Applied Learning

Algorithms for
Intelligent IoT
Applied Learning
Algorithms for
Intelligent IoT

Edited by
Pethuru Raj Chelliah
Usha Sakthivel
Susila Nagarajan
First Edition published [2022]
by CRC Press
6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742
and by CRC Press
2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN
© 2022 Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, LLC
The right of Pethuru Raj, Usha Sakthivel, and Susila Nagarajan, to be identified as the authors
of the editorial material and the authors for their individual chapters, has been asserted in
accordance with Sections 77 and 78 of the Copyright, Designs and Patents Act 1988.
Reasonable efforts have been made to publish reliable data and information, but the author
and publisher cannot assume the responsibility for the validity of all materials or the
consequences of their use. The authors and publishers have attempted to trace the copyright
holders of all the materials reproduced in this publication and apologize to copyright holders
if permission to publish in this form has not been obtained. If any copyright material has not
been acknowledged, please write and let us know so that we may rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted,
reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other
means, now known or hereafter invented, including photocopying, microfilming, and
recording, or in any information storage or retrieval system, without a written permission
from the publishers.
For permission to photocopy or use material electronically from this work, access
www.copyright.com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood
Drive, Danvers, MA 01923, 978-750-8400. For works that are not available on CCC, please
contact [email protected]
Trademark notice: Product or corporate names may be trademarks or registered trademarks
and are used only for identification and explanation without intent to infringe.
ISBN: 978-0-367-63594-7 (hbk)
ISBN: 978-1-032-11321-0 (pbk)
ISBN: 978-1-003-11983-8 (ebk)

DOI: 10.1201/9781003119838

Typeset in Caslon
by MPS Limited, Dehradun
Contents

CONTRIBUTOR LIST ix

CHAPTER 1 CONVOLUTIONAL NEURAL NETWORK IN COMPUTER

VISION 1
D. AISHWARYA AND R. I. MINU

CHAPTER 2 TRENDS AND TRANSITION IN THE MACHINE

LEARNING (ML) SPACE 33
SRUTHI ANAND, N. SUSILA, AND USHA SAKTHIVEL

CHAPTER 3 NEXT-GENERATION IOT USE CASES ACROSS

INDUSTRY VERTICALS USING MACHINE LEARNING

ALGORITHMS 57
T. R. KALAIARASAN, SRUTHI ANAND,
V. ANANDKUMAR, AND A. M. RATHEESHKUMAR

CHAPTER 4 A PANORAMIC VIEW OF CYBER ATTACK

DETECTION AND PREVENTION USING MACHINE
LEARNING AND DEEP LEARNING APPROACHES 77
ESTHER DANIEL, N. SUSILA,
AND S. DURGA

v
vi CONTENTS

CHAPTER 5 REGRESSION ALGORITHMS IN MACHINE LEARNING 99

USHA SAKTHIVEL, NEHA SINGHAL,
PETHURU RAJ CHELLIAH, AND ASHWINI R.
MALIPATIL

CHAPTER 6 MACHINE LEARNING-BASED INDUSTRIAL INTERNET

OF THINGS (IIOT) AND ITS APPLICATIONS 123
M. SURESH KUMAR, S. RACHEL,
AND M. V. KAVISELVAN

CHAPTER 7 EMPLOYEE TURNOVER PREDICTION USING SINGLE

VOTING MODEL 153
R. VALARMATHI, M. UMADEVI, AND T. SHEELA

CHAPTER 8 A NOVEL IMPLEMENTATION OF SENTIMENT

ANALYSIS TOWARD DATA SCIENCE 175
VIJAYALAKSHMI SARAVANAN, ISHPREET SINGH,
EMANUEL SZAREK, JEREON HAK,
AND ANJU S. PILLAI

CHAPTER 9 CONSPECTUS OF K-MEANS CLUSTERING

ALGORITHM 193
USHA SAKTHIVEL, A. P. JYOTHI, N. SUSILA,
AND T. SHEELA

CHAPTER 10 SYSTEMATIC APPROACH TO DEAL WITH INTERNAL

FRAGMENTATION AND ENHANCING MEMORY SPACE

DURING COVID-19 215
MOHAN APARNA, R. MAHESWARI
AND J. V. THOMAS ABRAHAM

CHAPTER 11 IOT AUTOMATED SPY DRONE TO DETECT AND

ALERT ILLEGAL DRUG PLANTS FOR LAW

ENFORCEMENT 241
GOTLURU ARUN KUMAR, PULUGU YAMINI,
AND R. MAHESWARI

CHAPTER 12 EXPOUNDING K-MEANS-INSPIRED NETWORK

PARTITIONING ALGORITHM FOR SDN CONTROLLER
PLACEMENT 265
J. PUSHPA AND PETHURU RAJ CHELLIAH
 CONTENTS vii

CHAPTER 13 AN INTELLIGENT DEEP LEARNING-BASED WIRELESS

UNDERGROUND SENSOR SYSTEM FOR IOT-BASED

AGRICULTURAL APPLICATION 291

PRISCILLA RAJADURAI
AND G. JASPHER W. KATHRINE

CHAPTER 14 PREDICTING EFFECTIVENESS OF SOLAR POND

HEAT EXCHANGER WITH LTES CONTAINING CUO
NANOPARTICLE USING MACHINE LEARNING 325
K. KARUNAMURTHY, G. SUGANYA, M. ANANTHI,
AND T. SUBHA

INDEX 345
Contributors

J. V. Thomas Abraham V. Anandkumar

School of Computer Science and Department of Information
Engineering (SCOPE) Technology
Vellore Institute of Technology Sri Krishna College of
Chennai, Tamil Nadu, India Engineering and Technology
D. Aishwarya Coimbatore, India
SRM Institute of Science and
Technology Mohan Aparna
Kattankulathur, Tamil Nadu, School of Computer Science and
India Engineering (SCOPE)
Vellore Institute of Technology
Sruthi Anand Chennai, Tamil Nadu, India
Department of Information
Technology Pethuru Raj Chelliah
Sri Krishna College of RJIL College of Engineering
Engineering and Technology and Technology
Coimbatore, India Bengaluru, India

ix
x CONTRIBUTORS

Esther Daniel M. V. Kaviselvan

Department of Computer Science Department of Mechanical
and Engineering, Technology Engineering
Karunya Institute of Technology Sri Sairam Engineering College
and Sciences Chennai, Tamil Nadu, India
Coimbatore, India
Gotluru Arun Kumar
S. Durga School of Computer Science and
Department of Information Engineering (SCOPE)
Technology Vellore Institute of Technology
Karunya Institute of Technology Chennai, Tamil Nadu, India
and Sciences
M. Suresh Kumar
Coimbatore, India
Department of Information,
Jereon Hak Technology
M.S. Student Sri Sairam Engineering College
Rochester Institute of Chennai, Tamil Nadu, India
Technology
R. Maheswari
Rochester, New York, USA
School of Computer Science and
A. P. Jyothi Engineering (SCOPE)
Department of Computer Vellore Institute of Technology
Science and Engineering VTU Chennai, Tamil Nadu, India
Chennai, India
Ashwini R. Malipatil
T. R. Kalaiarasan Department of CSE
Department of Information Sri Krishna College of
Technology Engineering and Technology
Sri Krishna College of Coimbatore, India
Engineering and Technology
R. I. Minu
Coimbatore, India
SRM Institute of Science and
K. Karunamurthy Technology
Vellore Institute of Technology Kattankulathur, Tamil Nadu,
Chennai, Tamil Nadu, India India

G. Jaspher W. Kathrine Anju S. Pillai

Karunya Institute of Technology Amrita Vishwa Vidyapeetham
and Sciences Coimbatore, India
Coimbatore, India
 CONTRIBUTORS xi

J. Pushpa T. Sheela
Jain University Department of Information
Bengaluru, India Technology
Sri Sairam Engineering College
S. Rachel
Chennai, Tamil Nadu, India
Department of Information
Technology Ishpreet Singh
Sri Sairam Engineering College M.S. Student
Chennai, Tamil Nadu, India Rochester Institute of
Technology
Priscilla Rajadurai
Rochester, New York, USA
St. Joseph’s Institute of
Technology Neha Singhal
Chennai, India Department of ISE
Sri Krishna College of
A. M. Ratheeshkumar
Engineering and Technology
Department of Information
Coimbatore, India
Technology
Sri Krishna College of T. Subha
Engineering and Technology Sri Sairam Engineering College
Coimbatore, India Chennai, India

Usha Sakthivel G. Suganya

Dept. of Research and Vellore Institute of Technology
Innovation Chennai, Tamil Nadu, India
Rajeswari College of
N. Susila
Engineering
Department of Information
Bengaluru, India
Technology
and
Sri Krishna College of
Dept. of Computer Science &
Engineering and Technology
Engineering
Coimbatore, India
Visvesvaraya Technological
University Emanuel Szarek
Belagavi, Karnataka, India M.S. Student
Rochester Institute of
Vijayalakshmi Saravanan
Technology
Faculty Member
Rochester, New York, USA
Rochester Institute of
Technology
Rochester, New York, USA
xii CONTRIBUTORS

M. Umadevi Pulugu Yamini

Department of Computer School of Computer Science and
Science and Engineering Engineering (SCOPE)
SRM Institute of Science and Vellore Institute of Technology
Technology Chennai, Tamil Nadu, India
Chennai, Tamil Nadu, India

R. Valarmathi
Department of Computer
Science and Engineering
Sri Sairam Engineering
College
Chennai, Tamil Nadu, India
 1
CONVOLUTIONAL NEURAL
NETWORK IN COMPUTER VISION
D. AISHWARYA1 AND R.I. MINU2
1
Research Scholar, SRM Institute of
Science and Technology,
Kattankulathur, Tamil Nadu, India
2
SRM Institute of Science and
Technology, Kattankulathur,
Tamil Nadu, India

Contents

Introduction 2
Convolutional Neural Network (CNN) 4
Distinctive Properties of CNN 5
Activation Functions for CNN 6
Loss Function 8
Datasets and Errors 9
Bias and Variance 9
Overfitting and Underfitting 10
Understanding Padding and Stride 11
Padding 11
Stride 12
Parameters and Hyper Parameters 13
CONV Layer 13
Filter 13
Feature Map 14
Convolution Operation 14
Key Points about Convolution Layers and Filters 16
Pooling Layer 16
Key Points about Pooling Layer 16
Types of Pooling 17
Forward Propagation 18
Calculating the Parameters 18

DOI: 10.1201/9781003119838-1 1
2 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Activation Shape and Size 19

Backward Propagation 20
Optimizers 21
Other Ways to Improve the Performance of CNN 21
Image Data Augmentation 21
Deeper Hidden Layers 22
Early Stopping 22
Application of CNN 22
Image Classification 22
RCNN and Object Detection 25
Region-Based Convolutional Neural Network (R-CNN) 26
Fast R-CNN 27
Faster R-CNN 28
You Only Look Once (YOLO) 29
Transfer Learning 29
Transfer Learning for CNN 29
Fine-Tuning or Freezing? 30
Neural Style Transfer (NST) 30

Introduction

Emergence of artificial intelligence (AI) has made machines “Smart.”

By “Smart,” we mean the ability of the machine (possibly a com­
puter) to process the given data with rationality, like an intelligent
human being, and take a suitable decision. AI achieves this in­
telligence with the help of its subsets of algorithms (see Figure 1.1).
AI is a broader field that includes machine learning (ML), artificial
neural network (ANN), and deep learning (DL) as its subsets. AI is
achieved by ANN, which is a network of artificial neurons resembling a
biological neuron and the nervous system. ML enables the machines to
process the data, make decision, and optimize the results based on ex­
perience. ML primarily requires a feature engineer to do the feature en­
gineering process. Also, ML works its best for structured dataset (mostly
comprises numerical data) but is proven inefficient for unstructured types.
To process the unstructured and continuous data like video,
speech, etc., we need more complex neural network. Here comes in
the DL with better results on processing the unstructured data. One
 CNN IN COMPUTER VISION 3

Figure 1.1 Subsets of Artificial Intelligence

of the biggest advantages of DL is that it features engineering au­

tomatically, that is, no manual feature selection is required. Also, ML
requires manual intervention when the algorithm is struck with op­
timization problem, which is also automated in DL. DL plays the
key role in implementing and visualizing the actual AI. Deep
learning, however, requires a large amount of dataset for the purpose
of training, which in turn requires a deep network of layers of neu­
rons to process the features and hence the name “deep learning.”
DL comprises many algorithms for different types of data. Of those,
convolutional neural network (CNN) is mainly used to process images
and solve computer vision problems. The idea of CNNs can be traced
back to 1980, when K. Fakushima proposed “neocognition” – a hier­
archical multi-layered neural networks, which is self-organized and
unaffected by the translation operation in the dataset (shift in positions
of patterns). It was primarily created for pattern recognition.
The first standard CONVNet was introduced by Yann LeCunn in
1998, in his research article titled “Object Recognition with Gradient
Based Learning.” It is important to note that Yann LeCunn has cited
Fakushima’s article in his research work, and the CONVNet of
LeCunn has a structure similar to neocognition with back propa­
gation added to the network. Thenceforth, multiple improvements
and the extension were made to it.
4 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Convolutional Neural Network (CNN)

Convolutional neural network is primarily used for operations like

image classification, segmentation, and other computer vision op­
erations. CNN is also called shape/space invariant artificial neural
network (SIAAN) based on its shared weights and translation in­
variance properties. The key idea behind CNN is to capture the local
patterns through convolution operation.
In general, each CNN architecture has a common working flow. A
CNN consists of multiple layers of convolution and pooling and a fully
connected ANN. A layer in CNN may comprise one convolution layer
and one pooling layer together as a single layer if the count of con­
volution layer and pooling layer is same. Most commonly used type is
two convolution layers followed by a single pooling layer.
The number of layers in a CNN architecture depends on the com­
plexity of the application. The output of the last layer is passed to an
ANN. The convolution layers and the pooling layers always perform the
same operation of learning the input image. The fully connected networks
perform the required operations, such as regression, classification, etc.
At each convolution layer, a local pattern is obtained by applying
filters throughout the image. Filters can also be mentioned as kernels.
The Initial/lower-level layers of the CNN obtain the elementary
pattern of the image like edges and lines. The higher-level layers of
CNN acquire the patterns recognized by the previous layer, and
through the concept of spatial locality, higher-levels obtain larger
patterns, as shown in Figure 1.2. The last layer of CNN would

Figure 1.2 Typical Operation of Conv Layer

 CNN IN COMPUTER VISION 5

generate the complete image, and the ANN attached at last would
give the required output depending on the problem for which the
particular neural network was developed.

Distinctive Properties of CNN

The following properties of CNN make it unique from the other DL

algorithms:
• A CNN contains volume of neurons as shown in Figure 1.3,
i.e., the neuron in CNN layers are arranged in three dimensions
(depth, height, and width).
• Shared Weights: When a filter is applied over a part of an
image, same set of parameter values of filter is used for all parts
of the image.
• Receptive Fields: Each neuron in the 3-D structure of a CNN is
subjected to a defined region of the input data, and this allows
the neurons to learn the elementary patterns of the input. This
region of input, associated to a particular neuron, is called the
local receptive field as in Figure 1.4.

Figure 1.3 Volume of Neurons

Figure 1.4 Local Receptive Field

6 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

• Local Connectivity: The local connectivity pattern between the

neurons is essential to exploit spatial locality. This local connectivity
pattern enables the network to create the representation of local
pattern and then assembles the representation of larger area.
• Pooling: Pooling layer is used to reduce the size of the feature map.
A feature map is the output obtained from convolution operation.
Before describing CNN in detail, it is important to know few other
terminologies and concepts that would be applied to CNN.

Activation Functions for CNN

An activation function is applied to the output of the convolution

layer of the CNN architecture to decide if the output is precise and if
it should be passed as an input to the next layer or should it be
dismissed.
Y = Activation ( (weight input ) + bias )

The resulting value after the application of activation function ranges

usually from 0 to 1 or from –1 to 1, depending upon the function
applied. The activation function may either be linear or non-linear.
For any neural network or DL algorithm, linear activation func­
tions are hardly used, since their output would be a direct multiple of
the input and the resulting value may be of any range.
The most commonly used non-linear activation functions include
the following:
• Sigmoid or Logistic Activation Function: It is used when
probability of output is required, since the resulting value ranges
from 0 to 1 (see Figure 1.5).
• Tanh or Hyperbolic Tangent Activation Function: The range
of the hyperbolic tanh function is from –1 to 1. It is often used
for classification between two classes.
Tanh is shown in Figure 1.6.
• Rectified Linear Unit (ReLU) Activation Function: CNN ar­
chitecture most commonly uses the ReLU activation function.
ReLU allows the neural network to converge quite easily and
 CNN IN COMPUTER VISION 7

Figure 1.5 Sigmoid Function

Figure 1.6 Tanh Activation Function

allow back propagation as well. One problem with ReLU is that

the gradient of the function becomes 0, when the input be­
comes 0 or negative. This condition prevents back propagation
and is known as dying ReLU problem.
• Leaky ReLU: Also the parametric ReLU (PReLU) prevents
dying ReLU problem by introducing a small slope to keep the
updates alive.
8 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 1.7 ReLU and Leaky ReLU Functions

As shown in Figure 1.7, ReLU works as

R(x) = (0, x)

But, PReLU/Leaky ReLU function adds the slope value of a to the input

ax if x < 0
R (x ) =
x if x 0

If a = 0.1, it is called Leaky ReLU and if a is any suitable value, it is

called PReLU.
• Softmax Activation Function: The most commonly used activa­
tion function for multiclass classification is Softmax. It is used only
for the output layer unlike the other types of activation function.

Loss Function

The optimization algorithm requires the error value of the current

iteration to be estimated repeatedly. This loss function also called
error function can be used to estimate the error value of the model
thereby, enabling it update the weights of the parameters, so that the
accuracy can be improved and the error be decreased in the next
iteration. Loss functions are selected based on the problem for which
the network is modeled.
• Regression Problem – Loss Functions used are
1. Mean Squared Logarithmic Error Loss (MSLE)
2. Mean Squared Error Loss (MSE)
3. Mean Absolute Error Loss (MAE)
 CNN IN COMPUTER VISION 9

• Binary Classification – Loss Functions used are

1. Binary Cross-Entropy
2. Squared Hinge Loss
3. Hinge Loss
• Multiclass Classification – Loss Functions used are
1. Multiclass Cross-Entropy Loss
2. Kullback–Leibler Divergence Loss
3. Sparse Multiclass Cross-Entropy Loss
• Segmentation – Loss Functions used are
1. Weighted Cross Entropy
2. Focal Loss
3. Dice Loss
4. Boundary Loss

Datasets and Errors

The dataset for a neural network is split into three different subsets:

Training Data: It is a set of data, comprising 80% of total data,

used for learning and modeling the neural network. Training data is
used to fit the parameters of the classifier. The error (based
on accuracy) arises when a trained model is run over the same
training data.
Validation Data: It is a subset of training data held unused, to es­
timate the accuracy of the model trained using training data based on
which the parameters are tuned. The error that arises when the
trained model is run over the held out validation data is known as
validation error. It provides an estimate of test error.
Test Data: It is the remaining 20% of total data, used to assess the
performance of the fully trained neural network. The error that arises
when the trained model is run over the test data is known as test error.

Bias and Variance

Bias is an error arising due to erroneous assumption of the model.

Higher value of bias causes larger value of training error.
10 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Example: The model might wrongly assume the features to error

and miss the essential details of the data while modeling.
Variance is also an error that occurs when the model begins to
consider noise as an essential feature and models it. High Variance
causes high validation error.

Total Error: The sum of the bias error and variance error is the total
error.
Trade-Off between Bias and Variance: For a model to work per­
fectly, low-bias and low-variance errors are the ideal conditions.

Overfitting and Underfitting

Overfitting occurs when the model trains on the training data very
well. Overfitting causes the model to make wrong predictions. False
positives increase due to overfitting. High variance and low bias result
in overfitting, also known as generalization error.
Underfitting occurs when the model generates accurate predictions
on the training set. Underfitting causes false negatives to increase.
High variance and high bias result in under fitting.
Appropriate Fit or the statistical fit is the ideal state between
overfitting and underfitting (Figure 1.8).

Figure 1.8 Types of Fit

 CNN IN COMPUTER VISION 11

Understanding Padding and Stride

The two main parameters that decide the behavior of the CNN are
padding and stride. Both these parameters have impact on the size of
the data.

Padding

While processing the images in the convolutions layer, there are

possibilities that the features on the edges are lost. Also called the
pixel loss at the edges, it might hold the key features required from
the images. For the first layer of the convolution, this loss might
seem to be less, but for a deeper CNN, the loss is significant and
affects the prediction accuracy of the network. To prevent this, the
sides of the images can be padded with the empty pixels, i.e., with 0
intensity value, as shown in Figure 1.9. The optimal number of layers
of padding “P” is calculated using
F 1
P= ,
2
where, F is the size of the filter applied to the image. If the resulting
value of the above formula is in fraction, the division quotient is rounded
off to the next higher value. When we apply the equation, the padding
type is called same padding, since this helps in keeping the size of the
output image after convolution operation same as that of the input
image, which usually reduces by the effect of the stride. There is another
type of padding, known as valid padding. Valid padding means no
padding, i.e., no padding will be done. In valid padding, the part of the
image that does not fit the filter would be dropped off.

Figure 1.9 Padding

12 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Stride

Stride is an important component of convolution layer. It denotes

how many pixels are shifted by the filter, while it moves over
the input image during the feature extraction process (see
Figure 1.10). This process can also be stated as convolving. The value
of stride depends on the connection of the convolution layer and the
pooling layer, the filter size and the properties of images in data.
Usually the value of stride is 1. If the values of stride are higher (i.e.,
more than 1), the dimension of the output is reduced, leading to
lesser computation, but there could be loss of feature representation.
The value of stride must be an integer. The value of stride plays a key
role in trade-off between the accuracy and the running time of the
neural network.

Figure 1.10 Stride

 CNN IN COMPUTER VISION 13

Parameters and Hyper Parameters

Model parameter is a variable whose value is set automatically and is

estimated from the training dataset. The values are learned by fitting
the model to the data, i.e., they are estimated by the optimization
algorithm from the data. These parameter values are used by the
model to make predictions. The parameters are different for different
applications.
Example: Weights in neural network intercept value in linear
regression.
Hyper parameters are set manually before the initiation of the
leaning process. These hyper parameters are required for estimating
the model parameters. Hyper parameters play a vital role in ensuring
good accuracy from the neural network since it controls the learning
process of the model. Therefore, it is very important to select an ideal
set of hyper parameters and initiate them to optimal values. This
process is called hyper parameter tuning. In most cases, it is done by
trial and error method. This can also be done by grid search, random
search, and Bayesian optimization.
Examples: learning rate, size of neural network, no. of neuron,
epoch size, batch size, etc.

CONV Layer

The convolution layer, in short CONV Layer, extracts features from

an input image by applying a large number of filters in parallel, which
are specific to the dataset and the predictive modeling problem. The
input image is the RGB model, which comprises three channels, red,
blue, and green. The 3-D volume of neurons performs the convolution
operation. The weights for the neuron are nothing but the filters.

Filter

Filter can also be called kernels. Each filter contains a set of weights
that are for a specific purpose, like for edge detection, filter for line
detection, sharpening, Gaussian blur, etc. The size of the filter is
mostly 3×3 or 5×5, but it can also be changed based on the require­
ment. The filter uses Sliding Window Protocol. Since the size of the
14 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

filter is smaller than the input image, the filter starts at the start of the
input, and then slides across the entire input. The filter acts as a feature
detector and sifts the input. It retains the required information in
the feature map and removes the irrelevant details. For example, if a line
detector is the filter used, upon passing over input, the output contains
the information only about the lines that are present in the input.

Feature Map

The result of the convolution layer is a feature map. The feature map
is a result of repeated application of the same filter in different
patches of the input image. The resulting feature map contains the
location and intensity of the detected features of the image. The size
of the output feature map is given by
W F + 2P
Dimension of the Feature Map = + 1,
S
where, W – Size of the Input Data (Image)
F – Size of the Filter
P – Padding
S – Stride

Convolution Operation

Convolution operation between the input image and the kernel is the
key idea behind the convolution neural network. Convolution op­
eration is a linear operation that involves the dot product of the input
image with the kernel. Convolution operation is indicated by “*.”
Consider an input image of size 5×5, i.e., W = 5, with zero pad­
ding, P = 0, and the size of the filter is 3×3, F = 3 and Stride S = 1,
based on the formula given above for the dimension of the feature
map, upon substituting
Dimension of Feature Map = (5–3 + 0/1) + 1 = 3.

Hence the size of the feature map would be 3×3.

The feature map given above is calculated by sliding the filter
across the input and doing dot product.
 CNN IN COMPUTER VISION 15

Similarly, the filter slides next step, and when all the values have
been covered horizontally, it then slides downward, i.e., vertically to
the beginning of the next row, as shown in Figures 1.11 and 1.12.
This way the convolution process takes place. The output feature
map is then passed to the ReLU activation function. This ReLU
layer is comprised within the CONV Layer. The purpose of the
ReLU function is to remove the linearity in the feature map obtained
and make it more meaningful. It removes negative values from the
feature map by replacing them with zero.

Figure 1.11 Convolution Layer

Figure 1.12 Filter Slides throughout Image

16 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Key Points about Convolution Layers and Filters

• The depth of the filter must be same as that of the input to

ensure that all the input layers are processed. Depth denotes the
number of color channels.
• Deeper layers extract more features but are computationally
expensive.
• Filters are automatically initialized on the basis of their normal
or Gaussian distribution for each region.
• The number of filters in the initial layers should be less, and the
deep hidden layers have twice or the thrice the number of filters
in the initial layers.

Pooling Layer

Pooling layer is added either after a single layer of convolutional layer or

after two layers of CONV Layer; a single layer of pooling is added.
When the images of the dataset are very large, the corresponding
feature map also will be of higher dimension. This subsequently in­
creases the computational cost. Pooling layer removes the unwanted
parameters, thereby reducing the size of the feature map.
Pooling, also called sub-sampling or down sampling, reduces the
size of each output feature map, but pooling layer ensures that the
important features are retained. It does down sampling by removing
unnecessary details from the feature map. The key idea behind
pooling is that the relative location of the feature is enough and the
exact location of the feature is less important. Pooling is generally
done over 2×2 square values of the feature map.

Key Points about Pooling Layer

• Pooling reduces the spatial size.

• It reduces the parameter size.
• It reduces the computation and the memory required.
• It operates independently on every channel of the feature map.
• Stride decides how the elements are pooled in a way similar to
CONV layer.
 CNN IN COMPUTER VISION 17

Types of Pooling

1. Max Pooling: Maximum of the four values in the 2×2 square

selected from the feature map is the selected value as shown in
Figure 1.13.
2. Average Pooling: Average of the four values from the 2×2
square values of the feature map is chosen as the final value (see
Figure 1.14).
3. Sum Pooling: Sum of the four values from the 2×2 square values
of the feature map is chosen as the final value (see Figure 1.15).

Figure 1.13 Max Pooling

Figure 1.14 Average Pooling

Figure 1.15 Sum Pooling

18 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 1.16 Forward Propagation

Forward Propagation

The convolution neural typically has an input layer, hidden layers

consisting of CONV and POOL layers, fully connected ANN, and
the output layer, as shown in Figure 1.16. Softmax activation func­
tion is often preferred for generating the output. The convolution
neural network consists of training part and testing part. In the
training part, optimization can be done on the basis of the accuracy
obtained, which is carried out during the backward propagation.

Calculating the Parameters

Parameters are basically the learnable elements from the data. They are
weights that are learnt during the training process, for example see
Figure 1.17. These weight matrices are used to contribute to the pre­
dictive power of the model that is changed during the back propagation.

Figure 1.17 Number of Parameters

 CNN IN COMPUTER VISION 19

Steps in calculating the parameters

1. Input layer has nothing to learn. So the parameter count is 0 in
input layer.
2. CONV layer: The formula to find the number of parameters
learnt from the convolution layer is
No. of Parameters in CONV Layer = ((m n d ) + 1) k ),

where,
m – width of the filter
n – height of the filter
d – no. of filters in the previous layer
k – no. of filters and 1 indicates the bias term
3. Pooling layer: Learning does not take place in the pooling layer.
Hence, the number of parameters is 0.
4. Fully connected layer: The fully connected layer has the
highest number of parameters. The formula to calculate the
number of parameters in FC layer and the softmax activation
function is

No. of Parameters in FC Layer = (c p) + 1 c,

where,
c – no. of neurons in the current layer
p – no. of neurons in the previous layer and 1 is the bias term

Activation Shape and Size

Activations maps are just another name for the output feature map of
the convolutional layers. The activation shape is nothing but the
dimensions of the output of the particular layer in the format (height,
width, and no. of channels). Activation size is the product of the
height, width, and the number of channels.
The forward propagation computes the result of the learning
process and saves the intermediate values that are required for the
gradients computation. The gradients computation is used for cal­
culating the loss function and the optimization (see Figure 1.18).
20 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 1.18 Forward and Backward Propagations

Backward Propagation

Once the forward propagation is done and the output is predicted,

the loss function or the cost function evaluates the model. If the
computed value of the loss function is high, it means that the model
has not been trained properly. If the value is low, it denotes that the
model is working with good accuracy. The loss function is indicated
by L. The loss function is selected on the basis of the application for
which the CNN architecture is modeled.
Whenever, L is high, there’s a need to back propagate and alter the
weights used in each layer in order to optimize the working of the
model. The weights here refer to the filter values. Given a CNN and
the loss function value, the backward propagation computes the
gradients of the loss function with respect to the weights and biases
in the network. In gradient computation, differentiation is done,
which denotes the rate of change or difference between the expected
output and the actual output obtained. It is important to note that
the backward propagation of the convolution operation is also a
convolution operation with flipped kernels.
dz dz
In Figure 1.18, dx
and dy
are local gradients and are calculated
and saved in memory during the backward propagation itself. The
local gradients are calculated from the intermediate values stored
in the forward propagation. By applying the differentiation chain
rule and the Jacobean notations, the weights can be altered, and
then the model is again forward propagated to find the new
output. This process is carried out repeatedly until the required
accuracy is obtained. This process can also be called fine-tuning of
the CNN.
 CNN IN COMPUTER VISION 21

Optimizers

During the backpropagation, to minimize the value of the loss

function, we use optimizers. An optimizer moulds the neural net­
work to attain maximum accuracy. It is guided by the loss function.
Adam optimizer is most commonly used for the CNN architecture,
since it gives 99.2% accuracy.
Adam optimizer stands for adaptive moment estimation. Adam
optimizer adds the gradients of the previous layers to the current
layer under the concept of momentum.

Other Ways to Improve the Performance of CNN

While backpropagation and parameters tuning act as the primary

method to the performance and the accuracy of the CNN, there are three
other proven ways to improve the model, which are described below.

Image Data Augmentation When the amount of data used for training
is less, the model may not be able to learn the key features properly.
This results in lower accuracy of the model. This can be overcome by
the concept of augmentation. Wherein the image data, which are al­
ready used for the purpose of training, can be augmented with various
operations, like rotation, reflection, tilting, performance of random crop,
and alteration of the image properties as shown in Figure 1.19. These
images add the count to the dataset, enabling the model to learn better.

Figure 1.19 Augmentation

22 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 1.20 Early Stopping

Deeper Hidden Layers Having multiple convolution layers enables

better features extraction and better learning. This increases the ac­
curacy of the model. It is also important to note that, better accuracy
is provided by a deeper network and not essentially by a wider net­
work. By wider network, we mean large number of filters in a single
layer of CNN. The wider layers are better at learning the features of
the input images but deeper networks with hierarchical structure are
good in generalizing the learning process. This generalization would
improve the accuracy for a wide variety of dataset.

Early Stopping As the convolution neural network performs the

training process in multiple iterations, it is important to ensure that the
model does not start to overfit. To ensure this, the iteration has to be
stopped at an early stage when the loss function value is less. To find
the ideal early stopping point, the network must be fed with the vali­
dation set (see Figure 1.20).

Application of CNN

CNN is mainly used to solve the problems of computer vision like

image classification, object localization in an image, segmentation,
Object detection in the image, style transfer, reconstruction, color­
ization, etc. Of all these applications, the most commonly used ap­
plication is image classification.

Image Classification

Image classification is of two types: binary classification and multi­

class classification. Binary classification classifies object between the
 CNN IN COMPUTER VISION 23

two given data classes, whereas multiclass classification classifies the

object as one among the many data classes.

Binary Classification
Using python, tensorflow and keras libraries, we can classify
images using the CNN. Following are the steps to create a classifier:
1. Create a folder for the 2c classes of object in the python.
2. Open the python notebook and import the required libraries,
like sequential, Convolution2D, Maxpooling2D, Dense, etc.
3. The next step is to initialize the CNN architecture, using the
command model = Sequential(). Here, model is the name of the
architecture and we are creating it in sequential pattern.
4. Add the convolution layer to the convolutional layer, model.add
(Convolution2D(filters=32, input_shape = (64, 64), kernel_size =
(3,3,activation = ‘relu’))
Here, the parameters are no. of filters = 32, size of the filter –
3×3, Size of the input – 64×64 and the activation function
is ReLU.
5. Add the pooling Layer.
model.add(Max_Pooling2D(pool_size = (2, 2)).
Here, pooling window size is 2×2.
6. Similarly add consecutive layers of convolution layer and
pooling layer.
7. After adding the conv and pooling layers, the output is in the
form of a feature map. A feature map is a matrix. As we know,
the last layer of CNN architecture is a fully connected neural
network and it takes only an array of data as input, we convert
the matrix into a 1-D structure using flatten.
model.add(Flatten())
8. Add fully connected neural network to the sequential network.
model.add(Dense(units = 128, activation = ‘relu’), Here, units
indicate the number of neurons.
Output layer:
model.add(Dense(units = 1, activation = ‘sigmoid’)
Here, the activation used is Sigmoid, because the application is
binary classification and the output should either be 1 (if class 1
object is present in the image) or 0 (if class 2 object is present).
24 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

9. Now that the model has been developed, it can be compiled. At

the time of compilation, the loss function to be used and the
optimizer has to be specified.
model.compile(optimize = ‘adam’, metrics = [‘accuracy’], loss = ‘bi­
nary_crossentropy’)
Metrics is a list of elements, and multiple metrics can also be
optimized.
10. Now, the model is developed. Training Data has to be fed to
the network.
The data has to be rescaled to a fixed size and pre-processed
using the keras library.
from keras.preprocessing.image import ImageDataGenerator
train_datagen = ImageDataGenerator(rescale = 1./255, zoom_­
range = 0.2)
11. Feeding the training data and test data:
training_set=train_datagen.flow_from_directory(‘data/trai­
ning_set’, target_size=(64,64),batch_size=32, class_mode =
‘binary’)
test_set = test_datagen.flow_from_directory(‘data/test_set’, tar­
get_size = (64,64), batch_size = 32,class_mode = ‘binary’)
Here, directory refers to location where the dataset is stored,
and target size indicates the size to which the data would be
resized. Batch size indicates the number of data to be passed for
each iteration and class mode here is binary, since the appli­
cation is binary classification.
12. Training the model is done by,
model.fit_generator(training_set, samples_per_epoch = 2000,
nb_epoch = 15, validation_data = test_set, nb_val_samples = 200)
Here, generator generates the training process with training set,
samples per epoch indicates the number of batches i which the
data have to be fed to the network, nb_epoch indicates the total
number of iteration, validation data indicates the set of data,
which is used for testing purpose, and nb_val_sample indicates
the number of samples to yield before stopping each iteration.
13. To view the summary of the model,
model.summary()
This command gives the accuracy obtained and the number of
parameters generated by each layer.
 CNN IN COMPUTER VISION 25

RCNN and Object Detection

Object classification algorithm identifies whether a particular class

of object is present in a given input image or not. If the object is
present, it is equally important to identify the location of the
object in the image. Image detection is done to identify whether
an object is present and then locate the object(s) with the help of a
bounding box. Object detection is important for various applica­
tions, such as vehicle detection in self-driving cars, human activity
recognition in the surveillance feed, etc. But to implement this,
the normal CNN architecture is not efficient enough to compute
the bounding boxes. The example of bounding box is shown in
Figure 1.21.
The main reason for the regular CNN architecture to be not
able to compute the bounding box is that the sliding window
pattern of filtering is not efficient enough to identify the bounding
box. One more difficulty with object detection is that there could
be any number of objects of interest, and each might be of dif­
ferent aspect ratios and different locations. In order to identify the
location and identify the number of objects, using the traditional
CNN architecture would be computationally very expensive and
less efficient. Hence, instead of using the sliding window protocol,
selective search protocol is used. The various CNN-based archi­
tectures to do object detection are R-CNN, Fast R-CNN, and
Faster R-CNN and YOLO.

Figure 1.21 Bounding Boxes

26 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Region-Based Convolutional Neural Network (R-CNN)

Instead of selecting large number of sub-regions from image,

R-CNN uses the selective search, to extract just to 2000 sub-regions.
The Selective Search Protocol is given below:
a. Generate candidate sub-segmented candidate region.
b. Using Greedy algorithm, the similar regions of the image are
recursively combined to a larger image.
c. Use the combined region to produce the complete image.
The image proposals are warped into 4096 dimension squares and are
fed to the CNN. The CONV Net performs the feature extraction
process and then passes the identified features to the support vector
machine for classification and the offset (Location information) to
the bounding box regressor.
Limitations of R-CNN:
• Since selective search algorithm does not learn, there could be
possibilities of poor candidate region proposals.
• It takes around 47 seconds for the R-CNN to process a test
image.
• Training is also slow, since it takes 2000 candidates from each
image.
• It is computationally expensive.
The process of the R-CNN is shown in Figure 1.22

Figure 1.22 R-CNN

 CNN IN COMPUTER VISION 27

Fast R-CNN

To overcome the limitations of the R-CNN, the Fast R-CNN was

introduced. It is faster than the R-CNN, and hence the name. Fast
R-CNN does not use selective search algorithm. In the fast R-CNN
architecture, the input image is given as input to the CNN archi­
tecture. The CNN layer extracts the feature and produces the feature
map. It is important to note that the object proposals are extracted
using the selective search algorithm from the feature map. Using the
object proposal, the co-ordinates of the object region are identified in
the feature map and stored as region of interests projection. The
pooling layers, with the help of region of interests projections, pool
the feature map to fixed length feature vectors of the objects of in­
terest as object proposals, and feeds them to the fully connected
layers and the softmax for the classification tasks. The offsets of the
identified regions are given to bounding box regressor to create the
bounding box around the object.
The architecture of Fast R-CNN is shown in Figure 1.23.

Figure 1.23 Fast R-CNN

28 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Faster R-CNN

Though Fast R-CNN is faster than the R-CNN, it still uses the selective
search algorithm, which is a time consuming process. Similar to the Fast
R-CNN, in Faster R-CNN also, the input image is directly fed into the
CNN. Then the output feature map is separately fed into a pooling layer
and a special Region Proposal Neural Network. This network generates
the object proposals and is then passed to the region of interest pooling
layer. Now that the pooling layer processes on both the feature map and
the object proposals, it identifies the regions of interest, warps them to
squares, and passes these to the classifier. The classifier then classifies the
object and show cases the location using the bounding box. The use of a
separate region proposal network instead of selective search algorithm
makes fast R-CNN even faster than the other two variants (see
Figures 1.24 and 1.25).

Figure 1.24 Comparison of Execution Speed

Figure 1.25 Faster R-CNN

 CNN IN COMPUTER VISION 29

Figure 1.26 YOLO

You Only Look Once (YOLO)

The R-CNN architecture variants use separate algorithms and networks

to capture the location of the object. YOLO does not use any additional
network for boundary box regression. Instead, a single convolution net­
work is used to perform the bounding boxes and the class probabilities.
When an input image is fed into the YOLO architecture, it splits
the image into grids of S×S dimension. S is selected based on the size
of image. Within each grid, we take n boundary boxes. The boundary
boxes predict the class probability and the offset values. The
predicted probability is compared with the threshold and if it is
above the threshold value, it is located as a particular object (see
Figure 1.26).

Transfer Learning

Transfer Learning is a machine and DL technique that uses

knowledge gained in the previously trained model like features,
weights, etc. for training new model. Transfer learning aids to better
accuracy and overcome the difficulties in training such as setting
optimal hyper parameter values, overcoming training data shortage,
choosing the ideal optimization technique, etc.

Transfer Learning for CNN

Most CNN Architectures are not developed from scratch, instead the
weights and the layers of the previously modeled architectures that
30 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

are well trained over a larger dataset and have attained higher effi­
ciency are used in modeling CONV net for a new application.
Transfer learning can be applied in the following two ways for the
CNN architecture:
1. Pre-trained Model as Feature Extractor: This method uses
the feature extraction layers of the pre-trained algorithm,
but the classification or the prediction process is done by the
layers that are newly modeled for the task. This can also be
called as freezing the convolutional layer.
2. Fine-Tuning Pre-trained Models: This method fine tunes the
feature extraction layers of the existing pre-trained models and
may or may not use the final fully connected layers of the pre-
trained model.

Fine-Tuning or Freezing?

It is important to decide the type of the transfer learning to be used

while developing the CNN architecture. Few points to identify the
ideal method are:
• When the dataset for the new task is less, it is better not to
fine-tune.
• When the dataset for the new task is large, but similar to the
pre-trained model’s dataset, fine tuning might only produce a
small change in accuracy.
• When the dataset for the new task is less, but is different from
pre-trained dataset, it is better to train only the classifier or final
output layer.
• When the dataset for the new task is large, and different from
the pre-trained model’s dataset, fine-tuning but feature ex­
traction and the classifier or developing a new CNN archi­
tecture could yield better result.

Neural Style Transfer (NST)

Neural style transfer is one of the applications of image styling and

transformation technique. NST can be performed by the CNN.

Input Image + Style Image Styled output image

 CNN IN COMPUTER VISION 31

Steps in NST:
1. Both the input x and the style image a are converted to equal
dimensions.
2. A pre-trained convolution neural network can be used for the
style transfer. Transfer learning is done here. Ex. VGG16 ar­
chitecture can be used.
3. The style components like color, basic shapes, and patterns are
extracted from the style image, denoted as S(a)
4. The content component featuring the actual object and features
in the input image, other than color and styling elements, are
extracted, denoted as C(x).
5. After extracting the required components, they are fed into the
CNN and an output styled image y is obtained.
6. The losses for NST are specific. They are
1. Content Loss: The difference in content between the input
and the styled image, denoted as C (x ) C (y ) .
2. Style Loss: The difference in style factors between the style
image and styled image, denoted as S (a ) S (y ) .
3. Total Loss: The total loss is denoted by L.
L (y ) = C (x ) C (y ) + S (a ) S (y )

The total loss is a regularization loss.

7. Once the loss function is found, optimization has to be done
through back propagation, until the loss is minimized. For
NST, with large number of parameters, Limited-Memory
Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) optimization
algorithm works the best.

The NST is done by various mobile applications for photography

editing and art works. One such application is Prisma. Figure 1.27
shows an example of NST.
32 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 1.27 Neural Style Transfer

 2
TRENDS AND TRANSITION IN THE
MACHINE LEARNING (ML) SPACE
SRUTHI ANAND1, N. SUSILA1,
AND USHA SAKTHIVEL2
1
Department of Information
Technology, Sri Krishna College of
Engineering and Technology,
Coimbatore, India
2
Dean, Dept. of Research and
Innovation, Rajeswari College of
Engineering, Bengaluru, India &
Prof. & Head, Dept. of Computer
Science & Engineering,
Visvesvaraya Technological
University, Belagavi, Karnataka,
India

Contents

Introduction to Machine Learning 34

Motivations for Machine Learning Algorithms 36
Data Analytics and Machine Learning 37
Supervised Learning 38
Classification 39
Naive Bayes (NB) 39
Support Vector Machine (SVM) Classifier 40
Types of Learners 40
Eager Learners 40
Document Classification 41
Illustration 41
Logistic Regression 42
Practical Use Cases 42
Weather Prediction 42
Patient’s Health Monitoring 42
Unsupervised Learning 42
Gaussian Mixture Models 43

DOI: 10.1201/9781003119838-2 33
34 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Hidden Markov Models 43

Types of Unsupervised Learning 43
Applications of Unsupervised Learning 45
Process of Machine Learning 46
Gathering Input Data 48
Data Preparation 48
Choosing a Model 48
Training 48
Evaluation 49
Parameter Tuning 49
Prediction 49
Machine Learning – Practical Use Cases 49
ML in Banking 50
ML in Agriculture 51
ML in Health Care 51
ML in Automobile Industry 52
Implementation 52
Pandas 53
Matplotlib 53
Jupyter Notebook 53
Google Colab 53
scikit-learn 53
Conclusion 54
References 54

Introduction to Machine Learning

Machine learning (ML) is a subset of artificial intelligence. ML is a

branch of study where algorithms are not explicitly needed to be
programmed, rather it directly allows computers to train on a par­
ticular set of data and the use of statistical analysis to predict the
output. Using these, models are built which in turn enable computer
systems to involve in the process of decision-making. The concept of
ML is based on the idea that the computer programs can access and
get trained by themselves which they use to learn for future processes.
ML is one of the most commonly used technology floating around
that can be used in any domain. The technology as perceived by Alan
 TRENDS AND TRANSITION IN THE ML SPACE 35

Turing that is “Can machines think” is now transformed to “Can

machines do what humans think.” [1] The ability to solve complex
problems is not only the task of humans, but can also be embedded in
the systems giving them the capability to think and formulate solu­
tions based on analytical models.

The field of ML has benefited a number of applications, like

optical character recognition, that help in conversion of texts into
movable types, user behavior prediction, facial recognition, and re­
commendation systems for identification of the interests of the users
in different social media, as shown in Figure 2.1.
As quoted by Tom Mitchell – “A computer program is said to
learn from experience E with respect to some class of tasks T and
performance measure P, if its performance at tasks in T, as measured
by P improves with experience E.” For instance, let us assume that we
have a group of animals and we have to classify them and give the
count. So how do we proceed? The answer is – for a human eye, it is
easy to identify the animals based on different features, like their
eyes, shape, color, etc. But for a machine how can it understand? The
machine must be trained to identify the different features so that it
can classify them. ML now deals with processing different kinds of
data and identifying the features based on which the prediction can
36 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 2.1 Applications of ML

be made. ML helps in many complex tasks, like object recognition,

summarization of the predicted object, classification, and clustering,
and recommends and suggests the kind of systems to be used for
different purposes. ML techniques may be used to gain useful in­
sights from the data that can be used for decision-making.

Motivations for Machine Learning Algorithms

Have we ever thought of how these algorithms have evolved. Let’s

just take a look at things that are influencing us in our daily lives.
Considering the way we handle Google, the search engine processes
more than 40,000 searches per second. Have we thought how Google
brings us the most accurate result? There are many other examples
that we can quote, such as the prediction of a route in our Google
Maps, the Spam folder, the options of autocomplete, etc. Can we
imagine what if we do not have a system like this? The answer for all
these is machine learning. The key idea behind this is given the data,
 TRENDS AND TRANSITION IN THE ML SPACE 37

we can extract useful information from it. The idea behind ML is to

reduce the tasks of humans in terms of decision-making in complex
task processing, build ecofriendly systems by receiving feedbacks
instantly and develop technology that can work like humans.
The great motivation factor toward ML algorithms is the thought to
develop some technologies that could be derived from human behavior.
These technologies have been increasingly used in every industry
without realizing that they are being used. ML has evolved from pattern
recognition where computers were able to learn with the help of patterns
without being explicitly programmed. With the rise of ML algorithms,
complex calculations can be applied to big data for faster processing.

Data Analytics and Machine Learning

Data analytics is the field of science that is used in analyzing raw data
and draw conclusions from it. It draws patterns and is used in many
specialized systems that are integrated with ML algorithms. Data
analytics can help to quantify the data and track them to make smarter
decisions. Data analytics works with answering the queries based on
some existing data, while ML is not focused toward the queries, rather it
works with datasets in unstructured ways to get the output. ML is much
more toward asking the right questions, and data analytics provides a
way to answer those questions and find actionable data. Thus both data
analytics and ML are interconnected, with data analytics having micro
scope in specific applications and ML having macro scope.

Let’s define some important terms:

Artificial Intelligence – The ability of the machines to mimic

human behavior and exhibit the traits that are associated with the
human mind, such as self-learning and problem solving [2].

Machine Learning – A subset of artificial intelligence that involves

the study of different algorithms that learns and improves auto­
matically with experience. ML algorithms are broadly classified into
four types (Figure 2.2):
1. Supervised Learning
2. Unsupervised Learning
38 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Machine
Learning

Supervised Semi
Unsupervised Reinforcement
Learning Supervised
Learning Learning
Learning

Classification Clustering
Markov
1. Naive Bayes 1. K-Means Decision
2. SVM 2. KNN Process
3. Decision trees

Regression Dimensionality
Reduction
1. Linear
1.PCA
2. Logistic
2.SVD

Figure 2.2 Categories of ML

3. Semi-Supervised Learning
4. Reinforcement Learning

Supervised Learning

A model is built using a set of input data that are trained called training
data to make the system predict the output. The predicted data may not
always be correct and in this case, the data are still trained until the
model reaches a desired level of accuracy. The term supervised learning
is analogous to a teacher supervising the activities of the students [3].
This learning method takes an input data (X ) and is applied to an
algorithm to predict the output (Y ). In short, it is denoted as
Y = Function(X)

Supervised learning techniques are further classified into two, namely:

Classification: The process of classification is to predict if the output

belongs to a particular category viz {true or false}, {red, blue or
green}, {positive or negative} etc.
 TRENDS AND TRANSITION IN THE ML SPACE 39

Examples include classifying the emails into spam and not spam,
classifying documents as confidential or not, predicting the user
behavior and classifying as churn or not, classifying a tumor if it is
malignant or benign.

Regression: The process of regression is identifying and predicting

continuous values based on either a single value or a set of values.

For example, regression can be used to predict the salary of an em­

ployee based on age, calculate the BMI based on the weight, and
calculate the price of a car based on its year and the number of
kilometers it has run, etc.

Classification

Classification is a method to understand, recognize, and group based

on a particular label or category. There are a number of algorithms
for classification, such as naive Bayes classification, support vector
machine, decision trees, and random forests. These algorithms take
some input data and are trained such that the subsequent data may fit
into one of the categories [4]. Let us analyze and understand these
algorithms in the following descriptions:

Naive Bayes (NB)

It is a classification algorithm that is based on the Bayes theorem, by

the use of conditional property. Bayes theorem helps to determine
the probability of an event by considering the probability of another
event that has already occurred. Mathematically, it is written as:
P(A|B) = P(B|A). P(A)/P(B)

The probability of A can be determined assuming B has occurred.

Here B is the evidence and A is the hypothesis. These classification
algorithms can be used for categorizing a class of data, given the
features as inputs. This algorithm works by taking into consideration
that all the features and the input variables are not dependent on each
other. This algorithm treats all the features independently to obtain
the result. NB can be used in text analysis and pattern recognitions [5].
40 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Few examples that can be predicted using NB are as follows:

1. Finding out if a person has diabetes or not considering the age,
diet, pressure, gender, etc.
2. Identifying the weather conditions based on the temperatures,
humidity, etc.
3. Identifying a particular climatic condition if it is suitable for a
particular sport.
There are different classifications of NB such as Gaussian NB,
Multidimensional NB, and Bernoulli NB used in various dimensions.
This algorithm is generally very fast and can be used for both binary
class and multidimensional classifications. This NB classification
algorithm can be used to train small datasets and make predictions on
large volumes of data. Since the algorithm takes in all features in­
dependently to analyze the data, this process could be more complex
if it has multiple attributes [6].

Support Vector Machine (SVM) Classifier

SVM are algorithms that can be used to solve classification and re­
gression problems, but it is used for classification in majority of the cases.

Types of Learners

It is a simple algorithm, where the generalization of algorithm takes

place after a query is raised. Though they consume less time in
training, they result in increase of prediction time.

Example: K-Nearest neighbor

Eager Learners

In this algorithm, the execution happens immediately and produces

desirable output. Though they consume a lot of time in training, they
result in less predicting time. They can be optimized.

Example: NB, Decision trees

Example for Classification:
 TRENDS AND TRANSITION IN THE ML SPACE 41

Document Classification

Document classification is a process of defining documents and separating

them into distinct classes based on their contents. This classification will
be helpful for people who deal with enormous amount of contents.
The words from a document can be used as “features” to help in
the prediction of the classification of a document.

Illustration

Let’s consider two reference classes.

Reference class 1 Reference class 2

Welcome to machine learning Happy coding

To classify these documents, we start by taking all of the words in the

two reference documents and creating a table or vector from these words.
(Welcome,to,machine,learning,happy,coding) class
We can create a vector by assigning a number 1 if the word exists
in the training document and number 0 if it doesn’t exist, for each of
the training documents.

Welcome To machine learning Happy Coding

1 1 1 1 0 0 Class 1
0 0 0 0 1 1 Class 2

Now let’s assume a new document comes with the tag,

“Welcome to python learning,” for this document, the word vector
becomes like

Welcome to machine learning Happy Coding

1 1 0 1 0 0 Unknown
class

If we compare this vector for the document of unknown class to

the vectors representing our two reference document classes, it closely
resembles the vector for class 1 document,

Unknown class: <1,1,0,1,0,0>

42 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

CLASS 1: <1,1,1,1,0,0> (matching vectors: 5)

CLASS 2: <0,0,0,0,1,1> (matching vectors: 1)

We can possibly label the data as class 1 (maximum number of

matching vectors). This is one of the examples of statistical natural
language processing method.

Logistic Regression

This is a technique of ML algorithm, where one or more independent

variables results in a desired output. In this case, possibly we have only
two outcomes. The ultimate aim of regression is to achieve best fit. It
is based on predictive analysis. In this, we can assume the probability
with the help of log function.

Practical Use Cases

Weather Prediction

Weather prediction can be done using logistic regression. The pro­

cess begins with analyzing the previous data and records of weather
report, thereby predicting the desired weather specification for the
given day. But with the help of logistic regression, we can only
predict whether it will happen or not [7].

Patient’s Health Monitoring

If a patient is admitted in hospital, his health is to be monitored. The

severity of illness can be predicted with the help of his history in medical
report. Thereby, we can predict if his severity is high/low of any disease.
We can also predict risk factors of a specific disease he is suffering from.

Unsupervised Learning

Unsupervised learning is learning algorithms that do not require

labeled responses and can draw their own references. Cluster analysis
is the commonly used unsupervised learning algorithm that takes
Euclidean or probabilistic distance as a metric. Clustering algorithms
include the following:
 TRENDS AND TRANSITION IN THE ML SPACE 43

This clustering creates a multilevel cluster hierarchy by building

cluster trees.
It is the simplest of unsupervised algorithm. As the name suggests,
clustering divides the data based on certain patterns. These unsupervised
algorithms derive conclusions from the input vectors from datasets
without using labeled outcome. The process is partitioning the data into
distinct clusters based on the means that is the centroid of a cluster.
Unlike supervised learning, the outcome is not labeled; instead, it
generally groups the data. They are represented by data points, and the
cluster that is formed must be relevant.

Gaussian Mixture Models

This includes modeling the clusters using an approach known as soft

clustering. They are also probabilistic models that distribute the data
points using the soft computing approach. Gaussian mixture models
are like Gaussian distribution or the normal distribution. Gaussian
considers both mean and variance to update centroid, whereas
k-means consider only the mean.

Hidden Markov Models

This model uses observed data to recover the sequence of states.

These are used in bioinformatics, data mining, and pattern mining.
Unsupervised learning algorithms allow users to perform more
complex tasks compared to supervised learning.
Unsupervised learning is unpredictable compared with other learning
methods. Unsupervised learning algorithms include clustering, anomaly
detection, neural networks, etc.

Types of Unsupervised Learning

Unsupervised learning problems can be divided into clustering and

association.
1. Clustering

Clustering technique deals with observing certain structure or pattern

from uncategorized data. They process the data by finding the
44 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

clusters from the dataset. They also consider the granularity and
make adjustments based on it.
There are different types of clustering:
1.1 Exclusive (partitioning)
In this method, grouping of data is done belonging to one cluster only.
Example: k-means
1.2 Agglomerative
Here, all the data are clusters. In order to reduce the number of
clusters that are formed, union operation is performed.
Example: Hierarchical clustering
1.3 Overlapping
Here fuzzy datasets help in clustering the data. There may be two or
more clusters associated with separate degrees of membership. Here,
the data are associated with an appropriate membership value.
Example: Fuzzy C-Means
1.4 Probabilistic
This technique uses probability distribution to create the clusters.
Example: Following keywords
“man’s shoe.”
“women’s shoe.”
“women’s glove.”
“man’s glove.”
These can be clustered into two categories “shoe” and “glove” or
“man” and “women.”

Clusters can also be typed into:

Hierarchical clustering
k-means clustering
K-NN (k nearest neighbors)
Principal Component Analysis
Singular Value Decomposition
Independent Component Analysis
 TRENDS AND TRANSITION IN THE ML SPACE 45

2. Association
This enables association among the data that are available inside the
datasets. This technique is used to discover relationships between the
variables in very large databases.

Examples:

The cancer patients are grouped by their gene expressions.

Groups of shopper based on their browsing and purchasing histories.
Movies can be grouped by the rating given by the movie viewers.

Applications of Unsupervised Learning

1. Anomaly Detection
Anomaly detection is used to identify the uncommon items,
events, or observations from the normal data. In this case, the
system trains with large data which are normal instances. So, in
an unusual instance, the machine detects whether there is some
unusual behavior and claims it to be anomaly.
One example is credit card fraud detection, which can be solved
using anomaly detection techniques in ML. Here, the system
detects any unknown credit card transactions, which help in
preventing fraudulent activities.
2. Clustering
Clustering is the process of grouping the input dataset into dif­
ferent clusters or groups based on the criteria. Clustering can be
done when the exact information about the clusters is not known.
For example, in a supermarket, we can see how different items
are grouped and arranged. Also, e-shopping websites like
Amazon and Flipkart use clustering algorithms to group their
products according to the users’ needs.
3. Visualization
Visualization is the process of creating diagrams, images, graphs,
charts, etc., to deliver and communicate a piece of information.
For example, a football coach has some data of team’s perfor­
mance in a tournament, to find all the statistics about the matches
quickly. We can feed complex and unlabeled data to some vi­
sualization algorithm.
46 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

These algorithms can be represented as a 2-D or 3-D re­

presentation of the data as output in a plotted graph, which helps
in determining the winning formula (chances), correcting pre­
vious mistakes, and winning the trophy.
4. Dimensionality Reduction
Dimensionality reduction is the process of reducing the number
of random variables by using a set of principal variables. The
objective of this technique is to simplify the data without losing
too much information.
One of the methods to perform dimensionality reduction is
feature extraction, where all correlated features are merged to
one. This is a good approach as it can reduce the redundant data
and then can be trained for better accuracy of the system.
Reducing the dimensionality may result in loss of some in­
formation. Even if it speeds up the training, it will make the
system perform slightly worse.
So, one should use dimensionality reduction only if the training
is very slow. Otherwise, use the original data.
5. Social Media
In social media, we use unsupervised learning techniques to
detect frauds and to learn and detect the categories of people
based on different criterias.
Consider Facebook, for example, which uses face recognition; it
helps in identifying the people around the globe by identifying
each and every minute details of people living in different
geographical areas. Also, the news feed section is filtered on the
basis of their interests and likings.

Process of Machine Learning

There are seven different steps involved in the ML process (See

Figure 2.3):
1. Gathering Input Data
2. Data Preparation
3. Choosing a `Model
4. Training
Figure 2.3 The seven steps of the ML process
TRENDS AND TRANSITION IN THE ML SPACE
47
48 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

5. Evaluation
6. Parameter Tuning
7. Make Predictions

Gathering Input Data

This is the first step of ML process. In this step, we collect the data
required for processing. This includes both the quality and the
quantity, which in turn are dependent on how good the model can
be. For example, if we want to identify a particular object, the first
step would be to identify and gather inputs based on the color, shape,
size, texture, etc. Incorrect data would lead to incorrect representa­
tions and direct the model in a wrong way.

Data Preparation

After gathering the data, the next step is to prepare our data in terms
of ordering the data and removing the skews. This means our data
must be prepared in such a way that it is ready to be trained for the
model. The next important part of data preparation is to break the
data into training and testing data, where training data would con­
tribute to 80% and testing data would be 20%, which are used for the
evaluation. This step also concentrates in removing duplicates, dis­
carding incorrect data, and fine-tuning for better process.

Choosing a Model

Next is to identify the model that chooses the problem and the kind
of learning techniques to be employed. We already know that we
have a number of models that can be used for different purposes and
it is necessary to choose the right one.

Training

As the name suggests, training is the phase where our model ori­
ginally starts learning. The 80% of the data that we have allocated for
training are used to teach our model. To train a model, more
 TRENDS AND TRANSITION IN THE ML SPACE 49

experimentation must be done in order to implement. The usage of

weights, biases, and mathematical models help in making predic­
tions. Training is highly successful if the output is satisfying and the
outputs are obtained for different sets of input.

Evaluation

This phase actually identifies if the model is able to achieve profi­

ciency and accuracy. Here the model may come across various si­
tuations, which it would not have witnessed in its training phase.
Evaluation step helps to check if the goals are met, especially using
the data that were not used for training. This represents the per­
formance of the model in the real world.

Parameter Tuning

In this phase, we can tune the parameters by making certain mod­

ifications and check if we can improve the training to get better re­
sults. This may be the inclusion of some other parameters that were
not considered in the training and measure the performance. Another
way is to run our dataset for multiple times and gain more accuracy.

Prediction

The final step of ML process is prediction. This phase makes the model
ready for practical applications. This is the stage where the model starts
to perform its own conclusions without human interventions. This step
is actually the result of what a human is expected to do. The major
difference between human predictions and model prediction is that the
models can process huge amounts of data simultaneously, unlike hu­
mans. This process helps in reducing the human’s burden for decision-
making process and arriving at conclusions [6].

Machine Learning – Practical Use Cases

We have discussed the basics of ML above. Now there are a number of

trends that are making way in today’s market with the technology. As
reported by Statista, $28.5 billion were allocated in the technology of
50 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

ML during the first quarter of 2019, and we cannot even imagine the
increase its use currently [7]. A great number of ML technologies will
have a greater impact altering the basis of industries around the globe.
These include Regulation of Digital Data, Voice Assistants, Marketing,
Cyber Security, and just simply to every kind of industries [8]. As noted,
there are a wide variety of applications of ML in the fields like banking,
agriculture, healthcare, automobiles, and finance systems.

ML in Banking

The usage of ML banking applications has revolutionized the fi­

nancial world, making it better for people to use. The major areas are
to make credit decisions, fraud prevention, and personal assistants.
The banks usually collect a huge amount of data, by which they can
derive greater rules that help for credit scoring [7]. The data available
are trained using ML algorithms and eventually improve to make
decisions on applicants who are subjected to more credit scores. The
increase in automation has paved a way toward these technologies
where the risk involved can also be assessed in a few minutes, much
improved than human efforts. Every financial sector is very much
prone to frauds and preventing the frauds is a great deal in today’s
world. To detect frauds, the patterns of spending and the location of
the cardholder or the customer are analyzed and anomalies are de­
tected. In addition, the system can also find any kind of suspicious
activities that happen by considering the deviations or abnormalities
in the patterns [9].
 TRENDS AND TRANSITION IN THE ML SPACE 51

ML in Agriculture

The role of ML in Agriculture plays a pivotal role in our society as

Agriculture is one of the predominant fields in the global economy.
The lifestyle that we live need a great improvement and it is all
determined by the type of food we eat. Promoting and preserving our
economy is very much important. Thus growing healthy seeds and
preserving the crops, determining what kind of crop to be harvested
can all be analyzed by the use of ML algorithms. Many retailers use
this technology to create better crops. The other methodologies
which can be employed in the field of agriculture are determining the
nature of the crops to get the best yield, find bugs and diseases as­
sociated with the crops and also to predict when to water the crops. It
also helps to predict bug hunters. Agriculture robots are designed to
manage and carry out all essential activities from seed sowing, wa­
tering, monitor crops and detect diseases. The major part in em­
ploying AI and ML is to identify the soil, weather conditions and
also nurture the crops that will give good production and make the
crops healthy [10].

ML in Health Care

The application of ML in the field of healthcare helps patients to

analyze and predict the data in prior with the use of real-time data.
The real-time patient data are considered and taken from different
systems to predict the options of treatment. The major part of
52 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

healthcare involving ML is to identify diseases and diagnose the

disease. Another application of ML involves the personalized med­
icine predictions using the data documented earlier. IBM Watson
Oncology is one of the tools used for personalized medicine to track
patient history and leverage treatment options. Artificial Intelligence
and ML have a greater impact in predicting any kind of pandemic or
diseases that arise [11].

ML in Automobile Industry

ML helps car manufacturers to predict sales and even their manu­

facturing process. Vehicles use this concept to introduce driverless
cars. Also it can assist drivers toward their journey. It is possible to
predict and control the speed and movement of the cars, which can
also direct and notify any kind of emergencies or dangers and prevent
accidents. The usage of digital twin and ML helps in managing
automobiles and making any kind of operations that can be handled
and monitored virtually [12,13].

Implementation

There are quite a number of tools and packages for implementing

ML algorithms. The packages include pandas, matplotlib, pycharm,
Google Colab, scikit-learn, and keras.
 TRENDS AND TRANSITION IN THE ML SPACE 53

Pandas

It is an open-source Python library used for data analysis. These are

used for loading different file formats and manipulating the data.
These packages help in carrying out the important processes of ML,
like data preparation, cleaning, manipulation, modeling, and analysis.

Matplotlib

It is a 2-D python library used for plotting. It helps to visualize the

model and derive conclusions out of it. It can be used along with all
the different kinds of other packages for processing large amount
of data.

Jupyter Notebook

This includes web applications that are very interactive and can share
live code documents, and develop, execute, analyze, and discuss re­
sults. This has higher functionality and user scenarios.
It can also be used along with other tools and also supports dif­
ferent languages.

Google Colab

An interactive notebook that maintains the entire document and

helps to share the code and execute along with other users. It is also
similar to Jupyter Notebook. With Colab, we can make full use of the
libraries for analyzing and visualizing the data. It also helps to train
and test a large dataset and evaluate and make the predictions with
visualizations. As it is collaborated with Google, it makes use of
Google Servers and GPUS.

scikit-learn

It is the most commonly used Python library for ML. It consists of

many different tools capable of implementing all learning algorithms
and carrying out statistical modeling. scikit-learn has a lot of fea­
tures, like supervised learning algorithms, unsupervised learning
54 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

algorithms, and datasets and feature extraction. Installation of the

packages like Pandas, Numpy, and Scipy before using scikit-learn is
important.

Conclusion

Machine learning (ML) is a subset of artificial intelligence prevalent

in almost all applications. ML is a process where computer systems
mimic the human behavior and perform all activities faster than a
human. In general, supervised learning algorithms can help to solve
problems with lesser amount of data and can clearly label them.
Unsupervised learning gives better performance and results for
large datasets. It is an ML technique, where you do not need to
supervise the model. Unsupervised ML helps to find all kinds of
unknown patterns in our datasets. These learning algorithms are used
for unlabeled datasets. One of the main scopes of this algorithm is
Quantum computers, which are good at manipulating high-
dimensional vectors in large tensor product spaces. Development of
both supervised and unsupervised quantum ML algorithms will in­
crease the number of vectors and their dimensions exponentially than
classical algorithms. With the rapid increase of technology and ap­
plications, ML takes a greater role in today’s generation. It eventually
makes human life easier in different aspects, such as disease predic­
tions and monitoring, robots, driverless cars, agricultural field, and so
on. As technology is growing, ML will also take up the world in a
few years.

References
[1] https://www.digitalocean.com/community/tutorials/an-introduction-
to-machine-learning#:~:text=Machine%20learning%20is%20a
%20subfield%20of%20artificial%20intelligence%20(AI).&text=
Because%20of%20this%2C%20machine%20learning,has%20bene­
fitted%20from%20machine%20learning.
[2] https://medium.com/towards-artificial-intelligence/machine-
learning-algorithms-for-beginners-with-ython-code-examples-
ml-19c6afd60daa
[3] https://machinelearningmastery.com/a-tour-of-machine-learning-
algorithms/
 TRENDS AND TRANSITION IN THE ML SPACE 55

[4] https://monkeylearn.com/blog/classification-algorithms/
[5] https://www.machinelearningplus.com/predictive-modeling/how-
naive-bayes-algorithm-works-with-example-and-full-code/
[6] https://livecodestream.dev/post/7-steps-of-machine-learning/
[7] https://towardsdatascience.com/introduction-to-machine-learning-
for-beginners-eed6024fdb08
[8] https://towardsdatascience.com/types-of-machine-learning-
algorithms-you-should-know-953a08248861#:~:text=Supervised
%20learning%20algorithms%20try%20to,from%20the%20previous
%20data%20sets.
[9] https://machinelearningmastery.com/types-of-classification-in-
machine-learning/
[10] https://www.techradar.com/in/news/how-ai-and-machine-learning-
our-improving-the-banking-experience
[11] https://technostacks.com/blog/machine-learning-in-agriculture/
[12] https://www.flatworldsolutions.com/healthcare/articles/top-10-
applications-of-machine-learning-in-healthcare.php
[13] https://machinelearningmastery.com/supervised-and-unsupervised-
machine-learning-algorithms
 3
NEXT-GENERATION IOT USE
CASES ACROSS INDUSTRY
VERTICALS USING MACHINE
LEARNING ALGORITHMS
T. R. KALAIARASAN, SRUTHI ANAND,
V. ANANDKUMAR,
AND A. M. RATHEESHKUMAR
Department of Information
Technology, Sri Krishna College of
Engineering and Technology,
Coimbatore, India

Contents

Introduction 58
Sensors 58
Connectivity 59
Data Processing 59
User Interface 60
IoT in Industry 60
Machine Diagnosis and Prognosis System 61
Remote Monitoring and Production Control 62
Indoor Air Quality Monitoring System 63
Noise Monitoring System 64
Inventory Management System 65
IoT in Health Care 65
Pharmacy Management 67
Drug Monitoring 68
Smart Pills 69
Wearables 70
Fitness Trackers 70

DOI: 10.1201/9781003119838-3 57
58 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Smart Hearing Aids 70

Wearables to Monitor Diseases 71
Use of IoT for Physicians 72
Medical Alert System 72
Data Analytics for IoT 73
IoT Data Analytics Using Machine Learning Algorithms 74
Conclusion 75
References 76

Introduction

Internet of Things (IoT) refers to the interconnection of devices

embedded in our everyday life via the Internet. It also enables
them to send and receive the data. The scope of the IoT is not
limited to getting the devices connected/networked; it is much
more than that as depicted in Figure 3.1. Aznd IoT comprises
things that have unique identities and are connected to the
Internet. IoT does not use the existing technology, rather a
combination of knowledge base and experience is mandatory
(Figures 3.1 and 3.2) [1].
The IoT workflow consists of four fundamental elements as shown
in Figure 3.2:
1. Sensors
2. Connectivity
3. Data Processing
4. User Interface

Sensors

Sensors are nothing new; these are used by the organizations for a
considerable period of time. Sensors play an important role in pro­
viding solutions using IoT. Sensors are devices that can detect the
external environment. It is possible to collect data of almost every
possible condition using sensors. There is a wide range of IoT sensors
used to detect and measure various aspects.
 NEXT-GENERATION IOT USE CASES 59

Figure 3.1 Domain-Specific IoT

Figure 3.2 IoT Workflow

Connectivity

IoT connectivity is defined as connection between devices such as sensors,

routers, gateway, etc. There are a number of ways to connect to an IoT
device, such as Bluetooth, Wi-Fi, Zigbee, NFC, Ethernet, LRWAN, etc.

Data Processing

Generally, sensors emit data that do not have any meaning (raw and
unprocessed data). The information is inferred from the raw data by
filtering and processing the data. Data processing involves input,
process, and output.
Input: The data may be in the form of numbers, text, images,
or videos. All these forms of data can be converted into machine
understandable language.
60 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Process: Different techniques to process data like classification,

sorting, calculations, etc., are used to convert meaningful information
from the data received.
Output: Although the information is inferred in the process
phase, it is rendered into human understandable format in the
output phase. The output in the form of text, graph, table, image,
audio, video, etc., can be stored locally or in the cloud for further
processing.

User Interface

A minimal user intervention is required for IoT where it is ne­

cessary for the user to understand the data captured by the system.
A user interface is the point at which a user and a computer system
can interact [2].

IoT in Industry

The major driving force behind Industry 4.0 is Internet of Things

(IoT) [2]. IoT is used in the majority of industries, like manu­
facturing, logistics, oil and gas, aviation, and transportation. IoT 4.0
promotes and enhances automation, data collection, and analytics
that can be useful for increased process optimization. It is observed
that in 2016 for manufacturing operations in IoT, a total of $102.5
billion was spent. With respect to Industry 4.0, more focus in
transportation is given toward freight monitoring. The evolution of
IoT has enabled rethinking of almost all businesses to be moved
toward the digital world. The devices connected via technology work
independently to produce the output. The rapid growth of IoT has
changed today’s industries from factories to smart factories and smart
production systems and so on. The data are collected through sen­
sors, which contribute much toward identifying system failures for
immediate maintenance. As the system can be integrated, it is used
along with Machine Learning and Deep Learning systems to pre­
dict and recover from failures. These predictions contribute to
better optimizations and gain insights yielding higher growth and
productivity [3] (Figure 3.3).
 NEXT-GENERATION IOT USE CASES 61

Figure 3.3 IoT in Industry

As shown in Figure 3.3, IoT enables the connected operation that

has the ability to collect data from the connected devices, equipment,
and sensors. The data can be processed, analyzed, and stored in the
cloud for real-time process adjustments and optimization.

Machine Diagnosis and Prognosis System

Generally, mechanical equipment requires some manual main­

tenance, due to which there is an inability to diagnose equipment on
time which sometimes leads to accidents. There is a number of
mechanical equipments working from which the data can be col­
lected in a short period of time. Using the collected data to improve
the fault prediction becomes a challenging process [4].
Machine diagnosis refers to finding the cause of a fault in ma­
chine. Machine prognosis refers to predicting the performance of the
machine. The data can be collected and analyzed on the basis of the
performance of the predicted machine (Figure 3.4).

Figure 3.4 Machine Diagnosis and Prediction System

62 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

There are various types of sensors like temperature sensor, pressure

sensor, vibration sensor, acceleration sensor, etc. Data can be col­
lected from sensors on a timescale of a few milliseconds to seconds
that leads to a huge amount of data. Those data can be sent to the
cloud using cloud application and stored. The faults of certain devices
can be detected by monitoring the health index of these devices in
real time.
The data can be analyzed and monitored using various techniques
to predict the performance of a machine by analyzing the data. The
current operating conditions can be compared with normal operating
conditions to find any deviations upon which the performance of a
machine can be predicted.

Remote Monitoring and Production Control

There are many industries in production. Some of the industrial

processes cannot be monitored by humans directly or near the pro­
duction line. The industry may be small or large, some of the re­
stricted areas need to be monitored. The most important feature of
remote monitoring and production control in industry is the cen­
tralized supervision over the machines in the process of production.
Information obtained from the devices like sensors, camera, etc.
(Figure 3.5), provides a much clearer and faster insight into the actual
production field. The staff is able to analyze the data with its assis­
tance. This ensures that IoT is a pivotal technology in ensuring

Figure 3.5 Remote Monitoring and Production Control System in Manufacturing Industry
 NEXT-GENERATION IOT USE CASES 63

production automation, tracking the status of the staff, and in

monitoring the workers [5].

Indoor Air Quality Monitoring System

Quality of an air plays a vital role in industry. The quality of air

determines the performance of workers that also affects opera­
tional cost and influences the productivity and efficiency of an
industry [6].
The following are common pollutants that impact indoor air
quality:
• Volatile organic compounds (VOCs)
• Carbon monoxide (CO)
• Particulate matter (PM)
Volatile organic compounds (VOCs): VOC is the main contributor
to poor air quality. It emits gases from processes or products.
Carbon monoxide (CO): CO is a dangerous gas. It is impossible
to see or smell. Sources of CO are automobile exhaust, generator,
improper use of boilers, etc.
Particulate matter (PM): PM is a mixture of liquid and solid
particles suspended in air, such as dust, pollen, smoke, and soot.
Different types of sensors are used to measure the air quality. Sensors
are capable of detecting harmful toxic gases like carbon monoxide,
nitrogen monoxide, nitrogen dioxide, particulate matter, and many
more. Typically, the sensors can be placed across the industrial build­
ings rather than placing them on the ceiling or floor, wall mounting is
the best suitable place to locate the sensors [7] (Figure 3.6).

Figure 3.6 Indoor Air Quality Management System

64 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Noise Monitoring System

Noise pollution is one of the major reasons for an impact on health.

Noise generated by various industries like plants, mills, refineries,
shipping docks, utilities, and more affects the surrounding environ­
ment, as well as workers, in the industry. By controlling and mon­
itoring this noise, an industry can establish a working relationship
which will benefit all [8].
This noise monitoring system helps in producing the noise map
(Figure 3.7). The noise map is the graphical representation of the
sound level of an environment. Number of noise monitoring sta­
tions is deployed at different locations of industry. Sound-level
sensors used to monitor the data on the noise level from the noise
monitoring station (Figure 3.8) will be collected and stored on
server or cloud. The collected data can be analyzed and aggregated
to generate the noise maps.

Figure 3.7 Noise Monitoring System

Figure 3.8 Noise Monitoring Station

 NEXT-GENERATION IOT USE CASES 65

Inventory Management System

Inventory management is a complex process specifically for large

organizations. In this modern business environment, inventory is very
important, as the uncertainty can happen at any point of time up­
stream or downstream. It is utmost important to keep the right level
of inventory at the right time in the right place with right price and
right time duration [9].
Inventory management using RFID tags will help the industry to
maintain the right inventory level. RFID is based on radio wave
technology to maintain and monitor the inventory. RFID fully differs
from barcode. By using barcode, the products can be scanned and
data can be returned, whereas in RFID, tracking can be done using
RFID readers attached with warehouses or shelves. This will make
them more powerful and useful for operators trying to find the stocks
quickly and easily. IoT systems enable the remote monitoring of
inventory using the data collected by using RFID readers.
Any materials or products can be tracked and fitted with RFID tags.
There are two types of tags 1. Passive tag (Integrated chip and
Antenna) 2. Active tag (Integrated Chip, Antenna and Battery). Each
RFID tag is assigned with a Unique Identification Number (UIN). A
RFID reader will extract UIN numbers and update in the database.
RFID reader sends energy to the tag for power. RFID tag sends
data to the reader. And the reader will decode the data and send the
data to the local server. A local server performs aggregation of data.
An integration server will make the data available in enterprise
resource planning (ERP) [10] (Figure 3.9).

IoT in Health Care

The role of IoT is increasing in the field of life sciences and medical
health which is the need for today’s society. It is explored for a
number of purposes enabling end-to-end integration. The tech­
nology fosters the medical environment to connect patients and
monitor them, which in fact improves systems as such. IoT can be
integrated with any other technology to enhance health care systems
starting from virtual online consultations to smart pills and wearables.
66 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 3.9 Inventory Management Using RFID

These technologies help people to lead a healthy life by keeping track

of their lifestyle and eating habits that can enable the medical
practitioners and patients to be connected through. Having a de­
centralized and dephysicalization kind of treatment is what is ex­
pected today. According to the reports by Statista, it is observed that
there are about 26 billion connected IoT devices and by 2025, it is
expected to grow to 75 billion. With respect to the global health care
market, it is assumed that it will hit $543.3 billion per year. The
current scenario in healthcare is the reactive sick care approach,
where the patient is treated only after identification of the disease.
But with the growth of IoT, it could be a proactive, predictive, or
preventive approach.
 NEXT-GENERATION IOT USE CASES 67

Figure 3.10 IoT in Health Care

The impact of IoT is that it can have a face-to-face commu­

nication to address the issues through the internet. This is extended
up to providing health care advice and treatments even when they are
on the way to the medical centers. It not only benefits the patient, but
also notifies and informs the health officials to be prepared for the
treatment prior to the patient reaching the hospital. Hence, it can be
well understood that IoT plays a very vital role in health care [8].
There are different applications that contribute to IoT in health care,
such as pharmacy management, smart pills, and drug discovery and
management (Figure 3.10).

Pharmacy Management

In the field of drug manufacturing, it is very critical to make sure that

there is no issue with the drug that is manufactured in terms of
equipment and composition used in that drugs. This involves
tracking the preparation of the drug process. There can be pitfalls due
to the failure in the assets of the system, like mechanical damage of
68 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

the equipment, lack of power supply, and lack of environment. To

address these issues, IoT paves a greater way using sensors to monitor
and track the working conditions of the equipment used for the
manufacturing process. The machineries can be monitored for
pressure, heat exchange, pH, compressors, and pumps. The data
received from these sensors help to keep track of the machines,
monitor the safety critical measures, and ensure safety of the work­
place [11].

Drug Monitoring

Drug monitoring and preserving are one of the most important

tasks in the field of medicines. It is well known that the drugs
manufactured must be preserved in certain conditions concentrating
on the environmental conditions, like temperature, humidity, light,
carbon dioxide level, light, etc. Hence, IoT makes it possible to
monitor the parameters and keep the pharmacists updated and alert
them for any deviations. This will help to automatically control the
leakage of any toxic substances and damage to the environment.
Another function that IoT can perform in this field is to place
sensors on the drug packages and monitor if it the package reaches
the patients on time without any other serious issues. As it is
known, there are issues arising due to the illegal transport of
medicines and thus the quality may be compromised. Next is the
issue in the process of supply chain management necessary to
monitor the way in which the drugs are shipped. To ensure that this
process is carried out, route details must be taken care of. This can
include smart labels and tags used for packaging and GPS to
continuously monitor how the vehicle travels. During shipment, it
is also possible to keep track of the temperature and maintain
ambience such that it would contribute to safely record and monitor
the stability of the drug transported [12] (Figure 3.11).
The methods and techniques mentioned contribute to a greater
extent in supply chain management. They also ensure that the patients
or their attenders are informed about the conditions and details about
the drug and how it is to be utilized. These are the advantages that IoT
provides in terms of pharma application. Since these operations do not
 NEXT-GENERATION IOT USE CASES 69

Figure 3.11 Drug Management

provide full transparency, it is difficult to ensure safety and security

at times, which can also lead to discarding of drugs or patient
mistreatment linking to some false electronic health records (EHR).

Smart Pills

Smart pills (Figure 3.12) are just like normal medicines monitored
even after they have been taken. The term smart indicates that the
pill is able to monitor the temperature, pressure, and keep track of
the activities inside the body. It might sound similar to a pill cam
also. This pill is used to treat the problems in the intestine, treat
cancer, identify bacterial infection, etc. Most of the time, this pill
continuously helps us to identify the issues in the intestinal region.
There are a number of situations where a person may suffer from
continuous vomiting or a kind of uneasiness in the stomach region.
In this case, the smart pill enables the doctors to collect data as the
pill travels through the body. This is done in order to identify
the process that takes place when the food is swallowed. It does not
take any kind of photos inside our human system but it monitors the
movement. The smart pill is developed using either web or mobile

Figure 3.12 Smart Pills

70 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

applications that could track the process. Smart pills include camera,
PCBs, antenna, and battery [13].

Wearables

IoT, as the name suggests, is the interconnection of devices, al­

lowing the devices to communicate with each other by means
of sensors. These sensors are connected to the bluetooth or their
mobile applications, which help in analyzing and predicting
the collected results. The sensors are embedded in wearables or
as a smart device, or implanted onto some garments or fashion
accessories or even imprinted as tattoos. Some examples of
wearable technology include wristbands, watches, smart clothing,
etc [14].

Fitness Trackers

Fitness trackers are smart devices for monitoring the activities of

humans that include heart rate, breathing, pulse rate, and various
body parameters. The fitness tracker measures the movements using
a 3-D accelerometer. The data will be tracked every time the fitness
tracker is worn on the body. These data are analyzed and shared
with the device connected and displayed via a monitor. The
movements are categorized in different factors and information is
generated based on the different actions. The person makes use of
the application and taps on it in which huge information is available
and can be tracked [15].

Smart Hearing Aids

Hearing aids are much needed technology of this present era.

These help to filter noise and sounds and manage hearing loss.
These can easily adjust to the environment, thereby able to classify
noise and sound without any disturbances. They are directly able to
communicate to any of our smart device like TV or mobile phones
without the need for any other intermediary devices. These devices
 NEXT-GENERATION IOT USE CASES 71

not only help in rectifying or managing hearing impairment, but

also in managing calls, stream audio, and easy understanding of the
conversation. The main advantage is that these can also improve
the quality of the sound received. These are usually placed at the
back of our ears, which comprise a microphone, a sensor, and an
amplifier to engage in the process. These aids are highly helpful for
people who suffer from hearing loss due to ageing, a disease, or an
injury [16].

Wearables to Monitor Diseases

Nowadays we come across a number of diseases that are not easily

identifiable. Hence, these kinds of technology can help identify
and alert in case of any abnormalities. It is highly used in detecting
Alzheimers, Parkinsons, and Depression. Wearable ECG moni­
tors help to detect pace, distance, elevation, and the different
activities. In case our level increases, it can notify us by sending an
alarm or a message. There are wearable blood pressure monitors
who keep track of blood pressure and our daily activities. These
are some of the common events. Apart from this, depression can
also be controlled using IoT. It aids in identifying the mood
of a person and also pacifies them both in normal and abnormal
conditions. The solution could be the use of a Fitbit or a smart
clothing that reduces their anxiety or disturbances to a greater
extent (Figure 3.13).

Figure 3.13 Wearables

72 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 3.14 Monitoring Patient Using Wearables

Use of IoT for Physicians

Physicians benefit more with this technology. The data from the
patient wearables help them to identify the treatment process, check
if they are adhering to the treatments, and monitor them. It also
promotes patient doctor interaction even in case of emergencies
(Figure 3.14).
IoT sensors are being used in different medical equipments like
nebulizers, wheelchairs, oxygen cylinders, glucose monitoring, and
insulin detection. Wearable devices can transmit data directly to the
practitioners that foster speedy diagnosis. It contributes to a greater
extent in remote monitoring of patients even if the patient may not
be able to communicate directly to the patients. Majority of the
medical devices can be connected via this technology which may
include wearables, labels that can be worn by patients or their
attenders to reduce the admission process associated with it [17].

Medical Alert System

In this modern world, seeking assistance to aged persons is difficult.

Many aged people fear losing their dignity. As a result of this, use of
fall detection devices and other medical alert systems has been on the
rise. Fall detection devices can prevent senior people from falling. In
this device, a camera is placed to track senior people across the room.
In addition to fall detection, medical alert systems are available for
remote monitoring. This medical alert system is a wearable, which can
be worn by an aged person as clothing or jewelry. It feeds into a series
of connected sensors, which measure the health and welfare of the
 NEXT-GENERATION IOT USE CASES 73

Figure 3.15 Medical Alert Device

person. If an aged person remains motionless for too long a period or

falls out of bed, an alert is sent to the family members or friends who
can help immediately [18] (Figure 3.15).

Data Analytics for IoT

IoT device will emit data, a raw data or an unprocessed data. Nothing
can be done if only raw data are available. The information is inferred
from the raw data by filtering and processing the data. From in­
formation, knowledge is extracted using organizing and structuring of
information and decisions can be taken (Figure 3.16).

Figure 3.16 Data – Information – Knowledge

74 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

A variety of devices can be connected to Internet to share data.

These data have no meaning without analysis. IoT data analytics is
the analysis of huge amount of data that can be generated by various
connected devices. IoT data are highly unstructured data, which
make it difficult to analyze with traditional analytics. The data from
the IoT devices will have some false readings and corrupted messages
that must be cleaned up before analysis. IoT data are often mean­
ingful only in the content of additional third-party data.
Example: In agriculture, rainfall data are added to help farmers so
that they can decide when to water their crops along with moisture
sensor data.
The data produced are effectively collected, analyzed, and stored.
IoT data analytics improve decision-making and many more benefits,
like improved equipment maintenance, operations optimization, and
so on. The combination of IoT and data analytics has already been
proven to be beneficial in various domains.

IoT Data Analytics Using Machine Learning Algorithms

With the emerging technologies in the field, IoT combines with cloud
computing and data analytics for predictive maintenance and decision-
making. IoT enables mechanical devices to be diagnosed and main­
tained with large amount of data collected instantly at different in­
tervals of time. The data collected must be utilized effectively so that it
increases the rate of fault prediction. The data received from the
sensors are related to different equipments, and incorporating the
concept of machine learning (ML) will enable efficient data analysis.
ML algorithms tend to make good predictions that help in the process
of easy decision-making across major fields, like health care, banking
and financial sector, industrial automotives, etc.
The process of ML involves collecting huge datasets and im­
plementing a model to train the datasets. The algorithm takes set of
samples as a training set. With the categorization of learning models,
viz supervised, unsupervised, and reinforcement, the appropriate
models can be used for training. Supervised learning involves training
the data and mapping it to the corresponding targets derived from
the past experiences. Unsupervised learning techniques involve
training the dataset without the need to supervise the data and it
 NEXT-GENERATION IOT USE CASES 75

works on its own to draw conclusions. This is done by identifying

patterns in the data and categorizing them, which is used in drawing
conclusions. Having such models provide great insights in the field of
IoT to model the data received from sensors. For instance, in the
application of machine diagnosis, it is highly important to monitor
the machine and also foresee if the machine needs some maintenance
to meet the demands. Integrating IoT and ML generally has the
following steps:
1. IoT Sensors are embedded to the devices.
2. Data from the sensors are loaded to the cloud.
3. Using an appropriate ML model, data are processed in the
cloud using training and testing data separately.
4. The model evaluates on huge datasets to arrive at conclusions or
come up with hypothesis.
5. This result is tested and validated.
6. This model is then evaluated on testing data and conclusions
are drawn on the basis of the inferences that it has got from the
trained data.
With these processes,companies benefit by applying them on dif­
ferent applications that help in giving detailed root cause analysis,
fault predictions, maintenance routines, etc.

Conclusion

Internet of Things (IoT) is a technology in which every device is

connected to some real-world objects. The basic idea behind this
technology is the use of sensors to monitor and analyze data to make
life more easy and comfortable. These can be used in almost every
application from industry, health care, smart home, smart city, etc.
This chapter focuses on the domains where IoT is prevalent in today’s
market. There are also certain issues with respect to using this with the
expectation to be fully connected to a network every time and since
every object is connected via the internet, it may be prone to security
attacks. With this technology, there is no doubt that in the coming
years, every organization may drastically move toward it. Hence, it
requires the data to be available always and need to be stored for future
retrieval and processing. It is the next step of an information age where
76 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

it reduces human involvement and transforms it to machine controlled

systems. The other important fact is that this technology will grow
only if the computing power increases. With many numbers of devices
connected, it is also essential to also care of the processing power and
the network speed in which it can communicate.

References
[1] ArshdeepBahga and Vijay Madisetti, “Internet of Things – A Hands -
On Approach”, Universities Press, 2015, ISBN: 978-81-7371-954-7.
[2] Anandkumar, V., T. R. Kalaiarasan, and S. Balakrishnan. “IoT Based
Soil Analysis and Irrigation System.” International Journal of Pure and
Applied Mathematics 119, no. 12 (2018): 1127–1134.
[3] https://www.i-scoop.eu/internet-of-things-guide/industrial-internet-
things-iiot-saving-costs-innovation/
[4] https://www.sciencedirect.com/science/article/pii/S1569190X19301145
[5] https://www.scnsoft.com/blog/iot-in-manufacturing
[6] Sruthi Anand, S. Lavanya, and T. Eswari, “A Smart Monitoring
System to Sense Short Circuit and Gas Leakage Using IOT”,
International Journal of Advance Research in Computer Science and
Management Studies, vol. 5, no. 7, ( July 2017), 1–6.
[7] https://www.iotone.com/usecase/indoor-air-quality-monitoring-iaq/u19
[8] http://www.larsondavis.com/applications/environmentalnoisemonitoring/
industrialnoise
[9] N. Susila Dr., Sruthi Anand, Dr. Granty Regina Elwin J. Dr.
Tsujatha, “Technology Enabled Smart Waste Collection and
Management System Using IoT”, International Journal of Pure and
Applied Mathematics, vol. 119, no. 12 (2018), 1283–1295.
[10] https://www.peerbits.com/blog/warehouse-smart-inventory-
management-solution.html
[11] https://www.ns-healthcare.com/analysis/iot-in-healthcare/
[12] https://www.pharma-iq.com/manufacturing/articles/the-role-of-iot-
in-pharma-manufacturing-and-distribution
[13] https://www.iottechtrends.com/how-smart-pills-change-healthcare/
[14] https://www.businessinsider.in/science/latest-trends-in-medical-
monitoring-devices-and-wearable-health-technology/articleshow/702
95772.cms
[15] http://iotlineup.com/device/sync_smartband
[16] https://www.sdglobaltech.com/blog/10-brilliant-examples-of-wearables-
in-healthcare
[17] https://www.finoit.com/blog/the-role-of-iot-in-healthcare-space/
[18] https://www.wirelesswatchdogs.com/blog/iot-applications-in-healthcare
 4
A PANORAMIC VIEW OF CYBER
ATTACK DETECTION AND
PREVENTION USING MACHINE
LEARNING AND DEEP LEARNING
APPROACHES
ESTHER DANIEL1, N. SUSILA2,
AND S. DURGA1
1
Assistant Professor, Department of
Computer Science and Engineering,
Karunya Institute of Technology and
Sciences, Coimbatore, India
2
Professor, Department of
Information Technology, Sri Krishna
College of Engineering and
Technology, Coimbatore, India

Contents

Introduction 78
State of the Art Techniques 80
Android-Based Malware Detection 80
Ransomware Malware Detection 83
Windows-Based Malware Detection 88
Linux-Based Malware Detection 91
Malware Detection Tools Used 92
Conclusion 94
References 95

DOI: 10.1201/9781003119838-4 77
78 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Introduction

The malware is the code that performs malicious actions, such as

causing computer operations disruptions, stealing sensitive in-
formation, accessing unauthorizedly, spying, flooding with spam,
causing distributed denial of service attacks, and holding files for
ransom. A concise glance at the historical backdrop of malevolent
programming advises us that the presence of malware dangers has
been with us since the beginning of computing. The most punctual
archived infection showed up during the 1970s. It was known as the
Creeper Worm and was a trial self-recreating program that re-
plicated itself far off frameworks and showed the message: “I’m the
creeper, catch me if you can.” Afterward, in the mid-1980s, Elk
Cloner showed up, a boot-division infection that focused on Apply
II PCs. From these basic beginnings, a gigantic industry was con-
ceived and, from that point forward, the battle against malware has
never halted [1].
On looking into the statistics of the total number of malware
since 2011 to date, we can see a drastic rise in the number of
malware created. Due to the enormous rise in malware, the need for
malware discovery has extended radically with the end goal that
5,638,828 distinct hosts were assaulted in 2018, according to the
report given by Kaspersky labs. In a further article by Juniper
Research, it was realized that more than 33 billion records will be
taken by digital criminals in 2023 alone. According to Symantec,
mobile malware variants increased by 54% in 2017, while up to
600% in IoT [2].
The Trojan malware was the most common type of malware.
Different types of malware [3,4] can affect our devices and hinder
privacy and performance.
1. Trojans: This malware camouflage itself as a legitimate app and
gets installed on a device. This app can steal sensitive in-
formation, spy on activity, gain access to the device, and
download other malware.
2. Keyloggers: This malware is designed to record or capture
keystrokes when typed in an android device. These apps are
readily available on the internet. Most of the apps masquerade
 A VIEW OF CYBER ATTACK DETECTION 79

as parental control apps and some developers promote these

apps as means to surveillance on other phones.
3. Ransomware: Ransomware is mostly for computer-based at-
tacks, still, Ransomware affects android devices also. Ransomware
encrypts all the data in a device and demands payment to decrypt
the files.
4. Spyware: Spyware can be installed as a root kit or as an app,
which monitors all the data including messages, phone calls,
photos, videos, and location, and it can even use the camera for
live streaming.
5. Riskware: It is an app that is not classified as malware, but as a
potentially unwanted app. Riskware utilizes system resources in
an undesirable manner and also poses security risks.
6. Adware: Adware steals your cookies information and gives
advertisements based on it. Adware shows ads even when the
device is not connected to the internet or when not using any
apps. Adwares are considered as potentially unwanted apps,
which get downloaded from third-party sites.
7. Viruses: Viruses wreck or even alter your touchy information
present in the framework, which bombs the CIA triad.
8. Root kit: The assailant can do whatever he needs if he has a
root kit infused in your framework. It stays disguised and is
exceptionally hard to distinguish.
9. Backdoors: If backdoors are introduced, the interloper will
never require any sort of confirmations to get into your fra-
mework. These can infuse infection, worms, and Trojan ponies
when indirect access is introduced.

Given below are the percentages of various malware attacks on

Windows systems.
Trojans are the most used malware to attack a windows system.
When compared with other malware, Trojan attack percentage is
around 64%, which is a big score. Virus is the second most used
malware attack taking 15% when compared with all other mal-
wares. Worms come in third place with 7% of attacks. These are
the top three attacks that we have to be worried about.
Ransomware is a deadly malware attack but when comes to the
80 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

attack percentage, it has 0.93%, which is very less as compared

with the top three.
The traditional method of identifying malware depends on
signature-based malware detection. But nowadays, the malware
attacks that occur are mostly zero-day malware, which increases the
difficulty in detecting and classifying malware. Zero-day malware
is a specific kind of malware or malicious software that has only
recently been discovered where we wouldn’t be able to classify or
detect these malware [5]. AI comes into the picture and gives hope
such that we are still able to detect and classify these types of
malware. Various machine learning and deep learning algorithms
are used to detect these types of malware and hopefully give a
better result.

State of the Art Techniques

Android-Based Malware Detection

The android malware gained popularity and cybercrime kept in-

creasing in the android environment. Most of the malwares began to
be sold on the dark web, which is spread in the android community
affecting thousands of devices. The first Android OS was launched
back in 2008, ever since android OS has been a popular choice for
smartphones. Now, 2.5 billion active Android users are present
around the world. Google made the Android OS an open source for
the smartphone manufacturers, thus the OS was made more custo-
mizable as per the manufacturers’ need. Popular android manu-
facturers, such as Samsung, One Plus, Xiaomi, Oppo, Huawei, etc.,
made their skin for android OS and applied it over the OS archi-
tecture. Stock Android (i.e., The original OS) has been used by
Google Pixels, Nokia, and some manufacturers. The use of skin over
the stock OS ensured better performance specifications for the
smartphone model; however, some manufacturers violated the
privacy concerns of the users by collecting user statics data without
proper concern from the users.
The first malware for android was reported in 2010,
‘DroidSMS’ a Trojan, an SMS fraud app. The app made auto-
matic subscriptions on the affected phones. The process of
 A VIEW OF CYBER ATTACK DETECTION 81

subscription was made in the background without the users

knowing about it. The same year, another Trojan ‘TapSnake’ was
released. The app leaked the GPS location of devices over HTTP,
making the location of the devices known by other devices through
that GPS Spy app.
Classification of detection techniques based on the deep learning
methods implemented for android malware detection over the years
are highlighted briefly.
Mohammed K. Alzaylaee et al. [6] in their paper, “DL-Droid:
Deep learning-based android malware detection using real de-
vices,” propose a deep learning system that uses dynamic analysis
along with stateful input generation for detecting malicious
Android applications. The experiment used 30,000 applications
composed of benign and malware, which were presented on real
devices. The study analyzed the code coverage and also compared
the detection performance of the models used. The study pro-
duced a 97.8% detection rate for dynamic features and 99.6% for
dynamic and static features, and the study achieved an accuracy
of 98.5%.
Deep android malware detection by McLaughlin et al. [7] im-
plements a convolution neural network for malware detection on
android devices. The method was implemented by performing static
analysis on the raw opcode sequence of the malware. The opcodes
were obtained from the dissembler program and the malware in-
dicative features were learned accordingly from the opcodes. The
operation for training the network takes is easier to approach than
the other n-gram methods used before, as this implementation re-
moves explicit enumeration of millions of n-grams for training. The
network structure allows the usage of long n-gram features, which is
not feasible with other methods.
S. Hou et al., in their work “Deep4MalDroid,” use a method
known as Component Traversal, which automatically executes the
app code routine for android completely. A weighted graph has been
derived from the extracted Linux kernel system calls and a deep
learning framework has been used to identify a new malware app.
This method outperforms signature-based detection techniques
which were used previously as anti-malware app [8].
82 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

DroidDeepLearner [9] uses deep learning to characterize and

identify malware based on behavioral patterns. The proposed method
emphasize on autonomous learning and solve issues without human
intervention. The method uses both dangerous application pro-
gramming interface (API) and risky permission to build the Deep-
Belief Network (DBN) model. The experimental results have shown
better accuracy than Support Vector Machine (SVM)-based
solutions.
Droid-sec by Zhenlong Yuan et al. emphasizes on ML-based
detection techniques utilizing 200 extracted features from static and
dynamic analysis for android malware detection. From the study, it is
observed that the deep learning techniques outperform machine
learning techniques. The study produced a deep learning method for
malware detection with 96% accuracy [10].
E. B. Karbab et al. propose MalDozer, [11] a convolution
neural network (CNN)-based malware detection system for an-
droid. The proposed method uses less preprocessing for the as-
sembly process and thus allows for quick preprocessing. MalDozer
automatically extracts features starting from the API calls and
learns malicious and benign patterns. The study used 33,000
malware apps and 38,000 benign apps for evaluation of the model
and the results showed an F1-score of 96% with a false positive
of 0.06%–2%.
Deep Flow analyzes the data flow for malware detection. The
study extracted 323 features from android apps which included
malicious and benign apps [12]. The study applied the SUSI tech-
nique to categorize data flow features combined with DBN deep
learning model. The study achieved a higher F1-score and out-
performed traditional machine learning models.
Hou et al., in their paper, propose DroidDelver for deep learning-
based android malware detection. In this paper, features are re-
presented as statically extracted API calls from smali codes, which are
converted into code blocks. The proposed deep learning technique
uses DBN to analyze generated code blocks from API calls. The
study proposed an accuracy of 96.66% and out-performed other
machine learning and deep learning models [13]. DroidDelver to
detect the malware contains five major components: (1) Unzipper
and decompile are used to unzip the APK files and decompile the dex
 A VIEW OF CYBER ATTACK DETECTION 83

files to smali codes; (2) API call extractors are used to extract the API
calls from the smali codes; (3) API call block generators are used to
extract the API calls and the smali codes are used to generate the
blocks; (4) Deep Belief Network classifier is used as malware de-
tection for API call blocks; and (5) Malware detector performs the
above four components process.
The three levels of penetration technique for installation, activa-
tion, and running of malware on the android system are as follows:
(1) repackaging is the type used to install the malicious thing on the
android platform [14]. (2) Update is difficult to detect but downloads
the malicious activity at runtime. (3) Downloading is a traditional
attack technique that is needed by the users who download apps very
keenly. The application features, API calls, and permissions are used
in the framework to detect. The first step is an app analyzer that
decompresses the apk files of an app and extracts the class files and
AndroidManifest.xml. The second step is characterizing the app
based on its permissions and API calls. The third step is used for
feature extraction based on the Android Manifest and API calls from
the class files. By using a support vector machine, decision trees, and
bagging predictor’s classifiers the Apps can easily be classified as
benign or malware.
There exist different types of attacks on the android system [15].
The first step uses the static and dynamic analysis with code to insert
the malware application in apps. The second step uses the mod-
ification of the system to access the changes occurring on the system.
The third step involves virtualization to separate the domains of an
application to install the multiple instances of the android system on
the same device.

Ransomware Malware Detection

Ransomware is a type of malware that the attacker uses to extort

money from the victim, by encrypting the files in the victim’s system
and holding them “hostage,” until the victim provides the attacker
with the demanded money. The attacker essentially blocks the vic-
tim’s access to the files or will threaten the victim with the exposure
of the files.
84 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Ransomware can enter a system or network in several ways; the

most common method is through phishing emails or links. Once a
download is done from these spam emails or phishing links, the
downloaded file then executes the Ransomware throughout the
system. This malware is capable of spreading through a network in
which the infected system is connected, thus effectively propagating
through all the systems in the network and infecting them all. More
sophisticated Ransomware is capable of spreading and working by
itself, without human activity. Such Ransomware attacks are popu-
larly known as drive-by attacks. In the past few years, the number of
Ransomware attacks has reached an alarming rate.
According to Purplesec, a cyber security services company based in
Washington, the number of new strains of Ransomware increased by
46% and the number of attacks by 41% in 2019 [16]. Mostly, bud-
ding or small businesses fall victim to this kind of attack, as they
usually can’t afford to invest in security. And the rise in Ransomware
attacks can be owed to the fact that Ransomware services are pro-
vided on the dark web and other unsavory parts of the internet for a
minimal price, and it requires almost no amount of technical
knowledge to deploy. Therefore, it has become of the utmost im-
portance to protect our systems from these kinds of malicious attacks
by identifying and detecting the attacks. Here, we look at the various
deep learning algorithms employed in detecting the Ransomware and
conduct a thorough study to determine the pros and cons of each
algorithm concerning the detection and identification of these
attacks.
Smith Maniath et al. [17] has suggested the use of long-short term
memory for identifying Ransomware by analyzing the behaviors
of API calls. They suggest that analyzing the API call sequence of
a process can explain the behavioral patterns of that process.
Ransomware families have some common traits, such as short-term
connection to Command and Control Center, deletion of Shadow
Volumes, and employement of a huge amount of file system opera-
tions. By analyzing the API calls dynamically with the use of LSTM,
a deep learning method, we can classify a binary as Ransomware.
They have created a model of three layers that has 64 LSTM nodes
in each layer and gained the highest training accuracy of 99.9%.
While testing, they have achieved the highest accuracy of 96.67%,
 A VIEW OF CYBER ATTACK DETECTION 85

when compared to the previous work of implementing deep learning

in malware analysis.
A Deep Neural Network (DNN) model that detects Ran-
somware tweets, is proposed by R. Vinayakumar et al. [18]. DNN
consists of basically three types of layers, an input layer, one or more
hidden layers, and an output layer. They have analyzed almost 25
Ransomware families from 2010 to 2017, used this model to comb
through the social media sites for malicious tweets, and classified
them to the corresponding identified Ransomware family. This
model can automatically and continuously analyze the online posts,
thus giving enabling incident management with better resources
and procedures to mitigate the threat. They have suggested a
Ransomware triage method that consists of three steps, namely,
data collection and preprocessing, optimal features extraction, and
classification algorithm.
Text Encoding is aepplied at the data collection and preprocessing
stage, Bag of Words and Keras Embedding are implemented in the
second stage, and DNN in the third stage. The proposed model was
able to give a higher accuracy rate of 78.9% than the predecessors and
has an average error rate of 27.8%.
Seong II Bae et al. [19], have proposed a model that identifies
Ransomware, malware, and benign files using machine learning al-
gorithms. The API invocation method although popular is not sui-
table for three-class detection; therefore, a new model has been
suggested. Input vectors are generated from the n-gram sequences
derived from the processed API sequence. The input vectors are
represented as 1 if n-gram appears in the n-gram sequence and 0 if it
does not appear. The weights of each element are decided using CF-
NCF (Class Frequency Non-Class Frequency). Mainly a combina-
tion of six machine learning algorithms is used to create the testing
model i.e., Random Forest (RF), Logistic Regression (LR), Naïve
Bayes (NB), SGD (Stochastic Gradient Descent), KNN (K-Nearest
Neighbor), and Support Vector Machine (SVM). The proposed
method has proved to have an accuracy rate of 98.65%, which is
relatively higher than the previous results. And experimental results
also have proven that CF-NCF has much better performance than
TF-IDF for the detection of Ransomware.
86 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

An extensive review of the existing Ransomware detection tech-

niques and has been provided and the Ransomware lifecycle and its
properties have been mentioned [20]. In addition to it, the pros and
cons of each detection technique along with a proposed detection
model are also provided. As a result of their extensive review, they
have concluded that regression-type algorithms were the most pre-
ferred for Ransomware detection method, and the most important
evaluation metric is considered to be the accuracy rate and true po-
sitive rate. In the proposed technique, they have suggested the im-
plementation of a hybrid technique, which is a combination of
regression and rule-based algorithms.
A regression algorithm can build a prediction model based on the
predictor-outcome relationship. And a rule-based model produces a
rule for its prediction model. They have stated that the correct
combination of algorithms can complement one another, thus
leading the hybrid algorithms to outperform individual ones.
Jinsoo Hwang et al. [21], have proposed a two-layer Ransomware
detection model that incorporates Markov Model in the first stage
and Random Forest Model in the second stage. Initially, they analyze
the API sequences at runtime and filter out 303 unique APIs that
represent the general API list. They have then implemented Markov
Model to these API sequences to filter out the Ransomware calls.
This gave a moderate error rate of 0.9744 ± 0.0159, and a fairly good
false-positive ratio but a poor false-negative ratio around 20%. To
eliminate this disadvantage, the Random Forest Algorithm is applied
in the second stage. The aim was to find a threshold value that
provides a false-positive ratio below 5% and a false-negative ratio as
low as possible, which concludes 0.2 as the threshold value. For the
given threshold value, the two-stage model gives an accuracy of
97.28%, 4.83% FPR, and 1.47% FNR.
Netconverse [22] is a machine learning evaluation study for dy-
namic detection of Ransomware in Windows network. This model
has three phases, data collection, feature extraction, and im-
plementation of machine learning classifier. A total of six machine
learning classifiers are implemented, namely, Bayes Network,
Multilayer Perceptron, J48, K Nearest Neighbors, Random Forest,
and Logistic Model Tree. Two classifiers, namely, KNN and LMT
were found to have an increase in model creation but the rest of the
 A VIEW OF CYBER ATTACK DETECTION 87

algorithms took decreased time rate to process owing to the decrease

in attribute numbers. The J48 algorithm proved to be the best al-
gorithm among the six, with the highest detection rate of 97.1%,
followed by the LMT classifier that has a detection rate of 96.8%.
Ibrar Yaqoob et al., in this paper, have thoroughly explored the
various aspects of a Ransomware attack, such as the various types of
common Ransomware, the rise, the challenges of Ransomware at-
tacks, security and the aspects in IoT [23]. A brief yet exhaustive
account of the latest and most innovative security development
against Ransomware in the field of IoT is given. Several case studies
are also included to alert the readers on the various security aspects
and threats of IoT devices and the different ways in which
Ransomware can attack and exploit these vulnerabilities. They have
also included several countermeasures and security details to secure
the IoT devices and several research challenges and lack of proper
Ransomware identification and detection systems are also discussed.
Rakshit Agrawal etal, have proposed a model in this paper that
involves the comparison between the LSTM and ARI-LSTM in
detecting and identifying Ransomware threads [24]. Attended
Recent Inputs (ARI) is an enhanced neural cell to involve learning
from Ransomware sequences. This cell processes the sequences of
input and at the same time analyzes the history of sequences also.
Here the implementation of ARI with LSTM has been developed, in
which the LSTM is incorporated with the ARI mechanism and the
resulting neural network sequence is used for sequence learning in
Ransomware. In the experimental results, it has been proven that
ARI-LSTM provides a much better accuracy ratio than LSTM.
ARI-LSTM has an accuracy rate of 0.97 when compared to the 0.87
of LSTM.
Hanqi Zhang etal., in this paper, incorporates the use of N-gram
sequencing, Term Frequency–Inverse Document Frequency (TF-
IDF), and a comparison of five classification vectors to find the best
one for Ransomware detection and identification [25]. The TF-IDF
is calculated for each N-gram sequence, which is transformed from
opcode sequences of the Ransomware samples. The TF-IDF vectors
are calculated to select the featured N-gram so that they can exhibit
better discrimination among the Ransomware families.
88 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

These vectors are then fed to the classification Machine Learning

Classification Algorithms, namely, Decision Tree, Random Forest,
K-Nearest Neighbor, Naïve Bayes, and Gradient Boosting Decision
Tree. Among these algorithms, Random Forest has been proved to
have the highest accuracy rate of 99.3%, while discriminating be-
tween Ransomware and trusted software and a recall of 99.8%. This
method has also gained a 91.43% accuracy rate in detecting wannacry
malware.

Windows-Based Malware Detection

The most recognized malware discovery techniques can be arranged

into three principal strategies:
1 . Static Analysis: It extract features from the file’s structure
without running the file.
2 . Dynamic Analysis: The file is made to run and the features are
extracted while the file is running.
3 . Hybrid Analysis: This is essentially the mixture of both dynamic
and static analyses.
In the further segments, we will be glancing through different ML
and DL algorithms that had the option to identify and classify
malware with the features acquired from the investigation stage.
Ihab Shahadat et al. [26], created various Machine Learning
models for prediction, which includes both binary classification and
multi-classification; for data splitting, cross-validation is used for the
feature selection which minimizes the features. The feature is ex-
tracted using a heuristic strategy which is nothing but dynamic
analysis. They have attained an accuracy of 98.2% in the Decision
tree for binary classification and 95.8% in Random forest binary
classification.
These authors [27] gathered the dataset obtained from both static
and dynamic analysis required and stored it in Apache Spark. The
data stored are pre-processed and the necessary features were ex-
tracted and sent to five diverse base classifiers. At that point the
outcome is utilized to plan two distinctive outfit learning plans, for
example, Weighted Voting and Stacking, which give a decent result.
Big Data is utilized to deal with the malware dataset. We have
 A VIEW OF CYBER ATTACK DETECTION 89

utilized a fair dataset in this model. This model gives an exactness of

99.5% utilizing WeightedVoting_RACA.
This model build [28] firstly converts the binary malware file into
a Gray-scale picture and utilized Visual-AT ML visualization to
classify malware. It deals with adversarial examples and gives a su-
perior precision rate. Adversarial examples are taken care of in this
model. Visual-At gives an exactness of about 97.73%.
These researchers have fabricated a multi-modular profound
learning system known as HYDRA that comprises four fundamental
segments: API-Based parts, Mnemonics-based segments, Byte-
based segments, and the element combination and characterization
component. The impediments of end-end learning have been limited
utilizing this method. The double substance can likewise be included
with the current model which may give a superior result. About
99.75% is the precision identified while actualizing Multi-modal
Deep Neural Network [29].
MalFCS [30] primarily has three stages. The double malware
documents are changed over into entropy graphs, and the highlights
are removed from the entropy graphs which at that point went into the
SVM classifier for malware characterization. It utilizes entropy graphs
instead of a grayscale picture. It can utilize GAN to create great
manufactured entropy graphs and lift model execution. It can acquire
astounding characterization execution with the exactness of 0.997 and
1 individually on the Mailing dataset and Microsoft dataset.
The model developers fabricated a model known as IMCFN
which can be predominantly isolated into two sections as malware
picture age where the double malware records are changed over into
picture and CNN calibrating utilizing back-proliferation procedure
which predicts whether the created picture is malware. When con-
trasted and CNN, IMCFN gives a superior result. The trial in-
vestigation demonstrated IMCFN gives an exactness of 97.56% [31].
A feature-hybrid malware variants detection using CNN-based
opcode embedding and BPNN based API embedding model com-
prises of two principal sub-divisions [32], in the first division the
Opcode is created from.exe record, and afterward utilizing an n-gram
model an opcode n-gram is produced. At that point, PCA is utilized
to lessen measurement and afterward went into a concealed layer and
Softmax classifier is utilized to characterize. In the subsequent
90 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

division, The API calls are produced from the.exe document and a
vector recurrence is utilized to change API calls into vectors and the
PCA is executed to choose the most required highlights and is sent
into the neural network for preparation. At that point, a classifier is
utilized to group malware. This is a component half and half malware
variation model. It can include more static elements. This metho-
dology achieves an exact pace of 95%.
Lu et al.’s model [33] utilizes the malware dataset and is imported
in the cuckoo sandbox. They arranged an API call sequence utilizing
a bidirectional LSTM model. Programming interface call affiliation
and recurrence are examined utilizing the AMHARA algorithm. At
that point, RF and LSTM are joined and anticipated, which give a
superior outcome. ML and DL are both consolidated in this model.
Only a little dataset is utilized to prepare the model. The exactness
gave in this joined model is 96.7%.
Suyeon Yoo et al., proposed a half and half model under the mix of
Random Forest and Deep Learning (MLP) utilizing 12 hidden
layers to distinguish whether the document is malware or kind [34].
The discovery depends on casting a ballot on the output gave by these
two models. It has high TP and low FP, which are significant in
Malware Detection. The accuracy in this model is just 85.1% which
can be improved. It provides a TP pace of 85.7% and FP pace of
16.1% with a presentation of 60.9s per record.
A MalDy structure [35] is built which initially collects the behavioral
report from sand-boxing and creation runtime and convert those report
into a BoW utilizing NLP. At that point, they utilize different AI
models to classify. Right off the bat, there is a dangerous location model
and n number of models for n no. of malware family order. This
structure is versatile in any environment. The model is reliant on the
execution condition report system. MalDy accomplishes over 94.1% of
F-score in android and accomplishes a precision pace of 94% on Win32.
Built-up a lot of regression models to separate the connections
between the highlights of cutting edge malware known as Stuxnet
and afterward ML calculations are utilized to foresee malware de-
pendent on those features. This model can anticipate progressed
malware as same as Stuxnet. Just features related to Stuxnet are
utilized for expectation. At the point when utilized RF the outcome
has R2 = 0.8203 for Stuxnet malware [36].
 A VIEW OF CYBER ATTACK DETECTION 91

Linux-Based Malware Detection

Linux-based Malware or malicious software is used to harm the host

operating system or to steal the sensitive data of the user, organi-
zation, and companies. It helps to gather the information illegally.
Malware detection is used to detect the presence of malware on a
host operating system or used to detect any malicious activity in the
system. Malware detection is mandatory for the system to not lose
the sensitive information of any user. Malware detection is performed
using different techniques but in this survey, we are using machine
learning algorithms to detect the malware on Linux operating sys-
tems. Nowadays malware is a great problem for personal data. Many
approaches are used to identify the malicious executable linkable files
(ELF). Linux operating system is open-source software and released
under General Public License(GNL).
The three-fold approach for the proposed method is: (1). logging
of system call traces; (2). methods for the elimination of redundant
features; and (3). Usage of prediction models [37]. This method is
proposed to give good computational efficiency without losing ac-
curacy. Kruskal–Wallis test is a rank test based on the differences of
independent medians. It is used to describe whether the two classes
have an equal median or not. If the two classes have an equal median,
then it is not in classification. Deviation From Poisson (DFP) is used
to select the robust vector for classification. The parameters of the
system are accuracy and feature length.
The malware sample dataset is taken from VX-Heavens and benign
samples are taken from directories of Linux which are in Executable
linkable files format [38]. System call logging is executed using the
Linux platform “strace.” By using arguments and return values ex-
traction of system calls is done after this preprocessing of system calls is
done for features. After the system call extraction, the data set is di-
vided into training level and testing level. The training step is used for
classification models and the testing test is used for prediction models.
After this step feature analysis and selection, the level is taken by two
steps: preparing different feature categories and feature selection.
Feature categories are union, intersection, discriminating features for
malware, and discriminating features for benign. Feature selection is a
class discrimination measure (CDM) that is used to improve the
92 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

classification for multiple data sets, the main concept of this is the odd
ratio and elimination of sparse features (ESF).
Machine learning systems are based on one-sided perceptron and
a kernelized one-sided perceptron for the datasets to clean the data
and the malware files [39] Gavrilut et al. framework takes three
datasets: training dataset, testing dataset, and scale up dataset.
Cross-validation is done to choose the correct values from the
parameters. The parameters of this paper are feature selection
measures and cross-validation with the cascade one-sided perceptron
algorithm.
To produce security and privacy for mobile users, the parallel
machine learning technique is used because it uses multiple algo-
rithms in a single unit. Supervised algorithms like Decision Trees,
Simple Logistics, Naive Bayes, and Part, Ridor are used to detect the
malicious on the mobile platform [40]. API calls related, command
related, and permissions are used to extract the datasets.

Malware Detection Tools Used

Malware such as are viruses, worms, Trojan horses, spyware, adware,

and Ransomware were discussed. Malware is a threat to the system
and spreads due to vulnerable or carelessness of users. The various
tools used for malware detection are listed below [41–43].

TOOL USED DESCRIPTION

PEiD It is an instrument which distinguishes the greater part of the normal
packers, cryptors, and compilers for a PE (Portable Executable)
records.
HashCalc HashCalc calculates many cryptographic hashes for your files.
Dependency Walker Dependency Walker is used to list the imported and exported functions of
a portable executable file. It also displays a recursive tree of all the
dependencies of the executable file
CFF Explorer CFF Explorer was designed to make PE editing as easy as possible, but
without losing sight on the portable executable’s internal structure
Resource Hacker Resource Hacker is used to add, modify, or replace most resources within
Windows binaries including strings, images, dialogues, menus,
Version-info, and Manifest resources
DIE Detect It Easy, or abbreviated “DIE,” is a program for determining types
of files.
(Continued)
 A VIEW OF CYBER ATTACK DETECTION 93

TOOL USED DESCRIPTION

rEMnux rEMnux is a lightweight Linux distribution for assisting malware analysts
in reverse-engineering malicious software.
JavaSnoop JavaSnoop attempts to attach to an existing process (like a debugger)
and instantly begin tampering with method calls, run custom code, or
just watch what’s happening on the system.
Valgrind Valgrind is a programming tool for memory debugging, memory leak
detection, and profiling.
Zeek Zeek sits on a “sensor,” hardware, software, virtual, or cloud platform
that quietly and unobtrusively observes network traffic.
diStorm diStorm is very robust and mature and used widely all over the world. It is
also the fastest disassembler in the world and is still highly
maintained and updated by its creator.
PEview PEview is a viewer for PE files. PEview is a lightweight program, being a
small standalone executable around 70 kb in size. For determining
basic PE information
Bintext A little, quick, and amazing book extractor that will be quite compelling to
developers. It can separate content from any sort of record and
incorporates the capacity to discover plain ASCII text, Unicode text, and
Resource strings, giving helpful data to everything in the discretionary
“progressed” see mode. Its extensive separating forestalls undesirable
content being recorded
MD5deep md5deep is a set of programs to process MD5,SHA-1, SHA-256, Tiger, or
Whirlpool message digests on a discretionary number of documents.
IDA pro IDA Pro is primarily a multi-platform, multi-processor disassembler that
translates machine-executable code into assembly language source
code to debug and reverse engineering. It can be used as a local or as
a remote debugger on various platforms.
FileAlyzer This is also an application that is used to access the PE file header but it
is also equipped with a VirusTotTab to submit to the Virus total for
analysis and functionality to unpack UPX and PECompact packed files.
Procman Procmon, or Process Monitor, is a free tool developed by Windows
SysInternals and is used to monitor the Windows file system, registry,
and process activity in real time.
Process Explorer Process Explorer is used to monitor the running processes and shows you
which handles and DLLs are running and loaded for each process.
RegShot Regshot is a great open-source utility to monitor your registry for changes
by taking a snapshot that can be compared to the current state of your
registry.
ApateDNS ApateDNS is a tool for controlling DNS responses and acts as a DNS
server on your local system. ApateDNS will spoof DNS responses to DNS
requests generated by the malware to a specified IP address on UDP
port 53.
94 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

For continuous live data feeds, the deep learning architectures

such as generative adversarial network, shallow deep learning feature
extraction, gradient boosting algorithm, Recurrent Neural Network
(RNN), and DNN are used. The Convolution Neural Network
(CNN)-based architectures are used for image classifications and
analysis.
The deep learning architectures mimics the human brain thinking
process and develops the design patterns for decision-making that
improve the accuracy of malware detections and measures to prevent
it. The various deep learning algorithms and architectures suitable for
cyber applications were elaborated.

Conclusion

This chapter discussed the malware types found in real time speci-
fically for android and Windows operating systems. The base for
malware detection starts with static and dynamic analysis. Other than
this, signature-based malware detection gives a better result. But
nowadays, zero-day malware is drastically increasing which paves the
way to machine learning and deep learning technologies. This work
also illustrated the demand for real-time dynamic identification of
malware instead of signature-based detection techniques. Various
deep learning and machine learning techniques used for the detection
of malware in android and other operating systems such as Windows
and Linux were analyzed.
The significance and methods adopted in each technique were
elaborated in-depth. Ransomware malware was studied in detail.
From the state-of-art methods, it can be concluded that the com-
bination of the algorithms, such as Random Forest, Logistic
Regression, Naïve Bayes, Scholastic Gradient Descent, K-Nearest
Neighbor, and Support Vector Machine, has yielded the highest
accuracy rate (98.65%) in identifying and detecting Ransomware.
The need for malware detection and classification has increased
now in this technology era filled with the Internet. We have dis-
cussed the various tools utilized by machine learning and deep
learning techniques for the detection and classification of various
malware.
 A VIEW OF CYBER ATTACK DETECTION 95

References
[1] Gilbert, Daniel Mateu, Carles, Planes, Jordi, “The rise of machine
learning for detection and classification of malware: Research devel-
opments, trends and challenges”, Journal of Network and Computer
Applications, Vol. 153, 2020, pp. 1–22, ISSN 1084-8045.
[2] Symantec 2018 Internet Security Threat Report. Tech. rep. Symantec
Corporation, https://www.symantec.com/content/dam/symantec/
docs/reports/istr-_23-_executive-_summary-_en.pdf.
[3] Singh, A., Handa, A., Kumar, N., and Shukla, S. K., Malware
classification using image representation, in: International Symposium
on Cyber Security Cryptography and Machine Learning, Springer.
pp. 75–92, 2019.
[4] https://www.geeksforgeeks.org/malware-and-its-types
[5] https://www.techopedia.com/definition/29741/zero-day-malware
[6] Alzaylaee, Mohammed K., Yerima, Suleiman Y., Sezer, Sakir, DL-
Droid: Deep learning based android malware detection using real
devices, Computers & Security, Vol. 89, 2020, 101663, ISSN 0167-
4048, https://doi.org/10.1016/j.cose.2019.101663.
[7] McLaughlin, Niall, et al. “Deep Android malware detection,”
Proceedings of the Seventh ACM on Conference on Data and Application
Security and Privacy. 2017.
[8] Hou, S., Saas, A., Chen, L., and Y. Ye, “Deep4MalDroid: A Deep
Learning Framework for Android Malware Detection Based on
Linux Kernel System Call Graphs,” 2016 IEEE/WIC/ACM
International Conference on Web Intelligence Workshops (WIW), Omaha,
NE, 2016, pp. 104–111, doi: 10.1109/WIW.2016.040.
[9] Wang, Z, Cai, J., Cheng, S., and Li, W., “DroidDeepLearner:
Identifying Android malware using deep learning,” 2016 IEEE 37th
Sarnoff Symposium, Newark, NJ, 2016, pp. 160–165, doi: 10.1109/
SARNOF.2016.7846747.
[10] Yuan, Zhenlong, et al. “Droid-sec: Deep learning in android malware
detection,” Proceedings of the 2014 ACM conference on SIGCOMM.
2014.
[11] Karbab, E. B., Debbabi, M., Derhab, A., and Mouheb, D., “MalDozer:
Automatic framework for android malware detection using deep
learning,” Digital Investigation, vol. 24, 2018, pp. S48–S59.
[12] Zhu, D., Jin, H., Yang, Y., Wu, D., and Chen W., “DeepFlow: Deep
learning-based malware detection by mining Android application for
abnormal usage of sensitive data,” 2017 IEEE Symposium on
Computers and Communications (ISCC), 2017.
[13] Hou, S., Saas, A., Ye, Y., and Chen, L., “DroidDelver: An Android
Malware detection system using deep belief network based on API
call blocks.” Web-Age Information Management Lecture Notes in
Computer Science, 2016, vol. 9998, pp. 54–66.
[14] Peiravian, N., and Zhu, X., “Machine learning for Android malware
detection using permission and API calls,” 2013 IEEE 25th
96 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

International Conference on Tools with Artificial Intelligence, Herndon,

VA, 2013, pp. 300–305, doi: 10.1109/ICTAI.2013.53.
[15] Abikoye, Oluwakemi, Gyunka, Benjamin, and Oluwatobi, Akande,
“Android Malware Detection through Machine Learning
Techniques: A Review. International Journal of Online and Biomedical
Engineering (iJOE), vol. 16, 2020, p. 14. doi: 10.3991/ijoe.v16i02
.11549.
[16] https://purplesec.us/resources/cyber-security-statistics/ransomware/
[17] Maniath, S., Ashok, A., Poornachandran, P., Sujadevi, V. G., Au,
P. S., and Jan, S., Deep learning LSTM based ransomware detection.
In 2017 Recent Developments in Control, Automation & Power
Engineering (RDCAPE) (pp. 442–446). IEEE.
[18] Vinayakumar, R., Alazab, M., Jolfaei, A., Soman, K. P., and
Poornachandran, P. (2019, May), “Ransomware triage using deep
learning: Twitter as a case study.” In 2019 Cybersecurity and
Cyberforensics Conference (CCC), pp. 67–73. IEEE.
[19] Bae, S. I., Lee, G. B., and Im, E. G. (2020), “Ransomware detection
using machine learning algorithms,” Concurrency and Computation:
Practice and Experience, vol. 32, no. 18, e5422.
[20] Kok, S., Abdullah, A., Jhanjhi, N., and Supramaniam, M. (2019),
Ransomware, threat and detection techniques: A review. Int. J.
Computer Science and Network Security, vol. 19, no. 2, 136.
[21] Hwang, J., Kim, J., Lee, S., and Kim, K. (2020). “Two-Stage
Ransomware Detection Using Dynamic Analysis and Machine
Learning Techniques.” Wireless Personal Communications, vol. 112,
no. 4, 2597–2609.
[22] Alhawi, O. M., Baldwin, J., and Dehghantanha, A. (2018).
“Leveraging machine learning techniques for windows ransomware
network traffic detection.” In Cyber Threat Intelligence (pp. 93–106).
Cham, Germany: Springer.
[23] Yaqoob, I., Ahmed, E., ur Rehman, M. H., Ahmed, A. I. A., Al-
garadi, M. A., Imran, M., & Guizani, M. (2017). “The rise of
Ransomware and emerging security challenges in the Internet of
Things.” Computer Networks, vol.129, 444–458.
[24] Agrawal, R., Stokes, J. W., Selvaraj, K., & Marinescu, M. (2019, May).
“Attention in recurrent neural networks for ransomware detection.” In
ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP), pp. 3222–3226, IEEE.
[25] Zhang, H., Xiao, X., Mercaldo, F., Ni, S., Martinelli, F., & Sangaiah,
A. K. (2019). “Classification of Ransomware families with machine
learning based on N-gram of opcodes.” Future Generation Computer
Systems, 90, 211–221.
[26] Shhadat, Ihab, Bataineh, Bara’, Hayajneh, Amena, and Al-Sharif,
Ziad A., “The use of machine learning techniques to advance the
detection and classification of unknown malware.” Procedia Computer
Science, vol. 170, 2020, pp. 917–922, ISSN 1877-0509.
 A VIEW OF CYBER ATTACK DETECTION 97

[27] Gupta, Deepak, and Rani, Rinkle, “Improving malware detection

using big data and ensemble learning.” Computers & Electrical
Engineering, vol. 86, 2020, Article ID: 106729, ISSN 0045-7906.
[28] Liu, Xinbo, Lin, Yaping, Li, He, and Zhang, Jiliang, “A novel method
for malware detection on ML-based visualization technique.” Computers
& Security, vol. 89, 2020, Article ID: 101682, ISSN 0167-4048.
[29] Gibert, Daniel, Mateu, Carles, and Jordi Planes, “HYDRA: A
multimodal deep learning framework for malware classification.”
Computers & Security, vol. 95, 2020, 101873, ISSN 0167-4048.
[30] Xiao, Guoqing, Li, Jingning, Chen, Yuedan, and Li, Kenli,
“MalFCS: An effective malware classification framework with auto-
mated feature extraction based on deep convolution neural networks.”
Journal of Parallel and Distributed Computing, vol. 141, 2020,
pp. 49–58, ISSN 0743-7315.
[31] Vasan, Danish, Alazab, Mamoun, Wassan, Sobia, Naeem, Hamad,
Safaei, Babak, and Zheng, Qin, “IMCFN: Image-based malware
classification using fine-tuned convolution neural network architecture.”
Computer Networks, vol. 171, 2020, 107138, ISSN 1389-1286.
[32] Zhang, Jixin, Qin, Zheng, Yin, Hui, Ou, Lu, and Zhang, Kehuan, “A
feature-hybrid malware variants detection using CNN based opcode
embedding and BPNN based API embedding.” Computers & Security,
vol. 84, 2019, pp. 376–392, ISSN 0167-4048.
[33] Xiaofeng, Lu, Fangshuo, Jiang, Xiao, Zhou, Shengwei, Yi Jing, Sha,
and Lio, Pietro, “ASSCA: API sequence and statistics features
combined architecture for malware detection.” Computer Networks,
vol. 157, 2019, pp. 99–111, ISSN 1389-1286.
[34] Yoo, Suyeon, Kim, Sungjin, Kim, Seungjae, Byunghoon Kang, Brent,
“AI-HydRa: Advanced hybrid approach using random forest and
deep learning for malware classification,” Information Sciences,
vol. 546, 2021, pp. 420–435, ISSN 0020-0255.
[35] Karbab, ElMouatez Billah, and Debbabi, Mourad, “MalDy: Portable,
data-driven malware detection using natural language processing and
machine learning techniques on behavioral analysis reports,” Digital
Investigation, vol. 28, Supplement, 2019, pp. S77–S87, ISSN
1742-2876.
[36] Bahtiyar, Şerif, Yaman, Mehmet Barış, and Altıniğne, Can Yılmaz,
“A multi-dimensional machine learning approach to predict advanced
malware.” Computer Networks, vol. 160, 2019, pp. 118–129, ISSN
1389-1286
[37] asmitha, K. A., and Vinod, P., “Linux malware detection using non-
parametric statistical methods,” 2014 International Conference on
Advances in Computing, Communications and Informatics (ICACCI),
New Delhi, 2014, pp. 356–361.
[38] Asmitha, K. A., and Vinod, P., “A machine learning approach for linux
malware detection,” 2014 International Conference on Issues and
98 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Challenges in Intelligent Computing Techniques (ICICT), Ghaziabad,

2014, pp. 825–830, doi: 10.1109/ICICICT.2014.6781387.
[39] Gavriluţ, D., Cimpoeşu, M., Anton, D., and Ciortuz, L., “Malware
detection using machine learning.” 2009 International Multiconference
on Computer Science and Information Technology, Mragowo, 2009,
pp. 735–741, doi: 10.1109/IMCSIT.2009.5352759.
[40] Muthuswamy, Sujithra (2017), “Detection of malware applications in
mobile devices using supervised machine learning algorithms.”
International Conference on Business Analytics and Intelligence, IIM
Bangalore, December 2017.
[41] Aslan, Ömer (2017), Performance comparison of static malware analysis
tools versus antivirus scanners to detect malware. In International
Multidisciplinary Studies Congress (IMSC).
[42] https://www.hackingtutorials.org/malware-analysis-tutorials/basic-
malware-analysis-tools/
[43] https://www.hackingtutorials.org/malware-analysis-tutorials/dynamic-
malware-analysis-tools/
 5
REGRESSION ALGORITHMS IN
MACHINE LEARNING
USHA SAKTHIVEL1,
NEHA SINGHAL2,
PETHURU RAJ CHELLIAH3,
AND ASHWINI R. MALIPATIL4
1
Dean Research and Innovation, Rajeswari College of Engineering,
Bengaluru, India, Prof. & Head, Dept. of CSE, Visvesvaraya
Technological University Belagavi, Karnataka, India
2
Dept. of ISE, Sri Krishna College of Engineering and Technology,
Coimbatore, India
3
Chief Architect, RJIL, Bengaluru
4
RRCE, Dept. of CSE, Sri Krishna College of Engineering and
Technology, Coimbatore, India

Contents

Introduction 100
Background Study 100
Different Types of Regression Algorithms in Machine
Learning 103
Simple Linear Regression 103
Significance of Linear Regression 105
Simple Linear Regression Model in Health-Related
Data Analysis 105
Polynomial Regression 106
Polynomial Regression Model in COVID-19 Growth
Pattern Analysis 108
Support Vector Machine Regression 108
Decision Tree Regression 110
Random Forest Regression 111
Case Study on Air Quality Prediction Using Machine
Learning Regression Techniques 111
Conclusions and Future Work 121
References 121

DOI: 10.1201/9781003119838-5 99
100 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Introduction

In machine learning (ML), various kinds of algorithm are used for the
prediction of results based on the provided data. The relationship that
exists among the input data and the expected output is used for fore­
telling results based on the prediction. Regression algorithm is a pre­
diction method that works on supervised learning techniques. These
supervised algorithms allow machines to train and learn so as to include
some kind of automation in outcomes. The predicted output is calcu­
lated on the basis of continuous numerical values in regression pro­
blems. Regression algorithms are categorized in linear and non-linear
algorithms. Regression algorithms are used in finance forecasting,
market trend analysis, time series prediction, and even for drug response
modeling for COVID-19 situations. Regression algorithms use con­
tinuous numerical analysis on the patterns to perform the predictions of
future outcomes. Some of the popular types of regression algorithms are
simple liner regression, polynomial regression, support vector regression,
decision tree regression, and random forest regression.

Background Study

Punal Pahwa et al. [1]. The paper works on the analysis of stock
exchange marketing where the regression method is used as a strong
predictor in the ML area. The regression method provides a statistics
accuracy in the investment and stock exchange. Stock market fully
depends on the prediction of the shares in the present and future.
Hence, ML algorithm with the regression methods predicts the fu­
ture stock price for the business.
Zhixia Yang et al. [2]. The paper works on multiclass classification
in ML with the ordinal regression method. Multiclass classification
works on the construction of decision tree. Support vector machines
are used as tools for the classification. The paper represents the kernel
ordinal regression for three class methods.
Ekta Ghambir et al. [3]. The paper work on the analysis and pre­
diction of the spread of novel Corona virus, which helps in taking
future precaution and necessary steps. With the implementation of ML
model with polynomial regression algorithm, it was possible to predict
the spread of virus in the next 60 days in India. Regression algorithm is
 REGRESSION ALGORITHMS IN MACHINE LEARNING 101

a supervised ML algorithm, which reads the previous data, analyze it,

and provide the accurate prediction on the live data. The extraction and
selection of data need to be done for the best results from this model.
Shen Rong et al. [4]. The paper analyzes the iced product affected
by the change in the temperature conditions. Linear regression model
works on the relationship between the independent and dependent
data. In this work, forecast temperature is taken as the independent
variable and the iced product as the dependent variable. The im­
plementation of leaner regression is done on the Python 3.6 platform
with various packages. The prediction results given by the regression
model help the company to have a control over the production and
sale of the iced products in the market.
Ki-Young Lee et al. [5]. The paper works on the comparison of
linear regression and artificial neural networks with the help of deep
learning and linear regression algorithm through the supervised learning.
AI works on two methods: supervised and unsupervised. Supervised
is when the human provides the data and logical interference.
Unsupervised is when the computer takes the data and performs the task
with the supervision of the previewed data. The paper works to predict
the future industry profit using the regression model using neural AI.
Khushbu Kumari et al. [6]. The paper explains about the analysis of
linear regression model. Regression model is a statistical model that
compares the relationship between the dependent and independent
variables of the real-time dataset. Here, the patients having blood
pressure are taken as the variable for the prediction of variation in the
blood pressure with different variations in the patient body conditions.
The linear regression model predicts the increase in the BP of the
person with the change in his daily routine of his health condition.
Karunya Rathan et al. [7]. The paper speaks on the crypto cur­
rency, which is also called Bitcoin and is considered the asset of
internet. The increasing demand of Bitcoin on the internet platform
has made this paper to work out with the prediction of price rate of
the Bitcoin day to day. Their study includes ML algorithm with the
decision tree and regression algorithm. The comparison was made
between the two algorithms, and the outcome of the regression al­
gorithm is found to be more accurate than the decision tress.
Sovojit Manna et al. [8]. The paper works on the flight delay
prediction by using the ML algorithm. Gradient booster decision tree
102 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

is used for prediction of air traffic delay. Gradient booster decision tree
can handle the regression model more effectively. The proposed fea­
ture can be used by the public, tourist, and airline agencies to obtain
information about the delay in the flight in a very accurate manner.
AI Amin Biswas et al. [9]. Used a random forest regression
method for detection of road accidents and number of causalities.
The regression model first analyzes the dataset and then predicts the
testing dataset or the real-time dataset. It is seen that the regression
model provides a good performance.
Tessy Badriyah et al. [10]. The paper analyzes the stroke disease
classification using ML. With the help of image processing techni­
ques, the images are collected as the dataset. The Random forest
algorithm uses the testing dataset and the real data for the prediction.
Random forest algorithm proved to depict the highest accuracy.
K Vijyakumar et al. [11]. The proposed system works on the early
prediction of diabetes in patients. The system is developed by the concept
of ML algorithm with the Random Forest algorithm. The regression
models are used for the classification and comparison of the dataset for
the prediction. Hence, the algorithm depicts high accuracy in the results.
Sunakshi Mamgain et al. [12]. Author have proposed a model that
works on the prediction of the popularity of cars in the market by
considering the parameters such as price, maintenance, safety,
mileage, etc. With the introduction of ML algorithm, Random
Forest algorithm, and regression technique for prediction of the
popularity of cars on the scaling 1–4.
Giorgio Biagetti et al. [13]. The paper works on the analysis of
Bernstein polynomials for nonlinear system identification. Bernstein
polynomial has a property that the co efficient is the value of
the function to be approximated at point in fixed grid. Bernstein
polynomial is particularly suitable to solve multivariate regression
problems. The results show the accuracy of the technique for the
identification of non-linear systems and provide a better performance
with respect to standard techniques.
Swapanil Sayan Saha et al. [14]. Authors are concerned about the
adulteration of fruits, vegetables, and fish with the hazards of formalin.
The study works on the detection of formalin used at a high rate to keep
the fruits, vegetables, and fish fresh for longer time, which causes health
issues. The system model incorporates polynomial regression, linear
 REGRESSION ALGORITHMS IN MACHINE LEARNING 103

regression, and Levenberg–Marquardt algorithm, which are supervised

ML algorithm. The system provides an accurate prediction of the level
of concentration of formalin at various levels of temperature. By the
application of logistic regression and support vector machine, the system
can identify artificially added and naturally formulated formalin.
Ching-Seh (Mike) Wu et al. [15]. The study focuses on the im­
plementation of model with different ML techniques, such as regression
and neural networks, which can predict the analysis of customers on the
Friday sale: the customers that spent in the past and predicting the
customers for the future with the accuracy to help the retail shop person
plan how to control the crowd, manage his billings, and increase his sale.
Junheung Park et al. [16]. The study focuses on the psychological
wellness indices with various parameters, such as anger, depression,
stress, fatigue, etc. By predicting these indices, we can increase the
practicality of psychological wellness and implement it in the medical
system. The prediction model is developed by using four ML algo­
rithms, namely, support vector regression, generalized regression neural
networks, multilayer perception, and k nearest neighbor regression.

Different Types of Regression Algorithms in Machine Learning

The regression algorithms in ML are categorized in different types as

given below:
1. Simple Linear Regression
2. Polynomial Regression
3. Support Vector Regression
4. Decision Tree Regression
5. Random Forest Regression

Simple Linear Regression

Linear regression is a measurable strategy that was presented by

Sir Francis Galton in 1984 [6]. Linear regression is perhaps the most
well-known and fascinating kind of regression strategy. Straightforward
linear regression is a factual strategy that empowers to close and study
connections between two distinctive quantitative factors. Linear regres­
sion is a direct model, a model that finds a direct connection between the
104 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

information factors (A) and the single yield variable (B). Here, we foresee
an objective variable (B) dependent on the information variable (A). A
straight relationship should exist between target variable and indicator,
thus comes the name linear regression. When there is a solitary info
variable (A), the strategy is known as a straightforward linear regression.
When there are different information factors, the technique is alluded as
numerous linear regressions. Consider anticipating the compensation of a
representative dependent on his/her age. We can without much of a
stretch distinguish that there is by all accounts a relationship between
representative’s age and compensation (more the age, more is the
compensation). Probably, the most mainstream utilizations of linear
regression calculation are in monetary portfolio expectation, compensa­
tion gauging, land forecasts, and in rush hour gridlock in showing up at
ETAs. The speculation of linear regression is
A = x + yB

A addresses pay, B is representative’s age, and x and y are the coef­

ficients of the condition. In this way, to anticipate A (compensation)
given B (age), we need to know the estimations of x and y (the
model’s coefficients).
While preparing and assembling a regression model, these coef­
ficients are learned and fitted to prepare information. The point of
the preparation is to locate the best-fit line with the end goal that
cost work is limited. The expense work helps in estimating the
blunder. During the preparation cycle, we attempt to limit the
mistake among real and anticipated qualities and in this manner,
limiting the expense works. In Figure 5.1, the red focuses are the
information and the blue line is the anticipated line for the pre­
paration information. To get the anticipated worth, these informa­
tion focuses are projected on to the line.
To sum up, our point is to discover such estimations of coefficients
which will limit the expense work. The most widely recognized ex­
pense work is mean squared error (MSE), which is equivalent to the
normal squared contrast between a perception’s genuine and antici­
pated qualities. The coefficient esteems can be determined utilizing
the Gradient Descent approach, which will be examined later in
detail. To give a short agreement, in Gradient plunge, we start for
certain arbitrary estimations of coefficients, process the inclination of
 REGRESSION ALGORITHMS IN MACHINE LEARNING 105

Figure 5.1 Linear Regression Graph

cost work on these qualities, update the coefficients, and compute the
expense work once more. This cycle is rehashed until we locate a base
estimation of cost work.

Significance of Linear Regression

The meaning of linear regression model is important for the ac­

companying underneath referenced reasons:

Descriptive – It helps in investigating the strength of the asso­

ciation between the outcome (dependent variable) and independent
predictor variables.
Adjustment – It adapts with the impact of covariates or anomalies.
Predictors – It helps in foreseeing and assessing the future danger
factors that influence the dependent variable.
Extent of prediction – It helps in investigating the degree of
progress in the free information variable and the reliant yield variable.
Prediction – It helps in evaluating the new outcomes.

Simple Linear Regression Model in Health-Related Data Analysis In

natural or health-related clinical information, linear regression model
is frequently used to portray connections between two factors or
among a few factors through measurable assessment. For instance, to
know whether the probability of having high systolic BP (SBP) is
affected by variables, for example, age, tallness, and weight, direct
106 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

relapse is utilized. The variable to be clarified, i.e., SBP, is known as

the needy variable, or on the other hand, the reaction factors that
clarify it, age, tallness, weight, and gender are, called free factors.

Polynomial Regression

Non-linear information is generally experienced in a day-by-day life.

Think about a portion of the conditions of movement as con­
centrated in material science and physics.
Projectile Motion: The stature of a projectile is calculated as
h =–½ gt2 + ut + ho
Condition of movement under free fall: The distance went by an article
in the wake of falling unreservedly under gravity for t seconds is ½ gt2.
Distance traveled by a consistently accelerated body: The distance
can be determined as

ut + 1/2at 2,

where,
g = acceleration due to gravity
u = beginning speed
ho = initial stature
a = acceleration
Notwithstanding these models, Non-linear patterns are additionally
seen in the development pace of tissues, the advancement of infection
pandemic, dark body radiation, the movement of the pendulum, and
so forth. These models obviously demonstrate that we can’t generally
have a linear connection between the independent and dependent
attributes. Henceforth, linear regression is a helpless decision for
managing such nonlinear circumstances. This is the place where
polynomial regression acts as the hero!!
Polynomial regression is an incredible method to experience the
circumstances where a quadratic, cubic, or a more serious level nonlinear
relationship exists. The basic idea in polynomial regression is to add
forces of every autonomous independent variable as new attributes and
afterward train a linear model on this extended assortment of highlights.
In polynomial regression, we change the first highlights into
polynomial highlights of a given degree and afterward apply linear
 REGRESSION ALGORITHMS IN MACHINE LEARNING 107

Figure 5.2 Polynomial Regression Graph

regression on it. Think about the above straight model: Y = a + bX is

changed into something like

Y = a + b X + c X2
It is as yet a linear model; however, the curve is currently quadratic as
opposed to a line. scikit-learn gives polynomial features class to change
the highlights. Consequently, if we attempt to utilize a straightforward
linear regression in the given chart, then the linear regression line will not
fit well overall. It is extremely hard to fit a linear regression line in the
diagram given in Figure 5.2 with a low estimation of mistake. Thus, we
can attempt to utilize the polynomial regression to fit a polynomial line so
that we can accomplish a base blunder or least expense work (Figure 5.2).
On the off chance that we increment the degree to a high worth,
the curve becomes overfitted as it learns the noise and outliers in the
information also.
Advantages of Using Polynomial Regression:
• Polynomial gives the best estimation of the connection between
the dependent and independent factors.
• A broad range of functions can be fit under it.
• Polynomial fundamentally fits a wide scope of the curve.

Disadvantages of Using Polynomial Regression:

• The presence of a couple of anomalies in the information can
truly influence the aftereffects of the nonlinear examination.
• These are excessively touchy to the exceptions.
108 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

• In expansion, there are lamentably fewer model approval ap­

paratuses for the recognition of anomalies in nonlinear regres­
sion than there are for linear regression.

Polynomial Regression Model in COVID-19 Growth Pattern Analysis

Coronavirus is spreading inside such a gigantic scourge for the globe.
This plague taints a ton of people in Egypt. The world wellbeing as­
sociation states that COVID-19 could be spread starting with one
individual then onto the next at an exceptionally quick speed through
contact and respiratory splash. Nowadays, Egypt and all nations
worldwide ascend to a compelling advance to examine this illness and
kill the impacts of this pestilence. In this paper, the genuine in­
formation base of Coronavirus for Egypt has been broken down from
February 15, 2020 to June 15, 2020, and anticipated with the quantity
of patients that will be tainted with Coronavirus, and assessed the
scourge last size. A few relapse examinations models have been applied
for information investigation of Coronavirus in Egypt. In this in­
vestigation, we’ve applied seven relapse examination-based models that
are outstanding polynomial, quadratic, third-degree, fourth-degree,
fifth-degree, sixth-degree, and logit development individually for the
COVID-19 dataset. Consequently, the remarkable fourth-degree,
fifth-degree, and sixth-degree polynomial relapse models are as­
tounding, unlike fourth-degree model that will help the public au­
thority in setting up its methodology for one month. What’s more! We
have applied the notable logit development relapse model and acquired
the accompanying epidemiological bits of knowledge: right off the bat,
the plague pinnacle might actually reach on June 22, 2020 and final
season of pestilence on September 8, 2020. Besides, the last all-out size
for cases 1.6676eþ05 cases. The activity of legislature of inter-evention
over a generally long stretch is important to limit the last plague size.

Support Vector Machine Regression

In ML and data analytics, support vector machines (SVMs, also support-

vector organizations) are supervised learning models with associated
learning algorithms that analyze and investigate pattern regression ana­
lysis based on classifications. It was proposed at AT&T Bell Laboratories
by Vapnik et al. In Support Vector Machine Regression (SVR), we
 REGRESSION ALGORITHMS IN MACHINE LEARNING 109

distinguish a hyperplane with most extreme edge to such an extent that

the greatest number of information focus is inside that edge. SVRs are
practically like the SVM arrangement calculation. SVM is another most
remarkable calculation with solid hypothetical establishments dependent
on the Vapnik–Chervonenkis hypothesis, as characterized by oracle docs.
This supervised ML algorithm has solid regularization and can be uti­
lized both for classification or regression challenges. They are char­
acterized by usage of kernels, the sparseness of the solution, and the
capacity control gained by acting on the margin, or on number of support
vectors, etc. The capacity of the system is controlled by parameters that
do not depend on the dimensionality of feature space. Since the SVM
algorithm operates natively on numeric attributes, it uses a z-score
normalization on numeric attributes. In regression, SVM algorithms use
margin of loss function to solve regression problems.
Rather than limiting the mistake rate as in straightforward linear
regression, we attempt to fit the blunder inside a specific edge. Our
goal in SVR is to fundamentally consider the focuses that are inside
the edge. Our best fit line is the hyperplane that has the most ex­
treme number of focuses. In SVM algorithm, “c” margin is always
considered as a variation margin.
In Figure 5.3, red (top and bottom) lines represent the patterns
with the margin “c” in the positive and negative axes margin.
SVM regression algorithms calculations have discovered a few
applications in the oil and gas industry, order of pictures, and text
and hypertext arrangement. In the oil fields, it is explicitly utilized for

Figure 5.3 Support Vector Regression with Margin

110 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

investigation to comprehend the situation of layers of rocks and make

2-D and 3-D models as a portrayal of the earth.

Decision Tree Regression

Decision trees can be utilized for classification and characterization

just as regression. In Decision trees, at each level, we need to dis­
tinguish the division property. On account of regression, the ID3
calculation can be utilized to distinguish the parting hub by lessening
the standard deviation (in order data classification is utilized).
A decision tree is worked by apportioning the information into
subsets containing examples with comparable qualities (homo­
genous). Standard deviation is utilized to compute the homogeneity
of a mathematical example. On the off chance that the mathematical
example is totally homogeneous, its standard deviation is zero.
The steps for finding the splitting node are briefly described below:
1. Calculate the standard deviation of the target variable using the
below formula.
N
1
= (xi µ)2
N i =1

2. Split the dataset on various attributes and figure the standard

deviation for each branch (standard deviation for target and in­
dicator). This worth is deducted from the standard deviation
before the split. The outcome is the standard deviation decrease.
SDR(T , X ) = S (T ) S (T , X )

3. The attribute with the biggest standard deviation decrease is

picked as the partition node.
4. The dataset is partitioned depending on the estimations of the
chose trait. This interaction is run recursively on the non-leaf
branches until all information is handled.
To maintain a strategic distance from overfitting, the coefficient of
deviation (CV) is utilized, which chooses when to quit spreading. At
last, the normal of each branch is allotted to the connected leaf hub (in
relapse mean is taken though in grouping method of leaf nodes is taken).
 REGRESSION ALGORITHMS IN MACHINE LEARNING 111

Random Forest Regression

Arbitrary backwoods are a gathering approach where we consider the

forecasts of a few choice regression trees.
1. Select An arbitrary focuses.
2. Identify n where n is the quantity of choice tree regressors to be
made. Repeat Stages 1 and 2 to make a few relapse trees.
3. The normal of each branch is appointed to the leaf node in
every decision tree.
4. To foresee yield for a variable, the normal of the multitude of
forecasts of all choice trees is mulled over.

Arbitrary Forest forestalls overfit (which is regular in choice trees)

by making irregular subsets of the highlights and building more
modest trees utilizing these subsets.

Case Study on Air Quality Prediction Using Machine Learning Regression

Techniques

Air quality checking and forecast have got perhaps the most funda­
mental movement in numerous mechanical and metropolitan territories.
The nature of air is exceptionally influenced because of the dif­
ferent types of poisons. With expanding air contamination, pro­
ductive air quality observing models is to be actualized; these
models gather data about the centralization of air toxins. We
propose a general and viable way to deal with care of the three
issues: expectation, insertion, and highlight examination. These
three issues were settled already utilizing three unique models;
however, now in the proposed framework, we can tackle these
three issues in a single model called the Deep Air Learning
(DAL). This methodology is related to unlabeled spatio-fleeting
information to improve the presentation of insertion and the ex­
pectation of air quality. We assess our methodology by the ex­
aminations dependent on ongoing information sources acquired
by the Karnataka State Pollution Control Board (KSPCB), India.
The point of this exploration is to examine different computerized
reasoning and ML-based procedures for checking and anticipating
112 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 5.4 Air Quality Prediction Model

the air quality. Python language is used for the development of the
experimentation.
Figure 5.4 represents the air quality prediction system in a detailed
manner by representing how the data from database are captured,
validated, and processed by the prediction system till it reaches the
final step, i.e., generating the query result.
Figure 5.5 shows the representation of how python code imports
libraries. Figure 5.6 represents certain columns of parameter values
for different index variables.
Figure 5.7 represents the AQI range with various AQI
categories defined that are considered during the experimentation work.
Figure 5.8 represents data labeling. In Figure 5.9, data graphs are
plotted after the data analytics.
Figure 5.10 represents the data split for testing and training datasets.
Figure 5.11 shows the air quality analysis, its parameters, and its
range taken for the analysis.
In Figure 5.12, AQI category, pollutants, breakpoints, and health
impacts are represented after the data analysis.
REGRESSION ALGORITHMS IN MACHINE LEARNING 113

Figure 5.5 How Python Code Imports Libraries

114 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 5.6 Columns of Certain Parameters

Figure 5.7 Air Quality Index Category and Range
REGRESSION ALGORITHMS IN MACHINE LEARNING
115
116 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 5.8 Data Preprocessing and Labeling

Figure 5.9 Data Analysis Stage and Data Graph
REGRESSION ALGORITHMS IN MACHINE LEARNING
117
 118

Figure 5.10 Data Split into Testing and Training Dataset

APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT
Figure 5.11 Air Quality Analysis, Parameters, and Its Range
REGRESSION ALGORITHMS IN MACHINE LEARNING
119
 120

Figure 5.12 AQI Category, Pollutants, Breakpoints, and Health Impacts

APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT
 REGRESSION ALGORITHMS IN MACHINE LEARNING 121

Conclusions and Future Work

Machine learning (ML) is a very incredible and evitable tool. ML

algorithms are very useful for data analysis and future information
forecast. In this work, various ML regression algorithms are dis­
cussed in detail. The algorithms are discussed with various available
real-time problem scenarios. This work considers only the supervised
ML-based methods. The case study elaborates the regression algo­
rithm for the air pollution control analysis with various samples of
real datasets. The proposed work is useful for future air quality and
pollution control based on data analysis techniques and shows the
effectiveness for the future research domain.

References
[1] Pahwa, K., and Agarwal, N., “Stock Market Analysis Using Supervised
Machine Learning,” 2019 International Conference on Machine
Learning, Big Data, Cloud and Parallel Computing (COMITCon),
Faridabad, India, 2019, pp. 197–200, doi: 10.1109/COMITCon.2019.
8862225.
[2] Yang, Zhixia, Deng, Naiyang, and Tian Yingjie, “A Multi-class
Classification Algorithm Based on Ordinal Regression Machine,”
International Conference on Computational Intelligence for Modelling,
Control and Automation and International Conference on Intelligent
Agents, Web Technologies and Internet Commerce (CIMCA-IAWTIC’06),
Vienna, 2005, pp. 810–815, doi: 10.1109/CIMCA.2005.1631568.
[3] Gambhir, E., Jain, R., Gupta, A., and Tomer, U., “Regression Analysis of
COVID-19 Using Machine Learning Algorithms,” 2020 International
Conference on Smart Electronics and Communication (ICOSEC),
Trichy, India, 2020, pp. 65–71, doi: 10.1109/ICOSEC49089.
2020.9215356.
[4] Rong, Shen, and Bao-wen, Zhang. “The Research of Regression
Model in Machine Learning Field,” MATEC Web of Conferences,
Vol. 176, No. 01033 (2018), https://doi.org/10.1051/matecconf/
201817601033, IFID 2018.
[5] Lee, Ki-Young, Kim, Kyu-Ho, Kang, Jeong-Jin, Choi, Sung-Jai, Im,
Yong-Soon, Lee, Young-Dae, and Lim, Yun-Sik, “Comparison and
Analysis of Linear Regression & Artificial Neural Network.”
International Journal of Applied Engineering Research, ISSN 0973-
4562 Vol. 12, No. 20 (2017), pp. 9820–9825.
[6] Kumari, Khushbu, and Yadav, Suniti, “Linear Regression Analysis
Study,” Journal of the Practice of Cardiovascular Sciences, Vol. 4, No. 1
(January–April 2018), pp. 33–36.
122 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

[7] Rathan, K., Sai, S. V., and Manikanta, T. S. (2019). Crypto-

Currency Price Prediction Using Decision Tree and Regression
Techniques. 2019 3rd International Conference on Trends in Electronics
and Informatics (ICOEI). doi: 10.1109/icoei.2019.8862585.
[8] Manna, S., Biswas, S., Kundu, R., Rakshit, S., Gupta, P., and
Barman, S. (2017). “A Statistical Approach to Predict Flight Delay
Using Gradient Boosted Decision Tree.” 2017 International
Conference on Computational Intelligence in Data Science(ICCIDS).
doi: 10.1109/iccids.2017.8272656
[9] Biswas, A. A., Mia, M. J., and Majumder, A., “Forecasting the
Number of Road Accidents and Casualties Using Random Forest
Regression in the Context of Bangladesh.” 2019 10th International
Conference on Computing, Communication and Networking Technologies
(ICCCNT), 2019, doi: 10.1109/icccnt45670.2019.8944500
[10] Badriyah, T., Sakinah, N., Syarif, I., and Syarif, D. R. Machine
Learning Algorithm for Stroke Disease Classification. 2020
International Conference on Electrical, Communication, and Computer
Engineering (ICECCE), 2020, doi: 10.1109/icecce49384.2020.9179307
[11] VijiyaKumar, K., Lavanya, B., Nirmala, I., and Caroline, S. S. (2019).
Random Forest Algorithm for the Prediction of Diabetes. 2019 IEEE
International Conference on System, Computation, Automation and
Networking (ICSCAN). doi: 10.1109/icscan.2019.8878802
[12] Mamgain, S., Kumar, S., Nayak, K. M., and Vipsita, S. (2018). Car
Popularity Prediction: A Machine Learning Approach. 2018 Fourth
International Conference on Computing Communication Control
and Automation (ICCUBEA). doi: 10.1109/iccubea.2018.8697832
[13] Biagetti, G., Crippa, P., Falaschetti, L., and Turchetti, C. (2017).
Machine learning regression based on particle bernstein polynomials
for nonlinear system identification. 2017 IEEE 27th International
Workshop on Machine Learning for Signal Processing (MLSP).
doi: 10.1109/mlsp.2017.8168148
[14] Saha, S. S., Siraj, M. S., and Habib, W. B. (2017). FoodAlytics: A
Formalin Detection System Incorporating a Supervised Learning
Approach. 2017 IEEE Region 10 Humanitarian Technology Conference
(R10-HTC). doi: 10.1109/r10-htc.2017.8288898
[15] Wu, C. M., Patil, P., and Gunaseelan, S., “Comparison of Different
Machine Learning Algorithms for Multiple Regression on Black
Friday Sales Data,” 2018 IEEE 9th International Conference on
Software Engineering and Service Science (ICSESS), Beijing, China,
2018, pp. 16–20, doi: 10.1109/ICSESS.2018.8663760.
[16] Park, J., Kim, K., and Kwon, O., “Comparison of Machine Learning
Algorithms to Predict Psychological Wellness Indices for Ubiquitous
Healthcare System Design,” Proceedings of the 2014 International
Conference on Innovative Design and Manufacturing (ICIDM), Montreal,
QC, 2014, pp. 263–269, doi: 10.1109/IDAM.2014.6912705.
 6
MACHINE LEARNING-BASED
INDUSTRIAL INTERNET OF THINGS
(IIOT) AND ITS APPLICATIONS
M. SURESH KUMAR1, S. RACHEL1,
AND M. V. KAVISELVAN2
1
Department of Information
Technology, Sri Sairam Engineering
College, Chennai, India
2
Department of Mechanical
Engineering, Sri Sairam
Engineering College, Chennai, India

Contents

History of Internet of Things 124

Scope of IoT 125
Features of IOT 126
Leading Types of IoT 126
Advantages of IoT 126
Machine Learning and Deep Learning 127
Applications of ML 127
IIoT – Industrial Internet of Things 128
History of IIoT 129
Advantages of IIoT 130
IIoT in Industry 4.0 130
IoT and IIoT 131
The Future of IIoT 132
IIoT Use Cases 132
What Is Smart Manufacturing? 134
Assets Tracking 135
Industrial Example 135
Digital Twins 136

DOI: 10.1201/9781003119838-6 123

124 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Industrial Example 137

Advantages of Digital Twins 137
Smart Metering 138
Advantages of Smart Metering System 139
Disadvantages of Smart Metering System 140
Smart Farming 140
How Smart Farming Can Be Done? 141
Data Collection 142
Informed Action and Planning 142
Things to Be Considered before Implementing Smart
Farming 142
Advantages of Smart Farming: 144
Livestock Management Using IoT 144
Monitoring the Health of Cattle 144
Monitoring Reproductive Cycle and Calving 145
Fine-Tuning Feeding 146
Maximizing Milking 146
Connected Vehicles 147
Conclusion 150
References 151

History of Internet of Things

The Internet of Things [1] was started in the early 1980s, but it was not
publicly called until 1999. The first example of an IoT is a Coca Cola
machine, based at Carnegie Mellon University. Local programmers can
link it to the refrigerator appliance through the internet and check to
see if a drink was available and if it was cold before the ride.
By 2013, IoT had developed into a device using different tech­
nologies, ranging from wireless networking, embedded devices,
networks of wireless sensors, GPS, control systems, etc.
Kevin Ashton, Executive Director and Visiting Engineer, is known
as the Father of IoT, at the Massachusetts Institute of Technology
(MIT). He led the work on coding for the next decade. The Internet
of Things word was invented by Jeremy Cowan. The first IoT system
connected to a computer via TCP/IP protocol was toaster.it.
 MACHINE LEARNING-BASED IIOT 125

Figure 6.1 Applications of IoT

IoT works in nearly every industry, namely, financial services,

mobile banking, health care, etc. IoT shocked us with so many things
we had never imagined before. Everything changed, from automatic
home security to smart transportation. Here, we will identify the
sectors that are booming with IoT.
It is a matter of fact today that a vast number of users have surpassed
the bulk of IoT products on this planet. There are actually 7.62 billion
people on our planet, but to your horror, by 2021, with a growing IoT
system graph, nearly 20 billion IoT smart devices will be up and
running with a surge in demand for 5G networks (Figure 6.1).

Scope of IoT

The rapid growth of technology such as embedded systems, wireless

sensor technology, machine learning (ML), and automation plays
vital role in IoT development. This emerging technology has a
greater scope in many technological fields, including the following:
• Automotive and transportation industry
• Telecommunication
• Health care
• Logistics
• Supply chain
• Agriculture
• Wearable devices
126 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

• Smart cites and smart home

• Smart retail.

Features of IOT

Following are the seven crucial characteristics of IOT:

• Connectivity
• Things
• Data
• Communication
• Intelligence
• Action
• Ecosystem

Leading Types of IoT

• LPWAN (Low-Power Wide AreaNetwork)

• Cellular (3G/4G/5G)
• Zigbee and other meshprotocols
• Bluetooth and BLE
• Wi-Fi
• RFID (Radio Frequency Identification).

Advantages of IoT

IoT is going to bring many revolutionary changes in this new tech­

nological era, and it has many advantages, such as great number of new
business opportunities, new capabilities to predict and act, improved
monitoring, fine-tuned services and products, improved control of
operation process, and the most essential, starting of new revenue.
There are also many other advantages to this technology.
Eventhough IoT has many advantages, it has few disadvantages also.
The main three disadvantages are:
• Breach of privacy
• Over-reliance on technology
• Loss of jobs
 MACHINE LEARNING-BASED IIOT 127

Machine Learning and Deep Learning

Machine Learning (ML) is a subset of artificial intelligence (AI)

technology. It is simply defined as the study of computer algorithms
that can be updated automatically by self-learning process. ML im­
proves the dataset by self-learning process. ML is used to collect the
various types of data while working and all the data are analyzed and
organized by a cloud storage system. From these gathered real-time
data, ML improves its efficiency and provides accurate actions de­
pending on the data analyzed. The learning process of this system is
majorly divided into three types:
• Supervised learning
• Unsupervised learning
• Semi-supervised learning
All these processes improve the idea of the system with less human
intervention. In supervised learning, the example input and the goal
of operation are provided to the system. Thus it will be easy for the
system to collect the real-time data from the process of operation and
improve the idea of the system.
In unsupervised learning, the example input and the goal of the
process will not be provided to the system. In this type of learning, ML
will improve its process only by gathering the real-time information
from the operation. This type of learning is quite difficult to analyze
the data collected and execute the preferred action on time.
Semi-supervised learning is the moderate learning process of ML
technology. In this learning process, some of the example inputs and goal
of operation will be provided to the system; thus, it obtains the other
needed data from the time of operation. This is easier than unsupervised
learning and harder than supervised learning process (Figure 6.2).

Applications of ML

• Image recognition
• Speech recognition
• Traffic prediction
• Product recommendations
128 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 6.2 Applications of Machine Learning

• Self-driving cars
• Email spam and malware filtering
• Virtual personal assistant
• Online fraud detection

Except the above-mentioned applications, ML also has wide range of

applications in our day-to-day life. Nowadays, ML is one of the most
awaited technologies for changing the industrial world and it will also
create a great impact on the Industry 4.0. ML has the ability of cor­
recting its own problems by self-learning and finding an appropriate
solution on its own, thus providing all possible solutions for many
problems without the intervention of human being (Figure 6.2).

IIoT – Industrial Internet of Things

The Industrial Internet of Things (IIoT) relates to the web of in­

tegrated devices, such as sensors and robots, and others, like produc­
tion and energy storage, through a network of industrial applications
for computers. This interconnectivity facilitates data storage and
sharing services. The transmitted data are then processed, potentially
 MACHINE LEARNING-BASED IIOT 129

Figure 6.3 Industrial IoT Applications

promoting changes in the industry’s competitiveness and performance,

as well as other economic benefits.
Industrial organizations are increasingly investing in IIoT to en­
hance operational visibility and insights, which will help streamline
manufacturing processes. Taking the complexity out of connecting,
deploying and managing devices within the field are vital to IoT
success (Figure 6.3).
Industrial modernization and therefore the shift to smart manu­
facturing have sparked innovations in automation and robotics and
therefore the Industrial Internet of Things (IIoT). Introducing ar­
tificial intelligence (AI), interoperability and secure connectivity
enable real-time monitoring, control and optimization of processes,
resources and systems – all of which result in greater productivity,
increased safety, and reduced cost.

History of IIoT

The history of the IIoT began with Dick Morley’s invention of the
programmable logic controller (PLC) in 1968, which was used by
General Motors in their automatic transmission manufacturing division.
130 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

These PLCs allowed individual components within the manufacturing

chain to be better controlled. Honeywell and Yokogawa implemented
the world’s first DCSs in 1975, the TDC 2000 and hence the Centum
system, respectively. These DCSs were subsequent steps in facilitating
scalable process management in a facility, with the added benefit of
contingency redundancies by spreading control throughout the whole
system, preventing a single point of failure in a central space.
The current conception of the IIoT arose after the emergence of
cloud technology in 2002, which allows information storage for
historical trends, and so the event of the Open Platform Com-
munication (OPC) Unified Architecture protocol in 2006, which
enabled secure and remote communications between devices, pro­
grams, and data sources without the requirement of human inter­
vention or interface. The OPC server is a software program that
converts the hardware communication protocol employed by a PLC
into the OPC protocol.

Advantages of IIoT

The phrase Industrial Internet of Things, pertaining to the economic

branch of the IoT, is commonly located within the manufacturing
industries. Improved efficiency, analytics, and the development of the
workplace are future advantages of the economic IoT. The potential
for growth by introducing IIoT is expected to reach $15 trillion of
global GDP by 2030. Through IoT apps comes the possibility of
automation. Smart offices use a number of connected devices to
track, control, and handle multiple activities within an organization.
They may also assist staff in managing their schedules in order to be
more efficient, using their time efficiently and successively.
Enabling a $800 billion global market by 2024, IIoT is going to be
a serious catalyst for connected products, services, and solutions.

IIoT in Industry 4.0

Indudtry 4.0 is a phrase coined in Europe. It means as equivalent as

IIoT and refers to the fourth technological revolution. The term is
interchangeable with IIoT and is now recognized globally.
 MACHINE LEARNING-BASED IIOT 131

Figure 6.4 Number of IoT devices in 2015–2025

In order to form a more comprehensive and better integrated

environment for businesses specializing in manufacturing and pro­
viding chain management, IIoT or smart manufacturing blends
physical development and practices with smart digital technologies,
ML, and enormous data (Figure 6.4).
Sensor data, computer communications, and automation systems
are referred to in the IIoT. It is more descriptive than the general
phrase Business 4.0, which covers the industry’s entire digital
revolution.
IIoT use cases are the latest features known in Business 4.0,
supporting the smart industry. In the economic focus of Industry 4.0,
these capabilities are basically what we often call usage cases for IoT
or IIoT:
• track and trace
• structural health monitoring
• remote diagnosis
• control
They’re all typical use cases.

IoT and IIoT

In terms of availability, intelligent and wired computers have an

identical main character. Its general uses are the only distinction
132 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 6.5 Market of IoT devices

between the two. Though IoT is most widely used for consumer use,
IIoT is used for distribution, supply chain control, and management
applications for industrial purposes.
List of the simplest IoT companies is given below:
• Vates
• ScienceSoft (USA and Europe)
• Oxagile (New York, USA)
• Style Lab IoT Software Company (San Francisco, CA)
• HQ Software Industrial IoT Company (USA and Europe)
• PTC (Boston, MA)
• Cisco (San Jose, CA)
• ARM IoT Security Company (Cambridge, Cambs) (Figure 6.5).

The Future of IIoT

As the IIoT progresses, it creates unparalleled opportunities for

companies to simplify processes, maximize consumer service, and ac­
celerate substantial sales gains from the SMB to the global business. In
order to tap into the business advantages of the linked facility, the
long-term IIoT would focus more on predictive maintenance, better
system connectivity, and cheaper access for enterprises of all sizes.

IIoT Use Cases

The IIoT is the network of a multitude of industrial devices linked by

communication technology, culminating in a machine that, as never
 MACHINE LEARNING-BASED IIOT 133

Figure 6.6 Statistics of Power Consumption of IoT devices

before, can track, capture, evaluate, share, and provide new insights.
Both these experiences then help guide strategic decisions for busi­
nesses that are smarter and quicker.
Technological changes in our daily lives are indispensable in this
current era. With the help of emerging new technologies, our lives
will become smarter and smarter, alongwith the industrial world. In
industrial technology, Industry 4.0 is the latest trend in automation
and data sharing. Cyber-physical networks, the IoT and cloud
computing, are included (Figure 6.6).
By implementing the IoT in manufacturing industries, a massive
change on the quality, quantity, and rate of work per hour will take
place. Large numbers of industrial applications are involved in the
process of implementing smart manufacturing. Some of the major
industrial use cases are discussed below:
• Smart metering
• Assets tracking
• Digital twins
• Smart agriculture
• Smart automobile manufacturing
• Smart farming
• Fleet management
• Livestock management
• Connected vehicles
134 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 6.7 Smart Manufacturing Market by Other Regions

What Is Smart Manufacturing?

Generally, we are familiar with the word manufacturing industry, which

means an industry or from which a product is produced or made on a
large scale. Our traditional manufacturing process involves large number
of steps, from choosing the raw materials to packaging the desired
product. All this process needs large mechanical machines and man
power. Later, computer devices are also involved in manufacturing.
Nowadays, this kind of manufacturing process is getting older and
boring. Thus the world is looking forward and is intent toward the
new processes of manufacturing, which perfectly suit the hustle and
bustle of day-to-day life.
“Smart manufacturing” is the solution for all the problems. By
implementing smart manufacturing, most of the problems get solved,
and there will be a massive change in the development of the industry
in all aspects (Figure 6.7).
In future, smart manufacturing will have a great impact on the
manufacturing industries. According to the IDC research, by 2025,
the improvement in operational devices by IoT applications could be
more than $470 billion per year.
By the following industrial use cases, we can easily implement
smart manufacturing in any field and make the industrial world much
smarter.
 MACHINE LEARNING-BASED IIOT 135

Figure 6.8 Asset Tracking and Locations of IoT

Assets Tracking

Tracking assets refer to the way physical assets are tracked. This can
be done either by scanning barcode labels attached to the assets or by
using tags that broadcast their location, such as GPS, RFID, or BLE.
This technique can also be used to monitor an individual or animal
(endangered species).
Assets tracking is a systematic approach that allows the company
to track and manage its assets. The best way to improve the man­
agement of assets is RFID (radio frequency identification) tech­
nology. RFID tags transfer data from an RFID tag to a reader using
electromagnetic fields. Listeb below are the systematic approach or
steps in an industry to set up asset tracking:
• First, determine a reliable person or team to be responsible for
the assets.
• The team would figure out the life cycle of those assets.
• Routine tracking of those assets are mandatory.
• By understanding the functions of the assets in particular opera­
tions, we can implement the automated asset tracking (Figure 6.8).

Industrial Example

In a manufacturing industry, large number of assets are involved in

the process. So it is essential to track those assets for effective
136 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

monitoring, for example, lift truck, which is used to carry the pro­
ducts from the manufacturing unit to the warehouse.
If the truck is fixed with an RFID tag, the RFID reader, placed at
the exit of the unit, will read and send the data to cloud and software
whenever the asset moves out of the manufacturing unit. Similarly,
when the asset enters the warehouse, the reader placed at the en­
trance will read the data and send to the cloud. All the collected data
in a cloud system are analyzed, which helps in improving the process.
Assets tracking helps the industry to improve the maintenance,
accuracy, accountability, and customer services, and also to lower the
administrative expenses. By implementing the assets tracking, we can
improve the planning for further growth of the industry.
Given below is the list of top sevenasset management softwares:.
1. ServiceNowITSM
2. Nifty
3. Monday.com
4. Softchoice IT Asset Management
5. NinjaRMM
6. Freshservice
7. Spiceworks IT Asset Management Software
BlackRock, with managed assets amounting to US$6.52 trillion, was
the largest asset manager globally in 2019.

Digital Twins

A digital twin may be a replica of a physical entity that is living or not

living. It’s the Virtual Information Construct group. The potential or
physical product of production from the micro atomic level to the
macro geometric level is fully described. We can gather any mer­
chandise information from its digital twin as accurately as by phy­
sically inspecting the merchandize to obtain the data. It’s a virtual,
dynamic representation of an object.

Later, the concept was divided into the following types:

▯ =>the digital twin prototype (DTP)
▯ =>the digital twin instance (DTI)
▯ =>the digital twin aggregate (DTA)
 MACHINE LEARNING-BASED IIOT 137

In order to understand a physical product, the DTP consists of

designs, analyses, and processes. There is a DTP before a physical
product exists.
The DTI is the digital twin of each individual instance of the
merchandise.
The DTA is an aggregation of DTIs whose data and knowledge
will be used for physical product questioning, forecasting, and learning.
The digital twin is supposed to be an up-to-date and accurate copy
of the physical object’s properties and states, including shape, position,
gesture, status, and motion. Simply digital twin means the connectivity
between the physical and virtual world. These are connected by the
bridge called data. The information would be obtained by using
thousands of sensors from a physical device. All these data are then
converted to virtual replica. The continuous flow of real-time data
from physical entity to its virtual replica merges the truth and virtual.
Digital twin is a platform that brings all the experts together providing
powerful analysis, insight, and diagnostics. By using digital twins, we
can gain more information about the efficiency, safety, performance,
and other physical entity at an early stage.

Industrial Example

If we want to know the efficiency of a windmill in power production,

we usually make a prototype of that and analyze the things. Then we
create a model of the windmill to examine the details of all criteria.
These are the methods we have been following for a long time. But
nowadays, there is no need for making the prototype and the model.
Digital twins make all the things easier. We can extract any necessary
information of the windmill by its digital twin. Not only the effi­
ciency, but also the safety risks of an aircraft engine can be known by
making a digital twin of it.

Advantages of Digital Twins

Digital twins are boon with the product idea. Their growth will be a
continuous process in product development. Thet also predict errors
in products, serve as a virtual template during assembly, and remain
138 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 6.9 Technologies and Operations of Digital Twins

connected for the entire life cycle. They protect reputation and
prevent financial cost (Figure 6.9).
Further examples of industry applications are:
• Aircraft engines
• Wind turbines
• HVAC control systems
• Locomotives
• Buildings
• Utilities (electric, gas, water, and wastewater networks)
• Large structures, e.g., offshore platforms, offshore vessels, etc.

Smart Metering

Nowadays, the demand for electricity is increasing sharply due to a

number of developmental processes. Maintaining the electricity
properly is becoming the greatest challenge. Electricity leakages,
tamping, and theft are the major issues in maintaining the bills. We
also need manpower to take reading from the electric meters of the
users. It is a time-taking process and sometimes the readings taken by
the EB person may be wrong or inaccurate, which may lead to extra-
charged bills for the users. To eliminate all these circumstances, a
new technology has been emerging with the help of IoT called smart
metering technology. Smart metering is one of the most popular
domains of IoT. Smart metering offers advanced metering infra­
structure (AMI) via smart meters. It is the smart monitoring and
reading of resource usages not only for electricity, but also for water
 MACHINE LEARNING-BASED IIOT 139

and gas. We can get a reading from a machine, rooms, and indoor or
outdoor environments. Once we install the smart meter, the re­
spective meter readings will automatically be sent to the energy
supplier, which means that we will receive accurate, not estimated,
bills. As all smart meters are certified by the Office for Product Safety
& Standards, you can be sure your smart meter is accurate. Smart
meters use a secure national communication network (called the
DCC) to automatically and wirelessly send your actual energy usage
to your supplier. Smart meters send only the reading of the meter to
your supplier; your personal details, such as name, contact, and bank
details, are neither stored nor transmitted to the internet by smart
meters. Smart meters don’t need the internet to communicate.
Instead, they work by using two wireless networks:
1. HAN (home area network)
2. WAN (wide area network)
These networks are very secure and similar to the mobile commu­
nication used to send and receive data. It nearly takes 3 hours for the
installation of smart meters for both gas and electricity. If we wish to
set installation for any one, it will take nearly 60 to 90 minutes and
the gas and power must be switched off for a certain while after the
installation is complete. Smart meters are not only used for providing
accurate bills of resources, but also provide daily usage details of all
the resources with estimated cost; thus, we can reduce and be aware
of our daily usage with the help of in-home display. In-home display
or IHD is a simple electric device which has a touch screen display. It
can be paired with our smart meter and it showcases the needed real-
time usage data of our resources in the display.

Advantages of Smart Metering System

• Eliminates the estimated manual monthly meter readings

• Monitors the real-time electric usage data
• Enables the efficient use of power resources
• Provides responsive data for balancing electric loads while re­
ducing blackouts
• Enables dynamic and accurate pricing
140 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

• Avoids the capital expense of building new power plants

• Helps to optimize and know the profit with existing resources.

Apart from the above-mentioned advantages, smart metering does

many more great things in a smarter way.

Disadvantages of Smart Metering System

• Most smart meters turn dumb whenever we change the supplier.

• Poor network signal prevents the smart meter from working.
• Smart meter stops sending readings due to poor signal, which
leads to delayed bills.
• The terms in new smart monitor are hard to understand by
elders and vulnerable people.
• Although smart meters do not send our personal details, it
poses a risk to security.
• Existing meters are hard to access.

All these mentioned disadvantages stand like a barrier to implement

smart metering systems all over the world.

Smart Farming

As the world population increases drastically, the needs of the

people also increase simultaneously. Industrial revolutions had
played a major role in the transformation of traditional methods to
modern solutions. From the past three industrial revolutions,
we have made the industrial fields to work in a faster and smarter
way by implementing the modern technologies. But in agriculture,
we are not able to achieve the major changes like those in
large industrial factories. Through the IoT, we can achieve the
greatest changes in the agricultural fields by smart farming
(Figure 6.10).
Using various types of sensors and IoT devices for monitoring
and maintaining the data about the crops and environment is
known as smart farming. By this smart farming methods, farmers
 MACHINE LEARNING-BASED IIOT 141

Figure 6.10 Agriculture IoT

will come to know the exact details of their crop at anytime

and any where. They will be well-informed about the progressed
data of their crops. The farmers can get the details like intensity
of sunlight, weather conditions, humidity, temperature, and
moisture of the soil, thereby automating the irrigating system
in their fields.

How Smart Farming Can Be Done?

There are three major steps involved in the smart farming method:
1. Data collection
2. Data analytics
3. Informed action and planning
142 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Data Collection

First, we need to collect the necessary real-time data of our field by

using various types of sensors and other IoT devices. It will help us be
more informed about the field. So the sensors must be placed in and
around the field at some selected locations, thereby we can get the
overall mixed average data of the land. All the collected data are
uploaded into a cloud storage for future use.
All the collected data have to be compared and run through dif­
ferent prediction models to get a clear strategic plan for farmers.
These analyzed data would be very helpful for the farmers to achieve
an optimal growth and yield of the crops. These analyzed data, also
stored in the cloud for future use, would be compared with the next
set of newly collected data for the next season.

Informed Action and Planning

The farmer have received all the essential actionable insights by this
stage. Based on the mentioned factors, he can make a clear plan and
decision and achieve profit and higher productivity level by taking
appropriate actions.
By following the above-mentioned steps with the help of sensors
and other IoT devices, massive revolutionary change can be achieved
in agriculture.

Things to Be Considered before Implementing Smart Farming

1. Hardware
We would like to configure the sensors for your device to create
an IoT solution for agriculture. Our choice will depend on the
data categories we want to collect and thus the purpose of our
solution. The standard of selected sensors is, in any case, crucial
to the success of the product. The accuracy of the data collected
and its reliability will depend on it.
2. The Brain
The core of each smart agriculture solution should be data
analytics. If you can’t add it up,the collected data itself will be of
 MACHINE LEARNING-BASED IIOT 143

Figure 6.11 Estimation of IoT Device Shipments

little help. Thus, you want to have powerful data analytics

capabilities and apply predictive algorithms and ML to support
the collected data in order to get actionable insights.
3. Maintenance
Maintenance of your hardware could be a primary challenge for
IoT products in agriculture, as the sensors we used in your farm
for data collection can easily be damaged. Thus, you want to
make sure the hardware is durable and easy to keep up with.
Otherwise, more often than you’d like, you’ll have to replace
your sensors.
4. Mobility
To be used within the field, smart farming applications should
be tailored. A business owner or farm manager should be able to
access the data via a sensitive phone or PC on site or remotely.
In addition, each connected device should be independent and
have adequate wireless range to communicate with the opposite
devices and transfer data to the central server.
5. Infrastructure
To check that your application for smart farming is doing well.
We’ll be sure it can handle the load of data. A solid internal
infrastructure is what you would need. In addition, the internal
networks must be stable. If you have not protected your device,
144 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

then stealing your records, or even taking charge of your au­

tonomous tractors, would be very easy for others (Figure 6.11).

Advantages of Smart Farming:

• Increased production
• Water conservation
• Real-time data and production insight
• Lowered operation costs
• Increased quality of production
• Accurate farm and field evaluation
• Improved livestock farming
• Reduced environmental footprint

Livestock Management Using IoT

In crop production, the use of livestock and its sub-product manure is

significant. Livestock is an energy source that provides animal power
for draught, while manure improves soil structure and fertility, as well
as water retention. Enhancing energy and nutrient cycling is en­
vironmentally friendly for both uses. Farmers can monitor their live­
stock very closely by using IoT in the field of livestock management.
Wearables play the vital role in tracking each and every movement
of animal. Wearables with connected sensor can track every real-time
data and send it to the cloud platform. By using this technology,
farmers can improve the standard of livestock breeding and also
improve their financial economy in a good manner.

Monitoring the Health of Cattle

Farmers in India suffer a loss of nearly ₹20,000 crore per year due to
the adverse effect of animal illness that leads to death. IoT gives
farmers a clear solution to those crises and improves better health
care in livestock management. IoT offers the farmers a low-cost and
low-bandwidth sensor-connected wearables to track the health of
cattle. The sensors in the wearables can monitor the blood pressure,
heart rate, respiratory rate, digestion, temperature, and other vitals
that alert a farmer at the first sign of illness (Figure 6.12).
 MACHINE LEARNING-BASED IIOT 145

Figure 6.12 IoT in Samrt Farming

Generally, cattle may easily affect disease when they are in a

herd. It may lead to enormous loss by spreading to every individual.
IoT can easily detect those diseases in a herd and allow the farmer
to find out the cattle with chance of disease and isolate it to avoid
spreading. IoT also detects the illness of the individual cattle due to
feed issue or other parameters and alerts farmers before it becomes
critical. For example, a cow’s temperature could rise if it is affected
by any disease; a small temperature rise rather than normal is en­
ough for IoT to alert the farmer before there is a change in the
behavior of the cattle.

Monitoring Reproductive Cycle and Calving

IoT can also be very useful in the reproductive cycle. Before calving,
there must be slight rise in the temperature of cow before 8 hours.
IoT can alert the farmers about the calving very before, thus in­
creasing the efficiency and productivity of cattle farming. IoT can
also inform the farmer when the cow needs to breed and when it is
ready for artificial insemination etc. By having these data, a farmer
can easily avoid various losses in a herd. If the cow is in a farm during
the time of calving, IoT informs the farmer whether she started
calving and sends the location of the cow, thus protecting the new
born calf from other problems.
Location tracking of cattle in a herd gives enormous benefits to the
farmer. Even it is an important feature in health and reproductive
146 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

cycle of livestock management. The location of an individual cattle

and herd can be traced using connected sensors fixed in the wearables
which is placed around the neck of the cattle. By tracking the lo­
cation of herd farmers can come to know, where their cattle are
grazed and they also know the area of pastureland.
Farmers can establish and optimize grazing patterns with the data
provided from the tracking of the movement of each animal, as well as
the migration of the herd. It also allows the farmer to find the sick
animal and allow it to be more handled and to isolate the herd that goes
into heat from the human form. Overall, it makes it easy for farmers to
handle and control their cattle that graze on multiple acres of land.

Fine-Tuning Feeding

Fine-tune feeding is the process of giving a perfect diet to the cattle;

it contains all the essential nutrients necessary for the proper growth
and lactation of the cattle. By tracking the location of the grazing
area, we can easily calculate how frequently they feed. We also come
to know about the grazing pattern of the herd, thereby allowing the
farmer to feed his herd in an appropriate proportion. By monitoring
the specific behavior of the herd such as grazing, socializing, lying
down, and chewing the cud, farmer can avoid wasting food or
overfeeding the cattle, knowing the frequency of those activities.

Maximizing Milking

IoT also gives dairy farmers a fantastic operation. Robots will literally
increase the workload of farmers by reminding farmers when the cow
needs to be milked by installing the IoT on a farm, which will lead to
increased milking sessions. It can also control any person on a farm’s
milking speed, quality and quantity of milk produced, quantity of
feed taken, and number of steps taken every day. From that data, by
deciding the required complement for each person and paying more
attention to the cow that produces high quality and quantity of milk,
farmers can improve milking. As a result, lactation frequently raises
the farm’s financial profits at the same time.
Ultimately, farmers are the one to take care of their herds, but IoT
can optimise and simplify many major processes in herd farming. IoT
 MACHINE LEARNING-BASED IIOT 147

easily monitor the location, fertility, health condition, temperature,

digestion, movements, grazing patterns, lactation and calving cycle of
each and every individual in the farm. By knowing all the data farmers
can ultimately increase the productivity, efficiency and revenue of their
farm in a progressive manner.
Undoubtedly, apart from agriculture herd, farming has also con­
tributed a huge role in satisfying the needs of people. Implementing
the IoT in livestock management will be very effective for the im­
provement in the life cycle of cattle, as well as human being.

Connected Vehicles

Connected vehicles are one of the most promising products of IoT.

Connected vehicles are simply a revolutionary thing at present. It is an
upgrade version of ordinary traditional type of vehicle. Connected ve­
hicles can communicate bidirectionally with other systems, both outside
and inside the vehicle. Connected vehicles can share the internet access
and data with other vehicles for the efficient driving and safety purpose.
By IoT vehicle, we can communicate with each other and understand
each and every movement of other vehicles (Figure 6.13).
Connected vehicles can share the data all around the vehicle
(360°). They collect data from nearby or passing vehicle and respond
with required action automatically. Prospectively connected vehicles
are able to connect using “dedicated short-range communication

Figure 6.13 Smart Tracking

148 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

(DSRC)” radios. They operate in FCC-granted 5.9-GHz band.

They have very low latency. Connected vehicles were introduced in
early 1996 for the safety purposes, but now they are used for various
purposes. Variety of connectivity and usage are enhanced in every
upgrade new versions.
Following are the five different types of connectivity available for
different purposes:
• V2V – vehicle to vehicle
• V2I – vehicle to infrastructure
• V2C – vehicle to cloud system
• V2P – vehicle to pedestrian
• V2X – vehicle to everything
Vehicle to vehicle technology offers exchanging the information
(data) between the vehicles around it. It shares the information about
the current real-time position and speed of the vehicle with accuracy.
It is used for the purpose of safety and it avoids unnecessary traffic
convention.
Vehicle to infrastructure technology offers the communication of
vehicle with the infrastructure of the vehicle. It shares the informa­
tion generated by the vehicle about the infrastructure to the driver.
This V2I technology is used to communicate the information about
safety, mobility, and environment information, such as weather
conditions, wind speed, and temperature to the driver.
Vehicle to cloud technology uses to share the information about
the vehicle to the cloud system. This allows the vehicle to use in­
formation from other cloud system of other industries, smart home, etc.
Vehicle to pedestrian technology provides a connection between
vehicle to the infrastructure, environment, and other vehicle. This
technology generates information about the environment and shares
the generated information to other vehicles for the safety and mo­
bility. Thus this technology provides a communication between ve­
hicle and pedestrians on the road.
Vehicle to everything technology is an ultimate goal of con­
nectivity of connected vehicles. It easily enhances communication
with infrastructure, environment, pedestrian, and other vehicles. This
technology also offers the connection between cars, rails, ships, and
airplane. Thus, it does everything in a well precise manner.
 MACHINE LEARNING-BASED IIOT 149

Figure 6.14 Connected Vehicle Prediction in US

Connected vehicles primarily use 3G/4G data connection for sharing

the information and communications between vehicle and mobile de­
vices, smart assistants, and websites all over the world. It is estimated that
over 28.5 million units of connected cars were sold worldwide in 2019
(Figure 6.14).
• Improves safety on road
• Avoids traffic congestion
• Is Eco friendly
• Provides Urban and suburban mobility
• Avoids accidents
• Provides Infotainment
• Enables enhanced communication
• Emergency calling

Out of these above-mentioned points, connected vehicles have

many advantages on real-time applications.
Connected vehicles are going to make a great change on normal
roads and traffic. Our new generation is going to experience new
things on roads. They connects everything about the vehicle to our
hand. We can get all the real-time data about the vehicle, such as
speed, quantity of the fuel in the fuel tank, air pressure in the wheel,
environmental changes, etc.
If the vehicle is supposed to be met with an accident, the cloud
system automatically sents a emergency call to a common center
150 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

with the current location. This helps the driver to get quick
emergency help from nearby hospital or health center, thereby in­
creasing the chance of life of road users. If any accident or repair
takes place to the vehicle on the lane of driving, the vehicle sends
the appropriate information to other vehicles riding on the same
lane before it comes closer. Thus it helps other vehicles to change
other lane of the road.
Connected vehicles also communicate with signals on the junction
and make appropriate actions with time. If a physically challenged
person wants to cross the road, the information will be sent to the
vehicle crossing the road so that it can be stopped and the person can
cross the crossing. The same information is also transferred to all
vehicles moving behind the first one thus avoiding the unwanted
accident and traffic. Connected vehicles will improve the standard of
driving and the safety factor from all sides. They contribute more to
the goodness of the society.

Conclusion

The Internet of things (IoT) is awaiting to give a new revolutionary

way of lifestyle to us. Many of us get panic or wondered, when we hear
the words such as Internet of Things, artificial intelligence, machine
learning, data analysis, block chain, cloud computing etc. We think
that these technologies are very complicated, but in reality, they are the
simplest things that are going to make ultimate changes in our day-to-
day life, if only we are ready to shake our hands with them. It is
expected that in the year of 2020, nearly over 50 million devices are
going to be connected with IoT in worldwide. Industry 4.0 is one of
the ultimate goals of this decade and IoT playing a vital role in the
implementation of industry 4.0. IoT has many industrial applications,
as we discussed above. They are the key way to the industrial re­
volution. All the routine ways of traditional manufacturing are digi­
talized by smart manufacturing. The standard of all manufacturing
industries will reach the peak. Thus smart metering, smart agriculture,
fleet management, logistics maintenance, digital twins, cloud com­
puting, assets tracking, connected vehicles, and other number of new
things through IoT pave a new version of lifestyle for human being
and machines. IoT has some drawbacks and disadvantages, but when
 MACHINE LEARNING-BASED IIOT 151

we compare with benefits, it looks like a piece of dust in the sky. IoT
and IIoT are never going to end simply; they are emerging as new and
upgrading every day by new technologies with the help of various types
of sensors, tracking devices, RFID tags, computing software, etc.
So make a bond with IoT and make everything you want within
finger tips. IoT is never be a period, it will always remain as a comma.

References
[1] Ismail, Yasser, “Introductory Chapter: Internet of Things (IoT)
Importance and Its Applications,” in Yasser Ismail (ed.), Internet of
Things (IoT) for Automated and Smart Applications. IntechOpen.com,
November 27, 2019, doi: 10.5772/intechopen.90022.
[2] Tracy, Phillip, “The Top 5 Industrial IoT Use Cases,” 5 minute read,
IBM.com, April 19, 2017, https://www.ibm.com/blogs/internet-of-
things/top-5-industrial-iot-use-cases
[3] DdGubbia, J., Buyyab, R., Marusic, S., and Palaniswami, M., “Internet
of Things (IoT): A vision, architectural elements, and future directions,”
Future Generation Computer Systems, vol. 29, September 2013.
[4] EeFlexeye, “Lord of the Things: Why Identity, Visibility and
Intelligence are the Key to Unlocking the Value of IoT,” 2014, https://
coe.flexeyetech.com/.
[5] Xu, L., He, W., and Li, S., “Internet of Things in Industries: A
Survey,” IEEE Transactions on Industrial Informatics, vol. 10, no. 4,
November 2014, pp. 2233–2243.
[6] IEEE Internet of Things Journal, http://standards.ieee.org/innovate/iot/
[7] Gilchrist, Alasdair, Professional and Applied Computing, eBook Packages,
Berkeley, CA: Apress, 2016, https://doi.org/10.1007/978-1-4842-2047-4.
[8] Elangovan, Uthayan, Smart Automation to Smart Manufacturing: Industrial
Internet of Things, Manufacturing and Processes Collection, New York:
Momentum Press, 2019, ISBN:1949449262,9781949449266.
 7
EMPLOYEE TURNOVER
PREDICTION USING SINGLE
VOTING MODEL
R. VALARMATHI1, M. UMADEVI2,
AND T. SHEELA3
1
Department of Computer Science
and Engineering, Sri Sairam
Engineering College, Chennai
2
Department of Computer Science
and Engineering, SRM Institute of
Science and Technology, Chennai
3
Department of Information
Technology, Sri Sairam Engineering
College, Chennai

Contents

Introduction 154
Role of Supervisors 154
Role of HR Department 154
Related Work 155
System Implementation 156
Data Preprocessing 157
Materials and Methods Used 157
Logistic Regression 157
Naive Bayes Classification 161
Voting Classifier 163
Results and Discussions 164
Conclusion 173
References 173

DOI: 10.1201/9781003119838-7 153

154 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Introduction

Switching of employees working in a company to another company is

called attrition. Employees change company due to lack of monetary
benefits, job pressure, work–life balance, lack of promotional op­
portunities, etc. The employee attrition rate differs with age group,
gender, marital status, department of work, and job role. When the
attrition rate goes high, companies and employers face various pro­
blems. Selecting the proper employees with matching educational
qualification and work expertise from market is a challenge.
Recruiting freshers or replacements of existing employees need
training and time to get the expertise in the domain, tools used,
documents to be prepared, etc. Client service gets affected during the
transition period. Service quality goes down due to the newcomers
trying their best in the new work areas. Due to this, some clients offer
the work assigned to some other companies with lesser attrition rate.

Role of Supervisors

Role of supervisors in boosting the employee morale plays the main

part. Supervisors have to recommend the right candidates for pro­
motion and onsite works and rate the employees correctly in annual
appraisals. New employees learn the work ethics and get trainings
from the supervisor mainly. Depending on employee preferences and
personal commitments, supervisors can recommend or decide the
department transfers, work location transfers, and working hour
change.

Role of HR Department

The Human Resources (HR) department works as a brain center of

any company. HR department has to devise right company policies
and take decisions so that the employees get the work–life balance
and correct annual remunerations. HR plays the main role in re­
cruiting right candidates, when vacancies arise, so that the super­
visors’ burden and time spent on training get reduced. Companies’
annual leave policies, working hours, and facilities, like transport,
health insurance, and provident fund, are decided by the HR
 EMPLOYEE TURNOVER PREDICTION 155

department only. Also, recreational programs and employees’ en­

gagement facilities are conducted by HR department, which decides
the attrition rate in a company.

Related Work

The proposed work chooses the top 10 features using CART ana­
lysis. The imbalanced dataset is split into 60% and 40%, thereby
reducing the bias and improving the sensitivity and specificity of the
ensemble model built [1]. The authors have addressed the employee
turnover issue with six different machine learning (ML) algorithms,
namely, SVM (RBF Kernel), XGBoost, logistic regression, Naive
Bayesian, Random Forest with Depth controlled, and LDA and
KNN. As predicted, the algorithms perform with different measures,
such as AUC, running time, and memory utilization. The experi­
mental results shows that XGBoost outperforms all other algorithms
in terms of accuracy and memory utilization. Logistic regression took
a minimum of 52 seconds to run the prediction algorithm [2]. The
author demonstrated the prediction of employee turnover by ana­
lyzing varying size (small, medium, and large) and complexity of the
human resource datasets with 10 different supervised classifiers. The
author has provided guidelines for using the statistical methods and
recommended gradient boosting for analyzing large datasets, as it
takes less time for data preprocessing and ranks the features auto­
matically and reliably [3]. Employee attrition dataset created by IBM
Watson predicts the employee turnover [4].
The researchers proposed a weighted quadratic random forest
method. The author followed a two-step process. In the first step, the
dimensions are reduced and important features are selected. In the
second step, F-measure is applied on the selected features for each
decision tree. Two most important features that influence employee
turnover identified by the author are monthly income and overtime
[5]. The algorithm lacks in ranking the best features [6–9] in these
studies. The researchers studied the impact of pay satisfaction and
school achievement on principals and found a negative correlation
with principal turnover [10].
The authors investigated the hospital nurse turnover and found job
satisfaction, age, and stress to be the major factors that influence the
156 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

nurse turnover [11]. Rotation forest approach helps in improving

individual classifier’s accuracy and diversity. Diversity is achieved
through features axes and accuracy through principal components [12].
The researchers [13] stated that rotation forest is a better and accurate
approach for the real valued attributes compared to SVM, NN, and
tree-based ensembles. They evaluated the performance on three dif­
ferent datasets containing about 200 classification problems based on
four parameters, such as errors, ROC (receiver operating character­
istic), negative likelihood, and balanced error. The Naive Bayes model
[14] was employed on HR dataset to predict the attrition. The authors
[15] used dynamic partite graph to model the job records.

System Implementation

The system is fragmented into a number of segments for better un­

derstanding. The integral segments within the systems are as follows:
• Data preprocessing
• Predicting the employee turnover using logistic regression and
Naive Bayes classification using grid search
• Combining the logistic regression and Naive Bayes classifica­
tion as a single voting model using grid Search to predict the
employee turnover by averaging the two models
The workflow is depicted in Figure 7.1.

Logistic
Regression
Data Employee
Employee Voting
Preprocessing turnover
Dataset Classifie
prediction
Naïve Bayes

Figure 7.1 System Flow Diagram

 EMPLOYEE TURNOVER PREDICTION 157

Data Preprocessing

The datasets are preprocessed to improve the accuracy of the model.

The HR dataset and employee satisfaction level are merged. Label
encoding is applied to convert names to numeric.

Materials and Methods Used

Logistic Regression Logistic regression is a binary classification model

to find the probability of a certain event existing, such as tail/head,
0/1, or unhealthy/healthy. The binary classification model is a pro­
blem with two class values, such as True/False and Buy/Sell. It is a
statistical model to predict the categorical dependent variable by
finding the value for it using a logit function. In 1972, Nelder and
Wedderburn created this method, which is the part of generalized
linear model (GLM).
To predict the continuous valued quantities, like price of a house,
linear regression model which is a linear function of input values (like
the square feet of the house) is used. To predict the discrete valued
quantities like predicting the pixel intensities representing a “0” digit
or a “1” digit, logistic regression model which is a logistic function of
input values is used. So it is a simple classification algorithm for
learning to make such decisions. It is one of the supervised ML
models. In supervised learning, the response variable (B) is re­
presented in terms of input variable (A) that maps the input to the
output.
B = f(A )

When the response variable (B) is of a categorical type like “green”/

“red,” “good”/“bad,” then classification model is used, whereas when
the response variable (B) is continuously valued like “dollars” or
“weight,” then linear regression model is used.
This model is used to calculate a value for the response variable
with only values like “married”/“unmarried” or “male”/“female” by
using the given independent variables. So, it is one type of linear
regression model with categorical value for the predictor, which uses
log of odds as dependent variable.
158 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Different Categories of Logistic Regression

1. Binary Logistic Regression

When the dependent variable (target) has merely two probable out­
comes, like “happy” or “sad,” then binary logistic regression is used.
2. Multinomial Logistic Regression
When the dependent variable (target) has three or more possible re­
sponses without ordering like the color that is preferred more, “Red” or
“Blue” or “Green,” then multinomial logistic regression is used.
3. Ordinal Logistic Regression
When the dependent variable (target) has three or more possible
responses with ordering, like the grade value of the marks be (“A,”
“A+,” “B,” “B+,” “F”), then ordinal logistic regression is used.

Deriving Logit Function The response variable Y which is a func­

tion of predictor/independent variable X is represented as follows:
Y= + (X ) (7.1)

where, α is a constant term which will be the probability of the event

happening when no other factors are considered. Equation (7.1) is a
link function that represents the binary outcome Yes or No. We can
create a dummy variable P to indicate if the observation is a Yes or a
No (i.e., P = 1 / P = 0). It is done as follows using the two things:
i. The probability of getting Yes (P)
ii. The probability of getting No (1 – P), where 0≥P≤1.
The probability value should be positive and to make it positive, we
find the exponent value of Equation (7.1) as:

P=e + (X) (7.2)

To confine P to 0≥P≤1, divide P by the value which is at least one
value greater than P:

e + (X)
P= + (X)
(7.3)
1+e
 EMPLOYEE TURNOVER PREDICTION 159

By Equation (7.1), Y = α + β (X) and by combining Equations (7.1),

(7.2), and (7.3), we can write P as follows:
ey
P= (7.4)
1 + ey

where, P is the Logit function which gives the probability of getting

“Yes.” Now the probability of getting “No” 1 – P is written as follows:
ey
1 P=1 (7.5)
1 + ey
By dividing Equations (7.4) and (7.5), we can get
ey
P P 1 + ey
= = ey
(7.6)
Q 1=P 1 1 + ey

By simplifying Equation (7.6), we can obtain

P P
= = eY (7.7)
Q 1 P

Taking logarithmic on Equation (7.7), we can obtain

P
Log(e Y ) = Log ( 1 P)
P
(7.8)
Y = Log ( 1 P)

By substituting (7.1) in (7.8), we will get

P
Log = + (X). (7.9)
1 P
P
where 1 P is the odd ratio.
It is a link function used in logistic regression. The logarithmic
value on the dependent variable is used to mode a non-linear asso­
ciation. Whenever the logarithmic value of the odd ratio is positive,
the probability of success is always more than 50%.

Performance of Logistic Regression Model The various performance

validation parameters of logistic regression model are null deviance
and residual deviance, confusion matrix, and AUC-ROC curve.
160 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

1. Null Deviance and Residual Deviance

Null deviance is the measure of the outcome predicted by the model
without considering any input, whereas residual deviance is the mea­
sure of the outcome predicted by the model in consideration with the
independent variable. Lower the value, better the performance.
2. Confusion Matrix
It is the representation of actual class outcome versus predicted class
outcome in table form, shown below, which is used to measure other
metrics, like accuracy, precision, recall, and F-Measure.
A
Specificity/TN Rate =
A+B
B
FP Rate =
A+B
D
Sensitivity/TP Rate =
C+D
c
FN Rate =
c+d

Actual class outcome Predicted class outcome

Yes No
Yes True Positive – TP (D) False Negative – FN (C)
No False Positive – FP (B) True Negative – TN (A)

TP and TN represent the data items that are correctly predicted,

whereas FP and FN represent the data items that are not correctly
predicted, i.e., the actual class contradicts with the predicted class.
True Positives (TP) – It shows the acceptable predicted values, i.e.,
actual class outcome and predicted class outcome have the same value.
True Negatives (TN) – It shows that predicted values are invalid, i.e.,
actual class outcome and predicted class outcome have different values.
False Positives (FP) and False Negatives (FN) – It shows that
predicted values are invalid, i.e., actual class outcome and predicted
class outcome have contradicted values.
Accuracy – It is a percentage over the total number of observations
that are predicted correctly to the total number of observations.
Higher the accuracy, better the model. If the model has symmetric
 EMPLOYEE TURNOVER PREDICTION 161

datasets consisting of equal number of false positive values and false

negative values, then accuracy is considered as a great measure of
performance. It is written in Equation (7.10) as:
A+D
Accuracy = (7.10)
A+B+C+D
Precision – It is the percentage of the total number of correctly
predicted observations to the total number of positively predicted ob­
servations. Whenever the precision is high, the false positive rate is low.
Precision = D/B + D
Recall (Sensitivity) – It is the percentage of the total number
of positively predicted observations to all the observations in actual
class – yes.
Recall = D/C + D
F1 Score – The trade-off between precision and recall is re­
presented as F-Score in Equation (7.11). Whenever there is an uneven
class distribution, this score is usually more practical than accuracy.

F1 Score = 2 (Recall Precision)/(Recall + Precision) (7.11)

3. ROC Curve (Receiver Operating Characteristic Curve)

It gives the tradeoff between the true positive and false positive. As
the model is concerned about the success rate, assume that the P
value is greater than 0.5. ROC curve represents all possible values of
P which is higher than 0.5.

Naive Bayes Classification Naive Bayes classification is a an ML al­

gorithm based on Bayes Theorem for predictive analysis. It is a
classification model based on conditional probability. The condi­
tional probability is a measure of the probability of an occurrence of
an event A after an occurrence of another event B. It calculates the
probabilities of discrete words like spam or not, married or un­
married, and so on. These Probabilities are also called likelihoods. It
consists of the following two parts:
1. Naive
2. Bayes
162 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Bayes Theorem
It is a theorem for calculating conditional probability of an event.
Consider events A and B. Probability of an event A to occur is P(A)
and probability of an event B to occur is P(B). Bayes theorem shows
the relationship between the probability of an event A before oc­
curring in the event B. Probability of an event A after getting the
event B, P(A/B), is as follows:
B
A P ( A ). P (A )
P = , (7.12)
B P (B )

where A is hypothesis, B is evidence, P(A/B) is conditional prob­

ability of A given B which tells that how often A occurs given that B
occurs. It is posterior probability, i.e., probability of the event after
the evidence has occurred. P(B/A) is conditional probability of B
given A which tells that how often B occurs given that A occurs. P
(A) is prior probability, i.e., probability of the event before evidence
has occurred.

Dataset
The dataset used for the classification consists of:
• Set of independent variables called input features
• Dependent/output variable
Each attributes in the input features are mutually independent. There
is always a strong independence among the features in the feature
matrix. Hence, it is called Naive. Also each attributes in the input
features have an equal contribution in predicting the output class.

Categories of Naive Bayes Classification

Multinomial Naive Bayes This is the multivariate event model to
categorize a text/article into games, news, technical, cine field, etc.

Bernoulli Naive Bayes When the dependent variable (target) has

merely two probable outcomes like “Happy” or “sad,” the classifica­
tion model used is Bernoulli Naive Bayes. It is the binomial model as
the input independent variables take only binary values “Yes”
or “No.”
 EMPLOYEE TURNOVER PREDICTION 163

Gaussian Naive Bayes When the independent variables have con­

tinuous values rather than discrete values, Gaussian Naive Bayes
classification can be used by sampling them using Gaussian dis­
tribution or normal distribution.

Benefits of Using Naive Bayes

• Straightforward method and high-speed classification model

• Suitable for multi-class classification
• Effectively works with outsized datasets
• Performs better than the other models, like logistic regression
• A smaller amount of training data is required
• Performance is better for categorical data than continuous data
• Low computation cost
• Most suitable for text classification problems

Areas of Applications

• Superior than other algorithms and so it is used in all kinds of

binary classifications
• Naive Bayes along with collaborative filtering are used in re­
commended systems
• It is also used in disease prediction based on health parameters
• Found its application in face recognition
• Naive Bayes is used in prediction of weather reports based on
atmospheric conditions (temp, wind, clouds, humidity, etc.)
• Can be used in real-time prediction
• Can be used in multiclass prediction

Voting Classifier The predictions of multiple classifiers are combined.

Multiple classifiers are used to train and validate the dataset. Finally,
the predictions are taken from the majority voting. Hard voting and
soft voting are the different categories of majority voting.

Hard Voting In hard voting, the final predicted class label will be
the label predicted by most of the classifiers. For example, assume
there are three different classifiers and two different classes, Dog and
Cat. If Dog is predicted by two classifiers and Cat is predicted by one
classifier, final prediction Dog will be chosen.
164 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Soft Voting In soft voting, the final predicted class label will be the
probabilities of all the classifiers chosen for the problem. For ex­
ample, take three classifiers M1, M2 and M3 and two classes Dog
and Cat. Assume

M1 >{0.4, 0.7}
M2 >{0.2, 0.8}
M3 >{0.7, 0.3}
Probability of Class Dog = 0.33 0.4 + 0.33 0.2 + 0.33 0.7
= 0.429
Probability of Class Cat = 0.33 0.7 + 0.33 0.8 + 0.33 0.3
= 0.594

The probabilities predicted by the combination of classifier are

[42.9%, 59.4%]. The class Cat which is having prediction probability
of 59.4% will be chosen. The above is an example of soft voting
classifier with equal weights.

Results and Discussions

All the three models are executed in Python 3.0 and scikit-learn
Libraries.

In [1]:
import numpy as np
import pandas as pd
emp_data=pd.read_csv(r‘D:\Research\Employee attrition\hr_data.csv’)

In [2]:
emp_data.head()

Out [2]:
EMPLOYEE_ NUMBER_ AVERAGE_ TIME_SPEND_ WORK_ LEFT PROMOTION_ DEPARTMENT SALARY
ID PROJECT MONTLY_ COMPANY ACCIDENT LAST_
HOURS 5YEARS
0 1003 2 157 3 0 1 0 sales low
1 1005 5 262 6 0 1 0 sales medium
(Continued)
 EMPLOYEE TURNOVER PREDICTION 165

EMPLOYEE_ NUMBER_ AVERAGE_ TIME_SPEND_ WORK_ LEFT PROMOTION_ DEPARTMENT SALARY

ID PROJECT MONTLY_ COMPANY ACCIDENT LAST_
HOURS 5YEARS
2 1486 7 272 4 0 1 0 sales medium
3 1038 5 223 5 0 1 0 sales low
4 1057 2 159 3 0 1 0 sales low

In [3]:
emp_data.shape

Out [3]:
(14999, 9)

In [4]:
emp_data.size

Out [4]:
134991

In [5]:
sat_data=pd.read_excel(r‘D:\Research\Employee attrition\
employee_satisfaction_evaluation.xlsx’)

In [6]:
sat_data.head()

Out [6]:
EMPLOYEE # SATISFACTION_LEVEL LAST_EVALUATION
0 1003 0.38 0.53
1 1005 0.80 0.86
2 1486 0.11 0.88
3 1038 0.72 0.87
4 1057 0.37 0.52

In [7]:
empatt_data= emp_data.set_index(‘employee_id’).join(sat_data.set_index(‘EMPLOYEE #’))
166 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

In [8]:
empatt_data.describe()

Out [8]:
EMPLOYEE_ID NUMBER_ AVERAGE_ TIME_SPEND_ WORK_ LEFT PROMOTION_ SATISFACTION_ LAST_
PROJECT MONTLY_ COMPANY ACCIDENT LAST_5YEARS LEVEL EVALUATION
HOURS

count 14999.0000 14,999.000 14,999.0000 14999.0000 14999.0000 14999.0000 14999.0000 14972.0000 14972.0000
mean 45424.6275 3.803054 201.050337 3.498233 0.144610 0.238083 0.021268 0.612830 0.716125
std 25915.9001 1.232592 49.943099 1.460136 0.351719 0.425924 0.144281 0.248714 0.171138
min 1003.00000 2.000000 96.000000 2.000000 0.000000 0.000000 0.000000 0.090000 0.360000
25% 22872.5000 3.000000 156.000000 3.000000 0.000000 0.000000 0.000000 0.440000 0.560000
50% 45448.00 4.000000 200.000000 3.000000 0.000000 0.000000 0.000000 0.640000 0.720000
75% 67480.500000 5.000000 245.000000 4.000000 0.000000 0.000000 0.000000 0.820000 0.870000
max 99815.000000 7.000000 310.000000 10.000000 1.000000 1.000000 1.000000 1.000000 1.000000

In [9]:
from sklearn.preprocessing import LabelEncoder
lab=LabelEncoder()
n=lab.fit_transform(empatt_data[‘salary’])

In [10]:
empatt_data[‘salary_num’]=n
empatt_data.drop([‘salary’],axis=1,inplace=True)
z=lab.fit_transform(empatt_data[‘department’])
empatt_data.drop([‘department’],axis=1,inplace=True)
X=empatt_data.drop([‘left’],axis=1)

In [11]:
X.head()

Out [11]:
NUMBER_ AVERAGE_ TIME_ WORK_ PROMOTION_ SATISFACTION_ LAST_ SALARY_NUM
PROJECT MONTLY_ SPEND_ ACCIDENT LAST_ LEVEL EVALUATION
HOURS COMPANY 5YEARS
0 2 157 3 0 0 0.38 0.53 1
1 5 262 6 0 0 0.80 0.86 2
2 7 272 4 0 0 0.11 0.88 2
3 5 223 5 0 0 0.72 0.87 1
4 2 159 3 0 0 0.37 0.52 1

In [12]:
y=empatt_data[‘left’]
 EMPLOYEE TURNOVER PREDICTION 167

In [13]:
y.head()

Out [13]:
01
11
21
31
41

Name: left, dtype: int64

In [14]:
#import the libraries
import matplotlib.pyplot as plt
%matplotlib inline
import seaborn as sns
from sklearn.preprocessing import StandardScaler
# Split data into Training and Testing Dataset
from sklearn.model_selection import train_test_split
# Data modeling
from sklearn.metrics import confusion_matrix,accuracy_score,roc_curve,classification_report
from sklearn.metrics import roc_auc_score
from sklearn.linear_model import LogisticRegression
from sklearn.naive_bayes import GaussianNB

In [15]:
from sklearn.model_selection import train_test_split
X_train_data, X_test_data ,y_train_data ,y_test_data =train_test_split (X, y, test_size=0.10,
random_state=0)

Logistic regression has different hyper parameters solver, penalty

and C. Different solvers result in different types of performance:
penalty helps in regularization and C helps in controling the penalty
strength. These parameters help to optimize logistic regression.
Logistic regression can be evaluated with different combination of
parameters and the best hyper parameters can be found. Grid search
is an exhaustive search strategy that considers all parameter combi­
nations while sampling in the search space.
168 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

In [16]:
#Model 1-Logistic Regression
from sklearn.model_selection import GridSearchCV
from sklearn.linear_model import LogisticRegression
log_reg = LogisticRegression()
params = {‘C’: [1e-4, 1e-3, 1e-2, 1e-1, 0.5, 1., 5., 10., 15., 20., 25.]}
#Grid Searchwith 10-fold cross validation
log_model1 = GridSearchCV( log_reg , param_grid=params,cv=10, n_jobs=-1)
log_model1.fit(X_test_data, y_test_data)
#The Best Hyper parameters of Logistic Regression
print(“Best Hyper Parameters:\n”, log_model1.best_params_)
#Prediction
lr_predict= log_model1.predict(X_test_data)
#import metrics
from sklearn import metrics
#Accuracy
lr_acc_score = accuracy_score(y_test_data, lr_predict)
print(“Accuracy:”,metrics.accuracy_score(lr_predict, y_test_data))
#Confusion Matrix
print(“Confusion Matrix:\n”,metrics.confusion_matrix (lr_predict, y_test_data))
#Classification Report
print(classification_report (y_test_data,lr_predict))
#ROC Score
print(“ROC_AUC_SCORE”)
roc_auc_score(y_test,lr_predict*100)

Output:
Best Hyper Parameters:
{‘C’: 25.0}
Accuracy: 0.77
Confusion Matrix:
[[1052 256]
[ 89 103]]
precision recall f1-score support
0 0.80 0.92 0.86 1141
1 0.54 0.29 0.37 359
accuracy 0.77 1500
macro avg 0.67 0.60 0.62 1500
weighted avg 0.74 0.77 0.74 1500
ROC_AUC_SCORE
Out[16]:
0.6044531625730252

The parameter var_smoothing helps to tune GaussianNB

 EMPLOYEE TURNOVER PREDICTION 169

In [17]:
#Model 2-Naive Bayes Classifier
from sklearn.naive_bayes import GaussianNB
nb_c = GaussianNB()
nb_params = { ‘var_smoothing’:np.logspace(0,1,num=100)}
#Grid Search with 10-fold cross validation
nbmodel = GridSearchCV(nb_c, param_grid=nb_params,cv=10, n_jobs=-1)
#Learning
nbmodel.fit(X_test_data ,y_test_data)
#The Best Hyper parameters are
print(“Best Hyper Parameters:\n”,nbmodel.best_params_)
#Prediction
nbpred=nbmodel.predict(X_test_data)
#Accuracy
nb_acc_score = accuracy_score(y_test_data, nbpred)
print(“Accuracy:”,metrics.accuracy_score(nbpred, y_test_data))
#Confusion Matrix
print(“Confusion Matrix:\n”,metrics.confusion_matrix (nbpred, y_test_data))
#Classification Report
print(classification_report(y_test_data, nbpred))
#ROC Score
print(“ROC_AUC_SCORE”)
roc_auc_score(y_test_data, nbpred*100)

Output:
Best Hyper Parameters:
{‘var_smoothing’: 1.0}
Accuracy: 0.7606666666666667
Confusion Matrix:
[[1141 359]
[0 0]]
precision recall f1-score support
0 0.76 1.00 0.86 1141
1 0.00 0.00 0.00 359
accuracy 0.76 1500
macro avg 0.38 0.50 0.43 1500
weighted avg 0.58 0.76 0.66 1500
ROC_AUC_SCORE
Out[17]:
0.5

Accuracy of logistic regression (LR) and Naive Bayes (NB) clas­

sifier are separately found in the above-mentioned modules. The
single voting model that combines the majority voting of LR and NB
is shown below:
170 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

In [18]:
#Model 3-Voting Classifier
from sklearn.ensemble import VotingClassifier
mod1 = LogisticRegression()
mod2 = GaussianNB()
vc = VotingClassifier(estimators=[(‘lr’,mod1),
(‘nb’,mod2)],
voting=‘soft’,n_jobs=-1)
params = {‘weights’:[[1,2],[1,2]]}
grid_vc = GridSearchCV(param_grid = params,cv=10, estimator=vc)
grid_vc.fit(X_test_data ,y_test_data)
print(grid_vc.best_params_)
grid_predict=grid_vc.predict(X_test_data)
#Accuracy
grid_acc_score = accuracy_score (y_test_data, grid_predict)
print(“Accuracy:”,metrics.accuracy_score(grid_predict, y_test_data))
#Confusion Matrix
print(“Confusion Matrix:\n”,metrics.confusion_matrix(grid_predict,y_test_data))
#Classification Report
print(classification_report(y_test_data,grid_predict))
#ROC Score
print(“ROC_AUC_SCORE”)
roc_auc_score(y_test_data, grid_predict*100)

Soft voting is used which produces the output as the average

probability of the two models. The combined voting method shows
the highest accuracy of 84.93% when compared to LR and NB run
separately.

Output:
{‘weights’: [1,2]}
Accuracy: 0.848
Confusion Matrix:
[[1021 108]
[120 251]]
precision recall f1-score support
0 0.90 0.89 0.90 1141
1 0.68 0.70 0.69 359
accuracy 0.85 1500
macro avg 0.79 0.80 0.79 1500
weighted avg 0.85 0.85 0.85 1500
ROC_AUC_SCORE
Out[18]:
0.7969967213434924
 EMPLOYEE TURNOVER PREDICTION 171

Visualization of true positive rate and false positive rate:

In [19]:
#Reciever Operating Characterstic Curve
log_reg_fpr,log_reg_tpr, log_reg_threshold = roc_curve(y_test_data,lr_predict)
nb_c_fpr, nb_c_tpr, nb_c_threshold = roc_curve(y_test_data, nbpred)
rf_fpr,rf_tpr,rf_threshold = roc_curve(y_test_data, grid_predict)
sns.set_style(‘whitegrid’)
plt.figure(figsize=(8,5))
plt.title(‘Reciever Operating Characterstic Curve’)
plt.plot(log_reg_fpr,log_reg_tpr,label=‘Logistic Regression’)
plt.plot(nb_c_fpr,nb_c_tpr,label=‘Naive Bayes’)
plt.plot(rf_fpr,rf_tpr,label=‘VC’)
plt.plot([0,1],ls=‘--’)
plt.plot([0,0],[1,0],c=‘.5’)
plt.plot([1,1],c=‘.5’)
plt.ylabel(‘True positive rate’)
plt.xlabel(‘False positive rate’)
plt.legend()
plt.show()

Figure 7.2 shows the ROC-AUC curve obtained for the three
models.

Figure 7.2 ROC-AUC Curve

172 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

In [20]:
model_ev = pd.DataFrame({‘Model’: [‘Logistic Regression’,‘NaiveBayes’,‘VC’], ‘Accuracy’:
[lr_acc_score*100,
nb_acc_score*100,grid_acc_score*100]})
model_ev

Out [20]:
MODEL ACCURACY
0 Logistic Regression 77.000000
1 Naive Bayes 76.066667
2 VC 84.800000

In [21]:
model_gr = pd.DataFrame({‘Model’: [‘LR’,‘NB’,‘VC’], ‘Accuracy’: [lr_acc_score*100,
nb_acc_score*100,grid_acc_score*100]})
colors = [‘blue’,‘red’,‘green’]
plt.figure(figsize=(5,5))
plt.title(“ Accuracy of three models using Voting Model”)
plt.xlabel(“Algorithms”)
plt.ylabel(“Accuracy”)
plt.bar(model_gr[‘Model’],model_gr[‘Accuracy’],color = colors)
plt.show()

The accuracy of the three models is compared in Figure 7.3.

Logistic regression shows an accuracy of 77%. Naive Bayes classifier

Figure 7.3 Accuracy of the Three Models

 EMPLOYEE TURNOVER PREDICTION 173

shows an accuracy of 76.07%. Voting classifier, the combination of

logistic regression and Naive Bayes, shows the highest accuracy
of 84.80%.

Conclusion

This work focused on predicting the employee turnover using a single

voting model. Data preprocessing was done at the initial stage. The
employee turnover was predicted with two different classifiers, lo­
gistic regression and Naive Bayes separately. Then the two classifiers
were combined to form a single voting model to improve the per­
formance of the model. The accuracy of the model with logistic
regression and Naive Bayes is found to be 77% and 76.07%, re­
spectively. The accuracy of the single voting model is found to be
84.80%. The experimental results show that the combination of the
two classifiers outperforms when the models were run separately.

References
[1] S. Karande, and L. Shyamala, “Prediction of employee turnover
using ensemble learning.” In Y. C. Hu, S. Tiwari, K. Mishra, and M.
Trivedi (eds), Ambient Communications and Computer Systems.
Advances in Intelligent Systems and Computing, vol. 904. Singapore:
Springer, 2019.
[2] P. Ajit, and R. Punnoose, “Prediction of employee turnover in or­
ganizations using machine learning algorithms.” Algorithms, vol. 4,
no. 5, p. C5, 2016.
[3] K. Arai et al. (Eds.), “Employee turnover prediction with machine
learning: A reliable approach.” IntelliSys 2018, AISC 869,
pp. 737–758, 2019. Springer Nature Switzerland AG 2019.
[4] McKinley Stacker IV, IBM Watson analytics. Sample data: HR
employee attrition and performance [Data file]. Retrieved from
https://www.ibm.com/communities/analytics/watson-analytics-blog/
hr-employee-attrition/ (2015)
[5] Xiang Gao, Junhao Wen, and Cheng Zhang, “An Improved Random
Forest Algorithm for Predicting Employee Turnover.” Mathematical
Problems in Engineering, vol. 2019, Article ID 4140707, 12 pages,
2019, doi: 10.1155/2019/4140707.
[6] T. H. Feeley, J. Hwang, and G. A. Barnett, “Predicting employee
turnover from friendship networks.” Journal of Applied Communication
Research, vol. 36, no. 1, pp. 56–73, 2008.
174 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

[7] P. A. Gloor, A. Fronzetti Colladon, F. Grippa, and G. Giacomelli,

“Forecasting managerial turnover through e-mail based social network
analysis.” Computers in Human Behavior, vol. 71, pp. 343–352, 2017.
[8] W. C. Hong, P.F. Pai, Y. Huang, and L. Yang, “Application of
support vector machines in predicting employee turnover based on job
performance.” In Proceedings of the International Conference on
Advances in Natural Computation, pp. 668–674, Springer, Berlin,
Germany, 2005.
[9] H.W. Kao, S.W. Lin, and S.Y. Wan, “Applying decision tree to
predict nursing turnover – A case study in a public hospital.” The
Journal of Taiwan Association for Medical Informatics, vol. 21, no. 4,
pp. 15–29, 2012.
[10] H. Tran, “The impact of pay satisfaction and school achievement on
high school principals’ turnover intentions.” Educational Management
Administration and Leadership, vol. 45, no. 4, pp. 279–290, 2016.
[11] L. J. Labrague, D. Gloe, D. M. McEnroe, K. Konstantinos, and P.
Colet, “Factors influencing turnover intention among registered
nurses in Samar Philippines.” Applied Nursing Research, vol. 39,
pp. 200–206, 2018.
[12] J. Rodriguez, L. Kuncheva, and C. Alonso, “Rotation forest: A new
classifier ensemble method.” IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 28, no. 10, pp. 1619–1630, 2006.
[13] A. Bagnall, M. Flynn, J. Large, J. Lines, A. Bostrom, and G. Cawley,
“Is rotation forest the best classifier for problems with continuous
features?” School of Computing Sciences, University of East Anglia,
April 2020.
[14] Riyanto Jayadi, Hafizh M. Firmantyo, Muhammad T. J. Dzaka,
Muhammad F. Suaidy, and Alfitra M. Putra, “Employee performance
prediction using Naive Bayes.” International Journal of Advanced
Trends in Computer Science and Engineering, vol. 8, no. 6,
pp. 3031–3035, November–December 2019.
[15] X. Cai et al., “DBGE: Employee turnover prediction based on
dynamic bipartite graph embedding.” IEEE Access, vol. 8,
pp. 10390–10402, 2020, doi: 10.1109/ACCESS.2020.2965544.
 8
A NOVEL IMPLEMENTATION OF
SENTIMENT ANALYSIS TOWARD
DATA SCIENCE
VIJAYALAKSHMI SARAVANAN1,
ISHPREET SINGH2, EMANUEL SZAREK2,
JEREON HAK2, AND ANJU S. PILLAI3
1
Faculty at Rochester Institute of Technology,
Rochester, New York, USA
2
M.S. Students, Rochester Institute of Technology,
Rochester, New York, USA
3
Amrita Vishwa Vidyapeetham, Coimbatore, India

Contents

Introduction 175
Related Work 178
Background of Sentiment Analysis 180
Proposed Methodology 182
Speech to Text Model 182
Implementation of Sentiment Analysis 183
Architecture of Sentiment Analysis 183
Results 186
Future Proposals of Speech-to-Text Sentiment Analysis 188
Conclusion and Future Directions 189
References 190

Introduction

Sentimental analysis (SA) is used to study the emotions of the

people, attitude of a person on a particular event/topic, and general
behavior during conversation. SA is also known as opinion mining,
which refers to text mining and natural language processing (NLP).

DOI: 10.1201/9781003119838-8 175

176 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Essentially data scientists can make use of these tools to analyze the
sentiment of large networks.
Communication is the essential component to progress on cor­
porate level and personal level. There are numerous ways of com­
munication, such as speech, hand signals, text, etc., among which
speech is deemed the most significant medium for human interac­
tion. The world is moving to digital era and has influenced the
communication techniques too. Speech recognition systems, such as
Amazon’s Alexa, Apple’s Siri, Google assistant, and many more, are
ubiquitous. There may be variations in the semantics, but the central
idea remains the same. The progressive advancements in technology
made all these advents possible. When phone calls, messages, SMS,
etc., became an integral part of human life with the invention of
mobile phones, the innovation in SMS technology via speech re­
cognition brought a tremendous change, by which voice messages are
reborn as text messages. The translations such as text-to-speech
(TTS) and speech-to-text (STT) have made user interaction with
devices possible, assisting disabled people and encouraging users to
carry out more compelling local and remote services.
STT has continued to get better as the world is rapidly advancing
with technological innovations. As compared with many existing
technologies, speech recognition was left to the imagination of the
science fiction community. The speech recognition is deeply rooted
in a research carried out by Bell Labs’ researchers back in 1952; it was
designed to recognize only numbers [1]. This system was discovered
by Alexander Graham Bell through a process that converts sound
waves into electrical impulses and is one of the novel innovations
until date [2]. Since then, speech recognition has been continuing to
grow and spread its reach. By 2011, easier, real-time, and faster way
of communication with Apple devices was possible by the launch of
Siri and to date, Amazon’s Alexa and Google’s Home are the most
preferred voice-based virtual assistants used by the people across the
globe. It has transformed the way the companies run their businesses.
But, companies lack in finding a reliable means of computing the
human quality of spoken conversations with their customers to help
employees with the way they handle inventory. During inventory
management, certain words are most likely being re-used repeatedly.
These data can hold valuable information because language is very
 A IMPLEMENTATION OF SENTIMENT ANALYSIS 177

important depending on how you use it. Words carry meaning and
we can distinguish how someone may be feeling depending on what
words they are using.
The speech converted to text finds further more purposes such as
SA. SA is a combination of methods, techniques, and tools to re­
cognize and group opinions using NLP to decide whether the
speaker’s attitude toward a specific topic/product is positive, negative,
or neutral [3]. It is very helpful in comprehending the mood of
humans in many cases. In a long conversation, people render their
opinions about the topic, products, social issues, movies, etc. The
identification of the sentiment of the speaker can enhance the re­
trieval of the content to improve the usefulness and collecting the
combined sentiment of a group of people on alike topic can help in
setting up the general sentiment. Sentiment detection using text is an
evolved research area and can tremendously influence the inventory
by giving emphasis on the product reviews. Based on the emotional
prosody studies, the tones of every human’s voice can be dis­
tinguished by its intensity, pitch, loudness, speech rate, pauses,
timbre, and many such parameters and whose deviations reveal dif­
ferent information from the speaker to the listener.
There are wide varieties of SA applications for social goods, such
as fighting for the pandemic, preparing people for future work,
promoting inclusivity, and protecting the environment. In this work,
we approached SA for social goods in a different way such that one
can utilize our analytic results in assistive technologies for students
with learning difficulties, and person with disabled arms. We use
STT tools to describe the efficiency of how our data science algo­
rithm outperforms with our real-time application domain, inventory
management systems. Further, this SA algorithm can be im­
plemented effectively in assistive devices.
The proposed work implements STT conversion with SA to
detect the condition of machine/module in an industry for ma­
chinery maintenance. Consider a case in which many laborers work
in an industry where different mechanical parts are fabricated and
there is a need to guarantee that every module is in working con­
dition and in stock. To incorporate an efficient method for finding
the working condition of the module, STT technique can be em­
ployed in place. By having an STT application, the details about the
178 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

parts could be sent directly to a text document. We then used the

obtained document to differentiate between the parts. Adding SA
to this case was found to be interesting and fetching accurate re­
sults. The outcome of the SA will help the workers to be aware of
parts needed to be considered on a concrete level, rather than simple
guesses and predictions, thereby improving the production and
down time.
The paper is organized as follows: the section “Related Work”
presents a detailed literature summary about the related work done in
the field, followed by background of SA in the section “Background
of Sentiment Analysis”. The section “Proposed Methodology” ex­
plains the methodology adopted to implement STT and SA algo­
rithms. Results are presented in the section “Results” and the future
application prospects of STT and sentiment algorithm are discussed
in the section “Future Proposals of Speech-to-Text Sentiment
Analysis”. The paper is concluded in the section “Conclusions and
Future Directions” along with the future direction of work.

Related Work

The vast literature shows that there is a lot of interest shown by the
researchers on SA. SA identifies the sentiment conveyed in a text and
upon analysis finds out whether the text expresses positive, negative,
or neutral sentiment. Kralj Novak et al. [4] conducted a Twitter
message SA based on emoji use. All messages were labeled with
either a positive, neutral, or negative label. If emojis would occur
several times, all occurrences were counted. They found that the use
of popular emojis could be used to classify a twitter message on
sentiment. A comparison was also made for twitter messages in
different languages and in all languages, there were emojis that can be
used to classify a message as positive, neutral, or negative.
Besides STT, an algorithm could be trained to recognize unique
voices and unique words. As Saurav Bhandari et al. [5] proposed, a
speech detection system could be used for unlocking and starting
cars. The algorithm would recognize the specific voice of the user and
recognize the designated word (or sentence) for unlocking and
starting the car. The idea was to increase the security and safety
against theft. To enhance safety, the proposed system could be
 A IMPLEMENTATION OF SENTIMENT ANALYSIS 179

expanded with face recognition sensor and software. The entire

system could work on a Raspberry Pi. Also, there are many SA
implementations using algorithms such as Naïve Base, support vector
machine, decision tree, maximum entropy, and many more in the
literature [6–8]. Mostafa Shaikh et al. [9] proposed a method to
classify the input sentences as subjective or objective from a given
document. Classical machine learning (ML) techniques are applied
for the subjective parts and objective parts are discarded. A multi-
modal approach to identify the sentiment of products based on audio
and text was proposed by Abburi et al. [10]. A real-time speech
emotion recognition based on deep neural network (DNN) to
identify emotions from a one-second raw speech spectrograms was
presented by Fayek et al. [11]. SA applied to political debates was
considered in a research by Salah et al. [12]. A maximum entropy
modeling approach to identify sentiment using natural audio streams
from YouTube was presented by Kaushik et al. [13].
STT has been around for quite some time and has had many
different applications. One primary use is of educational purpose,
that is to make teaching easier [14–18]. In the work by Shadiev et al.
[19], different ways of STT translations that helped the educational
field are presented. Clearly, there are many ways in which STT can
be used to enhance the learning capabilities of educators. Due to the
nature of online classrooms, some teachers may feel disconnected
with students. It may be hard to process what the students are
feeling. A possible future study could be carried out to use a com­
bination of the STT with SA to see if there can be different tech­
niques applied to help the students learn better.
Social media has become a sort of Wild West in our society. There
are not a lot of moderators or controls in place. For most cases,
people may say or express whatever they want. This creates a massive
forum of both positive and negative opinions, which are ripe with
data. This should be very helpful for those attempting to test the
waters of a new product. An article that did SA to evaluate social
media posts using YouTube gamers, such as PewDiePie, found some
interesting results. Whenever content was posted, usually some sort
of repost from YouTube, on Facebook, it was generally not very well
received and most of the comments were negative [20]. Market
analysts can get a better picture of what brings people together and
180 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

use it to their advantage to create advertisements based on what

brought on these happy comments.
The application for SA is endless. Ben-Porat et al. have conducted
research to predict action from free text [21]. A collection of text
including text messages, Facebook posts, tweets, etc. is analyzed as an
input into the algorithm to predict an action made in a game [21]. The
study proved successful and there was a correlation between free text
and the actions a person would take in a game. The results were biased
toward certain genders and the data analyzed were primarily college-
aged students but still opened the possibility with a larger dataset and a
robust ML algorithm that can successfully analyze and predict a user’s
behavior. This can be a very useful tool for marketers along with ad­
vertisers to place the proper advertisements for users.
Going beyond sentiment, we can also look at the data collected
from STT algorithms. Research done by Gillick and Bamman et al.
predicted the lyrics for a given audio input [22,23]. The project takes
an audio source and removes any speech leaving background sound.
From the background audio input into the algorithm we can create
musical lyrics based on the tones. Data produced from the STT
application can be used to train and create lyrics. This can assist
musicians and vocalists writing lyrics for their music. Perhaps one day
we’ll see producers creating beats and an algorithm developing lyrics
to become the next biggest virtual artist.

Background of Sentiment Analysis

Sentimental analysis identifies the sentiment in the expressed text and

then it analyzes the opinion, which is shortly referred as SA. Speech
recognition has a lot of different factors. Based on the uttered words,
words are classified as isolated word, connected word, continuous
speech, and spontaneous speech. Isolated word is what it sounds like,
when a single word is spoken. There will be silence, a word will be
spoken, and then silence again. Connected word is having two words
instead of one. One would utter a word and then a second word right
after it with minimal pause. Continuous speech is described as talking
naturally and these are considered as one of the difficult tasks because
identification of utterance boundaries is done through unique sound
and special methods [24]. Finally, spontaneous speech is impulsive
 A IMPLEMENTATION OF SENTIMENT ANALYSIS 181

speech that you would think of on the spot and just start yapping away.
This feature is good at detecting when the user is stringing different
words together and when he/she is not.
Depending on the place we grew up, everyone’s speech dialect is
different. Sometimes, for humans it can be challenging to understand
each other. This becomes more challenging for the speech software to
decode. Humans have such distinct voices and dialects that make it
harder for the software to understand what is being said. It is due to
these differences that there are two types of speaker models. (1) There
are speaker independent models, which are built for the masses, so that
they are able to understand a large array of speakers effectively. But is
hard to implement and is costly, and it renders reduced accuracy in
comparison with speaker-dependent systems. One of the factors for its
wide usage is its flexibility. (2) Speaker-dependent systems are the
opposite of the independent systems and are built specifically for one
target individual. These systems lack flexibility but are preferred as they
provide easy development and more accuracy, and are cheaper. For
single speaker model, it exhibits high accuracy, while for multi-speaker
model, the speaker-dependent system produces less accuracy [24].
When it comes to opinion, people usually have a stance on whether
they agree or support something, or whether they disagree and are
against something. It seems like the more the opinion is supported, the
stronger the support/hate is. With SA,it tries to differentiate between
the two and put them into a positive review or negative review. This
can be done in two different ways: direct or comparison opinions.
With direct opinion, we get the view of the product as positive or
negative directly. For example, “Life style of people at rural area is
poor” expresses a direct opinion. Meanwhile, for comparison, we
compare the subject with other objects of similar class. For example,
“The life style of people in village-A is better than that of people in
village-B” expresses a comparison.
In Mostafa et al.’s research [9], authors presented the sentence in
each document labeled with only subjective part and discarded ob­
jective part. After that, they applied the classical ML algorithms on
subjective parts, which is time consuming at the sentence level and
not an easy job to test them. To perform STT on SA, we have used
the following methods: – logistic regression and comparison are
made to find the efficiency of the proposed algorithm.
182 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Proposed Methodology
Speech to Text Model

STT model is built on the concept of an on-line STT engine [25].

The primary objective of any speech translation system is to make it
possible for a machine to listen, recognize, and produce actions based
on the spoken words. The user voice at run-time is captured via a
microphone to process the sampled speech. This allows recognizing
the uttered word by the user. The identified word is then conerted
into output as a text. For large volume of input speech signal, the
identified words can be stored in a file for further analysis and use.
Large and complex systems can adopt different choices for data entry.
The STT algorithm was built around the Google STT Python li­
brary. The Google API allows us to keep the algorithm very light­
weight and scalable. The Google API does STT efficiently allowing
the developers to focus on adding additional features to the algorithm
and making the product user friendly, as shown in Figure 8.1. In the
proposed work, the STT function in Python is used and the users
microphone is used as input.
Additional features such as SA and widgets were added to allow
users to use the algorithm with ease and allow engineers to look at the
data and respond accurately. The STT and SA algorithm is a user-
friendly codebase that can be implemented across multiple devices.
There are some negative impacts of the proposed algorithm, since
it relies on Google STT API. If Google had any outages or down­
time, the STT would cease to function. Since STT relies on an API,
the ability to customize the library and improvements are limited for
future development. STT algorithm is very sensitive and if there is
loud background noise, the model will exhibit reduced performance
by producing inaccurate texts from the processed sound. SA has its
own set of negatives with a small training set that was custom built
for engineers. Due to the small training set, the error rate contains

Speech Google’s Cloud Text file

Speech input
Processing STT API at work output

Figure 8.1 General Representation of the STT Model

 A IMPLEMENTATION OF SENTIMENT ANALYSIS 183

biases and has an error rate. Overtime data collected can be used to
improve the SA algorithm. The SA algorithm can also be improved
to a more robust algorithm with a higher accuracy rate. Other data
mining features can be added in the future.

Dataset: https://www.kaggle.com/c/tensorflow-speech-recognition-challenge
Methods: The software used is Python. Specifically, the libraries libROSA and SciPyis are used
to process the audio signals.

Implementation of Sentiment Analysis

SA is also referred as opinion mining, as it combines volumes of

opinion on a topic/product. Text mining is a technique used for
opinion mining. The various processes involved in SA are shown in
Figure 8.2. The SA algorithm takes input from online user and
processes the raw data to extract the features from the input signal.
From the extracted features, adjectives are identified to detect the
emotions of the user involved in the conversation.
In the proposed work, SA is implemented using a logistic regression
algorithm. Logistic regression is a classification model and is one of the
dominant algorithms for classification used in industry. The steps of
logic regression implementation are shown in Figure 8.3.
The input to the algorithm is maintenance data set from a ma­
chinery maintenance industry. This maintenance data are the input
vector x. Vector of weights w are created to adjust accordingly to
predict the output. Through the adjusting of weights, the algorithm
learns to classify the output correctly. The net input is referred as the
dot product of input vector and weight. Sigmoid function is used to
fix the net input in a range of [0,1], for classification of the input as
either positive or negative. Defining the threshold function helps in
accurate classification of the output.

Architecture of Sentiment Analysis

Architecture of SA exhibits the steps involved in concluding reviews

from the user input. Figure 8.4 shows the detailed steps of archi­
tecture of SA.
 184

Positive

Conclusive
Processing of Raw Sentiment
Opinion
Online User Data Model

Database
Negative

Figure 8.2 General Process of Implementation of SA

APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT
 Predicted
Sigmoid Threshold
Net input Output
Vector of activation function
Input function
weights function
vector

Figure 8.3 Steps of Implementation of Logic Regression Algorithm

A IMPLEMENTATION OF SENTIMENT ANALYSIS
185
186 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Engineers Speech to text assessment

Part Opinion/Information Retrieval

Stored STT File (Noted in the parts)

Classification of Parts of Text

Positive Part Negative Part

Review Review

Summarization of Reviews

Figure 8.4 Architecture of SA

Results

In this work, the STT translation with SA is employed to report the

module/part failure by mechanics for an easier and accurate classification
of their findings. The negative classified reports could be reviewed by
engineers to identify patterns. SA is implemented using a logistic re­
gression algorithm. In order to train the algorithm, a training set and test
file are created containing example messages about maintenance, as
shown in Figure 8.5 and Figure 8.6. The training set consists of 75
positive and 75 negative messages. The test set consists of 50 positive
and 50 negative messages. The data are cleaned and vectorized in order
to train the logistic regression algorithm.
An important step is to find the right hyper parameter C. Several
values for C have been tried and the best fit was chosen, as shown in
Table 8.1.

Part 1 appears to have no issues

Part 2 seems like it has a few scratches but nothing worth removing
Part 3 is very sharp and can be used next week
Part 4 needs to be buffered but other than that no issues
Part 5 looks new

Figure 8.5 Snapshot of Positive Training Set

 A IMPLEMENTATION OF SENTIMENT ANALYSIS 187

Part 152 crappy ridges and needs new bearings

Part 153 makes me question who approved thiscrap
Part 154 was just morejunk
Part 155 feels like it was put together in seconds and feels super cheap

Figure 8.6 Snapshot of Negative Training Set

Table 8.1 Accuracy for Different C Values

C VALUE ACCURACY
0.01 0.76
0.05 0.80
0.25 0.82
0.50 0.84
1.0 0.84
1.5 0.86
2.0 0.86
3.0 0.86

All values higher than C = 1.5 do not add to more accuracy. The
value of C = 1.5 is used to train the algorithm. After training
the algorithm, it was tested and evaluated based on the accuracy. The
accuracy of the algorithm is shown in Figure 8.7. The accuracy of
the proposed model is found to be 81% and can be considered as
sufficient for practical use.
Upon training and testing, the logistic regression algorithm is
integrated with the STT translation algorithm. This algorithm uses
the STT function in Python and uses the user’s microphone as input.
The output is both written to a .txt file and put through the SA. The
.txt output can be used to fill in the mechanics’ work order and the
SA can be sent to the engineering department. Other uses may be
applicable too. To make the STT algorithm even more accessible, it

Final Accuracy: 0.81

Figure 8.7 Final Accuracy

188 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

[‘part is looking good’, ‘no problems here’, ‘doors function very well’]

Figure 8.8 Final Model Input Example

was supplied with a button to record the user’s microphone for a set
amount of time. Using the slider, the amount of time can be adjusted.
In the figure, the slider was set to 40 seconds.
The STT algorithm was used to record multiple sentences and the
output was put through the SA. In Figure 8.8, a few example lines
are shown.
The algorithm classified all lines as positive as it should. A mes­
sage gets classified as positive by a “1” and negative by a “‘0.” Multiple
tests were conducted and the algorithm seems to classify the mi­
crophone most of the time correctly.
In our initial experimentation and testing, we were very successful
in running the STT algorithm. The STT itself is powered by the
Google STT API. The API is very powerful and accurate in taking
the speech input and giving a rapid output of text. By combining the
STT with our trained SA, we were able to get an accuracy rate of
81%. The output of the SA will help provide mechanics the most
critical situations to work on.

Future Proposals of Speech-to-Text Sentiment Analysis

The application of the STT SA is proposed for mechanics to report

their findings while maintaining machinery. The STT algorithm makes
it easier for mechanics to report the condition of the machines rather
than typing it on a keyboard. It is useful when they are not close to a
computer or even when the computer is near, as their hands could be
dirty due to maintenance or repairing of the machinery. The SA can
then be used by engineers who oversee maintenance. By focusing on the
analysis, the machine conditions can be classified as negative or positive
and it becomes easier to know the status of the machinery. This tool can
go hand in hand with real-time monitoring tools that monitor the status
of machinery through the sensor data. Combining the SA of findings
and the real-time monitoring tools could provide a good understanding
of the status of the machinery and will improve the production.
 A IMPLEMENTATION OF SENTIMENT ANALYSIS 189

Other than the maintenance, there are more applications for the
STT SA software. When you pay at a restaurant by card, the card
reader could perhaps ask your opinion and record it. The file is sent
to a computer, which runs the speech file through the software and
puts both the text and the SA in a database. Some people might feel
embarrassed to give their honest opinion when the waitress is
standing beside them. The waitress leaving the card reader on the
table and coming back after the customer has paid could fix this.
The same could be applied when rating an app or an Uber driver. The
rating with stars or numbers should not necessarily have to be operated
by speech. However, when there is a comment box, this could be re­
placed by speech. The sentiment of the message can be used to rate the
app or the Uber driver as an addition to the rating by stars or numbers.

Conclusion and Future Directions

This works presents the generalized model of STT on SA. In this

proposed work, we have implemented a simple system called inventory
management systems to perform the above-mentioned tasks to show
how this can be utilized for social goods, such as assistive devices. This
system works well with smaller datasets. We are currently working on
a larger dataset and increasing scalability of the proposed system.
The goal of the proposed work is to assist engineers to rapidly
record the system status and equipping sentiment analysis (SA) to
assess the status of the mechanical modules/parts. This is achieved by
implementing a speech-to-text (STT) translation algorithm to
output the corresponding texts recorded by the mechanic. The STT
algorithm is integrated with SA algorithm implemented using logic
regression classifier to group the output as either positive or negative
review. With the assessment provided by SA, mechanics can engage
the negative observances and fix the issues. The proposed work can
also assist with analytics and record the number of negative situations
engineers encounter each day.
This work is an amazing asset for engineers to use. The software
developed is very lightweight and user friendly providing a high rate of
accuracy. We can continue improving the algorithms by increasing the
rate of accuracy of the observances. As the input data grow, they can be
collected and added to the training set to improve our algorithm.
190 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Another improvement that can be made is looking at a more robust SA

algorithm, such as a neural network to provide a higher accuracy rate.
Creating a more robust algorithm will be beneficial for mechanics to
handle critical situations. We can also include analytics and data mining
functionality with more data to improve the accuracy of producing SA
output. Data mining at minimum would include information about the
parts that are continuously failing and being replaced or the number of
issues mechanics are dealing with on a daily basis. Another improvement
is adding multi-device support for the proposed work with the possibility
of rolling out mobile-ready applications that mechanics can use.

References
[1] Kikel, Chris. (2019). “A Brief History of Voice Recognition Technology,”
Total Voice Technologies. https://www.totalvoicetech.com/a-brief-
history-of-voice-recognition-technology/.
[2] Ghai, Wiqas, and Singh, Navdeep. (2012). “Literature Review on
Automatic Speech Recognition,” International Journal of Computer
Applications, vol. 41, no. 8, pp. 42–50. https://doi.org/10.5120/5565-
7646.
[3] Mäntylä, Mika V., Graziotin, Daniel, and Kuutila, Miikka. (2018).
“The Evolution of Sentiment Analysis – A Review of Research Topics,
Venues, and Top Cited Papers,” Computer Science Review, vol. 27,
pp. 16–32. https://doi.org/https://doi.org/10.1016/j.cosrev.2017.10.002.
[4] Kralj, Novak P., Smailović, J., Sluban, B., and Mozetič I. (2015).
“Sentiment of Emojis,” PLoS ONE, vol. 10, no. 12, e0144296. doi:
10.1371/journal.pone.0144296. https://journals.plos.org/plosone/
article?id=10.1371/journal.pone.0144296.
[5] Bhandari, Saurav, Jain, Saiyam, Jambhale, Sagar, and Khapare,
Ajinkya. (2017). “Voice Recognition System for Automobile Safety.”
International Research Journal of Engineering and Technology, vol. 4, no.
2, https://www.irjet.net/archives/V4/i2/IRJET-V4I2303.pdf.
[6] Pang, B., and Lee, L. (2014). “A Sentimental Education: Sentiment
Analysis Using Subjectivity Summarization Based on Minimum Cuts,”
In Proceedings of the 42nd Annual Meeting on Association for
Computational Linguistics, p. 271, Association for Computational
Linguistics.
[7] Pang, B., and Lee, L. (2005). “Seeing Stars: Exploiting Class
Relationships for Sentiment Categorization with Respect to Rating
Scales,” In Proceedings of the 43rd Annual Meeting on Association for
Computational Linguistics, pp. 115–124, Association for Computational
Linguistics.
 A IMPLEMENTATION OF SENTIMENT ANALYSIS 191

[8] Pang, B., Lee, L., and Vaithyanathan, S. (2002). “Thumbs Up?:
Sentiment Classification Using Machine Learning Techniques,” In
Proceedings of the ACL-02 Conference on Empirical Methods in Natural
Language Processing, vol. 10, pp. 79–86, Association for
Computational Linguistics.
[9] Shaikh, M., Prendinger, H., and Mitsuru, I. (2007). “Assessing
Sentiment of Text by Semantic Dependency and Contextual Valence
Analysis,” Affective Computing and Intelligent Interaction, ACII 2007,
LNCS 4738, pp. 191–202.
[10] Abburi, Harika, Gangashetty, Suryakanth V., Shrivastava, Manish,
and Mamidi, Radhika. (2017). “Audio and Text Based Multimodal
Sentiment Analysis Using Features Extracted from Selective Regions
and Deep Neural Networks,” MS Thesis. International Institute of
Information Technology, Hyderabad, India.
[11] Fayek, H. M., Lech, M., and Cavedon, L. (2015). “Towards Real-
Time Speech Emotion Recognition Using Deep Neural Networks,”
2015 9th International Conference on Signal Processing and
Communication Systems (ICSPCS), Cairns, QLD, pp. 1–5, doi: 10.11
09/ICSPCS.2015.7391796.
[12] Salah, Saleh, and Ibrahim, Zaher. (2014). “Machine Learning and
Sentiment Analysis Approaches for the Analysis of Parliamentary
Debates,” Thesis.
[13] Kaushik, L., Sangwan, A., and Hansen, J. H. L. (2013). “Sentiment
Extraction from Natural Audio Streams,” 2013 IEEE International
Conference on Acoustics, Speech and Signal Processing, Vancouver, BC,
pp. 8485–8489, doi: 10.1109/ICASSP.2013.6639321.
[14] Nisbet, P., and Wilson, A. (2002). “Introducing Speech Recognition
in Schools,” Edinburgh, UK: CALL Centre, University of
Edinburgh.
[15] Nisbet, P., Wilson, A., and Aitken, S. (2005). “Speech recognition for
students with disabilities,” Proceedings of the Inclusive and
Supportive Education Congress, ISEC 2005 Conference, Delph, UK:
Inclusive Technology.
[16] Nisbet, P., Wilson, A., and Balfour, F. (2008). Introducing Speech
Recognition in Schools: Using Dragon Naturally Speaking, Edinburgh,
UK: CALL Centre, University of Edinburgh.
[17] Petta, T. D., and Woloshyn, V. E. (2001). “Voice Recognition for
On-line Literacy: Continuous Voice Recognition Technology in
Adult Literacy Training,” Education and Information Technologies, vol.
6, no. 4, pp. 225–240.
[18] Ranchal, R., Taber-Doughty, T., Guo, Y., Bain, K., Martin, H.,
Robinson, J., and Duerstock, B. (2013). “Using Speech Recognition
for Real-Time Captioning and Lecture transcription in the class­
room,” IEEE Transactions on Learning Technologies, vol. 6, no. 4,
pp. 299–311.
192 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

[19] Shadiev, Rustam, Hwang, Wu-Yuin, Chen, Nian-Shing, and Huang,

Yueh-Min. (2014). “Review of Speech-to-Text Recognition
Technology for Enhancing Learning,” Journal of Educational Technology
& Society, vol. 17, no. 4, pp. 65–84. Accessed April 8, 2020.
www.jstor.org/stable/jeductechsoci.17.4.65.
[20] Poecze, Flora, Claus Ebster, and Christine Strauss. (2018). “Social
Media Metrics and Sentiment Analysis to Evaluate the Effectiveness
of Social Media Posts,” Procedia Computer Science, vol. 130,
pp. 660–666. https://doi.org/10.1016/j.procs.2018.04.117.
[21] Ben-Porat, Omar, and Sharon Hirsch. (2020). “Predicting Strategic
Behavior from Free Text,” Journal of Artificial Intelligence Research.
https://arxiv.org/pdf/2004.02973.pdf.
[22] Gillick, Jon, and David Bamman. (2019). “Breaking Speech
Recognizers to Imagine Lyrics,” NeurIPS 2019, Workshop on Machine
Learning for Creativity and Design, December 15. https://arxiv.org/
pdf/1912.06979.pdf.
[23] Kaushik, Abhishek, Anchal Kaushik, and Sudhanshu Naithani. (2015).
“A Study on Sentiment Analysis: Methods and Tool,” International
Journal of Science and Research, vol. 4, no. 12, pp. 287–292. https://
pdfs.semanticscholar.org/c151/dfad8c1bf88b0afc716758c77d533
ded7dd0.pdf.
[24] Khilari, Prachi, and Bhope, V. P. (2015). “A Review on Speech to
Text Conversion Methods,” International Journal of Advanced Research
in Computer Engineering and Technology, vol. 4, no. 7, pp. 3067–3072.
http://ijarcet.org/wp-content/uploads/IJARCET-VOL-4-ISSUE-
7-3067-3072.pdf.
[25] B. Raghavendhar Reddy, E. Mahender. (2013). “Speech to Text
Conversion Using Android Platform”, International Journal of
Engineering Research and Applications, ISSN: 2248-9622, vol. 3, no. 1.
 9
CONSPECTUS OF K-MEANS
CLUSTERING ALGORITHM
USHA SAKTHIVEL1, A. P. JYOTHI2,
N. SUSILA3, AND T. SHEELA4
1
Dean, Dept. of Research and Innovation, Rajeswari College
of Engineering, Bengaluru, India & Prof. & Head,
Dept. of Computer Science & Engineering,
Visvesvaraya Technological University, Belagavi,
Karnataka, India
2
Associate Professor, Department of CSE, Visvesvaraya
Technological University Belagavi, Karnataka, India
3
Prof. & Head, Department of IT, SKCET, Coimbatore,
India
4
Prof. & Head, Department of IT, SSEC, Chennai,
India

Contents

Introduction 194
Background Study 194
Overview of K-Means Clustering 198
An Insight about the Operational Working of K-Means
Clustering 199
Steps Involved in K-Means Clustering Algorithm 199
Real-Time Project Implementation on Synchronic Health
Fitness Detector of Combatants in Armed Conflicts Using K-
Means Clustering Algorithm 199
Proposed Work 200
Design Methodology and Implementation 201
Combatant’s Segment 201
Chief of Armed Force’s Segment 201
Army Control Room’s Segment 201
Results 201
Real-World Use Cases of K-Means Clustering 201

DOI: 10.1201/9781003119838-9 193

194 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

K-Means Clustering Algorithm Used as a Solution for

Following Business Problems 209
Purchaser Division 209
Conveyance Enhancement 209
Archive Arranging and Gathering 210
Client Maintenance 210
Rebate Examination 210
Comparisons of K-Means Clustering with Other
Unsupervised Algorithms 210
Pros and Cons 211
Conclusions and Future Work 212
References 212

Introduction

The expression “K-means” was originally utilized by James

MacQueen in 1967 as a major aspect of one of the manuscripts. The
standard calculation was likewise utilized in Bell Labs as a feature
method in beat program tweak in 1957. It was likewise distributed by
E. W. Forgy.
The majority of unsubstantiated erudition-based purposes use the
sub-domain as clustering. It is the way toward gathering information
tests together into clusters dependent on a specific element that they
share – precisely the motivation behind unaided learning in any case.
We are given an informational index of things, with specific high­
lights, and qualities of these highlights. The undertaking is to classify
those things into K gatherings. To accomplish this, we will utilize the
K-means calculation, an unsupervised learning algorithm. To compute
that similitude, we will utilize the Euclidean separation as estimation.

Background Study

Tajunisha et al. [1]. Execution analysis of K-means with distinctive

instatement techniques for high-dimensional information utilizes
principal component analysis (PCA) to measure decrease and
 CONSPECTUS OF K-MEANS CLUSTERING ALGORITHM 195

discover starting bunch communities. The variable with the most

elevated Eigen esteem is determined utilizing PCA and is taken as
first head segment along which apportioning is done, on the premise
of which K subsets are shaped and K middle qualities are taken as
beginning K communities.
Bouhmala et al. [2]. Joined genetic algorithm (GA) and K-means
(GAKM) are used to improve the nature of bunches framed and
accelerate their hunt procedure. The presentation of GAKM is tried
over the datasets, for example, iris, glass, and so on, which have been
taken from machine learning vault. The trial results have demon­
strated that GAKM joins quicker while contrasting with standard
GA. Despite the fact that this calculation neglected to catch the best
nature of bunches, it is unsatisfactory for expanding both homo­
geneity and heterogeneity inside same bunches and with various
groups individually.
Rouhollah et al. [3]. Anticipated a model called GA grouping for
improving K-means calculation. The calculation has been performed
on notable datasets to be specific iris, raw petroleum. The test results
give bunching standard μ; the lower estimation of μ gives better
bunch arrangement of information contrasted with customary
K-means bunching calculation.
Jyothi A. P. et al. [4]. The paper gives an inside and out ex­
amination of the current grouping methods alongside their adequacy
in vitality proficiency.
N. Kaur et al. [5]. This paper considered up gradation of the
customary K-means by presenting ranking technique. The creator
presents ranking technique to conquer the inadequacy of more
execution time taken by customary K-means. The ranking method
is an approach to discover the event of comparative information
and to improve search viability. The device used to execute the
improved calculation is Visual Studio 2008 utilizing C#. The focal
points of K-means are likewise investigated in this paper. The
creator discovers K-means as quick, powerful, and simple justifi­
able calculation. Also, the groups are non-various leveled in nature
and are not covering in nature. The procedure utilized in the
calculation takes understudy marks as informational index and
afterward, the starting centroid is chosen. Euclidean separation is
196 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

then determined from centroid for every information object. At

that point, the edge esteem is set for every informational collection.
Positioning method is applied straightaway and at last, the groups
are made dependent on least separation between the information
point and the centroid.
Sandeep Rana et al. [6]. This paper proposed another hybrid
sequential grouping approach. They have utilized particle swarm
optimization (PSO) and K-means calculation in grouping for in­
formation bunching. This methodology was proposed to beat the
disadvantages of the two calculations, as well as to improve the
grouping and abstain from being deteriorated. Four sorts of in­
formational indexes have been tried so as to get relative outcomes.
For examination reason, extraordinary calculations, for example,
PSO, K-means, hybrid K-means PSO, what’s more, hybrid K-means
+ GA were thought of. The proposed calculation produces pro­
gressively precise and vigorous bunching results.
Jyothi A. P. et al. [7]. The paper presents a novel logical de­
monstration of CFCLP that uses information conglomeration by a
novel combinatorial methodology utilizing straight programming for
developing up with another bunching instrument. An advanced
scientific demonstration is utilized for planning and its calculation
being actualized on MEMSIC hubs to find that CFCLP is equipped
for keeping up great harmony between vitality sparing execution and
information conveyance among the sensor hubs for longer cycles.
Bara’a Ali Attea et al. [8]. The paper found that exhibition of
bunching calculations debases with an ever-increasing number of
covers among bunches in an informational index. These realities
have inspired to build up a Fluffy Multi-Target Molecule Swarm
Streamlining System (FMOPSO) in an inventive style for in­
formation bunching, which can convey more viable outcomes than
cutting-edge grouping calculations. To find out the predominance
of the proposed calculation, a number of measurable tests have been
completed on an assortment of numerical and all out genuine in­
formational indexes.
Jyothi A. P. et al. [9]. The paper presents an epic multi-stage
enhancement demonstration for tackling huge scope vitality ad­
vancement issue and along these lines, defined conditional ideal
bunching setup is examined.
 CONSPECTUS OF K-MEANS CLUSTERING ALGORITHM 197

Chetna Sethi et al. [10]. The paper proposed a linear PCA-based

crossover K-means bunching and PSO calculation PCA-K-PSO. In
PCA-K-PSO calculation, the quick intermingling of K-means cal­
culation and the worldwide look through the capacity of PSO are
consolidated for bunching enormous information sets utilizing linear
PCA. Better grouping outcomes can be acquired with PCA-K-PSO
when contrasted with common PSO. This was viably evolved to
make its utilization for productive bunching of high-dimensional
informational collections.
Rui Xu et al. [11]. “Review of Clustering Algorithms.” Data ex­
amination assumes a key job for understanding different marvels.
Pack examination, unrefined investigation with little or on the other
hand no earlier information, comprises examination created over a
wide assortment of networks. The assorted variety, on one hand,
furnishes us with various instruments. Then again, the abundance of
alternatives creates disturbance, an audit gathering calculations for
informational records appearing in measurements, software de­
signing, and artificial intelligence (AI), and layouts their applications
in a couple of benchmark informational lists, the voyaging sales rep
issue, and bioinformatics, another field drawing in concentrated
endeavors. A couple of solidly related focuses, closeness measure, and
pack approval are additionally examined.
Jyothi A. P. et al. [12]. The paper presents a novel procedure
called EPCA. The strategy gives two-level group streamlining by
giving numerous credits to exact determination of cluster head.
EPCA gives less unpredictable activity, quicker reaction, and better
sustenance toward longer recreation adjusts.
Xiaohui et al. [13]. The paper utilized PSO for archive bunching.
K-means calculation is most regularly utilized apportioning calculation
for bunching enormous datasets, yet it produces neighborhood ideal
arrangement. Rather than confined looking through property of K-
means, PSO performs globalized search utilizing whole arrangement
space. Creators utilized PSO, K-means, and half breed PSO bunching
calculation on four report datasets, which are acquired from text retrieval.
Jyothi A. P. et al. [14]. The proposed framework in this paper
presents one kind of topology control component utilizing a novel
idea of interstellar direction toward upgrading the grouping execu­
tion in WSN. Receiving an expository exploration philosophy, the
198 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

proposed framework presents two interstellar-based topology control

framework, which focuses on the most extreme sparing of asset
utilization of the group head. The reproduced result of the ex­
amination shows that the proposed geography control framework
offers huge vitality protection execution in contrast with the current
various leveled grouping plan in WSN.
Osama Abu Abbas et al. [15]. The paper shows comparison of
K-means clustering algorithm with other unsupervised algorithms,
namely, HC, SOM, and EM clustering algorithms.

Overview of K-Means Clustering

K-means clustering belongs to a category of unsupervised learning, a

sort of unaided realizing, which is used when you have unlabeled
data. The target is to calculate the bunches in the records, to get a
count of get-togethers indicated by the variable K. The methodology
of K-means clustering is depicted in Figure 9.1.

Before K-Means

After K-Means

Cluster 2
Cluster 1
is
Ax
Z-

Cluster 3

Figure 9.1 K-Means Clustering Methodology

 CONSPECTUS OF K-MEANS CLUSTERING ALGORITHM 199

An Insight about the Operational Working of K-Means Clustering

The scheme follows a fundamental and covenant methodology, por­

traying the given enlightening file by a meticulous figure of bundles
predetermined previously. The indispensable proposal portrays K cen­
troids, individual in every bundle. These ought to be set in an astuteness
way because different territories cause different results. Thus, the su­
perior choice bundles them normally far off from one another.
The accompanying phase takes each immediate have a position
toward a known enlightening record and accomplice it to the nearest
centroid. Exactly when no aim is forthcoming, the basic advance and
an early on groupware are carried out.
As of now, we need to re-figure K new centroids (KNCs) as barycenter
of the gatherings approaching, because of the point of reference ad­
vancement. After refiguring these KNCs, another blend must be prepared
in the midst of comparative instructive file centres and the nearest NC.
An encircle is made. In view of this, we observe that the
K centroids transform their regions slightly at a time in anticipation
of nil changes. Finally, centroids do not shift to any further extent.

Steps Involved in K-Means Clustering Algorithm

Following are the steps involved:

• K centroids are formed arbitrarily.
• K-Means allocate each information point in the information set
so that they are near the centroid.
• Then K-means reconfigure the centroids by averaging out the
information and pronouncing the fresh centroid with respect to
all information points assigned to that centroid’s cluster, re­
sulting in reduction of the overall intra-cluster dissent compared
to the preceding step.
• Rehash Steps 2 and 3 until the centroids don’t shift to any
further extent.

Real-Time Project Implementation on Synchronic Health Fitness Detector

of Combatants in Armed Conflicts Using K-Means Clustering Algorithm

The aim of this project is to track the location of combatants in

armed conflicts and to provide synchronic health fitness monitoring
using K-means clustering algorithm.
200 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Proposed Work

The aim of the proposed work is to form clusters. In this project

called troops are formed using K-means clustering algorithm. The
chief of armed forces acts as cluster head, who responsible for data
aggregation from all combatants in that troop. Use of K-means
clustering algorithm plays a vital role in this project and acts as a very
important proposal in providing aid to the army personnel in case of
dread and emergency situations.
The project results help in locating the combatant and checking
and updating his fitness parameters to chief of armed forces and onto
the army control room. Chief (armed forces) of one troop can
communicate with chief (armed forces) of another troop by using the
NRF module. Combatants in one troop can communicate and also
provide aid in case of any crisis.
The collected data by chief of armed forces will be uploaded on the
cloud to army control room to provide necessary aid to the combatant.
Figure 9.2 shows the communication model of the proposed work
consisting of combatant segment, chief of armed forces segment and
army control room.

Army Control
Room
Cl
d
ou

ou
Cl

Troop 1 Troop 2

Chief of Chief of
Armed NRF Module Armed
Forces Forces
e
ZigBee
ZigBee

Be
ee
Zi
ee

Zig
gB
B

Zig
Zig

ee

Combatant Combatant Combatant Combatant Combatant Combatant

Figure 9.2 Communication Model of the Proposed Work

 CONSPECTUS OF K-MEANS CLUSTERING ALGORITHM 201

Each segment is discussed in the “Design Methodology and

Implementation” section.

Design Methodology and Implementation

Combatant’s Segment This segment has a suitable sensory unit con­

nected by a set of connections. It is used to check the fitness
of combatant, spot his position, and detect any hazards in war zone.
The obtained signals will be converted into required format using
ADC, which is compared with the typical conditional signals. There
is a crisis if incongruity between the two occurs.

Chief of Armed Force’s Segment This segment has data aggregation

facility to collect data from combatant’s unit and remove any re­
dundant data. The aggregated data will be sent to the army control
unit through cloud using K-means clustering algorithm.

Army Control Room’s Segment This segment has a computer system

with cloud infrastructure facility. The data available on cloud will be
displayed on the computer system.
Figure 9.3 depicts the flow chart of location tracking and updating
fitness status of combatant to control room for initiating rescue operation.
Figure 9.4 shows the circuit connection implementation of the
project work carried out by interfacing Arduino Uno, NRF Module,
and Zigbee.

Results

Following are the few sample results of the project. See Figures 9.5–9.8.

Real-World Use Cases of K-Means Clustering

Following are few interesting use cases of K-means clustering

algorithm:
1. Record Classification
It groups archives in various classes depending on labels, sub­
jects, and the substance of the report. This is an extremely
202 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Start

Initiate and
configure GPS

Send location
of the
commandant
via Cloud

Configure Configure
temperature No
Heart No
sensor beat sensor

No
No

If S>=60 No If 68>=h>=85 No
Yes

Yes

Send alert message and current location of the

commandant to the control room via Cloud

Rescue operation initiated

Stop

Figure 9.3 Flow of the Proposed Model

 CONSPECTUS OF K-MEANS CLUSTERING ALGORITHM 203

Figure 9.4 Circuit of the Project Work

Figure 9.5 Display of Commandant Fitness Parameters Like Pulse Rate and Temperature
204 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 9.6 Hardware Output to Track Any Hazardous Conditions in Armed Conflict

Figure 9.7 Display of Gas Sensor Output to Track War Zone Hazards
 CONSPECTUS OF K-MEANS CLUSTERING ALGORITHM 205

Figure 9.8 Display of Heart Beat Tracker

standard order issue and K-implies is an exceptionally appro­

priate calculation for this reason. The underlying preparing of
the reports is expected to speak to each record as a vector and
utilization term recurrence.
2. Conveyance Store Optimization
It improves the procedure of high-quality conveyance utilizing
automobile drone by a blend of K-intends to locate the ideal
figure of dispatch areas and a hereditary calculation to unravel
the automobile course as a voyaging sales rep issue.
3. Recognizing Offense Localities
With information identified with wrongdoings accessible in
explicit territories in a city, the class of wrongdoing, the terri­
tory of wrong doing, and the relationship between the two can
give quality understanding into wrongdoing-inclined zones
inside a town or an area.
4. Client Segmentation
Grouping assists advertisers in improving their client database,
taking a shot at target regions, and fragmenting clients de­
pending on previous purchase, interests, or movement checking.
5. Dream Group Statistics Study
Breaking down participant details has consistently been a basic
component of the donning scene, plus with expanding rivalry,
206 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

AI has a basic task to carry out here. As an intriguing activity,

on the off chance that you might want to make a dream draft
group and similar to recognize comparable troupe dependent
on team member details, K-means is capable of be a valuable
alternative.
What is the issue: Who would it be advisable for you to have
in your group? Which players will perform best for your group
and permit you to beat the opposition? The test toward the
beginning of the period is that there is less information acces­
sible to assist you with distinguishing triumphant players.
6. Detection of Fraud or Criminal Activity
AI has a basic work in extortion finding. This has several ap­
plications in car, medicinal services, and protection mis­
representation location. It uses previous chronicled information
on deceitful scenarios, based on its closeness to groups showing
fake examples. The ability to recognize cheats is predominant
and necessary.
What is the issue: You need to investigate false driving de­
velopment. The test is how might you perceive what is valid and
which is counterfeit?
How batching functions: By probing the global positioning
system details, the computation can gather proportional prac­
tice. Considering the attributes of the gathering, we can ex­
emplify them into authentic and fake.
7. Rideshare Data Analysis
The openly accessible Ola ride dataset gives data on traffic, time
of travel, pickup points, etc. Dissecting this information is
precious with respect to Ola, as well as in providing knowledge
required for smooth travel.
8. Digital Profiling Criminals
Digital profiling is the way toward gathering information from
people and gatherings to recognize critical relationship. The
possibility of digital profile is acquired from unlawful profile,
which gives data on the examination distribution to arrange the
sorts of crooks available at the wrongdoing sight.
9. Call Record Detail Analysis
It uses the data caught by telecom organizations during the call,
SMS, and web movement of a client. These data give more
 CONSPECTUS OF K-MEANS CLUSTERING ALGORITHM 207

noteworthy bits of knowledge about the client’s needs when

utilized with client socioeconomics. We can bunch client ex­
ercises for 24 hours by utilizing the unaided K-implies grouping
calculation. It is utilized to comprehend fragments of clients
regarding their utilization by hours.
10. Programmed Grouping of Information Technology Alerts
Huge endeavor information technology framework research
units create large amount of organized messages. Since these
relate to outfitted issue, they should be physically checked for
priority during procedure execution. Combination of in­
formation can give understanding into classifications of
alarms and intervening to fix, and can help in disappointment
forecasts.
11. Identifying Fake News
Fake news is not a new phenomenon, but it is one that is
becoming prolific.
What the issue is: Counterfeit reports are being prepared and
these reach at a rapid pace for novelty advancements, for in­
stance, online medium. The subject selected was thought as of
tardy during the 2016 U.S. official mission. The phrase false
report was suggested to an extraordinary amount of epoch.
The manner wherein the computation works is by captivating
the stuff of the not genuine report, the body, looking at the
vocabulary utilized, and subsequently gathering them. These
gatherings are what assist the calculation with figuring out which
pieces are authentic and which are phony information. Few
terminologies are exposed usually in sensationalized, misleading
content articles.
12. Spam Channel
You know the junk envelope in your email inbox? It is where
messages that have been distinguished as spam by the calculation.
Many AI courses, for example, Andrew Ng’s renowned
Coursera course, utilize the spam channel to act as an illustration
of unaided learning and grouping.
What the issue is: Spam messages are, best case scenario, an
exasperating piece of cutting edge showcasing methods, and even
from a pessimistic standpoint, an illustration of individual’s
208 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

phishing for his own information. To try not to get these messages
in your fundamental inbox, email organizations use calculations.
The reason for these calculations is to signal an email as spam
effectively or not.
How grouping functions: K-means bunching strategies have
demonstrated to be a compelling method of distinguishing spam.
The mechanism is by taking a gander at the various areas of the
email. The data are then combined.
Then we have the option to arrange the gatherings to differ­
entiate from spams. Recalling grouping for the gathering cycle
improves the accuracy of the channel to 97%. This is a great in­
formation for people who are not passing up #1 announcement
and offers.
13. Showcasing and Business Sales
Marketing and attracting customers in showcasing its product
are essential for a large-scale company.
What the issue is: For sales to get the unsurpassed revenue
for your advancing speculation, it is vital that you intend folks in
the right approach. If you fail to understand the situation, you
hazard not doing any deal, or all the further awful, destroying
your clients’ faith.
Clustering computations can gather together persons with
relative attributes and probability to purchase.
14. Classifying Network Traffic
Imagine you want to understand the different types of traffic
coming to your website. You are particularly interested in un­
derstanding which traffic is spam or coming from bots.
What the problem is: As an ever-increasing number of ad­
ministrations start to utilize APIs on your application, or as your
site develops, it is significant you know where the traffic is
coming from. For instance, you need to have the option to ob­
struct tactless traffic and twofold down on zones driving devel­
opment. Be that as it may, it is difficult to tell which will be with
regards to ordering the traffic.
How bunching functions: K-implies grouping is utilized to
gather qualities of the traffic sources. At the point when the
bunches are made, you would then be able to arrange the traffic
types. The cycle is quicker and more exact than the past auto class
 CONSPECTUS OF K-MEANS CLUSTERING ALGORITHM 209

technique. By having exact data on traffic sources, you can

develop your site and plan limit adequately.
15. Record Examination
There are various reasons why you would need to run an
examination on a report. In this situation, you need to have the
option to sort out the reports rapidly and productively.
What the issue is: Imagine you are restricted as expected and
need to coordinate data held in archives rapidly. To have
the option to finish this task, you need to comprehend the topic
of the content, contrast it and different archives, and order it.
How gathering functions: Hierarchical packing has been
used to take care of this topic. The estimate can take a gander
at the content and assemble it into a variety of matters.
Implementing this procedure will help in gathering, as well
as coordinating, comparable reports rapidly utilizing the
qualities perceived.

K-Means Clustering Algorithm Used as a Solution for Following

Business Problems
Purchaser Division

When you have finished group examination, you can recognize

portions of the shopper. To do this, you can utilize segment, psy­
chographic and conduct information just as execution information to
bunch your purchasers for a specific item class.
Next, you can profile your groups to comprehend your buyers
better and depict them as per the factors utilized for bunch ex­
amination. You can use this data to tailor your showcasing messages,
item collections, and general shopping experience to address your
client’s issues and increment your ROI.

Conveyance Enhancement

Retailers and providers have hoped to enhance their conveyance

procedure utilizing k-implies grouping. The conveyance courses and
examples of trucks and automatons have been checked to locate the
ideal dispatch areas, courses, and goals for the organization.
210 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Archive Arranging and Gathering

You can aggregate electronic documents as per classification, labels,

substance, or recurrence of utilization by K-implies grouping. The
calculation sees each archive as a vector and the recurrence of specific
terms to arrange and aggregate the records.

Client Maintenance

You can utilize K-implies bunching to break down and bunch client
stir to distinguish and profile your buyers depending on maintenance.
You can utilize factors, for example, recurrence of buys, how as of late
the purchaser visited the store, normal spends per outing. and con­
tainer arrangement to examine and anticipate degrees of consistency
of specific client portions.

Rebate Examination

You can utilize K-implies bunching to scrutinize various gatherings

of customers as indicated by their rebate buy practices. Clients might
be bound to buy packaged items, items with an ordinary low-value
technique, or items at a bargain before expiry. You can utilize the
calculation to distinguish buying designs among your customers and
use these data to settle on choices with respect to estimating and
limiting time methodologies.

Comparisons of K-Means Clustering with Other Unsupervised Algorithms

In this section, a comparison of K-means clustering algorithm with

other unsupervised algorithms, namely, hierarchical clustering algorithm
(HC), self-organization map algorithm (SOM), and expectation max­
imization (EM) clustering algorithm is discussed.
Table 9.1 shows comparison of all the above-mentioned four al­
gorithms with respect to performance and accuracy parameters by
keeping number of clusters (K) equal to 32.
Table 9.2 shows comparison of all the above-mentioned four al­
gorithms with respect to data types by keeping the number of clusters
(K) equal to 32.
 CONSPECTUS OF K-MEANS CLUSTERING ALGORITHM 211

Table 9.1 Comparison of SOM, K-Means, EM, and HC Algorithms with Respect to Parameters
Performance and Accuracy Keeping K = 32
PARAMETERS SOM K-MEANS EM HC
Performance 78 84 84 87
Accuracy 830 910 898 850

Table 9.2 Comparison of SOM, K-Means, EM, and HC Algorithms with Respect to Data Types
Keeping K = 32
DATA TYPES SOM K-MEANS EM HC
Random 830 910 898 850
Ideal 798 810 808 829

Table 9.3 Comparison of SOM, K-Means, EM, and HC Algorithms with Respect to Data Size
Keeping K = 32
DATA SIZE SOM K-MEANS EM HC
36,000 830 910 898 850
4000 89 95 93 91

Table 9.3 shows comparison of all the above-mentioned four al­

gorithms with respect to data size by keeping the number of clusters
(K) equal to 32.

Pros and Cons

Pros:
1. If K is small and variables are many, then the speed of com­
putation will be more than the hierarchical clustering.
2. If clusters are globular, it will produce clusters that are tighter
than the hierarchical clustering.

Cons:

1. Prediction of K is difficult.
2. Diverse preliminary partitions will outcome in diverse con­
cluding clusters.
212 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

3. It will not operate well with clusters of dissimilar dimension and

dissimilar concentration.

Conclusions and Future Work

Bunch investigation issue has consistently captivated researchers

as it bargains with gathering of articles having normal properties.
K-means is a basic and adaptable algorithm to trial when you are
beginning with grouping. K-means grouping is one of the most
popular clustering algorithms and usually the main thing practi­
tioners apply when settling batching tasks to get an idea of the
structure of the dataset. The goal of K-means is to aggregate data
centers into distinct non-overlapping subgroups.
By and large, K-means calculation is done with Euclidean separation
to acquire the groups. K-means calculation is simple to actualize, however,
not productive for covering informational indexes. Test perceptions have a
place with measurable populaces. One method of tackling this issue is to
present factual separation quantifies rather than Euclidean separation.

References
[1] Tajunisha and Saravanan. “Performance Analysis of k-Means with
Different Initialization Methods for High Dimensional Data.
International Journal of Artificial Intelligence & Applications (IJAIA),
vol. 1, no. 4, October 2010.
[2] N. Bouhmala, A. Viken, and J. B. Lonnum. “Enhanced Genetic
Algorithm with K-Means for the Clustering Problem.” International
Journal of Modeling and Optimization, 2015, pp. 150–154.
[3] Rouhollah Maghsoudi, Arash Ghorbannia Delavar, Somayye
Hoseyny, Rahamatollah Asgari, and Yaghub Heidari. “Representing
the New Model for Improving K-means Clustering Algorithm Based
on Genetic Algorithm.” The Journal of Mathematics and Computer
Science, 2011, pp. 329–336.
[4] A.P. Jyothi and Usha Sakthivel. “Trends and Technologies Used for
Mitigating Energy Efficiency Issues in Wireless Sensor Network.”
International Journal of Computer Applications. vol. 111. 2015,
pp. 32–40, 10.5120/19521-1150.
[5]
[6] Sandeep Rana, Sanjay Jasola, and Rajesh Kumar. A Hybrid
Sequential Approach for Data Clustering Using K-Means and
Particle Swarm Optimization Algorithm. International Journal of
Engineering, Science and Technology, vol. 2, no. 6, 2010, pp. 167–176.
 CONSPECTUS OF K-MEANS CLUSTERING ALGORITHM 213

[7] A. P. Jyothi, and Usha Sakthivel. “CFCLP – A Novel Clustering

Framework Based on Combinatorial Approach and Linear
Programming in Wireless Sensor Network.” In 2017 2nd International
Conference on Computing and Communications Technologies (ICCCT),
pp. 49–54. IEEE, 2017.
[8] Bara’a Ali Attea. A Fuzzy Multi-objective Particle Swarm Optimization
for Effective Data Clustering. Springer, July 2010, pp. 305–312.
[9] A. P. Jyothi, and S. Usha. (2019). MSoC: Multi-scale Optimized
Clustering for Energy Preservation in Wireless Sensor Network.
Wireless Personal Communications. 10.1007/s11277-019-06146-y.
[10] Chetna Sethi and Garima Mishra. A Linear PCA Based Hybrid K-
Means PSO Algorithm for Clustering Large Dataset. International
Journal of Scientific & Engineering Research, vol. 4, no. 6, June 2013,
pp.1559–1566. International Journal of Computer Applications (0975 –
8887), vol. 156, no. 8, December 2016.
[11] Rui Xu and Donald Wunsch. Survey of Clustering Algorithms. IEEE
Transactions on Neural Networks, vol. 16, no. 3, May 2005.
[12] A. P. Jyothi, and Usha Sakthivel. EPCA: Energy Preservation Using
Clustering Approximation in Sensor Network. ERCICA-2016
Springer Book, pp. 547–557. 10.1007/978-981-10-4741-1_47.
[13] C. Xiaohui, E. P. Thomas, and P. Paul. Document Clustering Using
Particle Swarm Optimization. Swarm Intelligence Symposium,
pp. 185–191, Pasadena, CA, USA, 2005.
[14] A. P. Jyothi, and S. Usha Interstellar Based Topology Control
Scheme for Optimal Clustering Performance in WSN. Int J Commun
Syst., vol. 33, 2020, p. e4350. https://doi.org/10.1002/dac.4350.
[15] Osama Abu Abbas, Comparisons between Data Clustering
Algorithms. The International Arab Journal of Information Technology,
vol. 5, no. 3, July 2008.
 10
SYSTEMATIC APPROACH TO
DEAL WITH INTERNAL
FRAGMENTATION AND
ENHANCING MEMORY SPACE
DURING COVID-19

MOHAN APARNA, R. MAHESWARI,

AND J. V. THOMAS ABRAHAM
SCOPE, Vellore Institute of Technology,
Chennai, Tamil Nadu, India

Contents

Introduction 216
Internal Fragmentation 216
Best-Fit Approach and Memory Bank 216
Linear and Lasso Regression 217
Literature Survey 217
Dynamic Memory Allocation and Internal Fragmentation 217
Internal Fragmentation Standards and Compatibility of
Systems 218
Processor Allocation and Internal Fragmentation 218
ML and Cloud Storage 219
Best-Fit Approach and Internal Fragmentation 219
Storage and Dependability Study 220
Lasso Regression Analysis and Cloud Computing 220
Data Placement and Machine Learning 221
Memory Resource Management 221
Existing Technique 222
Overview of the Existing Technique with Illustration 222
Proposed Technique – Optimized Approach to Deal with
Residual Space in Internal Fragmentation 223
Flow Chart of the Proposed Technique 223
Schematic of Memory Bank Approach 225

DOI: 10.1201/9781003119838-10 215

216 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Memory Bank Algorithm 226

Linear Regression 229
Lasso Regression 229
Implementation Using Open Source Software – Scilab and
Jupyter (Python) 230
Pseudo Code of the Proposed Technique 230
Using Scilab 230
Using Jupyter (Python) 231
Graphical Representation 231
Inference 1 231
Inference 2 232
Inference 3 232
Software Packages Used 234
Results and discussions 234
Conclusion and Future Work 236
References 237
Appendix 10.A 240

Introduction
Internal Fragmentation

The world is jeopardized by technology. With the escalating technology,

there is a demand of storage space in the operating system. For the users
and operating system developers, the system will run efficiently only if
there is an efficient way to manage the memory. There is a frequent need
to move and remove the processes from and to the memory. While doing
so, there exists the issue of fragmentation. After some point of time, it
becomes difficult for the system to accommodate processes into the
memory blocks. One of the significant causes for internal fragmentation
is the allocation of large memory space for a process that is smaller in size.
This leads to empty memory spaces during contiguous memory alloca­
tion. Internal fragmentation can be best viewed as the difference between
the desired memory size and the size of the memory that is allocated.

Best-Fit Approach and Memory Bank

The main reason for internal fragmentation is the memory block in

comparison with the size of the process to be accommodated is
 SYSTEMATIC APPROACH WITH FRAGMENTATION 217

bigger. In this scenario, a portion of the memory is left unused and

other processes cannot occupy this memory space as well. A memory
bank can be created in the memory management system, which
would resemble the main memory, but it is a virtual memory that is
also divided into blocks. When a process occupies a memory block
using best-fit approach, the residual memory can be moved to this
memory bank. Accumulation of such residual spaces will be enough
for any other process to occupy the space in the memory bank. If the
space is found to be adequate to accommodate one process, then that
process can be stored in this memory space. Only if the size of the
process is the same as the size of the memory block, a process is
stored. The remaining space in the memory bank is given to the next
block to accommodate another process in the queue. Hence, this
proves to be an efficient and optimized way to prevent the wastage of
memory space in internal fragmentation and also solves the lack of
storage issue. Hence, this research focuses on preventing the wastage
of memory space, which occurs in internal fragmentation.

Linear and Lasso Regression

Linear regression is unpretentious and most widely used analytical

method for predictive modeling. It provides an equation, where the
features are categorized as dependent variables, on which our target
variable depends, and independent variables. Lasso regression could
also be a kind of linear regression that uses shrinkage. The Lasso
procedure encourages sparse models. Lasso means least absolute
shrinkage and selection operator. Lasso regression aims at minimizing
the prediction errors by implementing certain constraints on the model.

Literature Survey
Dynamic Memory Allocation and Internal Fragmentation

In the field of network and communications, the need for dynamic

memory allocation is escalating at a greater pace. According to
Stylianos Mamagkakis et al. the real-time operating systems use
state-of-the-art dynamic allocators [1]. The five types of memory al­
locators used so far have improved the dynamic memory allocation
218 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

process but failed to provide a permanent solution for internal

fragmentation. The existing memory allocators don’t fully serve the
purpose of allocating different sizes of memory blocks for the wired
and wireless applications in network. This process is prone to internal
fragmentation and also causes heavy traffic in run-time memory re­
quests, which is due to its reliance on dynamic memory allocation, as
stated by the author Voss M. [2]. Hence, the efficiency, performance,
and de-fragmentation techniques of these existing allocators in the
world of operating system are sparse [3].

Internal Fragmentation Standards and Compatibility of Systems

Christian Del Rosso states that the compatibility of a system when

internal fragmentation occurs is controlled by two significant sources,
which are configurational and competitive. When a study was done
within the DVB-H (Digital Video Broadcasting – Handheld) and T-
DMB (Terrestrial-Digital Multimedia Broadcasting) mobile digital
multimedia broadcasting standards, it was observed that the systems
were not compatible with the fragmentation issue that occurred.
Techniques like adoption of a single standard and internal standards
fragmentation were proposed [4]. This was found to be prevalent in the
world of mobile-operating systems as well. The issue of compatibility
further reduced the performance of the system. Developers found a
delay in the loading of addresses and writing to the memory [5]. The
compatibility issue was solved using optimized systems and new
techniques of memory management in the later stages of development.

Processor Allocation and Internal Fragmentation

Several strategies were introduced to limit the rate of fragmentation

in the systems. One such strategy is Adaptive Non-Contiguous
Allocation algorithm. This works on the basic principle of partition
binding and allocating processors. The over-spitting of requests in
parallel conditions is also avoided using this algorithm. According to
author Bani Ahmad, an optimal value for partition binding is chosen,
which will make this algorithm more flexible and tuneable. Using
this approach, parallel jobs can be allocated by the processors quickly
based on the partitioning bound values [6]. This strategy was found
 SYSTEMATIC APPROACH WITH FRAGMENTATION 219

to be compatible with the high loads in communication in com­

parison to other existing techniques of non-contiguous memory al­
location [7]. Contiguous memory allocations fail to reduce the rate of
internal fragmentation when larger implementations are considered.

ML and Cloud Storage

The existing machine learning (ML) algorithms used for securing data
would categorize a node as an unusual node depending on temporary
behavioral data. These systems do not distinguish if an unusual node is
a malevolent node or a broken node. Hence for solving this insecurity,
the work proposes a long short-term memory (LSTM) model [8], which
initially learns the behavior of a user and automatically trains itself and
stores the behavioral data. The model can easily categorize the user
behavior as standard or non-standard. It can also find if a non-
standard node is a broken node or a new user node or a tampered node
using the calculated trust factor. The proposed technique detects the
security attack and safeguards the cloud.

Best-Fit Approach and Internal Fragmentation

According to Rachael Chikoore et al., the genetic algorithms

proved as an efficient way to reduce the internal fragmentation
caused in the memory management of a system [9]. To produce
an optimal configuration, the workload was used as an input.
Rational approach was adopted to find the solutions through genetic
algorithms [10]. One of the optimized approaches was best-fit ap­
proach. In this approach, the process size was observed to be greater
than the memory size. The internal memory fragmentation was ef­
fectively solved using this approach but there was some wastage of
memory involved in this process. Though usage of this approach was
beneficial, it had its cons when considering evolution, complexity, and
adaptability. It also paved way for providing an optimized configura­
tion parameter for a system for free segregated lists. When compared
to the historical “brute-force” technique, the best-fit approach was
found to be a more heuristic, feasible, and optimal way to deal with
internal fragmentation.
220 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Storage and Dependability Study

In today’s world, the escalating need for data storage systems (DSS) has
been turning out as a nightmare as they are attacked by soft errors that
are often found in controllers used for storage. The functionalities and
stack comprising hardware and software, detection, and correction of
errors in DSS stand out when compared to the architectures that follows
general-purpose computing. The existing systems or techniques did not
address the system-level effects of these soft errors on DSS. This study
gives a clear perspective of the effects of the errors in the controllers used
for storage and the whole of system dependability is analyzed. In the
proposed technique [11], several functions related to storage controllers
operating on a full stack of storage system were implemented and a new
framework was developed. The impact of soft error in storage systems is
accurately calculated using the metric storage system vulnerability factor
(SSVF). The results of this analysis helped to draw conclusions that
there is data loss when 40% of cache is occupied by user data, and it
differs according to the storage controller configuration. It is concluded
through the study that data unavailability is as a result of errors which
are detectable and cannot be recovered in the cache memory.

Lasso Regression Analysis and Cloud Computing

Cloud computing provides services as and when the need arises for
the users. All the resources in cloud are stacked in the data centers,
which give rise to optimal resource allocation methods. The existing
solutions include greedy algorithms and ML algorithms. As greedy
algorithms were NP-hard in nature, it couldn’t provide the best
solution, but ML provides nearly optimal solution. In the proposed
technique [12], ridge regression and lasso regression is im­
plemented on the resource needs of the user dataset. It was con­
cluded that the proposed algorithms gave better results than linear
regression because of the correct selection of features and ridge
regression has the ability to solve multicollinearity issue. The ana­
lysis of the models is done by obtaining root mean squared error
(RMSE) and the graphs are plotted. It was observed that the
proposed algorithms aided virtual machines to get drain in terms of
CPU utilization and storage.
 SYSTEMATIC APPROACH WITH FRAGMENTATION 221

Data Placement and Machine Learning

As cloud computing usage is accelerating, several applications re­

quire higher performance, which creates a demand on the storage
space available. The data obtained from various sources will contain
different frequencies of read/write, retention time, and data of
different sizes. A solution to this is hybrid storage system. The
underlying issue was the placement of data in hybrid storage based
on run-time status and several other properties of data and the
system used for storage. In this work [13], Archivist – an ML-
assisted data placement mechanism for hybrid storage systems – is
introduced to minimize the file access latency. To know the access
patterns of the input data, ML approach is followed. Furthermore,
a data placement algorithm is implemented with an aim to enhance
the data on the hybrid storage mediums. This is done by exactly
finding out similar properties of data and storage medium features.
The experimental results help in drawing conclusions that the
performance of the system is improved up to 49% for accessing files
compared to baseline.

Memory Resource Management

To increase the performance of a system, the number of cores present

should be increased. Several studies indicate that memory manage­
ment has been a consistent problem which caused a deprecating
performance especially in Android devices. The existing techniques
focuses on improving efficiency by more allocation of resources and
very few studies focus on dealing with internal fragmentation. In the
proposed technique, intra-thread behavior [14–34] was focused. In
consequence, a memory management based on thread behavior
(MMBT) was introduced. But it was found that the proposed
technique lacked unified optimization program interface and good
architecture. To solve this issue, thread-oriented memory management
layer (TOMML) was proposed. This uses the microkernel archi­
tecture pattern and was able to fulfil the user’s need for selecting
plug-ins to attain diverse optimization goals. Initially the efficiency of
TOMML was observed through theoretical simulation and experi­
mentation. As a result, the percentage of memory allocation
222 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

efficiency was increased by 12–20%. In comparison to previous ex­

perimental studies, this technique also saves power by allowing the
user to select the plug-in.

Existing Technique
Overview of the Existing Technique with Illustration

Internal fragmentation has been a persistent issue faced in the world

of operating system. Best-fit approach has been adopted for decades
and accepted as the best solution for internal fragmentation.
Consider two processes, P1 and P2, of size 25 MB and 30 MB,
respectively, as shown in Figure 10.1. The process goes from the
secondary memory to the main memory and occupies a block in the
main memory. P1 and P2 are accommodated in 1000 MB of space
where Process 1 occupies only 25 MB and process P occupies only 30
MB. In each of the above cases, 475 MB of memory space is being
wasted in fragment 1 occupied by P1 and 470 MB of space is
being wasted in fragment 2 where P2 is accommodated. To combat
this wastage of space, theories were proposed and researchers found
ways to make use of the block of space in the memory which is large
enough to accommodate the process but not the largest available
memory space. This gave rise to the method followed till date called
the Best-Fit approach. In this approach, the process of allocating the

Figure 10.1 Internal Fragmentation

 SYSTEMATIC APPROACH WITH FRAGMENTATION 223

smallest partition but large enough for the process is adopted.

However, the problem will not be resolved entirely but can be mi­
tigated to some extent. The illustration of the existing system in
Figure 10.1 also resulted in the lack of ability of the system to store
the incoming processes in the main memory as a result of sub­
stantial or no available memory space. The best-fit approach en­
sured an immediate but not an optimized solution to the problem
of internal fragmentation, faced by the operating system developers
for several years. Here, different colors are used for the purpose of
illustration to distinguish between the wasted memory space and
the block of space allocated to the processes P1 and P2. Hence, this
research work provides an optimized approach to avoid this wastage
of memory, thereby finding an effective solution for internal
fragmentation.

Proposed Technique – Optimized Approach to Deal with Residual Space

in Internal Fragmentation
Flow Chart of the Proposed Technique

The flow chart depicted in Figure 10.2 gives a clear perspective of

the ptechnique of memory bank approach. Initially, the process is
taken as input. When there is a request to move the process from
the secondary memory to the main memory, the process first goes to
the page table. The Page table maps the logical address and the
physical address. The valid and invalid bits are set as “v” and “iv,”
respectively, if the page is found and when the page is not found in
the memory. In the main memory, the processes are stored ac­
cording to the best-fit block approach. After implementing the
approach, if there is some residual space left in the memory blocks,
then that space is accumulated and goes to the memory bank. The
memory bank which can be viewed as an instance of main memory
can be used to store the incoming processes which cannot be fitted
into the main memory. In the memory bank, space is allocated
according to the size of the incoming process. This process is re­
peated and the rest of the process in the queue, waiting for the
allotment of a memory block, can be easily stored in the subsequent
blocks in the memory bank. The sequential execution of the
224 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 10.2 Flow Chart Depicting the Proposed Technique

 SYSTEMATIC APPROACH WITH FRAGMENTATION 225

memory bank technique as depicted in Figure 10.2 also enhances

the memory management and the general working of a specific
operating system. The space allocation in the proposed approach is
terminated once the memory bank is full and has no more blocks of
space to further accommodate the processes. The residual space
computation is an indication of the available space in the memory
bank. This helps the developer to determine if the threshold ca­
pacity is attained. It also proves to be useful in finding out if there is
a possibility to store the further incoming processes in the memory
bank. Thus, the technique helps in the prevention of wastage of
space after implementing best-fit approach and no process is left
behind without allotting a specific memory block.

Schematic of Memory Bank Approach

Memory management plays a significant role for an operating system

to function effectively. Initially, when the user intends to move a
process from the secondary memory to the main memory, the process
moves to the page table. Page table helps in mapping the logical
address with the physical address. When a process requests to access
the data in the main memory, the operating system stores the
mapping of virtual addresses to physical address. The operating
system takes the help of Page table to execute this, as it stores the
mappings. Each mapping is known as “page table entry”(PTE),
which is referred as page number.
The process, if requested, goes to the memory. Here, the pro­
cesses are prone to undergo fragmentation. In the schematic dia­
gram depicted in Figure 10.3, “internal fragmentation” occurs.
Internal fragmentation is viewed as difficult to regain or retrieve
compared to other. The process here is larger than the memory
space but the fragmentation is solved using the existing “best-fit
Approach.” During this approach, the process occupies the memory
block that is equal to or greater than the size of the process. In
Figure 10.3, three processes occupy memory space, which is greater
than the size of the process. Hence, there exists residual space in the
blocks of main memory. These residual blocks are collected and
stored in a replica of the main memory called memory bank. Memory
226 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 10.3 The Proposed Technique is Depicted in the Form of Schematic Diagram

bank can be best described as a virtual instance of memory which

can hold the processes. These processes can be retrieved as and
when needed by the developer. The memory bank depicted here is
used to store the residual space which has occurred due to internal
fragmentation after implementing the best-fit approach. Once the
main memory is full, there occurs a lack of storage issue. Hence, the
incoming processes are not able to get stored in the main memory.
This memory bank approach can be effectively used to resolve the
storage issue as well. The incoming processes which are not able to
accommodate in main memory is pushed to this memory bank.
Now, the residual space which we have already stored in the
memory bank can be used to store these processes. Here once the
process size is equal to the memory block size, the storage is stopped
and the rest of the space is used for the next incoming process. This
approach has now resolved the wastage of memory which prevailed
in the existing technique of the best-fit approach in an optimized
manner. It has also increased the performance of operating system
through effective memory management.

Memory Bank Algorithm

Input: psize_in_GB (Process size), process_name, p_no (Process

number), frame_no, vi, memory_block_size
 SYSTEMATIC APPROACH WITH FRAGMENTATION 227

Output: Remaining space in memory bank, graphical representa­

tions of Best-fit and Memory bank approaches with process size and
process no.

Algorithm 10.1 bestfitfunction(psize_in_GB, memory_block_size,

size_process, size_block)

bf_allot is set to zeros (1,size_process);

bf_flag is set to zeros (1,size_block)
for i=1:size_process
k is set to -1;
for i=1:block_size
if psize_in_GB(i) is less than or equal to memory_block_size(j) &&bf_flag(j)
is 0 then
if k==-1 then
k is set to j;
elseif(memory_block_size(j) is less than memory_block_size(k))
k is set to j;
end if
end if
end for
if(k is -1)
bf_allot(i) is set to 0; // if no block can accommodate the process, set allotted
block number as 0
else
bf_allot(i) is set to k; // store the selected index in the best fit array in the index
position i
bf_flag(k) is set to 1;// set the status of the selected block as busy endfunction

Algorithm 10.2 frag_memorybankfunction(psize_in_GB, pro­

cess_name, p_no, frame_no, vi,memory_block_size, total_size)

Print processes in the secondary memory;

Print page table entries;
for i =1: n
if psize_in_GB(i) is equal to memory_block_size(j) then
Print “process i stored in memory successfully without fragmentation”;
break;
228 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

elseif psize_in_GB(i) is less than memory_block_size(j) then

Print“Internal fragmentation occurred memory is allocated using best-fit
approach”;
residual_space is memory_block_size(j)-psize_in_GB(i);
end elseif
end if
end for
memory_bank_size is residual_space+memory_block_size(i)
if total_size is greater than 15 then, Greater than RAM size
Print “ RAM size is full. Storage Unavailable! Cannot store process”;
end
if psize_in_GB(i) is less than or equal to memory_bank_size then
Print “Process i of size fitted into the memory bank successfully”;
end
mb_size is memory_bank_size-psize_in_GB(i);
return residual_space_in_memory_bank;
endfunction

Algorithm 10.3 function display(allot, allotsize, size_process,

psize_in_GB)

for i=1:size_process, printing process no., process size, block no., block size
if allot(i) is 0 then
Print “Not Allocated” with psize_in_GB(i)
else
Print(i,psize_in_GB(i),allot(i),allotsize(i))
end
end
endfunction

Algorithm 10.4 Main_function

size_process is size(psize_in_GB); Determining the number of processes and

blocks
size_process is size_process(2);
size_block is size(memory_block_size);
size_block is size_block(2);
 SYSTEMATIC APPROACH WITH FRAGMENTATION 229

To call the function, defined in best fit.sci, for best fit allocation
bf_allot = bestFit(psize_in_GB, memory_block_size, size_process, size_block)
bf_allotsize is zeros(1,size_process), bf_allotsize - size of the selected blocks
for best fit
for i=1:size_process, storing the allocated block size for each process
if bf_allot(i) is not 0 then, checking if any block is selected
bf_allotsize(i) is set to memory_block_size(bf_allot(i)), storing the size of the
selected block for best fit
else
bf_allotsize(i) is set to 0, store size as 0 if no block is selected
end
end
func = frag (psize_in_GB,process_name,p_no,frame_no,vi,memory_block_
size,total_size);
Plot the 2d graph with plot2d() for best-fit and memory bank approach

Linear Regression

Algorithm 10.5

Determine the columns x, y as dependent and independent variables.

regression_model = LinearRegression()
regression_model.fit(x,y)
y_predicted = regression_model.predict(x)
evaluate the model
rmse = mean_squared_error(y, y_predicted)
r2 = r2_score(y, y_predicted)
print ‘Slope:’, regression_model.coef
print ‘Intercept:’, regression_model.intercept
print ‘Root mean squared error: ’, rmse
print ‘R2 score: ’, r2
Plot the graph with the regression line

Lasso Regression

Algorithm 10.6

Determine the dependent and target variables in the Data frame

model = Lasso(alpha=1.0)
230 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

model evaluation method:

cv = RepeatedKFold(n_splits=10, n_repeats=3, random_state=1)
evaluate model:
scores = cross_val_score(model, x,y, scoring=‘neg_mean_absolute_error’,
cv=cv, n_jobs= (-1))
force scores to be positive:
scores = absolute(scores)
print ‘Mean Absolute Error (MAE):’ mean(scores), std(scores)

Implementation Using Open Source Software – Scilab and

Jupyter (Python)
Pseudo Code of the Proposed Technique

Using Scilab

i. Declares the process size (in GB), process number, and process
name and assigns values initially.
ii. Displays the processes in the secondary memory.
iii. Determines the number of processes and blocks and store them
in size_process and size_block respectively.
iv. Best-Fit func() is defined and called with input parameters as
psize_in_GB, memory_block_size, size_process, size_block.
v. Declares a best-fit array (here bf_allot is used), and a flag array
bf_flag.
vi. Allocates blocks according to the best-fit approach using k as an
index variable to indicate the smallest block which can ac­
commodate a process. Initially, this value is set to zero.
vii. If (p_size_in_GB is less than the memory_block_size) then:
update k with the index position of the smaller block
Else set k as the index position of the block.
If no block can accommodate the process, set allotted block
number as 0.
viii. Store the selected index in the best-fit array in the index po­
sition i (process number) and set memory_block as busy.
ix. Call the function ‘internal_frag()’ with input parameters as
psize_in_GB,process_name,p_no,frame_no,vi,memory_block_
size,total_size.
x. Display the page table entries with valid and invalid bits.
 SYSTEMATIC APPROACH WITH FRAGMENTATION 231

xi. if (process size = = memory block size), process is stored in the

main memory successfully.
else if (block size is greater than the process size)
Print (‘Internal Fragmentation has occurred’)
if (memory block size is lesser than process size),
Print(‘Process cannot be stored in the memory block’).
xii. Compute the space in the memory bank. This is added with the
available space in the memory block that could not accom­
modate a process.
xiii. Return the space available in the memory bank.
xiv. If incoming process size is equal to or lesser than the
size in memory bank, then the process is stored in the
memory bank.
xv. The best-fit approach and memory bank approach are plotted as
graphs.

Using Jupyter (Python)

i. Load the dataset in the environment.

ii. Select the columns which are of interest in the data frame (i.e.)
the dependent variable, in this case revenue and target variable
type of industry.
iii. Initialize the linear regression and lasso regression models.
iv. Determine the y_predicted value for each of the models.
v. Evaluate the predicted model. In this case, root mean square
error is used for linear regression and mean absolute error for
lasso regression is computed.
vi. Print the error values after evaluation of the models and the y-
intercept.
vii. Plot the obtained linear regression in the form of a graph with
revenue of the company on the x axis and type of industry on
the y axis which is a binary value. Along the y axis, “0” denotes
small-scale industries using cloud and “1” denotes developed
industries using cloud.

Graphical Representation

Inference 1 In Figure 10.4, best-fit block approach is plotted using

the open source software Scilab. It is inferred that Processes 1, 2, 3,
232 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 10.4 Best-Fit Block Approach

and 4 are stored in the main memory according to the size of the
processes as shown in the y axis. Size of Process 4 is 5 GB but it is
stored in the memory block of size 7 GB. The residual space 2 GB
is wasted after implementing the best-fit block approach. Process 5
is not stored in the memory as there is no memory block size
which is greater than or equal to the size of the process.

Inference 2 From Figure 10.5, it is inferred that the residual space

after best-fit block approach is plotted as 2 GB for Process 2. Process
5 which was not plotted in the Figure 10.4 is now plotted as
the residual space 2 GB and the available memory block in main
memory 1 GB sums up to 3 GB. This memory size will be available
in the memory bank and Process 5 is plotted as it is stored in the
memory bank. Other processes are plotted with size 0 as they are
already stored in the memory using best-fit approach, as seen in
Figure 10.5.

Inference 3 From Figure 10.6, it can be inferred that the revenue of

an industry also plays as a crucial factor in adopting cloud infra­
structure facilities. Here the graph is plotted considering the type
of industries as two categories. The developing industries are de­
noted as “0” and the developed industries as “1” represent along the
 SYSTEMATIC APPROACH WITH FRAGMENTATION 233

Figure 10.5 Memory Bank Approach Displaying the Residual Space Available

Figure 10.6 Types of Industries Still Lagging in Adopting Cloud Infrastructure

y axis. The revenue of an industry is depicted in million dollars

along the x axis. From the graph depicted below, it can also be
observed that the large-scale companies are adopting cloud infra­
structure majorly compared to the developing industries due to the
234 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

factors like pandemic situation in certain regions, revenue, and size

of the industry.

Software Packages Used

1. Sklearn:
• The package sklearn in python helps in implementing many
ML algorithms.
• In this work, modules like train_test_split, DecisionTree-
Classifier, and accuracy_score is used.
2. NumPy:
• For faster and efficient arithmetic calculations, the numeric
python module NumPy is used.
• Large number of numpy arrays can be read and manipula­
tions are also performed using this package.
3. Pandas:
• This package is used for the purpose of reading from and
writing to different files.
• Data frames are helpful in performing manipulations.
4. RStudio – Scatter plot
To visualize a scatter plot, this package is used implementing
using the function plot (x, y).
5. Scilab – plot2d

plot2d is used to plot 2-D curves.

Results and discussions

Figure 10.7 shows the results obtained after implementing the

best-fit block approach. The list of processes in the secondary
memory, main memory, and page table entries is also displayed.
Figure 10.8 gives a clear perspective of the structure of the memory
bank. The residual space that is accumulated after best-fit block
approach and the space available after the process entering the
memory bank are displayed. The obtained results clearly show that
 SYSTEMATIC APPROACH WITH FRAGMENTATION 235

Figure 10.7 Output Displaying Secondary Memory, Main Memory, and Processes after
Implementing Best-Fit Block Approach

Figure 10.8 Output Displaying the Memory Bank Approach and the Space in the Memory Bank

the optimized technique is the memory bank approach as no space

issue was encountered. The residual space from the best-fit ap­
proach was reused effectively to store another process. Around 80%
of the number of processes were only stored using the existing
approach; hence, the process which is not stored even after using
the best-fit approach is stored in memory bank. The space wasted
in the previous technique is observed to be 20% which is again re-
used in the memory bank to store a process which required 20%
of the available storage space that is a 3-GB process was stored
using the technique of memory bank. Figure 10.9 gives a clear
perspective about the root mean square error obtained after pre­
diction of the model using linear regression as 0.195. Figure 10.10
236 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 10.9 Output Displaying the Results Obtained from the Evaluation of Linear Regression
Model

Figure 10.10 Results Obained from Lasso Regression

depicts that overfitting can be greatly avoided using lasso regression

as the mean absolute error is computed as 0.478 which is negli­
gible. Hence, the predicted two models are giving efficient results
using linear and lasso regression.

Conclusion and Future Work

The problem encountered in the best-fit approach technique is suc­

cessfully resolved using the proposed technique of using a memory
bank. This technique serves as an optimal way to deal with the residual
space obtained in internal fragmentation. Moreover, the prevailing
issue faced by the developers with the memory management that is the
storage issue is also effectively resolved. The memory space in the
memory bank is being re-used for the incoming processes, which is not
stored in the main memory. The future work of the proposed tech­
nique lies in the fact of how far it is used by the developers to enhance
memory management in smart phones. It can be further extended to
mobile operating systems (MOS) to solve the rising sparse in space issues
and would serve as the best solution for internal fragmentation. The
size of the memory bank can further be extended in the future ac­
cording to the design of the operating system. The proposed technique
also improves the system without any deterioration in the existing data.
From the ML predictions, it can be clearly observed that the devel­
oping industries, due to various external factors like lack of revenue,
 SYSTEMATIC APPROACH WITH FRAGMENTATION 237

location and size of the company during the pandemic situation, are
not majorly adopting for cloud infrastructure. Hence, in this COVID-
19 situation, the proposed technique of memory bank serves to
solve the memory storage issue considering even the small-scale and
developing industries.

References
[1] S. Mamagkakis, Christos Baloukas, David Atienza, Francky
Catthoor, Dimitrios Soudris, José Manuel Mendias, and Antonios
Thanailakis. 2005. “Reducing Memory Fragmentation with
Performance-Optimized Dynamic Memory Allocators in Network
Applications.” In: T. Braun, G. Carle, Y. Koucheryavy, and V.
Tsaoussidis (eds) Wired/Wireless Internet Communications. WWIC
2005. Lecture Notes in Computer Science, vol. 3510. Berlin,
Heidelberg: Springer,.
[2] M. Voss, R. Asenjo, and J. Reinders. “Scalable Memory Allocation.”
In: Pro TBB. Berkeley, CA: Apress, 2019.
[3] Nilesh Vishwasrao and Prabhudev Irabashetti. 2014. Dynamic
Memory Allocation: Role in Memory Management.” International
Journal of Current Engineering and Technology, vol. 4, no. 2,
April 2014.
[4] Christian Del Rosso. “Reducing Internal Fragmentation in
Segregated Free Lists Using Genetic Algorithms.” In Proc. of the 2006
international workshop on Workshop on interdisciplinary software en­
gineering research (WISER ’06). Association for Computing
Machinery, New York, NY, USA,. 2006.
[5] D. Atienza, S. Mamagkakis, and F. Catthoor, et al. “Dynamic
Memory Management Design Methodology for Reduced Memory
Footprint in Multimedia and Wireless Network Applications.” In
Proc. of IEEE/ACM DATE, 2004.
[6] S. Bani-Ahmad. 2011. “Processor Allocation with Reduced Internal
and External Fragmentation in 2D Mesh-Based Multicomputers.”
Journal of Applied Sciences, vol. 11, 2011, pp. 943–952.
[7] Lae Wah Htun, Moh Moh Myint Kay, and Aye Aye Cho. 2019.
Analysis of Allocation Algorithms in Memory Management Published in
International Journal of Trend in Scientific Research and Development
(ijtsrd), ISSN: 2456-6470, vol. 3, no. 5, August 2019.
[8] T. Nathezhtha and V. Yaidehi. “Cloud Insider Attack Detection
Using Machine Learning.” 2018 International Conference on Recent
Trends in Advance Computing (ICRTAC), Chennai, India, 2018,
pp. 60–65, doi: 10.1109/ICRTAC.2018.8679338.
[9] Rachael Chikoore, Okuthe P. Kogeda, and Manoj Lall. 2015. “An
Optimized Main Memory Management Dynamic Partitioning
238 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Placement Algorithm.” Pan African Conference on Science, Computing

and Telecommunications (PACT), July 27–29, Kampala, Uganda, 2015.
[10] D. Koch, C. Beckhoff, and J. Teich. 2009. “Minimizing Internal
Fragmentation by Fine-Grained Two-Dimensional Module Placement
for Runtime Reconfigurable Systems.” 17th IEEE Symposium on Field
Programmable Custom Computing Machines, Napa, CA.
[11] M. Kishani, M. Tahoori, and H. Asadi, “Dependability Analysis of
Data Storage Systems in Presence of Soft Errors.” In: IEEE
Transactions on Reliability, vol. 68, no. 1, pp. 201–215, March 2019,
doi: 10.1109/TR.2018.2888515.
[12] Harshala Shingne, Sehar Sountharrajan, M. Karthiga, and Rajan
Chinnasamy. “Lasso and Ridge Regression for Optimized Resource
Allocation in Cloud Computing.” Journal of Advanced Research in
Dynamical and Control Systems, vol. 12. 2020, pp. 1740–1747. 10.
5373/JARDCS/V12I2/S20201215.
[13] J. Ren et al. “Archivist: A Machine Learning Assisted Data
Placement Mechanism for Hybrid Storage Systems.” In: 2019 IEEE
37th International Conference on Computer Design (ICCD), Abu
Dhabi, United Arab Emirates, 2019, pp. 676–679, doi: 10.1109/
ICCD46524.2019.00098.
[14] Z. Zhu, F. Wu, J. Cao, X. Li, and G. Jia. “A Thread-Oriented
Memory Resource Management Framework for Mobile Edge
Computing.” IEEE Access, vol. 7, pp. 45881–45890, 2019, doi: 10.
1109/ACCESS.2019.2909642.
[15] D. Atienza, S. Mamagkakis, et al. “Modular Construction and Power
Modelling of Dyn. Mem. Managers for Embedded Systems.” In:
Proc. of LNCS PATMOS, 2004.
[16] C. Williamson. “A Tutorial on Internet Traffic Measurement.” Proc.
of IEEE Internet Computing, vol. 5, no. 6, p. 70–74,2001
[17] Eran N. Ben-Porath. “Internal Fragmentation of the News.”
Journalism Studies, vol. 8, no. 3, 2007.
[18] M. Norton. “A Structural Hermeneutics of The O’Reilly Factor.” Theory
and Society, vol. 40, p. 315, 2011.
[19] Joseph Yiu. 2015. The Definitive Guide to Arm® Cortex®-M0 and
Cortex-M0+ Processors (2nd Edition).
[20] M. S. Johnstone and P. R. Wilson. “The Memory Fragmentation
Problem: Solved?” Proceedings of the 1st International Symposium on
Memory Management, vol. 34, no. 3, October 1998.
[21] Z. Zhu, F. Wu, J. Cao, X. Li, and G. Jia. 2019. “A Thread-Oriented
Memory Resource Management Framework for Mobile Edge
Computing.” IEEE Access, vol. 7.
[22] Durgesh Raghuvanshi. “Memory Management in Operating System.”
International Journal of Trend in Scientific Research and Development
(ijtsrd), ISSN: 2456-6470, vol. 2, no. 5, August 2018.
 SYSTEMATIC APPROACH WITH FRAGMENTATION 239

[23] Amanjot Kaur Randhawa, and Alka Bamotra. “Study of Static and
Dynamic Memory Allocation.” International Journal of Innovative
Computer Science & Engineering, vol. 4, no. 3,pp. 127–131, 2017.
[24] Meenu, Vinay Dhull, and Monika. 2015. “Computational Study of
Static and Dynamic Memory Allocation.” International Journal of
Advanced Research in Computer Science and Software Engineering,
vol. 5, no. 8.
[25] Dharmender Aswal, Krishna Sharda, and Mahipal Butola. “DMA:
Dynamic Memory allocation.” IJIRT, vol. 1, no. 6, 2014.
[26] Dipti Diwase, Shraddha Shah, Tushar Diwase, and Priya Rathod.
2012. “Survey Report on Memory Allocation Strategies for Real Time
Operating System in Context with Embedded Devices.” IJERA
Internet Computing [Online], vol. 2, no. 3, pp. 1151–1156.
[27] Dinesh Kumar and Mandeep Singh. 2019. “Memory Management
in Operating System.” International Journal of Emerging Technologies
and Innovative Research, 6, 465–471.
[28] Tie Qiu, Heyuan Wang, Keqiu Li, Huan Sheng, Arun Kumar,
Sangaiah, et al. “A Novel Machine Learning Algorithm for Spammer
Identification in Industrial Mobile Cloud Computing,” IEEE
Transaction on Industrial Informatics, vol. 15, no. 4, pp. 1, April 2019.
[29] J. Miranda, N. Makitalo, J. Garcia-Alonso, J. Berrocal, T. Mikkonen,
C. Canal, et al. “From the Internet of Things to the Internet of
People.” IEEE Internet Computing, vol. 19, no. 2, pp. 40–47, 2015.
[30] T. K. Rodrigues, K. Suto, H. Nishiyama, J. Liu, and N. Kato.
“Machine Learning Meets Computation and Communication Control
in Evolving Edge and Cloud: Challenges and Future Perspective.”
IEEE Communications Surveys & Tutorials, vol. 22, no. 1, pp. 38–67,
First quarter 2020, doi: 10.1109/COMST.2019.2943405.
[31] V. Sze, Y.-H. Chen, T.-J. Yang, and J. S. Emer. “Efficient Processing
of Deep Neural Networks: A tutorial and Survey.” Proceedings of the
IEEE, vol. 105, no. 12, pp. 2295–2329, 2017.
[32] S. Gu, Q. Zhuge, J. Yi, J. Hu, and E. H.-M. Sha. “Data Allocation
with Minimum Cost under Guaranteed Probability for Multiple
Types of Memories.” Journal of Signal Processing Systems, vol. 84,
no. 1, pp. 151–162, 2016.
[33] X. Chen, E. H.-M. Sha, Q. Zhuge, W. Jiang, J. Chen, J. Chen, et al.
“A Unified Framework for Designing High Performance in Memory
and Hybrid Memory File Systems.” Journal of Systems Architecture,
vol. 68, pp. 51–64, 2016.
[34] X. Zhang, D. Feng, Y. Hua, and J. Chen. “A Cost-Efficient NVM-
Based Journaling Scheme for File Systems.” In: 2017 IEEE
International Conference on Computer Design (ICCD), pp. 57–64, 2017.
240 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Appendix 10.A

The source for the dataset used for the ML prediction model is Google
Cloud platform and Enlyft online platform as shown below:
 11
IOT AUTOMATED SPY DRONE TO
DETECT AND ALERT ILLEGAL
DRUG PLANTS FOR LAW
ENFORCEMENT
GOTLURU ARUN KUMAR,
PULUGU YAMINI, AND R. MAHESWARI
SCOPE, Vellore Institute of Technology, Chennai,
Tamil Nadu, India

Contents

Introduction 242
Objectives 243
Literature Survey 243
Existing System 244
Proposed System 244
Architectural Diagrams 244
Methodology 246
Technical Approach 246
VGG16 Architecture 247
Modules Involved 248
Description of Working Modules 248
The Steps Involved in the Image Classification VGG16
Model 248
Remote Interface Encryption Is Done by Using Vigenere
Cipher and MD5 Hashing Techniques 249
Vigenere Cipher 249
Encryption 250
Decryption 250
MD5 Hashing Techniques 250
Drone Frame 251
Implementation 252

DOI: 10.1201/9781003119838-11 241

242 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Implementation of Methodologies of the Neural Network

Models 252
Implementing VGG16 253
Import Packages 253
Import DataSet 253
Image Preprocessing 254
Import Target Image Input 254
Visualize the Plant in RGB 254
Visualize the Plant in HSV 255
Segmentation 257
Color-Based Segmentation 257
Segmentation Function 257
Computer Vision Model 258
Experimental Results 258
Remote Interface Screenshots 258
Conclusion 262
References 264

Introduction

In this era, there is no efficient way of detecting the psychoactive drugs

without manpower. Here the spy drone has a unique feature that
identifies the psychoactive drugs, as it contains an autopilot to identify
the location of the drug plants with the help of GPS, which can be
explicitly controlled remotely. The collected information will be stored in
the database managed by the drone controller. This paper discusses the
work on autopilot, image classification techniques to identify the drug
plants, and the information security concepts used in developing the
remote interface for securing and storing the collected information in the
database by the controller. A drone is passed through various areas of the
city to see the people roaming around. If the drone controller identifies
unwanted gathering or roaming, the location and the information will be
sent to the nearest police station. Now when it comes to the proposed
work, based on the existing work, the proposed system has been mod­
ified accordingly to identify the psychoactive drugs with the help of
a drone in which image classification techniques using a python
programming language to classify the drug plants in an area are used.
 IOT AUTOMATED SPY DRONE 243

GPS tracker is used to track the location of the identified drug plants by
the drone. The shortest path algorithm is used to send the drone to the
various locations and get back to the controller. This particular algorithm
is used for low consumption of energy and time. Cryptography tech­
niques like vigenere cipher and MD5 hashing algorithms are used for
data encryption in the database. Immediately, the police will take the
action. An automated spy drone is designed to reveal any persons/groups
planting and growing psychoactive drugs unofficially in their private land
and identify them. Whenever the spy drone system identifies these
psychoactive drugs, it immediately delineates the inspection officer by
storing the information in the database automatically.

Objectives

This paper aims to develop an eco-friendly smart spy drone that can
be used for large-scale sectors handled by the government to detect
the psychoactive drugs and can help in decreasing the growth of il­
legal drug production. The controller will control the drone with a
remote controller by sending the drone to specific locations to
identify the psychoactive drug plants. Drone detects the psychoactive
drugs grown in private lands and forest areas with the help of image
classification techniques. VGG16 network architecture has been
implemented for the image classification by the drone. Shortest Path
algorithm is used for sending the drone to different locations and to
return to the controller. It helps in saving energy and time con­
sumption. As the controller will be monitoring the location details of
the drone and the images captured by the drone, he or she will store
that information in a database provided in the remote controller. The
collected/stored information will be encrypted using cryptography
techniques to avoid data leakage. Vigenere cipher and MD5 hashing
techniques are used for the encryption. After storing the information
in a database, the controller sends the collected information to the
head office for further actions.

Literature Survey

In the literature survey, many articles had been referred to and had
some research on them. The working principles of VGG16 algorithm
244 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

in Keras are discussed in Aized Amin Soofi et al.’s work [1].

Cryptography techniques like enhanced vigenere cipher for data en­
cryption were presented by Aized Amin Soofi, Irfan Riaz, Umair
Rasheed in Wang et al.’s work [2], hashing algorithm such as MD5 by
Shweta et al. [3]. Working and possibilities of drone were discussed
by Piotr and Jacec [4] and working and functioning of VGG16
and VGG19 by [5]. Deep convolution networks for large-scale
images recognition were discussed by Karen Simonyan and Andrew
Zisserman [6].

Existing System

In this pandemic situation, drone technology is used for regular in­

spection whether people are roaming during a lockdown or not. A
drone is passed through various areas of the city to see the people
roaming around. If the drone controller identifies unwanted gath­
ering or roaming, the location and the information will be sent to the
nearest police station. Immediately, the police will take the action.

Proposed System

Based on the existing work, the proposed system has been modified
accordingly to identify the psychoactive drugs with the help of a
drone in which image classification techniques using a Python pro­
gramming language to classify the drug plants in an area are used.
GPS tracker is used to track the location of the identified drug plants
by the drone. The shortest path algorithm is used to send the drone
to the various locations and get back to the controller. This particular
algorithm is used for the low consumption of energy and time.
Cryptography techniques like vigenere cipher and MD5 hashing
algorithms are used for data encryption in the database.

Architectural Diagrams

Figure 11.1 elaborates the complete architecture of remote interface.

This interface is visualized and controlled by the controller where the
controller can register and login into the interface and this interface
provides services like Add, List, and Search.
 IOT AUTOMATED SPY DRONE 245

Figure 11.1 Remote Interface Architecture

The authorized person can add the details of the identified plants and
location into the database and if admin wants to see the location, then he/
she has to authenticate and retrieve data. The login details and location
are encrypted using vigenere cipher [2] and hashing technique like MD5
[3] algorithms are used for security. In the overall architecture of the spy
drone, there are mainly five modules: The first module is the commu­
nication module which is also a database used for the data transmission
from drone to the database. The second module is vision sensor. It helps
the drone for visualizing the specific location and after the visualization,
the identified plants will be classified into normal or psychoactive plants.
The next module is mission controller which involves a battery and the
remote controller. This remote controller is used to operate the drone by
the controller. Drone autopilot is the fourth module that is the drone with
an autopilot which has an auto mechanism to get away from distracting
objects. The last module is the GPS module that is used to identify the
locations and navigate the drone to the particular location. The overall
architecture of the spy drone is shown in Figure 11.2.
246 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 11.2 Architectural Diagram of Spy Drone

Methodology
Technical Approach

For image classification and identifying plants, ML/Dl algorithms

are used [7]. GPS tracker is used to tracking the location of the drone
and identified plants. The shortest path algorithm is used to get back
to the controller. This particular algorithm is used for low con­
sumption of energy and time. Cryptography techniques like vigenere
cipher and MD5 hashing techniques [5] are used for encrypting the
collected data to avoid data leakage.
The main steps involved in the machine learning (ML)
model are:
• Importing of necessary packages
• Importing of data set path
• Data prepossessing
• Data augmentation
 IOT AUTOMATED SPY DRONE 247

• Splitting of data
• Building of the model
• Training of the model
• Results of the accuracy
Study of each step of the ML model
For the image classification in Spy drone, all the necessary packages
are imported from Keras libraries in python. The next step is im­
porting the data set. The dataset will be stored in the respective
directory where the notebook file is stored. The path of the dataset
directory will be given for importing the dataset for further steps.
The collected data don’t need to be in the desired format.
Therefore, data prepossessing techniques are used. Data pre­
possessing is a process of preparing the raw data and making it
suitable for a machine learning model. For performing data pre­
possessing using python, we need to import some predefined
Python libraries, i.e., Pandas. Data prepossessing techniques in­
clude formatting of data, cleaning of data, and sampling of data.
Data augmentation is an integral process in deep learning. In deep
learning, we need large amounts of data and in some cases, it is not
feasible to collect thousands or millions of images, so data aug­
mentation comes to the rescue. There is no need to collect new
data; rather it transforms the existing data into the required forms
and stores as augmented data. Data augmentation operations in­
clude boundary boxing, max pooling, cropping, and extracting the
region of interest (ROI).

VGG16 Architecture

It is a convolution-based neural network model integrated with the

image net processing [6]. To date in the domain of visual archi­
tecture, VGG16 plays a vital role in its excellent performance. The
most important and unique feature in VGG16 is that it will not use
any hyper-parametric values in it. Rather it will have the convolu­
tional layers and the same type of max-pooling and padding will be
processed in this VGG16 modeled architecture. 3 × 3 layered filters
will be used in convolutional layers and 2 ×2 layered filters will be
used in max-pooling and padding for every time. Activation
248 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 11.3 VGG16 Architecture

functions like Relu or Softmax will act as fully connected layers in

this VGG16 architecture for providing output. In VGG16 [5], 16
represents the weights associated with the model architecture.
VGG16 has approximately 138m parameters in it, and this concludes
that the VGG16 neural network architecture is a very large network
model, as [8] represented in Figure 11.3.

Modules Involved

1. Image Classification
Captured image gets classified with the pre-trained model
VGG16. Model has been trained with several dataset plants and
it identifies the target plant.
2. Remote Control Interface
Remote controller contains an interface that contains features
like Add location, List, and search in database; thus, the person
who is going to access/control the drone gets authenticated if he
or she tries to get data and see the original data stored in it.
3. Drone Frame
Drone frame has been designed using the tinker-card tool,
which is the best tool for virtual simulation of IoT devices.

Description of Working Modules

The Steps Involved in the Image Classification VGG16 Model Import of

necessary packages, import of data set, image preprocessing, seg­
mentation, and computer vision model are represented in Figure 11.4.
Figure 11.5 represents the steps of various packages that were
downloaded from Keras, how dataset gets imported into the model
and visualizes the input images. Next image processing is done by
importing the target image and visualizing the image in the RGB
 IOT AUTOMATED SPY DRONE 249

Figure 11.4 Steps Involved in the Image Classification Process

Computer Vision Model

Figure 11.5 Image Classification Internal Process

and HSV format; thereby image gets segmented based on color and
visualizes the segmentation results in VGG16 model.

Remote Interface Encryption Is Done by Using Vigenere Cipher and MD5

Hashing Techniques

Vigenere Cipher It is a cryptography-based method which is used for

encrypting the text in alphabetical form. The form of poly-alphabetic

substitution is used in this method. This cipher is based on the
substitution of multiple alphabets. The plain text message will be
encrypted for generating the ciphertext using the vigenere table
250 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 11.6 The Vigenere Table

referred to Figure 11.6. When it comes to the vigenere table, it

consists of 26 rows and 26 columns which are filled with the al­
phabets 26 times in a looping format. That is, cyclically each alphabet
is shifted to the left from the previous row alphabets.

Encryption In the encryption process, from any of the rows, different

alphabets are used for generating the cipher text.

Decryption In the decryption process, the vice versa operation like en­
cryption will be performed. But the only difference is that the encryption
operation will be done on the plain text and the decryption operation will
be done on the cipher text generated to obtain the plain text.

MD5 Hashing Techniques A hashing technique accepts a message

input of any length and provides the message output of a fixed
length, which is also said as digest value. This digest value will be
used for plain text message authentication [8–10]. For providing a
secured cryptography hashing algorithm and for digital signatures
 IOT AUTOMATED SPY DRONE 251

Figure 11.7 MD5 Hash Technique Working

authentication, this MD5 hash technique is designed. This helps in

verifying data integration and avoids data corruption. As discussed
earlier, this MD5 hash functions, as shown in Figure 11.7, an input
message of any length and generates a fixed length of 128-bit output,
which is also known as the message digest value of the input message.
It was mainly designed for digital signatures authentication.
If some large files are highly confidential, then it must be com­
pressed in a cryptography secured manner before starting the en­
cryption process with a secret key of a particular user using
cryptography algorithms. This states the digital signature of the files.
So, for providing those digital signatures, this MD5 hash technique
will be very much useful.

Drone Frame Using the Tinker cad tool, a basic drone has been
designed and all modules get embedded within the drone, as shown
in Figure 11.8.
All the components required for the design of the drone along
with the processing unit such as Arduino is configured using the
Tinker cad tool and expected output of the design can be visualized
252 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 11.8 Design of Drone Simulation

by simulating the design in the Tinker cad tool. This may help us to
debug the circuit if the expected output gets varied from the observed
output.

Implementation

This section focuses on the implementation methodologies in per­

ception with neural network models, such as convolution and pooling
techniques, along with image classification implementation tech­
nique, such as VGG16. Image cleaning to its preprocessing and
image visualization using RGB and HSV color models are also
discussed here.

Implementation of Methodologies of the Neural Network Models

Figure 11.9 signifies the implementation methodologies using neural

network containing convolution and max pooling techniques.
Convolution: Convolutions are a type of operation that can be used
to learn representations from images. They include learnable kernel
slides over the image performing element-wise multiplication with the
 IOT AUTOMATED SPY DRONE 253

Figure 11.9 Implementation of Neural Network Models

input. The specification permits for parameter sharing and translation

in-variance. Given below is a continuously updated list of convolutions.
Pooling: Max pooling is an operation that calculates the utmost
value for patches of a feature map and uses it to form and create a
down-sampled feature map. It adds a little quantity of translation in-
variance – that means translating the image by a small quantity does
not considerably affect the values of most pooled outputs.

Implementing VGG16

In this model, VGG16 is implemented for image classification. The

following steps are involved in image classification process.

Import Packages In the first step, all the necessary packages for image
classification using VGG16 are imported from Keras library in
Jupyter Notebook with Python programming language.

Import DataSet Here, the dataset containing psychoactive drug

plants images should be imported into the Jupyter Notebook for
further data preprocessing of the dataset. The dataset consists of info
of 12 plant species. Also, the sample images from the dataset will be
displayed to check whether the data have been successfully imported
into the notebook. Figure 11.10 shows the sample visualization of
crop seedling images from the imported dataset.
254 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 11.10 Visualization of Crop Seedling Images

Image Preprocessing The images in the dataset will go under

preprocessing. In the imported data while importing it into the
notebook, there may be a chance of missing values and null values.
The dataset may also contain the noise data in it. For retrieving those
missed values and avoiding noise data, data preprocessing will
be done.

Import Target Image Input Processing the image to get rid of the
unnecessary pixels is depicted through Figure 11.11. Feeding this
picture to the model directly will result in a target leakage. The
classifier will learn the stones and the container’s pixels and build its
prediction upon these pixels as well. This is something that will be
avoided. Preprocessing starts with by doing object segmentation from
the input image which is based on the RGB colors using a Python
library called OpenCV, as it is the most popular computer vision-
based library for manipulating colors in an easy way.

Visualize the Plant in RGB Visualizing the small-flowered Cranesbill

image we opened above in RGB space to see the distributions of the
 IOT AUTOMATED SPY DRONE 255

Figure 11.11. Target Image as Input

color pixels. RGB is considered an “additive” color space, and colors

are visualized in red, green, and blue with the background color black.
From the given RGB visualization in Figure 11.12, green parts of the
image are mixed with red and blue values, segmenting our plant image
out in RGB space based on the ranges of RGB values. The solution is to
try another color space, and we go next with HSV color space.

Visualize the Plant in HSV HSV is the alternate approach for RGB
color model. HSV stands for hue, saturation, and lightness value, and
represents the mixing of different colors with saturation dimension
resembling. It defines a threshold value to help it become a binary
image with the color, as mentioned in Figure 11.13.
In HSV space, our green plant are much more localized and vi­
sually separable. The saturation and value of the greens do vary, but
they are mostly located within a small range along the hue axis. This
is the key point that can be leveraged for segmentation.
256 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 11.12 RGB Visualization

Figure 11.13 HSV Visualization

 IOT AUTOMATED SPY DRONE 257

Segmentation

Color-Based Segmentation To pick out a range of color, we can use a

color picking app online. I found the green color ranges in a forum:
• Minimum green(H=36, S=25, V=25)
• Maximum green(H=70, S=255,V=255)
The OpenCV function cv2.inRange() is used to ensure the threshold
values of the plant image ranges with the minimum and maximum
green values.
InRange() takes three parameters: the image, the lower range, and
the higher range. It returns a binary mask (and array of 1s and 0s) of
the size of the image where values of 1 indicate values within the
range and zero values indicate values outside [11] (Figure 11.14).
• colormin=(36, 25, 25)
• colormax=(70, 255,255)
• plot_mask(p1, coloring, color max)

The result of our initial threshold is not bad, since some pixels of the
plants were missed. Later, it has been manually adjusted with saturation
and brightness values to get a better mask [12] (Figure 11.15).
• new_colormin=(25,50,50)
• new_colormax=(80,255,255)
• plot_mask(p1, new_colormin, new_colormax)

Segmentation Function Now, a function that applies segmentation on

the rest of the images is created. This function will be used to feed

Figure 11.14 Segmentation of the Input Image

258 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 11.15 Segmentation of Image without Missing Values

the computer vision model with the segmented images instead of the
original ones.

Computer Vision Model

This shows plant images to a pre-trained VGG16 convolutional

neural network. It is trained with both original and segmented
images to find out the impact of the target leakage on training. It
involves VGG16 with original plant image.

Experimental Results

Training Images
Trained models get an input of the collected images and the model
performs the segmentation for the target image and then the image
gets converted into the RGB color model.
Once the RGB color model is not clear, the images are con­
verted and represented in the HSV color model. In this model,
HSV is a mixture of colors and can be visualized better than RGB
model. Images get classified after the segmentation, as shown in
Figure 11.16. After performing the model, the predicted results
are shown in Figure 11.17.

Remote Interface Screenshots

Services provided by the remote control interface are shown in the

following screen shots.
 IOT AUTOMATED SPY DRONE 259

Figure 11.16 Training Images

Figure 11.17 Images after Trained

260 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 11.18 GUI Registration Page

Figure 11.18 represents the registration page to the person who is

going to access the drone.
Figure 11.19 depicts the GUI for login into the application where
the person has to register to login into this module.
Figure 11.20 contains three options whether to add the location,
see the details in the list, or search for any data in the list.
Figure 11.21 contains the details of the authorizer who controls
the drone and location details which are in encrypted format and all
text fields must be filled.

Figure 11.19 GUI Login Page

Figure 11.20 Home Page of GUI

 IOT AUTOMATED SPY DRONE 261

Figure 11.21 Adding Collected Information

Figure 11.22 Added Details

In Figure 11.22, added details are viewed but the location and
details of the identified plants are not viewed. Only the person gets
authenticated and then the information is viewed.
Figure 11.23 is used for searching the details of the authorizer.
The security for this GUI application will be provided using
cryptographic algorithms like the vigenere cipher algorithm and
MD5 hash techniques. One of the most important things is that the
drone will be incorporated with the shortest path algorithm for
sending the drone to different specified locations and for returning to
the controller quickly. Figure 11.24 represents the entered location
saved in an encrypted format.
262 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 11.23 Search Module

Figure 11.24 Encrypted Data

Conclusion

In this work on “IoT-Automated Spy Drone to Detect and Alert

Illegal Drug Plants for Law Enforcement,” mainly three modules
that have been discussed. The first module is that there will be an
autopilot that is integrated with a camera, in-built GPS naviga­
tion, and the image classification application for classifying the
psychoactive drugs from a group of plants being grown in the
specified locations. The second module is the image classification
application, in detail. This application will be used for classifying
 IOT AUTOMATED SPY DRONE 263

the plant species that are captured by the drone using a camera.
For building this image classification application, a convolutional-
based neural network called VGG16 neural network model has
been used for classifying the psychoactive drugs from the group of
plant images. This VGG16 neural network model has been im­
plemented in the python programming language using the Jupyter
Notebook as a platform from the Keras libraries. The internal
processing of image classification has been discussed in this work.
Finally, the third module is about the remote control interface.
There should be a remote controller that is controlled by the
controller to control the drone. The unique feature of this remote
controller is that there will be an inbuilt GUI application in it for
the controller to store the collected information from the drone.
The collected information and user details must be secured to
avoid data leakage. Here, comes the cryptography algorithms used
for encrypting the gathered information and also for providing
confidentiality, authentication, and integrity for the application.
The controller after collecting the information will store the de­
tails of it in this GUI application, which is associated with a da­
tabase, so that he can retrieve the information in the future and
take proper actions on the identified landowners. The security for
this GUI application will be provided using cryptographic algo­
rithms like the Vigenere cipher algorithm and MD5 hash tech­
niques. One of the most important things is that the drone will
be incorporated with the shortest path algorithm for sending
the drone to different specified locations and for returning to the
controller quickly. This helps the drone to select the best and the
shortest path among all the possible paths to reduce energy and
time consumption. The output screenshots of the image classifi­
cation and the GUI application have been given in the experiment
results section. These are all the things that are discussed in this
work. Finally, to conclude that the main goal of this Spy Drone
application is detecting the psychoactive drugs in a specific loca­
tion and collecting the information that is required for taking
further actions using a drone. This helps the government sector
people to control the illegal activities of drug usage and production
efficiently without using more manpower.
264 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

References
[1] Aized Amin Soofi, et al., “An Enhanced Vigenere Cipher for Data
Security,” International Journal of Scientific & Technology Research, vol.
5, no. 03, pp. 141–145, 2016.
[2] Z.-p. Wang et al., “Single-Intensity-Recording Optical Encryption
Technique Based on Phase Retrieval Algorithm and QR code,” Optics
Communications, vol. 332, pp. 36–41, 2014.
[3] Shweta Mishra et al., Hashing Algorithm: MD5, Department of
Computer Science & Engineering, Echelon Institute of Technology,
Faridabad, India; JB Knowledge Park, Faridabad, India, 2013.
[4] Piotr Kardasz and Jacek Doskocz. “Drones and Possibilities of Their
Using”. Journal of Civil & Environmental Engineering, vol. 6,
Piłsudskiego, Wrocław, Poland, 2016.
[5] VGG16 and VGG 19 functions are referred from following the link
https://keras.io/api/applications/vgg/
[6] Karen Simonyan and Andrew Zisserman, “Very Deep Convolutional
Networks for Large-Scale Image Recognition,” In: International
Conference on Learning Representations, 2015.
[7] Ayon Dey, “Machine Learning Algorithms: A Review,” International
Journal of Computer Science and Information Technologies, vol. 7, no. 3,
2016, 1174–1179.
[8] Working of VGG16 model is referred from following the link
https://www.geeksforgeeks.org/vgg-16-cnn-model/
[9] Algorithm of MD5 hashing techniques are referred from following
the link http://practicalcryptography.com/hashes/md5-hash/
[10] Mary Cindy, Ah Kioon, Zhao Shun Wang, and Shubra Deb Das,
“Security Analysis of MD5 algorithm in Password Storage,” Applied
Mechanics and Materials, vol. 347–350, pp. 2706–2711, 2017.
[11] Vigenere Cipher, Eric Conrad, and Joshua Feldman, CISSP Study
Guide (3rd Edition), 2016.
[12] Jean-Paul Yaacoub, Hassan Noura, Ola Salman, and Ali Chehab,
“Security Analysis of Drones Systems: Attacks, Limitations, and
Recommendations,” Internet of Things, vol. 11, May 2020.
 12
EXPOUNDING K-MEANS-INSPIRED
NETWORK PARTITIONING
ALGORITHM FOR SDN
CONTROLLER PLACEMENT
J. PUSHPA1
AND PETHURU RAJ CHELLIAH2
1
Jain University, Bangalore
2
Site Reliability Engineering (SRE)
Division, Reliance JioInfocomm. Ltd.,
Bangalore, India

Contents

Introduction 266
Leveraging on Unsupervised Learning 267
k-Means Algorithm 270
Data-Set for k-Means Algorithm 272
Distance Formula of Centroid in k-Means Algorithm 273
Mechanism to Compute the Value of K 276
Properties of k-Means 278
Compare the k-Means with Fuzzy C Mean, K Mediods,
and k-Means++ 281
k-Means Clustering in Software Defined Network (SDN) 282
Exploring the k-Means Clustering in Resolving SDN
Challenges 284
SDN Network Partitioning of Saren Topology-Based using
k-Means 285
Conclusion 287
References 288

DOI: 10.1201/9781003119838-12 265

266 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Introduction

The power of Artificial Intelligence (AI) significantly brought the

new paradigm to resolve many challenges which were heaped with
inefficient solution. AI adopted many robust and sophisticated al­
gorithm and also the analysis of AI infuses in for many scenario
which gives accurate and efficient resultant values. In this chapter will
discuss about the one such algorithm that is k-means algorithm
which can be amalgamated by any technology for different kinds of
problems.
Before we discuss in detail about the k-means algorithm, will
discuss some of the scenario which direct to choose the kind of al­
gorithm/approach for analyzing/categorizing/resolving the problems.
In the field of electing leader such as master node in networking,
ranking the quality or grading the system are all not continuous or
common value. Those can be consider as for discrete data. Discrete
samples are countable/measurable which require linear regression
techniques. On the other hand if value have similarities such as
common symptom, same type of traffic, similar pattern or recursive
samples than it is continuous data. For continuous data which have
similarities which are not countable. For continuous samples which
are massive in volume with unlabeled data need to group with non
linear counter part by clustering. Clustering the data is one of the
prominent way to classify or group the values in which most of
clustering algorithm adopt the euclidean/correlation distance formula
for segregation non linear values.
If we look the idea behind clustering algorithm is basically origi­
nated for analysis purpose and its journey is from past many decades.
Our main objective is to discuss one of the traditional and efficient
clustering algorithm that is k-means clustering algorithm. Clustering
based problem are in all the fields from genetic related analysis to
computer network traffic analysis. Most of the researcher have opted
k-means algorithm for clustering. In this chapter will discuss on k-
means role in advance technologies, properties and its efficiency.
k-means algorithm also know as Llody’s algorithm which is much
influenced on vector quantization. Basically, this algorithm work as
all small K similar group moving towards centroid with minimal
distance and computing the distance is computed using Euclidean
 SDN CONTROLLER PLACEMENT 267

Distance formula. According to [1], k-means algorithm is the process

for partitioning N-dimensional population into K sets. Many defi­
nition or interpretation of k-means can be find in various field. J.
Andrew [2,3] has worked on augmented supervised and non su­
pervised clustering with k-means algorithm which has resultant to
the better convergence and classification. Integration of k-means
algorithm with other algorithm like LVQ algorithm [4] has increased
the performance of clustering.
k-means can not only integrate with different methodology but also
applied in the field of Genetics, Machine learning, Artificial in­
telligence and many more. Hence it can be declared as versatile algo­
rithm for clustering problems. Section 2 will give an origin and
hierarchical of k-means to which it belongs in machine learning. In
Section 3, we discuss the k-means algorithm and its working structure
along with its prime properties. k-means is applied in many technology
which conjugate with other methodologies for better efficiency, this
discussion helps to route our exploration for solving the problems.
And in last two section the discussion extend on the implementa­
tion and applications in which k-means plays an vital role.

Leveraging on Unsupervised Learning

Its a greater part to know new technologies and their application

basically as applied science to experience the benefit out of it.
Learning Data Science, Artificial Intelligence and Machine Learning
make the technology work in am very robust and acute way. There
are a plethora of use case which helps to understand in deeply about
the algorithms and approaches which helps to make the world
connected through IoT devices. In this section we learn more about
the Unsupervised learning and its importance.
IT is enriching in all the domain mainly fascinated to develop IoT
products which is widely in high demand. The technology behind
IoT is not only Arduino, Cloud Technology but also the intelligent
algorithm of machine learning which adopt either regression, clas­
sification or clustering algorithm.
These algorithms are categorized into Supervised, Semi Supervised
and Unsupervised learning, it basically depends on the type of the data
fed to the system.
268 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Machine Learning

Supervised Learning Semi Supervised Unsupervised Learning

Considering an automated system such as Face detection,

Weather forecasting, Stock Price prediction which are operating on
the trained data or predicting based on the previous knowledge in
these cases few information is already stored and based on it the
output will be predicted such kind of mechanism called Supervised
learning. In supervised learning if some situation arises like malicious
traffic or unknown symptom in the patient or any unusual load then
the prediction may fail to give the accurate output this can be
overcome by regress evaluation process with targeted result which has
no predefined label termed as Unsupervised learning. As discussed in
the previous section k-means is the most popular algorithm which
also comes under unsupervised learning due to it evaluation process
and the rate of convergence.
Before understanding in detail about the k-means algorithm of
unsupervised learning, will understand the basic characteristic of
unsupervised learning.
Unsupervised Learning is a technique to unveil the hidden pattern
or structure in the input data set. As shown in the Figure 12.1,
unsupervised learning methodology is classified into clustering, di­
mensional reduction and association.
Let us consider an example to get the difference between those
methodologies, as the internet connectivity is getting better the
malicious attack is the biggest challenge in networking. To overcome
this challenge the network traffic should be classified and identifies
the DDoS attack.
In this scenario not only the classification algorithm alone can be
used along with that data size should also matters because the net­
work traffic is collection of enormous packets with varying size and
also need to establish the relation to study the association to capture
the similarity.
 SDN CONTROLLER PLACEMENT 269

Classification Clustering
Data Size Reduction Dimensional Reduction
Similarity/Relation Association

I. Clustering: Data are been partitioned into distinct subset based

on the parameter such as distance from the centroid.
II. Dimensional Reduction: Reducing the raw data set into man­
ageable sized data set with integrity is the task of Dimensional
Reduction.
III. Association: Is a mechanism to identify the rule for establishing
the relation.
Relating the unknown value in a large data set is a complex process,
as depicted in the Figure 12.1. Classifying the large data set using the
mention methodology takes an deep analysis in which k-means is
categorized as clustering algorithm. But, some of the data set di­
mension value will be more in number which require mechanism to
simply and associate it based on the relativity.
Principal Components Analysis(PCA) play an important role in
segmentation to reduce the dimensional features. Assume if the data
contains more than 20 features, then clustering perhaps results to an
inefficient cluster. Hence, standardizing and reducing the low impact
features are needed. Further, before proceeding with k-means clus­
tering it is necessary to compute the PCA.
An unsupervised learning is a wide topic which is applied in most
of the technology, we are focusing on clustering methodology and
also discussing on most popular clustering algorithm which help us to
understand and differentiate with k-means algorithm.

Clustering

Unsupervised Learning Dimesional

Reduction

Association

Figure 12.1 Methodology of Unsupervised Learning

270 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Table 12.1 Clustering Algorithms

TYPE TASK EXAMPLE
k-means, K-Mediods Clustering the data points by Compression, Data Mining.
calculating the centroid distance.
Hierarchical Nestested clustering with levels Anomaly detection, Deep
Clustering Learning.
Density-Based Clustering based on density in which Text clustering, Image
Clustering the grouping will done based on clustering, Outlier detection.
the quantity.
Spectral Clustering Clustering based on graph. Community detection, image
segmentation, network.
Gaussian Mixture Clustering based on the probability Speech recognition, Signal
Models distribution for a given data set. mapping and positioning
Nearest Neighbors Grouping the data points which are Machine Learning
close using exhaustive search.
Hidden Markov Markov model for stochastic process Stock Market,Position of
Models based on current state Chromosome in genes

Clustering is a methodology which classify based on the require­

ment basically to analyze the pattern or organize the data else to
recognize the pattern. For implementing this clustering many stra­
tegic algorithm are discovered by researcher in which k-means is the
most popular clustering algorithm. As discussed in [5–7], the algo­
rithms comes under clustering is tabulated below in Table 12.1.

Clustering

k-Means Algorithm

The k-means algorithm is introduced by J. Macqueen [8] for the

method to classify and analyze the asymptotic behavior of data set,
the proposal has proved the optimal classification of sets which are
measured based on the relative distance and also concluded that as k
value should be minimal which will be than close to mean value, this
observation is made on multivariate with varying dimensional value.
k-means clustering is a partitioning clustering which partitions the
data set into k number of group which contain the equal size in both
the cluster and the objects of one cluster is not belong to other
clusters. It uses Euclidean/Squared Euclidean/Manhattan/cosine
distance algorithm for the distance between the centroid and the
 SDN CONTROLLER PLACEMENT 271

other nearing objects. This process will continue in an iterative way

to group all the objects closer to centroid, hence the objects can not
be in a single cluster till the final convergence.
Finding the k-means is depicted in the diagram Figure 12.1 given
in Figure 12.2.
As presented in flow chart, the flows are discussed below with few
question.

STEP 1: Indicate the considering the data set Ss = {x1, x2, x3, ….xn}:
In this step few things need to consider such as:
What kind of data set?
• Unsupervised data set is more suitable for k-means clustering
What is the limitation of data set?
• Big O(better performance) for small data set and number of iteration
increases for large data set which will non-optimal solution.

STEP 2: Initialize the value of K.:

To determine the optimal number of cluster there is no definite rule
to choose it. As discussed in datanovia [9], there are two startergies
a. direct method
b. statistical testing method, (will discuss it in detail in the next section)

STEP 3: Assign the Centroid to each cluster, basically euclidean distance

formula is applied to bring the data point closer to the centroid for
grouping. Centroid C = {c1, c2, c3, … ck}:
Euclidean Distance =arg min dist(ci,x)2,, where ci belongs to C.

STEP 4: Compute the new centroid which will be mean value in a group,
so new centroid will be emerged and regrouping may take place.

Ci = 1/ Si xi, xi S.
If reposition of centroid then repeat then step 3.:
else:
goto step 5

STEP 5: Once the cluster is static, k-means clustering is said to be

converged.:
The above steps not only explain the process in k-means but also
some of the parameter such as k value and distance formula
between the data points. Clustering algorithm not only aims to
group the data points but also need to consider some of the
272 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Start

Initialize the data set for clustering

S={n1,n2,n3,..nx}

Find the optimal the

value of K

Initialize the centroids for k-clusters

as C1,C2,..Ck

for(i=1;i<=x;i++)
If not Converged

Check ni is
Closer to Final Cluster formed
CI C2 CK

If Converged
Reposition the Centroid

Figure 12.2 k-Means Algorithm Flow Chart

coefficient factors such as ideal distance value between the data points
in group, the distance between the group, optimal number of cluster and
number of data point in a cluster. Will discuss those coefficient factor
as the properties of k-means algorithm and also see some the
example on it.: Many of the researcher have used the k-means to
resolve their problem for the certain kind of data set which kind of
Data set can use k-means algorithm is our next discussion.

Data-Set for k-Means Algorithm

• It is important to understand the pattern of data before applying

any algorithm. k-means gives better convergence when data set
 SDN CONTROLLER PLACEMENT 273

is in spherical / convexshaped structure and failed for non-convex

shaped.
• k-means clustering algorithm fails to give good convergence
when the data set contains outliers, the density spread of data
points across the data space is different and the data points
follow non-convex shapes.
• The K-means algorithm needs the value of K which determine
the number of cluster which has to entered manually.
• One of the feature scaling ensures that all the features get same
weight in the clustering analysis.

Distance Formula of Centroid in k-Means Algorithm

All the data points of a cluster in k-means algorithm is closer to the

centroid of that cluster, and the distance between centroid is com­
puted using Euclidean distance formula and there are other alter­
native formula also been used based on varying requirements as
shown in the Figure 12.2.
The k-means algorithm is an solution for many problems which
helps to reveal the hidden pattern, we learnt how the iterative process
helps to create the clusters in which two factors are playing an
prominent role one is distance between the centroid and the other
one is the value of k. In the Figure 12.3 various distance formula can
be associated with k-means for variety of problems.

Squared Euclidean
distance

Distance formula
between Centroid Manhattan
in K-mean Distance

Minkowski
Distance

Chebyshev
Distance

Figure 12.3 Distance Formula

274 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Let us consider few sample data which helps to understand the

applicability of those methods and direct to choose the appropriate
distance formula.

Case 1 k-means using Squared Euclidean distance formula.

Squared Euclidean distance formula is basically used to minimize the distance

within in the cluster. Jenks has come also used the concept of k-means using
squared deviation mean for cartography [ 10], and the squared distance is used in
many scenario’s.
k-means has proved over many complicated algorithm with a better
performance in classification, as discussed in [ 11] k-means is suggested to recover
the mixture of matrix which have more sparsity and also unlock the ICA- like
properties.
k-means is also gives effective result to classify the network traffic, as described
in [ 12] its been proved as effective and faster model to classify the network traffic.
Data set used in [ 12] by Erman J are two empirical packet traces, One is a
publicly available packet trace and other one is full packet trace such as Auckland
IV 2 and the University of Calgary respectively. The classification of internet
traffic to clustering using DBSCAN,k-means and Autoclass in which the building
time of model is comparitively less and more efficeint in k-means whereas
Autoclass is very time consuming and DBSCAN accuracy is very low. This
conclude that k-means is faster, accurate and efficient for clustering.

Case 2 k-means Clustering using Manhattan Distance.

Data mining is widely used in many field for extracting the data in high
dimensional data set at the same time challenges of searching and retrieving the
relevant data also triggering task. k-means with Manhattan distance or Taxicab
formula is well suited in such situation. As discussed in “On the Surprising
Behavior of Distance Metrics in High Dimensional Space” by Charu C. Aggarwal
[ 13], high dimensional data set with 2 or 3 D spatial system which is having
fractional matrix is very common in data mining in which applying Euclidean
distance which is best suited for integral distance is irrelevant for this hence
Manhattan distance has come out with best suitable and efficient distance
formula. Dibya in [ 14] has discussed the performance of Manhattan distance on
data set such as Iris and Wine Data set and the resultant proved that the
computation time is less compare with Euclidean and Cosine distance formula.
Hence Manhattan distance is more suitable for the data set which are in the
 SDN CONTROLLER PLACEMENT 275

grid format between a pair of coordinator. Though the computation is expensive

it is significantly resistant.

Case 3 k-means Clustering using Minkowski Distance.

Minkowski distance algorithm is used to measure the distance between two points
in a normal vector space as a generalization for both Manhattan and Euclidean
distance. It is an effective in fuzzy logic, networking, spatial data and many. In
[ 15], the performance of minkowski distance is proven on spatial distance for
measuring the distance on the basis of time.
Minkowski Distance has not only used for calculating the distance but the
convergence result with reduction of error and increased the accuracy with less
computation time as discussed in the field of image segmentation in [ 16].
Many researcher has used this distance formula and also extended their work
for improvement of this algorithm which is delimited for the value of p=1 and 2.
As we can see in “Fizzy clustering using Minkowski Distance” in [ 17] and
resultant with efficient and well applied to empirical applications.

Case 4 k-means Clustering using Chebyshev Distance.

k-means clustering is versatile due its performance with varying distance formula,
as discussed above we extended our discussion about chebyshev distance formula.
It gives the maximum absolute distance. Major applications are Chessboard and
warehouse logistics.
Chebyshev distance is used in many field such as in machine learning, IoT
devices and AI as this gives he distance of an object moving away or with
maximum distance.
Below are the some pattern of data representation for the above discussed
distance formula which help us to identifies the pattern before applying the
distance formula.

In this chapter, we have discussed four main distance formulae

with respect to k-mean which represents the qualitative behavior of
the different distance metrics for measuring proximity in high
276 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 12.4 Manhattan and Euclidean Distance

b>a, Chebyshev P=1 or 2,

Minkowski

Figure 12.5 Chebyshev Distance and Minkowski Distance

dimensions. The particular choice of distance metric which is used

from problems such as clustering, categorization, and similarity
search which depends on some notion of proximity. Every formula
have its own significant characteristic for the specific pattern
(Figures 12.4 and 12.5).
In the next section will discuss on mechanism to determine the
optimal number of k in k-means algorithm.

Mechanism to Compute the Value of K

In k-means algorithm k indicate the number of clusters, by default k

is been taken as random number which make be not appropriate and
significantly no valid point in it to over come that many indices are
used by researcher. Here, mainly two point to be noted one is about
the value of and another is to find whether in the data set is there any
possibility to form cluster or not which is decided based on null
hypothesis and alternative hypothesis [18], we are not covering in
detail about hypothesis in this chapter.
Before applying those method to find the value of K, identify the
dimentionality of data set because euclidean distance is not the
 SDN CONTROLLER PLACEMENT 277

Figure 12.6 Methods to Find K

suitable solution for multidimensional data set. Hence PCA should

be implemented on data to speed up the computation.
Determining the k-values have many method in which five main
popular methods are discussed below and represented in the
Figure 12.6 in which most of the methods are dependent on Within
Sum of Square (WSS) and Between Sum of Square (BSS) [19–21].

WSS = x Ci (x ci )2
BSS = i Ci (c ci )2

1. Elbow Curve: This method choose a range of candidate values

of k, then apply k-means clustering using each of the values of k
to draw the inertia. Represent the average distance of each point
to its centroid of the respective cluster. The value of inertia
decrease as k increases and stop declining after reaching sa­
turation value of K.
2. Silhouette Curve: It finds all the average distance within a
cluster and their distance with other cluster object. Once it
computes the minimum value will be taken. It is all about the
relation between cohesion and resolution.
3. Gap Statistics: In this methodology the value of within sum of
square cluster is compared with the different value of K.
4. 4.CH-Index: The CH-index is used to find the best value of k
using with-cluster-sum-of-squares and between-cluster-sum-
of-squares.
Ch is the ration of BSS/WSS.
278 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Table 12.2 Data Set

METHOD NAME DATA SET
1. Elbow Curve Customer Segmentation, News purity filteration, Wireless sensor network.
2. Silhouette Curve Classify the image based on color,Fuzzy clustering, Medical Images
3. Gap Statistics Online phenotype discover, Blind source seperation, spike sorting
4. CH Index Crude oil, Ionosphere and cancer dataset
5. DB Index Cereals data, Web navigations

5. DB-Index: It will be constructed based on the ratio between

within the cluster and between the cluster distance. It consider
the smallest DB value as optimal value. In most of the data set it
provides the acute value of K compare to any other algorithm.
Lets Compare these methods applicability in Table 12.2 [2].

Properties of k-Means

Properties of k-means algorithm: Clustering the labeled value or

supervised clustering can be done by any labeled value which is
common factor to differentiate the group such as gender for popu­
lation, Product name in mall, rating of a cloths or disease name in
patients are some of the example but clustering the unlabeled data set
or unsupervised clustering is quiet different and complex while
grouping due its nature of data such as networking traffic to find the
malicious data flow, chromosomes behavior, image compression,
Eruptions float and quality of soil and many more where there is no
weight labeled to classify the data set.
k-means algorithm is a most popular algorithm which is more
suitable for unsupervised clustering,as discussed in previous section
the data are not having confined value which needs to be segmented
based on multiple variables or a single variable. It comes under the
categories of partitioning algorithm which resultant with linear
time complexity which does not create any nested clustering hence
overlapping between the cluster is overruled out in this clustering
algorithm.
k-means algorithm can also conjugated with many other methods
for varying performance, will discuss in detail in the next section.
 SDN CONTROLLER PLACEMENT 279

Figure 12.7 k‐Means Clustering in Spherical Structure

Lets see the prominent features of k-means algorithm which

provides the base line to choose it for our respective problem either
for segmentation or compression.
Property 1: k-means clustering is better in creating the spherical
structure like cluster. As shown in the image Figure 12.7, k-means
clustering in a spherical structure gives more weight to bigger clusters
than smaller ones and the smaller cluster may also be neglected.
Property 2: Vector quantization based single processing problem
can opt the k-means due its basic characteristic which adopted
Llody’s algorithm especially for quantization. Consider a simple
example where image compression need to carried out in which the
similar pixel should be compressed by quantization, k-means is well
suited for this as shown in the below image the color with similar
density or pixel value can be easily grouped using k-means which give
better output for heavy weighted ones. Since k-means is fast and have
good trade off for the dataset which may not have large gap in the
dataset. Figure 12.8 depicts the example of colored compression
vector Quantization.
280 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 12.8 Vector Quantization [Each Data Point Will Be in the Voronoi Cell]

Figure 12.9 Mixing Matrix With k=6

Property 3: k-means in Linear independent component analysis,

ICA is basically used in deep learning which is used to disclose the
hidden values or random variable in data set. As discussed in [5], k-
means played the major role in mixing matrix for filtering the data
also the convergence rate is comparatively efficient. Consider the
example discussed in [5] as depicted in Figure 12.9, k-means when
integrated with ICA as come up as powerful tool to recover the
 SDN CONTROLLER PLACEMENT 281

Figure 12.10 Clustered Network Notes Using k‐Means

columns in mixing matrix and is also the best algorithm for dic­
tionary learning.
Property 4: Data points in a data set should be similar but it can be
different from the each clusters.
Consider an example of placing the router in the region based on the
number of host and which are grouped based on either distance or rate of
data transfer. Best part is to select the variable of distance than network
traffic which are varying with time. k-means group the host which are
with minimum distance between the host and plce the centroid to
maintain the minimum distance between the cluster as shown below.
Consider the centroid c1,c2,c3,c4 where c1 is close to c4 and c2 is
close to c3 then the connectivity looks as shown in the Figure 12.10.
As discussed above those 4 property plays a major role while
implementing the clustering using k-means algorithm.

Compare the k-Means with Fuzzy C Mean, K Mediods,

and k-Means++

In the previous section we learnt about k-means algorithm and its

properties, but in some of the research work k-means is been com­
pared with few other unsupervised algorithms which are overcome
some of the drawback of k-means algorithms. Will study there con­
cepts then compare with k-means.
a. Fuzzy C Mean: It is introduced by Bezdek[], this algorithm
data points can be in more than one cluster based on the
membership grade assigned to data points. This algorithm
282 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Table 12.3 Comparison of Unsupervised Clustering Algorithms

ALGORITHMS/ SQUARED QUICK ROBUST COHESION OF CLUSTERING
PERFORMANCE ERROR CONVERGENCE DATA POINTS ACCURACY
PARAMETER MEASURED IN CLUSTER
k-means Yes Average Average Completely Medium
Coupled
K-Mediod No Good Good Completely High
Coupled
k-means ++ Yes Good Good Completely Medium
Coupled
Fuzzy C Mean yes Low Average Partially Medium for
Coupled large
data set

combines with Gaussian and Expectation Maximization algo­

rithm for more accuracy.
b. K Mediods: Mediods is like centroid k-means in which the
datapoints should be minimum to this mediod to form cluster.
As compared with k-mean, it aims to reduce the dissimilarity in
data set by computing the pair wise distance perhaps time delay.
If the value between mediod and data point is more than the
swapping the data point to other mediod.
c. k-means ++: It overcomes the drawback of k-means algorithm
which fails to initialize the centroid. In this algorithm the
centroid chosen was initially in another cluster is the benefit in
reducing the run-time for convergence.

With the above small description we would like to compare those

3 algorithm with K-means with the reference [22,23] (Table 12.3).
The above table helps to conclude the k-means algorithm is using
squared distance error for measuring the accuracy and the data points
are tightly coupled with few points to also be consider with respect to
convergence but due its simplicity and optimal performance this can
be used for large data set also.

k-Means Clustering in Software Defined Network (SDN)

Controlling the network components such as switches, nodes and

also the gateways can be achieved using software by the new
 SDN CONTROLLER PLACEMENT 283

Figure 12.11 The Macrolevel SDN Architecture

technology called SDN. Internet of Thing and growing industries are

making the world digitized and making the job simple with the
power of internet from 2G to 5G.
A small introduction on SDN is as follows [24] (Figure 12.11),
• Northbound API’s are logical application such as Firewalls,
Routing algorithms, load balancer and so on which helps the
network components to communicate with high level components.
• Southbound API’s are used to communicate with switch fabrics,
flows, protocols and network virtualization protocol.
• Controller is placed in intermediary between northbound and
southbound application interface to manage, monitor and direct
the network component centrally.
As we new southbound API’s are the network components which will
scale as the connectivity demand increases at the same time challenges
with respect to scaling also place an major role in research which can
be smoothed by clustering the hosts are switches and controllers.
Many research is heading towards to this challenges as discussed in
[25–27] and many more, more interestingly the logical grouping of
these nodes are also a part of solution in it.
k-means clustering algorithm is implemented by research to place
the controller in optimal place and in [25], its is used to reduce the
latency and to group the incoming network [26]. k-means is not only
for placement but also used for detecting DDOS attack. Barki [27] as
concluded though the k-means is less in accuracy but consume less
time compared to other algorithms.
284 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Exploring the k-Means Clustering in Resolving SDN Challenges

The proliferation of SDN such as centralizing the networking de­

cision using the software component(Controller) has decreased many
issues and improved the network performance in terms of resource
utilization,processing the packets, managing the traffic and updating
the firewalls to switches centrally, along with the benefits many
challenges are branch out.
The challenges such as DDOS attack, Resource utilization such as
bandwidth, traffic management and intercommunication delay be­
tween the controllers. Many researcher have proposed the novel al­
gorithm and the mechanism to resolve and in most of the research
clustering technique is adopted to resolve the issues among which k-
means in been integrated by many. Lets see few of the studies to
know the insight of those research.

Case 1- To reduce the latency in SDN: Placing the SDN

controller in the appropriate place which helps to reduce
the latency, as discussed in Network Partitioning using k-
means [25] the latency compared to standard connectivity
and clustering using k-means there is an decrease in latency
for k=5,6 about achieved the latency around 2.3ms which is
smaller than non clustering mechanism.
Case 2- DDoS attack monitoring in SDN: Felipe A.
Lopes [28], has proven the k-means influence in mon­
itoring DDoS attack and the implemented to detect the
abnormal dataset using RYU application and successfully
found monitor the network traffic for the attack.
Case 3- Traffic Monitoring: Classifying the network
traffic based on the traffics and packet size. As discussed
in [29],the k-means plays a major rule in classifying the
traffic so as to place the controller by checking the sa­
turation point for placing it. k-means also proved the best
accuracy for traffic classification with the result of 90%
accuracy reducing the squared error [29]. As experi­
mented by Liu, the classification of network traffic for the
data from 1000 users created the classification with the
type of data transferred n which http protocol data
 SDN CONTROLLER PLACEMENT 285

Figure 12.12 k‐Means Clustering with SDN and 5G

transmission is high. Along with transformation steps
log data is considered and applied k-means results with
90% of accuracy with the k=80.Installing the SDN to
the same experiment is an novel approach for achieving
more accuracy.

The role of k-means clustering is one of the strategical step to make

the network more robust along with SDN and 5G network. As
depicted below in Figure 12.12, k-means can play a major role in
portioning the network.
As the new technology embed in current technology such as in­
tegration of 5G with SDN improves the Manageability, orchestra­
tion, services and resources. As Network Function Virtualization is
helps to partition the network traffic by creating a virtual LAN and
k-means can help in placing the controller in an efficient position as
well classify the traffic to find the malicious attack or to find the type
of traffic flowing toward network devices.

SDN Network Partitioning of Saren Topology-Based using k-Means

An small experiment conducted using mininet simulator to evidence

the k-means role in placing the controller for managing the switches.
In this chapter, a small implementation is worked on Saren
topologies dataset from Zootopology and computed the centroid for
placing the controllers for better performance in terms of reducing
delay is measured n millisecond.
286 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 12.13 k‐Means Clustering for Saren Topology

 SDN CONTROLLER PLACEMENT 287

Figure 12.14 RTT Time after Applying k-Means Algorithm

As discussed in the k-means algorithm first step is assigning the

dataset to an array as a delay vector for the 7 switches as mentioned in
gml file. Since the switches as very less in number we computed for
the value k=1,2 where the centroid value after applying euclidean
distance formula is 2.6908 and 1.46, 3.5 respectively which is re­
presented in the Figure 12.13.
We illustrated the effectiveness of k-means clustering and found
that k=2 result the reduction of latency compares to k=1 in k-
means and traditional approach. As the graph shows that when k=2
the delay is reduced around 15-20% and improve the response
time (Figure 12.14).

Conclusion

The k-means algorithm is very popular and simple for im­

plementation and is more suitable for the vector based data set. It is
primarily using euclidean distance algorithm which is used to
minimize the squared errors for more accuracy, through it started
the journey in 1967 its powerful concept make it as ubiquity. The
distance formula is varied for the suitable problem for better per­
formance. Its efficiency and applicability is plethora but tried an
small attempt with the example using SDN to highlight its ap­
plicability with best choice for partitioning the data points. Its also
proven by many researcher about the efficiency of k-means and also
we proved the improvement in network latency is reduced and in­
crease in response time intern.
Code for implementing k-means is available in sklearn,
288 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Sklearn---→Cluster----→k-means and refer the github given

below for the source code of k-means in Python.
[https://github.com/scikit-learn/scikit-learn/blob/0fb307bf3/
sklearn/cluster/_kmeans.py#L749]

References
[1] MacQueen, J. B. (1967). Some Methods for classification and
Analysis of Multivariate Observations. Proceedings of 5th Berkeley
Symposium on Mathematical Statistics and Probability.1. University
of California Press. pp. 281–297. MR0214227.Zbl0214.46201.
Retrieved 2009-04-07.
[2] Improved Clustering with Augmented k-means J. Andrew Howe,
Independent Researcher, Riyadh, arXiv:1705.07592[stat.ML], May 2017.
[3] Krishna, K., and M. Murty. 1999. Genetic K-Means algorithm. IEEE
Transactions on Systems, Man, and Cybernetics. Part B, Cybernetics:
a Publication of the IEEE Systems, Man, and Cybernetics Society
vol. 29 (3):pp. 433–439. doi: 10.1109/3477.764879.
[4] Li, Z., & Wang, H. (2019). Improving K-means method via
shrinkage estimation and LVQ algorithm. Communications in
Statistics - Simulation and Computation, vol. 50, pp. 1–16. doi:10.
1080/03610918.2019.1620274
[5] Vinnikov, Alon; Shalev-Shwartz, Shai (2014). “K-means Recovers
ICA Filters when Independent Components are Sparse”(PDF).
Proceedings of the International Conference on Machine Learning
(ICML 2014).
[6] “Supervised and Unsupervised Machine Learning Algorithms” by
byJason BrownleeoninMachine Learning Algorithms
[7] Reffered “Mathworks” for implementation, https://in.mathworks.com/
help/stats/cluster-analysis.html.
[8] MacQueen, J., 1967. Some Methods for Classification and Analysis
of Multivariate Observations. In: Cam, L. M. L., Neyman, J. (Eds.),
Proceedings of 5-th Berkeley Symposium on Mathematical Statistics
and Probability. Vol. 1. University of California, Berkeley, Berkeley,
USA, pp. 281–297.
[9] https://www.datanovia.com/en/lessons/determining-the-optimal-
number-of-clusters-3-must-know-methods/
[10] Jenks, George F. 1967. “The Data Model Concept in Statistical
Mapping”, International Yearbook of Cartography, vol. 7: pp. 186–190.
[11] Vinnikov, Alon; Shalev-Shwartz, Shai (2014). “K-means
Recovers ICA Filters when Independent Components are Sparse”.
Proceedings of the International Conference on Machine Learning
(ICML 2014).
[12] Erman, J., Arlitt, M., Mahanti, A.: Traffic classification using clus­
tering algorithms. In: Proceedings of the 2006 SIGCOMM
 SDN CONTROLLER PLACEMENT 289

Workshop on Mining Network Data, MineNet 2006, pp. 281–286.

ACM, New York (2006).
[13] Groenen, P. J. F., & Jajuga, K. (2001). Fuzzy clustering with squared
Minkowski distances. Fuzzy Sets and Systems, vol. 120(2), pp. 227–237.
doi:10.1016/s0165-0114(98)00403-5.
[14] Dibya Jyoti Bora, Dr. Anil Kumar Gupta. “Effect of Different
Distance Measures on the Performance of K-Means Algorithm: An
Experimental Study in MATLAB”, (IJCSIT) International Journal
of Computer Science and Information Technologies, Vol. 5 (2),
2014, pp. 2501–2506.
[15] M. K. Arzoo, A. Prof, and K. Rathod, “K-Means algorithm with
different distance metrics in spatial data mining with uses of
NetBeans IDE 8. 2,” Int. Res. J. Eng. Technol., vol. 4, no. 4,
pp. 2363–2368, 2017.
[16] “A K-Nearest Neighbor Technique for Brain Tumor Segmentation
Using Minkowski Distance” , Xie, Xiaozhen, Journal of Medical
Imaging and Health Informatics, Volume 8, Number 2, February
2018, pp. 180–185(6).
[17] Groenen, Patrick & Kaymak, Uzay & Rosmalen Joost. (2007). Fuzzy
Clustering with Minkowski Distance Functions. 10.1002/97804
70061190.ch3.
[18] Fuentes, Claudio, and George Casella. “Testing for the existence of
clusters.” SORT (Barcelona) vol. 33, 2 (2009): 115–157.
[19] In the article “Silhouette Analysis vs Elbow Method vs Davies-Bouldin
Index: Selecting the optimal number of clusters for KMeans clustering”,
GEORGIOS DRAKOS, 4th march 2020. https://gdcoder.com/
silhouette-analysis-vs-elbow-method-vs-davies-bouldin-index-selecting-
the-optimal-number-of-clusters-for-kmeans-clustering/
[20] Maulik, U., & Bandyopadhyay, S. (2002). Performance evaluation of
some clustering algorithms and validity indices. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence, 24(12),
1650–1654.doi:10.1109/tpami.2002.1114856
[21] Singh, A. K., Mittal, S., Malhotra, P., & Srivastava, Y. V. (2020).
Clustering Evaluation by Davies-Bouldin Index(DBI) in Cereal
data using K-Means. 2020 Fourth International Conference on
Computing Methodologies and Communication (ICCMC). doi:
10.1109/iccmc48092.2020.iccmc-00057.
[22] B. Simhachalam, G. Ganesan, “Performance comparison of fuzzy
and non-fuzzy classification methods”, Egyptian Informatics Journal
Volume 17, Issue 2, July 2016, Pages 183–188, ScienceDirect.
[23] Kalpit G. Soni,and Dr. Atul Patel, “Comparative Analysis of K-
means and K-medoids Algorithm on IRIS Data”, International
Journal of Computational Intelligence Research ISSN 0973-1873
Volume 13, Number 5 (2017), pp. 899–906.
[24] Pushpa. J. Research and P. Raj, “Topology-based analysis of
290 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

performance evaluation of centralized vs. distributed SDN controller,”

2018 IEEE International Conference on Current Trends in Advanced
Computing (ICCTAC), Bangalore, 2018, pp. 1–8, doi: 10.1109/
ICCTAC.2018.8370394.
[25] G. Wang, Y. Zhao, J. Huang, Q. Duan, and J. Li, “A k-means-based
network partition algorithm for controller placement in software
defined network,” in 2016 IEEE Int. Conf. on Communications
(ICC), pp. 1–6, May 2016
[26] Z. Lv et al., “An Optimizing and Differentially Private Clustering
Algorithm for Mixed Data in SDN-Based Smart Grid,” in IEEE
Access, vol. 7, pp. 45773–45782, 2019, doi: 10.1109/ACCESS.2019.
2909048.
[27] Barki, L., Shidling, A., Meti, N., Narayan, D. G., & Mulla, M. M.
(2016). Detection of distributed denial of service attacks in software
defined networks. 2016 International Conference on Advances in
Computing, Communications and Informatics (ICACCI). doi:10.
1109/icacci.2016.7732445.
[28] https://github.com/felipealencar/sdn-ddos-monitor
[29] Liu, Y., Li, W., & Li, Y.-C. (2007). Network Traffic Classification
Using K-means Clustering. Second International Multi-Symposiums on
Computer and Computational Sciences (IMSCCS 2007). doi:10.1109/
imsccs.2007.52.
 13
AN INTELLIGENT DEEP
LEARNING-BASED WIRELESS
UNDERGROUND SENSOR SYSTEM
FOR IOT-BASED AGRICULTURAL
APPLICATION
PRISCILLA RAJADURAI1
AND G. JASPHER W. KATHRINE2
1
St. Joseph’s Institute of Technology,
Chennai, India
2
Karunya Institute of Technology
and Sciences, Coimbatore, India

Contents

Introduction 292
System Overview 292
Scope of the Project 293
Literature Survey 294
Underground Communications Networks 294
WUSN in Crop Production 295
Test Bed Development for WUSN 296
Soil Moisture Measurement 297
Underground Mining 297
Underground Mine Communications 297
Occupational Health Hazard in Mining 298
Wireless Underground Sensor 298
Agriculture Field Monitoring 298
Spatio-Temporal Soil Moisture Measurement 299
Channel Model and Analysis for Wireless Underground
Sensor Networks in Soil Medium 299
System Analysis 300
Existing System 300
Proposed System 300

DOI: 10.1201/9781003119838-13 291

292 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

System Requirements 301

Software Requirements 301
Hardware Requirements 301
Module Description 302
System Architecture 318
Architecture Description 318
Wireless Underground Sensor Design Architecture 318
Components of WUSN 319
Cloud 319
Arduino 319
pH Sensor 320
Soil Moisture Sensor 320
Transmitter and Receiver 320
Conclusion 321
References 323

Introduction
System Overview

Wireless underground sensor networks (WUSNs), which consist

of wireless sensors buried underground, are a natural extension of
the WSN phenomenon and have been considered as a potential
field that will enable a wide variety of novel applications that were
not possible before. Compared to the current underground sensor
networks, which use wired communication methods for data re­
trieval, WUSNs have several remarkable merits, such as conceal­
ment, ease of deployment, timeliness of data, reliability, and
coverage density. The realization of wireless underground com­
munication and networking techniques will lead to potential ap­
plications in the fields of agriculture, border patrol, assisted
navigation, sports field maintenance, intruder detection, and in­
frastructure monitoring. Many of these novel applications will be
possible because WUSNs can provide localized and real-time data
about a specific soil region and its surrounding area. Despite its
potential advantages, several open research problems exist to make
WUSNs feasible.
 WIRELESS UNDERGROUND SENSOR SYSTEM 293

The main challenge is the realization of efficient and reliable un­

derground wireless communication between buried sensors. Aspects
such as temperature, weather, soil composition, soil moisture, and soil
homogeneity directly impact the connectivity and communication
success. In fact, underground communication is one of the few fields
where the environment has a significant and direct impact on the
communication performance. Hence, characterization of the wireless
underground channel is essential for the proliferation of communica­
tion protocols for WUSNs. In this chapter, we introduce theoretical
models to characterize the underground wireless channel. The models
developed in this chapter characterize not only the propagation of EM
wave in soil, but also other effects on the communication related to
multipath effects, soil composition, water content, and burial depth.
The results obtained from this formalization reveal that underground
communication is severely affected by frequency and soil properties,
and more specifically by the volumetric water content (VWC) of soil.
In addition to nodes buried underground, the existence of
aboveground nodes is also necessary for WUSN due to data retrieval
purposes. Accordingly, two communication types coexist in WUSNs.
Underground-to-underground communication refers to the in­
formation exchange between buried sensors for network management
and data relay purposes. The data in the network are then collected
through underground-to-aboveground communication. This type of
communication also includes command and control information
from aboveground stations to underground. The classification also
highlights significant differences on the challenges related to each
option. Underground-to-aboveground communication presents
better quality than underground-to-underground, since certain por­
tion of the communication takes place over the air. On the other
hand, underground-to-underground communication usually presents
more difficult challenges for the design and the deployment of
WUSNs. In this chapter, we focus on underground-to-underground
communication for WUSNs.

Scope of the Project

The scope of this work is to show that the data can be transferred
through soil wirelessly, to help the farmers to measure the moisture
294 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

of the soil wirelessly and intimate the farmers through monitor, and
to help maintain the soil moisture in the agriculture field. It also
helps to water the soil when automatically the moisture of the soil
reduces to certain level. It is done using WUSN sensor and the
embedded C is needed in Arduino.

Literature Survey

Wireless underground sensor networks constitute one of the pro­

mising application areas of the recently developed WSN techniques.
WUSN devices are deployed completely below the ground and do
not require any wired connections. Each device contains all necessary
sensors, memory, a processor, a radio, an antenna, and a power
source. This makes their deployment much simpler than existing
underground sensing solutions. Wireless communication within a
dense substance such as soil or rock is, however, significantly more
challenging than through air.

Underground Communications Networks

In agriculture, WSNs have been widely promoted as a means to

improve on-farm yield and profitability through the provision of real-
time or on-demand sensed data. Recent developments in commu­
nication techniques for WSN have resulted in a vast number of
applications in agriculture. WSN is used mainly for management
decisions of irrigation water resources, storage management of agri­
cultural products, time determination of the crop harvest, character­
istics of crop growth, forecast of fertilizer demand [1], and so on.
The WSN was a kind of network whose energy was limited,
reasonable clusters were determined, and the energy model used in
communication was given according to the number of nodes and
characteristics of regional distribution [2]. Collection node system of
farmland information based on WSN was designed [3].
The main difference between WUSNs and the terrestrial WSNs is
the communication medium. The propagation characteristics of
electromagnetic (EM) waves in soil and the significant differences
between propagation in air prevent a straightforward characterization
of the underground wireless channel [4]. The effects of variations in
 WIRELESS UNDERGROUND SENSOR SYSTEM 295

soil moisture are investigated through field measurement results. The

theoretical analysis and the simulation results prove the feasibility of
wireless communication in underground environment and highlight
several important aspects in this field [5].
Dohare et al. [6] presented a comprehensive review of existing
approaches for mine monitoring has. For the base protocol, most
works have relied on Zigbee for its ease of deployment, low data rate
(250 kb/s), substantial range and most importantly low power con­
sumption when compared to other technologies, such as Wi-Fi,
Bluetooth, and ultra-wideband communication.
Yarkan et al. [7] proposed that after a recent series of unfortunate
underground mining disasters, the vital importance of communications
for underground mining is underlined one more time. Establishing
reliable communication is a very difficult task for underground mining
due to the extreme environmental conditions. Until now, no single
communication system exists, which can solve all of the problems and
difficulties encountered in underground mine communications.
Donoghue [8] reviewed occupational health hazards in mining
and found occupational exposure limits (OELs) for hazardous sub­
stances. The physical, compound, natural, ergonomic, and psycho­
social word-related wellbeing perils of mining and related
metallurgical forms are depicted. Mining remains a critical modern
segment in many parts of the world and albeit significant advance has
been made in the control of word-related wellbeing dangers; there
remains space for further hazard decrease.
Kiran Lakkaraju et al. described that the number of attacks against
large computer systems is currently growing at a rapid pace. Despite
the best efforts of security analysts, large organizations are having
trouble keeping on top of the current state of their networks. A tool
called NVisionIP is designed to increase the security analyst’s si­
tuational awareness. As humans are inherently visual beings,
NVisionIP uses a graphical representation of a class-B network to
allow analysts to quickly visualize the current state of their network.

WUSN in Crop Production

Agnelo et al. [9] studied the influence of the communication per­

formance between the terrestrial nodes and the underground nodes in
296 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

some factors, including antenna bandwidth of WSN nodes in 433

MHz frequency, the buried depth of nodes in the soil and water
content of the soil. The temperature change in different position of
feed warehouse was monitored through WSNs which were in­
troduced by Green et al. [10]. Among all these technologies, the
agriculture domain is mostly explored concerning the application of
WSNs in improving the traditional methods of farming. Wireless
sensor units (WSUs): Each WSU, deployed on-field, has four dif­
ferent type of components: application-specific sensors, processing
unit, radio transceiver, and battery power.
X. Q. Yu et al. [11] conducted experimental measurements with
commodity sensor motes at the frequency of 2.4 GHz and 433 MHz,
respectively. Experiments are run to examine the received signal
strength of correctly received packets and the packet error rate for a
communication link. The tests show the potential feasibility of the
WUSN with the use of powerful RF transceivers at 433-MHz fre­
quency. Moreover, a classification for WUSN communication is
shown. Finally, the effects of burial depth, internode distance, and
volumetric water content of the soil on the signal strength and packet
error rate in communication of WUSN are concluded.

Test Bed Development for WUSN

Xin Dong et al. [12] estimated the distortion of distributed soil

moisture measurement using WUSNs. The main focus of the work is
to analyze the impact of the environment and network parameters on
the estimation distortion of the soil moisture. More specifically, the
effects of rainfall, soil porosity, and vegetation root zone are in­
vestigated by exploiting a rainfall model, in addition to the effects of
sampling rate, network topology, and measurement signal noise ratio.
Spatio-temporal correlation is characterized to develop a measurement
distortion model with respect to these factors. The evaluations reveal
that with porous soil and shallow vegetation roots, high sampling rate
is required for sufficient accuracy. In addition, the impact of rainfall on
the estimation distortion has also been investigated. In a storm, which
carries on a large area and lasts for a long time, the estimation dis­
tortion is decreased because of the increase in spatial correlation.
Moreover, only few closest sensors are needed to estimate the values of
 WIRELESS UNDERGROUND SENSOR SYSTEM 297

an interested location. These findings are utilized to guide the design

of WUSNs for soil moisture measurement to reduce the density of
network and the sampling rate of the sensors, but at the same time
maintain the performance of the system.
Underground is more complicated than the terrestrial environ­
ment, which contains not only air but also sand, rocks, and water
with electrolyte. It is challenging to realize wireless communication
in such complex environments. Underground communication solu­
tions, while mostly shielded from the environment, require the ability
to communicate through soil and adjust its parameters to adapt to
dynamic changes in soil with the use of temperature, soil moisture,
electric conductivity, and water potential sensors [13].

Soil Moisture Measurement

Ritsema Kuipers and Kleiboer [14–23] developed a new stand-alone

wireless embedded network system recently for continuous mon­
itoring of soil water contents at multiple depths. This information on
the technical aspects of the system, including the applied sensor
technology, the wireless communication protocols, the gateway sta­
tion for data collection, and data transfer to an end user Web page for
disseminating results, is provided to targeted audiences. Results from
the first test of the network system are presented and discussed,
including lessons learned so far and actions to be undertaken in the
near future to improve and enhance the operability of this innovative
measurement approach.

Underground Mining

Underground Mine Communications Serhan Yarkan et al. proposed that

after a recent series of unfortunate underground mining disasters, the
vital importance of communications for underground mining is un­
derlined one more time. Establishing reliable communication is a very
difficult task for underground mining due to the extreme environ­
mental conditions. Until now, no single communication system exists
which can solve all of the problems and difficulties encountered in
underground mine communications. However, combining research
with previous experiences might help existing systems improve, if not
298 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

completely solve all of the problems. In this survey, underground mine

communication is investigated. Major issues which underground
mine communication systems must take into account are discussed.
Communication types, methods, and their significance are presented.

Occupational Health Hazard in Mining M. Donoghue described the

physical, chemical, biological, ergonomic, and psychosocial occupa­
tional health hazards of mining and associated metallurgical pro­
cesses. Mining remains an important industrial sector in many parts
of the world and although substantial progress has been made in the
control of occupational health hazards, there remains room for fur­
ther risk reduction. This applies particularly to traumatic injury ha­
zards, ergonomic hazards, and noise. Vigilance is also required to
ensure exposures to coal dust and crystalline silica remain effectively
controlled.

Wireless Underground Sensor

Agriculture Field Monitoring Per Agnelo R. Silva et al., the WSN is

now a day widely used to build decision support for overcoming
many problems in real world. One of the most interesting fields
having an increasing need of decision support systems is precision
agriculture (PA). The purpose of this paper is to design and develop
an agricultural monitoring system using WSN to increase the pro­
ductivity and quality of farming without observing it for all the time
manually. Temperature, humidity, and water levels are the most
important factors for the productivity, growth, and quality of plants
in agriculture. The temperature, humidity, and water level sensors are
deployed to gather the temperature and humidity values. The sensor
has to transmit the gathered information through the wireless
communication network to the data server (cloud). The IoT gateway
is in charge of the communication between the remote-controll serial
devices and central control system. The farmers or the agriculture
experts can observe the measurements from the web simultaneously.
With the continuous monitoring of many environmental parameters,
the grower can analyze the optimal environmental conditions to
achieve maximum crop productiveness, for the better productivity,
and to achieve remarkable energy savings.
 WIRELESS UNDERGROUND SENSOR SYSTEM 299

Spatio-Temporal Soil Moisture Measurement Xin Dong et al. in­

vestigated the estimation distortion of distributed soil moisture
measurement using WUSNs. The main focus of this paper is to
analyze the impact of the environment and network parameters on
the estimation distortion of the soil moisture. More specifically, the
effects of rainfall, soil porosity, and vegetation root zone are in­
vestigated by exploiting a rainfall model, in addition to the effects of
sampling rate, network topology, and measurement signal noise ratio.
Spatio-temporal correlation is characterized to develop a measure­
ment distortion model with respect to these factors. The evaluations
reveal that with porous soil and shallow vegetation roots, high
sampling rate is required for sufficient accuracy. In addition, the
impact of rainfall on the estimation distortion has also been in­
vestigated. In a storm, which carries on a large area and lasts for a
long time, the estimation distortion is decreased because of the in­
crease in spatial correlation. Moreover, only few closest sensors are
needed to estimate the values of an interested location. These find­
ings are utilized to guide the design of WUSNs for soil moisture
measurement to reduce the density of network and the sampling rate
of the sensors but at the same time maintain the performance of
the system.

Channel Model and Analysis for Wireless Underground Sensor Networks in

Soil Medium Mehmet C. Vurana et al. WUSNs constitute one of
the promising application areas of the recently developed WSN
techniques. The main difference between WUSNs and the terrestrial
WSNs is the communication medium. The propagation character­
istics of electromagnetic (EM) waves in soil and the significant dif­
ferences between propagation in air prevent a straightforward
characterization of the underground wireless channel. To this end, in
this paper, advanced channel models are derived to characterize the
underground wireless channel and the foundational issues for effi­
cient communication through soil are discussed. In particular, the
underground communication channel is modeled considering not
only the propagation of EM waves in soil, but also the other effects,
such as multipath, soil composition, soil moisture, and burial depth.
The propagation characteristics are investigated through simulation
results of path loss between two underground sensors. Moreover,
300 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

based on the proposed channel model, the resulting bit error rate is
analyzed for different network and soil parameters.

System Analysis
Existing System

WSNs have vast number of applications. Among these applications,

WUSNs are a high promising field. WUSNs consist of wireless sen­
sors, which are buried underground. WUSNs will enable a wide
variety of novel applications that were not possible with current wired
underground monitoring techniques. Compared to the current un­
derground sensor networks, which use wired communication methods
for network deployment, WUSNs have several remarkable merits,
such as concealment, ease of deployment, timeliness of data, reliability,
and coverage density.
WUSNs can be used to monitor soil conditions maintaining
parameters such as water content, mineral content, salinity, and
temperature at optimal levels. Real-time knowledge of soil conditions
is also useful for landscaping, where WUSNs can be combined with
automatic sprinkler systems such that grass, trees, and flowers are
watered only when needed.

Proposed System

WUSNs constitute one of the promising application areas of the

recently developed WSN techniques. WUSN is a specialized kind of
WSN that mainly focuses on the use of sensors at the subsurface
region of the soil, that is, the top few meters of the soil. For a long
time, this region has been used to bury sensors, usually targeting
irrigation and environment monitoring applications, although
without wireless communication capability; WUSNs promise to fill
this gap and to provide the infrastructure for novel applications. The
main difference between WUSNs and the terrestrial WSNs is the
communication medium. In fact, the differences between the pro­
pagation of electromagnetic (EM) waves in soil and in air are so
significant that a complete characterization of the underground
wireless channel was only available recently.
 WIRELESS UNDERGROUND SENSOR SYSTEM 301

In this project, we propose WUSN-based communication for

agricultural land areas eliminating existing communication channels
like wired and wireless sensor networks. The agricultural sensors like
soil moisture and pH level of water of different parts of the agri­
cultural land areas are obtained using WUSN circuit, which is
emerged into the soil. Using soil as the medium of communication,
the sensor values are transmitted and received in the base station
suing WUSN transmitter and receiver. Once the base station receives
the data, we uploading the agricultural land features into the cloud
for backup and easy access by authorized persons.

System Requirements

A 64-bit PC can handle larger amounts of information than a 32-bit

system. Since it can use a RAM of 4 GB, a 64-bit computer can be
more responsive when running lots of programs at once. 1 GB of RAM
is considered low and most desktop and laptop computers comes stock
with at least 2 GB, but usually more. Most softwares define two sets of
system requirements, i.e., minimum and recommended.

Software Requirements A software requirement specification is a de­

scription of a software system to be developed. It lays out functional
and non-functional requirements and it also describes the operating
system and tools used in the system and they are:
• MPLAB IDE
• Embedded C

Hardware Requirements Hardware specifications are technical de­

scription of the computer’s components and capabilities, such as
processor speed, model, manufacturer, etc. So the hardware com­
ponents required for the proposed system are:
• PIC microcontroller
• Soil moisture sensor
• Humidity sensor
• LCD
• WUSN TX and RX
302 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

Module Description

• Interfacing sensors
• Programming microcontroller
• Data transmission through soil
• Visual basic
• Internet of things

Interfacing Sensor – Humidity Sensor

General Description A humidity sensor senses and measures both
moisture and air temperature. The sensor is composed of two metal
plates and contains a non-conductive polymer film between them.
This film collects moisture from the air, which causes the voltage
between the two plates to change. These voltage changes are
converted into digital readings showing the level of moisture in
the air.

Product Description
Humidity measurement can be done using electronic hygrometers.
Electronic-type hygrometers or humidity sensors can be broadly
divided into two categories, namely, capacitive sensing effects and
resistive sensing effects. Resistive-type humidity sensors pick up
changes in the resistance value of the sensor element in response
to the change in the humidity. Thick film conductor of precious
metals like gold and ruthenium oxide is printed and culminated
in the shape of the comb to form an electrode. Then a polym-
eric film is applied on the electrode; the film acts as a humidity
sensing film due to the existence of movable ions. Change
in impedance occurs due to the change in the number of mo­
vable ions.

Features

• Input voltage: 5 v
• Output: analog(0–5 v)
• High performance
• Long-term stability
• Close tolerances
• Low cost
 WIRELESS UNDERGROUND SENSOR SYSTEM 303

Applications

• Air conditioners
• Climate control for greenhouses
• Storage and warehouses
• Meteorological applications

Soil Moisture Sensor

General Description
Use the soil moisture sensor just as you would a traditional soil
moisture meter with the additional advantages of automated data
collection, graSoil moisturizing, and data analysis. Typical activities
using our soil moisture sensor include acid-base titrations, studies of
household acids and bases, monitoring of soil moisture change
during chemical reactions or in an aquarium as a result of soil
moisture to synthesis, investigations of acid rain and buffering, and
analysis of water quality in streams and lakes.

Product Description
A soil moisture sensor, shown in Figure 13.1, is a device that mea­
sures the hydrogen-ion concentration (soil moisture) in a solution,
indicating its acidity or alkalinity. In addition to measuring the soil
moisture of liquids, it can also measure the moist and light level. The
soil moisture sensor has an inbuilt meter to measure the light in­
tensity. The soil moisture of a solution indicates how acidic or basic
(alkaline) it is. The soil moisture term translates the values of the
hydrogen ion concentration – which ordinarily ranges between about
1 and 10 × – 14 g equivalents per liter – into numbers between 0
and 14.

Features

• Supply voltage: 5 VDC

• Soil tester moisture – light – soil moisture
• Output: analog

Applications

• Acid-base titrations
• Analysis of water quality in streams and lakes.
304 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 13.1 Moisture Detector Sensor

Programming Microcontroller

• Only 35 single-word instructions to learn.

• All single-cycle instructions except for program branches, which
are two-cycles.
• Operating speed: DC – 20 MHz clock input DC – 200 ns
instruction cycle.
• Up to 8K × 14 words of Flash program memory, Up to 368 × 8
bytes of data memory (RAM), Up to 256 × 8 bytes of
EEPROM data memory is shown in Figure 13.2.
 WIRELESS UNDERGROUND SENSOR SYSTEM 305

Figure 13.2 Pin Diagram of 877a

• Pin out compatible to other 28-pin or40/44-pin

• PIC16CXXX and PIC16FXXX microcontrollers are used in
the microcontroller design as shown in Figure 13.3.

Perisoil Moistural Feature

• Timer0: 8-bit timer/counter with 8-bit prescaler

• Timer1: 16-bit timer/counter with prescaler, can be in­
cremented during sleep via external crystal/clock

Figure 13.3 Microcontroller Design

306 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

• Timer2: 8-bit timer/counter with 8-bit period register, prescaler

and postscaler
• Two capture, compare, PWM modules
• Capture is 16-bit, max. resolution is 12.5 ns
• Compare is 16-bit, max. resolution is 200 ns
• PWM max. resolution is10-bit
• Synchronous serial port (SSP) with SPI™ (master mode) and
I2C™ (master/slave)
• Universal synchronous asynchronous receiver transmitter
(USART/SCI) with
• 9-bit address detection
• Parallel slave port (PSP) – 8 bits wide with external RD, WR
and CS controls (40/44-pin only)
• Brown-out detection circuitry for brown-out reset (BOR)

Analog Features

• 10-bit, up to 8-channel analog-to-digital converter(A/D)

• Brown-out reset (BOR)
• Analog comparator module with:
• Two analog comparators
• Programmable on-chip voltage reference (VREF) module
• Programmable input multiplexing from device inputs and
internal voltage reference
• Comparator outputs are externally accessible

Special Microcontroller Features

• 100,000 erase/write cycle enhanced Flash program memory

typical
• 1,000,000 erase/write cycle data EEPROM memory typical
• Data EEPROM retention > 40 years
• Self-reprogrammable under software control
• In-Circuit Serial Programming™ (ICSP™) via two pins
• Single-supply 5-V ICSP
• Watchdog timer (WDT) with its own on-chip RC oscillator for
reliable operation
• Programmable code protection
• Power-saving sleep mode
 WIRELESS UNDERGROUND SENSOR SYSTEM 307

• Selectable oscillator options

• In-Circuit Debug (ICD) via two pins

CMOS Technology

• Low-power, high-speed Flash/EEPROM technology

• Fully static design
• Wide operating voltage range (2.0 V to 5.5 V)
• Commercial and industrial temperature ranges
• Low-power consumption

Pin Diagram

Device Overview This document contains device-specific in­

formation about the following devices:
• PIC16F873A
• PIC16F874A
• PIC16F876A
• PIC16F877A
PIC16F873A/876A devices are available only in 28-pin packages,
while PIC16F874A/877A devices are available in 40-pin and 44-pin
packages. All devices in the PIC16F87XA family share common
architecture with the following differences:
• The PIC16F873A and PIC16F874A have one-half of the total
on-chip memory of the PIC16F876A andPIC16F877A
• The 28-pin devices have three I/O ports, while the 40/44-pin
devices have five
• The 28-pin devices have 14 interrupts, while the 40/44-pin
devices have 15
• The 28-pin devices have five A/D input channels, while the40/
44-pin devices have eight
• The parallel slave port is implemented only on the 40/44-pin
devices
Block diagrams of the PIC16F874A/877A devices are provided in
Figure 13.2. Additional information may be found in the PICmicro®
Mid-Range Reference Manual (DS33023), which may be obtained
308 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

from your local microchip sales representative or downloaded from

the microchip website. The reference manual should be considered
a complementary document to this data sheet and is highly re­
commended reading for a better understanding of the device archi­
tecture and operation of the Perisoil Moistural modules.

Memory Organization There are three memory blocks in each of the

PIC16F87XA devices. The program memory and data memory have
separate buses so that concurrent access can occur and is detailed in this
section. The EEPROM data memory block is detailed in the
Programming Microcontroller section and shown in Figure 13.4.

Data Memory Organization The data memory is partitioned into

multiple banks which contain the general purpose registers and the
special function registers (SFRs) as shown in Figure 13.4. Bits RP1
(Status<6>) and RP0 (Status<5>) are the bank select bits. Each bank
extends up to 7 Fh (128 bytes). The lower locations of each bank are
reserved for the SFRs. Above the SFRs are general purpose registers,
implemented as static RAM. All implemented banks contain SFRs.

Figure 13.4 Memory Organization of a Controller

 WIRELESS UNDERGROUND SENSOR SYSTEM 309

Some frequently used SFRs from one bank may be mirrored in an­
other bank for code reduction and quicker access.

I/O ports: Some pins for these I/O ports are multiplexed with an
alternate function for the Perisoil Moistural features on the device. In
general, when a Perisoil Moistural is enabled, that pin may not be
used as a general purpose I/O pin. Additional information on I/O
ports may be found in the PICmicro™ Mid-Range Reference
Manual (DS33023).

PORTA and the TRISA Registers PORTA is a 6-bit wide, bidir­

ectional port. The corresponding data direction register is TRISA.
Setting a TRISA bit (= 1) will make the corresponding PORTA pin
an input (i.e., put the corresponding output driver in a high-
impedance mode). Clearing a TRISA bit (= 0) will make the cor­
responding PORTA pin an output (i.e., put the contents of the
output latch on the selected pin). Reading the PORTA register reads
the status of the pins, whereas writing to it will write to the port
latch. All write operations are read-modify-write operations.
Therefore, a write to a port implies that the port pins are read; the
value is modified and then written to the port data latch.
Pin RA4 is multiplexed with the Timer0 module clock input to
become the RA4/T0CKI pin. The RA4/T0CKI pin is a Schmitt
Trigger input and an open-drain output. All other PORTA pins
have TTL input levels and full CMOS output drivers. Other
PORTA pins are multiplexed with analog inputs and the analog
VREF input for both the A/D converters and the comparators. The
operation of each pin is selected by clearing/setting the appropriate
control bits in the ADCON1 and/or CMCON registers. The
TRISA register controls the direction of the port pins even when
they are being used as analog inputs. The user must ensure that the
bits in the TRISA register are maintained set when using them as
analog inputs.

PORTB and the TRISB Registers PORTB is an 8-bit wide, bidir­

ectional port. The corresponding data direction register is TRISB.
Setting a TRISB bit (= 1) will make the corresponding PORTB pin an
input (i.e., put the corresponding output driver in a high-impedance
310 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

mode). Clearing a TRISB bit (= 0) will make the corresponding

PORTB pin an output (i.e., put the contents of the output latch on the
selected pin). Three pins of PORTB are multiplexed with the in-
circuit debugger and low-voltage programming function: RB3/PGM,
RB6/PGC and RB7/PGD. “ Each of the PORTB pins has a weak
internal pull-up. A single control bit can turn on all the pull-ups. This
is performed by clearing bit RBPU (OPTION_REG<7>). The weak
pull-up is automatically turned off when the port pin is configured as
an output. The pull-ups are disabled on a power-on reset.
This interrupt can wake the device from sleep. The user, in the
interrupt service routine, can clear the interrupt in the following
manner:
a. Any read or write of PORTB. This will end the mismatch
condition.
b. Clear flag bit RBIF.
A mismatch condition will continue to set flag bit RBIF. Reading
PORTB will end the mismatch condition and allow flag bit RBIF to
be cleared. The interrupt-on-change feature is recommended for
wake-up on key depression operation and operations where PORTB
is only used for the interrupt-on-change feature. Polling of PORTB
is not recommended while using the interrupt-on-change feature.
This interrupt-on-mismatch feature, together with software config­
urable pull-ups on these four pins, allow easy interface to a keypad
and make it possible for wake-up on key depression.

PORTC and the TRISC Register: PORTC is an 8-bit wide, bi­

directional port. The corresponding data direction register is TRISC.
Setting a TRISC bit (= 1) will make the corresponding PORTC pin
an input (i.e., put the corresponding output driver in a high-
impedance mode). Clearing a TRISC bit (= 0) will make the cor­
responding PORTC pin an output (i.e., put the contents of the
output latch on the selected pin). PORTC is multiplexed with several
per Soil Moistureral functions. PORTC pins have Schmitt Trigger
input buffers. When the I2C module is enabled, the PORTC<4:3>
pins can be configured with normal I2C levels, or with SMBus levels,
by using the CKE bit (SSPSTAT<6>). When enabling Perisoil
Moistural functions, care should be taken in defining TRIS bits for
 WIRELESS UNDERGROUND SENSOR SYSTEM 311

each PORTC pin. Some Perisoil Moistural override the TRIS bit to
make a pin an output, while other Perisoil Moistural override the
TRIS bit to make a pin an input. Since the TRIS bit override is in
effect, while the Perisoil Moistural is enabled, read-modify write
instructions (BSF, BCF, and XORWF) with TRISC as the desti­
nation should be avoided. The user should refer to the corresponding
Perisoil Moistural section for the correct TRIS bit settings.

PORTD and TRISD Registers PORTD is an 8-bit port with

Schmitt Trigger input buffers. Each pin is individually configurable
as an input or output. PORTD can be configured as an 8-bit wide
microprocessor port (parallel slave port) by setting control bit,
PSPMODE (TRISE<4>). In this mode, the input buffers are TTL.

PORTE and TRISE Registers PORTE has three pins (RE0/RD/

AN5, RE1/WR/AN6, and RE2/CS/AN7) which are individually
configurable as inputs or outputs. These pins have Schmitt Trigger
input buffers. The PORTE pins become the I/O control inputs for
the microprocessor port when bit PSPMODE (TRISE<4>) is set. In
this mode, the user must make certain that the TRISE<2:0> bits are
set and that the pins are configured as digital inputs. Also, ensure
that ADCON1 is configured for digital I/O. In this mode, the input
buffers are TTL.
The TRISE register which also controls the parallel slave port
operation. PORTE pins are multiplexed with analog inputs. When
selected for analog input, these pins will read as “0”s. TRISE controls
the direction of the RE pins, even when they are being used as analog
inputs. The user must make sure to keep the pins configured as in­
puts when using them as analog inputs.

Features of Microcontroller Microcontroller is the simplest com­

puter processor which is used as the “brain” of the future system.
Depending on the need, the manufacturer can extend the memory,
converters, timers, input/output lines, etc. which are packaged in a
standard format. Simple software can be used to control the mi­
crocontroller. Hence microcontroller has become man's invisible
companion.
312 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

Their incredible simplicity and flexibility conquered us a long time

ago and if you try to invent something about them, you should know
that you are probably late, someone before you has either done it or at
least has tried to do it.
The following things have had a crucial influence on development
and success of the microcontrollers:

• Powerful and carefully chosen electronics embedded in the

microcontrollers can independently or via input/output de­
vices (switches, push buttons, sensors, LCD displays, relays,
etc.), control various processes and devices, such as industrial
automation, electric current, temperature, engine perfor­
mance, etc.
• Very low prices enable them to be embedded in such devices in
which, until recent time it was not worthwhile to embed any­
thing. Thanks to that, the world is overwhelmed today with
cheap automatic devices and various “smart” appliances.
• Prior knowledge is hardly needed for programming. It is suffi­
cient to have a PC (software in use is not demanding at all and is
easy to learn) and a simple device (called the programmer) used
for “loading” ready-to-use programs into the microcontroller.
• Even though there is a large number of different types of mi­
crocontrollers and even more programs created for their use
only, all of them have many things in common. Hence if one is
handled, then all microcontroller devices can be handled.

Inside a Microcontroller
In a microcontroller, all the operations within the microcontroller are
performed at high speed and quite simply, but the microcontroller itself
would not be so useful if there are no special circuits which make it
complete. In continuation, we are going to call your attention to them.

Read Only Memory (ROM)

• Read only memory (ROM) is a type of memory used to per­

manently save the program being executed. The size of the
program that can be written depends on the size of this memory.
 WIRELESS UNDERGROUND SENSOR SYSTEM 313

• ROM can be built in the microcontroller or added as an ex­

ternal chip, which depends on the type of the microcontroller.
Both options have disadvantages.
• If ROM is added as an external chip, the microcontroller is
cheaper and the program can be considerably longer. At the
same time, a number of available pins are reduced as the mi­
crocontroller uses its own input/output ports for connection to
the chip.
• The internal ROM is usually smaller and more expensive, but
leaves more pins available for connecting to Perisoil
Moistural environment. The size of ROM ranges from 512 B
to 64 KB.

Random Access Memory (RAM)

• Random access memory (RAM) is a type of memory used for

temporary storing data and intermediate results created and
used during the operation of the microcontrollers.
• The content of this memory is cleared once the power supply
is off. For example, if the program performs an addition, it is
necessary to have a register standing for what in everyday life
is called the “sum.”
• For that purpose, one of the registers in RAM is called the
“sum” and used for storing results of addition. The size of RAM
goes up to a few KBs.

Electrically Erasable Programmable ROM (EEPROM)

• The EEPROM is a special type of memory not contained in all

microcontrollers.
• Its contents may be changed during program execution (similar
to RAM), but remains permanently saved even after the loss of
power (similar to ROM).
• It is often used to store values, created and used during op­
eration (such as calibration values, codes, values to count up to
etc.), which must be saved after turning the power supply off.
• A disadvantage of this memory is that the process of pro­
gramming is relatively slow. It is measured in milliseconds.
314 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

Memory Organization of Microcontroller

Special Function Register (SFR)

Special function registers are part of RAM memory. Their purpose is
predefined by the manufacturer and cannot be changed therefore.
Since their bits are soil moisture connected to particular circuits
within the microcontroller, such as A/D converter, serial commu­
nication module etc., any change of their state directly affects the
operation of the microcontroller or some of the circuits.
For example, writing zero or one to the SFR controlling, an input/
output port causes the appropriate port pin to be configured as input
or output. In other words, each bit of this register controls the
function of one single pin.
Program Counter: Program Counter is an engine running the
program and points to the memory address containing the next in­
struction to execute. After each instruction execution, the value of the
counter is incremented by 1.
However, the value of the program counter can be changed at any
moment, which causes a “jump” to a new memory location. This is
how subroutines and branch instructions are executed.
After jumping, the counter resumes even and monotonous auto­
matic counting +1, +1, +1 and so on.

Central Processing Unit (CPU) As its name suggests, this is a unit

which monitors and controls all processes within the microcontroller
and the user cannot affect its work.
It consists of several smaller subunits, of which the most im­
portant are:

Instruction decoder is a part of the electronics which re­

cognizes program instructions and runs other circuits on
the basis of that. The abilities of this circuit are expressed in
the “instruction set” which is different for each micro­
controller family.

Arithmetical logical unit (ALU) performs all mathematical and lo­

gical operations upon data.
 WIRELESS UNDERGROUND SENSOR SYSTEM 315

Accumulator is an SFR closely related to the operation of ALU. It

is a kind of working desk used for storing all data upon which some
operations should be executed (addition, shift, etc.).
It also stores the results ready for use in further processing. One of
the SFRs, called the Status Register, is closely related to the accu­
mulator, showing at any given time the “status” of a number stored in
the accumulator (the number is greater or less than zero etc.). A bit is
just a word invented to confuse novices at electronics. Joking aside,
this word in practice indicates whether the voltage is present on a
conductor or not. If it is present, the appropriate pin is set to logic
one (1), i.e., the bit’s value is 1.
Otherwise, if the voltage is 0, the appropriate pin is cleared (0), i.e.
the bit’s value is 0. It is more complicated in theory where a bit is
referred to as a binary digit, but even in this case, its value can be
either 0 or 1.

Data Transmission through Soil The concept of WUSN was in­

troduced in wireless communication lab at the Georgia Institute of
Technology in 2006. WUSNs have been investigated in many con­
texts recently, but research reports of WUSNs in agricultural appli­
cation are few. I put forward the concept of WUSN for the first time
and analyzed path loss of the underground wireless channel under the
different soil composition. Besides, the key issues that should be
considered were put forward in the WUSN communication archi­
tecture and WUSN design who studied the channel model of elec­
tromagnetic wave transmission in soil, analyzed path loss, bit error
rate, maximum transmission distance, water content test error of
electromagnetic wave multipath transmission, etc. under the main
influence factors such as the soil composition, soil volumetric water
content and node burial depth, node distance and sensor frequency
studied the effect of antenna bandwidth of wireless sensor network
node at 433 MHz frequency.
The buried depth of nodes in the soil 15 cm and 35 cm, and water
content of the soil 9.5% and 37.3% on communication between
aboveground node and underground node developed the near surface
WUSN system used for golf course which included acquisition
nodes, sink nodes, relay nodes, and a gateway node. The gateway
node controlled data storage and transmission of the sink node could
316 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

connect with computer or GPRS module through the RS232 in­

terface. Like sink node, gateway node could be remotely controlled
through DDI researched wireless signal attenuation of Zigbee
wireless transceiver module by using soil column.
In the wireless underground sensor network AG-UG and UG-
AG communication, sink node is deployed on the ground surface
that is perpendicular to underground node, underground node
burial depth changes every 10 cm within 10 cm to 100 cm, soil
water content in the range of 5% to 30%, a total of six levels, the
influence characteristics of node burial depth, and soil water
content on received signal strength are measured in AG-UG and
UG-AG communication. AG-UG and UG-AG communication
test model.
In the actual WUSN communication, is not only communication
that exists between the sink node and underground node, but also
communication between underground nodes is essential, and is par­
ticularly important. In UG-UG communication, transmitting node
and the receiving node burial depth are fixed 40 cm, soil water content
changes from 5% to 30%, separated in six levels, the horizontal inter-
nodes distance changes from 100 cm to 1000 cm, at 10 levels, and the
influence characteristics of horizontal inter-nodes distance and soil
water content on received signal strength are measured in UG-UG
communication.

Visual Basic Visual Basic is engineered for productively building

type-safe and object-oriented applications. Visual Basic enables
developers to target Windows, web, and mobile devices. As with
all languages targeting the Microsoft.NET Framework, programs
written in Visual Basic benefit from security and language
interoperability.
Visual Basic (VB) is a programming environment from Microsoft
in which a programmer uses a graSoil Moistureical user interface
(GUI) to choose and modify preselected sections of code. VB also
includes advanced features – concepts and structures which allow
programs to be adapted for use with the Internet. Visual Basic will
not run on operating systems other than Windows and on machines
with non-Intel compatible processors.
 WIRELESS UNDERGROUND SENSOR SYSTEM 317

Internet of Things (IoT) The Internet of things (IoT) is the inter-

networking of physical devices, vehicle, buildings, and other
items—embedded with electronics, software, sensors, actuators, and
network connectivity that enable these objects to collect and ex­
change data. In 2013, the Global Standards Initiative on Internet of
Things (IoT-GSI) defined the IoT as “the infrastructure of the
information society.” The IoT allows objects to be sensed or con­
trolled remotely across existing network infrastructure, creating
opportunities for more direct integration of the physical world into
computer-based systems, and resulting in improved efficiency,
accuracy, and economic benefit in addition to reduced human in­
tervention. When IoT is augmented with sensors and actuators, the
technology becomes an instance of the more general class of cyber-
physical systems, which also encompasses technologies such as
smart grids, smart homes, intelligent transportation and smart ci­
ties. Each thing is uniquely identifiable through its embedded
computing system but is able to interoperate within the existing
Internet infrastructure. Typically, IoT is expected to offer advanced
connectivity of devices, systems, and services that goes beyond
machine-to- machine (M2M) communications and covers a variety
of protocols, domains, and applications. The interconnection of
these embedded devices (including smart objects) is expected to user
in automation in nearly all fields, while also enabling advanced
applications like a smart grid, and expanding to areas such as smart
cities. “Things,” in the IoT sense, can refer to a wide variety of
devices, such as heart monitoring implants, biochip transponders on
farm animals, electric clams in coastal waters, automobiles with
built-in sensors, DNA analysis devices for environmental/food/
pathogen monitoring, or field operation devices that assist fire­
fighters in search and rescue operations. Legal scholars suggest to
look at “Things” as an “inextricable mixture of hardware, software,
data, and service.” These devices collect useful data. Current market
examples include home automation (also known as smart home
devices) such as the control and automation of lighting, heating
(like smart thermostat), ventilation, air conditioning (HVAC)
systems, and appliances such as washer/dryers, robotic vacuums, air
purifiers, ovens or refrigerators/freezers that use Wi-Fi for remote
monitoring.
318 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

System Architecture
Architecture Description

We developed the magnetic induction (MI) waveguide technique

for the wireless communications in WUSNs. The MI waveguide
consists of a series of relay coils between two underground
transceivers. The communications are achieved by magnetic in­
duction between two adjacent coils. These relay coils do not
consume extra energy and the cost is neglectable. The MI wave­
guide can solve the two major problems of the EM technique in
soil medium:
1. By using MI waveguides, the feasible communication range
between two transceivers can achieve nearly 100 meters.
2. The MI channel conditions remain constant, since the soil medium
causes little variation in the attenuation rate of magnetic fields.

Wireless Underground Sensor Design Architecture

As shown in Figure 13.5, the data from any soil can be collected
through WUSN and the analysis of the soil can be used to predict the
crops and the time for watering. A sprinkler system can also be
combined in this proposed scheme to automate watering of crops
based on the water content of the soil.

Figure 13.5 Proposed Architecture

 WIRELESS UNDERGROUND SENSOR SYSTEM 319

Components of WUSN

Cloud Cloud computing is the on-demand availability of computer

system resources, especially data storage (cloud storage) and computing
power, without direct active management by the user. The term is
generally used to describe data centers available to many users over the
Internet.

Arduino Arduino is an open-source platform used for building

electronics projects as shown in Figure 13.6. Arduino consists of both
a physical programmable circuit board (often referred to as a mi­
crocontroller) and a piece of software, or IDE (integrated develop­
ment environment) that runs on your computer, used to write and
upload computer code to the physical board.
The Arduino platform has become quite popular with people just
starting out with electronics, and for good reason. Unlike most
previous programmable circuit boards, the Arduino does not need a
separate piece of hardware (called a programmer) in order to load
new code onto the board – you can simply use a USB cable.
Additionally, the Arduino IDE uses a simplified version of C++,
making it easier to learn to program. Finally, Arduino provides a
standard form factor that breaks out the functions of the micro-
controller into a more accessible package.

Figure 13.6 Arduino

320 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

pH Sensor A pH sensor is one of the most essential tools that’s ty­

pically used for water measurements. This type of sensor is able to
measure the amount of alkalinity and acidity in water and other
solutions. pH sensors and transmitters are used in many industries
such as chemicals, water and wastewater, food and beverage, phar­
maceuticals, power plants, primaries and oil and gas. The selection of
the sensor depends on the application. Choose the glass sensor’s
diaphragm and reference system according to your needs.

Soil Moisture Sensor A soil moisture sensor measures the quantity of

water contained in a material, such as soil on a volumetric or
gravimetric basis. To obtain an accurate measurement, a soil tem­
perature sensor is also required for calibration. The soil moisture
sensor uses capacitance to measure the water content of soil by
measuring the dielectric permittivity of the soil, which is a function
of the water content. Simply insert this rugged sensor into the soil to
be tested, and the volumetric water content of the soil.

Transmitter and Receiver The WUSN transmitter is an electronic

device which produces radio waves with an antenna. Transmitter
itself generates a radio frequency alternating current, which is applied
to the antenna. The transmitter is buried in the soil and gets en­
ergized when the data are passed. The receiver is the destination of
the message. The receiver’s task is to interpret the sender’s message,
both verbal and nonverbal, with as little distortion as possible. The
process of interpreting the message is known as decoding.

Classification Based on Soil As shown in Table 13.1, the moisture

content of the soil and its retention capabilities can confirm the type
of soil. This needs no further analysis and the crops can be decided
on the basis of soil type.
Table 13.2 gives the names of crops that can be grown in farms
based on the soil type.

Table 13.1 Soil Type

TYPES OF SOIL HIGH WATER RANGE NORMAL BELOW THE LEVEL
Fine (clay) 80–100 60–80 below 60
Medium (loamy) 88–100 70–88 below 70
Coarse (sandy) 90–100 80–90 below 80
 WIRELESS UNDERGROUND SENSOR SYSTEM 321

Table 13.2 Crops for Soil Types

SOIL TYPE CROP
Clay soil Beans, cauliflower, cabbage
Loamy soil Oil seed, sugarcane, wheat
Sandy soil Maize, millets, wheat

Table 13.3 Result Obtained

DAYS WATER LEVEL
1 49
2 48
3 49
4 40
5 37
6 35
7 30
8 20
9 15
10 39

Result The proposed architecture was implemented and the soil

moisture was continuously monitored for 10 days. Based on the re­
sults obtained, as shown in Table 13.3, the analysis was done.
Based on the results obtained in Table 13.3, it is confirmed that
the soil under consideration is clay type and the crops as shown in
Table 13.2 for clay soil can be grown to get a better yield. The time
taken for transmitting and receiving of the information from the
sensors is a few seconds, as shown in Figure 13.7.

Conclusion

To conclude, the successful deployment of an IoT-enabled large-

scale urban-area WUSN has been implemented. It covers the wide
area, which is a typical urban environment with various buildings and
human activities. All the sensor nodes are buried underground
without interfering with the aboveground activities. Despite the ex­
treme environment, it is deployed in underground, urban environ­
ment, and extremely cold winters.
322 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

LogID DATA Logdate LogTime

1 Moisture:_PH: 07/13/2020 23:06:58

2 Moisture:_PH: 07/13/2020 23:07:02
3 Moisture:499_PH:0 07/13/2020 23:07:12
4 Moisture:499_PH:0 07/13/2020 23:07:22
5 Moisture:498_PH:0 07/13/2020 23:07:32
6 Moisture:498_PH:0 07/13/2020 23:07:42

7 Moisture:498_PH:0 07/13/2020 23:07:52

8 Moisture:496_PH:0 07/13/2020 23:08:02

9 Moisture:496_PH:0 07/13/2020 23:08:12
10 Moisture:496_PH:0 07/13/2020 23:08:22
11 Moisture:496_PH:102 07/13/2020 23:08:32
12 Moisture:496_PH:97 07/13/2020 23:08:42
13 Moisture:496_PH:111 07/13/2020 23:08:52
14 Moisture:496_PH:122 07/13/2020 23:09:02
15 Moisture:496_PH:131 07/13/2020 23:09:12
16 Moisture:496_PH:140 07/13/2020 23:09:22
17 Moisture:496_PH:145 07/13/2020 23:09:32
18 Moisture:496_PH:151 07/13/2020 23:09:42
19 Moisture:496_PH:155 07/13/2020 23:09:52
20 Moisture:496_PH:160 07/13/2020 23:10:02
21 Moisture:496_PH:163 07/13/2020 23:10:12
22 Moisture:496_PH:166 07/13/2020 23:10:22
23 Moisture:496_PH:167 07/13/2020 23:10:32
24 Moisture:496_PH:0 07/13/2020 23:10:42
25 Moisture:496_PH:0 07/13/2020 23:10:52
Figure 13.7 Time Analysis
 WIRELESS UNDERGROUND SENSOR SYSTEM 323

WUSN has proven its value in real-time soil property monitoring.

More importantly, we also implemented a cloud-based open plat­
form, so the sensing results are open to public access. In this way,
proposed system provides long-term continuous monitoring of the
soil properties.

References
[1] E. Berman, G. Calinescu, C. Shah, and A. Zelikovsky., “Power ef­
ficient monitoring management in sensor networks,” in Proceedings of
IEEE Wireless Communication and Networking Conference (WCNC04),
Atlanta. USA, 2004.
[2] L. Li and X. M. Wen, “Energy efficient optimization of clustering
algorithm in wireless sensor network,” J. Electron. Inform. Technol.,
vol. 30, no. 4, pp. 966–969, 2008.
[3] Y. H. Cai, G. Liu, L. Li, and H. Liu, “Design and test of nodes for
farmland data acquisition based on wireless sensor network,” Chinese
Soc. Agric. Eng., vol. 25, no. 4, pp. 176–178, 2009.
[4] M. C. Vuran and Ian F. Akyildiz, “Channel model and analysis for
wireless underground sensor networks in soil medium,”in Physical
Communication, vol. 3, no. 4, pp. 245–254, December 2010.
[5] M. C. Vuran and I. F. Akyildiz, “Channel model and analysis for
wireless underground sensor networks in soil medium,” Physical
Communication Journal, vol. 3, p. 254, 2010.
[6] Yogendra S. Dohare, Tanmoy Maity, P. S. Das, and P. S. Paul,
“Wireless communication and environment monitoring in under­
ground coal mines – Review,” IETE Tech. Rev., vol. 32, no. 2,
pp. 140–150, 2015.
[7] Serhan Yarkan, Sabih G¨uzelg¨oz, H¨useyin Arslan, and Robin R.
Murphy, “Underground mine communications: A survey,” IEEE
Communication Survey & Tutorials, vol. 11, no. 3, Third quarter 2009.
[8] A. M. Donoghue. “Occupational health hazards in mining: An
overview.” Occup. Med. (Lond), vol. 54, no. 5, pp. 283–289, 2004, doi:
10.1093/occmed/kqh072, PMID: 15289583.
[9] R. Agnelo, C. V. Silva, and C. Mehmet, “Communication with above
devices in wireless underground sensor networks: An empirical
study,” in Communications (ICC). IEEE International Conference
Proceedings, IEEE Commun. Soc., Cape Town, pp. 23–27, 2010.
[10] O. Green, E. S. Nadimi, V. Blanes, R. N. Jorgensen, and C. G.
Sorensen CG, “Monitoring and modeling temperature variations
inside silage stacks using novel wireless sensor networks,” Comput.
Electron Agric., vol. 69, no. 1, pp. 149–157, 2009.
[11] X. Q. Yu, Z. L. Zhang, and W. T. Han, “Evaluation of commu­
nication in wireless underground sensor networks,” in IOP Conference
324 ALEARNING ALGORITHMS FOR INTELLIGENT IOT

Series: Earth and Environmental Science, vol. 69, 3rd International

Conference on Advances in Energy, Environment and Chemical
Engineering, Chengdu, China, 2017.
[12] X. Dong, M. C. Vuran, and S. Irmak, “Autonomous precision
agriculture through integration of wireless underground sensor net­
works with center pivot irrigation systems,” Ad Hoc Networks Journal,
vol. 11, pp. 1975–1987, 2013.
[13] X. Zhang, A. Andreyev, C. Zumpf, M. C. Negri, S. Guha, and M.
Ghosh, “Thoreau: A subterranean wireless sensing network for
agriculture and the environment,” in IEEE Conference on Computer
Communications Workshops (INFOCOM WKSHPS), pp. 78–84, 2017.
[14] Coen J. Ritsema, Henk Kuipers, and Leon Kleiboer, “A new wireless
underground network system for continuous monitoring of soil water
contents,” Water Resources Research, vol. 45, 2010, doi: 10.1029/2008W
[15] I. F. Akyildiz and E. P. Stuntebeck, “Wireless underground sensor
networks: Research challenges,” Ad Hoc Networks Journal, vol. 4,
pp. 669–686, 2006.
[16] M. C. Vuran and A. R. Silva, “Communication through Soil in
Wireless Underground Sensor Networks Theory and Practice, Sensor
Networks: where Theory Meets Practice. Springer, 2009.
[17] A. R. Silva and M. C. Vuran, “Development of a Testbed for
Wireless Underground Sensor Networks,” EURASIP Journal on
Wireless Communications and Networking, vol. 2010, p. 620307, 2010.
[18] X. Dong, M. C. Vuran, and S. Irmak, “Autonomous precision
agriculture through integration of wireless underground sensor net­
works with center pivot irrigation systems,” Ad Hoc Networks Journal,
vol. 11, pp. 1975–1987, 2013.
[19] A. Markham, N. Trigoni, D. W. Macdonald, and S. A. Ellwood,
“Underground localization in 3-D using magneto-inductive tracking,”
IEEE Sensors Journal, vol. 12, pp. 1809–1816, 2012.
[20] X. Dong and M. C. Vuran, “Spatio-temporal soil moisture mea­
surement with wireless underground sensor networks,” in Ad Hoc
Networking Workshop, 2010. Med-Hoc-Net 2010. 9th IFIP Annual
Mediterranean, Juan-les-pins, France, 2010.
[21] C. J. Ritsema, H. Kuipers, L. Kleiboer, E. Elsen, K. Oostindie, J.G.
Wesseling, J. Wolthuis, and P. Havinga, “A new wireless under­
ground network system for continuous monitoring of soil water
contents,” Water Resources Research Journal, vol. 45, pp. 1–9, 2009.
[22] I. F. Akyildiz, Z. Sun, and M. C. Vuran, “Signal propagation tech­
niques for wireless underground communication networks,” Physical
Communication Journal (Elsevier), vol. 2, pp. 167–183, 2009.
[23] X. Dong and M. C. Vuran, “A channel model for wireless under­
ground sensor networks using lateral waves,” in IEEE Globecom 2011,
Houston, TX, USA, 2011.
 14
PREDICTING EFFECTIVENESS OF
SOLAR POND HEAT EXCHANGER
WITH LTES CONTAINING CUO
NANOPARTICLE USING MACHINE
LEARNING
K. KARUNAMURTHY1, G. SUGANYA1,
M. ANANTHI2, AND T. SUBHA2
1
Associate Professor, Vellore Institute
of Technology Chennai, Tamil Nadu,
India
2
Sri Sairam Engineering College,
Chennai, Tamil Nadu, India

Contents

Introduction 326
Solar Pond 327
Low-Temperature Energy Storage System (LTES) for
Solar Pond 328
Background Study 329
Experimental Setup 329
Experimentation 331
Performance Parameters of Solar Pond 332
Modeling Using Machine Learning 332
Data Collection and Preprocessing 334
Data Cleaning 334
Normalization 335
Outlier Analysis 335
Correlation Analysis 336
Multi-Linear Regression 338
Splitting Dataset into Training and Testing Data 340
Fitting and Prediction 340

DOI: 10.1201/9781003119838-14 325

326 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Results and Discussion 341

Accuracy 341
Goodness of Fit 341
Future Works 342
References 343

Introduction

Solar energy is clean, free, and available abundantly [1]. It is the

primary duty of humanity to be socially responsible to adopt the
sustainable development goals (SDGs) of the United Nations, and
this is possible if mankind utilizes this renewable source of energy
for the progress of society. The solar radiation striking our planet is
174 PW (174 × 1012 kW) at its upper atmosphere and the average
amount of solar energy absorbed by the landmasses, oceans, and
atmosphere is approximately 121.8 × 1012 kW. The rate at which
solar energy is received outside the earth’s atmosphere perpendi­
cular to the solar beam at the earth’s mean distance from the sun is
1.373 kW/m2.
Solar energy is utilized by two methods, such as solar photo­
voltaic (SPV) and solar thermal technology (STT) [2]; the appli­
cations of STT has a broader range than SPV. STT is used to
collect and store solar energy in the form of (i) low-temperature
heat (< 100°C) or (ii) high-temperature heat (up to 500°C). The
various solar collectors used are flat plate collectors (FPC), con­
centrated solar thermal collectors (CSCs), and evacuated tube
collectors (ETCs). FPC and ETCs are used for low-temperature
applications like domestic hot water supply, space heating of the
residential, commercial and institutional building, and for industrial
process heat. CSCs provide comparatively high-grade thermal en­
ergy, which can even operate heat engines. A solar pond is a type of
collecting and storing large amounts of low-temperature heat with
less expense in a salt gradient pool of water. Solar ponds provide
thermal energy for longer periods in the temperature range of 70°C
to 100°C.
 PREDICTING EFFECTIVENESS OF HEAT EXCHANGER 327

Solar Pond

Solar ponds (SPs) are water bodies with varying salinity gradient
called halocline, i.e., the salinity of the pond increases with depth.
These solar ponds possess three different distinct zones viz., upper
convective zone (UCZ), non-convective zone (NCZ), and lower
convective zone (LCZ). These zones have different densities due to
the presence of salinity, and they store the thermal energy by ar­
resting the convective effects. The solar radiation reaching the
bottom of the pond heats the water of the LCZ as the density of the
LCZ is higher, the bulk motion of water to the top surface of the
pond is prevented, thereby the thermal energy is trapped within the
LCZ itself. This collected and stored thermal energy at the LCZ is
used for various engineering applications. Thus solar ponds are one
of the most effective ways of capturing and storing solar energy [3].
The applications of solar ponds include industrial process heating,
desalination, greenhouse heating, crop drying, electric power gen­
eration, refrigeration, and air conditioning. A typical schematic
diagram of a solar pond is represented in Figure 14.1.
The stored energy of the solar pond is used for the above-said
various applications by incorporating internal or external heat ex­
changers. The main concern with this low-temperature energy sto­
rage system (LTES) is its poor thermal efficiency of 15% to 25% and
the period of operation of the solar pond. The thermal efficiency of a
solar pond is defined as the ratio of total energy extracted from the
solar pond with the help of the heat exchanger to the total solar
radiation falling on the surface of the solar pond. There are different

Figure 14.1 Typical Sketch of a Solar Pond

328 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

works reported by researchers in the past literature. The effect of

insulation thickness of the side wall on the SP efficiency was studied
[4]. The current research work carried out is to increase the per­
formance of the solar pond by providing twisted tapes in the flow
passage of the heat exchanger, and by placing encapsulated PCM
blended with CuO nanoparticles in the LCZ as an LTES system to
increase the period of operation.

Low-Temperature Energy Storage System (LTES) for Solar Pond

There is a great potential in various energy-related applications for

phase change material (PCM)-based thermal energy storage (TES)
[5]. The availability of solar energy is limited to the day, and hence
the solar pond can operate only during the day time. Thus LTES is
mandatory to operate the solar pond in the absence of solar radiation
and to increase the period of operation in a day. LTES tries to match
the demand and supply of the solar energy. The two ways by which
thermal energy can be stored are sensible heat storage and latent heat
storage. However, latent heat energy storage is preferred as the en­
ergy density of storage is more. In this research, the LTES is coupled
with the LCZ of the solar pond; this is achieved by introducing
encapsulating PCM in copper capsules. There are wide varieties of
PCM available; however, in this experimentation, paraffin was se­
lected as the PCM, as it is readily available in the market and also the
physical and chemical properties of paraffin do not change even after
thousands of charging and discharging cycle. The three processes in
the thermal cycle of LTES are (i) charging, (ii) storing, and (iii)
discharging. Figure 14.2 illustrates these three processes.
Low High
temperature LTES LTES LTES temperature
medium medium

High Low
temperature temperature
medium medium

Charging Storage Discharging

Figure 14.2 Processes Involved in LTES

 PREDICTING EFFECTIVENESS OF HEAT EXCHANGER 329

Background Study

Usage of PCM to increase thermal conductivity is discussed by

many researchers. Shu-Rong et al. proposed the method of
blending multi- walled Carbon nanotubes in different proportions
with paraffin to increase the thermal conductivity [6]. In this re­
search, CuO nanoparticles are added with to the paraffin. The in­
crease in the rate of convective heat transfer by using swirl tapes is
analyzed through a numerical study for 3-D heat transfer and flow
behaviors in circular tubes with loose-fit multiple channel [7].
Karunamurthy et al. discussed the effect of Al2O3 nanoparticles to
increase conductivity [8].
Machine learning (ML) techniques are profoundly used to re­
duce a number of experiments conducted physically thereby redu­
cing budget, manpower, and natural resources. Use of linear
regression for predicting output values where the values are con­
tinuous with greater correlation between predictor and response
variables is discussed by researchers. Since the amount of data
collected is relatively low, researchers suggest the use of linear re­
gression for prediction [9,10]. Christopher et al. compared the
advantages of using linear regression with neural networks and vice
versa [11]. Other applications of the method are discussed by
Mahesh Gadhavi et al. [12]. A relative study of various kinds of
applications using regression strengthens the selection of model for
the proposed study.

Experimental Setup

The experimental setup fabricated is as per Huseyin Kurt et al.

(2006) and it is presented in Figures 14.3 and 14.4.
Laboratory model solar pond of three numbers of dimension 50
cm × 60 cm × 60 cm is fabricated using 0.15-cm galvanized iron
sheet. A heat exchanger made up of a copper tube of 1.2 cm outer
diameter and 1.05 cm inner diameter is placed in the LCZ of the
solar ponds with five passes. The three different solar ponds used for
experimentation are (i) ordinary SP, (ii) SP with LTES (PCM
blended CuO nanoparticle), and (iii) SP with LTES (PCM blended
CuO nanoparticle) and swirl tape. Experiments are conducted and
330 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 14.3 Solar Pond Experimental Setup

Figure 14.4 Schematic Sketch of Heat Exchanger in LCZ

swirl tape of optimum twist ratio of 6.36 is placed in the flow passage
to augment the rate of heat transfer. Swirl tapes provide secondary
flow and turbulence, which enhances the rate of heat transfer.
The PCM paraffin (n-Tricosane C23H48) of shell make is chosen
for thermal energy storage and the same is encapsulated and placed in
the LCZ of the solar pond as LTES for experimentation. However,
the thermal conductivity of paraffin PCM is 0.214 W/mK (solid) and
0.15 W/mK (liquid). This is poor and to increase the thermal con­
ductivity, CuO nanoparticles are added.
There is a limit for the percentage of CuO nanoparticles that can
be blended to the PCM because the addition of too many CuO
nanoparticles with PCM will also lead to agglomeration which is not
 PREDICTING EFFECTIVENESS OF HEAT EXCHANGER 331

Figure 14.5 (a) Swirl Tapes and (b) Encapsulated PCM

preferred. In this aspect, experiments are conducted to identify the

optimum proportion of CuO nanoparticles that can be added to the
PCM. Encapsulated PCM blended with CuO nanoparticles and
swirl tapes used are represented in Figure 14.5(a) and (b).

Experimentation

Experiments were conducted on all the three solar ponds under solar
irradiated condition. Heat transfer fluid used is water, which is al­
lowed to flow through the flow passage of the heat exchanger under
laminar and turbulent flow conditions with Reynolds’s number of
1804 and 9020, respectively. The objective of the study is to have
higher outlet temperature of the heat exchanger, indicating better
performance. The various parameters related to the outlet water
temperature are ambient temperature, water inlet temperature, solar
irradiation, flow rate, and LCZ temperature.
The temperatures such as ambient, inlet, outlet, LCZ, NCZ, and
UCZ were measured using “K”-type thermocouple connected to
digital temperature indicator. The solar radiation reaching the solar
pond is measured using pyranometer, and the salinity of the LCZ,
NCZ, and UCZ is measured daily using pyranometer to ensure the
existence of halocline. Sampling vents were provided to get the
sample from LCZ, NCZ, and UCZ. The readings are recorded every
30 minutes in a day from 8:00 AM to 6:00 PM over a period of 41
days during May–June 2019. The average values are tabulated for
analyzing the solar pond performance parameters, such as outlet
water temperature, temperature rise, rate of heat transfer, efficiency
of solar pond, and effectiveness of heat exchanger.
332 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Performance Parameters of Solar Pond

i. Rate of heat transfer

Q = m cp(To Ti), (14.1)

where
Q – rate of heat transfer (W)
Q’ – maximum heat transfer (W)
m – mass flow rate (kg/s)
Cp – specific heat capacity (J/kg °C)
To – outlet temperature (°C)
Ti– inlet temperature (°C)
TLCZ – temperature of the LCZ (°C)
ii. Maximum rate of heat transfer (Q’) Q’ = mCp(TLCZ–Ti)
iii. Effectiveness of heat exchanger
Q
= (14.2)
Q’

iv. Efficiency of solar pond

Useful heat carried away by flow medium
=
Radiation incident Surface area of solar pond

Modeling Using Machine Learning

Intelligence, the need for the era, is expected not only from human,
but also from machines. ML is the field of study that gives computers
the competence to be intelligent without being explicitly programmed.
Applications of ML range from day-to-day application like automatic
switching on/off of household motors to more complex applications
like controlling satellite operations.
In course of this, ML finds an integral part in Industry 4.0 trans­
forming applications in mechanical engineering domain to be more
intelligent using techniques like predictive maintenance, demand
prediction, quality control, etc. Generally, numerous experiments
would be carried out on solar pond and the input parameters supplied
and output obtained will be stored as a relational dataset during
 PREDICTING EFFECTIVENESS OF HEAT EXCHANGER 333

experimentation. This collected dataset will be analyzed to identify

input values that can end up in optimal performance. But, this process
needs exhaustive repetition of experiments and will lead to waste of
materials, money, time and human resources. Also, the feasibility of
witnessing the optimality of the solution cannot be granted.
ML techniques can be a boon in such cases where the machine can
learn the pattern that exists in the dataset resulted from limited ex­
perimentation. The use of ML techniques for predicting perfor­
mance parameters is discussed by researchers [9]. Once patterns are
understood through proper modeling, any number of predictions is
possible without physical experimentation. ML algorithms witness to
reduce wastage of resources, budget, and time. Figure 14.6 represents
the process of modeling solar pond performance prediction using
multi-linear regression.

Figure 14.6 ML Modeling

334 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Data Collection and Preprocessing

The collected data are stored in a Microsoft excel document where

the columns represent the features and rows represent data for every
experimentation. Around 820 recordings are created over a period of
41 days. Ambient temperature, water inlet temperature, solar radia­
tion flux, and LCZ temperature are considered to be input para­
meters and outlet water temperature is recorded as output parameter.
LCZ and water temperatures are recorded for all three ponds.
Choice of right features and reliable and complete data are most
important for accurate prediction of output. Measuring instruments
such as Pyranometer, thermocouples and Pyranometer used for re­
cording the values are made of electronic components and hence are
prone to errors. Also, data are taken at different time periods and the
possible environmental changes may also affect recording of data.
The format and range of data also vary from one feature to another
feature and hence visualization also becomes difficult. This step
prepares the collected data to be suitable and accurate for the inputs
of the models selected for modeling. The entire process is pro­
grammed using Jupyter notebook and the data are stored using
Microsoft Excel in a comma separated values (CSV) format.

Data Cleaning

During data collection, there is a possibility for some values to be

missed. This may happen due to non-availability of data or some data
validation rule or malfunctioning of some device or due to human
error. But regardless of the reason, the dataset should be made
complete by processing these missing values. Several approaches are
used by researchers, namely, removal of missing data, filling up with
some static value, filling up with some statistical data, etc.
The following steps are performed for cleaning the data:
• Experimentation values (rows) in which three or more values
are missing are removed. Using this approach, a total of 18 rows
were removed.
• For each feature, when fewer values are missing, mean is cal­
culated and is used to fill the missing values. Figure 14.7
 PREDICTING EFFECTIVENESS OF HEAT EXCHANGER 335

Figure 14.7 Missing Values (before and after Filling Up).

represents the status of missing values in the original dataset

and after filling up with mean.

Normalization

Data for each attribute are collected in different ranges and units. ML
modeling algorithms work on numeric data and hence variables are
measured at different scales. Temperatures are measured in degree
Celsius and Incident solar radiation flux is measured in Watt/m2. The
range of values corresponding to various features does not contribute in
the identical way to the machine model fitting and might end up
creating a bias to the overall system. Also, visualization becomes very
difficult with varying ranges of data. Hence, normalization is applied
to convert ranges of all data between 0 and 1. Min-Max normalization,
as in Equation 14.3, is used in our modeling.
x min(x )
xnorm = (14.3)
max(x ) min(x )

Outlier Analysis

During data collection, there might be situations in which the elec­

tronic components may malfunction resulting in random data. Though
recorded by human being, such randomness may go unnoticed leading
to improper training. Hence outliers, the observations that lie at an
unusual distance from other values in the collected data are to be
336 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 14.8 Box Plot for Outlier Analysis

Figure 14.9 Outlier Analysis for “Solar Radiation Flux”

identified and removed. Box plots are used to perform this analysis
where the data that lie outside of 1.5 * interquartile range on both the
ends are considered as outliers and are removed. Figure 14.8 represents
the various ranges identified using box plot and Figure 14.9 represents
outlier analysis for inlet solar radiation flux.

Correlation Analysis

Since the values of output attributes is continuous that is spread over

a range, regression modeling is required. Linear regression (LR) is
the most simple and commonly used technique for regression [9].
Logically, LR models the relationship between the input and output
variables by fitting a straight line with the following assumptions:
1. Recordings are independent among them.
2. Relationship between input and output remains linear.
3. Very little or no correlation among the input attributes.
 PREDICTING EFFECTIVENESS OF HEAT EXCHANGER 337

Figure 14.10 Scatter Plots for Correlation Analysis

Assumption 1 is satisfied since the experimentation is done at

different timings and the relationship is not considered. Assumptions
2 and 3 are observed using correlation analysis, a method that is used
to calculate the strength of relationship between different features.
Scatter plots are used to visualize the correlation between features
graphically. The correlation between independent features to output
feature is visualized using scatter plot and is presented in Figure 14.10.
To analyze the numeric distribution of data, heat map is a better
graphical representation tool in which data values are represented as
colors. Variations from white to black indicate the distribution of
range of data. Correlation values can be visually understood using heat
map. Figure 14.11 presents a heat map representing the correlation

Figure 14.11 Heat Map Representing Correlation Values

338 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

between the features. It is evident from the figure that all the outlet
temperatures are positively correlated with maximum correlation with
all input variables and hence linear regression is chosen for modeling
the dataset.

Multi-Linear Regression

Predictive modeling is very useful for forecasting future conclusions

and assessing metrics that are impractical or difficult causing wastages
in physical measure. When the output data are continuously varying
in a finite range, regression models are the best choice. Many re­
gressive models are proposed by research community that varies in
understanding differing patterns in data. Linear regression is used for
predicting the values of output variables in situations where training
data are less [10].
Linear regression is a basic supervised regressive method that ar­
rives at a straight line with constant slope. It’s used to predict values
within a continuous range, (e.g., sales, price) rather than trying to
classify them into categories. Since, the number of independent
variables is more than one, multi-linear regression is chosen for this
task. Regression is an eager learning method that abstracts the data
during the training itself. Regression method returns a set of straight
line to a graph such that the line passes through the maximum
number of data points of the collected dataset. This line fitted is
called the hypothesis and is used to predict the dependent value for
any given set of independent values (x1, x2, …, xn ).
The output of multi-linear regression model is represented by
Equation 14.4.
ypred = m1 x1 + m2 x2 + …+mn xn + m 0 + c (14.4)

where, m1, m2, …, mn indicate the impact of each attribute over the
output.
x1, x2, …, xn indicate the values of independent variables.
m 0 refers to the common bias that exists in the dataset.
c represents the error of the model.
ypred refers to the predicted output.
 PREDICTING EFFECTIVENESS OF HEAT EXCHANGER 339

The coefficients m1, m2, …, mn measure the effect of each in­

dependent variable in association with the effects of all the other
independent variables in the model. Thus, these coefficients
measure the marginal effects of the independent variables. The
difference between the actual and predicted output is termed as
the error or deviation. The average of all errors in the dataset is
termed as mean error of the dataset caused by the model. This
error is to be reduced for the model to be more accurate and be
reliable for the given dataset. The least squares method offers a
way of choosing the coefficients properly by minimizing the sum
of the squared errors. Least squares plot is used to check the
distribution of data with respect to an approximate line fitting.
Figure 14.12 depicts the least squares plot for various input
parameters to output parameter.
The accuracy of the model is evaluated using various metrics
like mean absolute error (MAE), mean square error (MSE), or
root mean square method (RMSE). We have used RMSE as the
standard error analysis metric. Equation 14.5 defines the RMSE
calculation.
n
i =1 (x i x i ) 2
Root Mean Square Error = (14.5)
n

where xi represents the actual value of input attribute.

x i indicates the predicted value of dependent variable(s).

Figure 14.12 Least Squares Line Plot

340 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Splitting Dataset into Training and Testing Data

The dataset is divided into two random sets of data, namely, training
and testing using 80 to 20 proportion. The first set comprising 80%
of the data termed as training dataset is used to fit the model. The
training set covers the set of inputs with known output that the
model uses to learn to generalize other data. The second set com­
prising 20% of the data, termed as testing dataset is used to check
whether the fitting is accurate and acceptable. The package train-
test-split in the scikit-learn library of Python is used to divide the
dataset. This package takes care of random splitting of rows to avoid
bias that may be caused due to environmental conditions.
The dataset is divided into input and output features termed as “X”
and “Y,” respectively. Using the train-test-split function of scikit-
learn library, X and Y are divided into four subsets, namely, X_train,
X_test, Y_train, and Y_test. The dimensions of these are (642,4),
(642,), (160,4), and (160,). Three such datasets representing three
different solar ponds were taken for modeling.

Fitting and Prediction

Finding the best estimates of the coefficients by applying a specific

model is called “fitting” the model. This is performed using the training
dataset and the process is termed as learning or training the model.
“linear_model” of scikit-learn library is used to train the training data
using multi-linear regression. Once learnt or trained with the patterns
of the data, the model returns a set of four coefficients.
Coefficients: [0.46025584 1.04712856 0.00551125 0.07416891]

Intercept: 14.436684821577494

The equation can then be written as in Equation 14.6.

Outlet water temperature = 0.460255841x1 + 1.04712856x2

+ 0.00551125x3 0.07416891x4
14.436684821577494 (14.6)
 PREDICTING EFFECTIVENESS OF HEAT EXCHANGER 341

Figure 14.13 Sample Analysis of Actual vs. Predicted Value

The outlet water temperature for the test data is then predicted using
the “predict” function of the scikit-learn library. The predicted value
of testing data along with original output value in the dataset is
compared. Figure 14.13 depicts a sample comparison.

Results and Discussion

Accuracy

The accuracy is used as an important measure to find out the correctness

of prediction [11]. It is defined to be the ratio between numbers of
correct predictions to total number of predictions. Accuracy of the
prediction is tested visually using a 3-D graph that enables visualization
with more than two features. Figure 14.14 depicts original values of
outlet water temperature and predicted values of predictor variables,
outlet water temperature with ambient and inlet water temperature.
MAE, MSE, and RMSE are calculated and are given below. The
values are found to be less than the assumed threshold and are satisfactory.

Goodness of Fit

To understand the goodness of fit with respect to variations in data,

R2 value, coefficient of determination is calculated. This value reflects
342 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

Figure 14.14 Actual vs. Predicted Values

the proportion of variation in the independent variables as defined by

the regression model. If the predictions are close to the actual values,
the values of R2 will be close to 1. On the other hand, if the predic­
tions are unrelated to the actual values, then R2 will be close to 0. The
value obtained for the proposed model is 0.9928512303131685 and
hence is close to 1.

Future Works

Solar ponds are still in amateur stage and research is to be carried out
for augmenting its performance and to commercialize. The materials
for thermal energy storage at the LCZ, non-corrosive (materials)
layers for the side wall, and the bottom of the pond have lot of scope.
This is a low-temperature energy storage method and needs opti­
mization of process parameters in order to commercialize for gen­
erating electric power. There are many process parameters involved
 PREDICTING EFFECTIVENESS OF HEAT EXCHANGER 343

with the operation of solar pond. The above work uses only few
parameters and handled dataset taken only for a certain period of
time. Also, the collinearity between the predictor variables is not
properly realized by our model.
Future works may concentrate on handling large volume of data
with proper analyzing on relationship between the independent
variables. Also, metrics such as accuracy derived from ML modeling
will not be useful for adding business value to the stated problem by
researchers. The optimal values of input parameters can be obtained
from applying optimization techniques.

References
[1] The Solar Vision: A 100-day Agenda of Solar Energy Industries
Association. https://www.seia.org/initiatives/about-solar-energy.
[2] https://energypedia.info/wiki/Portal:Solar.
[3] Abhishek Kumar, and Ranjan Das, “Effect of peripheral heat con­
duction in salt-gradient solar ponds,” Journal of Energy Storage,
November 2020. 10.1016/j.est.2020.102084.
[4] Assad H. Sayer, Hazim Al-Hussainia, and Alasdair N. Campbell,
“New theoretical modelling of heat transfer in solar ponds,” Solar
Energy, vol. 125, pp. 207–218, 2016. 10.1016/j.solener.2015.12.015.
[5] Binjian Nie, Anabel Palacios, Boyang Zou, Jiaxu Liu, Tongtong
Zhang, and Yunren Li, “Review on phase change materials for cold
thermal energy storage applications,” Renewable and Sustainable
Energy Reviews, vol. 134, 2020. 10.1016/j.rser.2020.110340.
[6] Shu-Rong Yan, Rasool Kalbasi, Aliakbar Karimipour, and Masoud
Afrand, “Improving the thermal conductivity of paraffin by in­
corporating MWCNTs nanoparticles,” Journal of Thermal Analysis
and Calorimetry, 2020. 10.1007/s10973-020-09819-0.
[7] Suttisak Kunlabud, Varesa Chuwattanakul, Vichan Kongkaitpaiboon,
Pitak Promthaisong, and Smith Eiamsa-ard, “Heat transfer in tur­
bulent tube flow inserted with loose-fit multi-channel twisted tapes as
swirl generators,” Theoretical and Applied Mechanics Letters, vol. 7, no.
6, pp. 372–378, 2017. 10.1016/j.taml.2017.11.011.
[8] Karunamurthy, K., Rajesh, M. R., Vijaypal, B., and Kumar, A.,
“Thermal conductivity and charging & discharging characteristics of a
thermal energy storage system blended with Al2O3 nanoparticles,”
Nano Hybrids and Composites, vol. 17, pp. 10–17. 2017. 10.4028/
www.scientific.net/nhc.17.10.
[9] D. C. Montgomery, E. A. Peck, and G. G. Vining, Introduction to
linear regression analysis, John Wiley & Sons, 2015.
344 APPLIED LEARNING ALGORITHMS FOR INTELLIGENT IOT

[10] K. Karunamurthy, R. Manimaran, and M. Chandrasekar, “Prediction

of solar pond performance parameters using artificial neural network,”
International Journal of Computer Aided Engineering and Technology,
vol. 11, no. 2, pp. 141–150, 2019. 10.1016/j.taml.2017.11.011
[11] Christopher Osita Anyaeche, and Desmond Eseoghene Ighravwe,
“Predicting performance measures using linear regression and neural
network: A comparison,” African Journal of Engineering Research, vol.
1, no. 3, pp. 84–89, 2013.
[12] Mahesh Gadhavi, and Chirag Patel, “Student Final Grade Prediction
Basedon Linear Regression,” Indian Journal of Computer Science and
Engineering, vol. 8, no. 3, pp. 274–279, 2017.
Index

Note: Italicized page numbers refer to figures, bold page numbers refer to
tables

accuracy, 160–161 model, 112

activation functions, 6–8 parameter column, 114
Adam optimizer, 21 Python language in, 112
adaptive non-contiguous allocation Alexa, 176
algorithm, 218–219 Alzaylaee, Mohammed K., 81
advanced metering infrastructure AMHARA algorithm, 90
(AMI), 138 Android OS, 80–81
adware, 79 anomaly detection, 45
agglomerative clustering, 44 ApateDNS, 93
Agrawal, Rakshit, 87 Arduino, 319
Ahmad, Bani, 218 ARI-LSTM, 87
air quality prediction, 111–112 arithmetical logical unit (ALU), 314
air quality analysis, parameters, artificial intelligence (AI), 2
and its range, 119 defined, 37
AQI category, pollutants, subsets of, 3
breakpoints, and health Ashton, Kevin, 124
impacts, 120 asset management software, 136
data analysis stage and data assets tracking, 135, 135
graph, 117 association, 45
data pre-processing and Attended Recent Inputs (ARI), 87
labeling, 116 average pooling, 16
data split into testing and training
dataset, 118 backdoors, 79
index category and range, 115 back-propagation neural networks
libraries, 113 (BPNN), 89

345
346 INDEX

backward propagation, 20–22. computer vision, 258

See also convolutional neural concentrated solar thermal collectors
networks (CNNs) (CSCs), 326
Bag of Words, 85 confusion matrix, 160–161
Bayes theorem, 162–163 connected vehicles, 147–150
Bernoulli Naive Bayes 3G/4G data connection in, 148
classification, 162 advantages of, 148–149
best-fit approach, 216–217, 219, prediction in the US, 149
222–223 vehicle to cloud technology, 148
Bhandari, Saurav, 178 vehicle to everything
bias, 9–10 technology, 148
Big Data, 88 vehicle to infrastructure
binary classification, 23–24 technology, 148
Bintext, 93 vehicle to pedestrian
BlackRock, 136 technology, 148
vehicle to vehicle technology, 148
call record data analysis, 206–207 conveyance store optimization, 205
carbon monoxide, 63 CONVNet, 3
CFF Explorer, 92 convolution, 252–253
Chebyshev distance, 275, 276 convolution layer, 13–16
CH-index, 277 convolution operation, 14–15
class discrimination measure feature map, 14
(CDM), 91 filter, 13–14
classification, 38–39 convolutional neural networks
document, 40–41 (CNNs), 1–32
Gaussian Naive Bayes, 163 activation functions of, 6–8
Naive Bayes, 39–40, 161 application of, 22–31
network traffic, 208–209 image classification, 22–24
record, 201–205 neural style transfer, 30–31, 32
support vector machine object detection, 25–29
(SVM), 40 transfer learning, 29–30
client segmentation, 205 backward propagation, 20–22
cloud computing, 220, 319 bias, 9–10
cloud storage, 219 convolution layer, 13–16
clustering, 43–44, 45 convolution operation, 14–15
agglomerative, 44 feature map, 14
algorithms, 269, 270 filter, 13–14
exclusive, 44 datasets and errors, 9
k-means. See k-means clustering deeper hidden layers, 22
overlapping, 44 definition of, 4
probabilistic, 44 early stopping, 22
component traversal, 81 Fast R-CNN, 27
 INDEX 347

Faster R-CNN, 28 Deep4MalDroid, 81

forward propagation, 18–19 deep-belief network (DBN), 82–83
hyper parameters, 13 deeper hidden layers, 22
image data augmentation, 21 Dependency Walker, 92
layers, 4–5 Deviation From Poisson (DFP), 91
loss function, 8–9 DIE, 92
optimizers, 21 digital profiling, 206
overfitting, 10 digital twins, 136–138
padding, 11 dimensionality reduction, 46
parameters, 13 diStorm, 93
pooling layer, 16 document classification, 40–41
properties of, 5–6 dream group statistics, 205–206
region-based, 26 DroidDeepLearner, 82
stride, 12, 12 DroidDelver, 82–83
underfitting, 10 DroidSMS, 80–81
variance, 9–10 drone frame, 251
You Only Look Once, 29 drug monitoring, 68–69
correlation analysis, 336–338, 337 dynamic analysis, 88
Coursera, 207
Covid-19 growth pattern eager learners, 40
analysis, 108 early stopping, 22
Cowan, Jeremy, 124 elbow curve, 277
Creeper Worm, 78 electrically erasable programmable
criminal activity detection, 206 ROM (EEPROM), 313
elimination of sparse features
data analytics (ESF)., 92
for IoT, 73–75 Elk Cloner, 78
and machine learning, 37 employee turnover prediction,
data cleaning, 334–335 153–173
data storage system (DSS), 220 attrition rate, 154
datasets, 162 datasets, 155
datasets and errors, 9 HR department’s role,
DB-Index, 278 154–155
DDOS attack, 284 supervisor’s role, 154
decision tree (DT), 110 logistic regression, 157–173
decryption, 250 logit function, 158
dedicated short-range machine learning algorithms
communication (DSRC) for, 155
radios, 147–150 overview, 154
Deep Flow, 82 system implementation, 156–157
deep learning (DL), 127–128 data preprocessing, 157
deep neural network (DNN), 85, 179 workflow, 156
348 INDEX

encryption, 250 Google Pixels, 81

EPCA procedure, 197 Google STT Python library, 182
error function, 8–9 gradient descent approach, 104
Euclidian distance, 266–287
evacuated tube collectors HashCalc, 92
(ETCs), 326 health monitoring, 41
exclusive clustering, 44 hearing aids, 70–71
executable linkable files (ELF), 91 hidden Markov model, 43
home area network (HAN), 139
Facebook, 179 Hou, S., 81, 82
fake news identification, 207 HSV (hue, saturation, lightness
Fakushima, K., 3 value), 255, 256
false negatives, 160 Hwang, Jinsoo, 86
false positives, 160 hybrid analysis, 88
Fast R-CNN, 27. HYDRA, 88
See also convolutional neural hyper parameters, 13
networks (CNNs)
Faster R-CNN, 28. IDA pro, 93
See also convolutional neural image data augmentation, 21
networks (CNNs) IMCFN model, 88
FileAlyzer, 93 indoor air quality monitoring
fitness trackers, 70 system, 63
flat plate collectors (FPC), 326 Industrial Internet of Things (IIoT),
fluffy multi-target molecule swarm 128–137
streamlining system advantages of, 130
(FMOPSO), 196 applications of, 129
F-measure, 155 future of, 132
forward propagation, 18–19. history of, 129–130
See also convolutional neural Industry 4.0, 130–131
networks (CNNs) market of devices, 132
fraud detection, 206 use cases, 132–137
fuzzy C mean, 281–282 Industry 4.0, 130–131
information technology alerts, 207
Galton, Francis, 103–106 instruction decoder, 314
gap statistics, 277 internal fragmentation, 215–236
Gaussian mixture models, 43 best-fit approach, 216–217, 219,
Gaussian Naive Bayes 222–223
classification, 163 data placement, 221
genetic algorithm and K-means and data storage system, 220
(GAKM), 195 defined, 216
Google Assistant, 176 and dynamic memory allocation,
Google Colab, 53 217–218
 INDEX 349

lasso regression, 217, 220, industrial, 128–137

229–230 in industry, 60–65
linear regression, 217, 229 indoor air quality monitoring
memory bank approach, 216–217, system, 63
225–226 inventory management system,
algorithms, 226–230 65, 66
flow chart, 223–225, 224 machine diagnosis and
graphical representation, prognosis system, 61–62
231–233, 232, 233 noise monitoring system, 64
pseudo code, 230–231 remote monitoring and
software packages, 233 production control, 62–63
memory resource management, livestock management using,
221–222 144–147
optimized approach to deal with and machine learning, 127–128
residual space, 224 number of devices, 131
and processor allocation, 218–219 power consumption of
and system compatibility, 218 devices, 133
Internet of Things (IoT), scope of, 125–126
123–151, 317 sensors, 58
advantages of, 126 smart farming, 140–144
applications of, 125 smart metering, 138–140
companies, 132 types of, 126
connected vehicles, 147–150 use cases
connectivity, 59 assets tracking, 135, 135
data analytics for, 73–75 digital twins, 136–138
data processing, 59–60 smart manufacturing, 134, 134
defined, 58 user interface, 70
features of, 126 workflow, 59
Global Standards Initiative, 317 inventory management system,
in healthcare, 65–73 65, 66
drug monitoring, 68–69
fitness trackers, 70 JavaSnoop, 93
hearing aids, 70–71 Jinsoo Hwang, 86
medical alert system, 72–73 Jupyter, 231
pharmacy management, Jupyter Notebook, 53
67–68, 69
k mediods, 282
for physicians, 72
Karbab, E.B., 82
smart pills, 69–70
Keras, 249
wearables, 70, 71
Keras Embedding, 85
wearables to monitor diseases,
keyloggers, 78–79
71, 72
history of, 124–125
350 INDEX

k-means clustering, 193–212, in software defined network,

270–271 282–287
background studies, 194–198 DDOS attack monitoring, 284
datasets, 272–273 latency, 284
distance formula of centroid in, traffic monitoring, 284
273, 273–274 as solution for business problems
flow chart, 271–272, 272 archive arranging and
future work, 214 gathering, 210
versus fuzzy C mean, 281–282 client maintenance, 210
versus k mediods, 282 conveyance enhancement, 209
in linear independent component purchaser division, 209
analysis, 280–281 rebate examination, 210
operational working of, 199 in spherical structure, 279, 279
versus other unsupervised steps in, 199
algorithms, 210–211, 211 synchronic health fitness detector
overview, 198 of combatants in armed
properties of, 278–281 conflicts using, 199–201
pros and cons, 211–212 Army control room’s
real-world use cases of, 201–209 Segment, 201
call record data analysis, Chief of Armed Force’s
206–207 segment, 201
client segmentation, 205 combatant’s segment, 201
conveyance store using Chebyshev Distance.,
optimization, 205 275, 276
criminal activity detection, 206 using Manhattan distance,
digital profiling, 206 274–275, 276
dream group statistics, using Minkowski distance,
205–206 275, 276
fake news identification, 207 using squared Euclidean distance
fraud detection, 206 formula, 274
information technology value of k, 276–277
alerts, 207 and vector quantization, 279, 280
network traffic classification,
208–209 Kruskal–Wallis test, 91
recognition of offense
lasso regression, 217, 220, 229–230
localities, 205
leaky ReLU, 7–8
record classification, 201–205
LeCunn, Yann, 3
record examination, 208–209
likelihoods, 161
rideshare data analysis, 206
linear regression, 103–106, 217, 229
showcasing and business
in health-related data analysis,
sales, 208
105–106
spam channel, 207–208
 INDEX 351

significance of, 103–106 choosing model, 48

livestock management using data preparation, 48
cattle health monitoring, 144–145 evaluation, 49
fine-tune feeding, 146 gathering input data, 48
milking maximization, 146 parameter tuning, 49
reproductive cycle and calving prediction, 49
monitoring, 145–146 training, 48–49
Llody’s algorithm, 266 regression algorithms, 99–121
local connectivity, 6 types of, 37–38
Logistic Model Tree (LMT), 86–87 supervised learning, 37–42
logistic regression, 42, 157–173 unsupervised learning, 42–46
binary, 158 MacQueen, James, 193, 270
multinomial, 158 MalDozer, 82
ordinal, 158 MalDy, 90
performance of, 159–161 MalFCS, 88
logit function, 158 malware, 79–80
long short-term memory (LSTM), detection, 77–94
84, 87, 219 Android-based, 80–83
loss function, 8–9 ransomware, 83–88
lower convective zone (LCZ), 327 tools, 92–94
low-temperature energy storage Windows-based, 88–91
system (LTES), 328, 328 history, 78
Linux-based, 91–92
machine diagnosis and prognosis mobile, 78
system, 61–62 statistics, 78
machine learning (ML), 33–54 types of, 78–79
algorithms, 36–37 zero-day, 80
applications of, 36, 49–52, Manhattan distance, 274–275, 276
127–128, 128 Maniath, Smith, 84
agriculture, 50 Matplotlib, 53
automobile industry, 52 Max pooling, 16, 253
banking, 50 McLaughlin, Niall, 81
health care, 51–52 MD5 hashing, 250–251, 251
categories, 38 MD5deep, 93
and data analytics, 37 mean absolute error (MAE), 339
in data analytics for IoT, 74–75 mean square error (MSE), 104, 339
and deep learning, 127–128 medical alert system, 72–73
defined, 34, 37 mediods, 282
implementations, 52–54 memory bank, 216–217
and IoT, 127–128 memory bank algorithm, 226–229
overview, 34–36 memory management based on
process of, 46–49
352 INDEX

thread behavior Naive Bayes classification,

(MMBT), 221 39–40, 161
memory resource management, applications of, 163
221–222 benefits of using, 163
MEMSIC hubs, 196 Bernoulli, 162
microcontroller Gaussian, 163
analog features, 306 multinomial, 162
central processing unit, 314 voting classifier, 163–164
CMOS, 307 hard voting, 163
data memory organization, 308 soft voting, 165
design, 305 natural language processing
electrically erasable programmable (NLP), 175
ROM, 313 neocognition, 3
features of, 311–313 Netconverse, 86
I/O ports, 308 network traffic classification,
memory organization, 308, 308, 208–209
314–315 neural style transfer, 30–31, 32
perisoil moistural feature, Ng, Andrew, 109
305–306 N-gram sequencing, 87
pin diagram, 307–308 noise monitoring system, 64
PORTA register, 309 non-convective zone (NCZ), 327
PORTB register, 309–310 normalization, 335
PORTC register, 310–311 Novak, Kralj, 178
PORTD register, 311 null deviance, 160
PORTE register, 311 NumPy, 234
program counter, 314
random access memory, 313 object detection, 25–29
read only memory, 312–313 occupational exposure limits
special features, 306–307 (OELs), 295
special function register, 314–315 odd ratio, 92
TRISA register, 309 Opcode, 88
TRISB register, 309–310 Open Platform Communication
TRISC register, 310–311 (OPC) protocol, 129
TRISD register, 311 OpenCV library, 254
TRISE register, 311 optimizers, 21
Minkowski distance, 275, 276 outlier analysis, 335–336, 336
Mitchell, Tom, 35 overfitting, 10
Morley, Dick, 129 overlapping clustering, 44
multi-linear regression, 338–340
padding, 11
multinomial Naive Bayes
page table entry (PTE), 225
classification, 162
Pandas, 53, 234
 INDEX 353

parametric ReLU (PReLU), 7–8 record examination, 208–209

particle swarm optimization rectified linear unit (ReLU)
(PSO), 196 activation function:, 6–7
particulate matter, 63 region-based convolutional neural
patient monitoring, 42 networks (R-CNNs), 26
PEiD, 92 regression, 39
PEview, 93 logistic, 42
PewDiePie (YouTube gamer), 179 regression algorithms, 99–121
pharmacy management, 67–68, 69 in air quality prediction, 111–112
physicians, 72 decision tree regression, 110
polynomial regression, 106–108 polynomial regression, 106–108
advantages and disadvantages, random forest regression, 111
107–108 simple linear regression, 103–106
in Covid-19 growth pattern studies on, 100–103
analysis, 108 support vector machine
graph, 107 regression, 108–110
pooling, 253 RegShot, 93
pooling layer, 6 rEMnux, 92, 93
precision, 161 remote monitoring and production
principal component analysis (PCA), control, 62–63
195, 269 residual deviance, 160
probabilistic clustering, 44 Resource Hacker, 92
Process Explorer, 93 RFID (radio frequency
Procman, 93 identification), 135
programmable logic controller RGB color space, 255, 256
(PLC), 129–130 rideshare data analysis, 206
Purplesec, 84 root kit, 79
root mean squared error (RMSE),
radiofrequency identification 220, 339
(RFID), 65, 66 RStudio, 234
random access memory (RAM), 313
random forest, 111 scikit-learn, 53–54, 340
ranking technique, 195 Scilab, 230–231, 234
ransomware, 79–80, 83–88 SDN controller placement, 265–287
read only memory (ROM), 312–313 DDOS attack monitoring, 284
recall (sensitivity), 161 k-means clustering in, 282–287
receiver, 320 latency, 284
receiver operating characteristic network partitioning, 285–286
(ROC) curve, 161 traffic monitoring, 284
receptive fields, 5 unsupervised learning in, 267–270
recognition of offense localities, 205 sensors
record classification, 201–205 humidity, 302
354 INDEX

interfacing, 302 maintenance, 143

IoT, 58 mobility, 143
pH, 320 steps in, 141–142
soil moisture, 320 smart machines, 2
WUSNs, 303–304 smart manufacturing, 134, 134
sentiment analysis, 162, 175–190 smart metering, 138–140
architecture, 183, 186 smart pills, 69–70
background of, 180–181 social media, 46, 179
future of, 189–190 software defined network (SDN)
implementation of, 183, 184 DDOS attack monitoring in, 284
logic representation k-means clustering in, 282–287
algorithm, 185 latency, 284
negative training set, 187 network partitioning, 285–286
overview, 175–178 traffic monitoring, 284
positive training set, 187 solar collectors, 326
speech-to-text, 176–178, solar energy, 326
182–183, 188–189 solar photovoltaic (SPV), 326
studies on, 178–180 solar ponds, 325–343
Seong II Bae, 85 accuracy, 341
Shaikh, Mostafa, 179 correlation analysis, 336–338, 337
shape/space invariant artificial neural data cleaning, 334–335
network (SIAAN). datasets, 340
See convolutional neural defined, 327
networks (CNNs) fitting and predictions,
shared weights, 5 340–341, 342
sigmoid or logistic activation future works, 342–343
function:, 6 goodness of fit, 341–342
silhouette curve, 277 low-temperature energy storage
simple linear regression, 103–106 system for, 328, 328
in health-related data analysis, machine learning modeling,
105–106 332–333, 333
significance of, 105 multi-linear regression, 338–339
Siri, 176 normalization, 335
Sklearn, 234 outlier analysis, 335–336, 336
smart farming, 140–144 zones, 327
advantages of, 143 solar thermal technology (STT), 326
data analytics in, 142–143 spam channel, 207–208
data collection in, 142 special function register (SFR),
hardware, 142 314–315
informed action and speech recognition, 176
planning, 142 speech-to-text (STT) translation,
infrastructure for, 143–144 176–178, 182–183
 INDEX 355

spy drone, IoT-automated, 241–263 tanh or hyperbolic tangent activation

architectural diagram, 245–246, 246 function, 6
existing system, 244 TapSnake, 81
image classification VGG16 Term Frequency–Inverse Document
model, 248–249, 249 Frequency (TFIDF), 87
color-based segmentation, 257 test data, 9
computer vision, 258 text encoding, 85
datasets, 253 text-to-speech (TTS)
image preprocessing, 254 translation, 176
packages, 253 thread-oriented memory
segmentation function, management layer
257–258 (TOMML), 221
target image input, 254 Tinker, 251
visualization in HSV, 255, 256 traffic monitoring, 284
visualization in RGB, training data, 9
254–255, 256 transfer learning, 29–30
implementation methodologies, transmitter, 320
252–253 Trojans, 78
machine learning model, 246–247 true negatives, 160
overview, 242–243 true positives, 160
proposed system, 244 Twitter, 178
training images, 258–261, 259
VGG16 architecture, 247–248 underfitting, 10
drone frame, 248 unsupervised learning, 42–46,
image classification, 248 267–270
remote control interface, 248 anomaly detection, 45
spyware, 79 applications of, 45–46
squared Euclidean distance association, 45, 269
formula, 274 clustering, 43–44, 45, 269
stacking, 88–89 dimensionality reduction, 46
static analysis, 88 dimensionality reduction
status register, 315 algorithms, 269
storage system vulnerability factor methodology, 269
(SSVF), 220 social media, 46
strace, 91 types of, 42–46
stride, 12, 12 visualization, 45–46
sum pooling, 16 upper convective zone (UCZ), 327
supervised learning, 38
Valgrind, 93
classification, 38–39
validation data, 9
regression, 39
variance, 9–10
support vector machine (SVM),
vector quantization, 279, 280
108–110
356 INDEX

vehicle to cloud (V2C) features, 302

technology, 148 humidity sensor, 302
vehicle to everything (V2X) interfacing sensor, 302
technology, 148 and Internet of Things, 317
vehicle to infrastructure (V2I) literature on, 294–295
technology, 148 microcontroller, 304–313
vehicle to pedestrian (V2P) modules, 302
technology, 148 and occupational health hazard in
vehicle to vehicle (V2V) technology, 148 mining, 298
vigenere cipher, 249–250 overview, 292–293
Vinayakumar, R., 85 product description, 302
viruses, 79 sensors, 303–304
Visual Basic, 316 soil moisture measurement, 297
visualization, 45–46 channel model and analysis,
volatile organic compounds, 63 299–300
voting classifier, 163–164 sensors, 303–304
hard voting, 163 spatio-temporal, 299
soft voting, 165 system analysis, 300–302
VX-Heavens, 91 existing system, 300
hardware requirements, 301
wearables, 70, 71 proposed system, 300–301
to monitor diseases, 71, 72 software requirements, 301
weather prediction, 41 system requirements, 301
weighted voting, 88–89 system architecture, 318–321
wide area network (WAN), 139 versus terrestrial WSNs,
wireless underground sensor 294–295, 299
networks (WUSNs), test bed development for,
291–323 296–297
and agriculture field underground mine
monitoring, 298 communications, 297–298
applications, 303 and Visual Basic, 316
components of, 319–321 worms, 79
Arduino, 319
cloud, 319 Yaqoob, Ibrar, 87
pH sensor, 320 Yoo, Suyeon, 90
receiver, 320 YouTube, 179
soil moisture sensor, 320 Yuan, Zhenlong, 82
transmitter, 320
in crop production, 295–296 Zeek, 93
data transmission through soil, zero-day malware, 80
315–316 Zhang, Hanqi, 87
design architecture, 318 Zigbee, 295