk: binary robust invariant scalable keypoints. In:
2011 IEEE International Conference on Computer Vision (ICCV), pp. 2548–2555. IEEE (2011)
17. Lindeberg, T.: Feature detection with automatic scale selection. Int. J. Comput. Vision 30(2),
79–116 (1998)
18. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vision
60(2), 91–110 (2004)
19. Mikolajczyk, K., Schmid, C.: Scale and affine invariant interest point detectors. Int. J. Comput.
Vision 60(1), 63–86 (2004)
Feature Detection and Tracking for Visual Effects … 311

20. Moravec, H.P.: Obstacle avoidance and navigation in the real world by a seeing robot rover.
Technical report, Stanford Univ CA, Dept of Computer Science (1980)
21. Perona, P., Malik, J.: Scale-space and edge detection using anisotropic diffusion. IEEE Trans.
Pattern Anal. Mach. Intell. 12(7), 629–639 (1990)
22. Radke, R.J.: Computer Vision for Visual Effects. Cambridge University Press (2013)
23. Romero-Ramirez, F.J., Muñoz-Salinas, R., Medina-Carnicer, R.: Speeded up detection of
squared fiducial markers. Image Vis. Comput. 76, 38–47 (2018)
24. Rosten, E., Drummond, T.: Machine learning for high-speed corner detection. In: European
Conference on Computer Vision, pp. 430–443. Springer (2006)
25. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.R.: Orb: an efficient alternative to sift or surf.
In: ICCV, vol. 11, p. 2. Citeseer (2011)
26. Tissainayagam, P., Suter, D.: Assessing the performance of corner detectors for point feature
tracking applications. Image Vis. Comput. 22(8), 663–679 (2004)
27. Wang, J., Olson, E.: Apriltag 2: efficient and robust fiducial detection. In: 2016 IEEE/RSJ
International Conference on Intelligent Robots and Systems (IROS), pp. 4193–4198. IEEE
(2016)
Spectral Image Recognition Using
Artificial Dynamic Neural Network
in Information Resonance Mode

Ivan Peleshchak, Roman Peleshchak, Vasyl Lytvyn, Jan Kopka,

Mariusz Wrzesien, Janusz Korniak, Janusz Kolbusz, and Pawel Rozycki

Abstract This paper shows a new type of artificial neural network with dynamic
(oscillatory) neurons that have natural frequencies. Artificial neural network in the
mode of information resonance implements a new method of recognition of multi-
spectral images. A computer experiment was performed to recognize multispectral
images by a dynamic neural network based on the resonance effect.

Keywords Neural network · Multispectral images · Resonance

1 Introduction

Currently, intensive research is underway on the use of neural networks to solve a wide
class of Data Mining problems (identification of non-stationary chaotic processes,
clustering, intelligent control, diagnostics of biosystems, forecasting, emulation
and recognition of multispectral images). In particular, hybrid neuro-fuzzy systems
[1], adaptive neuro-fuzzy systems [2], wavelet-neuro-fuzzy networks [3] and adap-
tive wavelet-neuro-fuzzy systems with W-neurons are increasingly used [4, 5]. In
wavelet-neuro-fuzzy networks, wavelet neuron design is sufficiently similar to the
standard formal neuron with N inputs, but instead of the usual synaptic weight λik ,
W Si
wavelet-synaptic weight λik (i = 1, 2, ..., N ) are used, in which not only the weight
s but also the scaling and displacement parameters of the wavelets are tuned.

I. Peleshchak (B) · V. Lytvyn

Lviv Polytechnic National University, Lviv 79000, Ukraine
e-mail: [email protected]
V. Lytvyn
e-mail: [email protected]
R. Peleshchak · J. Kopka · M. Wrzesien · J. Korniak · J. Kolbusz · P. Rozycki
University of Information Technology and Management in Rzeszow, 35-225 Rzeszow, Poland
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 313

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_22
314 I. Peleshchak et al.

It should be noted that in all of these artificial neural networks, the neurons were
considered as static, that is, without its own dynamics (self-oscillation frequency of
the neuron ω0i = 0), without taking into account the accumulation of impulses N0k
in neurons that trigger when N0k ≥ Nck (Nck —threshold pulses is k-th neuron) and
without taking into account non-locality in time.
In [6, 7] the authors suggested that neurons of the human brain generate refer-
ence waves of different frequencies (ω0i , i = 1, 2, ..., k). With the help of reference
waves with a certain frequency, we “poll” neurons, that is, analyze the input image
at different frequencies and compare the responses to input images that were previ-
ously memorized. In this case, each neuron operating on the principle of “question–
answer” answers only the question to which “knows” the answer, as it performs, for
example, the Grossberg neuron. The basis of human brain neurons is the microtubules
of the cytoskeleton [8, 9]. The microtubule subunits of the cytoskeleton are tubulin
molecules. The structure of such a molecule is a dimer, that is, it consists of two parts
(α and β—tubules) connected by a thin jumper. The approximate resonant frequen-
cies of tubulin molecules, which are subunits of the microtubule of the cytoskeleton,
are ω0ir es
= 1011 –1013 Hz [9]. Thus, it can be assumed that the resonant frequency
of biological neurons will be approximately the same as the resonant frequency of
tubulin molecules.
Therefore, the urgent task is to develop an artificial neural network with dynamic
(oscillatory) neurons that have natural frequencies (ω0i , i = 1, 2, ..., k). With the
help of this type of artificial neural network we realize a new method of recog-
nition of multispectral images (multispectral electromagnetic signals) on the basis
of information resonance. These multispectral signals include: charged heterophase
plasma nanoparticles, high-temperature industrial aerosol emissions, ionized dust
and gas formations in the atmosphere and space, products of meteorite destruction
during their interaction with the Earth’s atmosphere, and the accumulation of cosmic
conditions.

2 Mathematical Model of Oscillatory Neural Network

The mathematical model of neural network architecture with dynamic (oscilla-

tory) neurons is described by the dependence according to which the values of
the initial state vector Ym = {y1 (ω1 ), y2 (ω2 ), ..., ym (ωm )} ⊂ Y can be esti-
mated from the known values of the vector of the input characteristics Vn =
{V1 (ω1 ), V2 (ω2 ), ..., Vn (ωn )} ⊂ V (Vn (ωn ), n = 1, N —the n-th component of the
vector of the input spectral characteristic). For the recognition of multispectral input
images Vn = {V1 (ω1 ), V2 (ω2 ), ..., Vn (ωn )} ⊂ V , we offer a three-layer perceptron
with dynamic neurons. It has an input layer containing the number of neurons equal
to the number of input attributes; hidden layer; a output layer in which the number
of neurons is equal to the number of decision classes.
Spectral Image Recognition Using Artificial Dynamic … 315

The hidden layer consists of M dynamic neurons m = 1, M (with its own oscil-
lation frequency (ω0i , i = 1, 2, ..., M)), which receives information from the first
layer, which has N inputs.


N
Vm = Ṽnm · ω̃nm
2
· sin2 (ω̃nm · τ ) + Ṽ0m (1)
n=1

where Ṽnm is the weight coefficient of the n-th input characteristic entering the input
of the m-th dynamic neuron of the hidden layer; Ṽ0m —threshold.
The functioning of the m-th dynamic neuron is described by the inhomogeneous
oscillatory differential equation of the 2nd order.

d 2 Um dUm  N
+ 2 δ̃ + ω̃ 2
0m U m = Ṽnm ωnm
2
sin2 (ωnm τ ) (2)
dτ 2 dτ n=1

 
where 2δ̃(Nk , Nck ) = δ tanh Nk −N σ2
ck
is the attenuation parameter of the m-th neuron;
Nk ; Nck ; σ 2 is the number of pulses that arrive at the m-th neuron; the threshold
value of the mth neuron pulse; the variance.
The mathematical model (2) can be used to investigate the collective behavior
of ensembles of neurons interconnected by synaptic connections λ jm . To do this,
the second and third additions must be replaced Um → Uαm , where Uαm = Um +
N
λ jm U j ; α = 1, 2, ..., N [10].
j=1
The output signal of the m-th dynamic neuron Um (τ ) is described by the expression
that is the solution of Eq. (2)


N 
N
Um (t) = C̃nm + λ̃nm |sin(2ω̃nm τ + ϕ̃n )| (3)
n=1 n=1
 2 2 2

16Ṽnm δ̃ 2 ω̃nm
4
+ Ṽnm ω̃nm 2
ω̃0m − 4ω̃nm
C̃nm =     (4)
2ω̃0m
2 2 + ω̃2 − 4ω̃2 2
16δ̃ 2 ω̃nm 0m nm

Ṽnm ω̃nm
2
2
λ̃nm =   (5)
ω̃0m
2
− 4ω̃nm 2 2 + 16δ̃ 2 ω̃2
nm
316 I. Peleshchak et al.

ω0m
2
ω̃m
r es
= (6)
2 ω0m
2
− 2δ 2

Ṽnm ω0m
2
λ̃rmes = (7)
16|δ| ω0m
2
− δ2

where ω̃nm is the frequency of the nth spectral image arriving at the m-th oscillatory
neuron in the hidden layer.
It should be noted that in order to have oscillations it is necessary that in formula
(6) ω̃mr es
.
In biological neurons, after the action of the membrane of the neural cell, the
refractory period begins (relaxation time tr = 2–30 ms), so in an artificial neuron
it is necessary to enter a nonlocal time—delay tr . This application also allows for
smoothing out random Um (τ ) signal jumps. Nonlocality can be performed by both
the adder and the neuron itself, if the nonlocality is set linearly then it can be written

tr

Ũm (τ ) = e−α1 t Um (τ − t)dt (8)

0

where Um (τ − t) is the solution (3) of Eq. (2); 0 < α1 < 1.

The activation function in the hidden layer of a dynamic neural network is taken
as a sigmoid function

1
z m = f (Um ) = (9)
1 + e−βm Um

where Um is defined by formula (3); βm is the coefficient that determines the slope
of the activation function f (Um ).
The output layer of neurons has a threshold activation function. These neurons
are used to determine the specific class to which the input spectral images belong.
⎧
⎪ 
M

M ⎪
⎨ 1, i f vmk z m − v0k ≥ 0
yk = ϕ vmk z m − v0k = m
, k = 1, K (10)
⎪
⎪ 
M
m=1 ⎩ 0, i f vmk z m − v0k < 0
m

where vmk is the coefficients of weight between the neurons of the hidden layer and
the neurons of the output K-th layer.
Spectral Image Recognition Using Artificial Dynamic … 317

Substituting formula (5) into formula (3) with respect to |sin(2ω̃nm τ + ϕ̃n )| ≤ 1,
we obtain a non-resonant amplitude value at the output of the m-th neuron, when N
multispectal signals arrive at the input of this neuron


N
Um = Ṽnm xn (11)
n=1

ω̃nm
2
where xn = .
2 ( 2
)
2 2
ω̃0m −4ω̃nm +16δ 2 ω̃nm
2

To obtain the resonant amplitude value of the signal, it is necessary to put (7) in
the formula (3).
A necessary condition for recognition of multispectral images by the resonant
effect is the inequality |ω̃nm − ω̃r es | ≤ ε, where 0 < ε  1.
The architecture of a three-layer neural network with dynamic neurons consists
of 6 input spectral images (sensor layer), a hidden layer with 10 dynamic neurons,
and an output dynamic neuron (Fig. 1).

Fig. 1 Architecture of a three-layer dynamic neural network in Matlab

318 I. Peleshchak et al.

3 The Algorithm of Training a Three-Layer Neural

Network and Its Testing in Matlab

The algorithm for learning and evolving a dynamic neural network is based on
the scalable connect gradient method (TRAINSCG). The training minimizes the
objective function.

 N    2
 1  K 
 yk ω̃nm , P − ykr es ⇒ min (12)
N K k=1 n=1 P

   
where P = Ṽnm , ω̃0m , δ, ω̃nm − ω̃m
r es 
< ε, νmk , ν0k , βm is a vector of parameters
that are tuned during neural network training;
ykr es
 —required
 output values of the neural network;

yk ω̃nm , P is the k-th output value of the neural network when the input value
is ω̃nm .
Construction of arrays of input–output pairs for training neural network.
1. Creating an element (xn , yk )

delta=0.01;
eps=0.1;
wnm=[0.4791 0.5801 0.6811 0.7821 0.8831 0.9841] ;
V0k=0.46;
for k=1:6
Vn(k)=0.2;
end;
Xn3=(wnm-eps).^2./(16*delta*sqrt((wnm-eps).^2-delta^2))
for i=1:10
Vmk(i) =0.8;
beta(i)=0.1;
end;

SUr=sum(Vmk./(1+exp(-beta.*sum(Vn.*Xn3)))-V0k)

if SUr < 0
Yr3=0;
else
Yr3=1;
end;
Yr3
save D:\mat1\Xn3
save D:\mat1\Yr3
Spectral Image Recognition Using Artificial Dynamic … 319
 
2. Write this element (xn , yk ) to array X , Y

load('Xn.mat')
load('Xn1.mat')
load('Xn2.mat')
…
load('Xn29.mat')
XnS(:,1)= Xn'
XnS(:,2)= Xn1'
XnS(:,3)= Xn2'
…
XnS(:,30)= Xn29'
XnS
save D:\mat1\XnS
load('Yr.mat')
load('Yr1.mat')
load('Yr2.mat')
…
load('Yr29.mat')
YrS= [Yr Yr1 Yr2 … Yr29];
YrS
save D:\mat1\YrS

Neural Network Testing.

To train the neural network, 20 pairs of input-outputs were selected, of which 13
were resonance elements and 7 non-resonance elements. All pairs were successfully
saved (Fig. 2).
For validation, 5 elements were selected, 2 of them with resonance and 3 without
resonance. All were successfully recognized.
For testing examples, 5 items were selected without resonance. All of them were
successfully recognized.
In particular, another test was conducted with the other 20 items, all of which
were successfully recognized (Fig. 3).

4 Conclusions

Mathematical model and architecture of artificial neural network with dynamic

neurons for recognition of multispectral images are constructed.
An algorithm for training a three-layer neural network with dynamic neurons has
been developed.
320 I. Peleshchak et al.

Fig. 2 Test result of dynamic neural network

The results of testing the neural network for the recognition of multispectral
images based on the resonant effect.
Spectral Image Recognition Using Artificial Dynamic … 321

Fig. 3 Second test with other examples

References

1. Wang, L.: Adaptive Fuzzy Systems and Control. Design and Stability Analysis, 352 pp. Prentice
Hall, New Jersey (1994)
2. Jang, J., Sun, C., Muzutani, E.: Neuro-Fuzzy and Soft Computing: A Computational Approach
to Learning and Machine Intelligence, 614 pp. Prentice Hall, Inc., Upper Saddle River, N.J.
(1997).
3. Abiyev, R., Kaynak, O.: Fuzzy wavelet neural networks for identification and control of
dynamic plants—a novel structure and a comparative study. IEEE Trans. Ind. Electron. 55(8),
3133–3140 (2008)
4. Bodyanskiy, Y., Pliss, I., Vynokurova, O.: Hybrid wavelet-neuro-fuzzy system using adaptive
W-neurons. Wissenschaftliche Berichte, FH Zittau/Goerlitz 106, 301–308 (2010)
5. Vinokurova, E.: Hybrid adaptive neuro-fuzzy and wavelet-neuro-fuzzy systems of computa-
tional intelligence in signal processing problems in noise. Adapt. Autom. Control Syst. 15(35),
113–120 (2009)
6. Kholmansky, A.: Simulation of brain physics. Quantum Magic 3(3), 3126–3155 (2006)
7. Smith, K.: Sensory Systems Biology, Moscow BINOM. Laboratory of Knowledge, 583 pp.
(2005)
8. Hameroff, S., Penrose, R.: J. Conscious. Stud. 1, 91–118 (1994)
322 I. Peleshchak et al.

9. Slyadnikov, E.: Physical model and associative memory of the dipole system of the cytoskeleton
microtubule. J. Tech. Phys. 7, 77–86 (2007)
10. Lytvyn, V., Vysotska, V., Peleshchak, I., Rishnyak, I., Peleshchak, R.: Time dependence of the
output signal morphology for nonlinear oscillator neuron based on Van der Pol model. Int. J.
Intell. Syst. Appl. (IJISA) 10(4), 8–17 (2018)
U-Net Based Model for Obfuscated
Human Faces Reconstruction

Elmoukhtar Zemmouri , Youssef Yakoubi, Mohammed Douimi ,

and Adil Saadi

Abstract Protecting the identity of users in social networks and photographic media
has become one of the most important concerns, especially with the breakthrough
of internet and information technology. Thus several techniques, such as blurring
and pixelization, have been introduced by privacy-protection technology designers to
make the features of a user’s face unrecognizable. However, many researchers pointed
out the ineffectiveness of these techniques. In this paper with deal with the problem
of reconstructing obfuscated human faces. We propose a Deep Neural Network based
on U-Net and trained on Labeled Faces in the Wild (LFW) dataset. Our model is
evaluated using two image quality metrics PSNR and SSIM. Experimental results
show the ability of our model to reconstruct images that have underwent different
forms of obfuscation (blurring and pixelization) with great accuracy, outperforming
previous work.

Keywords Deep neural network · Image reconstruction · Obfuscation · U-Net ·

VGG-16

1 Introduction

With the increasing use of mobile devices coupled with ubiquitous availability of
Internet connectivity, photos end up by being shared online at a very large scale.
This is further amplified by the popularity of social networks and the immediacy

E. Zemmouri (B) · Y. Yakoubi · M. Douimi · A. Saadi

Moulay Ismail University, ENSAM Meknes, Morocco
e-mail: [email protected]
Y. Yakoubi
e-mail: [email protected]
M. Douimi
e-mail: [email protected]
A. Saadi
e-mail: [email protected]
© Springer Nature Switzerland AG 2021 323
T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_23
324 E. Zemmouri et al.

they offer in content sharing. Unfortunately, such growth have raised privacy con-
cerns which have been discussed and studied by the research community. In these
photographic media, privacy can be breached by identifying human faces, license
plates and contents of computer screens.
Several privacy protection technologies were introduced in order to address the
digital privacy concerns. Most of these technologies rely on image obfuscation.
This is the process of making the sensitive parts of the image unrecognizable by
humans. Sometimes full obfuscation is performed on images and videos (using solid
rectangles), which blocks all the information. But usually, partial obfuscation is used
in order to preserve useful information of photographs that are not privacy-breaching,
thus achieving a good balance between privacy and utility.
In this work we consider the most two common image obfuscation techniques:
the blurring and the pixelization (also known as mosaicing).
For the first technique, we used the Gaussian blur (also known as Gaussian
smoothing). Mathematically, this is achieved by convolving the image by a Gaussian
kernel with parameter σ . The larger σ is, the higher level of obfuscation and the
less details in the image. Since the image is stored as a collection of discrete pixels,
the values in the kernel are obtained by a discrete approximation to the Gaussian
function in order to perform the convolution operation. In the context of our work,
we will examin three levels of blurring σ = 4, σ = 6 and σ = 8. Figure 1 illustrates
the effect of the Gaussian blurring with the three different levels by varying the
parameter σ .
For the second obfuscation technique, a pixelized image is obtained by superpos-
ing a rectangular grid over the original image and averaging the color values of the
pixels within each grid cell [3]. The size of the grid as a parameter (denoted by n)
can be varied to determine the level of obfuscation. The larger the grid cell, the more
pixels will be averaged together, so the more obfuscated the transformed image will
be. For the purpose of our work, we will fix the grid cell size at three different levels:
8 × 8, 12 × 12 and 16 × 16 pixels. Figure 2 illustrates the effect of the pixelization
with three different levels by varying the parameter n.

Original Blurred σ = 4 Blurred σ = 6 Blurred σ = 8

Fig. 1 A sample image from the LFW dataset and its obfuscated version using Gaussian blurring
with three different levels
U-Net Based Model for Obfuscated Human Faces Reconstruction 325

Original Pixelized n = 8 Pixelized n = 12 Pixelized n = 16

Fig. 2 A sample image from the LFW dataset and its obfuscated version using pixelization with
three different levels

Neverthless, many researchers have examined the effect of those techniques and
suggested that blurring and pixelization may not be the best techniques to protect
privacy, and do not provide sufficient guarantees of anonymity.
In this paper, we will illustrate how deep neural networks can be a challeng-
ing adversarial technology against these privacy protection techniques. Since deep
convolutional network have recently known a great success in many computer
vision tasks like image recognition, image segmentation and object detection, many
researchers have applied this new approach for image reconstruction.
Our approach consists of training an artificial neural network, and more precisely
a convolutional neural network based on U-Net architecture [11], using pairs of
original and obfuscated images. The obfuscated images were fed to the network
that tries to minimize the distance between its output and the target images that are
the original images. This allows the network to learn the transformation done by the
obfuscation techniques. We have examined different loss functions and see how other
loss functions perform better than the traditional Euclidean distance also known as
the mean squared error.
The rest of the paper is organized as follow: in Sect. 2, we will examin some pre-
vious approaches that demonstrated the ineffectiveness of pixelization and blurring
as a privacy protection techniques. In Sect. 3, we will describe the architecture of
our model. In Sect. 4 we present experimental implementations and results. Finally
some conclusions.

2 Related Work

Image restoration or reconstruction refers to the process of reducing blur or noise or

both in the image. This is usually an ill-posed problem in the sense of Hadamard,
which means that, generally, it does not have a unique solution or the solution’s
behavior does not change continuously with the initial data.
In the area of image processing and computer vision, a number of different tech-
niques have been introduced for digital image restoration. Traditional restoration
techniques are oriented towards modeling the degradation and applying the inverse
326 E. Zemmouri et al.

process in order to recover the original image. Among the deblurring techniques
for instance, we can find the inverse filter, the Wiener filter, the Lucy-Richardson
algorithm and so many others. We refer the reader to this paper [8] for more image
restoration techniques.
The authors of [1] have proposed an approach performing an attack against pri-
vacy of pixelized videos, assuming that the same image has been recorded in different
frames at different positions. As the same subject moves in a video, the pixelization
squares change position. Therefore averaging different pixels at different times. Pro-
cessing these video frames can be helpful in obtaining a reconstruction of the original
image. Their experiments were limited to gray-scale videos. They also made some
assumptions on the video such as there is no zooming or horizon rotation and the
only allowed movements for the camera are horizontal/vertical translation or slight
rotation.
The problem was formalized as to find an image such that, if shifted and pixelized
in multiple frames, will give an output sequence that is as close as possible to the
input video. Their procedure leads to a linear system P x = b where x represents the
image they attempt to reconstruct. Obviously, this system is very unlikely to have a
solution for real-world videos that are so far from satisfying the assumptions made
by the authors on artificial videos. However by following an iterative process, they
demonstrated that the proposed approach can recover images from pixelized videos.
This example illustrates some limitations of the classic approaches that makes
restrictive assumptions that are in most cases far from real-world conditions.
Recently, with the success of artificial neural networks in many computer vision
applications like object detection and image recognition, several researchers have
applied this new approach based on convolutional neural networks in image recon-
struction.
In their paper [9], McPherson et al. have shown how to train artificial neural
networks to successfully classify images that have underwent different forms of
obfuscation. In order to demonstrate how their approach can defeat these techniques,
they have applied it on four different datasets which are widely used for face, object,
and handwritten digit recognition. The classic MNIST dataset of black and white
handwritten digits, the CIFAR-10 dataset of colored images in 10 classes of animals
and vehicles (truck, frog, deer …) and the AT&T dataset of black and white human
faces. Finally the FaceScrub dataset of face images of celebrities collected from the
internet.
A similar work as ours was done in this paper [15] where the author have trained a
deep convolutional neural network on the Labeled Faces in the Wild dataset using two
techniques of obfuscation: pixelization and Gaussian blurring. The model architec-
ture is a convolutional neural network with an input and output size of 112 × 112 × 3.
Images were sent through 4 convolutional layers. The first layer consists of 64 9 × 9
filters. The following three layers consists of 64, 128, and 256 4 × 4 filters respec-
tively with stride 2. Then, the images were sent through 3 residual blocks. Each block
is 2 convolutional layers with 256 3 × 3 filters. Then it goes through 4 transpose con-
volution layers. The first three layers have 256, 128, and 64 4 × 4 filters respectively
with stride 2. The last convolutional layer has 3 9 × 9 filters. All convolution and
U-Net Based Model for Obfuscated Human Faces Reconstruction 327

transpose convolution layers are followed by batch normalization and ReLU activa-
tion, except the last layer which is followed by a tanh activation and then scaled to
the range [0, 255].
Because this last work deals directly with the problem of reconstructing obfuscated
faces, we will use it as a reference for comparison in the experiments section.

3 Methodology

As we have mentioned, our model was inspired by the U-Net architecture (see Fig. 3).
This network was developed for biomedical image segmentation and has achieved
good performance on different biomedical segmentation applications [11]. Image
segmentation is the process of partitioning an image into multiple segments (sets
of pixels). The goal of segmentation is the separation of touching objects of the
same label or class. In our work we have adjacted U-Net for the task of image
reconstruction.

Fig. 3 U-Net architecture as proposed by [11] with an image size of 32 × 32 pixels. A blue box
represents a multi-channel feature map where the number of channels is on top of the box and the
size is at the lower left edge. A white box represents a copied feature map. The arrows represent
the different operations
328 E. Zemmouri et al.

3.1 U-Net Architecture and Adaptation

U-Net is a Deep Convolutional Neural Network that takes as input an image and
outputs a segmentation map (a label for each pixel). It consists of two paths: a
contracting path to capture context, and a symmetric expanding path that enables
localization.
Each block of the contracting path applies two 3 × 3 convolution layers (unpadded
convolutions), each followed by a ReLU activation and a 2 × 2 max pooling operation
with stride 2 for downsampling.
For the expansive path, each block applies an upsampling followed by a 2 × 2
“up-convolution” layer, a concatenation with the correspondingly cropped feature
map from the contracting path, and two 3 × 3 convolutions, each followed by a
ReLU activation.
The final layer which applies a 1 × 1 convolution is used to map each 64-
component feature vector to the desired number of segments (classes).
In our work, we have modified this network so that it can output an image of the
same size as the input, which is 112 × 112 × 3 in our case. In the contracting path
we have used padded convolutions with the adequate padding to keep the size intact
after each convolution. We have also changed the last layer and the loss function to
one of the losses that we will define in the next subsection. The Sigmoid activation
function at the last layer makes sure that the output of our model is in [0, 1] range.

3.2 Loss Functions

In the domain of Machine/Deep Learning, loss functions or cost functions are used
to measure and minimize the error (that is the loss) between the model output and
the given terget value. In our experiments we used two loss functions.

3.2.1 Per-Pixel Loss Function

Per-Pixel Loss Function is a simple loss function that depends only on low-level
pixel information. It is simply defined by the squared normalized euclidean distance
between pixels of the target image y and pixels of the model output ŷ. Since we
have in our case images with size 112 × 112 × 3 the pixel loss is defined by the
expression:
1
 pi xel = y − ŷ22
112 × 112 × 3

Per-Pixel loss has many drawbacks as it does not accurately capture perceptual
qualities of images. This is why other loss functions that focus more on high-level
representations of the image have been proposed in the literature.
U-Net Based Model for Obfuscated Human Faces Reconstruction 329

3.2.2 Perceptual Loss Function

Perceptual Loss Function is a function used to compare two different images that look
similar by capturing perceptual differences between the two images (for example an
image that is shifted by only a few pixels). Perceptual loss functions measure image
similarities more robustly than per-pixel loss functions [5]. In fact, they focus on
differences between high-level features instead of differences between pixels. These
high-level features are generally extracted from a pretrained convolutional network
called the loss network. Perceptual losses have been recently used for tasks like style
transfer and super-resolution [5, 7].
Let M denote our model and  = {φ1 , φ2 , . . . , φn } a loss network, where φi (I )
denotes a feature map obtained at the ith layer of shape Wi × Hi × Ci of an image I .
Given an obfuscated image x, it’s original (ground truth) y and ŷ = M(x) the output
image of our model. The perceptual loss at the ith layer between the model output
and the original image is defined by:
1
 per cep = φi (y) − φi ( ŷ)22
Wi × Hi × Ci

In this work, we choose the VGG-16 [14] as a loss network pretrained on Imagenet
dataset [12]. And we captured feature representations from the relu2_2 layer of the
VGG-16. Figure 4 depict the overall architecture.
So, the transformation network in our approach consists of the proposed adaptation
of U-net network. Whereas for Trinidad’s approach [15], it consists of a simple CNN
as detailed in Sect. 2.

3.3 Image Quality Metrics

Digital images are subject to many forms of distortions including noise, blurring,
ringing and compression. These types of distortion occur during image acquisition,
processing, storage and transmission which may result in a degradation of the visual

Fig. 4 Model overview. The transformation network is our proposed U-Net based architecture.
The loss network remains fixed during the training
330 E. Zemmouri et al.

quality of the image. Many quality metrics have been introduced in order to compare
image processing algorithms and define objective measures of quality. We can cite
two different types of metrics:
– Full-reference quality metrics that assume the existence of the original image
(without distortion) which we call the reference image, and compare it to the target
image (generally the reconstructed version of the distorted image). These metrics
include The mean squared error (MSE), the peak signal-to-noise ratio (PSNR)
and the structural similarity (SSIM) index.
– No-reference quality metrics, also known as blind quality assessment, are used
when no reference image is available. Hence they use some statistical features
of the image to assess image quality. An example of these no-reference image
quality assessment is the Blind/Referenceless Image Spatial QUality Evaluator
(BRISQUE) [10]. A BRISQUE evaluator is trained on a set of images with known
distortions, and can only evaluate quality of images with the same type of distortion.
In our experiments, reference images are available, so we will use PSNR and
SSIM metrics to evaluate the quality of our reconstruction algorithm.

3.3.1 Peak Signal-to-noise Ratio (PSNR)

PSNR is a traditional metric that is used to measure the quality of image reconstruc-
tion. It’s defined via the Mean Squared Error. Given an image I with size m × n and
its noisy approximation J with the same size, the mean squared error between I and
J is defined by the formula:

1 
m−1 n−1
M S E(I, J ) = (I (i, j) − J (i, j))2
mn i=0 j=0

The PSNR is expressed (in dB) by this formula:

 M AX 2I 
P S N R(I, J ) = 10 · log10
M S E(I, J )

where M AX I is the maximum possible pixel value of the image I.

Generally, the two metrics MSE and PSNR are popular and widely used because
of their simplicity, mathematical and physical meaning, but also they are easy to use
in the context of optimization. However many studies have shown that both of them
do not correlate well with subjective human perception of quality [2, 13, 17].

3.3.2 Structural Similarity (SSIM)

The structural similarity index was first developed by Zhou Wang and Alan Bovik
in 2001. The current version of this metric was published in [18]. Unlike MSE and
U-Net Based Model for Obfuscated Human Faces Reconstruction 331

PSNR metrics, the SSIM is a perception-based model that combines local structure,
luminance, and contrast into a single local quality score. It is calculated using a
sliding window on the images. The measure between two local windows x and y of
same size N × N is given by this formula:
(2μx μ y + c1 )(2σx y + c2 )
SS I M(x, y) =
(μ2x+ μ2y + c1 )(σx2 + σ y2 + c2 )

where:
– μx the average of x
– μ y the average of y
– σx2 the variance of x
– σx2 the variance of y
– σx y the covariance of x and y
– c1 = (k1 L)2 , c2 = (k2 L)2 two variables to stabilize the division with weak
denominator
– L the dynamic range of the pixel-values (typically this is 2#bits per pi xel−1 )
– k1 = 0.01 and k2 = 0.03 by default.
In order to evaluate the overall image quality we use a mean SSIM over the local
windows: If I is the original image and J is its distorted version, M the number of
local windows, then the SSIM is given by:

1 
M
SS I M(I, J ) = SS I M(xi , yi )
M i=1

3.4 Dataset, Preprocessing and Training

In this work, we used the Labeled Faces in the Wild dataset (LFW), published in
2007 [4]. It contains more than 13,000 images of faces collected from the web. Each
face has been labeled with the name of the person in the picture. 1680 of the people
pictured have two or more distinct photos in the dataset. The LFW images are the
result of detection of the Viola-Jones face detector [16]. Hence the faces are usually
(but not always) centered and are at the same scale. The images were also cropped
to a fixed size.
We follow the same preprocessing steps on this dataset as in this paper [15]. This
allows us to compare numerical results with this work. The preprocessing steps are
as follow.
Firstly, we croped each image from 250 × 250 to 112 × 112, maintaining the
center. This slightly reduces the computational complexity, but the main reason of
this cropping process is to focus on the facial region of the picture and ignore all the
background.
332 E. Zemmouri et al.

Secondly, we obfuscated each image of the entire dataset using two obfuscation
techniques: blurring and pixelization (Figs. 1 and 2).
Finally, we split our datasets (the original and obfuscated) into 8000 training
examples and the remaining images into a validation and testing examples. The
model takes as input the obfuscated image and outputs a reconstructed version of
it by minimizing the error between the output of the model and the target image
(the original image). We train our model on only one technique and one level of
obfuscation at a time.

4 Experiments and Results

In all our experiments we used the Adam optimizer [6]. Our initial runs have shown
that a learning rate of 10−3 , as used in [15], is too high and works only with β1 = 0.5
but generally with this learning rate, our model was easy to diverge. Hence we set
our learning rate to 10−4 and keep the other parameters of the Adam optimizer as
suggested by its authors: β1 = 0.9, β2 = 0.999 and  = 10−8 .
We have tested a batch size of 8, 16 and 32, but the initial runs have shown that
a batch size of 8 was the best in terms of the PSNR and SSIM metrics. Since our
metrics plateau, on the validation set, at around 20 epochs and the resulting images
are about the same after this number of iterations, we trained our model for only 20
epochs.
For the hardware configuration, we used an Nvidia GeForce GTX 980 Ti GPU. It
consists of 2816 Cuda core and 6 Go memory. A training session of 20 epochs run
in approximately 50 min.

4.1 Evaluation of Per-Pixel Loss Implementation

In our first implementation we have used per-pixel loss function. We evaluated the
performance of our architecture in terms of PSNR and SSIM on the blured dataset
with σ = 4 and on the pixelized dataset with n = 8.
Figure 5 shows the results of this experiment on two sample images from the
training and test sets. Table 1 depict the results in terms of mean PSNR and mean
SSIM. These results show that the metrics on the training set are only slightly better
compared to the test set, hence our model is not overfitting and can achieve good
performance on both sets. The resulting images illustrate the ability of the network
to reconstruct a very similar image to the original from its obfuscated version. But
we can also see that thin lines and fine details of faces are generally lost. Notice also
that the output images for the blurring technique are visually more pleasing and their
corresponding metrics are better compared to the pixelized dataset.
U-Net Based Model for Obfuscated Human Faces Reconstruction 333

Training Set Original Blurred Output Pixelized Output

PSNR/SSIM 23.44 / 0.80 31.69 / 0.94 21.51 / 0.70 27.79 / 0.89

Test Set

PSNR/SSIM 23.02 / 0.76 30.67 / 0.94 20.27 / 0.65 25.24 / 0.87

Fig. 5 Results for the training and the test sets using the per-pixel loss. The PSNR/SSIM were
mentioned for each example. Levels of obfuscation are σ = 4 for blurring and n = 8 for pixelization

Table 1 Mean PSNR/SSIM on training and test set

Dataset Training set Test set
Blurred (σ = 4) 31.58/0.88 31.33/0.88
Pixelized (n = 8) 27.30/0.81 25.54/0.78

4.2 Evaluation of Perceptual Loss Implementation

In our second implementation, we used the same datasets and parameters as above.
We changed the loss function by introducing VGG-16 as a loss network in our
architecture (see Fig. 4). We evaluated the performance using the same metrics.
Figure 6 shows that the use of perceptual loss results in low values of the PSNR
metric. This was expected since the PSNR is defined via the mean squared error
which is exactly the function that we minimize when using the pixel loss. The SSIM
does not necessarily goes down, which gives us an idea about how perceptual loss
may recovers the structural information of the images. Results also show that the
models trained using the perceptual loss outputs images with sharper edges with
more details.
We compared the performance of our model with the model proposed by Trinidad
in [15]. Table 2 shows a results of the comparison in terms of the mean PSNR/SSIM
metrics on the validation set between our neural network architecture and the
Trinidad’s architecture.
Figure 7 depicts a comparison of the evolution of the PSNR and SSIM metrics
during training between the two architectures.
334 E. Zemmouri et al.

Original Blurred pixel percep pixelized pixel percep

PSNR/SSIM 23.33/0.72 28.91/0.89 28.25/0.91 20.83/0.60 25.52/0.82 23.88/0.83

PSNR/SSIM 23.22/0.75 29.93/0.91 28.77/0.92 21.17/0.67 25.05/0.84 24.16/0.84

Fig. 6 Comparing the results of per-pixel loss and perceptual loss for the training set (first row) and
the test set (second row). The PSNR/SSIM are mentioned for each example. Levels of obfuscation
are : σ = 4 for blurring and n = 8 for pixelization

Table 2 Mean PSNR/SSIM comparison on the validation set

Dataset Trinidad Our architecture
Blurred (σ = 4) 27.28/0.8337 31.28/0.8758
Pixelized (n = 8) 24.78/0.7555 25.51/0.7797

Fig. 7 Evolution of PSNR and SSIM during training. First row is the training on the blurred dataset
with σ = 4. The second row is the training on the pixelized dataset with n = 8
U-Net Based Model for Obfuscated Human Faces Reconstruction 335

4.3 Evaluation with High Obfuscation Levels

In the last experiment, we tested our architecture on high obfuscation levels. Results
have shown that the pixel loss perform badly on highly obfuscated dataset. We only
present the results obtained using the perceptual loss at relu2_2 layer of VGG-16.
Figure 8 shows that even with highly obfuscated dataset, our model can still output
recognizable faces that are similar to the original. Notice again that the models trained

Blurred Blurred
Original (σ = 6) Output (σ = 8) Output

PSNR/SSIM 20.82 / 0.67 25.58 / 0.90 19.41 / 0.62 23.90 / 0.86

PSNR/SSIM 21.72 / 0.67 25.73 / 0.86 20.70 / 0.63 24.76 / 0.85

pixelized pixelized
Original (n = 12) Output (n = 16) Output

PSNR/SSIM 17.27 / 0.52 20.61 / 0.81 16.23 / 0.46 20.51 / 0.79

PSNR/SSIM 18.94 / 0.54 21.61 / 0.78 18.04 / 0.53 20.95 / 0.76

Fig. 8 Results on training set (first and third rows) and test set (second and fourth rows) of high
obfuscation levels experiment. Blurring with σ = 6 and 8. Pixelization with n = 12 and 16. The
PSNR/SSIM are reported for each example
336 E. Zemmouri et al.

on the pixelization technique outputs images with different facial features (looking
younger due to the absence of wrinkles) and find it harder to recover an identical
face as the original.

5 Conclusion

In this work, we have examined the effectiveness of traditional image obfuscation

techniques in protecting the human identity. First we introduced the most two com-
mon techniques of obfuscation which are pixelization and blurring that are widely
used in social media in order to protect the user from directly viewing offensive
content while browsing the internet. These technologies are also used also used
for hiding objects and human faces in photographic media as recently proposed by
Youtube. As many papers and studies have suggested their inefficiency, we have
examined the ability of artificial neural networks to challenge these techniques. We
have inspired our model from the U-Net architecture, which is a convolutional neural
network designed for biomedical image segmentation. Generally our trained models
have shown good performance in reconstructing accurate images of faces that have
underwent a process of obfuscation and recovering what seems unrecognizable by
humans. We have tested several loss functions and show how perceptual loss gives
better results than the per-pixel loss function. Our proposed architecture was a deep
convolutional neural network, that outperformed (visually and in term of metrics) the
architecture proposed in a prior work. Next we have trained our architecture on highly
obfuscated images and results have shown that our model can output accurately recon-
structed and recognizable human faces from highly obfuscated and unrecognizable
faces. This work show furthermore fundamental concerns about privacy protection
on social media, suggesting the need of more efficient techniques.

References

1. Cavedon, L., Foschini, L., Vigna, G.: Getting the face behind the squares: reconstructing
pixelized video streams. In: 5th USENIX Workshop on Offensive Technologies, WOOT’11,
August 8, 2011, San Francisco, CA, USA, Proceedings. pp. 37–45 (2011)
2. Girod, B.: What’s Wrong with Mean-Squared Error?, pp. 207–220. MIT Press, Cambridge,
MA, USA (1993)
3. Hill, S., Zhou, Z., Saul, L., Shacham, H.: On the (in)effectiveness of mosaicing and blurring
as tools for document redaction. PoPETs 2016(4), 403–417 (2016). https://doi.org/10.1515/
popets-2016-0047, https://doi.org/10.1515/popets-2016-0047
4. Huang, G.B., Mattar, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild: a database
for studying face recognition in unconstrained environments. Tech. Rep. (2007)
5. Johnson, J.M., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-
resolution. In: European Conference on Computer Vision (2016)
U-Net Based Model for Obfuscated Human Faces Reconstruction 337

6. Kingma, D.P., Ba, J.: Adam: a method for stochastic optimization. In: 3rd International Con-
ference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015,
Conference Track Proceedings (2015). http://arxiv.org/abs/1412.6980
7. Ledig, C., Theis, L., Huszar, F., Caballero, J., Aitken, A.P., Tejani, A., Totz, J., Wang, Z., Shi,
W.: Photo-realistic single image super-resolution using a generative adversarial network. In:
2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 105–114
(2016)
8. Maru, M., Parikh, M.C.: Image restoration techniques: a survey. Int. J. Comput. Appl. 160(6),
15–19 (2017). https://doi.org/10.5120/ijca2017913060, http://www.ijcaonline.org/archives/
volume160/number6/27077-2017913060
9. McPherson, R., Shokri, R., Shmatikov, V.: Defeating image obfuscation with deep learning.
CoRR abs/1609.00408 (2016)
10. Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spa-
tial domain. Trans. Img. Proc. 21(12), 4695–4708 (2012). https://doi.org/10.1109/TIP.2012.
2214050, https://doi.org/10.1109/TIP.2012.2214050
11. Ronneberger, O., Fischer, P., Brox, T.: U-net: convolutional networks for biomedical image
segmentation. CoRR abs/1505.04597 (2015), http://arxiv.org/abs/1505.04597
12. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A.,
Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: Imagenet large scale visual recognition
challenge. Int. J. Comput. Vis. 115(3), 211–252 (2015). https://doi.org/10.1007/s11263-015-
0816-y, https://doi.org/10.1007/s11263-015-0816-y
13. Sheikh, H.R., Sabir, M.F., Bovik, A.C.: A statistical evaluation of recent full reference image
quality assessment algorithms. IEEE Trans. Image Process. 15(11), 3440–3451 (2006). https://
doi.org/10.1109/TIP.2006.881959
14. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recog-
nition. In: International Conference on Learning Representations (2015)
15. Trinidad, J.C.: Reconstructing obfuscated human faces (2017). http://cs231n.stanford.edu/
reports/2017/pdfs/223.pdf
16. Viola, P., Jones, M.J.: Robust real-time face detection. Int. J. Comput. Vis. 57(2), 137–
154 (2004). https://doi.org/10.1023/B:VISI.0000013087.49260.fb, https://doi.org/10.1023/B:
VISI.0000013087.49260.fb
17. Wang, Z., Bovik, A.C.: Mean squared error: love it or leave it? a new look at signal fidelity
measures. IEEE Signal Process. Mag. 26(1), 98–117 (2009). https://doi.org/10.1109/MSP.
2008.930649
18. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error
visibility to structural similarity. IEEE Trans. Image Process. 13(4), 600–612 (2004). https://
doi.org/10.1109/TIP.2003.819861
A Machine Learning Assistant
for Choosing Operators and Tuning
Their Parameters in Image Processing
Tasks

Issam Qaffou

Abstract Operator choosing and parameter tuning in image processing tasks is

still to be a challenging problem to achieve good results. This paper discusses the
formulation of a solution combining both machine learning and multi-agent system to
help users to choose best operators and attribute appropriate values to their parameters
in image processing applications. The aspect of cooperative learning makes this
solution faster and outperforms one agent learning. The empirical study shows that
our solution is effective on finding the optimal vision operators and their parameters’
values.

Keywords Reinforcement learning · Multi-agent system · Cooperation ·

Coordination · Computer vision · Parameter tuning · Operator selection

1 Introduction

The image processing library is rich in operators. For each type of application
(restoration, segmentation, object recognition, etc.), there are several possible opera-
tors available to users. To run an operator, we must affect values to its parameters. For
instance, to segment a texture by applying the vision operator GLCM (Gray Level
Co-occurrence Matrix) [1], we must assign a value to the sliding window, which is
an operational parameter for GLCM.
Usually, the process followed to accomplish an image processing task, is to choose
an operator, among others, to apply and give values to its parameters in order to
obtain a result. If the user is not satisfied, he gives other values until the result is
satisfactory. If he can’t find a good result with the chosen operator despite trying
all the parameters’ values, he chooses another operator. This process is repeated
until the user finds a satisfactory result. We can conclude that this process is tedious
and time consuming for tasks requiring the application of just one operator. For

I. Qaffou (B)
Faculty of Science Semlalia, LISI Laboratory, Computer Science Department, Cadi Ayyad
University Marrakech, Marrakesh, Morocco
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 339

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_24
340 I. Qaffou

tasks requiring the combination of several operators, the process becomes more
tedious and greedily time-consuming. Good results in image processing depend on
the chosen operator and the values assigned to their parameters. Therefore, making
good choices in terms of operators and parameters’ values is decisive in the field of
image processing. Moreover, it’s a real challenge for users. The absence of efficient
systems or methods that could assist users in their choices, pushes them to intervene
manually and proceed by trial and error. Unfortunately, this intervention is only
intuition and experience based, and it is limited because a user can stop as soon as
he finds a result that satisfies him without being the best. That is, there are other
operators or values not tested.
As an example, to detect boundaries in an image, we may apply one of the two
operators; “Shen-Castan” filter or “Deriche” filter. The parameter to fix for each
operator is α, the width of the filter. If we want to proceed manually to decide which
operator is more appropriate and what is the optimal value to attribute to α, a user
deprived of any knowledge about these filters and their width α, tries different values
until he reaches a satisfactory result. He may not succeed because he doesn’t know
beforehand that α  [0,1]. An expert, meanwhile, will minimize his tests because he
knows that α must be between 0 and 1, and the smaller is the value of α the best the
result will be for “Shen-Castan”, and the opposite for “Deriche”.
Proceeding manually is not an efficient solution. Thus, the need of a strong auto-
matic method that can substitute the user or at least that help him in his choices
becomes insistent.
To implement an automatic vision system that can overcome all the constraints
faced during the process of accomplishing an image processing application is not
easy. Indeed, the variation of images, the diversity of operators and techniques and
also the impact of users’ preferences makes the task of implementing a vision system
more complicated. Although, some works have been done with the attempts to help in
operator choosing and parameter regulating. In [2] we present an automatic method
based on reinforcement learning that tells the user, among several proposed opera-
tors, which is the best combination of operators with their best parameters’ values to
apply for a segmentation task. The proposed method uses one learning agent using
Q-learning algorithm. In [3], authors discuss an adaptive discrete cuckoo algorithm
that helps to choose among a set of parameters’ values those which are optimal.
In [4], B. Nickolay et al. propose a method based on evolutionary algorithms to
adjust parameters in an application of surface inspection. In [5], to generate bright-
ness membership functions for several images, Taylor used a reinforcement learning
framework which uses connectionist systems as approximators to set parameters.
Farhang et al. introduces in [6] a method to segment prostate in transrectal ultra-
sound images. He uses also a reinforcement learning scheme to adjust parameters.
Its strategy is to divide the input image into sub-images and works on each one in
order to reach a good segmentation.
A Machine Learning Assistant for Choosing … 341

Our goal in this paper is to propose a method that also solves the problem of param-
eter optimization and operator choosing but using a system of multi-reinforcement
learning agent. The agents are learning together and cooperating in order to achieve
the best result.
This paper is organized as follows. The general context and its related concepts
are presented in Sect. 2. The proposed solution is detailed in Sect. 3. Experimentation
and discussion are presented in Sect. 4 followed by a conclusion and future work.

2 General Context

To process an image, we need to select a combination of vision operators and fix their
parameter values and then apply it on the image. This process shows that it exists a
set of possible vision operators and different possible values for their parameters.
Our work starts by setting up a system building operators’ combinations for an
image processing application.
First of all, we need to know the type of image to process, the goal to reach; it
means witch kind of processing we want to accomplish and what are the constraints.
All these data will control the process of solution. To do so, a library of possible
operators and their parameters to adjust is needed by our system.
In fact, the choice of operators to apply and the values to assign to their parameters
depends on the type of the input image, and the final decision depends on an evaluation
between the obtained result and a reference represented by a ground truth. Figure 1
summarizes the general context of the proposed approach.

Fig. 1 General context of the proposed solution

342 I. Qaffou

The solution we propose is a new methodology combining both multi-agent

system and reinforcement learning. Our idea is to model the problem using a system
of intelligent agents where each one takes care of part of the overall problem. Next
subsections will present briefly the related concepts.

2.1 Multi-Agent System

Applying just one agent to do tasks may be not efficient and can need much time
to finish. To overcome these problems, we use multi-agent system (MAS) that can
proceed in parallel and cooperate to accomplish a required task. MAS is a group
of autonomous and intelligent agents that are independently connected within an
environment to achieve a common goal. The agents may cooperate or not, share
knowledge between them or not, all it depends on the goal and sometimes users’
preferences.
MAS-based approaches have been mainly used in different fields because of the
beneficial advantages given. One of the important benefits available by using MAS
technology [7] is speed and accuracy of the operation, due to the concept of parallel
and asynchronous computation. Usually, MAS outperforms one agent in the quality
of results and the speed to find them.

2.2 Reinforcement Learning

The basis of reinforcement learning (RL) is an intelligent agent who seeks by trial-
and-error a reward or a special signal, and, on one hand, forces itself to receive that
reward by exploring its environment. On the other hand, it must also exploit the
knowledge acquired through previous actions and feedback [8]. Figure 2 illustrates
the principal components of RL.
The agent, which is the decision maker of the process, tries an action recognized
by the environment, according to which it receives either a reward or a punishment
from its environment. The agent receives also information about the current state of
the environment. It acquires then knowledge of actions which generate rewards and
punishments and ends up by learning to execute those which are the most rewarding.
To be able to proceed by RL, the following concepts must be defined and taken
into account:
States: characterize the situation of the RL agent or its environment at all times.
A Machine Learning Assistant for Choosing … 343

Fig. 2 General interaction between RL component

Actions: each state is associated with a set of actions to be applied at a given time.
An action can take until the next instant.
The return: this is a reinforcement signal produced after each reaction of the agent.
It allows you to assess your choice of actions.
There are several algorithms in RL, such as Q-learning which constitutes the core
of our solution.

2.2.1 Q-learning Algorithm

Q-learning (QL) is proposed by Watkins in 1989 [9]. It is a form of RL algorithm

that can be used on-line, it means that it does not need a model of its environment.
It proceeds by learning an action-value function that gives the expected goal of
applying an action in a given state, and thereafter flowing a determined policy [10].
QL generates a table called Q-table. It contains Q-values representing the maximum
expected rewards of state-action pairs Q(s,a). The goal of the Q-table is to find the
best action at each state. By trying different actions in different states, the agent
learns which is the best by judging it using the long-term updating of the rewards
[11]. The simplest form of Q-learning is given by the following update.

Q(st , at ) ← Q(st , at ) + α[rt+1 + γ max a Q(st+1 , at ) − Q(st , at )]

where γ ∈ [0, 1[ is a discount factor.

344 I. Qaffou

2.3 Multi-Agent Reinforcement Learning

The extension of RL to MAS has aroused the interest of many researchers. The main
limitation is the non-Markovian character of MAS. In this paper, we are concerned
with independent learning agents defined by [12] as agents with access only to their
own actions. The goal is to maximize joint reinforcement.
In the case of independent learning agents, we must make sure that all the agents
will individually choose their action so that the resulting joint action is optimal:
this is the problem of coordination. There are many algorithms in literature which
simulate the behavior or distributed decision making based on multi agent learning.
This paper uses Q-learning hysteretic to model learning within the proposed MAS.

2.3.1 Hysteretic Q-learning.

In an SMA, the reinforcement perceived by an agent depends on the actions chosen

by the group; therefore, an agent carrying out an optimal action can still be punished
because of a bad choice of the group. It is, therefore, preferable for an agent to
give little importance to a punishment received after having chosen an action that
gave it satisfaction in the past. However, the agent should not be completely blind
to sanctions at the risk of remaining in sub-optimal equilibria or of not coordinating
on an optimal joint action. The idea of hysteretic Q-Learning is to use two learning
coefficients according to the result of a joint action. Based on QL, Hysteritic QL uses
two learning rates α et β to increase and decrease the Q-value of each agent. The
update function used in Hysteritic QL for an agent i executing an action ai in a state

s to find itself in a state s is:
 
δ ← r + γ max

Q i s  , a  − Q i (s, ai )
a

Q i (s, ai ) + αδi f δ ≥ 0
Q i (s, ai ) ←
Q i (s, ai ) + βδi f not

where α > β to basically get optimistic learners [13].

3 The Proposed Approach

The agents used in the proposed system cooperate to reach the optimal opera-
tors’ combination with optimal values for their parameters. Hysteretic Q-learning
is adopted to update the Q-values during the learning of the SMA.
We use two types of agent:
A Machine Learning Assistant for Choosing … 345

Fig. 3 Interaction between supervisor agent and learning agent

Learning agents: each agent takes care of an operator to adjust its parameters. These
agents are hysteretic, i.e. they give little importance to the received penalties, but are
not completely blind to them [14].
Supervisor agent: it assigns to each operators a learning agents to find its optimal
parameters’ values, and decides which is the best operators’ combination. Figure 3.
shows the architecture of the two types of agents.

3.1 Formulation of RL Within the MAS

The learning agents proceed by reinforcement learning to find the optimal values of
the parameters. The action space, the state space and the return function of must then
be defined.
The action space is made up of the group’s joint actions. A joint action is a
combination of individual agent actions. Indeed, for an agent i dealing with the
adjustment of the parameters of the operator OPi, an action is any change made to a
value of a parameter.
Individual actions: if an operator has, for example, k parameters ( p1 , . . . , pk ) and
each parameter pi has a set of possible values Vi = {vi1 , . . . , vi p }, an individual
action is then: (u 1 , u 2 , . . . , u k ) ∈ V1 x V2 x . . . x Vk .
The state of the environment is represented by the selected features from a
treatment. It is defined by all the agents.
The reward function is defined as a ratio of comparing the features of the result
image and those of the ground truth. The smaller is the ratio, the maximum the reward
will be.
346 I. Qaffou

4 Empirical Study

To prove the proposed approach, we apply it in an image segmentation task. Such

application needs three phases to be accomplished: pre-processing, processing and
post-processing.
Pre-processing: This phase consists of improving the quality of the image by
applying filters. In this phase, we propose three operators which are predefined in
Matlab by: ‘medfilt2’; ‘ordfilt2’ et ‘wiener2’. Only one parameter to adjust for each
one of these operators: the size of the filter which may take the one of the possible
values: 3 or 5.
Processing: It consists in detecting the contours contained in the image. We propose
only one operator in this phase: the operator ‘edge’ predefined in Matlab. We offer
two parameters for this operator to adjust; the first relates to the filter to choose
(sobel, prewitt, zerocross or log as predefined in Matlab) and the second relates to
the threshold to eliminate the contours of weak contrast [0.02, …, 0.09, 0.1]. Only
the pixels whose value is greater than a given threshold are retained as contours.
Post-processing: In this phase the image is refined by removing small objects. we
offer only one operator for this phase: the operator ‘bwareaopen’ predefined in
MATLAB.
‘bwareaopen’ has two parameters to adjust: the first defines the maximum size (in
pixels) of the objects to be deleted 5, 10, 15 or 20. The second defines connectivity:
the number of neighbors to take into account 4 or 8. For 2D images, connectivity is
by default equal to 8.
After classifying operators per phases, the possible combinations are:
Cop1 = (medfilt2, edge, bwareaopen).
Cop2 = (wiener2, edge, bwareaopen).
Cop3 = (ordfilt2, edge, bwareaopen).
To know which combination gives the best result, we must run the proposed MAS.
The learning agent must find the appropriate parameter values for each operator so
that the result segmentation is closest to the ground truth.

4.1 RL Component Adaptation

Actions: a value of a parameter represents an elementary action for a learning agent.

Thus, a joint action of the MAS is a combination of the values of the parame-
ters. Subsequently, the action space consists of all the possible combinations of the
parameters’ values.
(3, (‘sobel’, 0.02), (5,8)) is an example of a joint action in this study.
States: A state is defined by a set of features extracted from the image. In this
experimentation, a state is defined by three features:
A Machine Learning Assistant for Choosing … 347

S = [x1 , x2 , x3 ]

x1 is the ratio between the number of contours in the result image and the number
of contours in the ground truth;
x2 is the ratio between the number of white pixels in the result image and the
number of white pixels in the ground truth;
x3 is the ratio between the length of the longest contour in the result image and
the length of the longest contour in the ground truth;
Reward: The reward is calculated by an objective evaluation which consists in
comparing between the result image and its ground truth. Because we adopt, in
this experimentation, a contour segmentation approach, we use a weighted sum
of the three error measures which give global indices on the quality of a contour
segmentation result: the over-detection error, the under-detection error and the error
location [15].

D = a1 ∗ D1 + a2 ∗ D2 + a3 ∗ D3

where ai is a weight valuing Di .

And
  car d(Ir ) − car d(Ir ∩ Ir e f )
D1 = E R R sur Ir , Ir e f =
car d(Ir ) − car d(Ir e f )
 
cont
  car d Ir e f /r
D2 = E R R sous Ir , Ir e f =  
car d Ircont
ef
 
  car d Ir e f /r ∪ Ir/r e f
D2 = E R R loc Ir , Ir e f =
car d(Ir )

We applied our approach on fifty images of traffic signs that were used with their
ground truths. Figure 4 shows the results obtained for some random images taken
from the treatment process.
In the learning algorithm the number of episodes and stages, the values of epsilon,
gamma and alpha influence the results. The obtained results are for: α = 0.5, γ =
0.8, ε = 0.5, episodes = 200 and steps = 80.
After execution, the best results are always returned by the same combination
of operators (wiener2, edge, bwareaopen) with, respectively, the best values of the
parameters (5;(prewitt,3.000000e-002);(10,8)).
The curve in Fig. 5 shows that there is learning, and the MAS accumulates
experience, that’s why the number of steps decreases per episode.
348 I. Qaffou

Fig. 4 From left to right: the input images, the ground truths and the results obtained

5 Conclusion

Accomplishing an image processing by testing all the possible cases in terms of

operators choosing and parameter adjusting is almost impossible manually. In this
paper we proposed an automatic method that can help users to find optimal operators
and parameter’s values to apply. The proposed method tests all the possible cases by
proceeding by reinforcement learning and gives good result by using a multi-agent
system which outperform one agent processing. The Obtained results are satisfactory,
and will be compared to other proposed approach in terms of speed and accuracy.
A Machine Learning Assistant for Choosing … 349

Fig. 5 Learning curve

References

1. Haralick, R.M., Shanmugan, K., Dinstein, I.: Textural featuresfor image classification. IEEE
Trans. Sys. Man Cyber. 3(6), 610−621 (1973)
2. Qaffou, I., Sadgal, M., Elfazziki,A.: A multi-agents architecture to learn vision operators and
their parameters. Int. J. Comput. Sci. Iss. (IJCSI) 9(3), 140 (May 2012)
3. Benchikhi, L., et al.: A novel adaptive discrete cuckoo search algorithm for parameter
optimization in computer vision. Intel. Art 20(60), 51–71 (2017)
4. Nickolay, B., Schneider, B., Jacob,S.: Parameter optimization of an image processing system
using evolutionary algorithms. CAIP, pp. 637–644 (1997)
5. Taylor, G.W.: Areinforcement learning framework for parameter control in computer vision
applications. In: Proceedings of the First Canadian Conference on Computer and Robot Vision
(CRV’04), IEEE (2004)
6. Sahba, F., Tizhoosh, H.R., Salama, M.: Application of reinforcement learning for segmentation
of transrectal ultrasound images. BMC Medical Imaging (2008)
7. Nikos Vlassis.: A concise introduction to multiagent systems and distributed artificial intel-
ligence. Synthesis Lectures on Artificial Intelligence and Machine Learning, 1st edition
(2007)
8. Sutton, R.S., Barto A.G.: Reinforcement learning: an introduction. Adaptive computation and
machine learning. MIT Press, Cambridge, MA (1998)
9. Watkins, C.J.C.H.: Learning from delayed rewards. PhD thesis, Cambridge University (1989)
10. Manju, S., Punithavalli, M.: An analysis of QLearning algorithms with strategies of reward
functions. Int. J. Comput. Sci. Eng 3(2), 814–820 (2011)
11. Watkins, C.J.C.H., Dayan, P.: Q-learning. Mach. Learn 8, 279–292 (1992)
350 I. Qaffou

12. Claus, C., Boutilier, C.: The dynamics of reinforcement learning in cooperative multiagent
systems. In: Proceedings of the Fifteenth National Conference on Artificial Intelligence, p. 746–
752 (1998)
13. Matignon, L., Laurent, G.J., Le Fort-Piat, N.: Hysteretic q-learning: an algorithm for decentral-
ized reinforcement learning in cooperative multi-agent teams. In: Proc. of the IEEE Int. Conf.
on Intelligent Robots and Systems, pp. 64–69, San Diego, CA, USA (2007)
14. Swets, J.A.: Signal detection theory and roc analysis in psychology and diagnostics. Lawrence
Erlbaum Associates, Mahwah, NJ (1996)
15. Chabrier, S., Laurent, H., Rosenberger, C., Zhang,Y.J.: Supervised evaluation of synthetic and
real contour segmentation results. European Signal Processing Conference (EUSIPCO) (2006)
Convergence and Parameters Setting
of Continuous Hopfield Neural Networks
Applied to Image Restoration Problem

Joudar Nour-eddine, Ramchoun Hassan, Zakariae En-Naimani,

and Ettaouil Mohamed

Abstract In image processing, Hopfield neural networks (HNNs) have been suc-
cessively invested to solve the optimization model of image restoration. However,
in the discrete case, HNNs lead to the fluctuation behaviors due to the limitation
of the hard limit activators. The continuous model of HNNs generates non feasible
solutions owing to the trial-and-error setting values process of the model parameters.
To deal with these issues effectively, we suggest a new energy function with appro-
priate parameters settled by the stability analysis of the network. Performance of our
method is demonstrated numerically and visually by several computational tests.

Keywords Image restoration · Optimization · Artificial neural network ·

Continuous Hopfield neural network · Penalty function

1 Introduction

Image Restoration Problem (IRP) has started since the 50s after many studies carried
out to improve the quality of the images received in the space program [1]. Restoration
techniques are applied to remove:

J. Nour-eddine
Higher Normal School of Technical Educating, University Mohamed V Rabat, Rabat, Morocco
e-mail: [email protected]
R. Hassan (B)
Laboratory of Mathematical Modeling, Simulation and Smart Systems,
University of Moulay Ismail, Meknes, Morocco
e-mail: [email protected]
Z. En-Naimani
National School of Applied Sciences, University Mohammed First Oujda, Oujda, Morocco
e-mail: [email protected]
E. Mohamed
Faculty of Science and Technology, Department of Mathematics,
University Sidi Mohamed Ben Abdellah Fez, Fes, Morocco
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 351

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_25
352 J. Nour-eddine et al.

– System degradations, such as blur owing to optical system aberrations, atmospheric

turbulence, motion and diffraction.
– The electronic noise, which affects the measurement.

Several methods have been proposed in the literature to solve IRP, and they have
proved to a great extent their effectiveness in image restoration while preserving
as possible the image details. The filtering methods are one of the most effective
approaches that have been used widely to enhance the image quality. These tech-
niques include the Median filter [2, 3], Wiener filter [4], inverse and Kalaman filters
[5]. However, the Median filter allows a good noise suppression, but it deteriorates
the image features. Wiener filter is implemented only after the wide sens stationary
assumption has been made for images. Inverse filter generates a good restoration
when the image is not affected by noise. However, in noisy case, it produces bad
effects. Kalaman filter can deal with non-stationary images, but it is computationally
very intensive
Recently, a novel image processing technology based on variational methods is
motivating many mathematicians and image scholars to research on this challenging
topic [6, 7]. This technique has investigated deeply the gradient magnitude properties
of the image for deblurring and denoising [8].
Among the variational image restoration models, the Total Variational (TV)
model, proposed by Rudin and Fatemi, has been the most important and has been
properly applied in several applications [6]. Other extensions of TV technique have
been suggested using l2 , l1 norms to restore degraded images as proposed in [7, 8].
Image restoration problem has been also solved by Partial Differential Equations
(PDE) [9]. Motivated by the isotropic and anisotropic properties described by the
Heat and Perona equations, different studies have deeply investigated these PDEs to
rebuild the degraded images [9–11]. However, these equations of second order are
very weak since they tend to cause the processed image to exhibit staircase effects.
To avoid this drawback, recent studies have been proposed based on Four order PDE
(FOPDE), and they have proved their performance because of the high smoothing
ability [12]. The optimization filed has also contributed to image restoration using
the modelling techniques and resolution. For example, we may cite those that are
inspired by the biological behaviours like genetic algorithms [13], and artificial neural
networks ANNs [14–17].
The first use of ANNs in image restoration was due to Zhou and Chellapa [14].
Indeed, they have been the first who proposed to solve the combinatorial problem
of IRP by Hopfield neural network HNN [18, 19], in which they have mapped IRP
into the discrete HNN by comparing the Lyaponuv function with the energy function
associated with image restoration model. In spite of the efficiency of this model,
it still suffering from the instability during the neurons updates. To overcome this
problem, Paik and Katssagelos have proposed a Modified HNN model, in which they
suggested several improved algorithms using sequential, decoupled n-simultaneous
and partially asynchronous modes of updates [15]. In [16], authors have reduced the
number of iterations by feeding the neurons outputs only to higher-order neurons.
However, these methods are based on the discrete activation function, which often
Convergence and Parameters Setting of Continuous Hopfield Neural Networks … 353

limits the search space because of the fluctuation problem [23]. To avoid this draw-
back, we suggest in this chapter to use the continuous Hopfield neural network (CHN)
for image restoration problem. In the new process of resolution named (CHN-IRP),
we start by defining a new energy function associated with IRP utilizing the penalty
method. Next, we map IRP into the CHN, thus extracting the weights and the bias of
the network. Thereafter, we discuss the parameters of the energy function ensuring,
on one hand, the convergence of the feasible solutions, and the instability of interior
points on the other hand.
The paper is organized as follows. In the next section, we briefly describe the
Hofield neural networks including the discrete and the continuous models. Image
restoration problem and its resolution by discrete Hopfield neural network are dis-
cussed in Sect. 3. In Sect. 4, the continuous Hopfield neural network is used for solving
IRP and supported by suitable parameters extracted from the stability analysis of the
CHN. Section 5 presents experimental results of CHN-IRP. Finally, conclusions are
given in Sect. 6.

2 Hopfield Neural Networks

Hopfield neural network (HNN) was invented by Dr. John Hopfield in 1982 [18]. It
consists of a single layer that contains one or more fully connected recurrent neurons.
The Hopfield networks are commonly used for auto-association and optimization
tasks [17–19]. Hopfield has proposed two basic models: discrete and continuous
HNNs, the difference between the two models refers to the nature of the state variables
and time in these networks.

2.1 Discrete HNNs

In the discrete HNN, the output of each neuron is calculated through the following
formula: ⎛ ⎞

n
X i (t + 1) = sign ⎝ Ti j x j (t)⎠ (1)
j=1

Thus, the vector X is given as:

X = sign(T X − I ) (2)

where X , W , T and the sign function are:

354 J. Nour-eddine et al.
⎡ ⎤
x1
⎢ x2 ⎥
⎢ ⎥
– X is the activation vector of the n neurons: X = ⎢ . ⎥
⎣ .. ⎦
x
⎡ ⎤n
T11 T12 T13 . . . T1n
⎢ T21 T22 T23 . . . T2n ⎥
⎢ ⎥
– T is the weight matrix: T = ⎢ . .. .. . . .. ⎥
⎣ .. . . . . ⎦
Td1 wd2 Td3 . . . Tdn
– The sign function is defined as:

+1 if x > 0
sign(x) = (3)
−1 otherwise

As consequence, the Hopfield neural network can be represented by fully intercon-

nected network with undirected weights.

2.2 Continuous Hopfield Neural Networks

In 1982, Hopfield and Tank have proposed a neural network called Continuous Hop-
field neural network for the purpose of solving a wide variety of combinatorial
problems and realizing a self-associative memories [20–22].
The Continuous Hopfield neural network represents a strong extension of the
DHNN that is governed by the following differential equation [22]:

du u
= − + Tv + I (4)
dt τ
where u, v and I are, respectively, vectors of neuron states, outputs and biases. τ is
a time constant and the output function vi = g(u i ) is an hyperbolic tangent defined
by:   
1 ui
g(u i ) = 1 + tanh u 0 > 0, ∀i = 1, . . . , n (5)
2 u0

where u 0 denotes an input vector and the matrix T represents the weights of the
synaptic connections between neurons.
The evolution of this network towards an equilibrium point is equivalent to the
minimization of the Lyaponuv function [21]. This point is defined as follows:
Definition 1 If a vector ve satisfies v(t) = ve , ∀t ≥ t e for some t e ≥ 0, then, ve is
called an equilibrium point of the system (4).
Solving combinatorial problem via CHN requires the building of an energy function
associated with the problem. Indeed, the solution of the proposed energy function is
expressed by the neurones states of the equilibrium point.
Convergence and Parameters Setting of Continuous Hopfield Neural Networks … 355

For this purpose, Hopfield has introduced the energy function E on [0, 1]n , which
is defined by:
n 
1 1  vi −1
E(v) = − vt T v − (I )t v + g (v)dv (6)
2 τ i=1 0

when τ → ∞, the Lyaponuv function is associated with the cost function to be

minimized in the combinatorial problem. Thus, CHN will solve problems with the
following form:
1
E(v) = − vt T v − (I )t v (7)
2

If H = {v ∈ [0, 1]n } denotes an Hypercube of I R n , then, a point v ∈ H is an equi-

librium point for the C H N if, and only if the three following conditions are satisfied
[21]:
∂ E(v)
E i (v) = ≥ 0 ∀i = 1, . . . , n such that vi = 0 (8)
∂vi

∂ E(v)
E i (v) = ≤ 0 ∀i = 1, . . . , n such that vi = 1 (9)
∂vi

∂ E(v)
E i (v) = = 0 ∀i = 1, . . . , n such that vi ∈ ]0, 1[ (10)
∂vi

The conditions (8) and (9) can be dominated by the following equations:

E 0 (v) ≥ 0 (11)
−

−1
E (v) ≤ 0 (12)

−1
where E 0 (v) = min{i/vi =0} E i (v) and E (v) = max{i/vi =1} E i (v).
−

3 Image Restoration Problem

3.1 Optimization Model Associated with IRP

Practically, the image degradation can be, adequately, modelled by a linear blur
(motion, defocusing, atmospheric turbulence …) and an additive white Gaussian
noise. In this sense, the degradation model is given by:

y =h∗x +η (13)
356 J. Nour-eddine et al.

where x and y denote respectively the original and the degraded images, h is the blur
operator where its form is extracted from the point spread function (PSF). η is an
additive noise and ∗ is the convolution operator.
The restoration model is expressed as follows [9]:
 
min J (x) = |y − h ∗ x|2 dt + λ |∇x|2 dt (14)
x∈Ω Ω Ω

In the discrete variable, the problem (14) is expressed by:

1 1
min J (X ) = ||Y − H X ||2 + λ||D X ||2 (15)
x∈[|0,...,255|] N 2 2 2

where X , Y are the vectors of the original and degraded images. H is the Blur matrix.
In the case of the shift invariant blur, the matrix H is represented by a Toeplitz matrix
with a specific form depending on the choice of the boundary condition (BCs) used
in the problem [23]. The added term ||D X || is a high pass filter known by the
regularization term of Tikhonov. λ is a regularization parameter that measures the
trade off between the closeness to data and the prior knowledge of the solution. ||.||
is the euclidean norm.
Obviously Eq. (15) can be expressed as:
⎛ ⎞2
 
2 2
N N
Min J (X ) = ⎝Y p − Hp,i X i ⎠
X ∈[|0,...,255|] N 2 p=1 i=1
⎛ 2 ⎞2
1  ⎝
2
N N
+ λ D pi X i ⎠ (16)
2 p=1 i=1

In the next section, we aim to explain the details of the resolution of IRP by the
discrete Hopfield neural network.

3.2 Using Discrete Hopfield Neural Network for Solving IRP

In this part, we explain the principle steps that have been used to solve image restora-
tion problem based on discrete Hopfield network [14].
Let X be the Gray level function of an image of size N × N , and G its maximal
value. First, we transform X into a vector using the following formula:

X m = X (i−1)N + j (17)
Convergence and Parameters Setting of Continuous Hopfield Neural Networks … 357

In the second step, we represent each element of the image X by a simple sum scheme
(18), which is given by [24]:
G
Xi = vik (18)
k=1

where:
⎛ 2 ⎞ 
 N  G
1 if u ik ≥ 0
vik = f (u ik ) = f ⎝ Tik, jl v jl + Iik ⎠ = (19)
j=1 l=1
0 if u ik < 0

The function f (t) denotes the Heavside activation function.

This coding (simple sum) is a one-to-many mapping from Z+ to V , and the
numbers have degenerate representations. This scheme requires N bits to express a
number N and is not economical in the number of bits or neurons. However, it is
highly fault-tolerant because an error in a single bit does not cause a large error in
the number represented.
In the next step, we replace the data X in Eq. (16). Thus, we obtain:
⎛  ⎞2
  
2 2
N N G
Min E(v) = ⎝Y p − Hp,i vik ⎠
v∈{0,1} N 2 ×G p=1 i=1 k=1
⎛ ⎞2 
1  ⎝ 
N2 N2 G
+ λ D pi vik ⎠ (20)
2 p=1 i=1 k=1

which equals to:

1 
N2 N2 N2 G G
Min E(v) = Hp,i Hp, j vik v jl
2
v∈{0,1} N ×G 2 p=1 i=1 j=1 k=1 l=1

1 
2 2 2
N N N G G
+ λ D p,i D p, j vik v jl
2 p=1 i=1 j=1 k=1 l=1


N 
N 
1 2
2 2
G N2
− y p Hp,i vik + y (21)
p=1 i=1 k=1
2 p=1 p

Afterwards, we map the model (21) in a DHN by proposing a network with N 2 × G

neurons mutually interconnected. The suggested network is characterized by:
– The set of the network states:
v = {vi,k / i = 1, . . . , N 2 and k = 1, . . . , G}.
358 J. Nour-eddine et al.

– The weight matrix:

T = {Tik, jl / i, j = 1, . . . , N 2 and k, l = 1, . . . , G}
Using the identification between Eq. (21) and Lyapunov function (7), we obtain
 N 2 N 2
jl = − p=1 Hp,i Hp, j − λ
DH N
Tik, p=1 D p,i D p, j
Map = N 2 (22)
DH N
Iik = p=1 y p Hp,i

Two remarks that must be mentioned, the first is that the weights are independent
of the subscripts (k, l), which reduces the complexity by neglecting the repeated
terms. The second remark concerns the self-connections Tik,ik , which are not equal
to zero, and then, it contradicts the convergence criterion of Hopfield [19]. To solve
this problem, Zhou and al have proposed a decision rule, which depends on the
variations of the energy function ΔE and the neurons states vik . It is described as
follow [14]:
new
vik if ΔE due to state change Δvik is less than zero
vik = old (23)
vik otherwise

where ΔE = E new − E old and Δvik = vik

new
− vik
old
.
Image restoration using DHN is summarized in the following steps:
1. Set the initial state of the neurons.
2. Update the state of all neurons randomly and asynchronously according to the
decision rule (23).
3. Check the energy function, if energy does not change continue, otherwise go
back to step 2.
4. Construct the image using Eq. (17).
Although the D H N deals successively with image restoration problem, it suffers
from the fluctuation problem. In other words, the network state keeps wandering
around the state space near the minima of the energy function, thus generating an
oscillatory behaviour. To overcome this drawback, we propose to use the continuous
model of Hopfield neural networks. This is the subject of the next section.

4 Using Continuous Hopfield Neural Network for Solving

IRP

4.1 Mapping IRP into CHN

The main purpose of this section is to use the CHN with the aim of solving image
restoration problem. To this end, we define an appropriate energy function which
takes into a count the IRP particularities. To be precise, the choice of the parameters
Convergence and Parameters Setting of Continuous Hopfield Neural Networks … 359

of this function must ensure the feasibility of CHN equilibrium points. Based on
the energy function defined in Eq. (21), we add a new constraint that describes the
feasibility of the neurons states. The constructed model is:
⎧
⎪
⎪
min E(v)
⎪
⎪
⎨ subject to
(P) =  N 2 G (24)
⎪
⎪
⎪
⎪
⎩ i=1 k=1 (vik (1 − vik )) = 0
v ∈ [0, 1] N ×G
2

The role of such constraint is to force neurons states to take 0 or 1.

Next, we use the penalty function approach to build a new energy function asso-
ciated with (P). Thus, the energy function E is transformed to:

1 
N2 N2 N2 G G
Min E(v) = α Hp,i Hp, j vik v jl
v∈[0,1] N 2 ×G 2 p=1 i=1 j=1 k=1 l=1

1 
2 2 2
N N N G G
+ λ D p,i D p, j vik v jl
2 p=1 i=1 j=1 k=1 l=1


N 
N 
2 2
G
−α y p Hp,i vik
p=1 i=1 k=1

1  2 1 
N2 N2 G
+ α y + β (vik (1 − vik )) (25)
2 p=1 p 2 i=1 k=1

where the scaling parameters α, β, and λ are real parameters that we will set by the
network convergence analysis.
Comparing this equation with Lyaponuv function described by Eq. (7), we obtain
the following parameters:
 N 2 N 2
jl = −α p=1 Hp,i Hp, j − λ p=1 D p,i D p, j + β
CHN
Tik,
 N2 (26)
IikC H N = α p=1 y p Hp,i − 2β

4.2 Convergence Analysis and Parameters Setting

Convergence of CHN is an important task in the resolution of combinatorial prob-

lems. Indeed, in many cases, the adopted network converges toward an equilibrium
point which is not a feasible solution. Generally, the stability of CHN is analysed
over the three sets: feasible set HF , Hypercube infeasible corner set HC − HF and
interior points set H − HC (Fig. 1). In our case, any point of HC makes a feasible
360 J. Nour-eddine et al.

Fig. 1 Feasible and

infeasible sets

solution which means that HF and HC are identical and thereafter the stability anal-
ysis will be made only for HC and H − HC . Thus, we aim to ensure the following
objectives:
– Convergence of feasible solutions to an equilibrium point.
– Instability of interior points.

4.2.1 Convergence of Feasible Solutions v ∈ HF

For each feasible solution, the stability of the corner set is ensured from the equilib-
rium point criteria (11), (12).
First, to minimize the first term of Eq. (25), we impose the following condition:

α≥0 (27)

In the next conditions, we will focus on equilibrium points to achieve the stability
conditions. Thus, we have:
Theorem 1 If α, β and λ are three parameters verifying the following assertions:

α, λ ≥ 0
(S1) = (28)
β ≥ G × N 4 (α M1 + λM2 )

with M1 = max∀( p,i)∈{1,...,N }2 Hp,i and M2 = max∀( p,i)∈{1,...,N }2 D p,i .

Then, any point v is an equilibrium point.

Proof As stated above, the equilibrium point is ensured from the partial derivative
of the energy function. In this regard, we calculate the derivative on vik as follow:

1 
N2 N2 G
∂E
E ik (v) = = α Hp,i Hp, j v jl
∂vik 2 p=1 j=1 l=1

1  
2 2 2
N N G N
1
+ λ D p,i D p, j v jl − α y p Hp,i + β((1 − 2vik )) (29)
2 p=1 j=1 l=1 p=1
2
Convergence and Parameters Setting of Continuous Hopfield Neural Networks … 361

taking into account that the terms of H and D are positive, we obtain the following
assertion:
N2
min E i (v) ≥ 0 =⇒ β − 2α y p Hp,i ≥ 0
{i/vi =0}
p=1

which means that:

min E i (v) ≥ 0 =⇒ β ≥ 2α N 2 G M1 (30)
{i/vi =0}

On the other side, to ensure the negativity of max{i/vi =1} E i (v), we obtain from (29):


N 
N 2 2
G
max E i (v) ≤ 0 =⇒ α Hp,i Hp, j
{i/vi =1}
p=1 j=1 l=1


N 
N 
2 2
G
+λ D p,i D p, j − β ≤ 0 (31)
p=1 j=1 l=1

By maximizing, we obtain the following result:

min E i (v) ≤ 0 =⇒ −β + G N 4 (α M12 + λM22 ) ≤ 0 (32)

{i/vi =0}

Finally, it is obvious to see that 2αG N 2 ≤ αG N 4 . Consequently, the condition (30)

is covered by the equation (32).

4.2.2 Non Convergence of Infeasible Solutions v ∈ H − HF

Before going any further, we begin by giving the following definitions:

Definition 2 A point v ∈ H is said an interior point if, and only if:

∃vi,k ∈ v/vi,k ∈]0, 1[

Definition 3 Let v ∈ H be an interior point, v defines an edge point if and only if:

car d{(i, j) ∈ {1, . . . , n + 1} × {1, . . . , Nmax }/vi,k ∈ ]0, 1[} = 1

Definition 4 Let v ∈ H be an edge point such that va,b ∈ ]0, 1[, v is called the
adjacent corner of v if:
362 J. Nour-eddine et al.

vi, j if (i, j) = (a, b)
vi, j = (33)
1 or 0 if (i, j) = (a, b)

As mentioned in [25], interiors points not being edge points have a null probability
of being the convergence point of the CHN.
Any equilibrium interior point v ∈/ HC cannot be a local minimum and it is nec-
essary to analyze the second derivatives of the generalized energy function (25):

 
2 2
N N
∂2 E
= −Tik,ik = α H 2
p,i + λ D 2p,i − β (34)
∂vik 2 p=1 p=1

In this regard, non-convergence of any edge point is ensured through the following
Theorem [21]:
Theorem 2 Let v ∈ H − HC be an edge point such that vi,k ∈ ]0, 1[, If E i, j (v )
verifies: 
∈
/ [0, −Ti j,i j ] ∀ (i, j)/vi, j = 1
E i, j (v ) (35)
/ [Ti j,i j , 0] ∀ (i, j)/vi, j = 0
∈

Then, the stability of v is avoided.

Practically, these conditions are replaced by:
⎧
⎨ E 0 (v) ≤ Tik,ik
−
1 (36)
⎩−
E (v) ≥ −Tik,ik

Fortunately, these conditions are verified as shown below:


2
N
E (v) = −2α N G M1 + β ≤ Tik,ik = −α
0 2 2
Hp,i
−
p=1


N2
−λ D 2p,i + β (37)
p=1


N 
N 
2 2
G
−1
E (v) = α Hp,i Hp, j
p=1 j=1 l=1


N 
N 
2 2
G
+λ D p,i D p, j − β
p=1 j=1 l=1

 
2 2
N N
≥ Tik,ik = +α 2
Hp,i +λ D 2p,i − β (38)
p=1 p=1
Convergence and Parameters Setting of Continuous Hopfield Neural Networks … 363

In summary, solving the problem of image restoration by CHN is described in the

following algorithm:

Data:
• Degraded image Y of size N 2 ;
• The matrix H and D;
• Initialization of itermax;
Result: Restored image X
• Encode the image X using Eq. (18);
• Map IRP into CHN using Eq. (25);
• Extract weights matrix T and bias I from Eq. (26);
• Determine α, β and λ using Theorem 3.1;
• Initialize a random v0 ;
• Initialize the counter iter ← 1.;
• E old1 ← 0, E new = E old1 + 1;
while |E new − E old1 | > or iter ≤ iter max do
E old1 ← E new ;
for Each neuron vi,k do
E old2 ← 0;
while |E old2 − E new | ≥ do
E old2 ← E new ;
for each neuron, update the state using (19);
E new ← E(v)
end
iter ← iter + 1;
end
end
Algorithm 1: Image restoration using the continuous Hopfield network
(CHN-IRP)

5 Experiments

In this section, we present some numerical results to illustrate the performance of the
proposed method. The test images are shown in Fig. 2 with sizes of 256 × 256. All
experiments are carried out on Windows 7 32-bit and Matlab 9 running on a desktop
equipped with an Intel Core i2 CPU 2.20 GHz and 3 GB of RAM.
In most cases, restoration methods provide results which are almost similar and
this makes some difficulties to distinguish visually between the restored images,
therefore, the judgment becomes subjective. For this reason, we propose the known
measure PSNR (Peak to Signal Noise Ratio) to identify and to compare the effects of
image enhancement algorithms on image quality. The PSNR is defined as follows:
364 J. Nour-eddine et al.

Fig. 2 Test images. a IRM: spine. b House. c Boat

2552
P S N R = 10 log10 , (39)
MSE
MSE: is the Mean Square Error, which is defined as:
M N
i=1 j=1 (X 1 (i, j) − X 2 (i, j))2
MSE = , (40)
M×N

where X 1 and X 2 denote, respectively, the restored and the original images.
Convergence and Parameters Setting of Continuous Hopfield Neural Networks … 365

The performance of the proposed model is compared against some of the most
popular image filters, such as Median, wiener, FOPDE and Total variational method
TV. In order to fairly compare our method with these methods, the next Tables 1
and 2 show the PSNR of the restored images using these filters. The comparisons
Methods were applied on different noise densities = 0.01 to = 0.29 with increments
of 0.04. The parameters of the CHN are calculated based on Eq. (28).
As shown in these tables, denoising filters such as wiener and median filters
have provided a better PSNR in comparison with the other analytical methods that
have generated a lower PSNR because of the original features changes owing to the
isotropy phenomena. In addition, theses results reports that our algorithm produces
for all noise densities a good PSNR than those obtained by Median, wiener, FOPDE
and TV. Our results are supported by graphic result of the CHN-IRP applied to
“House” image. In fact, in all the noise densities, the curve of our methods (black
colour) outperforms the others curves with a considerable exceed (Fig. 3).
In the next experiment, we compare the efficiency of our model with the others
HNN-based filters including MHNN [15] and MRHNN [16]. For this purpose, we
use “Lena” (Fig. 4a) test image that is degraded by a shift invariant blur and 30% of
impulse noise (Fig. 4b).

Table 1 PSNR values of CHN-IRP and comparison filters for “IRM spine” test image at different
Gaussian noise density
Methods 0.01 0.05 0.09 0.13 0.17 0.21 0.25 0.29
Noisy 20.09 19.12 17.43 15.72 14.13 12.72 11.47 10.37
Median 29.36 24.80 20.46 17.52 15.29 13.55 12.05 10.78
FOPDE 28.83 24.29 20.27 17.42 15.22 13.48 12.01 10.78
Wiener 26.97 23.64 20.00 17.29 15.15 13.44 11.99 10.76
TV 7.40 7.40 7.41 7.41 7.41 7.42 7.42 7.42
CHN- 32.47 27.13 24.33 22.13 19.23 18.81 15.33 13.73
IRP

Table 2 PSNR values of CHN-IRP and comparison filters for “Boat” test image at different Gaus-
sian noise density
Methods 0.01 0.05 0.09 0.13 0.17 0.21 0.25 0.29
Noisy 20.02 19.09 17.48 15.77 14.21 12.81 11.65 10.61
Median 24.89 22.59 19.58 17.03 15.02 13.34 11.95 10.74
FOPDE 23.55 21.69 19.07 16.74 14.84 13.23 11.93 10.80
Wiener 26.16 23.21 19.85 17.20 15.15 13.45 12.09 10.92
TV 5.44 5.44 5.44 5.44 5.45 5.45 5.45 5.45
CHN- 31.51 26.73 24.05 21.87 19.08 18.31 15.52 14.23
IRP
366 J. Nour-eddine et al.

Fig. 3 PSNR-curves of CHN-IRP and comparison filters for “House” test image at different Gaus-
sian noise density

In Fig. 4, MHNN has restored the image with a small improvement at the deblur-
ring scale contrary to the denoising process where the restoration does not give a
much (Fig. 4c). Restoration by MRHNN outperforms MHNN as shown in, it has
generated an enhanced image in term of denoising and deblurring while preserving
as much as possible the image quality (Fig. 4d). In Fig. 4e, Our method has gener-
ated a good restored image which is close to the original image. Features are also
protected in term of discontinuities and contrast.
In the Fig. 5, we show the performance of our model applied to “Spine” image,
as well as its comparison with other methods from the literature.
As shown, the proposed CHN-IRP provides an encouraging outcome against the
others methods. Indeed, CHN-IRP allows a good restoration thanks to its ability to
protect the image information especially with this kind pictures. For more illustration,
in Fig. 5g, some details like the position of the heart is clear and can be easily marked
on the right of the picture which is not possible for the other restoration techniques.

6 Conclusion

In order to guarantee the feasibility of CHN in solving image restoration problem,

the stability conditions of CHN defined on a unit hypercube have been proved in this
paper. Using the penalty function, a suitable energy function associated with IRP has
been built. Afterwards, the equilibrium conditions have been used to form a system
of parameters that makes each obtained state a feasible solution of IRP. In addition,
the spurious solutions have been avoided by discussing the energy variations over
the set of interiors points. The efficiency of the proposed algorithm has been proved
through several numerical and visual tests.
Convergence and Parameters Setting of Continuous Hopfield Neural Networks … 367

Fig. 4 Comparison of “Lena” outcomes using the different methods. a Original. b Degraded by
blur and impulse noise ρ = 0.3. c MHNN. d RMHNN. e CHN-IRP
368 J. Nour-eddine et al.

Fig. 5 Outputs of “Spine”

image using the different
methods. a Original. b Noisy
with σ = 0.3. c Median.
d FOPDE. e wiener. f Tv.
g CHN-IRP
Convergence and Parameters Setting of Continuous Hopfield Neural Networks … 369

References

1. Alvarez, L., Guichard, F., Lions, P.-L., Morel, J.-M.: Axiomatisation et nouveaux opérateurs
de la morphologie mathématique. Comptes rendus de l’Académie des sciences. Série 1, Math-
ématique 315(3), 265–268 (1992)
2. Astola, J., Kuosmanen, P.: Fundamentals of Nonlinear Digital Filtering. CRC Press (1997)
3. Joudar, N.-E., Harchli, F., Abdelatif, E.-S., Ettaouil, M.: New adaptive switching scheme for
impulse noise removal: modelling and resolution by genetic optimization. Int. J. Signal Imaging
Syst. Eng. 10(6), 316–323 (2017)
4. Ozkan, M.K., Erdem, A.T., Sezan, M.I., Tekalp, A.M.: Efficient multiframe Wiener restoration
of blurred and noisy image sequences. IEEE Trans. Image Process. 1(4), 453–476 (1992)
5. Wu, W.-R., Kundu, A.: Image estimation using fast modified reduced update Kalman filter.
IEEE Trans. Signal Process. 40(4), 915–926 (1992)
6. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms.
Phys. D Nonlinear Phenom. 60(1–4), 259–268 (1992)
7. Chambolle, A., Caselles, V., Cremers, D., Novaga, M., Pock, T.: An introduction to total
variation for image analysis. Theor. Found. Numer. Methods Sparse Recover. 9, 263–340
(2010)
8. Dong, F.F., Liu, Z.: A new gradient fidelity term for avoiding staircasing effect. J. Comput. Sci.
Technol. 24(6) (2009)
9. Aubert, G., Kornprobst, P.: Mathematical Problems in Image Processing: Partial Differential
Equations and the Calculus of Variations. Springer Science & Business Media (2006)
10. Koenderink, J.J.: The structure of images. Biol. Cybern. 50(5), 363–370 (1984)
11. Perona, P., Malik, J.: Scale-space and edge detection using anisotropic diffusion. IEEE Trans.
Pattern Anal. Mach. Intell. 12(7), 629–639 (1990)
12. You, Y.-L., Kaveh, M.: Fourth order partial differential equations for noise removal. IEEE
Trans. Image Process. 9(10), 1723–1730 (2000)
13. Sakthidasan, K., Nagappan, N.V.: Noise free image restoration using hybrid filter with adaptive
genetic algorithm. Comput. Electr. Eng. 54(2016), 382–392 (2016)
14. Zhou, Y., Chellapa, R.: Image restoration using a neural network. IEEE Trans. Acoustics Speech
Signal Process. 36(7), 1141–1151 (1988)
15. Paik, J.K., Katsaggelos, A.K.: Image restoration using a modified Hopfield network. IEEE
Trans. Image Process. 1(1), 49–63 (1992)
16. Uma, S., Annadurai, S.: Image restoration using modified recurrent Hopfield neural network.
Int. J. Signal Imaging Syst. Eng. 1(3–4), 264–272 (2008)
17. Joudar, N., Ettaouil, M.: Circuits Syst. Signal Process. (2018). https://doi.org/10.1007/s00034-
018-0950-1
18. Hopfield, J.J.: Neural networks and physical systems with emergent collective computational
abilities. Proc. Natl. Acad. Sci. 79(8), 2554–2558 (1982)
19. Hopfield, J.J., Tank, D.W.: Neural computation of decisions in optimization problems. Biol.
Cybern. 52(3), 141–152 (1985)
20. Wen, U.-P., Lan, K.-M., Shih, H.-S.: A review of Hopfield neural networks for solving mathe-
matical programming problems. Eur. J. Oper. Res. 198(3), 675–687 (2009)
21. Talaván, P.M., Yáñez, J.: Parameter setting of the Hopfield network applied to TSP. Neural
Netw. 15(3), 363–373 (2002)
22. Talaván, P.M., Yáñez, J.: The generalized quadratic knapsack problem. A neuronal network
approach. Neural Netw. 19(4), 416–428 (2006)
23. Joudar, N., El Moutouakil, K., Ettaouil, M.: An original continuous Hopfield network for
optimal images restoration. WSEAS Trans. Comput. 14(2015), 668–678 (2015)
24. Takeda, M., Goodman, J.W.: Neural networks for computation: number representations and
programming complexity. Appl. Optics 25(18), 3033–3046 (1986)
25. Park, S.: Signal space interpretations of Hopfield neural network for optimization. In: IEEE
International Symposium on Circuits and Systems, pp. 2181–2184 (1989)
The IBN BATTOUTA Air Traffic Control
Corpus with Real Life ADS-B
and METAR Data

Kasttet Mohammed Saïd and Lyhyaoui Abdelouahid

Abstract The Ibn Battouta ATC (Air Traffic Control) communication corpus is
English corpus characterized with a strong Moroccan accent speech, under the
creation and concentrated on Tangier’s airport1. It synchronizes the voice, the ADS-
B (flight data broadcast by aircraft) and the METAR (Weather Report on Airport), as
trusted structured data frequently pronounced during the communication. The aim
of this work is to update the acoustic model and train the ASR (Automatic Speech
Recognition) engine like CMU Sphinx or IBM Kaldi for the purpose of improving
ATC capabilities, facilitating decision making and ensuring security. Currently we
have recorded more than five hours long audio files with silence elimination of
real life communication between pilots and controllers during takeoffs, approaches
and landings in Tangier airport, Gibraltar airport and in the way from the Moroccan
airspace to the Spanish one and vice versa. In addition to 10logging files from ADS-B
flight data and METAR reports. All audio files have been fully transcribed manually,
with time marking to indicate the start of transmission and its duration, by airlines
pilots and controllers working on Tangier airport. To the best knowledge, it’s the first
corpus that aligns speech data and structured data perfectly to all for a richer ASR
modeling. In this paper, we will describe the Tangier’s airport geographic specifi-
cation, and the techniques used to ensure quality of recording, data collecting, and
their transcription, using the META data annotation and the advantages of the ATC
phraseology as a reducing and controlling language, with synchronized flight data,
and weather reports as trusted structured data.

Keywords Speech corpus · Air traffic control · Phraseology · METAR · ADSB ·

Flight data · Speech recognition · Air traffic control communication

Tangier IBN BATTOUTA international airport, which is situated in the north of Morocco.

K. M. Saïd (B) · L. Abdelouahid

Laboratory of Innovative Technologies (LTI), Abdelmalek Essaadi University, Tangier, Morocco
e-mail: [email protected]
L. Abdelouahid
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 371

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_26
372 K. M. Saïd and L. Abdelouahid

1 Introduction

Air Traffic Control (ATC) is a service offered by air controllers who direct aircraft
flying on its assigned airspace or taxing on the ground. The initial task of ATC is
to avoid collisions and provide alerts and information for pilots; it’s manage and
organize the flow of air traffic.
The location of aircraft in an assigned airspace is monitored by air traffic
controllers using primary or secondary radar data. They communicate with aircraft
pilots using VHF (Very High Frequency) radio. In order to prevent collisions, Air
traffic controllers make vertical and horizontal separation between aircrafts. Pilots
must request authorization to change the flight level or heading.
On the one hand, Tangier’s airport and most airports in Morocco are not
equipped with surveillance radars, which adds more cognitive workload and stress
on controllers. This implies that such controllers need to permanently know the posi-
tion of aircraft and ensure separation especially when height traffic is situated. The
unique way to exchange information and requests or order instructions between the
controller and the pilot is VHF radio communication.
On the other hand Natural Language Understanding (NLU) and Automatic Speech
Recognition (ASR) are significantly well developed and many applications and
toolkits were delivered and can be used during human activities.
For all reasons, the purpose of introducing ASR and NLU is to assist and help
controllers by displaying flight data in visual interface after aircraft identification
instead of only listening technique for better performances, improve decision-making
and enhance safety.
However, the disadvantage of using a statistical ASR and/or NLU is that the
introduced errors may add risk. The Ibn Battouta communication corpus comes to
mitigate this risk by introducing a structured data set in parallel to the speech for
better processing capacities.

2 ATC Communication and Systems Data Characteristics

In this section we start by describing the specification of ATC communication then

we will present the structure of ADS-B data and METAR report.

2.1 Specificity of ATC Speech

Here we present the list of most common specifics of ATC speech:

• Special phraseology: ATC communications are messages exchanged between
pilots and controllers via VHF radio channel using a controlled language known
The IBN BATTOUTA Air Traffic Control Corpus … 373

Table 1 Example of ICAO phraseology

Message
PILOT TOWER AIRCRAFT Information Request Instruction
Tangier Tower RAM972C at holding point ready for departure
runway 28
CONTROLLER AIRCRAFT TOWER Information Instruction
RAM972C Tangier Tower Last wind 350° cleared for takeoff
09 Kt runway 28

as the ICAO1 phraseology, ATC is based on English but could be also in French
or Spanish if both pilot and controller are native speakers.
This property can be also extended to the pronunciation of some words inferior
to 4 letters, such as spelling of “navaids” (navigational aid) or aircraft registration
number that must be in phonetic alphabets. In Table 1 we subdivide the message
between the aircraft with call sign RAM972C and Tangier Tower as below:
• Native and Non-native speakers: All utterances are pronounced by controllers
or pilots who could be native speakers or non native speakers of the English
language, for example (Moroccan pilot and Gibraltar controller).
• Noise and voice quality: VHF transmission is mostly affected by natural relief that
generate noise and micro cut on voice communication. The standard phraseology
always tries to compensate it, however the ASR engine is very sensitive to noise
and results could be alterated.
For best performance we will only record and transcribe voice with normal signal
strength and natural noise.
• Height level of speech: The speech rate is higher; the controller must follow the
aircraft position and monitor the traffic flow. If more than two or three aircrafts
are presented to the controller airspace, he should quickly provide information.

2.2 Structured Data

Here we will describe and give examples of ADS-B data and METAR Report.
• ADS-B Data (ADS–B): Automatic dependent surveillance—broadcast is a tech-
nology of monitoring and surveillance that allows tracking an aircraft position
via satellite navigation and broadcasting flight information, it contains flight call
sign, speed, altitude, vertical speed, heading and GPS latitude and longitude.

1 https://www.icao.int.
374 K. M. Saïd and L. Abdelouahid

For our case, those data will be collected by our ADS-B receiver and will be used
as a replacement or alternative for secondary surveillance radar data. Table 2 shows
all data collected from ADS-B receiver.
• METAR is a weather information report for aerodrome periodically generated
each 30 min. It contains data for the wind direction and speed, temperature, dew
point, cloud cover and heights, visibility, and barometric pressure. It’s primordially
used by aircraft pilots and controllers to determine run way in use and flight rules
during takeoff or landing conditions.

Table 3 is a report of Tangier aerodrome was made, on 20/11/2019 at 17:30 UTC,

Wind 14 Kt from the North with gusts up to 25 Kt, Temperature 13 °C. Humidity 84%.
Pressure 1017 Hpa. Visibility: 5000 m. Broken clouds at a height of 2000 ft. No
significant changes in next 2 h.

3 Related Work

In last few years many speech corpus was created using Web and TV recording but the
real ATC speech corpus are rarely found. Only 7 ATC speech corpora in the literature,
half of them are not available (nnMATC, VOCALISE) because of commercial issue
or military information. A good description of state of arts of ATC speech corpus is
given by Van Nhan Nguyen [1].
He reviews five well-known ATC-related corpora including The Air Traffic
Control Complete LDC94S14A corpus, ATCOSIM Corpus of Non-Prompted Clean
Air Traffic Control Speech, HIWIRE corpus, Air Traffic Control Communication
Speech Corpus and Air Traffic Control Communication corpus. Beside the cited
corpora, one additional corpus is the AIRBUS-ATC Corpus. In the following we
present list of ATC corpora:
• The Air Traffic Control Complete Corpus

The LDC94S14A Air Traffic Control Complete corpus [2] is comprised of recorded
speech of air traffic voice communication between different pilots and controllers, it’s
published in 1994.It consists of three sub corpora, their transmissions were collected
from the following airports: Washington National (DCA), Logan International (BOS)
and Dallas Fort Worth (DFW). The data base contains approximately 70 h of pilot
and controller native English speakers. The transmissions collected via antennas and
radio receivers.
• The HIWARE Corpus

The HIWIRE database [3] is a noisy English speech corpus of communications

between controllers and pilots in military air traffic control. According to [3], the
database contains around 8100 English utterances pronounced by non native English
Table 2 Example of ADS-B data
Date Time Call Sign Radar Altitude Speed Head Vertical Latitude Longitude
The IBN BATTOUTA Air Traffic Control Corpus …

06/09/2019 07:31:59.47 RAM972C 6461 38000ft 450kt 270° 850ft/min 35.230865 -6.561527
375
376 K. M. Saïd and L. Abdelouahid

Table 3 Example of METAR Reporta

Aerodrome Day/Time Wind Direct Visibility Clouds Temp/Dew Presure
Speed
GMTT 201730Z 35014G25KT 5000 BKN020 13/11 Q1017 NOSIG

speakers (31 French, 20 Greek, 20 Italian, and 10 Spanish speakers). The data was
recorded in studio with a microphone and real noise recorded in an airplane cockpit
was artificially added to the data.
• The Air Traffic Control Communication Speech Corpus

The Air Traffic Control Communication Speech corpus [4] is a database of voice
communications between pilots and controllers from different control entities,
ground control tower control, approach control and area control. The most speech
data was recorded from the Czech Republic Air Navigation Services. The rest of the
speech data was recorded from Lithuania and Philippines airspace. According to [5];
the Air Traffic Control Communication corpus contains around 138 h of recordings
communications between pilots and air traffic controllers.
• The English TTS speech Corpora

The English TTS speech Corpora are small samples ATC-related corpora of air
traffic (pilot) messages. It contains Serbian accent [6] 3000 utterances, Taiwanese
accent [7] 1000 utterances, Czech accent [8] 1692 utterances and German accent [9]
1685 utterances. The corpora are a simulation communications between pilot and
controllers for speech synthesis system.
• The ATCOSIM Corpus

The ATCOSIM Corpus of Non-Prompted Clean Air Traffic Control Speech

(ATCOSIM) [10] is a simulation speech corpus of ATC operators. The ATCOSIM
corpus consists of direct recordings of operational air traffic speech in civil en
route control simulation. The ATCOSIM corpus contains ten hours recording of
ten speakers: six male and four female controllers who were either German or Swiss
nationality. Their native languages are German, Swiss German or Swiss French. The
ATCOSIM corpus is free of charge and online available for public.2
• AIRBUS-ATC Corpus

The AIRBUS-ATC corpus real life French accented corpus of ATC [11] which is
composed of 2160 paired audio plus transcription. It totalizes around 59 h English
utterances. It contains Approach and Tower communications.

2 https://www.spsc.tugraz.at/databases-and-tools/atcosim-air-traffic-control-simulation-speech-cor

pus.html.
The IBN BATTOUTA Air Traffic Control Corpus … 377

Following the seven reviewed ATC-related speech corpora, we removed restricted

ATC corpora and add the IBN Battouta ATC speech corpus with flight data char-
acteristics. We summarized the important corpora in Table 4 for comparison. It’s
important to note that none of these corpora provide the parallel and synchronized
structured data of flight information and weather.

4 New Concept for IBN BATTOUTA Corpus

In Figure 1 we describe the new architecture used to create the IBN BATTOUA
speech corpus. It consists of record audio speech with synchronized related aircraft
flight data because most utterances contain codes and information are pronounced
during conversation between pilots and controllers.
To ensure the synchronization, the record station is connected to the voice and data
sources at the same time, the VHF receiver controlled via RS232 [12] communication
provide voice communication between Pilot and ATC operating in VHF band 118-
137 Mhz. Flight information and weather data are downloaded respectively from
ADS-B receiver and NWS4 FTP server using internet connection.

4.1 Geographical Specification of Tangier

Tangier international airport called IBN BATTOUTA airport is located in Tangier

city that is a frontier city between Morocco, Spain and Gibraltar that allow us
to collect Flight data and ATC radio communication related to different airspace
controller and language spoken with native or non native accent. When entering or
exiting any country airspace we have new instructions exchanged and frequency
code transmitted. Figure 2 shows the airspace frontiers between Morocco, Spain and
Gibraltar.

4.2 Corpus Characteristics

IBN Battouta corpus contains three types of Air Traffic Control communications:
Regional Radar Control (CCR), approach (APPR) and tower (TWR) controls, all
coming from different airport and different regional control. It also contains ATIS
(Automatic Terminal Information Service) recordings from Tangier airport. Voice
and data collection it’s divided into 7sub corpora corresponding to different entities
with unique VHF frequency as described in Table 5.
 378

Table 4 Summary of features of ATC-related corpora

LDC94S14A HIWIRE ATCC ATCOSIM AIRBUS-ATC IBN Battouta
1994 2008 2014 2016 2018 2020
Recording situation Record content ATCO & PLT N/A ATCO & PLT Civil ATCO ATCO & PLT ATCO & PLT
Control entity APPR N/A En-route& APPR En-route ATIS & APPR En-route& APPR
Region USA (GR/IT/FR) DE/CH/FR Europe (FR) MAR, BR, ES
Speaking style operational Prompted text Operational Operational Operational Operational
Duration 70H 8099 Utterance 138H 10H 59H 05H
Recording setup bandwidth narrowband Wideband Narrowband Wideband Narrowband Narrowband
Transmission radio None Radio None Radio Radio
Radio noise high None Low None Low Low
Acoustical noise CO & CR CO (artificial) CO & CR CR CO & CR CO & CR
Signal source VHF radio Direct micro VHF radio Direct micro VHF radio VHF radio
Speaker property English mostly native Non-native Non-native Non-native Non-native Both
Number of speakers unknown (large) 81 Unknown (large) 10 18 Unknown (large)
flight data ADS-B no No No No No Yes
METAR no No No No No Yes
Available Free sample Payment Yes Yes No Free Sample
* ATCO: Air Traffic Controller * PLT: Pilot * CO: Cockpit * CR: Control Room
K. M. Saïd and L. Abdelouahid
The IBN BATTOUTA Air Traffic Control Corpus … 379

Fig. 1 New Architecture of IBN Battouta speech corpus

Fig. 2 Tangier geographic position

380 K. M. Saïd and L. Abdelouahid

Table 5 Structure of IBN

ICAO code Description VHF frequency Country
Battouta ATC corpus
GMMM Casablanca control 125.5 MHz Morocco
regional
ECS Sevilla control 134.8 MHz Spain
regional
GMTT Tangier airport 119.5 MHz Morocco
tower
GMTT Tangier airport 121.2 MHz Morocco
approach
GMTT Tangier airport 123.7 MHz Morocco
ATIS
LXGB Gibraltar airport 131.2 MHz Gibraltar
tower
LXGB Gibraltar airport 122.8 MHz Gibraltar
approach

5 Corpus Acquisition and Processing

5.1 Audio Recording

All raw audio files are recorded under VAD (voice active detection) using the
computer input sound card, the signal strength is provided by the VHF receiver
from ICOM Company allowing us to avoid record silent for hard disk space saving.
The recording is a sampling rate of 16 kHz in mono with a single channel which is
a CMU Sphinx3 ASR toolkit [10] recommendation for adapting acoustic model.

5.2 Flight Data ASB-B Log

The logging of aircraft flight data is provided by Air Nav Systems Company by
installing radar box receiver station connected to an Omni antenna 1090 MHz. aircraft
flight information is saved in txt format every 2 s. We add a program for calculating the
distance and radial between aircraft position and Tangier airport, basing respectively
on Haversine formula (great-circle distance between two points on a sphere ) as
indicated in Eq1, and the compass bearing (the horizontal angle between the direction
of an object and another object) by given their longitudes and latitudes.
    
φ2 − φ1 λ2 − λ1
d = 2r arcsin sin2 + cos(φ2 ) sin2
2 2

3 https://cmusphinx.sourceforge.net.
The IBN BATTOUTA Air Traffic Control Corpus … 381

The parameters in Eq1: Φ 1 and Φ 2 are latitudes and λ1 and λ2 are longitudes,
while r is the radius of the Earth around 3961 miles.

5.3 METAR Report

The meteorological report is collected from FTP server of The National Weather
Service (NWS) [12]. It’s a US governmental organization providing weather, water,
and climate data, forecasts and warnings. During record we add to METAR report
the date and time corresponding to VAD in Appendix A we show the Python 3 code
to get the METAR report from NWS ftp server.

5.4 Processing

We developed a special program to control the VHF receiver via RS232 [12] commu-
nication and display a menu of used frequencies list corresponding to their airport
code. The user can select the desired frequency and automatically start recording
on VAD. When a signal is detected it saves the current METAR report related to
selected airport. The ADS-B data is saved separately by Radar box software in a
log file each two seconds. We added a script to calculate the haversine distance and
compass bearing to Tangier airport, because ATC controller always request current
position from aircraft pilot when entering his airspace (Fig. 3).

Fig. 3 Screenshot Python 3 program for VHF Receiver

382 K. M. Saïd and L. Abdelouahid

Table 6 Time duration and

ATC unit Volume Utterance
number of utterances in Ibn
Battouta ATC corpus Casablanca control 90 min 550
Sevilla control 30 min 202
Tangier tower 40 min 280
Tangier approach 80 min 615
Tangier ATIS 40 min 95
Gibraltar tower 10 min 127
Gibraltar approach 10 min 132
Total 05H00 min 2001

5.5 Transcription

We based on Express Scribe Transcription Software4 for the transcription of English

utterance recorded in the corpus according to American-English spelling rules. One
of the authors of this paper is a private pilot PPL; they manually started transcription
with the help of airlines pilot and air traffic controller. We also contacted the real
speakers working in Tangier airport for validation of misunderstanding speaking
words. These diligence steps are carried out to ensure the high quality of the corpus.
In Table 6 we summarize the current ATC corpus volume and number of utterances.

6 Conclusion and Outlook

The IBN Battouta corpus is the first real life ATC corpus including flight data and
weather report. We proposed a new architecture to create voices utterance based on
hardware VHF and ADS-B receivers. We also developed new program for controlling
and recording voices without silent time duration. Furthermore, we added two codes
for calculating the arc distance and the compass bearing between aircraft position
and Tangier airport that mostly requested by ATC controllers to aircraft pilots.
This corpus will be used to contribute for ASR systems and training acoustic model
English Moroccan accented and developing economics systems for aviation radar
equipment’s. We estimate to record and transcribe around 10 h English utterance for
each airport frequencies.
A sample version of speech corpora can be provided under request, free of charge
for scientific research.

Acknowledgments The authors are grateful for the cooperation of air traffic controllers working
in Tangier airport and the captain Mountassir Noureddine airline pilot for his valuable help and
guidance. This work was funded by DETROIT TELECOM Company and ULTRA CONTROL
Technical Center.

4 https://www.nch.com.au/scribe/index.html.
The IBN BATTOUTA Air Traffic Control Corpus … 383

Appendix A

def get_metar_info(code):
temp_url ="https://tgftp.nws.noaa.gov/data/observations/
metar/stations/"+code+".txt"
request = requests.get(temp_url)
if request.status_code == 200:
url = urlopen(temp_url)
for i, line in enumerate(url):
if i == 1:
line = str(line)
end = line.__len__()-3
metar_info = line[2:end]
return metar_info
else:
return ''

References

1. Van Nhan Nguyen, Using linguistic knowledge for improving automatic speech recognition
accuracy in air traffic control (2016)
2. Godfrey, J.: Air traffic control complete ldc94s14a. Linguistic Data Consortium, Philadelphia
(1994)
3. Segura, J.C., Ehrette, T., Potamianos, A., Fohr, D., Illina, I., Breton, P.A., Clot, V., Gemello,
R., Matassoni, M., Maragos, P.: The hiwire database, a noisy and non-native english speech
corpus for cockpit communication. (2007). Online. http://www.hiwire.org
4. Sm´ıdl, Luboˇs, PavelIrcing.: Air traffic control communication (atcc) speech corpus. CLARIN
Annual Conference 2014 in Soesterberg. National Center for Biotechnology Information, The
Netherlands. (2014). https://www.ncbi.nlm.nih.gov
5. LuboˇsSm´ıdl.: Air traffic control communication. LINDAT/CLARIN digital library at Institute
of Formal and Applied Linguistics, Charles University in Prague (2011)
6. Jindˇrich Matouˇsek and Daniel Tihelka.: English TTS speech corpus of air traffic (pilot)
messages—serbian accent. LINDAT/CLARIN digital library at Institute of Formal and Applied
Linguistics, Charles University in Prague (2014)
7. Jindˇrich Matouˇsek and Daniel Tihelka.: English TTS speech corpus of air traffic (pilot)
messages—Taiwanese accent. LINDAT/CLARIN digital library at Institute of Formal and
Applied Linguistics, Charles University in Prague (2014)
8. Matoušek, J., Tihelka, D.: English TTS speech corpus of air traffic (pilot) messages—Czech
accent, LINDAT/CLARIN digital library at the Institute of Formal and Applied Linguistics
(ÚFAL). Charles University, Faculty of Mathematics and Physics (2015a)
9. Matoušek, J., Tihelka, D.: English TTS speech corpus of air traffic (pilot) messages—German
accent, LINDAT/CLARIN digital library at the Institute of Formal and Applied Linguistics
(ÚFAL). Charles University, Faculty of Mathematics and Physics (2015b)
10. Stefan Petrik, Konrad Hofbauer, Horst Hering.: The atcosim corpus of non prompted clean
air traffic control speech. In: Proceedings of the Sixth International Conference on Language
Resources and Evaluation (LREC’08), Marrakech, Morocco. European Language Resources
Association (ELRA). (May 2008). https://www.lrecconf.org/proceedings/lrec2008/
384 K. M. Saïd and L. Abdelouahid

11. Estelle Delpech, Marion Laignelet, Christophe Pimm, Céline Raynal, Michal Trzos, et al.:
A Real-life, French-accented Corpus of Air Traffic Control Communications. Language
Resources and Evaluation Conference (LREC), Miyazaki, Japan. ffhal-1725882 (May 2018)
12. Afritha Amelia, Julham, Bakti Viyata Sundawa, Morlan Pardede, Wiwinta Sutrisnoa,
Muhammad Rusdi: Implementation of the RS232 communication trainer using computers and
the ATMEGA microcontroller for interface engineering Courses, IOP Conf. Series: Journal of
Physics: Conf. Series 890 (2017)
Weed Recognition System for Low-Land
Rice Precision Farming Using Deep
Learning Approach

Olayemi Mikail Olaniyi, Emmanuel Daniya,

Ibrahim Mohammed Abdullahi, Jibril Abdullahi Bala,
and Esther Ayobami Olanrewaju

Abstract Precision farming helps to achieve maintainable agriculture, with an

objective of boosting agricultural products with minimal negative impact on the
environment. This paper outlines a deep learning approach based on Single Shot
multibox Detector (SSD) to classify and locate weeds in low-land rice precision
farming. This approach is designed for post-emergence application of herbicide for
weed control in lowland rice fields. The SSD uses VGG-16 deep learning-based
network architecture to extract a feature map. The adoption of multiscale features
and convolution filter enables the algorithm to have a considerable high accuracy
even at varying resolutions. Using SSD to train the weed recognition model, an
entire system accuracy of 86% was recorded. The algorithm also has a system sensi-
tivity of 93% and a precision value of 84%. The trained SSD model had an accuracy
of 99% for close-up high definition images. The results of the system performance
evaluation showed that the trained model could be adopted on a real rice farm to help
reduce herbicide wastage and improve rice production with low chemical usage.

Keywords Precision agriculture · Deep learning algorithm · SSD · Google

TensorFlow · Low-land rice

O. M. Olaniyi (B) · I. M. Abdullahi · E. A. Olanrewaju

Department of Computer Engineering, Federal University of Technology, Minna, Nigeria
e-mail: [email protected]
I. M. Abdullahi
e-mail: [email protected]
E. A. Olanrewaju
e-mail: [email protected]
E. Daniya
Department of Crop Production, Federal University of Technology, Minna, Nigeria
e-mail: [email protected]
J. A. Bala
Department of Mechatronics Engineering, Federal University of Technology, Minna, Nigeria
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 385

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_27
386 O. M. Olaniyi et al.

1 Introduction

Agriculture is an important sector of any society since every individual relies on food
for survival [1]. Precision agricultural practices help to achieve sustainable agricul-
ture and improve production yield [2, 3]. This technique shows a huge tendency to
lessen the number of chemicals used on farms to kill weeds. The concept of precision
farming helps apply inputs at the appropriate time, at an appropriate place, in the
correct amount, and in a proper manner, all which are targeted for improved crop
production. Hand weeding and use of hoes and cutlasses are manual means of weed
control on the farms which although, environmentally friendly, result in a high cost
of labourer acquisition and are very tedious in nature [4–6]. Herbicide usage domi-
nates the chemical method of weed control and is seen as an alternative approach to
weed control [7, 8]. However, researchers have frowned upon uniform spraying and
health concerns have been raised on the safety of the crops when consumed [5, 9,
10]. It is obvious that accurate information is needed to identify weeds from crops in
order to avoid damage to the crop itself. Therefore, the need arises for a weed control
technique that can obtain and analyse information about weeds from the farm, as
well as act accordingly to control the weeds on the farm. Watchareeruetai et al. [11]
suggested that automated weed control is a suitable technique for weed management
by adopting the advance in technology as a tool. Similarly, [12] stated that computer
vision abilities help to identify and classify weeds swiftly than manual techniques.
Faridi and Aboonajmi [13] and Liakos et al. [14] also suggested that application
of machine learning approach directly on farm would help reduce farmer’s stress,
manage cost, provide flexibility, accuracy, and improve quality of crop production.
Due to the development in the computer vision field, it is easier to use automated
techniques to acquire information, process them and then utilize the information to
separate crops from the weeds [12]. The system computations are based on geometric
characteristics such as aspect ratio, shape factor, area or length of the weeds and crops
[15, 16]. Colour images are also used to establish the quality of seeds, distinguish
weeds from crops, classify plant diseases and other damages that could be made to
plants [5, 17]. Textural features also form a powerful discriminate for weeds and
crops [18].
In low-land rice production, one of the major ways of weed control is the early
application of herbicides. The application which [19] suggested should be done
within 40 to 50 days of rice planting. These approaches discussed above addressed
broadleaved weeds for upland crops which are not suitable to be used in low-land
rice production. Rice is a grass species, whose weeds often looks similar to the
rice crop itself and the area of low-land is a water-logged area. These factors make
existing methods insufficient to address the weeds in that area. As a result of this
shortcoming, this research follows an object detection approach based on a deep
neural network algorithm called Single Shot multi-box Detector (SSD). The SSD is
one of the fastest algorithms for object detection and it is known for achieving good
result on targeted object [20]. The SSD which is based on VGG-16 architecture
helps to differentiate weeds from rice of low-land rice field using multiscale features
Weed Recognition System for Low-Land Rice Precision Farming … 387

and convolution filters. The remaining of the paper is organised as follows: Sect. 2
presents an overview of the SSD model used while the image recognition process is
presented in Sect. 3. Section 4 shows the result obtained from the simulation of the
process and Sect. 5 presents the conclusion made and new research directions.

2 Review: Automated Weed Control and Object Detection

Techniques

Several research works have been reported on intelligent and automated weed control
[5, 10, 18, 21, 22, 23] and real-time object detection techniques [20–28]. In this paper,
the review of intelligent and automated techniques is reported in Sect. 2.1, while the
review of real-time object detection techniques such as the SSD are reported in
Sect. 2.2.

2.1 Automated and Intelligent Weed Control

Several works exist in the area of automated and intelligent weed control. [21] used
fast Fourier transform to extract the shape of weeds in the cornfield. The shape
features were extracted and used as a discriminate between the weeds and the crop
of the cornfield. Similarly, [29] used morphological thresholding techniques, erosion
and dilation to differentiate between weed and plants. In addition, [18] adopted the
use of textural features to distinguish weeds from vegetables under varying lighting.
The researcher used Gray Level Co-occurrence Matrix to calculate textural features
which form the feature map. The map was then fed to into a Support Vector Machine
algorithm for weed classification. Other approaches have used colour features to
separate weeds from crops, by adopting separation of vegetative plants from soil as
seen in [5, 10, 22, 23].

2.2 Real-Time Object Detection: SSD Model

Single Shot Detector (SSD) is one of the fastest algorithms for real-time object
detection [20]. The network was targeted at solving the problem of region proposed
network of faster region with Convolution Neural Network (R-CNN) by using multi-
scale features to extract features map and convolutional features for detection [24].
The SSD model consists of two parts, extraction of features map using multiscale
features and application of convolution filters to detect objects. In SSD, localization
and classification are done once by running the algorithm through the image in one
forward pass to give a prediction which consists of an image with a bounding box
388 O. M. Olaniyi et al.

Fig. 1 Network architecture of SSD [24]

and the confidence score of detection. The SSD goes through matching strategies,
hard negative limiting, data augmentation and non-maximum suppression to remove
prediction duplicates. The network architecture as shown in Fig. 1 is based on VGG-
16 architecture which has 16 layers. The figure shows the convolution stages the
image goes through so that the features of images can be extracted. Other network
architectures are ResNet-101, Inception V2, Inception V3, Inception, ResNet and
MobileNet [25, 26–28]. VGG-16 is reported to have high performance on image
classification tasks and has improved results on problems that involve using transfer
learning.
Rather than use the original fully connected VGG layers, six supplementary layers
are added to permit feature extraction at multiple scales and gradually reduce the input
size at each successive layer. SSD runs the convolutional network on the image once
to get the features map, the features map is then used to predict the bounding box
and get the confidence level of the classifier. SSD can learn at different scales using
the varying size of filters. The convolution layer acts and operates at different scales
to enable it to locate the images at a fast speed. The box drawn has four parameters,
namely: the centre, width, height and the vector of probability also known as the
confidence level of the classifier. In SSD, predictions are classified as positive matches
or negative matches depending on the value obtained for Intersection Over Union
(IOU). IOU also known as Jaccard index compares ratio of area of overlap to area of
union. Prior which are bounding box of fixed size with close match to distribution
of original ground truth are selected in such a way that their IOU value is greater
than 0.5. The multibox starts with the prior as initials prediction and starts to regress
towards the ground truth boxes. However negative IOU are discarded. The objective
loss function of the model is given as the weighted sum of localization loss (loc) and
confidence (conf)

1 
L(x, c, lg) = L con f (x, c)+ ∝ L loc (x, l, g) (1)
N
Alpha (∝) is a weight term which is set to 1 for cross-validation of the model, with
a key reason for balancing the influence of the location loss. The criteria for choosing
∝ is to select an optimal value that lessen the value of the localization loss and brings
the prediction closer to the ground truth. N is the number of matched default boxes
Weed Recognition System for Low-Land Rice Precision Farming … 389

and if N = 0 loss is set to 0. The localization loss is the smooth L1 loss between
the predicted box and ground truth. The model is regressed to the coordinate of the
centre of x and the coordinate of the centre of y (cx, cy) of the bounding box(d). The
localization loss tells the mismatch between the ground truth box and the predicted
box. The positive match is worked with while the negative match is ignored. The
localization loss L loc is given as:
N   
L loc (x, l, g) = xikj smooth L1 lim − g mj (2)
i∈ posm∈{cx,cy,w,h}

 vcx 
j = gj
g cx − dicx / diw (3)
 
cy vcy cy
gj = gj − di /dih (4)

 
g wj
g wj = log (5)
diw

g hj
g hj = log (6)
dih

Confidence loss is a SoftMax loss over multiple classes. This loss is used to
penalise loss according to the confidence score of the corresponding class for every
positive match prediction, while in negative case “class 0” is used which shows no
object was detected. Confidence loss is given as:
N  p   
L con f (x, c) = − xikj log ci − log cio (7)
i∈ pos i∈N eg

p
c
where cio = i
p exp(ci )
p

The scale of the default boxes for each feature map is given as:

Smax − Smin
Sk = Smin + (k − 1)k[1, m] (8)
m−1

Since a lot of negative matches are made during prediction, due to the fact that
we make more predictions than classes of objects; an imbalance is therefore created
which has a negative influence on the training process. The process, known as hard
negative mining, helps the SSD to use negative sampling to know what makes up a bad
sampling. SSD then picks up the top loss and makes the ratio between the negative
and positive to be at almost 3:1, this then leads to faster and stable training. The
model accuracy is improved by using data augmentation process. Data is augmented
using either original image, sampling of a patch (with values of 0.1, 0.3, 0.5, 0.7 or
0.9), or randomly sample of a patch. This part is needed to handle colour distortion
and image variant sizes. The last step of the SSD model is non-max suppression and
390 O. M. Olaniyi et al.

Fig. 2 SSD implementation steps

it helps to remove duplicate predictions of the same image by sorting the predictions
via their confidence loss. Figure 2 shows the SSD algorithm.

3 Weed Recognition System Development

This section describes the methods used for the development and deployment of the
weed recognition system. The weed recognition process consists of image acquisi-
tion, labelling of image (data preparation), splitting images into testing and training,
generation of TensorFlow record, setting up of model configuration files, training
of the selected SSD model, exporting of the frozen inference graph and using of
the frozen inference graph to classify new images. The summary of the processes is
shown in Fig. 3.

3.1 Image Acquisition into Database

Images of seventeen (17) different types of low-land weed species of rice were
obtained from the teaching and research farm, Department of Crop Production,
Federal University of Technology (FUT), Minna, Niger State, Nigeria between 18th
May, 2018 to 20th August, 2018. Also, images were obtained from the Google server
about the same time. Images from the FUT teaching and research farm were taken
with a 13 Megapixel Tecno Camon CX air camera with a resolution of 4160 X 3120
pixels and a focus of 0.15 s. The images were taken at different time intervals of rice
growth, namely: one week, two weeks, one month and two months. A total of 3866
images were acquired into the database for the recognition process. Figure 4 shows
part of the images acquired from the at the FUT teaching and research farm while
Weed Recognition System for Low-Land Rice Precision Farming … 391

Fig. 3 Image recognition process

Fig. 4 Image acquired from the Cereal and Legume Crops practical plot, of teaching and research
farm Department of Crop Production, Federal University of Technology, Minna, Nigeria for data
preparation

Fig. 5 shows samples of the images acquired from the Google Server. Figure 6 shows
the process chart for the image acquisition process.
392 O. M. Olaniyi et al.

Fig. 5 Sample images of weed obtained from Google server

Fig. 6 Process chart of

image acquisition and START
storage system

Input
Image

Delete Image

Is weed?

YES
Trained
Store Image
Sample

STOP
Weed Recognition System for Low-Land Rice Precision Farming … 393

3.2 Image Annotation

After obtaining a dataset of the images, a process called data preparation was carried
out which is a process of putting the data into a suitable form which makes it easier
for use. The images of weeds obtained from Google server were resized to a size of
250 × 250 pixels using adobe Photoshop to minimise the size of the images and the
memory requirements. A graphical tool for data preparation called ‘Label Image’,
was used to draw a rectangular bounding box around the weeds in the image and the
boxes were labelled as ‘weed’. This process is called annotation. The essence of the
annotation is to draw the bounding box around the image so that the object of interest
(weed) can be separated from the background and other objects in the image. The
labelling shows the class the object belongs to while the bounding box determines the
location of the weed in the image through a process called localization as shown in
Fig. 7. This forms the next step in the object recognition process. The labelled image
was saved in XML format which contains the four parameters (xmin, xmax, ymin
and ymax) which describe a bounding box, needed for the object to be recognized.
The four coordinates represent the x coordinate of the centre, the y coordinate of the
centre, the height of the object and the width. The labelled image was saved in XML
format as a dataset for the training. After the first stage of data preparation the image,
the images obtained were divided into test and train. 90% of the data was used for
training while the remaining 10% was used for testing. A total number of 4860 and
540 datasets were obtained for training and testing respectively. 10 samples each of
the training and testing datasets are respectively shown in Tables 1 and 2.

Fig. 7 Weed annotation with labelling

394 O. M. Olaniyi et al.

Table 1 Sample of the training dataset

Filename Width Height Class Xmin Ymin Xmax Ymax
Aeschynomene indica L. (16).jpg 259 194 weed 7 4 259 162
Aeschynomene indica L. (17).jpg 259 194 weed 21 15 259 162
Aeschynomene indica L. (18).jpg 259 194 weed 6 5 259 159
Aeschynomene indica L. (19).jpg 194 259 weed 3 3 189 259
Aeschynomene indica L. (22).jpg 259 194 weed 48 1 259 186
Aeschynomene indica L. (23).jpg 259 194 weed 3 1 259 193
Aeschynomene indica L. (25).jpg 259 194 weed 1 3 258 194
Aeschynomene indica L. (26).jpg 259 194 weed 36 49 191 194
Aeschynomene indica L. (27).jpg 259 194 weed 15 6 235 192
Aeschynomene indica L. (28).jpg 259 194 weed 2 5 257 191

Table 2 Sample of the testing dataset

Filename Width Height Class Xmin Ymin Xmax Ymax
Commelina benghalensis L. 001.jpg 230 250 weed 1 1 230 249
Commelina benghalensis L. 002.jpg 250 166 weed 2 1 250 165
Commelina benghalensis L. 003.jpg 250 166 weed 4 5 245 166
Commelina benghalensis L. 004.jpg 250 140 weed 9 1 246 140
Commelina benghalensis L. 005.jpg 250 200 weed 3 2 249 200
Commelina benghalensis L. 006.jpg 250 179 weed 6 3 250 167
Commelina benghalensis L. 007.jpg 250 166 weed 41 12 220 166
Commelina benghalensis L. 008.jpg 250 187 weed 5 4 249 182
Commelina benghalensis L. 009.jpg 250 140 weed 56 5 250 137
Echinochloa colona (L.) Link 001.jpg 257 196 weed 11 7 257 193

3.3 Creating TensorFlow Records

Google TensorFlow (TF) is an open source software library for machine learning
application such as a deep neural network. The google TF was chosen for this
study because of its large number of trained libraries which are well recognised
for machine learning applications and object detection exercises. The TF creation
process is shown in Fig. 8. The models for training was obtained from GitHub
repository and saved. The test and train images together with respective XML were
converted into comma-separated value (CSV) format. The CSV format for the test
and train was then converted to TensorFlow record for training the SSD model.
Weed Recognition System for Low-Land Rice Precision Farming … 395

Identify Weed
Literature Review Meet Expert
and Rice

Fatkum Batch
Actual Image Image Acquisition
Downloader

XML Files
and Image
Image annotation Resize Image
250 X 250 pixels
LabelImg (250 x 250 pixels)

XML Files and Pascal voc format Python code and

Image (CSV) library

Pascal voc format Python code and

Tfrecord
(CSV) library

.record
Dataset
Tfrecord

Fig. 8 Dataset generation process

3.4 Paperspace

Since the model training cannot be done locally due to hardware constraints, a virtual
cloud computing platform was obtained on Paperspace. Paperspace is an online
cloud computing platform that gives limitless computing power. The platform is used
for deep learning, data exploration and gaming. Paperspace provides the Graphics
Processing Unit (GPU) necessary for the training of the deep learning algorithm.
GPU is necessary for training deep learning models because they require a lot of
computational power to run on, and considerable hardware to run efficiently [30]. In
the deep neural network training, two basic operations are performed, the forward
and backward pass operations. Weights of the neural network are updated on the basis
of the error obtained in the forward pass. For a typical convolution neural network
of 16 hidden layers, it has about 140 million parameters.
396 O. M. Olaniyi et al.

Table 3 Specification of the

Machine name Performance specifications
paperspace machine
Operating system Ubuntu 16.04 SSH/terminal only
RAM size 30 GB
Storage size 50 GB SSD
CPU 8 CPU cores plus GPU
Machine IP address 184.105.169.162
Support for AVX instruction NO

The parameters are the weights and bias. Thinking of the number of multiplication
applications, it would take typical system years to completely train the model [31].
Thus, using the GPU can help the model to be trained at once and in a shorter
amount of time. The server image of Ubuntu 16.04 Secure Shell (SSH)/terminal
was rented for this operation. The details of the server image are shown in Table
3. Basic python libraries such as NumPy, pillow, TensorFlow 1.5, pandas, CUDA
drivers from Nvidia, TensorFlow GPU were installed on the machine to enable the
SSD model training. The configuration of model selection and configuration files
modification was done locally using Python 3.6. The configuration files, the images,
the TF records and the SSD model training directory were copied inside a directory
and transferred to the Paperspace for the training.

3.5 Model Training

The model configuration and libraries loading were done off line. Table 4 gives the
system specification of the laptop used. In the configuration file, the class number,
class name, path to TF records and batch size was configured. A class batch size
of 12 was used to prevent the RAM from running out of memory due to large data
thrown at it for training. A library for object detection (“protocol buffer”) was loaded
into the model to configure the model for the training on Paperspace. The training
commenced with a high error rate of 7 but as the training started the error gradually
got reduced until it got to nearly 1. Different checkpoints were made for the training
model at every 1000 steps.

Table 4 Offline machine

System name HP Pavilion 15-au010wm
used for model configuration
and extraction RAM size 12 GB
Processor Intel® Core™ i7
Video graphics Nvidia® GeForce 940MX
Storage size 1 TB
Support for AVX instruction YES
Weed Recognition System for Low-Land Rice Precision Farming … 397

The training stopped at 26500 steps. The training results were obtained via google
tensor board. Checkpoints of the model were recorded under a trained model after
every 1000 steps. A frozen inference graph of the model was created and each
checkpoint of the training makes up the result of the training. After the training, the
trained network was downloaded to a core i7 system whose specification is in Table 4
because of its support for AVX instruction. The checkpoints together with a protocol
buffer text file was used to load the tensor board to see how the training process took
place and how the losses were reduced to ensure the data got trained.

3.6 Model Testing

For the neural network testing, the frozen inference graph which was named weed
inference graph was loaded with the help of google TensorFlow 1.10. 50 random
images from testing folder were used for testing and the result was plotted. The
confidence level of each of the tested and recognized images was recorded on the
images.

3.7 Model Deployment

The downloaded model was a frozen inference graph which was extracted using
a google TensorFlow 1.10. The Model was loaded into Raspberry Pi3 via Virtual
Network Computing (VNC) interface after quantization. The offline model was
tested with images and the integrated system was tested to note its performance.
The performance evaluation of the device was noted for accuracy, precision and
sensitivity.

4 Results

The objective function of the SSD model is obtained by the summation of the loss
function. The loss function is given as sum of the classification and localization. The
goal is to have a loss below 2 for images that are quite similar. The two of them
together shows how well the model has learnt. From Fig. 9 the losses encountered
by the model was reduced as the training steps increases and which shows that the
SSD model was actually learning as the steps increases.
Figure 10 shows the total loss obtained. The total loss which is a combination
of localization loss and confidence loss. The two of them together which shows the
objective function of the network as shown in Sect. 2. The graph above shows the
network learning process as it moves from 0.000 steps to 24000 steps. From the graph
398 O. M. Olaniyi et al.

Fig. 9 Training result from tensor board

Fig. 10 Learning process of the deep learning algorithm

obtained from the tensor board, the SSD model was learning as the steps increased
and reduction in loss signifies an improvement in the neural network algorithm.
Figures 11 to 13 show the results obtained from the deployment of the model on
images from the farm and stand-alone images. The image was able to achieve high
Weed Recognition System for Low-Land Rice Precision Farming … 399

Fig. 11 Result obtained from running the model on weed species of rice obtained from the Cereal
and Legume Crops practical plot, of teaching and research farm Department of Crop Production,
Federal University of Technology, Minna, Nigeria

accuracy for high definition images and considerable accuracy for low-resolution
images obtained from the school farm.

Fig. 12 More results from running the model on the image from the Cereal and Legume Crops
practical plot, of teaching and research farm Department of Crop Production, Federal University of
Technology, Minna, Nigeria
400 O. M. Olaniyi et al.

Fig. 13 Running the model on close range image of low land weed obtained from the Google
server

4.1 Developed Model Performance Evaluation

The neural network was evaluated for 50 different test samples comprising of weeds
and rice based on accuracy, sensitivity and precision as used by Aggarwal & Agrawal
[32] and Aibinu et al. [33]. Sensitivity measures how well the neural network was
able to detect positive case while accuracy measures how well the neural network was
able to detect correctly the positive cases and precision measures the repeatability of
the measured value. The metrics are represented mathematically as:

Accuracy = TP / (TP + FN) (9)

Sensitivity = FP / (TN + FP) (10)

Precision = TP / (TP + FP) (11)

where:
TP is true positive (number of items classified correctly)
FP is a false positive (number of negative cases classified correctly)
TN is true negative (number of correct cases classified wrongly)
FN is false negative (number of incorrect cases classified wrongly).
Weed Recognition System for Low-Land Rice Precision Farming … 401

5 Conclusion

The SSD model based on three parts which the base convolution using the VGG
16 as the base network, this helped to achieve the lower level features, the auxiliary
convolution during the forward pass allowed more feature extraction so as to create a
features map that constitute a representation of the image at different scale and size.
Also, the prediction convolution helps to identify objects in the features map
leading to the high accuracy of the model. Finally, the single shot multi- ‘box detector
trained model showed high accuracy value of 99, 98 and 89% for close up images
of weed species taken at high resolution and fairly good accuracy of 72, 69, 59%
for weeds found in between rice plants. The model was able to identify the weed
species associated with rice and did not misclassify the rice as a weed. In terms of the
overall performance, the Single Shot Detector (SSD) model showed 86% accuracy,
84% precision and 92% sensitivity.
With the images of weed taken at very close range, the confidence level of the
system becomes better. Application of this method in low land rice can reduce herbi-
cide wastage, its detrimental effect on the environment, increase yield, and a better
return of investment for the farmer.

References

1. Diao, X.: Economic importance of agriculture for sustainable development and poverty reduc-
tion: Findings from a case study of ghana Xinshen Diao, IFPRI, no. November, pp. 15–17
(2010)
2. Amin, M.S.M., Rowshon, M.K., Aimrun, W.: Paddy water management for precision farming
of rice. In: Current Issues of Water Management, InTech (2011)
3. Naresh, R.K., Kumar, Y., Chauhan, P., Kumar, D.: Role of precision farming for sustainability
of rice-wheat cropping system in western indo gangetic plains. Int. J. Life Sc. Bt. Pharm. Res.
1, 1–13 (2012)
4. Rodenburg, J., Johnson, D.E.: Chapter 4—weed management in rice-based cropping systems
in Africa. In: Sparks, D.L. (ed), Advances in agronomy, vol. 103, pp. 149–218. Academic Press
(2009)
5. Pulido-rojas, C.A., Molina-villa, M.A., Solaque-guzmán, L.E.: Machine vision system for
weed detection using 80, 124–130 (2016)
6. Ajala, A.S., Gana, A.: Analysis of challenges facing rice processing in Nigeria. J. Food Process
2015 (2015)
7. Hossain, M.M.: Recent perspective of herbicide: review of demand and adoption in world
agriculture. J. Bangladesh Agric. Univ. 13(1), 19–30 (2015)
8. Singh, V., Jat, M.L., Ganie, Z.A., Chauhan, B.S., Gupta, R.K.: Herbicide options for effective
weed management in dry direct- seeded rice under scented rice-wheat rotation of western Indo-
Gangetic Plains. Crop Prot. 81, 168–176 (2016)
9. Christensen S. et al.: Site-specific weed control technologies, pp. 233–241 (2009)
10. Burks, T.F., Shearer, S.A., Payne, F.A.: Classification of Weed species using colour texture
features and discriminant analysis 43(2), 441–448 (2000)
11. Watchareeruetai, U., Takeuchi, Y., Matsumoto, T., Kudo, H.: Computer vision based methods
for detecting weeds in lawns, pp. 287–296 (2006)
402 O. M. Olaniyi et al.

12. Tellaeche, A., Burgos-artizzu, X.P., Pajares, G., Ribeiro, A.: A vision-based method for weeds
identification through the bayesian decision theory. Pattern Recogn. 41(2), 521–530 (2018)
13. Faridi, H., Aboonajmi, M.: Application of machine vision in agricultural products. In: Proceed-
ings of the 4th Iranian International NDT Conference (IRNDT), (2017). Retrieved from http://
www.ndt.net/?id=21279
14. Liakos, K.G., Busato, P., Moshou, D., Pearson, S.: Machine learning in agriculture: A Review.
Sensors, 18(8), 2674 (2018). https://doi.org/10.3390/s18082674
15. Herrera, P.J., Dorado, J., Ribeiro, Á.: A novel approach for weed type classification based on
shape descriptors and a fuzzy decision-making method, pp. 15304–15324 (2014)
16. Lin, F., Zhang, D., Huang, Y., Wang, X., Chen, X.: Detection of corn and weed species by the
combination of spectral. Shape and Textural Features, pp. 1–14 (2017)
17. Yang, C.C., Landry, J., Ramaswamy, H., Ditommaso, A., Prasher, S.: Application of artificial
neural networks in image recognition and classification of crop and weeds. Can. Agric. Eng.
42(3) (2000)
18. Pulido, C., Solaque, L., Velasco, N.: Weed recognition by SVM texture feature classification in
outdoor vegetable crop images Reconocimiento de maleza por características de textura usando
2017, 68–74 (2017)
19. Khan, Y.N.: Weed detection in crops using computer vision. Cent. Robot. (2015)
20. Zheng, L., Fu, C., Zhao, Y.: Extend the shallow part of single shot multibox detector via
convolutional neural network. In: Proc. SPIE 10806, Tenth International Conference on Digital
Image Processing (ICDIP 2018). https://doi.org/10.1117/12.2503001.2018
21. Nejati, H., Azimifar, Z., Zamani, M.: Using fast fourier transform for weed detection in corn
fields. In: Systems, Man and Cybernetics, 2008. SMC 2008. IEEE International Conference
on, pp. 1215–1219 (2008)
22. Liu, H., Lee, S.H., Saunders, C.: Development of a machine vision system for weed detection
during both of off-season and in-season in broadacre no-tillage cropping lands. Am. J. Agric.
Biol. Sci. Publ. Online 9(92), 174–193 (2014)
23. Chang, C.L., Lin, K.M.: Smart agricultural machine with a computer vision-based weeding
and variable-rate irrigation scheme. Robotics 7(36), 1–17 (2018)
24. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C., Berg A.C.: SSD: Single shot
multibox detector. In: European Conference on Computer Vision (pp. 21–37). Springer, cham.
(2016)
25. Sun, C., Murphy, K.: Speed/accuracy trade-offs for modern convolutional object detectors
26. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception archi-
tecture for computer vision. In: 2016 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), Las Vegas, NV, pp. 2818–2826 (2016). https://doi.org/10.1109/CVPR.
2016.308
27. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.A.: The Impact of Residual Connections on
Learning, pp. 4278–4284
28. Szegedy, C., Com, S.G.: Batch Normalization: Accelerating Deep Network Training by
Reducing Internal Covariate Shift, vol. 37 (2015)
29. Vikhram, G.Y.R., Agarwal, R., Uprety, R., Prasanth, V.N.S.: Automatic Weed Detection and
Smart Herbicide Sprayer Robot, vol. 7, pp. 115–118 (2018)
30. Jenny.: Paperspace: Overview (2018)
31. Alzantot, M., Wang, Y., Ren, Z., Srivastava, M.B.: RSTensorFlow: GPU enabled tensorflow
for deep learning on commodity android devices. MobiSys. 7–12 (2017). https://doi.org/10.
1145/3089801.3089805
32. Aggarwal, N., Agrawal, R.K.: First and second order statistics features for classification of
magnetic resonance brain images. J. Signal Inform Process. (3), 146–153 (2012)
33. AIbinu, A.M., Salami, M. J. E., Shafie A.A.: Retina fundus image mask generation using
pseudo parametric modeling technique. IIUM Eng. J. 11(2), 163–177 (2010)
A Comparative Study Between Mixture
and Stepwise Regression to Model
the Parameters of the Composting
Process

Echarrafi Khadija, Ben Abbou Mohamed, Taleb Mustapha, Rais Zakia,

Ibtissam El Hassani, and El Haji Mounia

Abstract Composting is a good recycling operation for organic waste because it

reduces the volume of waste in landfills up to 41% compared to other biotechnolog-
ical solutions. In the present study, the Simplex-centroid design mixture is used to
optimize the mixing ratios of the substrate involved in the composting process, such
as green waste (GW), oil mill wastewater (OMW), waste solid oil mills (OMSW) and
poultry litter (PL). The main goal of this optimization is to define the final parameters
of the compost from the initial proportion of the different ingredients using a mathe-
matical forecasting model for an efficient composting process in terms of cost, quality
and timing. The performance of the response models; the organic matter content and
the carbon / nitrogen ratio as a function of the raw material flow rate were measured
using the deviation standard, the predicted R squared, the predicted sum of squares
(PRESS) and the correlation coefficient (R2). Indeed, they were respectively 2.26,

E. Khadija (B) · E. H. Mounia

Optimization of Industrial and Logistics Systems Team (OSIL), Laboratory of Research and
Engineering (LRI), Hassan II University, ENSEM, Casablanca, Morocco
e-mail: [email protected]
E. H. Mounia
e-mail: [email protected]
B. A. Mohamed
Faculty of Sciences, Laboratory of Natural Resources and Environment (NR&E), University Sidi
Mohamed Ben Abdellah, Taza, Morocco
e-mail: [email protected]
T. Mustapha · R. Zakia
Laboratory of Engineering, Electrochemistry, Modeling and Environment (LIEME), Faculty of
Sciences, University Sidi Mohamed Ben Abdellah, Fez, Morocco
e-mail: [email protected]
R. Zakia
e-mail: [email protected]
I. E. Hassani
Artificial Intelligence for Engineering Sciences (IASI) Team—ENSAM, My Ismaïl University,
Meknes, Morocco
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 403

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_28
404 E. Khadija et al.

42.71, 1441.71, 98.58 for the C/N and 8.45, 8.27, 6065.35, 87.02 for the organic
matter content. The best mixture satisfying the response constraints was 7.07% GW,
36.67% OMSW, 21.22% OMW and 35.03% PL.

Keywords Mixture design · Composting process · Optimization · Poultry litter ·

Olive mill wastewater · Olive mill solid waste · Green waste · Organic matter
content · C/N ratio

1 Introduction

In Morocco, the amount of waste in 2015 was 26.2 million tons and it will increase
to 30 million tons by 2030 [1]. This volume reveals the importance of reliable waste
management. According to certain waste management, green six-sigma, the best way
to minimize the volume of waste is to reduce its production at the source. The realiza-
tion of this proposal seems almost impossible, that’s why the search for some effective
alternatives for waste management is necessary, in particular for the substrate with a
high content of organic matter. According to the literature, there are huge biotechnolo-
gies that can safely handle this type of wastes, including the composting process [2].
This method is defined as a spontaneous biological decomposition and stabilization of
organic materials under a warm, moist and aerobic environment in order to produce a
valuable product that can be beneficial for the nutrition of agricultural land [3–5].The
choice of the composting process is justified by its ability to satisfy the three pillars
of sustainable development; it can reduce environmental impact, create employees
and make a sales product. However, the increasing consumer demands for sustain-
able products and the claims for environmental governance make the composting
process as the best choice for waste management. The final product takes a long time
to be produced (4 months) and its agronomic quality is variable due to the difficulty
of finding a compromise between biological, physical and chemical factors when
carrying out the process. Obtaining the final product in the best deadlines with the
required quality parameters need the realization of many experiences. Optimizing
the process remains the best way to achieve an efficient process that can take into
account the three constraints of optimization such as cost, quality and deadlines. To
do that, an optimization using a mixture design method is performed to obtain a
mathematical model of both responses of the process such as the Carbon/Nitrogen
(C/N) ratio and the content of organic matter (%OM). This document is structured
as follows. Section 2 presents a review of the literature on the terminology of the
composting process. Section 3 describes the materials and methods used in this study.
Section 4 presents an analysis and discussion of the results obtained. Finally, we draw
conclusions and perspectives for further research in Sect. 5.
A Comparative Study Between Mixture and Stepwise … 405

2 Literature Review

Olive Mill Wastewater OMW and Olive Mill Solid Waste (OMSW are non-valuable
by-products of olive presses which can cause serious danger to the environment due
to their high polyphenols content, but their physic-chemical composition maybe suit-
able for recycling operations. In this context, many waste treatment methods have
been proposed [6]. However, its valorization through the composting process could
be more reliable to be used as a soil amendment or to avoid erosion phenomena
because they are rich in the organic matter [3, 4]. The green waste is a by-product
resulting from the domestic or industrial vegetables processing and garden residues.
Its management by the composting process is advantageous in environmental and
financial terms. Poultry litter is produced from the mixture of poultry excrement,
spilled feed, feathers, and certain materials used in poultry litter such as wood shav-
ings. Its use as a substrate in the composting process could solve the problem of
waste disposal and produce a valuable product. The present study aims to opti-
mize the co-composting of the aforementioned substrates; OMW, OMSW, GW and
PL. The composting process is carried out in two temperatures ranges such as the
thermophilic phase which requires temperatures approximately up to 40 °C while
those thermophilic are reached in temperatures ranging from 45 to 70 °C [7]. The
composting process requires the existence of other conditions in terms of physico-
chemical parameters which must be carefully monitored as follows: moisture content
in the range of 50–60%, a pH value oscillating between 8–8.5 during the process but
the optimum is at 7, a C/N ratio between 10–15 in aerobic conditions, an organic
matter content between 16 and 22 and an electrical conductivity between 2.2 and 3
S·m − 1 [7–9]. According to the literature, the recycling of oil extraction processes
residues using composting technology is highly recommended [8], but they must be
co-composted in order to avoid the constitution of anaerobic conditions which can
hinder the normal evolution of the process due to their high moisture content [10].
Thus, the co-composting of two or more types of organic substrates is useful for
balancing the moisture content, the C/N ratio and avoiding the bad consequences
of composting a single material as follows: a long treatment time, a low degrada-
tion efficiency [11]. Following the practices applied in chemical, pharmaceutical and
food industries which use the mixture designs to optimize the formulations, recent
studies in the field of compost start using this statistical technique to optimize the
quantity of each individual substrate implied in the mixture [12]. The mixture design
is also used to analyze thoroughly the influence of different compounds constituting
the feedstock of the process and to predict the desired responses. Through previous
studies, the mixture design was used to predict C/N, moisture content and aeration
rate as response variables and the proportions of the substrate as input variables [11–
13]. This work aims to maintain the C/N ratio at the output of the process and to add
a new response variable which is the organic matter content. Therefore, the majority
of these previous studies used three input variables, but in our case study, four inputs
were used [14, 15].
406 E. Khadija et al.

3 Materials and Methods

a. Materials

The materials used in experiments were gathered from different sources: OMSW and
OMW are supplied by olive presses located in the region of Fez city, GW is collected
from the garden of the Faculty of Sciences “Dhar El Mahraz” and from certain
markets in the city of Fez. PL is provided by chicken breeders. The experiments
are carried out at the Engineering, Electrochemistry, Modeling and Environment
Laboratory located in the aforementioned Faculty. The ingredients collected are
well mixed according the proportions given by the design of the experiment. Thus,
the total of experiences was 23, each is put in a 10 kg bucket with monitoring in
terms of turning frequency and physicochemical parameters such as pH, temperature,
moisture content, organic matter content and C/N ratio to guarantee the favorable
conditions for the activities of microorganisms.
b. Mixture design

The mixture used in this study used OMSW, OMW, GW and PL as inputs and C/N
ratio and organic matter content as outputs of the process. The inputs of the mixture
design are the proportions of the ingredients implied in the mixture and not the
quantity of the mixture as depicted in Eq. 1. To this end, the Simplex-centroid design
with axial points called “pure blends” and 3 replicates of center point called “centroid
blend” is used. This experimental design is obtained using Minitab® software version
17. The preparation of the different blends was carried out by taking into account
the below constraint and a total load of 10 kg.
n
xi = 1 (1)
i=1

which.
n: refers to the number of input variables (in this case n = 4).
xi: refers to the proportion of each input variable.
c. Mathematical models

The main objective of a model based on mixture design is to predict the input variables
by varying the proportion of one, two or more compounds of the mixture. In this case,
the output variables are C/N ratio and organic matter content. Thus, the modeling
of the mixing surface with mathematical equations takes place. The use of these
mathematical equations makes the desired responses (C/N ratio and OM content)
depend only on the proportions of the ingredients involved in the blend regardless
of its quantities. The main distinction between the mixture design and the other
experiment designs is the absence of the constant term in the mixture design models.
In this comparative study, the models used are Scheffé canonical polynomials as
shown in Eqs. 2, 3, 4, 5 [16].
A Comparative Study Between Mixture and Stepwise … 407

Linear
n
Y = βi Xi (2)
i=1

Quadratic
n n−1 n
Y = βi Xi + βi j Xi X j (3)
i=1 i=1 j=i+1

Full cubic
n n−1 n
Y = βi Xi + βi j Xi X j
i=1 i=1 j=i+1
n−1 n
+ γ i j Xi X j(Xi − X j) (4)
i=1 j=i+1

Special quartic
n n−1 n
Y = βi Xi + βi j Xi X j
i=1 i=1 j=i+1
n−1 n
+ γ i j Xi X j(Xi − X j)
i=1 j=i+1
n−2 n−1n
+ βi jk Xi2X j X k
i=1 j=i+1 k= j+1
n−2 n−1 n
+ βi jk Xi X j2X k
i=1 j=i+1 k= j+1
n−2 n−1 n
+ βi jk Xi X j X k2 (5)
i=1 j=i+1 k= j+1

Which.
Y is the predicted value of the response,
X is the input variables; in this case, they are the proportions of substrate.
β and γ are the polynomial coefficients.

4 Results and Discussion

a. Initial characterization of composting materials

Table 1 shows the organic matter content, carbon and nitrogen contents and the C/N
ratio for each substrate used in the process. OMSW has the highest organic matter
content, followed by green waste. Green waste and poultry litter are considered a
good supplier of nutrients for compost because of their high nitrogen content, while
OMSW and green waste are considered a good source of organic carbon content.
The choice of C/N ratio and organic matter content as response parameters of the
408 E. Khadija et al.

Table 1 Initial ratios of organic matter content and C/N

Substrats Organic matter (%) Carbon content (%) Nitrogen content (%) C/N ratio
Poultry Litter 30,65 16,11 1,17 13,77
OMW 76,70 7,64 0,21 36,38
Green waste 80,12 47,83 1,45 32,98
OMSW 95,18 53,2 0,95 56

composting process because they are the main indicators of the quality of the final
compost.
b. Comparative study for the best fitting model

The samples were prepared according to the mixture design matrix. Indeed, the
proportions of the inputs involved in the mixture were generated randomly by
Minitab software 17. A simplex centroid design was used because there are no linear
constraints between the input variables. The design matrix is composed of 23 tests
with a triple of the central point in order to check the variability of the data (Table 2).
Responses variables such as C/N ratio and organic matter content were fitted to linear,
quadratic, full cubic and special quartic models according to mixture and stepwise
regression. To verify the goodness of the adjusted models, various features of the
model were generated such as the standard deviation (S), the correlation coefficient
(R-squared), the predicted R-squared, the adjusted R-squared and the predicted sum
of squares (PRESS). According to the literature, the best fitting model must have, on
the one hand, a small deviation of standard (S) and a low predicted sum of squares,
and a high value of the following parameters: Predicted R-squared, the coefficient
of correlation (R-squared) and adjusted R-squared on the other side [14]. Following
these guidelines, the appropriate model of C/N ratio and organic matter content is a
full cubic model using stepwise regression as shown in the Tables 3 and 4.
c. Mathematical model and analysis of variance

Equation (6) and (7) present the best mathematical models which can predict the
response variable with a low margin of error; the margin of error depends on the
quality of the features of the model. The variable x1, x2, x3 and x4 are the four
input variables which each represents, in the order, Green waste, Olive Solid Mill
Wastewater, Olive Mill Wastewater and Poultry Litter. Yom and YC/N respectively
represent the initial organic matter and C/N ratio of the mixture. The numbers in the
equations refer to the regression coefficients as depicted in Table 5.

Y O M = 74, 8 X 1 + 91, 7 X 2 + 69, 2 X 3 + 22, 8 X 4 − 179, 9 X 1X 2

− 144, 1 X 1X 3 − 228, 2 X 2X 3 − 80, 5X 2X 4 + 617, 9 X 1X 2X 3
(6)
A Comparative Study Between Mixture and Stepwise … 409

Table 2 Runs of mixture design and responses values

Runs Input variables (xij) Response variables (Y)
GW (%) OMSW (%) OMW (%) PL (%) C/N OM (%)
1 0,125 0,125 0,125 0,625 12,178 36,07
2 0,500 0,000 0,000 0,500 12,120 37,65
3 0,333 0,333 0,333 0,000 12,178 39,79
4 1,000 0,000 0,000 0,000 32,990 80,12
5 0,000 0,000 0,500 0,500 13,130 31,52
6 0,250 0,250 0,250 0,250 12,148 39,15
7 0,500 0,000 0,500 0,000 12,160 36,84
8 0,333 0,333 0,000 0,333 12,156 40,02
9 0,250 0,250 0,250 0,250 12,148 39,06
10 0,000 0,333 0,333 0,333 13,146 33,02
11 0,000 0,000 1,000 0,000 36,380 76,70
12 0,500 0,500 0,000 0,000 12,100 36,88
13 0,625 0,125 0,125 0,125 12,170 40,05
14 0,125 0,625 0,125 0,125 13,145 35,92
15 0,000 0,000 0,000 1,000 13,770 30,65
16 0,333 0,000 0,333 0,333 12,190 40,00
17 0,000 0,500 0,000 0,500 12,150 29,93
18 0,125 0,125 0,625 0,125 13,160 32,46
19 0,000 1,000 0,000 0,000 56,000 95,18
20 0,250 0,250 0,250 0,250 12,148 38,92
21 0,000 0,500 0,500 0,000 13,177 22,27

Table 3 Summary of model features for C/N ratio

Model Terms Standard Regression Predicated Adjusted Predicated
fitting deviation R Sq (%) regression regression Sum of
method (S) R-Sq(pred) R-Sq(adj) squares
(%) (%) (PRESS)
(%)
Mixture Linear 10,8975 19,78 0 5,63 5666,6
regression Quadratic 4,5458 90,97 0 83,58 4658,89
Full cubic 2,5319 98,98 0 94,91 27,692,1
Special 2,5448 98,97 0 94,85 51,593,8
quadratic
Stepwise Linear 10,8975 19,78 0 5,63 5666,6
Quadratic 4,8222 87,99 0 81,52 2557,51
Full cubic 2,2634 98,58 42,71 95,93 1441,73
Special 4,5464 90,14 23,22 83,57 1932,4
quadratic
410 E. Khadija et al.

Table 4 Summary of model features for organic matter content

Model Terms Standard Regression Predicated Adjusted Predicated
fitting deviation R Sq (%) regression regression Sum of
method (S) R-Sq(pred) R-Sq(adj) squares
(%) (%) (PRESS)
(%)
Mixture Linear 17,7221 19,25 0 5,01 14,719,9
regression Quadratic 10,2457 82,54 0 68,25 24,741,6
Full cubic 5,43,332 98,21 0 91,07 118,678
Special 4,95,035 98,52 0 92,59 191,983
quadratic
Stepwise Linear 17,7221 19,25 0 5,01 14,719,9
Quadratic 9,87,118 80,84 0 70,53 11,026,2
Full cubic 8,45,777 87,02 8,27 78,36 6065,35
Special 8,31,027 88,51 1,75 79,11 6496,64
quadratic

Table 5 Coefficient of regression and analysis of variance

C/N OM
Coefficient value p-value Coefficient value p-value
Coefficients X1 32,5 − 74,8 −
X2 55,7 − 91,7 −
X3 35,8 − 69,2 −
X4 13,1 − 22,8 −
X1X2 −129,1* 0,000 −179,9 0,001
X1X3 −84,1* 0,000 −144,1 0,003
X1X4 −39 0,007 − −
X2X3 −131,6* 0,000 −228,2* 0,000
X2X4 −90* 0,000 −80,5 0,042
X3X4 −41,8 0,005 − −
X1X2X3 231,4 0,013 617,9 0,034
X1X2X4 175,71 0,040 − −
X2X3X4 186,4 0,032 − −
X1X3X4 −139,1 0,016 v −
Features of the S 2,26 8,46
model fitting R-Sq 98,58 87,02
R-Sq(pred) 42,71 8,27
R-Sq(adj) 95,93 78,36
PRESS 1441,73 6065,35
A Comparative Study Between Mixture and Stepwise … 411

Y C/N = 32, 5 X 1 + 55, 7 X 2 + 35, 8 X 3 + 13, 1 X 4 − 129, 1 X 1X 2

− 84, 1 X 1X 3 − 39 X 1X 4 − 131, 6 X 2X 3 − 90 X 2X 4 − 41, 8 X 3X 4
− 41, 8 X 3X 4 + 231, 4 X 1X 2X 3
+ 175, 71X 2X 4 + 186, 4 X 2X 3X 4 − 139, 1X 1X 3X 4 (7)

According to the analysis of variance (Table 5), all 3rd degree terms and the two
2rd degree terms (X1X4, X3X4) of the C/N model are significant with a p-value
<0.05, which means that composting poultry litter and green waste humidified by
olive mill wastewater is recommended. In the organic matter model, all terms are
significant with a p-value <0.05 except the interaction X2X3 which is significant with
a p-value <0.001. Following the results of both models (C/N and organic matter),
olive solid mill wastewater should be removed.
The interactions marked by (*) are significant at 0.001 level, the rest of interactions
are significant at 0.05.
The accuracy of both models is corroborated by analysis of variance in which all
model terms are significant at 0.05.
c. Mixture proportion optimization

The optimized combination is obtained using the response optimization tool of the
Minitab software. The targets set for the two responses are respectively 12 and 30%
for the C/N ration and the organic matter, their constraints are 10 < C/N < 15 and 20
< %OM < 60. Thus, the combination of optimization of the feedstock at the start of
the process which satisfies the aforementioned constraints is 7.07% of green waste,
36.67% of wastewater from olive oil mill, 21.22% of waste oil mill oil and 35.03% of
poultry litter as shown in Fig. 1. This combination provides an organic matter content
of 30.0493 and a CN ratio of 12.0142. The desirability of the predicted responses is
respectively 0.9996 and 0.9993 for the C/N ratio and the organic matter content. The
overall desirability is 0.9995, which indicates that the optimal initial organic matter
content and the C/N ratio have been satisfied and are within acceptable limits.
d. Validation of the models

The experimental data have been adjusted in Eq. (6) and (7). To ensure the goodness
of the two models, the C/N ratio and the organic matter content, we replaced the
optimal proportions in the models which are 7.07% of Green waste, 36.67% of
OMSW, 21.22 of OMW and 35.30 of poultry litter. The results were found to be
11.64% C/N ratio and 30.06% organic matter content. The under or over estimation
of the two response parameters is due to the low values of predicted R-squared which
are respectively 42.71 and 8.27 for the C/N ratio and the organic matter content.
412 E. Khadija et al.

Fig. 1 Optimization combination plot

5 Conclusion

A satisfactory overall desirability of 0.9985 near to one has made it possible to define
the optimal formulation within the experimental domain. The adjustment model for
C/N ratio and %OM revealed fairly good satisfaction due to their low predicted R-
squared coefficients which are respectively 42.71 and 8.27 for C/N ratio and %OM.
The R-squared (R2) of 98.5 and 87.02, in order, for C/N ratio and %OM reveals
that the models selected correspond very well to the data. Our study would like to
emphasize the lack of use of the mixture design when the input parameters are greater
than 3 in the field of compost formulations and it tried to fill this gap. Our future
works consists of a comparative study of current results with other advanced methods
such as artificial neural networks and thus choose the most accurate model.

Acknowledgements The work was supported by funds from the Moroccan Ministry of Energy,
Mines, Water and Environment. The experiment part was achieved by the Laboratory of Engineering,
Electrochemistry, Modeling and Environment located in the faculty of science “Dhar Mehraz” in
Fes city.
A Comparative Study Between Mixture and Stepwise … 413

References

1. Le matin.: La stratégie nationale de réduction et de valorisation des déchets prévoit un gisement

total des déchets à l’échelle nationale de 39 millions de tonnes à l’horizon 2030 (2019)
2. Wang, L.K., Ivanov, V., Tay, J.-H., Hung, Y.-T.: Environmental biotechnology, vol. 10. Springer
Science & Business Media (2010)
3. Majbar, Z., Rais, Z., El Haji, M., Ben Abbou, M., Bouka, H., Nawdali, M.: Olive mill wastewater
and wine by-products valorization by co-composting. J. Mater. Environ. Sci.8(9), 3162–3167
(2017)
4. Majbar, Z. et al.: Co-composting of olive mill waste and wine-processing waste: an application
of compost as soil amendment. J. Chem. 2018 (2018)
5. Echarrafi, K., El Harhouri, H., Ben Abbou, M., Rais, Z., El Hassani, I., El Haji, M.: Mixture
design formulation for optimized composting with the perspective of using artificial intelligence
optimization algorithms. J. Appl. Sci. Environ. Stud. 1(2), 1–2
6. Kapellakis, I.E., Tsagarakis, K.P., Crowther, J.C.: Olive oil history, production and by-product
management. Rev. Environ. Sci. Bio/Technology 7(1), 1–26 (2008)
7. Miller, F.C.: Composting of municipal solid waste and its components. Microbiol. Solid Waste,
115–154 (1996)
8. Filippi, C., Bedini, S., Levi-Minzi, R., Cardelli, R., Saviozzi, A.: Cocomposting of olive oil
mill by-products: chemical and microbiological evaluations. Compost Sci. Util. 10(1), 63–71
(2002)
9. Zhu, N.: Effect of low initial C/N ratio on aerobic composting of swine manure with rice straw.
Bioresour. Technol. 98(1), 9–13 (2007)
10. Khalil, A., Domeizel, M., Prudent, P.: Monitoring of green waste composting process based
on redox potential. Bioresour. Technol. 99(14), 6037–6045 (2008)
11. Kumar, M., Ou, Y.L., Lin, J.G.: Co-composting of green waste and food waste at low C/N ratio.
Waste Manag. 30(4), 602–609 (2010)
12. Eriksson, L., Johansson, E., Wikström, C.: Mixture design—design generation, PLS analysis,
and model usage. Chemom. Intell. Lab. Syst. 43(1–2), 1–24 (1998)
13. Soares, M.A.R., Andrade, S.R., Martins, R.C., Quina, M.J., Quinta-Ferreira, R.M.: Organic
biowastes blend selection for composting industrial eggshell by-product: experimental and
statistical mixture design. Water Sci. Technol. 65(11), 1939–1945 (2012)
14. Abdullah, N., Chin, N.L.: Simplex-centroid mixture formulation for optimised composting of
kitchen waste. Bioresour. Technol. 101(21), 8205–8210 (2010)
15. Soares, M.A.R., Quina, M.J., Quinta-Ferreira, R.: Prediction of free air space in initial
composting mixtures by a statistical design approach. J. Environ. Manage. 128, 75–82 (2013)
16. Smith, W.F.: Experimental design for formulation. SIAM (2005)
Deep Learning Based Sponge Gourd
Diseases Recognition for Commercial
Cultivation in Bangladesh

Tahmina Tashrif Mim, Md. Helal Sheikh, Sadia Chowdhury,

Roksana Akter, Md. Abbas Ali Khan, and Md. Tarek Habib

Abstract Sponge gourd, called as “ ” in Bangladesh, is scientifically known as

Luffa cylindrical that belongs to Cucurbitaceae family. Sponge gourd or luffa gourd
is one of the most easily found vegetable in Bangladesh. It is an edible wild vegetable
that the plant can be seen anywhere around us during late summer till late autumn
in Bangladesh. Cooked sponge gourd or the curry is a bit sweetish in taste. Even
though sponge gourd is kind of a wild vegetable plant, in recent time a lot of people
are cultivating it in the countryside thinking of profit and commercial production since
there is a market demand for it. Despite of having every opportunity to commercial
benefit most of the farmers neglect the issue of insects and diseases attack on the
plant resulting on huge loss in the business. Also lack of proper knowledge of related
diseases, advance technology and trustable source the farmers lag behind the diseases
detection process to use pesticides or different methods of reducing diseases attack.
If necessary steps can be taken to prevent the insects and diseases attack at the very
beginning of cultivation, then the profits will increase as the crop yields increase.
This research paper attempts to detect the leaf and flower diseases of sponge gourd
using Convolutional Neural Network (CNN) and image processing techniques. CNN
and image processing are one of the most recently introduced technologies using in

T. T. Mim (B) · Md. H. Sheikh · S. Chowdhury · R. Akter · Md. A. A. Khan · Md. T. Habib
Department of Computer Science & Engineering, Daffodil International University, Dhaka,
Bangladesh
e-mail: [email protected]
Md. H. Sheikh
e-mail: [email protected]
S. Chowdhury
e-mail: [email protected]
R. Akter
e-mail: [email protected]
Md. A. A. Khan
e-mail: [email protected]
Md. T. Habib
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 415

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_29
416 T. T. Mim et al.

the agriculture sector in Bangladesh ensuring highest accuracy. This system will take
leaf images as input and after examining them healthy or detected diseases will be
shown as output which has diffrent true and false values for different diseases and
reached to the average accuracy of 81.52%.

Keywords Sponge gourd · Commercial · Cultivation · CNN · Image processing

1 Introduction

The people of Bangladesh are always health concious. Different types of vegetables
get priority in the meal list. With the increase in demand, the supply of vegetables
are in increasing in the market. Seasonal vegetables are sharing place with yearlong
vegetables. According to the report of The Daily Star newspaper, the annual demand
for the vegetables in Bangladesh is 13.25 million metric tons. But the supply of
vegetables in the market is only 3.73 million metric tons per year covering only one
third of the total demand [1]. Millions of tons of vegetables get wasted because of
diseases attack and lack of knowledge of reducing them. To overcome the production
constrains, if we can control the diseases and insects attack before the yield comes
then the production rate will boost up in general.
Sponge gourd is one of the most common choice of consumers in the market. Tons
of sponge gourd or Luffa Gourd are marketed daily. The price varies from place to
place and the rate of sponge gourd per kilo varies from 30 to 60 BDT depending on
weather. Sponge gourd is a creeping plant crop having low yeild. The green fruits
are used as vegetables and ripe fruits can be used for cleaning kitchen utensils and
also while bathing. For the cultivation of sponge gourd, soil enriched in biomass and
sunny weather are preferred [2].
Sponge gourd is a manifold rich vegetable. It provides natural insulin which
is very useful for diabetic healing and also helps to reduce and control excessive
cholesterol and body fat. To recover from jaundice, sponge gourd curry is very
effective. It is loaded with calcium, phosphorous, iron, fibers and vitamin-B. Since
Sponge gourd is a member of Cucurbitaceae family, it is enriched in anti-oxidant
and has antimicrobial, anticancer, cardioprotective and hypolipidemic qualities [2].
Two species of sponge gourd is found in Bangladesh, one is cultivated to eat as
vegetable and other one is found and known as wild sponge ground having non edible
bitter taste which is only used as ripe fruit to produce bath accessories. There are
many types of diseases of sponge gourd. To reduce them dry and dead leaves and
attacked leaves should be collected and floated in the ground. Otherwise there is a
heavy chance of diseases spreading amongst healthy plants. This research tends to
detect those leaf and flower diseases using latest technology to combine agriculture
sector with computer science. Here we are going to detect two leaf disease class
and one flower disease class along with two healthy leaf and flower class of sponge
gourd.
Deep Learning Based Sponge Gourd Diseases … 417

CNN cooperate with the pretrained models. There are more than 15 various CNN
models to pre-train the dataset according to the research requirements and subjects. In
this research work we have used AlexNet to pre-train our raw dataset for detecting the
diseases of sponge gourd leaf and flower diseases. About 8 class including 2 healthy
class of leaf and flower and 6 disease class has been taken in notice for experimenting
our proposed model on collected raw dataset. They are: Downy mildew, Fungal
Disease, Alternaria blight, Angular leaf spot, Anthracnose, Healthy leaf, Unhealthy
flower and Healthy flower.

2 Literature Review

This literature review section of this research paper is going to present the near past
related works done by some researchers on vegetable diseases prediction. We have
followed and studied their work to understand the processes and methods expressed
by them.
M. TURKOGLU, D. HANBAY [3] have worked on plant disease and pest detec-
tion using deep learning-based features. Their study in this research evaluated the
performance results of nine powerful architecture of deep neural network for plant
disease prediction. They have obtained features using deep feature extraction and
classified the work by support vector machine (SVM), Extreme Learning Machine
(ELM), and K-nearest neighbor (KNN) methods. CNN model-based transfer learning
is used to pretrain their system in their research work. The ResNet50 model and SVM
classifier produced the highest accuracy among these with an accuracy of 97.86%
and obtained the best accuracy for the AlexNet model of 95% using SVM classifier.
The highest accuracy for the VGG16 model was 95% by using the SVM classifier.
The highest accuracy for vgg19 model was 94.74% using ELM classifier.
M. Akila, P. Deepan [4] presented their work on deep-learning-based approach to
detect leaf diseases in many different plants using image of those plant leaves. The
goal was to find and develop the more suitable deep-learning methods that can easily
detect the diseases only from plant images. Three main friendly detectors: Region-
based Fully Convolutional Network (R-FCN), Faster Region-based Convolutional
Neural Network (Faster R-CNN), and Single shot Multibox Detector (SSD) have
been used in their research.
Sharada Prasanna Mohanty, Davia Hughes, and Marcel Salathe [5] proposed a
deep Convolution Neural Network to identify 14 crop species and 26 of their diseases.
Their trained model achieved an accuracy of 99.35% during the feasibility test.
Santosh Adhikari et al. [6] aimed to classify and detect the plant disease automat-
ically for the tomato plant. Image processing, adjusting image RIO, CNN have been
applied on their research work and an overall accuracy of 89% has been achieved on
their plant village dataset.
418 T. T. Mim et al.

Nadeem Ahmed et al. [7] predicted the potato late blight disease based upon
environmental factor in Faisalabad, Pakistan. A disease predictive model was devel-
oped using two years data of PLB disease severity and epidemiological factor. Their
proposed model explained up to 80% disease variability.
M. Donatelli et al. [8] proposed a five-stage roadmap to improve the simulation
of the impacts caused by plant disease and pest. Improve the quality and avail-
ability of data for model inputs, improve the quality and availability of data for
model evaluation, Improve the integration with crop model, improve the processes
for model evaluation and develop a community of plant pest and disease modelers
were proposed and staged by them.
With the invention of new technology, the processes to detect the plant diseases are
changing for better agricultural and economic growth. In recent year, deep learning,
which is especially widely use in image processing offers many new applications
related to precision agriculture. CNN have succeed to receive the trust and acceptance
of agricultural specialist through the whole world. In this research of ours, we have
tried to employ CNN to detect the leaves and flower diseases of sponge gourd in
Bangladesh with an accuracy of 81.52% on collected raw dataset.

3 Proposed Methodology

The purpose of this research is to commence a new concept of identifying the leaf
diseases of Sponge gourd in order to help the local farmers of Bangladesh to cultivate
and vend them commercially aiming to produce more thrive. For any kind of cultiva-
tion, healthy plant or farm is a must. Soil, weather, insects and diseases, cultivation
methods are the main barriers for any flatten form of cultivation. Thus, much consid-
eration of ranchers and time to time expert’s instruction is needed to be paid. But
the cultivation processes and methods for Sponge gourd is mostly neglected in our
country by labeling it a wild vegetable. Maximum cultivate it for hobby or jut to fulfill
their personal need. To cultivate it professionally, it is a matter of cost, carefulness
and time. In any case of insects or diseases attack ranchers need to seek the expert’s
recommendation which is a bit costly for the native ranchers to appoint an experi-
enced agriculture officer. Also, the exposure procedure is a lengthy process. Artificial
Intelligence saves us from this. Specially image processing system and CNN solve
all these hurdles with less effort, low cost and expedition in a more professional way.
Our proposed method woks taking the insect or disease attacked leaf image as an
input. After inspecting and comparing the input image with the pre-trained machine
an explanatory output is offered by the system which detects the disease name.
Only by clicking the insects attacked or diseased leaf’s picture and input them into
the system we have created, an informational output with the name of the disease
will be shown to the farmers. So, CNN can be the easiest procedure for the native
ranchers to perceive the disease name and provide the necessary pesticides in time
from escalating the diseases to the fruit.
Our proposed methodology is shown below in Fig. 1.
Deep Learning Based Sponge Gourd Diseases … 419

Raw Data Selected

Data Selection
Collection Images

Image
augmentation

Train Apply CNN Prepared

model algorithms Data

Fig. 1 Proposed methodology

3.1 Dataset

Dataset is a collection of essential text, image, numerical, questionnaires or any form

of data on which the proposed is applied on and executed by the researchers. So, to
acquire an accurate output and to determine or declare a numerical or analytical result
dataset is a must prerequisite for any research work. In this research work, CNN and
image processing methods are intended to administrate on our individually collected
raw dataset.
We have assembled about 6,000 leaf and flower images from various sponge gourd
with the help of android phones and digital cameras to distinguish the diseases to
carry out the proposed idea. To attain the best accuracy by using CNN, vast amount
of data is a must. So, we have built our own dataset of sponge gourd leaves and flower
according to the need of our proposed model.
In this research work we plan to detect 4 leaf diseases of sponge gourd along with
a healthy class and also the delicate flowers as well as the fresh flowers. We have
allowed.jpg,.gif,.bmp image formats in this research work.
Downy mildew.
Alternaria blight.
Anthracnose.
Healthy leaf.
Unhealthy flower and.
Healthy flower.
The sample dataset of this research is shown in Fig. 2.

4 Experimentation

This research section thoroughly describes the research methods, processes, models,
algorithms and equations we have used.
420 T. T. Mim et al.

Fig. 2 Collected dataset

4.1 Image pre-Processing

Our collected raw dataset has been breached into two individual section asserted as
test and train. We have put 20% image data in the test folder and 80% image data in
the train folder.
All assembled raw images have been processed and resized into (256*256) pixels
without compromising their aspect for pre-training them bestow to our proposed
method.

4.2 Model Architecture

CNN model architecture has been enforced to achieve substantial accuracy. CNN
model architecture is a vast part of deep learning algorithm in which input images
are trained and tested by passing through a sequential series of convolutional layers
with paramount filters. Minimal number of image processing is required for the
fasten computation process to work with CNN algorithm. In this research, CNN
algorithms have evaluated the visual imageries for assuring higher qualities to process
the images.
A CNN model always endorsed an input and an output layer. Our proposed model
consumed multi-layers CNN. RGB (Red, Green, Blue) color model is consumed for
the images in this research.
Since, our proposed CNN model used multi-layer, there are multiple layers of
convolutional layer as convolutional 2D and multiple ReLU activation function is
also used in this proposed methodology. Maxpooling 2D is used to convert the image
matrix in a shorten matrix. To activate the layers batch normalization is used. The
first layer contains (256,256,3) input shaped, (64 × 64) filters, (8 × 8) kernel, (1 ×
Deep Learning Based Sponge Gourd Diseases … 421

1) strides and same “padding”. Every layer followed the same formation.

ReLU (X ) = M AX (0, X ) (1)

Dense layer has followed the flatten layer using 0% of dropout. Last output layer
has used 17 units with activation softmax and sigmoid.

ez j
σ (Z ) = k f or j1, . . . . . . , k . . . . . . . . . . . . (2)
zk
k=1 e

1
ϕ(Z ) = ................ (3)
1 + e−z

The arrangement of the proposed CNN model is shown below in Table 1.

In his research, Adam algorithm has been used to reduce the learning rate. The
value of Adam is taken as 0.001.
The model architecture of our proposed CNN [2] model is given below in Fig. 3.

Table 1 Our proposed model

Layer (type) Output shape Parameters
conv2d_1 (Conv2D) (None, 256, 256, 64) 12352
conv2d_2 (Conv2D) (None, 256, 256, 64) 262208
max_pooling2d_1 (MaxPooling2) (None, 128, 128, 64) 0
conv2d_3 (Conv2D) (None, 128, 128, 32) 51232
conv2d_4 (Conv2D) (None, 128, 128, 32) 25632
max_pooling2d_2 (MaxPooling2) (None, 64, 64, 32) 0
batch_normalization_1(Batch) (None, 64, 64, 32) 128
conv2d_5 (Conv2D) (None, 64, 64, 16) 12816
conv2d_6 (Conv2D) (None, 64, 64, 16) 6416
max_pooling2d_3 (MaxPooling) (None, 32, 32, 16) 0
conv2d_7 (Conv2D) (None, 32, 32, 8) 1160
conv2d_8 (Conv2D) (None, 32, 32, 8) 584
max_pooling2d_4 (MaxPooling) (None, 16, 16, 8) 0
batch_normalization_2 (Batch) (None, 16, 32, 8) 32
flatten_1 (None, 2048) 0
dense_1 (Dense) (None, 512) 1049088
422 T. T. Mim et al.

Fig. 3 Model architecture of our proposed model

4.3 Image Enhancement

The dataset is artificially inflated by the image enhancement method. In this research
image enhancement is used to segregate the raw images into several augment. Image
enhancement is used as given below in this research work:
Lessening complexity exhibition of an image.
Facilitating and fluctuating the representation of an image.
Generating a considerable dataset by producing various shapes and angle from
an image.
For assuring the highest accuracy, our proposed model used image rotation in
range with the value of 40, width shift and height shift in the range with the value of
0.2, rescale as 1. /155, shear and zoom in the range with the value of 0.2. Horizontal
flip as True and fill mode as nearest is augmented in the proposed model for finding
the highest accuracy shown in Fig. 4.

(a) Original (b) rotation (c) width shift (d)height shift (e) zooms

(f) Original (g) rotation (h)width shift (i)height shift (j) zooms

Fig. 4 Image enhancement

Deep Learning Based Sponge Gourd Diseases … 423

4.4 Layer Visualization

Layer visualization of a CNN model displays the feature maps that are the output
given by the CNN model layers as well as different pooling layers and batch normal-
ization. It shows the data structure processes inside a CNN network. Layers of a CNN
can be obtained in different shapes and sizes such as circle, square, trigonal etc. [9].
In this research work, we have used square layer visualization by adding MaxPool
layer, Batch normalization, SoftMax, Dense along with CNN layers for better visual
displays. Figure 5. below shows every working CNN layers inside the network with
a real time visual outcome.

5 Result and Discussion

This section briefly exposes the practical findings of the studied dataset and the whole
outcome of the complete research work. Both graphical and numerical or statistical
explanation of our established model is disclosed in this section. How did the model
work during the research process and their output explanations are described without
any biasness.

Fig. 5 Layer visualization

424 T. T. Mim et al.

5.1 Statistical Analysis

The section of statistical analysis explains how the iterative CNN model worked after
every epoch with accuracy, loss and error rate with more detailed numeric values.
The very first accuracy rate acheived after the successful run of first epoch is 43.79%
and loss is 1.829. With more successful train epoch the accuracy rate improved and
the learning rate achieved after two successful completion of train step is 0.0002.
Following the similar procees, 6.78 loss and 78.40% of accuracy rate have been
accuired after the completion of seven successful run with a learning rate of 4.000.
An accuracy of 80.30% and 0.5952 loss with 1e-06 redueced learning rate have been
accuried after eleven successful training run. Similarly after thirty successful epoch
we have achieved an accuracy of training 81.52% and loss 0.5715.

5.2 Accuracy Graph

Statistical explanations of a research are usually condoned by the accuracy graph.

Precession of the accomplished numeric values or statistical explanations of an output
is checked by the model accuracy graph. The accuracy graph of this research model
displays that the uppermost graph of training and validation loss faced no issues of
underfitting or overfitting. Following the first graph the lower graph is displaying the
and training and validation accuracy after the closing epoch and it also does not has
no issues of noises. The accuracy graph of the proposed model is shown below in
Fig. 6.

Fig. 6 Accuracy graph

Deep Learning Based Sponge Gourd Diseases … 425

Alternaria 67 0 0 0 0 0
Blight

Anthracnose 33 41 7 0 0 5

Downy 3 0 91 0 0 0
Mildew

Healthy 2 0 0 44 29 5

Healthy 0 0 0 11 59 0
Flower

Unhealthy 5 16 1 0 0 58
flower

Alternaria Anthracnose Downy Healthy Healthy Unhealthy

Blight Mildew Flower flower

Predicted label
Accuracy = 0.7447, missclass =0.2453

Fig. 7 Confusion matrix

5.3 Confusion Matrix

Confusion matrix portrays the performance of the proposed classification model.

The confusion matrix of this research work describes how accurately accurate our
proposed model works. This is an extreme table of precision indicating the perfor-
mance of our proposed model on test dataset from where we obtain the True values.
Accuracy, Precision, Recall, F1 score can be easily uncovered by reviewing a confu-
sion matrix of any research. The proposed model of this research is working smoothly
as all the diagonal values of (4 × 4) matrix is larger than other values. The confusion
matrix shown in Fig. 7.

5.4 Classification Table

The classification report of our proposed methodology according to these formulas

is given in Table 2.
426 T. T. Mim et al.

Table 2 Performance of classification

Disease Precision Recall F1-score Support
Alternaria blight 0.61 1.00 0.76 67
Anthracnose 0.72 0.48 0.57 86
Downy mildew 0.92 0.97 0.94 94
Healthy leaf 0.80 0.55 0.65 80
Healthy flower 0.67 0.84 0.75 70
Unhealthy flower 0.85 0.72 0.78 80
accuracy 0.75 477
Macro avg 0.76 0.76 0.74 477
Weighted avg 0.77 0.75 0.75 477

6 Conclusion

The most applicable chalange for any new technology is “ acceptance”. An ingenious
idea has been introducing in this research to develop, improve and upgrade our raw
food industry both locally and globally. This research targeted to help the native
ranchers by enlightening the biggest problem of disease detection in commercial
sponge gourd cultivation. Nowadays advanced countries are implementing CNN
and image processing algorithms to make different disease prediction software and
application for vegetables and crops. In this research our proposed model of CNN
and image processing has exhibited its ability to detect the diseases ensuring the most
accurate peocess. Thus, we are hoping that native ranchers as well as agricultural
experts in Bangladesh will have faith in our system and broaden the agriculture sector
with the help of latest technology. Since mobile technology is way more available in
our country and most of the rural ranchers feel ease with the mother tounge “bengali”,
following this path we plan to manufacture a software and an android application
in Bengali language for the rural farmers having no knowledge in other language
except Bengali. We are hoping to see the sponge gourd name in export vegetable list
in near future over its commercial cultivation.

References

1. Correspondent,S.: Vegetable yield inadequate. The Daily Star, 2019. [Online]. Avail-
able: https://www.thedailystar.net/city/annual-vegetable-production-only-one-third-demand-
1519954. [Accessed: 10- Oct- 2019]
2. Author, B.: ?, Inshaschool.com. (2019). [Online]. Available:
https://www.inshaschool.com/2015/12/blog-post.html. [Accessed: 02- Nov- 2019]
3. Muammer TURKOGLU, Davut HANBAY Plant disease and pest detection using deep learning-
based features. Turk J ElecEng& comp sci 27, 1636–1652 (2019)
4. Akila, M., Deepan, P.: Detection and classification of plant leaf diseases by using deep
learning algorithm. International Journal of Engineering Research & Technology (IJERT)
Deep Learning Based Sponge Gourd Diseases … 427

nceISSN:2278–0181, ICONNECT-2018 Conference Proceedings

5. Mohanty, S.P., Hughes, D., Salathe, M.: Using deep learning for image-based plant disease
detection. This Manuscript was on April 15 2016
6. Adhikari, S., Shrestha, B., Baiju, B., Er. Saban Kumar K.C.: Tomato plant diseases detection
system using image processing. KEC Conference, ISBN 978–9937–0–4872–9, September 27,
2018
7. Ahmed, N., Aslam Khan, M., Ahmed Khan, N., Asif Ali, M.: Prediction of potato late blight
disease based upon environmental factor in Faisalabad, Pakistan. J Plant Pathol. Microbial. S:3
(2015). DOI:10.4172/2157-7471.S3-008
8. Donatelli, M., Magarey, R.D., Bregaglio, S., Willocquet, L., Whish, J.P.M., Savary, S.:
Modelling the impact of pest and disease on agriculture system. Agricul. System 155, 213–224
(2017)
9. CS231n Convolutional Neural Networks for Visual Recognition, Cs231n.github.io, 2019.
[Online]. Available: https://cs231n.github.io/convolutional-networks/. [Accessed: 02- Nov-
2019]
10. Habib, M., Majumder, A., Jakaria, A., Akter, M., Uddin, M. and Ahmed, F.: Machine vision
based papaya disease recognition. J. King Saud University—Comput. Inform. Sci. (2018)
11. AlexNet—ImageNet Classification with Convolutional Neural Networks. Neurohive.io.
(2019). [Online]. Available: https://neurohive.io/en/popular-networks/alexnet-imagenet-classi
fication-with-deep-convolutional-neural-networks/. [Accessed: 02- Nov- 2019]
12. A Comprehensive Guide to Convolutional Neural Networks—the ELI5 way. Medium.
(2019). [Online]. Available: https://towardsdatascience.com/a-comprehensive-guide-to-convol
utional-neural-networks-the-eli5-way-3bd2b1164a53. [Accessed: 02- Nov- 2019]
13. Accuracy, Recall, Precision, F-Score & Specificity, which to optimize on?Medium. (2019).
[Online]. Available: https://towardsdatascience.com/accuracy-recall-precision-f-score-specif
icity-which-to-optimize-on-867d3f11124. [Accessed: 02- Nov- 2019]
14. Confusion Matrix in Machine Learning—GeeksforGeeks.GeeksforGeeks. (2019). [Online].
Available: https://www.geeksforgeeks.org/confusion-matrix-machine-learning/. [Accessed:
02- Nov- 2019]
Exploitation of Vegetation Indices
and Random Forest for Cartography
of Rosemary Cover: Application
to Gourrama Region, Morocco

Hassan Chafik and Mohamed Berrada

Abstract The purpose of this study is to give an efficient and practical method
for mapping rosemary cover, which belongs to esparto grasslands. The approach
consists of merging two technics: remote sensing and machine learning. At first,
a set of vegetation indices was calculated using Sentinel 2A MSI image clipped
for the study area. In a second place, a correlation test was executed to eliminate
highly correlated indices. And so, three combinations were determined using the
holden indices. After that, the model was trained using terrain truth samples. In
the end, the model was applied to classify the image built, for the study area, by
stacking the indices. The model was run for the study area, then validated using
terrain sample points. Results are maps of rosemary cover densities corresponding
to each combination. The validation test showed a score of 80, 67.5, and 75,5%
respectively for combination 1, 2 and 3. This confirmed the selection of the first
combination.

Keywords Random forest · Spectral indices · Rosemary · Classification

1 Introduction

Rosemary is an evergreen shrub native to the Mediterranean and Asia region. It is

widely cultivated for its medicinal and kitchen utilities, in Portugal, Spain, France
Maghreb countries and China [1]. The use of rosemary was recorded since ancient
times, in Sumerian civilization, Chinese, Greeks, and Egyptians [1], which gave it
large popularity.
In Morocco, rosemary is prospering in arid to semi-arid climate. It coexists with
other shrubs forming what is called esparto grasslands. Those lands cover about

H. Chafik (B) · M. Berrada

Departement of Mathematics and Informatics, National High School of Arts and Crafts,
University of Moulay Ismail, Meknes, Morocco
e-mail: [email protected]
M. Berrada
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 429

T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9_30
430 H. Chafik and M. Berrada

3,255,714 hectares, which constitute about the third of the national forestall, Para
forestall, and esparto area (9069714 ha) [2]. The country was placed on the 12th in
2015, as an exporter of aromatic and medicinal plants, with about 4200 species.
The grandness of this resource makes very significant for countries, like Morocco,
and directions in charge, like the high charge for water, wood, and competing against
desertification, to finely map rosemary cover, and all esparto grasslands. This to
correctly estimate the production, and thus, to fair tax the private incomes.
The problematic, we treat in our work, discuss the problem caused by the weak
vegetation cover while discriminating it using remote sensing technics. This difficulty
is due to the weak density of leaf which allows the spectral influence of adjusted soil.
Thus, this work points to demonstrate an effective methodology for mapping rose-
mary and other esparto grasslands. The paper proposes a remote sensing—Random
Forest approach to build a classification model for rosemary cover based on spectral
indices.
The approach was inspired by various recent works like [3], who used random
forest classifier for tree cover mapping in drylands of sub-Saharan Africa. We find
also, the paper of [4], where they present a remote sensing—GIS-based approach
exploiting the spectral indices NDVI, SAVI and MSAVI. The same paper compared
the maximum likelihood and spectral angle mapper classification algorithms. The
work of Uyen [5] has also present a similar methodology using image segmentation
and random forest algorithm to classify an aerial image.
The methodology, discussed here, could be reproduced by a large user community,
either experts in remote sensing or administrative personnel of concerning directions.

2 Study Area

The study area (Fig. 1) belongs to Errachidia province, Draa Tafilalet region in the
southeast of Morocco. The zone is caught between 3,553,150 and 3546290 m in
latitude, and between 357,410 and 368230 m following the longitude (coordinates
are in WGS 1984 UTM zone 30 spatial reference).
The climate of the region is semi desertic. It is characterized by an important
thermic gap with a very high temperature in summer (mean = 31.5°) and a very low
temperature in winter (mean = 5°). Precipitation is weak and irregularly distributed
in space and time. The wind is recording an average speed of 57.6 km/h in spring
and summer.
Exploitation of Vegetation Indices and Random … 431

Fig. 1 The geographic situation of the study area in Morocco and the world

3 Material and Methods

3.1 Sentinel 2 MSI Images

The paper used Sentinel 2A MSI satellite image caught for the study area. The
image was downloaded from the USGS Earth explorer platform. The image contains
13 spectral bands at 10 to 60 m spatial resolution.
The selection of the image capture date is delicate for our study case. Scientific
papers are suggesting summer images for better discrimination between evergreen
and drought semi-deciduous vegetation [6]. In our area, precipitation is irregular.
Retaining only the years 2018 and 2019, we can notice this irregular behavior (Fig. 2).
Several periods of drought and precipitation are alternating during the twelvemonths.
432 H. Chafik and M. Berrada

60
50
40
30
20
10
0
august

december

august
january
february

may
june

september

november

january
february

may
june

september

november
july

july
april
march

october

march
april

october
Fig. 2 Graph demonstrating the variance of precipitation in the study area for the years 2018 and
2019

The drought period in summer 2018 was short (only June and July). While two
large ones were recorded from February 2018 to April 2018, from October 2018
to February 2019, and a short one in summer 2019 (from June to July), but with a
presence of weak precipitation (8 to 12 mm).
From this, the capture date chosen was 2019/04/01. Coming after three months
of drought and weak precipitation (less than 9 mm).
Another argument supporting our choice is that Google earth images are actualized
every first January, which means just three days before the capture of our satellite
image. Thus, Google earth images could be utilized for defining sample points and
validation points.

3.2 Spectral Indices

The foremost measure is to find which vegetation indices to use for better training
our model. Using the satellite image clipped for the subject region, we calculated a
set of eight different vegetation indices (Table 1):
• The Modified Chlorophyll Absorption in Reflectance Index (MCARI)
• The Normalized Differenced Vegetation Index (NDVI)
• The Normalized Differenced Water Index (NDWI)
• The Coloration Index (CI)
• The Transformed Differenced Vegetation Index (TDVI)
• The Brilliance Index (BI)
• The Moisture Stress Index (MSI)
• The Soil Adjusted Vegetation Index (SAVI)
Exploitation of Vegetation Indices and Random … 433

Table 1 Relative formulas for the vegetation indices used in this study
Index Formula
MCARI [7]      R E D edge 
R E D edge − R E D − 0, 2 R E D edge − G R E E N × RE D
N I R−R E D
NDVI [8] N I R+R E D
G R E E N −N I R
NDWI [9] G R E E N +N I R
R E D−B LU E
CI [10] RE D
TDVI [11] 1, 5 × √ N I R−R E D
N I R 2 +R E D+0,5

BI [12] N I R2 + R E D2 + G R E E N 2
SW I R
MSI [13] NIR
SAVI [14] S AV I = 1.5 × NIR−RED
NIR+RED+0.5

Then, we tested out the correlation between each of them using Pearson’s r coeffi-
cient. Considering that two highly correlated variables will contribute indifferently to
the explanation of the phenomenon, this test was introduced in order to keep only the
non-correlated indices. Pearson’s r test gave the correlation matrix between indices
(Table 2).
Based on the results of the correlation (Table 2), we picked up only indices showing
no or low correlation. Those indices are MCARI, NDVI, TDVI, IB and MSI.
The Modified Chlorophyll Absorption in Reflectance Index (MCARI) is an index
developed by [15] as the Chlorophyll Absorption Ratio Index (CARI), then was
modified by [7] to be the Modified Chlorophyll Absorption Ratio Index (MCARI).
MCARI is responsive to chlorophyll variation without incorporating near-infrared
wavelength.
The Normalized Differenced Vegetation Index (NDVI) developed by [8]. It is an
index based on measuring the gap between the red wave and the near-infrared wave
reflectance for a given body. It is a biophysical parameter that clearly correlates with
chlorophyll concentration [16].

Table 2 The correlation matrix for the vegetation indices used in this work
MCARI NDVI NDWI CI TDVI BI MSI SAVI
MCARI 1 0.22 −0.07 −0.03 0.004 −0.13 0.63 −0.03
NDVI 1 −0.7 0.46 0.62 0.29 −0.33 −0.73
NDWI 1 −0.89 −0.8 −0.68 0.36 0.87
CI 1 0.75 −0.79 −0.31 −0.79
TDVI 1 −0.89 −0.4 −0.98
BI 1 −0.33 −0.83
MSI 1 −0.43
SAVI 1
434 H. Chafik and M. Berrada

Table 3 The combinations

Combination 1 MCARI—NDVI—BI
tested in this work
Combination 2 MCARI—NDVI—BI—MSI
Combination 3 MCARI—TDVI—MSI

The Transformed Differenced Vegetation Index (TDVI) is an index presented by

[11]. It is as sensitive as the Soil Adjacent Vegetation Index (SAVI) to subjacent
bare soil, but it shows excellent linearity comparing to NDVI and SAVI and does not
saturate like them.
The Brilliance Index (BI) developed by [12]. This index is useful for distinguishing
vegetation cover from bare soil, but it is sensitive to shadow and humidity variations
[17].
The last one is the Moisture Stress Index (MSI). It was developed by Hunt [13]
as a simple ratio between short wave infrared and near-infrared.
We used those indices to build a set of combinations (Table 3). Each combination
will feed our model to find the most efficient one for classifying rosemary cover
(Fig. 3).

3.3 Random Forest Model

Random forest (RF) is a combination of predictors (algorithms). It was created by

[18] and proposed by [19] as a tree-based classifier. It consists of giving a random
vector to the RF which generates classifiers organized as trees, then it averages the
results. The random vector is sampled from the given vector in the input (the training
set).
Random forest classifier was chosen for its wide use in remotely sensed data classi-
fication. Besides, it is characterized by its attractivity due to its ability to manipulate
high dimensional and non-normally distributed data which makes it adequate for
image classification [20]. Several works have proven the utility of combining remote
sensing technics and automated technics such as random forest. We find the work
of Steven M.[21] where they show the potential of random forest in integrating data
from different sensor systems and GIS data. We have also the papers [3, 4] and [5]
cited in the introduction.
In our case, we provide the model with an input vector of 206 training samples
(Table 4). The vector is made up of points. To each point, the values of the indices
forming the combination were associated (explanatory variables). Besides, the vector
contains the class (the dependent variable) to which belongs each of the sample points.
For what concerns the architecture of our model, it is constituted of 50 trees. Every
tree holds the Gini impurity function used for measuring the quality of the split. This
function is a metric that measures, within the subsets, the homogeneity of the target
variable. In other words, it measures the probability of misclassifying an object into
a class [22]. Its formula is given by (1):
Exploitation of Vegetation Indices and Random … 435

Fig. 3 Organigram resuming the remote sensing approach

Table 4 An extract of the input vector provided to our model (combination 1)

Explanatory variables Class Significance of the class
ID NDVI MCARI IB
0 0.178716004 0.0091629 0.42524001 2 veg F
1 0.156408995 0.02317791 0.365152001 3 veg TF
2 0.168661997 0.007115124 0.329504997 2 veg F
3 0.138964996 0.021320885 0.235801995 4 sol nu
4 0.059999999 0.000545819 0.042502701 4 sol nu
5 0.176909998 0.007475095 0.288599014 2 veg F
6 0.156329006 -0.002273347 0.283845007 3 veg TF
436 H. Chafik and M. Berrada

Fig. 4 Organigram resuming the random forest approach

d
ui ui
i Gini (u) = 1− (1)
i−1
u1 u1

We set a minimal bit of samples required to split an internal node into 2 samples,
this forces the nodes to expand until they are pure or until all leaves contain less than
the defined minimum number of split samples.
After the training phase, we run the model to classify the image built by stacking
the indices relative to each combination and distinguish four classes: important rose-
mary cover (IR), weak rosemary cover (WR), very weak rosemary cover (VWR) and
bare soil (Bs). For the validation phase, we used a set of terrain truth samples from
the study area. We repeat those steps for each combination (Fig. 4).

4 Results and Discussion

Running the random forest model gave, as results, three classified images for the
study area, relative to the three defined combinations. In each case, the model has
attributed a set of weights to the parameters (Table 5).
Focusing on the combination 1, we see that the model has given different weights
to the parameters. The NDVI took the higher weight followed respectively by BI and
MCARI. This implies that the most determinant factor in describing the vegetation
cover is the gap between the red wave and the near-infrared wave. In the second
combination, the model coupled the parameters by giving them close values of weight
(MCARI and MSI; NDVI and BI), but we steel notice the importance of the NDVI.
Exploitation of Vegetation Indices and Random … 437

Table 5 Weights assigned to each index for each combination

Combination 1 Combination 2 Combination 3
Index Weight Index Weight Index Weight
MCARI 0.23 MCARI 0.12 MCARI 0.28
NDVI 0.40 NDVI 0.36 TDVI 0.36
BI 0.36 BI 0.32 MSI 0.35
MSI 0.18

For what concerns the third combination, TDVI and MSI were weighted equally
proving the importance of the infrared domain and the red-infrared gap.
On the maps (Figs. 5, 6 and 7), we have four classes (IR, WR, VWR, Bs). The
maps present an evident difference concerning the spatial distribution of the classes.
From the map in Fig. 6, we see that the model has affected many pixels to the Bs
class, while they were classified as WR or VWR in the other maps. In Fig. 7, we
notice that the model has classified many pixels, in the south of the area, as IR class.
Those results were validated, as explained in the methodology, by comparing
the observed values and the predicted values of the points. The validation test gave
different scores (Table 6). The score reflects the percentage of points that match.

Fig. 5 Map of the rosemary cover relative to the combination 1

438 H. Chafik and M. Berrada

Fig. 6 Map of the rosemary cover relative to the combination 2

Fig. 7 Map of the rosemary cover relative to the combination 3

Exploitation of Vegetation Indices and Random … 439

Table 6 The score of the

Training score Mean Validation
training, mean accuracy and
accuracy score
score of the validation for
each combination Combination 1 0.68 0.8 0.8
Combination 2 0.70 0.85 0.675
Combination 3 0.61 0.675 0.675

The combination 1 shows the higher score of validation, while the second combi-
nation presents the higher mean accuracy and the higher training score. From this,
we adopt the first combination which has given the best results.

5 Conclusion

The analysis of the potential of using vegetation indices for mapping rosemary cover
has revealed some tips concerning the use of the vegetation indices for detecting weak
vegetation cover. The results established that not any vegetation index is useful for
classification, simply there are combinations to find for well training the supervised
classification. The work demonstrated that within the eight indices tested, NDVI,
MCARI and BI are forming the best combination. The scores and the weights, given
by running the random forest model, explained that the red wave-infrared wave gap
and the infrared domain are determinant for detecting the different rosemary classes.

References

1. Sasikumar, B.: Rosemary. Handb. Herbs Spices Second Ed. 1, 452–468 (2012). https://doi.org/
10.1533/9780857095671.452
2. Blérot, P., Mhirit, O.: Le grand livre de la forêt marocaine. Mardaga (1999)
3. Adjognon, G.S., Rivera-Ballesteros, A., van Soest, D.: Satellite-based tree cover mapping for
forest conservation in the drylands of Sub Saharan Africa (SSA): application to Burkina Faso
gazetted forests. Dev. Eng. 4, 100039 (2019). https://doi.org/10.1016/j.deveng.2018.100039
4. Jarradi, S., Tounsi, K.: The use of satellite remote sensing and geographic information systems
in monitoring the dynamics of alfatières aquifers. Case of the delegation of Hassi el Frid of the
governorate of Kasserine in L ’ utilisation de la télédétection satellitaire et des s. CI, 3449–3458
(2018)
5. Nguyen, U., Glenn, E.P., Dang, T.D., Pham, L.T.H.: Mapping vegetation types in semi-arid
riparian regions using random forest and object-based image approach: a case study of the
Colorado River Ecosystem, Grand Canyon. Arizona. Ecol. Inform. 50, 43–50 (2019). https://
doi.org/10.1016/j.ecoinf.2018.12.006
6. Maselli, F., Rodolfi, A., Romanelli, S., Conese, C., Bottai, L.: Classification of Mediterranean
vegetation by TM and ancillary data for the evaluation of fire risk. Int. J. Remote Sens. 21,
3303–3313 (2000). https://doi.org/10.1080/014311600750019912
7. Daughtry, C.S.T., Walthall, C.L., Kim, M.S., De Colstoun, E.B., McMurtrey Iii, J.E.: Estimating
corn leaf chlorophyll concentration from leaf and canopy reflectance. Remote Sens. Environ.
74, 229–239 (2000)
440 H. Chafik and M. Berrada

8. Rouse Jr, J.W., Haas, R.H., Schell, J.A., Deering, D.W.: Monitoring the vernal advancement
and retrogradation (green wave effect) of natural vegetation (1973)
9. McFeeters, S.K.: Using the normalized difference water index (ndwi) within a geographic
information system to detect swimming pools for mosquito abatement: a practical approach.
Remote Sens. 5, 3544–3561 (2013). https://doi.org/10.3390/rs5073544
10. Escadafal, R., Belghit, A., Ben-Moussa, A.: Indices spectraux pour la teledetection de la degra-
dation des milieux naturels en Tunisie aride. Sous la direction de G. Guyot. Dans Actes du
6e Symp. Int. sur les Mes. Phys. signatures en Teledetect. Val d’Isere, Fr. 17–24 janvier 1994,
ISPRS-CNES. 6, 253_259 (1994)
11. Bannari, A., Asalhi, H., Teillet, P.M.: Transformed difference vegetation index (TDVI) for
vegetation cover mapping. Int. Geosci. Remote Sens. Symp. 5, 3053–3055 (2002). https://doi.
org/10.1109/igarss.2002.1026867
12. Escadafal, R., Bacha, S.: Strategy for the dynamic study of desertification. (1996).
13. Hunt, E.R., Jr., Rock, B.N.: Detection of changes in leaf water content using near-and middle-
infrared reflectances. Remote Sens. Environ. 30, 43–54 (1989)
14. Huete, A.R.: A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25, 295–309
(1988). https://doi.org/10.1016/0034-4257(88)90106-X
15. Kim, M.S., Daughtry, C.S.T., Chappelle, E.W., McMurtrey, J.E., Walthall, C.L.: The use of
high spectral resolution bands for estimating absorbed photosynthetically active radiation (A
par) (1994)
16. Sellers, P.J.: Canopy reflectance, photosynthesis and transpiration. Int. J. Remote Sens. 6,
1335–1372 (1985). https://doi.org/10.1080/01431168508948283
17. Levin, N., Tsoar, H., Maia, L.P., Sales, V.C., Herrmann, H.: Dune whitening and inter-dune
freshwater ponds in NE Brazil. CATENA 70, 1–15 (2007)
18. Ho, T.K.: Random decision forests. Proc. Int. Conf. Doc. Anal. Recognition, ICDAR. 1, 278–
282 (1995). https://doi.org/10.1109/ICDAR.1995.598994
19. Breiman, L.: RANDOM FORESTS-RANDOM FEATURES (1999)
20. Millard, K., Richardson, M.: On the importance of training data sample selection in Random
forest image classification: a case study in peatland ecosystem mapping. Remote Sens. 7,
8489–8515 (2015). https://doi.org/10.3390/rs70708489
21. Kloiber, S.M., Macleod, R.D., Smith, A.J., Knight, J.F., Huberty, B.J.: A Semi-Automated,
Multi-Source Data Fusion Update of a Wetland Inventory for East-Central Minnesota. USA.
Wetlands. 35, 335–348 (2015). https://doi.org/10.1007/s13157-014-0621-3
22. Laber, E., Murtinho, L.: Minimization of Gini Impurity: NP-completeness and Approximation
Algorithm via Connections with the k-means Problem. Electron. Notes Theor. Comput. Sci.
346, 567–576 (2019). https://doi.org/10.1016/j.entcs.2019.08.050
Author Index

A Dantan, Jean-Yves, 15
Abadi, Asmae, 1 Deshayes, Laurent, 45
Abadi, Chaimae, 1 Diany, Mohammed, 215
Abdelouahid, Lyhyaoui, 371 Douimi, Mohammed, 323
Abdullahi, Ibrahim Mohammed, 385
Ahmadi, Abdeslam, 149
Aissaoui, Hicham, 215 E
Akter, Roksana, 415 Eddayer, Omar, 253
Amhraoui, Elmehdi, 263 Elhaq, Saad Lissane, 59
Ayad, Hassan, 239 El-kaime, Hafsa, 59
Elkari, Abdeljalil, 239
En-Naimani, Zakariae, 351
B Etienne, Alain, 161
Bala, Jibril Abdullahi, 385
Battas, Ilham, 45
Behja, Hicham, 45 F
Benabdellah, Abla Chaouni, 103 Faquir, Sanaa, 173
Bendella, Fatima, 185
Benghabrit, Asmaa, 103, 125
Benghabrit, Youssef, 89 H
Bennani, Younès, 277 Habib, Md. Tarek, 415
Berdeddouch, Ayoub, 277 Hajji, Tarik, 173
Berrada, Mohamed, 429 Halmaoui, Houssam, 291
Boudnaya, Jaouad, 253 Haqiq, Abdelkrim, 291
Bouhaddou, Imane, 89, 103, 125 Hassani, Ibtissam El, 77
Bourezza, El Mehdi, 33 Hassan, Ramchoun, 351
Haytoumi, Amine, 253
Homri, Lazhar, 15
C
Chafik, Hassan, 429
Chennoufi, Mohammed, 185 I
Cherrafi, Anass, 89 Iathriouan, Zakaria, 173
Chowdhury, Sadia, 415 Ibtissam, El Hassani, 403

D J
Daniya, Emmanuel, 385 Jaara, Elmiloud, 173
© Springer Nature Switzerland AG 2021 441
T. Masrour et al. (eds.), Artificial Intelligence and Industrial Applications,
Advances in Intelligent Systems and Computing 1193,
https://doi.org/10.1007/978-3-030-51186-9
442 Author Index

Jourhmane, Mostafa, 199 P

Peleshchak, Ivan, 313
Peleshchak, Roman, 313
K
Khadija, Echarrafi, 403
Khan, Md. Abbas Ali, 415 Q
Khdoudi, Abdelmoula, 77 Qaffou, Issam, 339
Kolbusz, Janusz, 313
Kopka, Jan, 313
Korniak, Janusz, 313 R
Rhazzaf, Mohamed, 227
Rostami, Hamideh, 15
Rozycki, Pawel, 313
L
Labbi, Oulfa, 149
Lytvyn, Vasyl, 313 S
Saadi, Adil, 323
Saïd, Kasttet Mohammed, 371
M Semmouri, Abdellatif, 199
Manssour, Imad, 1 Sheikh, Md. Helal, 415
Masrour, Tawfik, 77, 173, 227, 263 Siadat, Ali, 161
Mazgualdi, Choumicha El, 77
Mim, Tahmina Tashrif, 415
Mjahed, Mostafa, 239 T
Mkhida, Abdelhak, 253 Taleb, Ismail, 161
Mohamed, Ben Abbou, 403
Mohamed, Ettaouil, 351
Moudoud, Brahim, 215 V
Moufaddal, Mariam, 125 Verde, Rosanna, 277
Mounia, El Haji, 403
Mousrij, Ahmed, 33
Mustapha, Taleb, 403 W
Wrzesien, Mariusz, 313

N Y
Nour-eddine, Joudar, 351 Yahyaouy, Ali, 277
Yakoubi, Youssef, 323

O
Olaniyi, Olayemi Mikail, 385 Z
Olanrewaju, Esther Ayobami, 385 Zakia, Rais, 403
Ouazzani Jamil, Mohammed, 173 Zekhnini, Kamar, 89
Oulhiq, Ridouane, 45 Zemmouri, Elmoukhtar, 323
Oultiligh, Ahmed, 239 Zouhri, Wahb, 15