10.1007@978 981 15 2774 6 PDF

Lecture Notes in Networks and Systems 109
Mihir Narayan Mohanty

Swagatam Das Editors
Advances in
Intelligent
Computing and
Communication
Proceedings of ICAC 2019
Lecture Notes in Networks and Systems
Volume 109
Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland
Advisory Editors
Fernando Gomide, Department of Computer Engineering and Automation—DCA,
School of Electrical and Computer Engineering—FEEC, University of Campinas—
UNICAMP, São Paulo, Brazil
Okyay Kaynak, Department of Electrical and Electronic Engineering,
Bogazici University, Istanbul, Turkey
Derong Liu, Department of Electrical and Computer Engineering, University
of Illinois at Chicago, Chicago, USA; Institute of Automation, Chinese Academy
of Sciences, Beijing, China
Witold Pedrycz, Department of Electrical and Computer Engineering,
University of Alberta, Alberta, Canada; Systems Research Institute,
Polish Academy of Sciences, Warsaw, Poland
Marios M. Polycarpou, Department of Electrical and Computer Engineering,
KIOS Research Center for Intelligent Systems and Networks, University of Cyprus,
Nicosia, Cyprus
Imre J. Rudas, Óbuda University, Budapest, Hungary
Jun Wang, Department of Computer Science, City University of Hong Kong,
Kowloon, Hong Kong
The series “Lecture Notes in Networks and Systems” publishes the latest
developments in Networks and Systems—quickly, informally and with high quality.
Original research reported in proceedings and post-proceedings represents the core
of LNNS.
Volumes published in LNNS embrace all aspects and subfields of, as well as new
challenges in, Networks and Systems.
The series contains proceedings and edited volumes in systems and networks,
spanning the areas of Cyber-Physical Systems, Autonomous Systems, Sensor
Networks, Control Systems, Energy Systems, Automotive Systems, Biological
Systems, Vehicular Networking and Connected Vehicles, Aerospace Systems,
Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power Systems,
Robotics, Social Systems, Economic Systems and other. Of particular value to both
the contributors and the readership are the short publication timeframe and the
world-wide distribution and exposure which enable both a wide and rapid
dissemination of research output.
The series covers the theory, applications, and perspectives on the state of the art
and future developments relevant to systems and networks, decision making, control,
complex processes and related areas, as embedded in the fields of interdisciplinary
and applied sciences, engineering, computer science, physics, economics, social, and
life sciences, as well as the paradigms and methodologies behind them.
** Indexing: The books of this series are submitted to ISI Proceedings,
SCOPUS, Google Scholar and Springerlink **
More information about this series at http://www.springer.com/series/15179

Mihir Narayan Mohanty •
Swagatam Das
Editors
Advances in Intelligent
Computing
and Communication
Proceedings of ICAC 2019
123
Editors
Mihir Narayan Mohanty Swagatam Das
Department of Electronics and Electronics and Communication
Communication Engineering, Institute Sciences Unit
of Technical Education and Research Indian Statistical Institute
Siksha ‘O’ Anusandhan Deemed Kolkata, West Bengal, India
to be University
Bhubaneswar, Odisha, India
ISSN 2367-3370 ISSN 2367-3389 (electronic)

Lecture Notes in Networks and Systems
ISBN 978-981-15-2773-9 ISBN 978-981-15-2774-6 (eBook)
https://doi.org/10.1007/978-981-15-2774-6
© Springer Nature Singapore Pte Ltd. 2020
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface
This issue of Lecture Notes in Networks and Systems is dedicated to the

International Conference on Intelligent Computing and Advances in
Communication (ICAC) held at the campus of Institute of Technical Education and
Research (Faculty of Engineering and Technology), Siksha ‘O’ Anusandhan
Deemed to be University, Bhubaneswar, from November 15 to 17, 2019. This
conference was organized by the Department Electronics and Communication
Engineering of Institute of Technical Education and Research (Faculty of
Engineering and Technology). This conference had three tracks, namely advances
in communication systems, intelligent systems, and signal processing. From 165
papers received from three countries, 63 were selected for inclusion into the con-
ference proceedings. Each paper was peer-reviewed by at least two reviewers.
The objective of the conference is to bring together experts from academic
institutions, industries, research organizations, and professional engineers for
sharing of knowledge, expertise, and experience in emerging trends related to the
computer, communication, and electrical topics. The aim of this international
conference is to cover all the issues on a single platform and provide international
forum for researchers to discuss the real-time problems and solutions to exchange
their valuable ideas and showcase the ongoing works which may lead to path-
breaking foundation of the futuristic engineering. This conference mainly aims at
advanced communication protocol, database security and privacy, advanced com-
puting system, energy saving, etc., on several updated techniques. The conference
offers a platform to focus on the inventive information and computing toward the
investigation of cognitive mechanisms and processes of human information pro-
cessing, and the development of the next-generation engineering and advanced
technological systems.
This conference was held in the Bansuri Guru Auditorium, Institute of Technical
Education and Research (Faculty of Engineering and Technology), Siksha ‘O’
Anusandhan Deemed to be University, Bhubaneswar, Odisha, India. Siksha ‘O’
Anusandhan is a Deemed University located in Bhubaneswar, Odisha. It was
originally founded as Institute of Technical Education and Research (ITER) in the
year 1996. The university has been at the forefront nourishing a learning ambience,
v
vi Preface
encouraging academic, research, and innovations since its inception in 2007. The
university is composed of nine degree-granting schools with 10,000 students. Many
of SOAU’s programs are nationally accredited for meeting high standards of aca-
demic quality, including engineering, medicine, pharmacy, business, nursing,
biotechnology, humanities, environment, nanotechnology, agriculture, and law. The
university was ranked 16th by National Institutional Ranking Framework (NIRF)
under the aegis of Ministry of Human Resource Development, Government of
India, and has been awarded ‘A’ grade by NAAC. It has established ten research
centers and 23 research laboratories to fulfill the need of faculties and students.
Faculty of Engineering and Technology is a constituent of SOAU having 13
departments with more than 400 faculty members. The Department of ECE is
continually working to provide quality research outputs in the areas of signal and
image processing, communication engineering, and microelectronics devices.
The editors thank the authors for extending their fullest cooperation in preparing
the manuscripts to the Springer Lecture Notes guidelines, taking aboard the addi-
tional review comments.
The editors would also like to convey their heartfelt thanks to Prof. M. R. Nayak,
President, Siksha ‘O’ Anusandhan Deemed to be University; Prof. Amit Bannerjee,
Vice-Chancellor, Siksha ‘O’ Anusandhan Deemed to be University; Prof. M. R.
Mallick, Director, ITER (FET), Siksha ‘O’ Anusandhan Deemed to be University;
Prof. P. K. Nanda, Dean (R&D), Siksha ‘O’ Anusandhan Deemed to be University;
and Prof. P. K. Sahoo, Dean, ITER (FET), Siksha ‘O’ Anusandhan Deemed to be
University, for their constant inspiration and motivation in all stages of the
conference.
Sincerely,
Bhubaneswar, India Mihir Narayan Mohanty

Kolkata, India Swagatam Das
Guest Editors
Contents
Children’s Age and Gender Recognition from Raw Speech Waveform

Using DNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Mousmita Sarma, Kandarpa Kumar Sarma, and Nagendra Kumar Goel
Inter-vehicular Communication in Urban Traffic Scenario . . . . . . . . . . 10
Suman Malik and Prasant Kumar Sahu
Distributed Estimation of IIR System’s Parameters in Sensor Network
by Multihop Diffusion LMS Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 20
Meera Dash, Trilochan Panigrahi, and Renu Sharma
Implementation of Raspberry Pi for Fault Detection
in Optic Fibre Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
K. P. Swain, S. R. Das, Sangram Kishore Mohanty, and G. Palai
Distributed Optical Fiber Sensing System Performance Improvement
Using Signal Processing Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Ramji Tangudu and Prasant Kumar Sahu
A Brief Understanding of IOT Health Care Service Model Over
Remotely Cloud Connected Environment . . . . . . . . . . . . . . . . . . . . . . . . 46
Subhasish Mohapatra and Smita Parija
Fusion-Based Multi-biometrics Authentication System with Intrusion
Detection and Prevention Anamoly . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Mehreen Kirmani and Dinesh Kumar Garg
Design of Dual Notch Band Antenna with Rejected Filter
with Defected Slot for WLAN and Satellite Applications . . . . . . . . . . . . 63
Sanjay Kumar Sahu, Bhukya Arun Kumar, and Kaibalya Kumar Sethi
Weighted Filtered-s LMS Algorithm for Nonlinear Active Noise
Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Tapasmini Sahoo and Kunal Kumar Das
vii
viii Contents
Comparison of Classifiers for Speech Emotion Recognition (SER)

with Discriminative Spectral Features . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Hemanta Kumar Palo, Debasis Behera, and Bikash Chandra Rout
Handwritten Odia Numerals Recognition: A Supervised Learning
Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
Suchismita Behera and Niva Das
An Implementation of Neural Network Approach for Recognition
of Handwritten Odia Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Sachikanta Dash and Rajendra Kumar Das
Recognition and Extraction of Rain Drops in an Image
Using Hough Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
A. Susmitha, Lipsa Dash, and Sunanda Alamuru
A Novel Technique for Target Recognition Using Multiresolution
Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Tamanna Sahoo and Bibhuprasad Mohanty
A Novel Technique of Shadow Detection Using Color Invariant
Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
Leeza Panda and Bibhuprasad Mohanty
Comparison of Performance Metrics of Star Topology and Ring
Topology in Wireless Sensor Network . . . . . . . . . . . . . . . . . . . . . . . . . . 122
Laxmipriya Moharana, Bedanta Kumar Biswal, Raman Raj,
and Suraj Naik
A Framework for Real-Time Lane Detection Using Spatial Modelling
of Road Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Pankaj Prusty and Bibhuprasad Mohanty
Mobile Robot Path-Planning Using Oppositional-Based Improved
Firefly Algorithm Under Cluttered Environment . . . . . . . . . . . . . . . . . . 141
Mohit Ranjan Panda, Susmita Panda, Rojalina Priyadarshini,
and Pradipta Das
Design and Modal Analysis of Few-Mode Fibers for Spatial Division
Multiplexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
Bhagyalaxmi Behera, S. K. Varshney, and Mihir Narayan Mohanty
Speech Enhancement Using Wiener Filter Based on Voiced Speech
Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Rashmirekha Ram, Abhisek Das, and Saumendra Kumar Mohapatra
Analysis of Different Parameter of FSO Communication System
by Using Robust Model Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Shubham Mahajan and Amit Kant Pandit
Contents ix
An Improved Grasshopper Optimization Algorithm for Solving

Numerical Optimization Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Puneet Mishra, Vishal Goyal, and Aasheesh Shukla
Design of Array Antenna for Body Area Network . . . . . . . . . . . . . . . . . 189
Sarmistha Satrusallya, Shaktijeet Mahapatra, and Mihir Narayan Mohanty
Optimal Load Condition for a Resonant Wireless Charging System . . . 194
Jagadish C. Padhi, Durga P. Kar, Praveen P. Nayk, Biswaranjan Swain,
Dipti Patnaik, and Satyanarayan Bhuyan
Design of Microstrip Antenna with Corner Truncation for UWB
Application and Its Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Swarnaprava Sahoo and Laxmi Prasad Mishra
Analysis of Vector-Controlled Onshore DFIG with STATCOM
for Power Compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
Sovit Kumar Pradhan, Radhakrishna Das, Debswarup Rath,
and S. S. P. B. K. Prasad
Dictionary Design for Block-Based Intra-image Compression . . . . . . . . 217
Arabinda Sahoo and Pranati Das
Performance Analysis of Fractional Order Low-pass Filter . . . . . . . . . . 224
Kumar Biswal, Madhab Chandra Tripathy, and Sanjeeb Kar
Chronic Disease Risk (CDR) Prediction in Biomedical Data
Using Machine Learning Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
Lambodar Jena, Soumen Nayak, and Ramakrushna Swain
A Novel Approach to Detection of and Protection from Sybil Attack
in VANET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
Binod Kumar Pattanayak, Omkar Pattnaik, and Sasmita Pani
Design of Monochromatic Photonic Filter in Near-Infrared Region
Using Plane Wave Expansion Method . . . . . . . . . . . . . . . . . . . . . . . . . . 248
K. P. Swain, Sangram Kishore Mohanty, P. S. Das, and G. Palai
Modeling, Analysis, and Control of Vehicle Suspension System Based
on Linear Quadratic Regulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
Akshaya Kumar Patra, Alok Kumar Mishra, Anuja Nanda,
Narendra Kumar Jena, and Bidyadhar Rout
Performance and Evaluation of Different Kernels in Support Vector
Machine for Text Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
Ashish Kumar Mourya, ShafqatUlAhsaan, and Harleen Kaur
High Gain Slot Antenna by Using Artificial Magnetic Conductor . . . . . 272
Prakash Kumar Panda and Debalina Ghosh
x Contents
Electrical Energy Storage: A Great Business Ahead . . . . . . . . . . . . . . . 278

Renu Sharma, Satish Choudhury, Bibekananda Jena, and Sonali Goel
Embedded Entropy-Based Registration of CT and MR Images . . . . . . . 289
Sunita Samant, Subhaluxmi Sahoo, and Pradipta Kumar Nanda
Power Factor Corrected SEPIC and Cuk
Converter—A Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
Alok Kumar Mishra, Akshaya Kumar Patra, Ramachandra Agrawal,
Lalit Mohan Satapathy, Shekharesh Barik, Samarjeet Satapathy,
and Jnana Ranjan Swain
An Efficient Classifier-Based Approach for Early Arrhythmia
Detection with Feature Reduction Using Ranker Search Algorithm . . . . 313
Monalisa Mohanty, Asit Kumar Subudhi, Pradyut Kumar Biswal,
and Sukanta Sabut
The Fractional Order PID Controller Design for BG Control in Type-I
Diabetes Patient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Akshaya Kumar Patra, Anuja Nanda, Santisudha Panigrahi,
and Alok Kumar Mishra
Mutual Fund Investment Method Using Recurrent Back Propagation
Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
Smruti Rekha Das, Debahuti Mishra, Pournamasi Parhi,
and Prajna Paramita Debata
Stabilizing and Trajectory Tracking of Inverted Pendulum Based
on Fractional Order PID Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
Akshaya Kumar Patra, Alok Kumar Mishra, Anuja Nanda,
Dillip Kumar Subudhi, Ramachandra Agrawal, and Abhishek Patra
Inter-area and Intra-area Oscillation Damping of Power System
Stabilizer Design Using Modified Invasive Weed Optimization . . . . . . . 347
Mohammad Salik, Pravat Kumar Rout, and Mihir Narayan Mohanty
IoT-Based Baby Swing Monitoring, Alerting and Security System . . . . . 360
Shuvendra Kumar Tripathy and Pradyut Kumar Biswal
Effective ECG Beat Classification and Decision Support System
Using Dual-Tree Complex Wavelet Transform . . . . . . . . . . . . . . . . . . . . 366
Nihar Kar, Badrinarayan Sahu, Sukanta Sabut, and Santanu Sahoo
Classical 2D Face Recognition: A Survey on Methods, Face Databases,
and Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
Manoj Kumar Naik and Aneesh Wunnava
Multi-level Authentication-Based Secure Aware Data Transaction
on Cloud Using Cyclic Shift Transposition Algorithm . . . . . . . . . . . . . . 384
Prasanta Kumar Bal and Sateesh Kumar Pradhan
Contents xi
A Novel Graphical User Interface-Based Toolbox for Optimization

and Design of Linear Antenna Array . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
Guru Prasad Mishra, Shibanee Dash, and Saumendra Kumar Mohanty
Wideband Patch Antenna Element for Broadcast Applications
in Lower 5G Bands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
Sonagara Abhishek M and Rahul Vishwakarma
Anonymized Credit Card Transaction Using Machine Learning
Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
B. K. Padhi, S. Chakravarty, and B. N. Biswal
Performance Comparison of Adaptive Algorithms for Smart Antenna
of Dipole Array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
Barsa Samantaray, Kunal Kumar Das, and Jibendu Sekhar Roy
Effect of Buffer Size on Performance of Wireless Sensor Network . . . . . 432
Manoranjan Das and Benudhar Sahu
Noise Cancellation Using a Novel Self-adaptive Neuro-fuzzy Inference
System (SANFIS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
Laxmipriya Samal, Debashisa Samal, and Badrinarayan Sahu
Frequency Regulation in an Islanded Microgrid with Optimal
Fractional Order PID Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
Narendra Kumar Jena, Subhadra Sahoo, Amar Bijaya Nanda,
Binod Kumar Sahu, and Kanungo B. Mohanty
Biometric Detection Using Stroke Dynamics . . . . . . . . . . . . . . . . . . . . . 458
Abhishek Das, Saumendra Kumar Mohapatra, and Laxmi Prasad Mishra
Hardware Partitioning Using Parallel Genetic Algorithm to Improve
the Performance of Multi-core CPU . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
Suryakanta Nayak and Mrutyunjaya Panda
An Internet of Healthcare Things (IoHT)-Based Healthcare
Monitoring System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475
Harleen Kaur, Mohd. Atif, and Ritu Chauhan
Botnet Detection Technology Based on DNS-Based Approach . . . . . . . . 483
Bhavya Alankar
Development of an IoT-Based Tourism Guide System . . . . . . . . . . . . . . 495
Vedanta Prusty, Abhisek Rath, Kshirod Kumar Rout,
and Sivkumar Mishra
Design and Simulation of a Standalone Photovoltaic System Using
Synchronous Boost Converter and Reduced Switch Five-Level
Inverter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504
Kshirod Kumar Rout, Sivkumar Mishra, and Pravat Kumar Biswal
xii Contents
Designing and Sizing of a Stand-alone Photovoltaic System:

A Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514
Kshirod Kumar Rout, Sivkumar Mishra, and Pravat Kumar Biswal
A Model for Optimizing Cost of Energy and Dissatisfaction
for Household Consumers in Smart Home . . . . . . . . . . . . . . . . . . . . . . . 523
Nilima R. Das, Satyananda C. Rai, and Ajit Nayak
A Circular Ultra-Wideband Antenna for Wearable Applications . . . . . . 532
Shaktijeet Mahapatra, Sarmistha Satrusallya, and Mihir Narayan Mohanty
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537

Editors and Contributors
About the Editors
Dr. Mihir Narayan Mohanty received his M.Tech. in Communication System

Engineering from Sambalpur University, Odisha, and his Ph.D. in Applied Signal
Processing from Biju Patnaik University of Technology, Odisha. He is currently
serving as a Professor at the Department of Electronics and Communication
Engineering, Institute of Technical Education and Research (FET), Siksha ‘O’
Anusandhan (Deemed University), Bhubaneswar, Odisha. With 24 years of
teaching experience, his research interests include intelligent signal and image
processing, digital signal/image processing, biomedical signal processing, micro-
wave communication engineering, and antennas. He has published over 300
research papers in national and international journals and conference proceedings,
and has received various national and international awards for his contributions. He
is an active member of numerous professional societies, e.g. the IEEE, IET, IETE,
ISTE, EMC & EMI Engineers India, etc. Dr. Mohanty is also an active reviewer for
the IEEE, Elsevier, Springer, and several international conferences.
Swagatam Das received his B. E. Tel. E., M. E. Tel. E (Control Engineering

specialization) and Ph. D. degrees, all from Jadavpur University, India, in 2003,
2005, and 2009 respectively. Currently serving as an Associate Professor at the
Electronics and Communication Sciences Unit of the Indian Statistical Institute,
Kolkata, India, his research interests include evolutionary computing, pattern
recognition, multi-agent systems, and wireless communication. Dr. Das has pub-
lished more than 300 research articles in peer-reviewed journals and international
conference proceedings. He is the founding co-editor-in-chief of Swarm and
Evolutionary Computation, an international journal from Elsevier. He has also
served as or is serving as an associate editor of various journals, including Pattern
Recognition, Neurocomputing, Information Sciences, IEEE Access, and so on.
xiii
xiv Editors and Contributors
He is an editorial board member of numerous other journals, including Progress in

Artificial Intelligence, Applied Soft Computing, and Artificial Intelligence Review.
Dr. Das is the recipient of the 2012 Young Engineer Award from the Indian
National Academy of Engineering (INAE), and of the 2015 Thomson Reuters
Research Excellence India Citation Award for the highest cited researcher from
India in the Engineering and Computer Science category for the period 2010 to
2014.
Contributors
Ramachandra Agrawal Department of CSIT, ITER, S‘O’A University, Deemed

to be University, Bhubaneswar, India;
EEE Department, ITER, S‘O’A (Deemed to be University), Bhubaneswar, India
Sunanda Alamuru SreeNidhi Institute of Science and Technology, Hyderabad,
India
Bhavya Alankar Department of Computer Science and Engineering, School of
Engineering Sciences and Technology, New Delhi, India
Bhukya Arun Kumar School of Electronics and Electrical Engineering, Lovely
Professional University, Phagwara, India
Mohd. Atif Department of Computer Science and Engineering, School of
Engineering Sciences and Technology, JamiaHamdard, New Delhi, India
Prasanta Kumar Bal Department of Computer Science & Application, Utkal
University, Bhubaneswar, India
Shekharesh Barik CSE Department, DRIEMS (Autonomous), Cuttack, India
Bhagyalaxmi Behera ITER, Siksha ‘O’ Anusandhan (Deemed to be University),
Bhubaneswar, India
Debasis Behera C. V. Raman College of Engineering, Bhubaneswar, Odisha,
India
Suchismita Behera Department of ECE, Siksha‚ ‘O’ Anusandhan (Deemed to be
University), Bhubaneswar, Odisha, India
Satyanarayan Bhuyan Department of Electronics and Communication
Engineering, SOA Deemed to be University, Bhubaneswar, Odisha, India
B. N. Biswal Department of Computer Science & Engineering, Bhubaneswar
Engineering Colleges, Bhubaneswar, Odisha, India
Bedanta Kumar Biswal Electronics and Communication Engineering, Faculty of
Engineering, Siksha ‘O’Anusandhan (Deemed to be University), Bhubaneswar,
India
Editors and Contributors xv
Kumar Biswal School of Electronics Engineering, Kalinga Institute of Industrial

Technology (DU), Bhubaneswar, India
Pradyut Kumar Biswal Department of Electronics Engineering, IIIT
Bhubaneswar, Bhubaneswar, India
Pravat Kumar Biswal Department of EEE, KIIT, Bhubaneswar, Odisha, India
S. Chakravarty Department of Computer Science & Engineering, Centurion
University of Technology & Management, Bhubaneswar, Odisha, India
Ritu Chauhan Centre for Computational Biology and Bioinformatics, Amity
University, Noida, U.P, India
Satish Choudhury Department of Electrical Engineering, ITER Siksha ‘O’
Anusandhan (Deemed to be University), Bhubaneswar, India
Abhisek Das ITER, Siksha ‘O’ Anusandhan (Deemed to be University),
Bhubaneswar, India
Abhishek Das ITER, Siksha ‘O’ Anusandhan (Deemed to be University),
Bhubaneswar, India
Kunal Kumar Das Electronics and Communication Engineering, Institute of
Technical Education and Research, Siksha ‘O’ Anusandhan Deemed University,
Manoranjan Das Institute of Technical Education and Research, SOA (Deemed
to be University), Bhubaneswar, India
Nilima R. Das Faculty of Engineering & Technology, Siksha ‘O’ Anusandhan
Deemed to Be University, Bhubaneswar, Odisha, India
Niva Das Department of ECE, Siksha‚ ‘O’ Anusandhan (Deemed to be
University), Bhubaneswar, Odisha, India
P. S. Das Department of Electronics and Communication Engineering, GITA,
Bhubaneswar, India
Pradipta Das Veer Surendra Sai University of Technology, Burla, Odisha, India
Pranati Das Department of Electrical Engineering, Indira Gandhi Institute of
Technology, Sarang, Odisha, India
Radhakrishna Das Department of Electrical Engineering, GIET, Bhubaneswar,
India
Rajendra Kumar Das Department of ENTC, DRIEMS Autonomous Engineering
College, Tangi, Cuttack, Odisha, India
S. R. Das Department of Electronics and Communication Enginerring, GITA,
Bhubaneswar, India
xvi Editors and Contributors
Smruti Rekha Das Department of Computer Science and Engineering, Gandhi

Institute for Education and Technology, Bhubaneswar, Odisha, India
Lipsa Dash New Horizon College of Engineering, Bengaluru, India
Meera Dash Department of Electronics and Communication Engineering, SOA
(Deemed to be) University, Bhubaneswar, India
Sachikanta Dash Department of CSE, DRIEMS Autonomous Engineering
College, Tangi, Cuttack, Odisha, India
Shibanee Dash Department of ECE, FET, ITER, S‘O’A Deemed to be University,
Bhubaneswar, India
Prajna Paramita Debata International Institute of Information Technology,
Dinesh Kumar Garg Department of Computer Science and Engineering, SSCET,
Pathankot, India
Debalina Ghosh School of Electrical Sciences, IIT Bhubaneswar, Bhubaneswar,
India
Nagendra Kumar Goel GoVivace Inc., McLean, VA, USA
Sonali Goel Department of Electrical Engineering, ITER Siksha ‘O’ Anusandhan
(Deemed to be University), Bhubaneswar, India
Vishal Goyal Department of Electronics and Communication Engineering, GLA
University, Mathura, India
Bibekananda Jena Department of Electrical Engineering, ITER Siksha ‘O’
Anusandhan (Deemed to be University), Bhubaneswar, India
Lambodar Jena Department of Computer Science and Engineering, Siksha O
Anusandhan (Deemed to Be University), Bhubaneswar, India
Narendra Kumar Jena Department of Electrical Engineering, Siksha ‘O’
Anusandhan Deemed to be University, Bhubaneswar, Odisha, India;
Department of EEE, ITER, S‘O’A University, Bhubaneswar, India
Durga P. Kar Department of Electronics and Communication Engineering, SOA
Deemed to be University, Bhubaneswar, Odisha, India
Nihar Kar Department of ECE, Faculty of Engineering, Siksha ‘O’ Anusandhan,
Sanjeeb Kar ITER, SoA, Deemed to be University, Bhubaneswar, Odisha, India
Harleen Kaur Department of Computer Science and Engineering, School of
Mehreen Kirmani Department of Computer Science and Engineering, SSCET,
Pathankot, India
Editors and Contributors xvii
Sonagara Abhishek M Department of Electronics and Communication

Engineering, ITER, Siksha ‘O’ Anusandhan (Deemed to be University),
Shubham Mahajan School of Electronics & Communication, Shri Mata Vaishno
Devi University, Katra, Reasi, J&K, India
Shaktijeet Mahapatra Department of ECE, ITER, Siksha ‘O’ Anusandhan
Suman Malik School of Electrical Science, Indian Institute of Technology
Bhubaneswar, Bhubaneswar, Odisha, India
Alok Kumar Mishra Department of EEE, ITER, S‘O’A University, Deemed to be
Debahuti Mishra Department of Computer Science and Engineering, Siksha
‘O’Anusandhan Deemed to be University, Bhubaneswar, Odisha, India
Guru Prasad Mishra Department of ECE, FET, ITER, S‘O’A Deemed to be
Laxmi Prasad Mishra Department of Electronics and Communication
Engineering, Institute of Technical Education and Research, Siksha ‘O’
Anusandhan (Deemed to be University), Bhubaneswar, Odisha, India
Puneet Mishra Department of Electrical and Electronics Engineering, Birla
Institute of Technology and Science, Pilani, Jhunjhunu, Rajasthan, India
Sivkumar Mishra Department of EE, CAPGS, BPUT, Rourkela, Odisha, India;
Department of EEE, CAPGS, BPUT, Rourkela, Odisha, India
Bibhuprasad Mohanty Department of Electronics and Communication
Engineering, ITER, Siksha ‘O’ Anusandhan University (Deemed to be University),
Bhubaneswar, India
Kanungo B. Mohanty Department of Electrical Engineering, NIT, Rourkela,
Rourkela, Odisha, India
Mihir Narayan Mohanty Department of ECE, ITER, Siksha ‘O’ Anusandhan
Monalisa Mohanty Department of Electronics and Communication, Institute of
Technical Education & Research, SOA Deemed to be University, Bhubaneswar,
India
Sangram Kishore Mohanty Department of ECE, ICE, GITA, Bhubaneswar,
India;
Department of ECE, ICE, Bhubaneswar, India
Saumendra Kumar Mohanty Department of ECE, FET, ITER, S‘O’A Deemed
to be University, Bhubaneswar, India
xviii Editors and Contributors
Saumendra Kumar Mohapatra ITER, Siksha ‘O’ Anusandhan (Deemed to be

University), Bhubaneswar, India
Subhasish Mohapatra Department of Computerscience Engineering, Adamas
University, Kolkata, India
Laxmipriya Moharana Electronics and Communication Engineering, Faculty of
Engineering, Siksha ‘O’Anusandhan (Deemed to be University), Bhubaneswar,
India
Ashish Kumar Mourya Department of Computer Science and Engineering,
School of Engineering Sciences and Technology, JamiaHamdard, New Delhi, India
Manoj Kumar Naik Department of ECE, ITER, Siksha O Anusandhan,
Suraj Naik Electronics and Communication Engineering, Faculty of Engineering,
Siksha ‘O’Anusandhan (Deemed to be University), Bhubaneswar, India
Amar Bijaya Nanda Department of Electrical Engineering, Siksha ‘O’
Anusandhan Deemed to be University, Bhubaneswar, Odisha, India
Anuja Nanda Department of EEE, ITER, S‘O’A University, Deemed to be
Pradipta Kumar Nanda Image & Video Analysis Laboratory, Department of
ECE, S O A, Deemed to be University, Bhubaneswar, India
Ajit Nayak Faculty of Engineering & Technology, Siksha ‘O’ Anusandhan
Soumen Nayak Department of Computer Science and Engineering, Siksha O
Anusandhan (Deemed to Be University), Bhubaneswar, India;
Department of Computer Science and Engineering, Indian Institute of Technology
(ISM), Dhanbad, India
Suryakanta Nayak Department of Computer Science and Application, Utkal
Praveen P. Nayk Department of Electronics and Communication Engineering,
SOA Deemed to be University, Bhubaneswar, Odisha, India
B. K. Padhi Department of Computer Science & Engineering, Centurion
University of Technology & Management, Bhubaneswar, Odisha, India
Jagadish C. Padhi Department of Electronics and Communication Engineering,
G. Palai Department of Electronics and Communication Engineering, GITA,
Bhubaneswar, India
Hemanta Kumar Palo Siksha O Anusandhan (Deemed to Be University),
Editors and Contributors xix
Leeza Panda Department of Electronics and Communication Engineering, ITER,

Siksha ‘O’ Anusandhan University (Deemed to be University), Bhubaneswar, India
Mohit Ranjan Panda School of Computer Engineering, KIIT University,
Mrutyunjaya Panda Department of Computer Science and Application, Utkal
Prakash Kumar Panda Department of ECE, ITER, SOA Deemed to be
Susmita Panda Institute of Technical Education and Research, SOA Deemed to
be University, Bhubaneswar, Odisha, India
Amit Kant Pandit School of Electronics & Communication, Shri Mata Vaishno
Devi University, Katra, Reasi, J&K, India
Sasmita Pani Department of Computer Science and Engineering, Government
College of Engineering, Keonjhar, Odisha, India
Santisudha Panigrahi Department of CSE, ITER, S‘O’A University,
Bhubaneswar, India
Trilochan Panigrahi Department of Electronics and Communication Engineering,
SOA (Deemed to be) University, Bhubaneswar, India
Pournamasi Parhi Department of Computer Science and Engineering, Siksha
‘O’Anusandhan Deemed to be University, Bhubaneswar, Odisha, India
Smita Parija Department of Eletronics, CV Raman College of Engineering,
Odisha, India
Dipti Patnaik Department of Electronics and Communication Engineering, SOA
Abhishek Patra Department of CSIT, ITER, S‘O’A University, Deemed to be
Akshaya Kumar Patra Department of EEE, ITER, S‘O’A University, Deemed to
be University, Bhubaneswar, India
Binod Kumar Pattanayak Department of Computer Science and Engineering,
Institute of Technical Education and Research, Siksha ‘O’ Anusandhan Deemed to
Omkar Pattnaik Department of Computer Science and Engineering, Government
College of Engineering, Keonjhar, Odisha, India
Sateesh Kumar Pradhan Department of Computer Science & Application, Utkal
Sovit Kumar Pradhan Department of Electrical Engineering, NIT, Silchar, India
xx Editors and Contributors
S. S. P. B. K. Prasad Department of Electrical & Electronics Engineering, GIET,

Bhubaneswar, India
Rojalina Priyadarshini School of Computer Engineering, KIIT University,
Pankaj Prusty Electronics and Communication Engineering, Faculty of
Bhubaneswar, India
Vedanta Prusty Department of ETC, IIIT, Bhubaneswar, Odisha, India
Satyananda C. Rai Department of IT, Silicon Institute of Technology,
Raman Raj Electronics and Communication Engineering, Faculty of Engineering,
Siksha ‘O’Anusandhan (Deemed to be University), Bhubaneswar, India
Rashmirekha Ram ITER, Siksha ‘O’ Anusandhan (Deemed to be University),
Bhubaneswar, India
Abhisek Rath Department of ETC, IIIT, Bhubaneswar, Odisha, India
Debswarup Rath Department of Electrical & Electronics Engineering, GIET,
Bhubaneswar, India
Bidyadhar Rout Department of EEE, Veer Surendra Sai University of
Technology, Burla, India
Bikash Chandra Rout Siksha O Anusandhan (Deemed to Be University),
Kshirod Kumar Rout Department of EE, IIIT, Bhubaneswar, Odisha, India;
Department of EEE, IIIT, Bhubaneswar, Odisha, India
Pravat Kumar Rout Department EEE, ITER, Siksha ‘O’ Anusandhan (Deemed
to be University), Bhubaneswar, India
Jibendu Sekhar Roy School of Electronics Engineering, Kalinga Institute of
Industrial Technology (KIIT), Deemed to be University, Bhubaneswar, Odisha,
India
Sukanta Sabut School of Electronics Engineering, KIIT Deemed to beUniversity,
Arabinda Sahoo Department of ECE, ITER, Siksha ‘O’ Anusandhan Deemed to
Santanu Sahoo Department of ECE, Faculty of Engineering, Siksha ‘O’
Anusandhan, Bhubaneswar, Odisha, India
Subhadra Sahoo Department of Electrical Engineering, Siksha ‘O’ Anusandhan
Editors and Contributors xxi
Subhaluxmi Sahoo Image & Video Analysis Laboratory, Department of ECE, S

O A, Deemed to be University, Bhubaneswar, India
Swarnaprava Sahoo Department of Electronics and Communication Engineering,
Institute of Technical Education and Research, Siksha ‘O’ Anusandhan (Deemed to
be University), Bhubaneswar, Odisha, India
Tamanna Sahoo Department of Electronics and Communication Engineering,
ITER, Siksha ‘O’ Anusandhan University (Deemed to be University),
Bhubaneswar, India
Tapasmini Sahoo Electronics and Communication Engineering, Institute of
Technical Education and Research, Siksha ‘O’ Anusandhan Deemed University,
Bhubaneswar, India
Badrinarayan Sahu Department of ECE, Faculty of Engineering, ITER, Siksha
‘O’ Anusandhan (Deemed to be University), Bhubaneswar, Odisha, India
Benudhar Sahu Institute of Technical Education and Research, SOA (Deemed to
be University), Bhubaneswar, India
Binod Kumar Sahu Department of Electrical Engineering, Siksha ‘O’
Anusandhan Deemed to be University, Bhubaneswar, Odisha, India
Prasant Kumar Sahu Indian Institute of Technology Bhubaneswar, Argul,
Odisha, India;
School of Electrical Science, Indian Institute of Technology Bhubaneswar,
Sanjay Kumar Sahu School of Electronics and Electrical Engineering, Lovely
Professional University, Phagwara, India
Mohammad Salik Department EEE, ITER, Siksha ‘O’ Anusandhan (Deemed to
Debashisa Samal Department of ECE, ITER, S‘O’A Deemed to be University,
Bhubaneswar, India
Laxmipriya Samal Department of ECE, ITER, S‘O’A Deemed to be University,
Bhubaneswar, India
Sunita Samant Image & Video Analysis Laboratory, Department of ECE, S O A,
Deemed to be University, Bhubaneswar, India
Barsa Samantaray ECE Department, Institute of Technical Education and
Research, Siksha ‘O’ Anusandhan Deemed to be University, Bhubaneswar, Odisha,
India
Kandarpa Kumar Sarma Department of Electronics and Communication
Engineering, Gauhati University, Guwahati, India
xxii Editors and Contributors
Mousmita Sarma Department of Electronics and Communication Engineering,

Gauhati University, Guwahati, India
Lalit Mohan Satapathy EEE Department, ITER, S‘O’A (Deemed to be
University), Bhubaneswar, India
Samarjeet Satapathy EEE Department, ITER, S‘O’A (Deemed to be University),
Bhubaneswar, India
Sarmistha Satrusallya Department of ECE, ITER, Siksha ‘O’ Anusandhan
Kaibalya Kumar Sethi Department of Electronics and Communication
Engineering, GITA, Bhubaneswar, India
ShafqatUlAhsaan Department of Computer Science and Engineering, School of
Renu Sharma Department of Electrical Engineering, Institute of Technical
Education and Research, SOA (Deemed to be) University, Bhubaneswar, India
Aasheesh Shukla Department of Electronics and Communication Engineering,
GLA University, Mathura, India
Asit Kumar Subudhi Department of Electronics and Communication, Institute of
Technical Education & Research, SOA Deemed to be University, Bhubaneswar,
India
Dillip Kumar Subudhi Department of CSIT, ITER, S‘O’A University, Deemed to
be University, Bhubaneswar, India
A. Susmitha New Horizon College of Engineering, Bengaluru, India
Biswaranjan Swain Department of Electronics and Communication Engineering,
Jnana Ranjan Swain EEE Department, ITER, S‘O’A (Deemed to be University),
Bhubaneswar, India
K. P. Swain Department of Electronics and Communication Engineering, GITA,
Bhubaneswar, India
Ramakrushna Swain Department of Computer Science and Engineering, Silicon
Institute of Technology, Bhubaneswar, India
Ramji Tangudu Indian Institute of Technology Bhubaneswar, Argul, Odisha,
India
Madhab Chandra Tripathy Department of Instrumentation & Electronics
Engineering, CET, BPUT, Bhubaneswar, India
Shuvendra Kumar Tripathy Trident Academy of Technology, Bhubaneswar,
India
Editors and Contributors xxiii
S. K. Varshney Department of Electronics and Electrical Communication

Engineering, IIT Kharagpur, Kharagpur, India
Rahul Vishwakarma Department of Electronics and Communication
Aneesh Wunnava Department of ECE, ITER, Siksha O Anusandhan,
Children’s Age and Gender Recognition
from Raw Speech Waveform Using DNN
Mousmita Sarma1(&), Kandarpa Kumar Sarma1,

and Nagendra Kumar Goel2
1
Department of Electronics and Communication Engineering, Gauhati
University, Guwahati, India
{mousmita.s,kandarpaks}@gauhati.ac.in
2
GoVivace Inc., McLean, VA, USA
nagendra.goel@govivace.com
1 Introduction
Speaker’s age and gender are two most important para-linguistic characteristics
extracted from speech signal, and there has been extensive research conducted in this
direction. Extraction of age and gender like information from children speech (below
14 years) is considered as a difficult task, and it has not been very well studied. Most of
the previous works are focused mainly on adult’s speech. A few works considered
children as a separate age group along with other age groups like young, adult, senior,
etc. for the design of classifiers [1, 2]. The reason for children’s speech not getting
much attention may be because greater extent of variability is involved in this age
group. This is mainly due to children’s shorter vocal tracts, higher fundamental fre-
quencies, developing cognitive issues relating to language acquisition before reaching
the adult stage, physiological variations in growth, etc. [3] and [4].
This work is focused on the design of end-to-end DNN for automatic extraction of
age and gender information from speech signal of children within the age group of
4–14 years. Initially, we design separate DNN for both the tasks, using age or gender
target in the output layer. However, age and gender are speaker characteristics which
may have shared interdependencies. Therefore, such tasks may better fit in a multi-task
learning setup and help to improve recognition accuracy. To explore such possibility,
we also design multi-task learning DNN, where we use speaker’s age and gender as
target in two separate output layers, where shared hidden representation is used to learn
both the objectives. The proposed multi-task learning DNN system leverages speaker’s
age information to improve gender identification accuracy and vice versa.
The DNN models developed in this work uses data-driven approach of filter
learning from time-delay waveform, which is previously described in [5, 6]. It uses
convolutional layer and network-in-network (NIN) layer for DNN front end, and the
learned feature representation is used by temporal context modeling. Thus, the DNN
learns features within the network and jointly optimizes age and gender identification
objective with the rest part of the network. Such networks are able to learn evidence of
features during training time. For temporal context modeling in the DNN, we use
interleaving time-delay neural network (TDNN) with unidirectional long short-term

M. N. Mohanty and S. Das (eds.), Advances in Intelligent Computing
and Communication, Lecture Notes in Networks and Systems 109,
https://doi.org/10.1007/978-981-15-2774-6_1
2 M. Sarma et al.
memory (LSTM) cell-based recurrent layer (TDNN-LSTM)-based configuration. The

configuration also uses time-restricted attention mechanisms at the end to make the
network attentive toward relevant features of speech. The proposed DNN is trained and
evaluated on OGI Kids’ corpus [7]. We provide a description of the proposed multi-
task learning DNN for gender recognition in Sect. 2. Section 3 explains experimental
setup which includes details of train and test datasets, network design consideration and
result and analysis. In Sect. 4, we summarize the findings of this work and subse-
quently conclude the description.
2 System Description
In this work, we design single-task and multi-task DNN models for age and gender
recognition from children’s speech as shown in Fig. 1a and b. The DNN has a front end
consisting of convolutional layer and NIN layers. Convolutional filter learns band-pass
filter like structure within DNN and generates time-delay features. In our previous work
on the domain of emotion recognition [6], we reported that this time-delay raw
waveform front end layer learns emotion-specific cues better than MFCC features. This
raw waveform front end is a four-layer block jointly trained with rest part of the
network. A 1 − d time convolution layer consisted of 100 filters of 31.25 ms dimen-
sion operates on 40 ms raw signal with step size 1.25 ms. Absolute logarithm of filter
outputs are computed using a logarithmic layer, and next filter outputs are aggregated
using two trainable NIN nonlinearity layers. More details of the raw waveform front
end are given in [5, 6].
Another crucial issue in speech-based application is temporal context modeling of
speech. Particularly for para-linguistic information extraction, it is important to pre-
serve long temporal context since such information perhaps mostly lies in a longer span
of time. In this work, we explore temporal convolution in the form of TDNN layers
along with LSTM layers for long temporal context modeling. Interleaving of temporal
convolution with unidirectional LSTM is reported to outperform bidirectional LSTM to
preserve future context. We use higher frame rate at lower layers of LSTM and TDNN
layer in the network, and we decrease layer frame rate with layer depth. The dimen-
sions of projection and the recurrence in LSTM layers are one-quarter of the cell
dimension. We found 128 cell dimension optimal for the current gender identification
task with the recurrence of dimension 32 and the LSTM output of dimension 64. The
LSTMs operate with a recurrence that spans 3 time steps. Time-restricted attention
layer is used as the last layer along with TDNN and unidirectional LSTM layer. Use of
an attention layer also helps the network to put less weight in non-speech or non-
relevant regions of the signal. Removing non-speech frames using fixed speech activity
detection (SAD) module beforehand may create a discontinuity in the temporal content.
However, DNN setups with an attention layer become more attentive toward the salient
regions of speech ignoring non-speech regions. The attention layer used has 12 heads, a
context of [−5, 2], a key dimension of 40 and value dimension of 60.
Children’s Age and Gender Recognition … 3
Fig. 1. a Single-task learning model. b Multi-task learning model
The output of attention layer is used as input to the output layers of the DNN. In the
single-task learning model (Fig. 1a), age or gender label are used as target. The age task
can be considered either as a regression or a classification problem. However, this work
models the 11 different age (4–14 years) of OGI Kids’ corpus as 11 different classes and
trains the network as a classification problem. Male and female gender is modeled as
separate output classes in a different output layer. Thus, the two output layers of the DNN
are trained using softmax layer with cross-entropy loss function. This gives the network
freedom to model any distribution over output. We train the network using per-frame
objective method, where the label of the class is repeated for each frame of the speech
chunk. On the other hand, we use speaker’s age and gender both as targets in two separate
output layers in the multi-task learning DNN model. It has three distinct parts: feature
extraction front end layers, temporal modeling layers and output layers. However, all
layers are jointly optimized with two separate output layers with gender and age target,
respectively. Therefore, the multi-task objective loss sums the loss of each task.
4 M. Sarma et al.
3 Experimental Details
This section describes the experimental details and results. All our experiments are
done using Kaldi toolkit. Following here is a detail description of experimental
considerations.
The proposed work is trained and tested on OGI Kids’ corpus [7]. We create train
and test partition of 1093 children for tenfold training. The entire dataset has children
from 4 years to 14 years age, i.e., total 11 different ages. We create ten test sessions
where every session contains on an average 100 non-overlapping speakers and against
every test session there is a train session made up of the speakers from the rest of the
test sessions.
To further increase the amount of data in the training set, we perform data aug-
mentation by means of amplitude and speed perturbation. Amplitude perturbation also
mitigates the effect of low and high amplitude variations of speech over ages.
Amplitude perturbation is applied by modulating the amplitude of the raw speech
signal. Each recording in the training data has been scaled with a random variable
drawn from a uniform distribution over [0.125, 2]. Speed perturbation is performed on
the amplitude modulated signals with speed factors 0.9, 1.0 and 1.1.
The performance of the gender identification tasks is reported using two parame-
ters, weighted accuracy (WA) which is the overall classification accuracy and
unweighted accuracy (UA) which is the average recall over the classes. On the other
hand, the performance of the age estimation is measured by computing the mean
absolute error (MAE) and root mean square error (RMSE) for every utterance of
evaluation datasets. The lower the MAE and RMSE, the better is the performance.
The MAE and RMSE is given by,
1X M
MAM ¼ jðEstimatedAge TrueAgeÞj ð1Þ
M i¼1
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u M
u1 X
RMSE ¼ t ðEstimatedAge TrueAgeÞ2 ð2Þ
M i¼1
where M is the total number of examples in the dataset.
3.1 DNN Design Consideration

One of the challenging issues in training speech-based task is to mitigate the effect of
variable lengths of training and test examples. It is important to minimize network
sensitivity to varying speech duration. One solution is to train the network on chunks of
different duration and making mini-batch size a function of chunk length. For example,
for the present work the length of training chunks is varied from 1 to 6 s, and the entire
utterance is used as an example if it is shorter than 6 s. Corresponding to it mini-batch
sizes used for training examples with length 100, 200, 300 are 128, 64, 32, respec-
tively, and so on. This way the total number of frames is almost equal in different mini-
batches. We call this training approach variable chunk and variable mini-batch.
However, we compare this variable chunk and variable mini-batch approach with fixed
chunk and fixed mini-batch approach as shown in Table 1. It can be observed that
using variable chunk and variable mini-batch network performs better.
Table 1. Effect of chunk length of training examples (test session 1 is used to derive this
comparison)
Training chunk length Age Gender
MAE RMSE WA UA
Fixed (100 frames) 1.61 2.42 75 74.24
Variable (100–600 frames) 1.51 2.33 82.69 82.30
We train DNNs with per-frame objective, i.e., label is repeated for every frame of
speech chunk. The learning rate of the training is set to gradually decrease from
6 10 − 4 to 6 10 − 5 over the course of 30 epochs. We use per-frame dropout
using the dropout schedule method, where the entire vector is forced to be zero or one.
The dropout schedule is expressed as a piece-wise linear function on the interval [0, 1],
where f (0) gives the dropout proportion at the start of training and f (1) gives the
dropout proportion after seeing all the data. A dropout schedule of the form 0, 0@0.20,
p@0.5, 0@0.75, 0 is used in this setup, where p is 0.3 in the results reported here.
Thus, the dropout probability is 0 at f (0), 0 at f (0.2), 0.3 at f (0.5), 0 at f (0.75) and 0 at
f (1). In DNNs trained with per-frame objective, we give some extra left context at the
time of decoding which provides flexibility to the network regarding number of frames
it sees in addition to what was provided during training and we evaluate the model
several times to tune this length of decode left time context. However, we also compare
the effect of a fixed length decode time left context with that where we provide an
effectively unlimited left context during decode time. Unlimited means the network is
allowed to reuse hidden state activations from previously computed chunk. As can be
seen from Table 2 unlimited left context does not help in the present multi-task
learning set up and finding an effective fixed length left context provides better
performance.
Table 2. Effect of decode time left context (test session 1 is used to derive this comparison)
Training chunk length Age Gender
MAE RMSE WA UA
20 frames 1.51 2.33 82.69 82.30
Unlimited left context 1.67 2.37 76.92 76.17
The speech segments are converted to 8000 Hz sample frequency and quantized at
16 bit per sample. Mean and variance normalization is done at the segment level and
low energy dither is added to the windowed signal to avoid runs of digital zeros. The
pre-processed raw speech utterances are passed to the input layer of the DNN.
6 M. Sarma et al.
Label smoothing regularizes our model based on a softmax with k output values by
replacing the hard 0 and 1 classification targets (one hot encoding) with targets of e/
k − 1 and 1 − e, respectively, where e is small constant.
3.2 Result and Analysis

Results of age and gender identification have been evaluated on ten test sessions of
OGI Kids’ corpus for both single-task and multi-task training. To evaluate every test
session, we use the model which is trained with rest of the test sessions data. The
single-task training of DNN follows the same process as shown in Fig. 1, which uses
only one output layer for the specific task. On the other hand, multi-task DNN uses an
auxiliary task trained in an extra output layer. The auxiliary task is gender for age
estimation and age for gender recognition task. The results are shown for files longer
than 30 s and files shorter than 30 s separately. It is observed that except test session 2,
all other test sessions achieve less MAE and RMSE using single-task training of DNN
for age estimation. In contrary to age estimation, we observe that for all test sessions,
multi-task training provides better WA and UA.
We summarize the results of age and gender recognition in Table 3, where average
performance has been shown. We achieved best age estimation MAE-RMSE pair of
0.91 and 1.43 using single-task DNN training and 1.13 and 1.72 using multi-task DNN
training for test set longer than 10 s. The results show that gender information is not
much helpful while estimating age. On the other hand, we obtain best gender recog-
nition WA-UA pair 55.63 and 50.97 using single-task DNN training and 72.67 and
70.55% using multi-task DNN training for test set longer than 10 s. The result clearly
tells that for gender recognition age information is highly useful which increases the
overall accuracy by 17.04%. Previous works by Safavi et al. 2014 and 2018 are also
focused on age and gender recognition from children’s speech and reported to obtain
78% accuracy on age independent gender recognition. However, we perform a tenfold
testing and obtain 82.69% highest accuracy among ten different test sessions and an
average 72.67% accuracy. On the other hand, in case of age recognition we attempt to
estimate exact age of the child which is a more difficult task compared to the work by
Safavi et al. 2014 and 2018 which have performed an age group classification. Table 3
also shows how the performance is affected by the duration of speech. In an obvious
manner, better performance is observed using longer duration of speech for both age
and gender recognition.
Table 3. Summary of results (averaged over all test sessions)

Model Duration of speech (s) Age Gender
MAE RMSE WA UA
Single-task 30 0.91 1.43 82.69 82.30
<30 1.13 1.72 59.31 50.33
Multi-task 30 1.39 2.04 72.67 70.55
<30 1.86 2.57 64.86 59.64
Table 4 reflects the gender recognition performance across the 11 age classes of
children using the multi-task DNN model. Although the observation here is not very
conclusive, it can be observed that (at least from the <30 s files results) for younger
children below 9 years accuracy is less. It clearly tells that gender discrimination in
case of younger children is more difficult. Table 5 shows the age estimation MAE and
RMSE across the two gender classes using the single-task DNN model. The MAE for
female and male is 0.85 and 0.89 for 30 s duration of speech and 1.21 and 1.61 for
<30 s speech duration. The observation here tells that, although age estimation per-
formance is not heavily biased toward one particular gender, MAE of male speakers is
little less. Perhaps, this is because the kids’ voice most of the time is like a female voice
and it is difficult even for human to discriminate male and female for kids younger than
7–8 years. Therefore, it is not very surprising that machine’s age estimation error is
more in case of male children.
Table 4. Result of gender identification across all ages for test session 1(multi-task learning
model)
Speaker’s age 30 s <30 s
WA UA %F %M WA UA %F %M
4 90 90 80 100 60.2 60.2 60 60.41
5 80 80 80 80 58.52 58.87 65.13 51.91
6 87.5 90 80 100 58.18 59.08 48.1 70.06
7 66.66 65 80 50 49.45 49.72 53.33 46.12
8 80 80 100 60 73.3 73.36 73.5 73.22
9 70 70 80 60 56.07 56.11 62.06 50.15
10 80 80 100 60 70.67 70.83 77.91 63.74
11 77.77 75 100 50 71.54 72.02 60.49 83.55
12 100 100 100 100 76.83 76.66 61.07 92.26
13 77.77 77.5 80 75 75.89 75.9 71.51 80.29
14 100 100 100 100 89.41 89.38 84.98 93.78
Table 5. Result of age identification across all genders for test session 1 (single-task learning
model)
Speaker’s gender 30 s <30 s
MAE RMSE WA UA
Female 0.85 1.28 1.21 1.88
Male 0.89 1.34 1.168 1.83
8 M. Sarma et al.
4 Summary of Findings and Conclusion
We described single-task and multi-task DNN performance for age and gender
recognition task. For single-task model, we train age and gender estimation objectives
separately with softmax output layer. On the other hand, for multi-task training we train
single DNN model with two different output layers for age and gender and jointly
optimize the objectives along with feature extraction and temporal modeling layers. In
order to exploit the feature learning and discriminating ability of multi-task learning
process of DNN, we used raw speech waveform in this work. A1 − d time convolu-
tional and NIN layer-based raw waveform front end have been used in the proposed
DNN, and these layers are jointly trained with the speech temporal modeling parts of
the DNN to optimize gender classification objectives. We use interleaving TDNN-
LSTM-based learning to preserve long temporal context of speech, and it uses time-
restricted self-attention mechanism to attend age- and gender-specific features of
speech. The results obtained throughout this study show that although children’s
gender recognition process is very much benefited by age information provided during
multi-task learning process, the vice versa is not true. We achieve less MAE in age
estimation task when we train the network with one output layer compared to the
network trained with gender auxiliary task. Therefore, further investigation of age
estimation process in multi-task learning DNN is required. We are not aware of any
previous works where multi-task learning process of DNN is utilized for para-linguistic
information extraction from children’s speech. The proposed work shall serve as a
baseline for DNN-based recognition system design from children’s speech.
References
1. Kumar N, Nasir M, Georgiou P, Narayanan SS (2016) Robust multichannel gender
classification from speech in movie audio. In: Proceeding INERSPEECH 2016, The 17th
annual conference of the international speech communication association. SanFrancisco,
USA, 8–12 September, 2016
2. Kabil SH, Muckenhirn H, Doss MM (2018) on learning to identify genders from raw speech
signal using CNNs. In: Proceeding INERSPEECH 2018, The 19th annual conference of the
international speech communication association. Hyderabad, India, 2–6 September, 2018
3. Safavi S, Russell M, Jancovic P (2014) Identification of age-group from childrens speech by
computers and humans. In: Proceeding INTERSPEECH 2014, The 15th annual conference of
the international speech communication association. Singapore, 14–18 September, 2014
4. Safavi S, Russell M, Jancovic P (2018) Automatic speaker, age-group and gender
identification from childrens speech. Comput Speech Lang 50:141–156
5. Ghahremani P, Manohar V, Povey D, Khudanpur S (2016) Acoustic modellingfrom the signal
domain using CNNs. In: Proceeding INTERSPEECH 2016, 17th annual conference of the
international speech communication association. San Francisco, CA, USA, 8–12 September
2016
6. Sarma M, Ghahremani P, Povey D, Goel NK, Sarma KK, Dehak N (2018) Emotion
identification from raw speech signals using DNNs. In: Proceeding INERSPEECH2018, The
19th annual conference of the international speech communication association. Hyderabad,
India, 2–6 September, 2018
7. Shobaki K, Hosom J, Cole (2000) The OGI kids speech corpus and recognizers. In:
Proceeding INTERSPEECH 2000, The 6th international conference on spoken language
processing, ICSLP 2000/INTERSPEECH 2000, Beijing, China, 16–20 October, 2000
Inter-vehicular Communication in Urban
Traffic Scenario
Suman Malik(&) and Prasant Kumar Sahu
School of Electrical Science, Indian Institute of Technology Bhubaneswar,

Bhubaneswar, Odisha 752050, India
sm38@iitbbs.ac.in
1 Introduction
Vehicular ad hoc network (VANET) aims at enhancing the transportation efficiency

and road traffic safety through the application of advance information processing,
communication and control techniques [1]. A key to implement the intelligent trans-
portation system (ITS) is traffic management and the road traffic state monitoring.
VANET is the most suitable solution to this regard. The unique features offered by
VANETs are higher mobility and higher density of mobile nodes [2]. VANET deals
with V2V and V2I communication link using wireless local area network (WLAN)
technologies [3]. Routing of data packet and selection of suitable wireless channel is a
major scientific challenging task in VANET because of higher mobility and dense
mobile nodes in urban environment [4]. Ad hoc on demand distance vector (AODV),
dynamic source routing (DSR) protocol and optimized link state routing (OLSR) are
generally employed in VANET to set up the communication link between V2V and
V2I in an ad hoc network scenario [5]. The wireless channel characteristics do play a
major role within the VANET disseminating of information between vehicles and
infrastructures. The channel impairments [6] especially path loss and fading effect also
affect the signal power during the propagation through free space channel. Many fading
channel models, like Rayleigh, Nakagami and Rician, were proposed as a wireless
channel corresponds to VANET [7].
With respect to the previously published paper, the main contribution of this paper
is outlined as follows;
• Primarily, ideal wireless channels are taken into consideration, and thereafter,
Rayleigh and Nakagami fading channel with higher network size in terms of larger
number of vehicles and random mobility is employed to VANET scenario for
performance analysis.
• We have also investigated the performance of the VANET, parameters like packet
loss ratio and average delay with different number of vehicles under higher mobility
condition.

https://doi.org/10.1007/978-981-15-2774-6_2
Inter-vehicular Communication in Urban Traffic Scenario 11
2 Related Work
Several investigations were reported in the scientific literature for studying and com-
paring the routing protocols to enhance the performance of VANETs [8–11]. The
significant characteristics of a routing protocol are to determine the effective way
between the source and destination for information dissemination in a reliable manner.
In the published literature, the performance analysis has been carried out using
metrics like PDR, delay, route length and overhead packets of VANET and by using
position-based routing approach in dense vehicular environment [8]. The overall per-
formance of VANET was studied using different routing protocols where wireless
channel and higher network size are not considered [9]. In [10], review and perfor-
mance analysis has been studied for VANETs’ protocols. However, the effect of
channel model and vehicular mobility on VANET was not considered in their studies.
Similarly, in the work of Chandel and Sharma et al. [11], cluster-based routing pro-
tocols were used to study the performance metric of VANET such as PDR, throughput,
jitter and average delay with respect to network size without considering the effect of
free space channel model, dense traffic and vehicular mobility.
3 VANET Model, Methodology and Simulation Setup
VANET testbed deployment is very expensive, and to avoid complicated and expen-
sive setup, software simulation platforms are in wide use by research community. We
have created and simulated the proposed model by using the NS2 and NetSim which
are event-oriented simulator which is used to carry out the simulation of the wireless
network and SUMO which is a road traffic simulator and is used to analyze the realistic
simulations of traffic mobility [12] (see Fig. 1). The main reason for selecting the
proposíed simulation environment is its capability to handle the distribution of vehicles
as well as the support for channel model.
3.1 Simulation Using NS2 and NetSim with SUMO Tools

The VANET model is simulated using NS2 using mobility trace that is generated using
SUMO. SUMO gains rapid research interest to the transportation engineer due to its
capability of multi-dimensional traffic simulation, traffic lights and for time schedule
generation. Here, we have considered ideal wireless channel model for VANET. The
step required to create and simulate the proposed scenario is shown in Fig. 2. NetSim
along with SUMO tool be used to perform the simulation of the proposed model where
realistic mobility-based traffic platforms are considered. It permits different vehicle
mobility modeling and simulation with different types of vehicle, traffic lights and
multi-lane roads with lane alternating features. The step required to create and simulate
the proposed scenario is shown in Fig. 3.
12 S. Malik and P. K. Sahu
Fig. 1. VANET simulation scenario a mobility model in NS2, b mobility model in NetSim,
c VANET scenario in NetSim
Fig. 2. Methodology flowchart for SUMO and NS2

Fig. 3. Methodology flowchart for SUMO and NetSim
3.2 Packet Loss Ratio

Packet loss ratio is the percentage of the MAC collisions packets loss or packet
dropped during total packet dissemination. Dense network increases the channel uti-
lization which is the cause of packet loss. It is expressed as [13]:
PLR ¼ Total packet dropped/Total packet send ð1Þ
3.3 Average Delay

Delay is the average time consumed by the node when the data packets delivered
successfully to the targeted vehicles. Smaller average delay should be maintained for
better network performance. Propagation, processing and transmission delay are the
main factor for average delay in VANET. Average delay is expressed as follows [14]:
Average Delay ¼ Time consumed to transfer packets successfully/

ð2Þ
Total no of packets delivered
4 Simulation Results
Figure 4 presents the PLR for the specified routing protocols and for different network
sizes. Here, network simulator NS2 tool is used for simulation using an ideal wireless
channel (see Table 1). We can observe that PLR is increasing with network size and is
changing with routing protocols. It is also noticed that packet drop is lower when less
vehicles are present in the VANET. DSR routing protocol provides the lower PLR with
respect to AODV and OLSR for same vehicles.
Fig. 4. Packet loss ratio versus routing protocol with different number of vehicles using NS2
Table 1. Simulation parameters

Parameters Values Parameters Values
Simulation tool NS2, NetSim 10.2 Bit rate 27 Mbps
Mobility model SUMO Link mode Half duplex
Simulation area Bhubaneswar city Transmitter power 40 mw
No of vehicles 30, 50, 80 link range 1 km
Simulation time 400 s Message type Unicast, broadcast
Fading models Rayleigh, Packet size 1420 bytes
Nakagami
Shadowing Lognormal model Frequency band 5.9 GHz
Vehicle speed 40−120 kmph Routing protocols DSR, AODV, OLSR
Performance PLR, average IEEE standard IEEE-1609, IEEE-
metric delay protocol 802.11p
In a similar way, Fig. 5a, b presents the same performance metric with respect to
total number of vehicles in VANET. Here, we have considered two fading channel
model, i.e., Rayleigh and Nakagami fading channel and NetSim-based simulation
platform. As shown in Figs. 4 and 5, packet loss ratio is significantly increased when
we changed the ideal wireless scenario to wireless fading channel model (Rayleigh and
Nakagami), due to more number of packets drop because of fading channel. This
observation shows that DSR could be a suitable routing protocol with Rayleigh fading
channel for application where lower packet loss is required over a large network.
(a) (b)
Packet loss ratio Vs Routing protocols Packet loss ratio Vs Routing protocols
with different network sizes (NetSim) with different network sizes (NetSim)
60 60
30Car 30Car
60Car 60Car
80Car 80Car
50 50
Nakagami Channel model Rayleigh Channel model

Packet loss ratio (%)

40 40
30 30
20 20
10 10
0 0
DSR AODV OLSR DSR AODV OLSR
Routing Protocols --> Routing Protocols -->
Fig. 5. Packet loss ratio versus routing protocol with different number of vehicles using NetSim
a Rayleigh channel model, b Nakagami channel model
Figure 6 demonstrates the relation between packet loss ratio and vehicular speed
with specified wireless channel model and routing protocol. Here, Rayleigh and
Nakagami fading channel model is considered with DSR routing protocol along with
variable vehicular mobility ranging from 40 kmph to 120 kmph to analyze the PLR in
VANET. We can observe that the PLR reached maximum of 50% at 120 kmph of
vehicle speed with the combination of DSR routing and Rayleigh fading channel,
whereas it becomes 54.2% with Nakagami fading channel (see Table 2). Hence, we can
conclude that PLR is increased with increase in vehicle speed for a particular number of
vehicles because of higher mobility of vehicles cause a fast alternate of topology and
packet drop in the network.
Figure 7 demonstrates the delay for three mentioned routing protocols under dif-
ferent number of vehicles. Two appropriate fading models, like Rayleigh and Nak-
agami fading channel, are introduced in VANET for simulation. It can be seen that the
average delay is stick to 0.8–1 ms and the maximum it can reach up to 1.3 ms. From
Fig. 7, we can determine that DSR is taking less time as compared to AODV and
(a) (b)
Packet loss Vs vehicular speed (NetSim) Packet loss Vs vehicular speed (NetSim)
55 55
30Car 30Car
50Car 50Car
50 80Car 50 80Car
DSR routing and

DSR routing and
45 45 Nakagami channel model
Rayleigh channel model

40 40
35 35
30 30
25 25
20 20
15 15
40 50 60 70 80 90 100 110 120 130 40 50 60 70 80 90 100 110 120 130
Car velocity (Kmph)--> Car velocity (Kmph)-->
Fig. 6. PLR versus vehicular speed a Rayleigh channel model b Nakagami channel mode
Table 2. Comparative analysis of PLR

Parameter DSR AODV OLSR Vehicles
PLR (%) (NetSim) Rayleigh 15.3 22 28 30
28 33 36 50
35.7 44 54 80
Nakagami 19.2 19.4 34.4 30
31 34.5 41 50
38.2 46.4 59.6 80
PLR (%) (NS2) 3.2 6 16 30
9.6 14.62 26 50
13.65 21.4 34.6 80
OLSR, irrespective of the vehicle’s number. This can be possible when the packet
transmits through shortest path by suitable routing protocol.
Figure 8a, b presents the average delay with respect to the vehicle speed under the
Rayleigh and Nakagami fading channel with DSR routing, respectively. It can be seen
that slow-moving vehicle exhibits a slight increase in average delay, and it slowly
decays with increase in vehicle speed. Further, it is also noticed that as the density of
the traffic increases (30–80 vehicles), delay also increases for both type of fading
channel. Under dense traffic scenario (around 80 vehicles) and slow-moving vehicles
(40 kmph), network suffers for 1 ms average delay under Rayleigh fading channel,
whereas for network utilizing Nakagami fading channel, the delay is of the order of
1.1 ms delay. Therefore, we can conclude that traffic density, vehicle speed and fading
channel have detrimental impact on VANET performance.
(a) Average Delay for routing protocol (b) Average delay for routing protocols
and network size (NetSim) and network size (NetSim)
1.5 1.4
30Car 30Car
50Car 50Car
80Car 1.2 80Car
1.25
Rayleigh Channel model Nakagami Channel model
1
1
Average Delay (ms)
Average Delay (ms)

0.8
0.75
0.6
0.5
0.4
0.25
0.2
0 0
DSR AODV OLSR DSR AODV OLSR
Routing Protocols --> Routing Protocols -->
Fig. 7. Average delay versus routing protocol with different number of vehicles using NetSim
a Rayleigh channel model, b Nakagami channel mode
(a) (b)
Average delay Vs vehicular speed (NetSim) Average delay Vs vehicular speed (NetSim)
1.2
30 Vehicles 30 Vehicle
. 1.2
50 Vehicles 50 Vehicle
1.15 80 Vehicle
80 Vehicles
1.1 1.1 DSR Routing and

DSR Routing and
Nakagami Fading channel
Rayleigh fading channel
1.05
Average delay (ms)
Average delay (ms)
0.9
0.95
0.9 0.8
0.85
0.7
0.8
0.75 0.6
40 50 60 70 80 90 100 110 120 40 50 60 70 80 90 100 110 120
Car velocity (Kmph)--> Car velocity (Kmph)-->
Fig. 8. Average delay versus vehicular speed a DSR routing and Rayleigh fading channel,
b DSR routing and Nakagami fading channel
5 Conclusion
In this work, the performance of vehicular ad hoc network using three existing
topology-based routing protocols was presented and analyzed. Simulation results show
that DSR performs better than AODV and OLSR routing protocol over ideal wireless
as well as for fading channel models. DSR has lower packet loss ratio and approxi-
mately minimum average delay compared to other protocols. The DSR also exhibits
reduced data delivery latency. Though, VANET would provide better platform and
effective communication between vehicles, but under higher vehicle density and higher
mobility condition, the network QoS parameter suffers due to network congestion and
bandwidth wastage.
Acknowledgements. This work was carried out in the FIST supported Optical Engineering
Laboratory of SES, IIT Bhubaneswar.
References
1. Cuomo F, Rubin I, Baiocchi A, Salvo P (2014) Enhanced VANET broadcast throughput
capacity via a dynamic backbone architecture. Ad-hoc Netw 21:42–59
2. Omar H, Zhuang W, Li L (2013) Ve-MAC: a TDMA-based MAC protocol for reliable
broadcast in VANETs. IEEE Trans Mob Comput 12(9):1724–1736
3. Lin YY, Rubin I (2015) Vehicular and messaging throughput tradeoffs in autonomous
highway systems. IEEE Global Commun Conf (GLOBECOM) 1–6
4. Al-Bayatti AH, Zedan H (2014) A comprehensive survey on vehicular Ad-Hoc network.
J Netw Comput Appl 37:380–392
5. Awang A, Husain K, Kamel N, Aissa S (2017) Routing in vehicular Ad-Hoc networks: a
survey on single- and cross-layer design techniques, and perspectives. IEEE Access 5:9497–
9517
6. Uzcátegui R, Marum GA (2009) WAVE: a tutorial. IEEE Commun Mag 47(5):126–133
7. Sandalidis HG, Tsiftsis TA, Karagiannidis GK (2009) Optical wireless communications with
heterodyne detection over turbulence channels with pointing errors. J Lightw Technol 27
(20):4440–4445
8. La D, Fuente MG, Labiod H (2007) Performance analysis of position-based routing
approaches in VANETs. In: 9th IFIP international conference on mobile wireless
Communications Networks, pp 16–20
9. Paul B, Ibrahim M, Bikas MAN (2011) Experimental analysis of AODV & DSR over TCP
and CBR connections with varying speed and node density in VANET. Int J Comp App
24(4)
10. Shakya M, Markam K (2016) Comparison of AODV and DSR routing protocols by using
group mobility model with varying pause time in vehicular Ad-Hoc network. Int J Hybrid
Info Technol 9(2):177–182
11. Chandel N, Gupta V (2014) Comparative analysis of AODV, DSR and DSDV routing
protocols for VANET City Scenario. Int J Recent and Innovation Trends in Comp Commun
2(6):1380–1384
12. Piorkowski M, Raya M, Lugo AL, Papadimitratos P, Grossglauser M, Hubaux JP (2008)
TraNS: realistic joint traffic and network simulator for VANETs. ACM Mob Comput
Commun Rev 12(1):31–33
13. Oliveria R, Montez C, Boukerche A, Wangham MS (2017) Reliable data dissemination

protocol for vanet traffic safety. Ad-hoc Netw 63:30–44
14. Malik S, Sahu PK (2019) A comparative study on routing protocols for VANETs. Heliyon,
5(8)
Distributed Estimation of IIR System’s
Parameters in Sensor Network by Multihop
Diffusion LMS Algorithm
Meera Dash1(&), Trilochan Panigrahi1, and Renu Sharma2

1
Department of Electronics and Communication Engineering,
SOA (Deemed to be) University, Bhubaneswar, India
meeradash@soa.ac.in
2
Department of Electrical Engineering, Institute of Technical
Education and Research, SOA (Deemed to be) University, Bhubaneswar, India
renusharma@soa.ac.in
1 Introduction
Wireless sensor networks (WSNs) are comprised of thousands of tiny wireless sensor
nodes which are distributed in an area for specific remote sensing activities [1]. Every
node then estimates the desired parameters out of the observed data of its own and also
depends on the data from all the wireless sensor nodes of the network. For this, each
sensor node after sensing the environment processes the data and then sends the data to
its predefined neighbors. To improve the desired parameter estimation at each node,
distributed algorithm plays an important role with information exchange between them
[2]. But, in less connected or sparse wireless sensor networks, there is a difficulty in
accessing information from the sensor network for estimating the parameters [3, 4].
FIR system’s estimation of the parameter using a distributed strategy is discussed in
literature [5, 6]. Mostly infinite impulse response (IIR) modeling is applied in cases of
naturally occurring systems in which the unknown system has feedback. Thus, having
reasonable length systems using FIR filter may not have an acceptable model of the
feedback system [7]. Therefore, an IIR adaptive filter is used to model an unknown
system.
But, the disadvantage of the IIR filter is, it is prone to have non-quadratic and multi-
modal error surfaces with regard to the coefficients of the filter [8]. However, the
general IIR LMS solution has been proved useful, and the local minima problem is less
severe and the IIR LMS algorithm extensively improves after incorporating themselves
in the censor network. Thus, the objective of the paper is to approximate the parameters
of the IIR system in sparse sensor network by employing diffusion cooperation. Fur-
ther, for sparse network where the nodes are connected to less number of neighbors,
multihop diffusion is proposed in literature [4, 9]. The number of hops to and fro
between the nodes impacts the overall network performance [10]. The algorithm is
simulated in MATLAB, and results are compared with and without multihop strategy
for standard IIR system.

https://doi.org/10.1007/978-981-15-2774-6_3
Distributed Estimation of IIR System’s Parameters … 21
2 Formulation of the Estimation Problem
A sensor node with k nodes is distributed in the area of interest to estimate system
parameters. The network topology is assumed as an undirected graph.
Let S ¼ fs1 ; s2 ; . . .; sK g be the number of nodes in the network. The node sk is
denoted as k for compact notation. Let an undirected graph G be defined by a set of
sensor nodes as vertices V and a set of edges E. Sensor nodes j and k are called
neighbors if they are connected by an edge. The links are assumed to be bidirectional.
The neighborhood of k is defined by
Nk ¼ fjjðj; kÞ 2 Eg ð1Þ
Each node is modeled as IIR system. Figure 1 shows the block diagram of the
system identification at each node and every node, k 2 S is in association with an
output dk ðiÞ, which is noisy, to xk ðiÞ, which is the input at time instant i.
Fig. 1. Block diagram of IIR system at kth node in a WSN
Noisy observations are assumed to be independent to space and time and to input
and output observations too. Each sensor k has a set of measured data dk ðiÞ to its
corresponding xk ðiÞ, i.e., the input vector at time instant i [11]. The IIR filter system is
modeled having N number of feedback and M number of feed-forward parameters,
respectively. The relationship between the output and input measured values are given
by equation as [8]
X
N X
M
dk ð i Þ ¼ an dk ð i nÞ þ bm xk ði mÞ þ vk ðiÞ ð2Þ
n¼1 m¼0
22 M. Dash et al.
where fan gNn¼1 and fbm gM

m¼0 are the poles and zeroes of the IIR system, and the noise
measurement is denoted by vk ðiÞ at the ith instant for the kth sensor with variance r2v;k
and mean as zero. Noise in the system is supposed as white Gaussian which is not
correlated to the input data xk ðiÞ. In compact form, the equation can be written as
h T T
i
dk ðiÞ ¼ aT dki1 þ bT xik þ vk ðiÞ ¼ dki1 xik w ð3Þ
where
a ¼ ½a1 ; a2 ; . . .; aN T ;
b ¼ ½b0 ; b1 ; . . .; bM T ; dki1 ¼ ½dk ði 1Þ; dk ði 2Þ; . . .dk ði N ÞT

and wT ¼ aT bT :
Modeling of the adaptive system is done as IIR type with an equal number of
parameters, respectively. Each sensor node estimates the output of the adaptive system
given by as
X
N X
M
yk ð i Þ ¼ âk;n ðiÞyk ði nÞ þ ^bk;m ðiÞxk ði mÞ ð4Þ
n¼1 m¼0
where IIR system output of the IIR adaptive system is represented by yk ðiÞ at the kth
node. The current estimate of the output is dependent on N outputs previously esti-
N M
mated. The estimated coefficients âk;n ðiÞ n¼1 and ^ bk;m ðiÞ m¼0 are at each sensor
node k, such that during training they are adjusted to achieve steady-state value after
the training is done. Equation (4) in compact form is given as
yk ðiÞ ¼ âTk yi1 ^T i

k þ bk x k ¼ w k uk ð5Þ
T i
T T T
where âk ¼ âk;1 ; âk;2; ; . . .; âk;N , ^bk ¼ ^bk;0 ; ^bk;1 ; . . .; ^
bk;M , and wTk ¼ ^
ak .
The input data vector is defined as
uik ¼ ½yk ði 1Þ; . . .; yk ði N Þ; xk ðiÞ; xk ði 1Þ; . . .; xk ði M Þ ð6Þ

Global parameter vector w is estimated using data dk ðiÞ; xik available at every
node using the diffusion mode of cooperation method [12] which is the objective of this
paper.
Diffusion LMS (DLMS) algorithm is used for approximation of the IIR coefficients.
By optimizing the following global cost function w can be estimated [13],
X
K 2
JðwÞ ¼ Edk ðiÞ uik ð7Þ
k¼1
2
where Ekk indicates the expectation operator and Jk ðwÞ ¼ E dk ðiÞ uik w indicates
the local cost function.
The optimal value is calculated by solving Eq. (7), [11].
X
K 2
^ ¼ arg min
w E dk ðiÞ uik w ð8Þ
w
k¼1
Desired response, dk ðiÞ, is approximated by using LMS algorithm [7]. To resolve

Eq. (7), the standard steepest-descent recursive algorithm given in (9) is used. To
resolve the difficulty, the equivalent implementation is introduced by using the adapt
then combine (ATC) strategy as in [5]. Therefore, the modified diffusion algorithm is
^ 0k ¼ winit
i ¼ 1; 2; . . .; k ¼ 1; 2; . . .; K; w
ek ðiÞ ¼ dk ðiÞ uik wi1
k
T
w ^ i1
^ ik ¼ w k þ lk ek ðiÞuk
i
ð9Þ
X
k
^ ik ¼
w ^ il
clk w
l2Nk
By following the diffusion cooperation algorithm, each sensor node cooperates

among each other (9) to find the IIR system parameter.
3 Distributed Estimation Using Multihop Diffusion Strategy
To upgrade the diffusion cooperation strategy, many authors have proposed different
methods in the literature. Mostly, many of the authors have tried to minimize the
communication overhead in the network because more energy is required for commu-
nication than computation [6, 14, 15]. But these methods failed to provide the desired
performance in a less connected sparse network. By adopting a partial diffusion [14], the
number of communications is reduced. In [15], energy-aware distributed diffusion
strategy is discussed. These methods are also not suitable for IIR system parameter
estimation problems. It is because the poles play an important role in IIR system and
slight deviation in the value during the learning process may cause instability.
The multihop DLMS is proposed in [3, 9] with constrained energy. As per the
energy of the nodes, the 1-hop neighbors and multihop neighbors are specified. Data
aggregation is a difficult process and takes more time as each sensor node has to find
the information of energy of the neighbors.
Here, we propose a multihop diffusion adapt then combine (ATC) strategy without
constraints. In comparison to combine then adapt (CTA), the ATC is more adaptive.
From each h-hop neighbors, every node can obtain the information. The h-hop sensor
nodes are defined as a set of nodes available at h-hop distance from itself. The 1-hop
network topology and 2-hop network topology are described in [3].
24 M. Dash et al.
The update equations for h-hop ATC (hATC) [3] are following the same steps as
that of one-hop strategy as in (9),
^ 0k ¼ winit
i ¼ 1; 2; . . .; k ¼ 1; 2; . . .; K; w
T
ek ðiÞ ¼ dk ðiÞ uik wi1
k w ^ i1
^ ik ¼ w k þ lk ek ðiÞuk
i
ð10Þ
X
k
^ ik
w ¼ ^ il
clk w
l2Nkh
where lk denotes the step size at sensor node k, and clk is combination weights
satisfying
X
clk 0; ckl ¼ 1; and clk ¼ 0 if l not belonging to Nkh ð11Þ
l2Nkh
Conventional ATC and hATC varies with each other with respect to the combi-
nation step [6, 16]. But, in distributed signal processing in sensor network, each sensor
is assumed to communicate only with 1-hop neighbors only. So, it is difficult to
combine the estimates from multihop neighbors in. This can be calculated by repeating
1-hop combination times in h-hop diffusion strategy [4]. The multihop algorithm can be
implemented with same communication protocol used in conventional diffusion LMS.
The algorithm also does not require additional storage capacity. But the algorithm
needs ðjNk j 1Þh communication cost in each iteration. This is times more than the
communication cost for DLMS ðjNk j 1Þ: The additional communication requirement
can be reduced by employing block LMS method in multihop diffusion LMS
algorithm.
4 Results Obtained by Simulation
Here, the performance analysis of diffusion LMS algorithm is done with and without
multihop for parameter estimation of IIR system with simulations. For sparse sensor
network, multihop diffusion LMS (hDLMS-IIR) is used. The simulations are done
using shift registers and regressor input. According to the model in (3), the desired data
are generated. For simulation purpose, the standard IIR 2nd, 3rd, and 4th order transfer
function of IIR system systems are considered [17, 18].
1:25 0:25z1
H2 ðzÞ ¼ ð12Þ
1 þ 0:32z1 0:412z2
0:6 0:4z1
H3 ðzÞ ¼ ð13Þ
1 þ 0:2z1 0:51z1 0:1z3
And
1 0:9z1 þ 0:812z2 0:73z3

H4 ðzÞ ¼ ð14Þ
1 þ 0:042z1 þ 0:28z2 0:21z3 þ 0:1z4
Here, we have used IIR adaptive filter of 2nd, 3rd, and 4th order at each sensor
node to model the corresponding orders of IIR system, respectively. The transfer
function for the adaptive IIR system models to estimate their parameters is
^bk0 þ ^bk1 z1

^ 2k ðzÞ ¼
H ð15Þ
1 âk1 z1 âk2 z2
^bk0 þ ^bk1 z1 þ ^

bk2 z2
^ 2k ðzÞ ¼
H ð16Þ
1 âk1 z âk2 z ^
1 2 ak3 z3
^bk0 þ ^bk1 z1 þ ^bk2 z2 þ ^bk3 z3

^ 4k ðzÞ ¼
H ð17Þ
1 âk1 z âk2 z ^
1 2 ak3 z ^
3 ak4 z4
The average over 500 independent experiments is plotted. Performance of diffusion

LMS algorithm in less connected network is discussed here. Let a sparse sensor net-
work with 20 sensor nodes. Network topology is given in Fig. 2. From the figure, it is
evident that the ILMS algorithm is not possible here to estimate the systems’ parameter
using distributed approach, whereas diffusion LMS may apply for the same. But the
edge nodes may not get sufficient information from the network because of less con-
nectivity. Therefore, multihop diffusion LMS algorithm is used here for IIR systems’
parameter estimation (hDLMS-IIR).
Fig. 2. Network topology used for sparse sensor network

26 M. Dash et al.
The parameters of each sensor node of the sensor network are given in Fig. 3.
Number of iterations to simulate for each system is taken 3000. The block size is
chosen as length of the optimum vector that is the sum of the feed-forward and
feedback parameters. For the simulations, block length L has been assumed to be 4, 5, 8
for 2nd, 3rd, and 4th order IIR systems, respectively. The parameter l (step size) is
chosen as 0.075 for DLMS-IIR and non-cooperative algorithms, whereas l is chosen as
0:075 L for BDLMS-IIR and BhDLMS-IIR algorithms to get similar performance.
This is based on the theory for block adaptive algorithm and also valid for distributed
block adaptive algorithms [6, 19].
(a) Noise Power Profile (b) Signal Power at Each Node
(c) Correlation Index αk at Each Node
Fig. 3. Sensor node’s parameters for less connected wireless sensor network
In literature, the performance of distributed block LMS and distributed LMS is

nearly the same if the step size of LMS is chosen L time and the step size for LMS
algorithm is without batch processing [6]. The MSE and MSD curves are plotted by
averaging 500 Monte Carlo simulations for 3000 iterations for 1-hop DLMS (DLMS-
IIR), 2-hop DLMS (hDLMS-IIR) and non-cooperation LMS algorithms, and then
block version of all the algorithms. The simulation results show that the performance of
the hDLMS-IIR provides best performance compared to the other algorithms. The MSE
and MSD simulation results for 2nd, 3rd, and 4th order IIR system are plotted in
Figs. 4, 5, and 6, respectively. From the figure, it is evident that the 2-hop hDLMS
algorithm performs well by providing less MSD and MSE performance with faster
convergence compared to that of 1-hop DLMS algorithm in a sparse WSN. The block
version of both 1-hop and 2-hop algorithms perform similar to that of non-block
algorithms for 2nd order IIR system with less communication overhead, Whereas in the
case of 3rd and 4th order system, 1-hop BDLMS algorithms’ performance fluctuates,
especially the MSD performance.
(a) MSE Performance (b) MSD Performance
Fig. 4. Average MSE and MSD performances by 1-hop DLMS-IIR, 2-hop hDLMS-IIR, non-
cooperative LMS-IIR, 1-hop BDLMS-IIR, 2-hop BhDLMS-IIR, and non-cooperative block
LMS-IIR for 2nd order IIR system
LMS-IIR algorithms for 3rd order IIR system
28 M. Dash et al.
LMS-IIR algorithms for 4th order IIR system
But the 2-hop BhDLMS algorithm gives similar to that of non-block algorithm.
There is an improvement of 5 and 10 dB in MSE and MSD performance of 2-hop
hDLMS and BhDLMS algorithm over 1-hop algorithms, respectively. We suppress IIR
in the name of the algorithm for compactness.
5 Conclusion
Distributed IIR systems’ parameter estimation in sparse WSNs using multihop DLMS
algorithm has been studied in this paper. In the wireless sensor network, every node is
assumed to be an adaptive IIR filter, and each and every node in the network collab-
orates among themselves by using diffusion cooperation method. MSE and MSD are
computed through simulation and compared the non-cooperative LMS algorithm for
benchmark IIR systems. Further to reduce the increase in communication overhead in
the network, block distributed multihop diffusion LMS algorithm is proposed where
each node shares data after processing a block of data. By doing this, the multihop
block diffusion LMS provides the best MSE and MSD performance with minimum
number of iterations.
References
1. Li Y, Thai MTE (2008) Wireless sensor networks and applications. Springer
2. Sayed AH (2014) Adaptive networks. Proc IEEE 102(4):460–497
3. Nayak M, Panigrahi T, Sharma R (2015) Distributed estimation using multi-hop adaptive
diffusion in sparse wireless sensor networks. In: International Conference on Microwave,
Optical and Communication Engineering (ICMOCE), Dec 2015, pp 318–321
4. Kong J-T, Lee J-W, Kim S-E, Shin S, Song W-J (2017) Diffusion LMS algorithms with
multi combination for distributed estimation: formulation and performance analysis. Dig
Signal Proc 71(Supplement C), 117–130
5. Cattivelli FS, Sayed AH (2010) Diffusion LMS strategies for distributed estimation. IEEE
Trans Signal Process 58(3):1035–1048
6. Panigrahi T, Pradhan PM, Panda G, Mulgrew B (2012) Block least-mean square algorithm
over distributed wireless sensor network. J Comput Netw Commun., vol. 2012, Hindawi
Publishing Corporation
7. Shynk JJ (1989) Adaptive IIR filtering. IEEE ASSP Mag 6(2):4–21
8. Agrawal N, Kumar A, Bajaj V, Singh GK (2017) High order stable infinite impulse response
filter design using cuckoo search algorithm. Int J Autom Comput 14(5):589–602
9. Hu W, Tay WP (2015) Multi-hop diffusion LMS for energy-constrained distributed
estimation. IEEE Trans Signal Process 63(15):4022–4036
10. Jerew O, Blackmore K (2014) Estimation of hop count in multi-hop wireless sensor
networks with arbitrary node densitym. Int J Wireless Mob Comput (IJWMC) 7(3):207–216
11. Schizas I, Mateos G, Giannakis G (2009) Distributed LMS for consensus-based in-network
adaptive processing. IEEE Trans Signal Process 57(6):2365–2382
12. Nedic A, Ozdaglar A (2009) Distributed subgradient methods for multi-agent optimization.
IEEE Trans Autom Control 54(1):48–61
13. Khalili, Rastegarnia A (2016) Tracking analysis of augmented complex least mean square
algorithm. Int J Adapt Control Signal Proc 30(1):106–114
14. Arablouei R, Werner S, Huang Y-F, Dogancay K (2014) Distributed least mean-square
estimation with partial diffusion. IEEE Trans Signal Process 62(2):472–484
15. Gharehshiran ON, Krishnamurthy V, Yin G (2013) Distributed energy-aware diffusion least
mean squares: game-theoretic learning. IEEE J Select Top Signal Proc 7(5):821–836
16. Xiao L, Boyd S, Lall S (2005) A scheme for robust distributed sensor fusion based on
average consensus. In: Proceedings of 4th international symposium on information
processing in sensor networks, Loss Angles, CA, pp. 63–70, Apr 2005
17. Scarpiniti M, Comminiello D, Parisi R, Uncini A (2015) Nonlinear system identification
using IIR spline adaptive filters. Signal Process 108:30–35
18. Majhi B, Panda G (2013) Distributed and robust parameter estimation of IIR systems using
incremental particle swarm optimization. Dig Signal Proc 23(4):1303–1313
19. Panda G, Mulgrew B, Cowan CFN (1986) A self-orthogonalizing efficient block adaptive
filter. IEEE Trans Acoust Speech, Signal Process 34(6): 1573–1582, Dec 1986
Implementation of Raspberry Pi for Fault
Detection in Optic Fibre Line
K. P. Swain1(&), S. R. Das1, Sangram Kishore Mohanty2,

and G. Palai1
1
Department of Electronics and Communication Enginerring, GITA,
Bhubaneswar, India
2
Department of ECE, ICE, GITA, Bhubaneswar, India
1 Introduction
Optical fibres are long cables made of glass with two layers called core and cladding
which works on the principle of total internal reflection where the light is propagated
through the length of the cable, reflecting back in the core by enabling the light to pass
through the cross-section area even in the bend regions. But, the signal is attenuated
while travelling for a certain distance which causes weakening of signal strength and
introduces noise in the signal. So repeaters are generally used to amplify the optical
signal and transmit to further stages. But, when there is a fault in the optical fibre, it is
usually not known to the concerned person about the fault details.
There are more than a few works are discussed in the literature for finding the fault
in the optic fibre line. One foremost technology like distributed temperature system
(DTS) is used in [1, 2] for real-time monitoring system. In [1], a cable monitoring
system for underground power system is discussed by DTS technology in real-time
basis and online. Here, DTS technology is applied by comparing the temperature of
conductor and cables in real time, and also a test-bed is erected to ensure the appli-
cability of the system towards the underground system. By this technology, the
operators can forecast the fault so that the issue can be resolved within a very short
span of time. This technology is also used in [2] for real-time temperature monitoring
system which is applied to many fields like power cable monitoring, underground
power system monitoring, power cable monitoring in tunnel, etc.
Optical time-domain reflectometer (OTDR) is used in [3] to identify different kinds
of fault like bend, crack, splice, etc. It also conversed a MATLAB-based GUI to
indicate the fault condition with time. OTDR is interfaced with computer screen to
show every measurement which can be further analysis. OTDR is worked on the
principle of Rayleigh backscattering (RBS) which detects an abrupt shift known as
point defects. In [4], limited perimeter vector matching (LVM) protocol is used to
locate single link failure in the optical network. Here, the optimization technique is
implemented to minimize the fault detection time in a static optical network. A photon-
counting OTDR with linearized Breyman iteration algorithm-based processing unit is
discoursed in [5] to find an optical fault. But in [6], minimum length probe cycle cover
algorithm is realized for fault detection and in which m2-cycle mechanism is found
superior among others. A single-mode fault detection is discussed in [7] by using

https://doi.org/10.1007/978-981-15-2774-6_4
Implementation of Raspberry Pi for Fault Detection … 31
approximation algorithm and GIS-based fibre optic monitoring system is illustrated in

[8], where alarms are used to indicate in case of any fault or degradation of optical
signal is occurred. Microcontroller is used in [9, 10], to observe the received power
from an optical fibre line by using a sensor circuit mostly a photodiode. In [9], message
is sent through GSM technology if any abrupt change is occurred in received voltage,
whereas in [10], a database is maintained by using MySQL to keep record of the
received voltage. Arduino Uno is used along with a LDR to sense the received voltage
in real-time basis and to display the fault status continuously through a LCD. Here, a
database is created by using Apache Web server, perl, MySQL and PHP which are very
popular package for Web application and XAMPP is also used for hosting the server
locally.
In the previous work, on optical fault detection system [10], Arduino was used to
acquire data from the optical fibres and processing was done in a different computer to
detect fault and update the data to the server. But sending data to the server requires an
Internet connection and setting up a computer altogether is a difficult task and some-
times not feasible. But in this work, we propose a system using Raspberry Pi to
implement the fault detection by constantly monitoring the received power through a
photodiode. Raspberry Pi used in this work is a mini-computer having multithreading
and multitasking capability which makes it as a superior processing power than
Arduino and also it can directly host Web server by eliminating the need of separate
computer for maintaining the database.
2 System Overview
See Fig. 1.
Fig. 1. Graphical representation of system overview
2.1 Sensor
As shown in Fig. 2, the sensor circuit consists of a photodiode along with an op-amp
(LM358) which acts as a comparator. At the non-inverting end of the op-amp, a
reference voltage is set by the voltage divider circuit comprising of R1 and R2 of value
20 K and 30 K, respectively, whereas capacitive coupled voltage is available at the
inverting end generated by C2 (100 nF) which is coming from the fibre line through
photodiode. Then, these voltages are compared to give an output to GPIO pin of
32 K. P. Swain et al.
Raspberry Pi. A Zener diode is also connected at the output to maintain 3.3 V for
protection of Pi as the GPIO pins are operated at 3.3 V.
Fig. 2. Circuit diagram of the sensor unit
2.2 Signal Processing Through Raspberry Pi

Raspberry Pi acts as a brain of the system where GPIO pins are exposed for direct
interfacing of variety of sensors and electronic systems. With higher voltages, there is a
high chance of damaging the processor as it works on 3.3v logic level. MySQL
database is installed on the Raspberry Pi to store real-time data. It features Wi-Fi and
Bluetooth which enables to connect with existing Wi-Fi networks with ease. The
interfacing and processing logic is written in python which gives us more computing
power and control to handle and store data simultaneously without any lag. The
complete circuit diagram of the proposed system is shown in Fig. 3 where GPIO 2 pin
is connected to the output pin of op-amp which is used to push the data in case of any
fault is raised.
2.3 Web Server (My SQL Server)

MySQL is database management software which stores data in the form of tables and
has the capability to form relations within tables. It was created by Oracle Corporation
which now powers applications right from low-end to high-end data science. It is a
relational database, which means tables within a database can refer to each other, using
foreign key concept. MySQL database provides users to query data from the tables
through commands known as structured query language or SQL for short. In this
optical fault detection system, MySQL is used to store the fault data for monitoring
Fig. 3. Complete circuit diagram of the proposed system
which is installed on Raspberry Pi. Data from the digital inputs is detected by the GPIO
of the Raspberry Pi and are pushed to MySQL database through python script. The
python script acts as a software interface between the Raspberry Pi GPIOs and data-
base. The script constantly listens for change in signal at the GPIO and as soon as the
signal change is detected, it stores the data along with the timestamp in the database.
3 Result and Discussion
The fault detection system includes a sensing unit, a comparator unit and data pro-
cessing unit, where signal is coming from the optical fibre and fell on the sensing unit,
which converts the optical signal to the electrical form. The electric signals are in the
form of voltages which represent the strength of a signal. To detect a fault, the average
of a signal is taken over a time period and its corresponding voltage is generated. For
benchmarking, samples of the signals with typical strength is recorded and stored for
comparison. The comparator circuit is developed using an op-amp (LM358) with
capacitive coupled input on the inverting terminal of it in which average voltage
required for comparison is provided by capacitor (C2 of 100 nF). The non-inverting
terminal of the comparator is given with a voltage reference through a resistor
arrangement (R1 and R2 of 20 K and 30 K, respectively). When the sensed signal
strength falls below the setpoint, the op-amp sends a high signal which represents a
fault in the system. During this period, timestamp of the event occurred and the fault
status is updated by Raspberry Pi. This data is recorded in the database and notification
is generated for the operator for inspection.
3.1 Physical Realization of the System

The physical implementation is realized in the laboratory by using two optical demo
kits of Scientech 2506 model with a two metre poly methyl methacrylate also known as
acrylic glass (PMMA) fibre cable. Here, the two kits are used to demonstrate optical
sender and receiver through the fibre cable. A sinusoidal signal is generated using a
function generator of Scientech 4061 (3 MHz) model which drives the sender kit for
converting the electrical signals into the optical signals and transmitted through the
optical fibre. The receiver has optical transducer which converts optical signal back to
electrical signals, and both the output of the receiver and sender are shown in the
oscilloscope as shown in Fig. 4. The receiver output is connected to the op-amp which
is configured as a comparator. The output of the op-amp is connected to the GPIO of
the Raspberry Pi where the Pi constantly listens for the interrupt signal from the op-
amp. As soon as a signal is generated, Raspberry Pi stores the signal information along
with the timestamp. The monitor in the image shows MYSQL database with the
records inserted during the fault. To test the system, the fault signals are generated by
reducing the gain of the amplifier at the sending end simulating an attenuated signal.
Fig. 4. Realization of the system in laboratory using Scientech 2506 kit
3.2 Database Apprising

The table shown in Fig. 5 is used to store the status of the signal strength coming from
the fibre line where the columns include timestamp, signal strength and its
corresponding remark. The values in the signal strength represent the fault present or
absent. The lowest value 1 represents the signal has fault and value 3 represents the
system is normally operational condition. To support this, remark column shows the
same and the timestamp column represents the data insertion time.
Fig. 5. Update of database by experimental method
4 Conclusion
An optical fault detecting system using Raspberry Pi is realized successfully in this

paper, which is used to signpost the fault status ensued in the fibre line through a local
server. Here, MySQL is installed in the Raspberry Pi which acts as a local server to
update the fault info in real-time basis by the aid of a comparator containing an op-amp.
Here, op-amp with a photodiode used together to compare the incoming line voltage
through photodiode available at inverting terminal with a reference voltage maintained
at non-inverting input of it. To end with, an experiment is also conducted successfully
in the laboratory by using two optical transceiver kit of model Scientech 2506 to verify
the result in the database by varying the voltage manually at the transmitting end.
References
1. Cho J, Kim J, Lee H, Kim J, Song I, Choi J (2014) Development and improvement of an
intelligent cable monitoring system for underground distribution networks using distributed
temperature sensing. Energies 7:1076–1094
2. Downe J, Leung HY (2004) Distributed temperature sensing worldwide power circuit
monitoring applications. IEEE ICED 1804–1809
3. Ab-Rahman MS, Chuan NB, Premadi A, Jumari K (2008) A new approach for identifying
fiber fault and detecting failure location. IEEE ICED 1–4
4. Khair MG, Kantarci B, Zheng J, Mouftah HT (2008) Optimization for minimizing fault
localization time in all-optical networks. IEEE ICTON 63–66
5. Lunglmayr M, Amaral GC, Linearized B (2018) Iteration for automatic optical fault analysis.
IEEE Trans Instrum Meas 1–13
6. Wang N, Wang R, Huang S (2009) A low cost fault detection mechanism based on cycle
cover algorithm for optical burst switching networks. IEEE ICCMC 545–549
7. Nayek P, Pal S, Choudhury B, Mukherjee A, Saha D, Nasipuri M (2005) Optimal monitor
placement scheme for single fault detection in optical network. IEEE ICTON 433–436
8. Akdemir OK, Dursun T, Arslan S, Benzer R, Akcayol MA (2016) A GIS-based novel active
monitoring system for fiber networks. Turk. J. Electr. Eng. Comput. Sci. 24:247–261
9. Baskaran G, Seethalakshmi R (2012) Intelligent fault detecting system in an optical fibre.
J. Theor. Appl. Inf. Technol. 39:178–187
10. Swain KP, Sahoo J, Prasad MVSV, Palai G (2015) Fault detection system in an optical fiber
using arduino. Int. J. Appl. Eng. Res. 10:31745–31749
Distributed Optical Fiber Sensing System
Performance Improvement Using Signal
Processing Techniques
Ramji Tangudu(&) and Prasant Kumar Sahu
Indian Institute of Technology Bhubaneswar, Argul, Odisha, India

{rt10,pks}@iitbbs.ac.in
1 Introduction
In this paper, the detailed study of Brillouin optical time-domain reflectometry

(BOTDR)-based distributed optical fiber sensing (DOFS) system is reported. Nowa-
days, optical fiber is being utilized for both communication and sensing purposes [1].
DOFS systems are more effective, cost-effective, and popular, due to the feature of
replacing thousands of point sensors and the associated complexity by using single
long-range fiber sensors typically of the orders of few kilometers [1]. The DOFS system
operation is based on backscattered signals, namely Rayleigh, Brillouin, and Raman.
Raman backscattered signal-based DOFS system is useful for temperature sensing
only, whereas Brillouin backscattered signal-based DOFS system is useful for both
temperature and strain sensing. Due to the inherent advantages of Brillouin sensing,
this system has higher potential applications in oil and gas pipeline leakage detection,
fire detection, high power cable monitoring, structural health monitoring (bridges,
dams, tunnels, and roads), as well as in defense and military applications [1]. The major
advantages of fiber-based sensing system include the electromagnetic interference
immunity, lightweight, design simplicity, and more survivability under harsh envi-
ronmental conditions [1]. The variation in temperature along the fiber will result in
fluctuations in frequency of Brillouin backscattered signal. From these parametric
fluctuations, we can estimate the actual temperature along the fiber. Typically, the value
of Brillouin frequency shift (BFS) for 1550 nm pumping wavelength is *11 GHz [1].
Brillouin backscattered signal-based DOFS system can further be classified as
distributed temperature sensing (DTS) system and distributed strain sensing
(DSS) systems. In this paper, we have focused on DTS system only. According to
literature survey, Brillouin backscattered signal-based DTS (BDTS) system is broadly
classified into two categories, such as spontaneous Brillouin scattering (SpBS) and
stimulated Brillouin scattering (SBS)-based BDTS system. Under SpBS type, BOTDR,
Brillouin optical correlation domain reflectometry (BOCDR) and Brillouin optical
frequency-domain reflectometry (BOFDR)-based DTS systems are grouped together
[2–4], whereas under SBS type, Brillouin optical time-domain analysis (BOTDA),
Brillouin optical correlation domain analysis (BOCDA) and Brillouin optical
frequency-domain analysis (BOFDA)-based DTS systems are grouped together [5–7].
In all analysis based DTS systems which requires the double ended accesses (probe and

https://doi.org/10.1007/978-981-15-2774-6_5
38 R. Tangudu and P. K. Sahu
pump lights) and stimulated Brillouin scattering are complex to design and costly
[5–7]. BOCDR-based DTS system is useful for short-range applications typically for
meter-long sensing only, whereas BOFDR-based DTS system is having high cost, due
to usage of electrical spectrum analyzer (ESA) and vector network analyzer (VNA) [4].
Song and Zhao [8] proposed a BOTDR system using Hilbert transform method for
sensing the temperature for a range up to 25 km with an input optical launch power of
22 dBm of peak power. Bernini et al. [9] proposed a combined scheme of BOTDR
technique and harmonic approach for obtaining 40 cm spatial resolution in temperature
sensing. The major feature of their proposed system is as follows: 1 mW of laser source
power over a sensing range of 35 m only. Soto and Sahu [2] have designed a dis-
tributed temperature sensing system using BOTDR and 127-bit simplex coding scheme
for 21 km of sensing range and the reported temperature resolution was of 3.1 K.
However, this method suffers from the limitations of design complexity and bandwidth
requirement when the coding length increases. In [10], combination of BOTDA
technique and optimized pulse using 127-bit simplex coding is proposed for getting
2.6 °C temperature resolution as well as Raman amplification being used for getting
120 km of sensing range in the presence of more relative intensity noise (RIN).
However, since the system is designed on the analysis-based technique, coding scheme
and amplification process, complexity, bandwidth requirement, and cost will be
more [10].
Minardo et al. [7] presented BOFDA technique-based system for temperature
sensing using polymer optical fiber (POF). However, the sensing ranging is limited to
20 m only. Pradhan and Sahu [11] proposed a BOTDR system using Fourier wavelet
regularized deconvolution (FourWaRD) scheme for achieving a 1.85 K temperature
resolution and 30 m spatial resolution using 400 ns laser pulse width. The system was
designed using evolutionary optimization schemes for achieving better SNR value.
Though the designed system offered a sensing range up to 50 km of fiber length, but the
spatial resolution was 30 m, which needs improvement. Li et al. [5] proposed a dif-
ferential pulse pair (DPP)-BOTDA technique for achieving 2 m of spatial resolution.
Here, they have used two pulse signals, the respective pulse widths are 80 and 100 ns,
respectively, and they have achieved a 40 km of fiber sensing length. This scheme was
based on SBS waveform. The major drawbacks of this scheme include higher degree of
system complexity and suffers from nonlinearity effects. Shlomi et al. [6] proposed a
system for temperature sensing operation with DPP-BOCDA technique, and the
achieved spatial resolution was 2.7 cm with a 1.9 MHz of BFS error. However, the
major drawbacks of this technique include higher system complexity and lower sensing
range. The sensing range was 43 m only.
Bao et al. [12] have presented a work on analysis of temperature and strain sensing.
The temperature and strain sensing was carried out from 8 °C to 50 °C (with Panda
fiber), from 0 °C to 43 °C (with Bow-Tie fiber), from 0 le to 408 le (with Panda
fiber), and from 0 le to 1021 le (with Tiger7 fiber), respectively. Because of using
polarization maintaining fiber (PMF), the polarization mode dispersion (PMD) effect
was removed. The system features a 20 cm spatial resolution with 2 ns pulse width and
PMF. The system offered achieved 15 cm of spatial resolution with pulse width of
1.5 ns and photonic crystal fiber (PCF). Employing PCF, they have obtained tem-
perature and strain resolutions of 1.3 °C and 15 le, respectively. From PMF, they
Distributed Optical Fiber Sensing System … 39
have obtained temperature and strain resolutions varied from 2 °C to 38 °C and varied
from 39 le to 598 le, respectively. These resolutions are obtained for 2 m of sensing
fiber only [12].
Soto et al. [13] presented a BOTDR system using 127-bit simplex coding tech-
nique. The system was designed for temperature and strain sensing over 53 km of
measurement range with 10 mW of laser source power. The reported SNR enhance-
ment was 7.4 dB with 127-bit simplex coding technique. In this work, the temperature
resolution is achieved as 8.8 K and strain resolution is achieved as 220 le. They have
achieved spatial resolution as 32 m using 320 ns of laser source pulse width [13]. Zou
et al. [14] proposed BOCDR technique with PMF. The system offered spatial resolu-
tion of 5 mm, 5 cm, and 16 cm using 18.525 MHz, 1.785 MHz, and 528.968 kHz of
modulation frequency. However, the sensing range was limited to 200, 56, and 5 m
only. The BFS error reported was ±3 MHz for 200 m and ±5 MHz for 56 m of
sensing range. For 200 m sensing range, temperature and strain resolutions are ±3 °C
and le, respectively. For 56 m sensing range, temperature and strain resolutions are
±5 °C and le [14].
Bernini et al. [15] proposed a BOFDA system, using peak search, inverse Fourier
transform (IFT), and iterative procedures. This system has been offered a 29 mm of
spatial resolution and has low spurious effects on the collected data (with iterative
algorithm). The FUT length was up to *5.5 m with 40 mW of laser input power.
Yamashita et al. [16] proposed a combination of BOCDA and temporal gating tech-
nique. They have obtained 5 cm of spatial resolution and almost zero error in strain
sensing. Due to the usage of temporal gating technique, their FUT length was obtained
as 500 m (here fiber is PMF). The amount of error in temperature sensing was 10 °C.
Hayashi et al. [3] proposed BOCDR system for sensing temperature and strain. The
system was designed based on per fluorinated graded index—polymer optical fiber
(PFGI-POF) and using heterodyne detection scheme. For 2 m of POF length, 34 cm
spatial resolution was obtained with 11.698 MHz modulation frequency, and for 1.3 m
POF length, 7.4 cm spatial resolution was obtained with 53.451 MHz modulation
frequency. The system offered a 1.3 m of FUT length for temperature sensing and 2 m
of FUT length for strain sensing. Since PFGI-POF fiber has 250 dB/km of attenuation
loss, the system required 27 dBm of laser input power. The achieved temperature and
strain resolution were ±3 °C and ±0.09%, respectively [3]. Wada et al. [17] proposed
a system based on OFDR technique and using PMF-PANDA. The system has been
offered 0.9 °C of temperature resolution and 7.5 le of strain resolution. Since the
spatial resolution was 1 mm, the length of FUT is restricted to 10 m only [17].
In order to minimize the specified drawbacks of BOTDR-based DTS system, we
have proposed two signal processing techniques, such as empirical mode decomposi-
tion (EMD) and lifting wavelet transform-optimum modified differential threshold
(LWT-OMDT). The purposes behind the usage of proposed (EMD) [18] signal pro-
cessing technique are: It decomposes any non-stationary and nonlinear signals. It gives
fast response. It contains an adaptive nature. This technique operates at the scale of one
oscillation; it means high locality is available. It offers multiresolution property [18].
The purposes behind the usage of proposed (LWT-OMDT) signal processing technique
are: Lifting scheme [19] does not require any complex mathematical calculation. This
technique is very easy to implement; it gives the quick response, and it does not depend
on Fourier transform. The modified differential thresholding (MDT) function produces

differentiable and continuous segmentation at all points of a signal. In order to get
optimum solution, we have optimized the MDT function using the modified particle
swarm optimization (MPSO) [19]. The MPSO algorithm majorly depends on a velocity
parameter only. MPSO algorithm contains very few number of control parameters,
better convergence, and simplest implementation for linear and nonlinear signals.
Unlike PSO algorithm, it consists of adaptive nature and rejects local optimum, due to
the weight factor variation for every iteration [19].
2 Theoretical Background
The considered BOTDR setup is illustrated in Fig. 1. In Fig. 1, the laser diode gen-
erates a series of optical pulses and these pulses are injected into an optical fiber. These
pulses are passed through a 3-port circulator as shown in the figure below. This
circulator’s second port is connected with 50 km of FUT or sensing fiber. When the
optical signal propagates through the FUT, the spontaneous Brillouin backscattered
signal will be generated along an entire FUT due to an interaction between the incident
optical pulses and the acoustic phonons in the sensing fiber. The frequency of this
Brillouin backscattered signal is different from an incident optical signal’s frequency.
This frequency difference is called as BFS.
1 2
Optical Source Cleaved
(Laser Diode ) Circulator
End
3
Sensing Range or FUT of 50 km
Pulse Generator
Avalanche Photo Digital Storage Personal

Diode Oscilloscope Computer
Fig. 1. BOTDR schematic setup
The Brillouin backscattered signal from the FUT will be transmitted toward the
circulator, which then directs the backscattered signal toward the avalanche photodiode
(APD) equipped with an amplifier. Since the APD device is used to convert an optical
signal into an electrical signal with gain, this electrical signal is displayed in time
domain with the help of a digital storage oscilloscope (DSO) and finally it will be
processed using a personal computer. In Fig. 1, the FUT under consideration is sub-
jected to temperature variation. The structure of FUT for temperature sensing is shown
in Fig. 2.
Heating Heating Heating

Unit 1 Unit 2 Unit 3
(HU 1) (HU 2) (HU 3)
1 2
Circulator
15.00 km to 25.50 km to 44.44 km to
3 16.00 km 32.50 km 48.88 km
Fiber under test (FUT) length of 50 km
Fig. 2. Structure of silica optical fiber (SOF) under test for temperature
In Fig. 1, the FUT under consideration is subjected to temperature variations. The

structure of FUT for temperature sensing is shown in Fig. 2. Brillouin frequency shift
#B ðT Þ is produced by the applied temperature on the sensing fiber or FUT. It can be
written as [1]:
#B ðT Þ ¼ #B ðTr Þ þ KT# ðT Tr Þ ð1Þ
KT# parameter denotes temperature coefficient of BFS in a silica single mode optical
fiber. T and Tr parameters denote non-room and room temperatures. For silica single
mode optical fiber, the value of KT# is 1.07 MHz/°C.
The mathematical background for the proposed (EMD) signal processing technique
is available in [18]. The mathematical background for the proposed (LWT-OMDT)
signal processing technique is available in our previously contributed work [19]. In the
proposed (EMD) signal processing technique, the total number of levels, 0 = 4. It
means the four number of empirical mode decompositions are required. Thus, the four
number of an intrinsic mode functions (IMF) are required. In the proposed (LWT-
OMDT) signal processing technique, the MPSO evolutionary algorithm is applied to
each level of LWT scheme. Here, the total number of levels, 0 = 3. It means the three
number of lifting wavelet decompositions are required.
3 Simulation Results and Discussions
In this paper, we have presented a BOTDR-based DTS system. The system response is
simulated using MATLAB. These results are presented for the proposed (EMD and
LWT-OMDT) and conventional (discrete wavelet transform (DWT)-soft threshold and
DWT-hard threshold) technique-based systems.
The simulation parameters of these systems are: optical launch power of 10 mW,
sensing fiber length 50 km, laser source pulse width of 400 ns, refractive index of silica
fiber core as 1.48, and pumping wavelength as 1550 nm. We have considered three
different temperatures or events generated by different heating units (HU) on the
sensing fiber for BOTDR-based DTS system. The locations of these heating units are
shown in Fig. 2. The temperatures from these HUs are 37.77 °C, 48.88 °C, and
66.66 °C, respectively, and the rest part of the sensing fiber is under room temperature.
In order to minimize the temperature error in BOTDR-based DTS system, we have
implemented conventional (DWT-hard threshold and DWT-soft threshold) [20] and
our proposed (EMD and LWT-OMDT) signal processing techniques.
By the conventional DWT-soft threshold technique, we have obtained an
improvement in temperature extraction over other conventional DWT-hard threshold
[20] technique as reported in Figs. 3 and 4. Though the system response has improved,
scope is there for further improvement.
80
With the conventional (DWT-Hard Threshold) Signal Processing Technique
True Value
70
60
Temperature( ° C)
50
40
30
20
0 5 10 15 20 25 30 35 40 45 50
Sensing Length (km)
Fig. 3. Extraction of temperature from BOTDR-DWT-hard threshold-based DTS system
In our proposed EMD signal processing technique, IMF has played a major role to
produce minimum error in extraction of temperature. From Fig. 5, it is clear that the
proposed BOTDR-EMD-based DTS system offers higher degree of temperature reso-
lution than the conventional BOTDR-DWT-hard threshold and BOTDR-DWT-soft
threshold based DTS systems. The EMD-based technique resulted in *1.14 °C of an
average temperature resolution value for BOTDR-based DTS system over a 50 km of
sensing range. However, the proposed LWT-OMDT signal processing technique pro-
duced minimum error in temperature extraction using an OMDT function and the
observed an average temperature resolution was of *0.80 °C as shown in Fig. 6.
The OMDT function value was obtained using MPSO algorithm.
80
With the conventional (DWT-Soft Threshold) Signal Processing Technique
True Value
70
60
Temperature( ° C)
50
40
30
20
0 5 10 15 20 25 30 35 40 45 50
Sensing Length (km)
Fig. 4. Extraction of temperature from BOTDR-DWT-soft threshold-based DTS system
70
With the first Proposed (EMD) Signal Processing Technique
65 True Value
60
55
Temperature(°C)
50
45
40
35
30
25
20
0 5 10 15 20 25 30 35 40 45 50
Sensing Length (km)
Fig. 5. Extraction of temperature from our proposed (BOTDR-EMD) DTS system

80
With the second Proposed (LWT-OMDT) Signal Processing Technique
True Value
70
Temperature( ° C)
60
50
40
30
20
0 5 10 15 20 25 30 35 40 45 50
Sensing Length (km)
Fig. 6. Extraction of temperature from our proposed (BOTDR-LWT-OMDT) DTS system
4 Conclusion
In conclusion, the proposed DOFS system has many potential applications ranging
from biomedical engineering to oil and gas pipeline monitoring, high power cable
monitoring, fire detection, to defense and aviation industries. The proposed BOTDR-
EMD-based DTS system offers an average temperature resolution of *1.14 °C,
whereas the BOTDR-LWT-OMDT-based DTS system offered a temperature resolution
of *0.80 °C.
Acknowledgements. This work was carried out in the FIST supported optical engineering
laboratory of School of Electrical Sciences (SES), Indian Institute of Technology Bhubaneswar.
References
1. Lee B (2003) Review of the present status of optical fiber sensors. J. Opti Fiber Technol
9:57–79
2. Soto MA, Sahu PK, Bolognini G, Pasquale FD (2008) Brillouin-based distributed
temperature sensor employing pulse coding. J Sens 8:225–226
3. Hayashi N, Mizuno Y, Nakamura K (2014) Distributed Brillouin sensing with centimeter-
order spatial resolution in Polymer optical fibers. J Lightw Technol 32:399–4003
4. Lombera RR, Minardo A, Bernini R, Mirapeix J, Higuera JML, Zeni L (2017) Experimental
demonstration of a Brillouin optical frequency-domain reflectometry (BOFDR) sensor. In:
Proceedings of the SPIE International conference on optical fiber sensors (SPIE OFS ‘25),
pp 1–4
5. Li Z, Yan L, Shao L, Pan W, Luo B, Liang J, He H (2016) Coherent BOTDA sensor with
single- sideband modulated probe light. J Photon 8:13804–13810
6. Shlomi O, Preter E, Dexin B, London Y, Antman Y, Zadok A (2016) Double-pulse pair

Brillouin optical correlation domain analysis. J Op Exp 24:26867–26876
7. Minard OA, Bernini R, Zeni L (2014) Distributed temperature sensing in polymer optical
fiber by BOFDA. Photon Technol Lett 26:387–390
8. Song MP, Zhao B (2005) Accuracy enhancement in Brillouin scattering distributed
temperature sensor based on Hilbert transform. J Opt Commun 250:252–257
9. Bernini R, Minardo A, Zeni L (2006) An accurate high-resolution technique for distributed
sensing based on frequency-domain Brillouin scattering. Photon Technol Lett 18:280–282
10. Taki M, Soto MA, Pasquale FD, Bolognini G (2011) Long-range BOTDA sensing using
optical pulse coding and single source bi-directional distributed Raman amplification. In:
Proceedings of the IEEE Sens (IEEE Sensors ‘10). 6127160 (2011) 382–385
11. Pradhan HS, Sahu PK (2014) High-performance Brillouin distributed temperature sensor
using Fourier wavelet regularized deconvolution algorithm. J Optoelectron 8:203–209
12. Bao X, Zou L, Yu Q, Chen L (2005) Development and applications of the distributed
temperature and strain sensors based on Brillouin scattering. In: Proceedings of the IEEE
Sens. (IEEE Sensors ‘04) 1210–1213
13. Soto MA, Bolognini G, Pasquale FD (2007) Enhanced simultaneous distributed strain and
temperature fiber sensor employing spontaneous Brillouin scattering and optical pulse
coding. Photon Technol Lett 21:450–452
14. Zou W, He Z, Hotate K (2010) Single-end-access correlation-domain distributed fiber-optic
sensor based on stimulated Brillouin scattering. J Lightw Technol 28:2736–2742
15. Belal M, Newson TP (2010) Enhanced performance of a temperature-compensated sub
meter spatial resolution distributed strain sensor. Photon Technol Lett 22:1705–1707
16. Bernini R, Minardo A, Zeni L (2012) Distributed sensing at centimeter-scale spatial
resolution by BOFDA: measurements and signal processing. J Photon 4:48–56
17. Yamashita RK, Zou W, He Z, Hotate K (2012) Measurement range elongation based on
temporal gating in Brillouin optical correlation domain distributed simultaneous sensing of
strain and temperature. Photon Technol Lett 24:1006–1008
18. Maheshwari S, Kumar A (2014) Empirical mode decomposition: theory and applications.
J Electron and Elect Eng 7:873–878
19. Tangudu R, Sahu PK (2018) Strain resolution enhancement in Rayleigh-OTDR based DSS
system using LWT-MPSO scheme. J Optik 176:102–119
20. Huimin C, Ruimei Z, Yanli H (2012) Improved threshold denoising method based on
wavelet transform. In: Proceedings of the Elsevier international conference on medical
physics and biomedical engineering (ICMPBE), pp 1354–1359
A Brief Understanding of IOT Health Care
Service Model Over Remotely Cloud
Connected Environment
Subhasish Mohapatra1 and Smita Parija2(&)

1
Department of Computerscience Engineering, Adamas University, Kolkata,
India
2
Department of Eletronics, CV Raman College of Engineering, Odisha, India
1 Introduction
Human body depicts a temple so we must take care of it properly. It is the vital
backbone for any developing nation like India. The most hopeful and awaited smart
city project health care is the vital aspect so the Government can’t ignore this cutting
edge application for underprivileged person. There are patient go in a parallel way. It
can provide tangible benefit to patient by accurate, reliable, and in a timely manner,
consequently it imparts healthcare service provider in a mode of reliability many
application-based technique are available for citizen health access parameter [1].
Here, the patient-specific information is passed by a sensor to gateway which is
connected over cloud. In this model, we can take advantage of public cloud which can
Send patient data to cloud data store after accessing the internet. Here web is
adhering to hoisting the health care application project where a patient can fill out the
necessary credential. A doctor can verify patient data securely and give his valuable
feedback over cloud. Here a citizen can collaborate access his claim from the company
database and bank for transferring a bill payment. The propound model in this paper
provide abstraction of citizen information and flexibility of cloud data processing
Sensed by IOT mechanism [2, 3].
Section 2 gives IOT application for hospitals. Section 3 discuss the motivation of
the work; Sect. 4 gives proposed System architecture, Sect. 5 explains the working
model and Sect. 6 presents the conclusion and future scope,
2 IOT for Hospitals
Here citizens can swipe the card which is supplied by unique Id authority the feature of
Unique Id for this cloud-based access is given in Fig. 2 Sect. 3 of this paper. Subse-
quently, an OTP is generated for the registered mobile no of the person credential is
send over the public kiosk of the hospital and a person can check in his detail. The
sensed data is transmitted over ZigBee network to intrahospital-based cloud. In the

https://doi.org/10.1007/978-981-15-2774-6_6
A Brief Understanding of IOT Health Care Service … 47
proposed method, ZigBee network is used as it is a low power, low cost, and long
battery life device [4–7].
For remote accessing, the sensed data can be viewed to web module. The benefit of
cloud-based model enhances documentation and quick dissemination of valuable
information to citizens, and secure service-centric data delivery as such claim settle-
ment with the bank. It also helps research group to study the patient history data for
further research in health sector without going pillar to post. For grasping the sweet of
this effort, every citizen can go for this ID card and every hospital including CHC, PHC
present at nook and corner of the country must supply with this sensed bar code reader
and ZigBee network to hoist application over cloud so this ambitious model will be
successful in future [8, 9].
3 Motivation
Now, IOT is transforming the future of every technology and has created a lot of
envision in a wide coverage, it enables more smart services that provide huge amount
of on-demand access to a shared pool of configurable resources, e.g., computer, cloud,
device, software, etc. So in this paper, we illustrate barcode sensing in healthcare
industry to hold massive amount of particular data at any place around us in a user-
friendly manner.
4 Proposed System Architecture
The underlying backbone of this model is cloud connection and sensing of data by
barcode reader and sending it over public kiosk cloud is depicted as below Fig. 1. The
most promising application of this model is it can provide data level security at entry
point and passing through firewall for hoisting the data over cloud sub-module data
store [10, 11]. Before to this validation of user credential is done by OTP (one time
password)-based mechanism is projected ass below Fig. 1.
5 Working Model
Figure 1 elaborates the cascading flow of this model how a citizen avail services
through Unique id with barcode, which will serve as an elementary IOT-based facil-
itation of quick service delivery over cloud [12]. Electronic services like checking of
claim status, banking transaction for payment of the bill, Research group can check the
medical history database, and viewing of an item to transact over cloud to citizen. Here,
the Fig. 1 diagram depicts Citizen2Health interface2cloud is illustrated as below:
48 S. Mohapatra and S. Parija
Unique Id
with bar
code
OTP
GENERATED
IF
OTP IS TIME OUT
SUCCESS NO
YES
Enter Public Kiosk
ENTER user id
& PASSWORD
NO
IF
PASSWORD STOP
IS
CORRECT
YES
PATIENT
HISTORY
Health service
interface kiosk
REGISTER
SERVICE
HOSPITAL
OFFERED
DATA
1.PATIENT CENTER
HISTORY
E BANK
F
N
2.REGISTRED I
C
HOSPITAL R
R Load
E
Y
3.BANK W balan CLOUD
P cer CLAIM
A COMPANY
T
4.CLAIM L
I
COMPANY L
O
N
5.RESERCH RESEARCH
GROUP GROUP
6.VIEW
ITEM
Fig. 1. Barcode sensing healthcare service application model for cloud-connected infrastructure
Service side of Citizen
Step-1-citizen trigger out the transaction by using an ID with barcode supplied by

certain unique identification authority.
Step1.1-citizen provide a unique set of attributes to Health interface for a sign of
verification. This has been discussed below.
Step1.2-Here citizen credential is checked by health interface and further chan-
nelization of the service to C-Health interface.
Service side of Public kiosk
Step 2-Public kiosk verifies the user input for authentication which is the Current
environment of the hospital sector connected to cloud over a gateway.
Scenario1-For the successful validation the transaction proceeds further to step3.
Scenario2-In case of unsuccessful validation the transaction is aborted. The system
timeout or negations of validation success, i.e., negative acknowledgement give a
status of failure transaction to citizen.
Step3-Transaction of further service request is processed for health interface that
deliberate the vita aspect of this paper here we have channelized the connected
service for the ease of citizen service. In this step, Citizen2Health interface2cloud
service delivery notifies the citizen and doctor about this model.
Figure 2 project the Unique ID card issued by unique identification authority to
citizen for accessing medical service and other vital aspects from the hospital sector.
Fig. 2. Pictorial representation of Unique id with barcode sensing feature
• Citizen ID-combination of character and integer special character type

• Name of citizen
• Date of birth of a citizen
• Date of issue
• Photo and signature
50 S. Mohapatra and S. Parija
This thing will be shown as soon as citizen touch this Id card in a barcode reader
machine that will be connected to cloud over gateway as soon as the credential is
checked it redirected to cloud by a gateway and OTP is generated for the request for a
secure transaction purpose.
And a person can login to health interface service kiosk.
After the citizen chooses the menu for service he desired for and views the
transaction over the cloud.
For security here we inculcate two-way data security paradigm as patient data grab
high attention in black market so input side citizen data-centric encryption and service
filtration on output side is the beauty of this model which enhances further data
security.
One submodule research group is added over for further research upon electronic
health record data that is stored in submodule patient history module.
The proposed system also has taken the optimistic approach of load balancing to
further protect the cloud data store and segregate the load among the submodule of
cloud as per user request selected in View item section in service offered menu in
Fig. 2.
6 Conclusion and Future Scope
In this paper, we propose an effective cloud-based model for better solutions over the
tedious traditional methods. The advantage of cloud can better serve especially for
hospital where many patients lacking the claim settlement over bank. Accessibility of
patient data over any hospital in the country will demarcate as a tremendous revolution.
Furthermore, every citizen must take advantage of the process of digitization incurred
automatic remote access of data and service. This model is adding future implemen-
tation scope of healthcare industry by transferring and storing messages over cloud. It
captures user authentication by barcode sensor to process and capture user authenti-
cation. This model implements a device to device (D2D) communication and device to
cloud (D2C) communication. The D2D exhibits communication by bluetooth or Zig-
Bee for establishing direct communication. Here, in this model, in Fig. 1, we gather
knowledge of implementation for D2C. Here IoT devices share with internet cloud
directly like an application service provider to exchange data and control message over
network. This model takes the advantage of Wi-Fi network for establishing a con-
nection between device and IP network that will absolutely connect to cloud service.
We as an author must not forget to check vulnerabilities of all cloud interface and API
for authentication and security. IOT to cloud communication is going through this
paradigm as given in Fig. 1.
Device to gateway and then to cloud infrastructure. A device is a thing that interacts
with the real-life world directly in this paper we introduce ID card of a patient that can
be read by barcode reader and data is sensibly passed to gateway for propagating it to
cloud infrastructure. Where the device information of patient should be captured in this
format.{Citizen ID: “dhsl888ja”, “Date of birth”: “Date of issue”, “Signature”:
“Photo”, Barcode}. This IoT-based model uses the hardware interface to exchange data
serially over the network we can general purpose input–output serial interface for
building this mechanism. IoT-based barcode scanner can be used by medical for dis-
semination of patient information to cloud.
References
1. Tochukwu C, Abraham I, Somtoochukwu I (2015) Cloud based health care information
system for rural clinics in Nigeria, Dec 2015
2. Abayomi-Alli AA, Ikuomola AJ, Robert IS, Abayomi-Alli OO (2014) An enterprise cloud-
based electronic health records system
3. Boyinbode O, Toriola G (2015) CloudeMR: a cloud based electronic medical record system.
Abouelmehdi K, Beni-Hessane A, Khaloufi H (2018) Big healthcare data: preserving
security and privacy. J Big Data 5(1). https://doi.org/10.1186/s40537-017-0110-7
4. Chiyan Z (2014) Cloud-based hospital information system as a service for grassroots
healthcare institutions. Singh M, Srivastava VM (2018) Multiple regression based cloud
adoption factors for online firms. In: 2018 International Conference on Advances in
Computing and Communication Engineering (ICACCE). https://doi.org/10.1109/icacce.
2018.8457722
5. Hitachi (2012) How to improve healthcare with cloud computing, Hitachi Data Systems,
White Paper, May, 2012, 1–20
6. Kavitha R, Kannan E, Kotteswaran S (2016) Implementation of cloud based Electronic
Health Record (EHR) for Indian Healthcare Needs, Jan 2016
7. Hu NZ, Lee CY, Hou MC et al (2013) A cloud system for mobile medical services of
traditional Chinese Medicine. J Med Syst 37(6):1–13
8. Karthikeyan N, Sukanesh R (2012) Cloud based emergency health care information service
in India. J Med Syst 36(6):4031–4036. https://doi.org/10.1007/s10916-012-9875-6
9. Xi AB, Qin BQ, Huang CL (2013) The regional medical business process optimization
based on cloud computing medical resources sharing environment. Cybern Inf Technol 13
(Special Issue):18–29
10. Muthaiyan M, Goel N, Prakash DS (2012) Virtual E-Medic: a cloud based medical aid. In:
Proceedings of World Academy of Science, Engineering and Technology (WASET), issue
71, p 1344
11. Min C, Tao L, Le Z (2015) Explore the development of medical and health tourism industry
in Wuzhishan City Hainan Province. Consume Guide 1:45–47
12. Zhang J, Gu Z, Zheng C (2010) The progress on cloud computing research
Fusion-Based Multi-biometrics Authentication
System with Intrusion Detection
and Prevention Anamoly
Mehreen Kirmani and Dinesh Kumar Garg(&)
Department of Computer Science and Engineering, SSCET, Pathankot, India
1 Introduction
The “Biometrics” combines the two Greek words “bio” that represents “the life” and
the word “metrics” that denotes “measuring something”. The biometric devices are
easily available from the last few decades on a very large scale that is mainly because
of the tremendous elevations in the field of computer science and technology. Bio-
metrics recognizes every person distinctly, depending upon one or more than one
inborn physical trait or the behavioural traits. Therefore, the biometric system is used to
make sure that the security and integrity of the personal or the business assets is well-
protected. The output of the verification step in biometric systems produces a matching
a score that is equated and then compared with a decision step. When the result of the
decision step is true, the user is considered as a genuine and a legitimate user; other-
wise, the user is treated as illegitimate and hence, rejected as an attacker [1]. Adding the
extra preventive steps helps to reduce the chances of the spoofing attacks, and our
system gets secured and it becomes less susceptible to attacks [2]. Every biometric trait
has got its pre-defined merits and demerits. Face is treated as the common form of
biometrics, due to its ease of use, and the user-friendliness [3]. On the other hand, Iris
identification is the most dependable and the precise form of the biometrics, available
nowadays [4]. Biometric determiners, therefore, is the easy way to secure those system,
which needs to be protected. The entire biometric community is going through the
difficult tasks of developing an efficient, and secure way, against such threats [5]. Over
a unimodal-biometric system, a multi-biometric system has an advantage of making
authentication systems trustworthy and difficult to attack, also offering the advantage of
user-choice and last not the least, enhancing the authentication-performance. Although
when it comes to implementing the design of multi-biometric systems, it can give rise
to some of the issues [6]. The security of the network is the challenging and one of the
most obvious applications of the biometrics. The most concerning element of the
network security is the authentication of the users. The authenticity of the user is fused
with the identification, and these two together, comprises of the whole purpose of
biometrics technology. The network which requires the users to identify and authen-
ticate themselves as legitimate users, therefore, is the main prospect for biometrics [7].
In the current society, demand for such a network is very high, and also, the use of such
a network, is growing at an exponential rate [8].

https://doi.org/10.1007/978-981-15-2774-6_7
Fusion-Based Multi-biometrics Authentication System … 53
1.1 Architecture of Biometrics System

The below block diagram illustrates the architecture of biometrics. It starts from
extracting the images from the sensors, moving towards extracting the features from the
specific biometric traits and moving towards matching and verifying the biometrics,
that is stored in the form of images, and therefore, deciding whether the user is a
legitimate or an illegitimate one [9]. The working of the whole biometrics system is
divided into two phases.
First is enrolling the individual’s phase, in which the system captures and processes
the biometric traits (one or more than one trait) and stores the unique values for every
individual in the database, on the basis of which access is granted to the individuals.
The block diagram shows the enrolling process (Fig. 1).
Fig. 1. Enrolling the biometric traits
Second phase, which is the verification phase, is a phase during which the security
system collects and processes the biometrics to match with the biometric templates
which are stored during the enrolment phase to grant or deny access to the system [8].
The given block diagram depicts the working of the verification phase (Fig. 2).
Fig. 2. Verifying biometric traits

54 M. Kirmani and D. K. Garg
1.2 Advantages of Biometric System Over Traditional Security Systems

Some the advantages due to which the biometric recognition systems are considered
over traditional security systems are as follows:
1. Security is high:
The traditional security viz. passwords comprise of the numbers, the alphabets and the
symbols, which are becoming easy to hack every day. There are Zillions of hacking
episodes that are taking place every year and because of this reason, people are losing a
lot of money. However, with the advent of the latest Biometric technologies, different
solutions are put forward, which are almost impossible to hack unlike passwords.
2. Accuracy is better:
The most usual traditional security systems such as the passwords or personal identi-
fication numbers (PINs), when hacked costs some amount of money and time, as well
as other resources and usually aren’t accurate. However, the biometric security systems
work by considering the individuals’ biometric traits, may that be physical or beha-
vioural, such as fingerprints, palm pattern, vein pattern, retina, etc. which always
behaves accurately at any time.
3. Accountability is required:
In other verification methods, apart from the biometric traits, anyone can make use of
the password or security PIN number to get the personal information of an individual,
which is considered highly risky. However, if we take the case of the biometric security
system, it always requires the interactions of particular individuals to login or passes as
the legitimate user in the security system that is employed, which will allow 100%
accountability for any sort of the activities.
4. Convenient or User-friendly:
It requires knowledge and is somewhat tiring to remember or pen-down the passwords.
People also are more prone to forget the passwords at some point in future due to the
uneasy situations in one’s life. In such cases, the biometrics, therefore, stands out to be
the most convenient solutions.
5. High Scalability:
The biometric security solutions are the scalable solutions for any kind of projects and
it can be used in almost every organization that is the banking, workforce-management
systems, or semi-government and government projects.
2 Problem Formulation and Methodology

2.1 Problem Formulation
The Spoofing attack is usually defined as an illegal-way by which a person(s) tries to
masquerade as an individual and, therefore, makes an attempt to get the access of a
system.
2.2 Existing Methodology

The exiting methodology presented an idea of fusing the biometric traits (iris and FKP)
together. A matching score is obtained by fusion, which is used to decide the authen-
ticity of an individual(s). The focus was laid on the detecting intrusions, while the
prevention of the intrusions was neglected. Many of the biometric systems undergo the
spoofing attacks that are mostly because they ignore verifying all the parameters.
The existing method monitors for the intrusion detections by getting through the log-
analysis of the system. The extra user-interface and steps enhance the security. The
Database is stored with existing collected featured images of iris and FKP of the user(s).
After the first matching-step, the framework performs the image acquisitions and
afterward, a matching-procedure is performed. In this research work, the biometric traits
which were studied in detail were iris and Fingerprint.
3 Proposed Work
The main aim of the proposed work is to use image-processing techniques for biometric
acquisition and identification process, for as many of the biometric determiners like the
knuckle-print, iris, and the face that are represented by the images. Depending upon the
data for the pre-processing, the algorithms have been used. To enhance the working of
the biometric system using the integrated or multi-modal approach, techniques are used
and are presented in this work. This research work is based on the concept that each
individual has to pass the various biometrics checkpoints in order to get access to the
system. Multi-modal Biometrics recognition systems include making use of more than
one biometric trait to login into the system, which in turn makes the environment more
reliable, and therefore, less prone to the spoofing attacks. Using the combination of the
face, iris and the finger-knuckle print, which are normally registered at the time of the
registration and then verified at the time of the login, followed by generating unique
random OTP’s to the email of individual guarantees the security to the greater extent.
Proposed Algorithm:
Step 1. Multi-Biometric trait acquisition.

Step 2. Extracting the features from the acquired images.
Step 3. Matching the biometric images.

Step 4. Generate the random OTP’s to the Email of the authorized and registered
user for logging in the system.
Step 5. Verification of the generated unique OTP’s, for detecting and preventing
Spoofing attacks.
Step 6. Encryption/decryption of the data files done using the AES algorithm.
Step 7. Transfer of the encrypted data files via an intermediary node, making use of
Dijkstra’s shortest path algorithm.
Step 8. Preventing and detecting the intrusions performed by intermediary node.
3.1 Unimodal Biometric System Versus Multi-modal Biometrics System

The mentioned table compares and gives a contrast between the unimodal security and
the multi-modal biometrics-security system (Table 1).
Table 1. Unimodal-biometric versus multi-modal biometric authentication system

S. Unimodal biometric system Multi-modal biometric system
no
1. Unimodal biometric systems make use of Multi-modal biometric systems integrate
only a single biometric trait, that may be a or make fusion of more than one
physiological trait, such as the iris, biometric trait of an individual that can be
fingerprint, face, etc. or a behavioural a combination of both the types of
trait such as typing rhythm of an biometric traits
individual, signature, etc.
2. A single biometric trait can be stolen as It is tough to steal the multiple biometric
nowadays individual’s information such traits of an individual and also impossible
as face images are easily available on to steal the internal organs of an
social networking sites individual, such as retina, iris, etc.
3. Security is less than expected by the It offers more security compared to the
organization Unimodal biometric system
4. Cannot be trusted Can be trusted as compared to the former
authentication systems
5. Because only one biometric trait is It is costly, but it can be a better option
included, it is less complex and costly for organizations that needs high security
and integrity
4 Results
Security, nowadays, is on the top of the list of every individual or an organization that
has some data to protect. This research is mainly themed on resolving the security
concerns related to individual(s) or organization. Every biometric trait of the user is
captured, matched, verified and only after then, the system generates a unique random
OTP’s for allowing access to the system (Fig. 3).
Fig. 3. Verification of multiple biometric traits

Once the verification of OTP is done, the individual gets an access to the system
and only then, the users will be able to communicate in an encrypted form, and
decryption is done by applying the unique hash values at the destination side, in order
to detect and prevent the attacks on the system (Figs. 4, 5 and 6).
Fig. 4. Generating and verifying the random unique OTP

Fig. 5. Encypting and transferring the data at terminals
The data travels from the source node to the destination through an intermediary
node, which is usually the shortest path to the destination. The path computation can be
done at any stage of the transfer to get the information about who will be acting as an
intermediary node (Fig. 7).
If the intermediary node performs an attack on the system, the data packet that was
sent is dropped and, an acknowledgment is sent to both the sender and receiver, about
data being compromised.
If there is no attack detection, then the corresponding data is successfully trans-
ferred to the destination, which at the terminal side applies the hash and decrypts the
corresponding received data (Fig. 8).
Fig. 6. Intermediary node
Fig. 7. Shortest path computation at the terminal end

Fig. 8. Destination node
5 Future Scope and Conclusion
This research paper helps to provide an approach that works on giving more impor-
tance to security of the system rather than extracting the information from the
resources. Extra user interaction, apart from biometric acquisition expects more
accurate intrusion detection and prevention. For the detection and prevention of the
spoofing attacks, simple pattern linking as well as matching mechanisms are taken into
account, therefore, making a system more reliable, trustworthy, and secured.
References
1. Wild P, Radu P, Chen L, Ferryman J (2015) Robust multimodal face and fingerprint fusion in
the presence of spoofing attacks. Pattern Recogn 50:17–25
2. Kaustubh D, Vishnu RD, Raut VM, Thakare (2016) Effective methodologies for preventing
face spoofing attacks. IJARSE 6(06), June 2016
3. Bagga M, Singh M (2016) Spoofing detection in face recognition: a review. In: 3rd
international conference on computing for sustainable global development (INDIACom),
pp 2037–2042
4. Kak N, Gupta R, Mahajan S (2010) Iris recognition system. (IJACSA) Int J Adv Comput Sci
Appl 1(1)
5. Galbally J, Marcel S, Fierrez J (2014) Biometric antispoofing methods: a survey in face
recognition. IEEE Access 2:1530–1552
6. Kiltter J (2018) Multibiometric for identity authentication: issues, benefits and challenges. In:
2nd IEEE international conference
7. Radhika P, Ramya G, Sadhana K, Salini R. Defending man in the middle attacks. Int Res J
Eng Technol (IRJET) 4(3). e-ISSN: 2395 -0056, Mar-2017, p-ISSN: 2395-0072
8. Gurumurthy S (2015) Design and implementation of biometrics in networks. JTASR 1(3)
eISSN-2454-1788, pISSN-2395-5600. https://doi.org/10.14260//2015/27july-sept 2015
9. Akbulut Y, Sengur A, Budak U, Ekici S (2017) Deep learning based face liveness detection in
videos. In: Conference on international artificial intelligence and data processing symposium,
September 2017
Design of Dual Notch Band Antenna
with Rejected Filter with Defected Slot
for WLAN and Satellite Applications
Sanjay Kumar Sahu1(&), Bhukya Arun Kumar1,

and Kaibalya Kumar Sethi2
1
School of Electronics and Electrical Engineering, Lovely Professional
University, Phagwara, India
sanjay.23393@lpu.co.in
2
Department of Electronics and Communication Engineering, GITA,
Bhubaneswar, India
1 Introduction
Filters at microwaves are extensively used in wireless communication as it adheres to

certain properties such as lightweight, low design cost, flexibility, compatibility with
integrated circuits and more importantly ease of fabrication. Day by day large numbers
of wireless networking techniques are emerging out and those need to be accommo-
dated within the limited spectrum. This stringent demand for modern communication
systems can only be filled if it can replace conventional methods by certain advanced
and versatile techniques with less investment. For example, the bands like 5.15–
5.35 GHz, as well as 5.725–5.775 GHz are unlicensed bands which are meant for
indoor use only. But to use it for outdoor purpose, it is required to facilitate the creation
of new innovative ideas which may come under wider wireless communication net-
works [1]. In this context, we may highlight two powerful methods in terms of defected
microstrip structure (DMS) and defected ground structure (DGS). Both techniques
assume a similar approach of creating some kind of defect either in the signal line or on
the ground plane to get something new to the conventional design [2–6]. The beauty in
this technique is that it shows resonance properties due to abrupt change in the current
path. Though both bring some kind of defect, DMS provides better results than DGS.
The advantageous part in DMS [1–6] over DGS is that in the former case the amount of
cross talk and ground noise is significantly reduced as per the filter design is concerned.
In this design approach, DMS is implemented which refers to band rejection process
which is extremely beneficial to reduce noise in local oscillator, satellite, and mobile
phone. Due to the presence of defects along signal path, electric, and magnetic field
[7–13] are redistributed and this leads to bringing a change in capacitance and
inductance. As a consequence that effective permittivity of the substrate is likely to
change.

https://doi.org/10.1007/978-981-15-2774-6_8
64 S. K. Sahu et al.
2 Geometry of the Proposed Structure
The bandstop filter is constructed on RT-Duroid substrate having 0.002 as loss tangent
value, depicted in Fig. 1. There are two U slot of different length are removed which is
shown clearly in Fig. 2. The dimension is available in table number 1. The resonance
frequency can be varied by changing the length (L) and thickness (LX in this case) of U
slot. Figure 3 is the equivalent diagram for U-slot in terms of R-L-C. The values of R,
L and C can be calculated from the equation given below.
Fig. 1. Structure of proposed filter with defected mechanism
fc 2
C¼ f 0 f 2c ð1Þ
200p
1
L ¼ ð2Þ
4Cp2 f 20
2Z0
R ¼ r
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð3Þ
2&1=jS11j2 ½2Z0 ðwc 1=wLÞ2 1
Resistance is normally a value which is estimated from the conductor and dielectric
losses in the defect.
Design of Dual Notch Band Antenna with Rejected Filter … 65
Fig. 2. Sectional view of U-slot
Fig. 3. Equivalent diagram of the defect in terms of R-L-C
Parameters Quantity Dimensions (mm)

W Width of the filter 25
L Length of the filter 25
h Substrate height 0.787
Wfeed Width of signal line 3.2
LX Width of the U slot 0.4
3 Result Analysis
The proposed structure is simulated using HFSS-13.0 to investigate the frequency

response characteristics. The results obtained are presented in terms of S21, Q-factor
and VSWR plot [14–16] about two resonant frequencies (5.8 GHz and 8.2 GHz). As
per the S21 plot is concern, it is found to be more than −20 dB shown in Fig. 4. For
same frequency, we have also calculated Q-factor by using the mathematical formula
as = 2Δ0
dB(S(1,1)) []
0 dB(S(2,1)) []
-5
S11,S21 in (dB)
-10
-15
-20
-25
-30
0 1 2 3 4 5 6 7 8 9 10
Frequency in (GHz)
Fig. 4. Frequency response plot (S11 and S21) of the filter
dB(VSWR(1)) []
25
20
VSWR in(dB)
15
10
0 1 2 3 4 5 6 7 8 9 10
Frequency in (GHz)
Fig. 5. Required VSWR plot in dB
Where 0 represents resonance frequency, whereas Δ defines the 3 dB bandwidth for

the corresponding frequencies. The value for Q is evaluated which is approximated to
15(14.87). The VSWR result depicted in Fig. 5 explains how good the filter for the
above said frequencies are. The graph of VSWR (≅ 25 dB) is a clear indication of
representing a bandstop filter.
From the result we obtained, it is observed that DMS technique shows very narrow
bandwidth characteristics for a band-reject response of approximately −20 dB. The
presence of U-slot fetches better performance in terms of selectivity with a smaller size.
This is one of the important aspects to remove the interference signal for a particular or
specific frequency application.
Design of Dual Notch Band Antenna with Rejected Filter … 67
4 Conclusion
A compact dual notch bandstop filter is reported just by employing defect along signal
line. The defect being the key factor to realize the above-said filter. An equivalent L-C
circuit model is also included in this design. The frequency response plots (S11 and
S21) are examined through HFSS-13.0 software. The results are quite good for upper
WLAN and X-band satellite communication applications. By varying different
dimensions of slots more number of stop band can be achieved. High quality factor,
better insertion loss makes this filter efficient for practical use.
References
1. Pozar DM (2010) Microwave engineering, 4th edn. Wiley
2. Salari MA, Abbasiniazare S, Manoochehri O (2012) The effect of electromagnetic waves on
multilayer orthogonal microstrip lines with and without defected microstrip structure. IEEE
Antennas Wireless Propag Lett 1:1206–1209
3. Liu HW, Zhang ZC, Wang S, Zhu L, Guan XH, Lim JS, Ahn D (2010) Compact dual-band
bandpass filter using defected microstrip structure for GPS and WLAN applications.
Electron Lett 21:1444–1445
4. Ding C, Guo YJ, Qin P-Y, Yang Y (2015) A compact microstrip phase shifter employing
Reconfigurable Defected Microstrip Structure (RDMS) for phased array antennas. IEEE
Trans Antennas Propag 63:1985–1996
5. Xiao JK, Zhu YF, Li Y, Ma JG, Li XW (2014) Controllable miniature tri-band bandpass
filter using defected microstrip structure. Electron Lett 50:1534–1536
6. Wang J, Zhao J, Li J-L (2014) Compact UWB bandpass filter with triple notched bands
using parallel U-shaped defected microstrip structure. Electron Lett 50:89–91
7. Dash RK, Sahu SK, Mishra CS, Sethi K, Palai G, Sahu S (2016) Realization of ‘non-linear
invisibility cloak’ using meta-material. Optik - Int J Light Electron Opt 127:9635–9639
8. Sahu SK, Sahu S, Mishra CS, Palai G (2016) Analysis of reflected frequency band of
metamaterial grating at THz frequency: a future application of filter. Optik Int J Light
Electron Opt 127:4547–4550
9. Lin C-C, Jin P, Ziolkowski RW (2012) Tri-band-notched Ultrawideband (UWB) antenna
using Split Semicircular Resonator (SSR) and Capacitively Loaded Loops (CLL). IEEE
10. Si L-M, Zhang Q-L, Hu W-D, Yu W-H, Wu Y-M, Lv X, Zhu W (2015) A uniplanar triple-
band dipole antenna using complementary capacitively loaded loop. IEEE Antennas
Wireless Propag Lett 14:743–746
11. Siddiqui JY, Saha C, Antar YMM (2014) Compact SRR loaded UWB circular monopole
antenna with frequency notch charactristics. IEEE Trans Propag 62:4015–4020
12. Panda AK, Satyabrat Sahu K, Mishra RK (2012) A compact triangular SRR loaded CPW
line and its use in highly selective wide band Bandpass filter for WiMax communication
system. In: International conference on computer and devices for communication
13. Sahu SK, Sahu S, Palai G (2017) Finite difference time domain approach to metamaterial
waveguide for IR application as cavity resonator. Optik - Int J Light Electron Opt 130:872–
876
14. Liu Q, Yu Y, He S (2013) Capacitively loaded, inductively coupled fed loop antenna with an
omnidirectional radiation pattern for UHF RFID tags. IEEE Antennas Wireless Propag Lett
12:1161–1164
15. Sarkar P, Ghatak R, Pal M, Padar DR (2012) Compact UWB bandpass filter with dual notch
band using open circuited stubs. IEEE Microw Wireless Compon Lett 9:453–455
16. Hong JS, Lancaster MJ (2001) Microstrip filters for RF/microwave applications. Wiley, New
York
Weighted Filtered-s LMS Algorithm
for Nonlinear Active Noise Control
Tapasmini Sahoo and Kunal Kumar Das(&)
Electronics and Communication Engineering, Institute of Technical Education

and Research, Siksha ‘O’ Anusandhan Deemed University,
Bhubaneswar 751030, India
kunaldas@soa.ac.in
1 Introduction
Active noise control has been an attractive field of research for many years. But
research in this area accelerated only during the last two decades. This is possible with
advancement in Digital Signal Processing and availability of fast processors with huge
memory. ANC is based on acoustically superimposing a noise with its antinoise. This
superposition is carried out with an objective to cancel the noise in real time. For
optimum cancellation of noise, antinoise should be equal in magnitude but opposite in
phase to that of noise. To generate antinoise an adaptive filter is used which adapt to
changes in acoustic paths and noise statistic. ANC systems can be feed forward,
feedback or hybrid in nature. Depending on number of error microphones employed to
record error signals they are further classified as single channel and multichannel
ANCs.
A single channel feedforwad ANC is shown in Fig. 1, where the path between
actual noise source and error microphone is called the primary path, the path between
the secondary loudspeaker and the error microphone is defined as the secondary path
[1]. Antinoise is produced by the controller, which consists of the adaptive filter and
update mechanism. The ANC system uses the reference microphone to pick up the
reference noise, processes it through the controller to cancel the noise. ANC also uses
an error microphone to measure the error which is subsequently used to update the
weights of the adaptive filter. Filtered-X LMS (FXLMS) algorithm, which is a modified
version of LMS algorithm, is used to adapt the adaptive filter. FXLMS algorithm is the
most popular algorithm for ANC, because of its simplicity and low computational
requirement. FXLMS algorithm is modified to develop filtered-error LMS (FELMS)
algorithm to reduce the computational complexity of multichannel ANCs [2].
When reference noise travels through certain structures the acoustic paths become
nonlinear. If the amplitude of the antinoise increases beyond the secondary loudspeaker
specifications, nonlinearity results. The presence of nonlinearity in the primary and
secondary paths reduces the noise cancellation capability of the FXLMS algorithm
drastically. In order to deal with nonlinearity some nonlinear adaptive controller have

https://doi.org/10.1007/978-981-15-2774-6_9
70 T. Sahoo and K. K. Das
been proposed. Multilayer artificial neural network, fuzzy logic, polynomial filter,
Volterra filter, genetic algorithm and particle swarm algorithms [3, 4], etc. have been
employed in nonlinear ANC. Recently, functional link artificial neural network
(FLANN) has become popular for nonlinear ANC [5]. This interest in FLANN is due
to its simple structure and easy to implement filtered-S LMS training algorithm. In
order to further improve noise cancellation capability of FLANN based ANC systems,
several modifications have also been suggested [6–9].
Fig. 1. Schematic diagram of an active noise control system
2 Filtered-s LMS Algorithm
A class of artificial neural networks called functional link artificial neural network
(FLANN) has appeared as a strong contender for identifying nonlinear systems.
FLANN, was proposed by Y. Pao and has a simple structure compared to other
multilayer neural networks [10]. In terms of computational complexity, it has a huge
advantage as it requires much less computation in comparison to multilayer neural
network using back propagation algorithm. In FLANN, the input is functionally
expanded using appropriate basis functions. The basis functions generally used for
functional expansion are trigonometric [5], Chebyshev polynomial [11] and Legendre
polynomial [12]. In [5] Das and Panda have employed FLANN in a nonlinear ANC
and developed a novel filtered-s LMS (FSLMS) algorithm for training the network.
Figure 2 represents block diagram of a nonlinear ANC using FLANN. The reference
input, xðnÞ, is nonlinearly expanded using trigonometric functions as defined below.
Weighted Filtered-s LMS Algorithm for Nonlinear Active … 71
Fig. 2. Block diagram of nonlinear ANC using FSLMS algorithm
sðnÞ ¼ ½xðnÞ; sinðpxðnÞÞ; cosðpxðnÞÞ; . . .; sinðpPxðnÞÞ; cosðpPxðnÞÞ;

xðn 1Þ; sinðpxðn 1ÞÞ; cosðpxðn 1ÞÞ; . . .; sinðpPxðn 1ÞÞ; cosðpPxðn 1ÞÞ; . . .; ð1Þ
xðn NÞ; sinðpxðn NÞÞ; cosðpxðn NÞÞ; . . .; sinðpPxðn NÞÞ; cosðpPxðn NÞÞ:
where P is the order of functional expansion and N is the memory length. The
trigonometrically expanded input signal is rearranged to obtain
sðnÞ ¼ ½s1 ; s2 ; s3 ; . . .; s2P þ 1

s1 ðnÞ ¼ ½xðnÞ; xðn 1Þ; . . .; xðn NÞ
s2 ðnÞ ¼ ½sinðpxðnÞÞ; sinðpxðn 1ÞÞ; . . .; sinðpxðn NÞÞ
ð2Þ
s3 ðnÞ ¼ ½cosðpxðnÞÞ; cosðpxðn 1ÞÞ; . . .; cosðpxðn NÞÞ; . . .;
s2P ¼ ½sinðpPxðnÞÞ; sinðpPxðn 1ÞÞ; . . .; sinðpPxðn NÞÞ;
s2P þ 1 ¼ ½cosðpPxðnÞÞ; cosðPpxðn 1ÞÞ; . . .; cosðpPxðn NÞÞ
This rearrangement facilitates us to restructure the network as a set of time-delay-

line filters, operating in parallel. The output of the time-delay-line filters are defined by
yi ¼ hi ðnÞ si ðnÞ; i ¼ 1; 2; . . .; 2P þ 1 and hi ðnÞ are impulse responses of the 2P þ 1
adaptive filters. The controller output is obtained by
X
2P þ1
y ð nÞ ¼ yi ðnÞ: ð3Þ
i¼1
^
The desired antinoise, dðnÞ, is produced by filtering the controller output through the
secondary path impulse response. The error recorded by the error microphone is obtained
^
eðnÞ ¼ dðnÞ dðnÞ. FSLMS algorithm is employed for updating all the filter weights.
The weight update equations can be compactly written as hi ðn þ 1Þ ¼ hi ðnÞ þ
0
lf eðnÞsi ðnÞ. The developed FLANN structure has a relatively simpler structure and has
lower computational load compared to multilayer artificial neural networks for nonlinear
ANCs. But this FLANN structure with FSLMS algorithm inherently slow in conver-
gence. This is because it is based on the LMS algorithm which has a slow convergence
capability. Many improvements in the algorithm have been proposed by researchers [13–
15]. But research has been mainly focused on reducing the steady-state mean square error
performance by taking care of the cross terms. In order to circumvent this drawback of
slow convergence, a new structure is proposed in this paper.
3 Weighted Filtered-s LMS Algorithm
In the proposed system a set of new adaptive weights, called combining weights are
introduced. The output of the adaptive filters is multiplied with combining weights. The
controller output is obtained as follows:
yðnÞ ¼ w1 ðnÞy1 ðnÞ þ w2 ðnÞy2 ðnÞ þ þ w2P þ 1 ðnÞy2P þ 1 ðnÞ ð4Þ
Fig. 3. Block diagram of the proposed weighted filtered-s LMS algorithm

The weights of the adaptive filters are adapted by the FSLMS algorithm and the
newly introduced combining weights are updated by the following equation, where lc
is the step size for newly introduced combining weights.
wi ðn þ 1Þ ¼ wi ðnÞ þ lc eðnÞs0i ðnÞ ð5Þ
4 Simulation Study
The performance of the proposed algorithm for nonlinear ANC is evaluated by con-
ducting a battery of simulation experiments. The noise cancellation capability of the
proposed algorithm is compared with that of the standard FSLMS algorithm. Mean
Square Error (MSE), defined by MSE ¼ 10 log10 ðE½eðnÞ2 Þ is considered as the per-
formance measure. In each experiment average, MSE of 100 independent trials is
plotted. The reference noise considered in the simulation experiments is a uniformly
distributed random noise within the −0.5 to 0.5. It is assumed that the secondary path is
accurately identified by system identification techniques (Fig. 3).
4.1 Experiment I
The primary and secondary paths are modelled as FIR filters of length 256 and 128,
respectively. The coefficients of the acoustic paths are obtained from the data attached
to [1]. The order of the FLANN filter is 1 and memory length is 256. lf ¼ 0:00005 and
lc ¼ 1 where lf and is the step size for FLANN filter and lc is the step size for
combining weights. The Average MSE is shown in Fig. 4.
Fig. 4. Average MSE for experiment I

4.2 Experiment II
Primary path and secondary path transfer functions considered in this experiment are
PðzÞ ¼ z5 0:8z6 þ 0:7z7 and SðzÞ ¼ z2 þ 1:5z3 z4 [5]. The order of the
FLANN expansion is 1 and memory length is 15. The step sizes are lf ¼ 0:00006 and
lc ¼ 0:4. The Average MSE is shown in Fig. 5.
Fig. 5. Average MSE for experiment II
4.3 Experiment III

In this simulation experiment, the reference noise at the error microphone is obtained
by first filtering the noise through a filter RðzÞ ¼ z3 0:3z4 þ 0:2z5 and then
computing dðnÞ, where uðnÞ ¼ xðnÞ rðnÞ and rðnÞ is the impulse response of the filter
RðzÞ [16].
The secondary path is having a transfer function SðzÞ ¼ z2 þ 1:5z3 z4 . The
order of FLANN expansion is 1 and memory length is 15. The step sizes are lf ¼
0:0005 and lc ¼ 0:4. The Average MSE is shown in Fig. 6.
4.4 Experiment IV
In this experiment, both the primary and secondary paths are considered as nonlinear
[17]. Here the desired noise signal dðnÞ and reference noise xðnÞ are related by
dðnÞ ¼ xðnÞ þ 0:8xðn 1Þ þ 0:3xðn 2Þ þ 0:4xðn 3Þ

0:8xðnÞxðn 1Þ þ 0:9xðnÞxðn 2Þ
þ 0:7xðnÞxðn 3Þ
The secondary path is defined by
^
dðnÞ ¼ yðnÞ þ 0:35yðn 1Þ þ 0:09yðn 2Þ 0:5yðnÞyðn 1Þ þ 0:4yðnÞyðn 2Þ
Where yðnÞ is the output from the controller. The order of the FLANN expansion is 1
and memory length is 15. The step sizes are lf ¼ 0:0003 and lc ¼ 1. The Aver-
age MSE is shown in Fig. 7.
Fig. 6. Average MSE for experiment III
Fig. 7. Average MSE for experiment IV

5 Conclusion
In this paper, we have modified the structure of FLANN by introducing a set of

adaptive weights. These weights are used to adaptively combine the output of subfilters
of NANC system. A suitable algorithm is also developed for updating these combining
weights. These modifications not only alleviate the drawback of slow convergence but
also achieve it without sabotaging the structural simplicity of FLANN. The developed
algorithm has emerged as a strong competitor of standard FSLMS algorithm as it
achieves faster convergence without sacrificing MSE and retain the low computational
load of FSLMS algorithm. This structural modification can also be incorporated into a
generalized FSLMS algorithm.
References
1. Kuo SM, Morgan DR (1996) Active noise control systems: algorithms and DSP
implementations. Wiley, New York
2. Wan EA (1996) Adjoint LMS: an efficient alternative to the Filtered-x LMS and multiple
error LMS algorithm. In: Proceedings of IEEE ICASSP, pp 1842–1845
3. Chang CY, Louh FB (2007) Enhancement of active noise control using neural based
Filtered-x algorithm. J Sound Vib 305:348–356
4. Tan L, Jing J (2001) Adaptive Volterra filters for active control of nonlinear noise processes.
IEEE Trans Signal Process 49(8):1667–1676
5. Das DP, Panda G (2004) Active mitigation of nonlinear noise process using novel Filtered-s
LMS algorithm. IEEE Trans Speech Audio Process 12(3):313–322
6. Sicuranzaan GL, Carini A (2012) A generalized FLANN filter for nonlinear active noise
control. IEEE Trans Speech Audio Process 19(8):2412–2417
7. Lee DC, Zhang J, Pang Y (2018) A generalized exponential functional link artificial neural
network with channel reduced diagonal structure for nonlinear active noise control system.
Appl Acoust 139:174–181
8. Lee DC, Zhang J, Pang Y (2018) A bilinear functional link artificial neural network filter for
nonlinear active noise control and its stability condition. Appl Acoust 132:19–25
9. Patra JC, Bal RN, Chatterji BN, Panda G (1999) Identification of nonlinear dynamic systems
using functional link artificial neural network. IEEE Trans Syst Man Cybern Part B 29(2):
254–262
10. Pao Y (1989) Adaptive pattern recognition and neural network. Addition-Wesley, MA
11. Patra JC, Kot AC (2002) Nonlinear dynamic system identification using Chebyshev
functional link artificial neural network. IEEE Trans Syst Man Cybern Part B 32(4):505–511
12. George NV, Panda G (2013) Active control of nonlinear noise processes using cascaded
adaptive nonlinear filters. Appl Acoust 74:117–222
13. George NV, Gonzalez A (2014) Convex combination of nonlinear adaptive filters for active
noise control. Appl Acoust 76:156–161
14. Xhao H, Zeng X, He Z, Li T (2012) Adaptive extended pipelined second order Volterra
filterfor nonlinear active noise controller. IEEE Trans Audio Speech Lang Process 20(4):
1394–1399
15. Zhao H, Zeng X, He Z, Yu S, Chen B (2016) Improved functional link artificial neural
network via convex combination for nonlinear active noise control. Appl Acoust 42:
351–359
16. Patel V, George NV (2015) Nonlinear active noise control using spline adaptive filter. Appl
Acoust 93:38–43
17. Zhou D, DeBrunner V (2007) Efficient adaptive nonlinear filter for nonlinear active noise
control. IEEE Trans Circuits Systems-I 54(3):669–681
Comparison of Classifiers for Speech Emotion
Recognition (SER) with Discriminative
Spectral Features
Hemanta Kumar Palo1(&), Debasis Behera2,

and Bikash Chandra Rout1
1
Siksha O Anusandhan (Deemed to Be University),
hemantapalo@soa.ac.in
2
C. V. Raman College of Engineering, Bhubaneswar, Odisha, India
1 Introduction
The ambiguity of human behavior in society increases day by day that belongs to the
study of human affective states. Among different modalities of expressive emotions
such as facial, gestures, and spoken conversation, recognition of emotions from the
speech is a complex domain of research as spoken words contribute only 7% in an
expressive behavior as compared to 55% and 38% due to facial expression and voice
intonation and semantic information, respectively [1]. Acoustic features representing
these speech emotions are most informative in the spectral domain. However, spectral
features including the MFCCs suffer from embedded pitch filtering and textural
dependency that result in poor SER accuracy. The absence of the desired sentence-level
information for speech emotion identification motivates researchers to explore the delta
(or differential) and double delta (or acceleration) coefficients as additional inputs to the
baseline features [2–5]. Nevertheless, the use of a fixed window in these techniques has
resulted in an equal time resolution for both high and low frequencies. For better
analysis at these frequencies, wavelet analysis with an adaptive window remains an
effective time–frequency tool in this field [6]. The spectral features extracted in the
wavelet domain have shown to outperform the baseline spectral features in the field
of SER.
There are several conventional classifiers applied in the field of SER. Nevertheless,
the suitability of a classification algorithm largely depends on the discriminative fea-
tures and the feature dimension. Three of the classifiers such as the GMM, PNN, and
RBFN have been emphasized in this work due to their following attributes: First, it is
possible to model any pattern using a sufficient number of RBF neurons in a single
hidden layer, unlike MLP, which requires more than one hidden layer [7, 8].

https://doi.org/10.1007/978-981-15-2774-6_10
Comparison of Classifiers for Speech Emotion … 79
Like MLP, the RBFN does not require back-propagating the error through multiple
hidden layers and hence is faster. Similar to RBFN, the PNN is a neural network-based
and is parallel structured. The network is faster by 200,000 to 1 than MLP and two to
three orders faster than RBFN [9]. It is simple, requires a single parameter adjustment
during the simulation, and can be implemented easily [10]. On the contrary, GMM is a
stochastic classifier that uses statistical parameters to model the desired emotion ade-
quately. Ease of learning, bootstrapping from flat data, and faster computation make the
classifier compatible for frame-level features [11].
The rest of the sections are organized as follows. The feature extraction techniques
proposed in this work are provided in Sect. 2. The simulated results and detailed
discussions on the outcomes of the proposed work using the chosen classifiers have
been elaborated in Sect. 3. The conclusion and further possible future directions have
been briefed in Sect. 4.
2 Proposed Feature Sets
The proposed method of feature extraction has been provided in this section.
The MFCC and wavelets have been used as the baseline features to extract the pro-
posed discriminant features with different dimensions. In the first phase, MFCC and
wavelets are used to simulate and compare the GMM, RBFN, and the PNN classifiers.
Subsequently, different variants of MFCC such as differential MFCC (delta MFCC or
DMFCC), wavelet MFCC (WMFCC), and wavelet differential MFCC (WDMFCC)
features are simulated and tested with these classifiers for the performance appraisal.
Finally, hybrid feature sets such as MFCC + DMFCC and WMFCC + WDMFCC
have been formed as a possible means to enhance the classification accuracy of the
GMM classifier.
The signal si ðmÞ is framed into 30 ms with 10 ms overlapping between frames. The
Hamming window has been used to obtain the windowed signal for spectral estimation
using FFT [12]. From each frame of an utterance, 16 numbers of MFCC coefficients are
extracted. For N frames per utterance and fifty utterances per emotion, the feature
matrix has a size 50N 16.
In wavelet analysis, a 3-level decomposition using Daubechies-4(dB-4) wavelets
has been used. For 16 bit/sample used in this work, and taking the detailed coefficients
D-6 into account, the feature dimension remains approximately 470 1 per utterance.
For 50 numbers of utterances per emotion used in this work, the feature size becomes
ð50 470Þ 1 ¼ 23500 1. The proposed derived feature extraction scheme is
shown in Fig. 1.
80 H. K. Palo et al.
Fig. 1. Proposed method of derive feature extraction
2.1 Wavelet-Based MFCC (WMFCC)

The wavelet-based MFCC (WMFCC) technique is herby explained below.
The signal at the ith level is decomposed into detailed di and approximation ai
coefficients, respectively. A three-level decomposition using Daubechies-4 or dB-4 has
been chosen. The Surrey Audio-Visual Expressive Emotions (SAVEE) database used
in this work has a sampling frequency of 44.1 kHz which is downsampled to 16 kHz
with a bandwidth of 8 kHz [13]. Thus, the sub-bands are segmented into (1) 0–1 kHz,
(2) 1–2 kHz, (3) 2–4 kHz, and (4) 4–8 kHz range using the filtering. The MFCC
coefficients are fed to these bands individually as shown in Fig. 1. From each sub-band
1–2 kHz and 2–4 kHz, five WMFCC coefficients are extracted. For each sub-band
0–1 kHz, 4–8 kHz, three WMFCC coefficients are fetched. The number of WMFCC
coefficients in 1–4 kHz frequency bands (middle bands) is more than other bands. The
aim is to extract suitable perceptual information from these middle bands as the voice
components are more prominent in this band. Sixteen WMFCCs per utterance are
collected in this way. For fifty utterances of each emotion, the size of the WMFCC
matrix becomes 50 16:
2.2 Differential MFCC (DMFCC)

The delta or differential MFCC features signify the dynamism involved or the trajec-
tories of the cepstral-based static MFCC feature concerning time [5]. Thus, including
the DMFCC coefficients with the MFCC coefficients will likely to enhance the
recognition performance of speech emotion. These velocity or time derivative features
of MFCC at time t can be extracted using the regression formula as given by
PK
kðMFCCði þ kÞ mfccði kÞÞ
DMFCCðtÞ ¼ k¼1
P ð1Þ
2 Kk¼1 k 2
where MFCC ðt þ kÞ and MFCC ðt kÞ are the static coefficients computed from
frame i. The value of K maybe 2 or 3 in practice. However, for delta MFCC K ¼ 2 has
been chosen here. For 16 MFCC features per frame extracted in this work, there are 16
DMFCC features per frame. The resultant feature size for fifty utterances per emotion
becomes 50 N 16.
2.3 Wavelet-Based Differential MFCC (WDMFCC)

While the wavelet analysis gives information on both the time and frequency content of
a signal, the conventional MFCCs provide information on the signal’s amplitude or
energy [14]. To accommodate the dynamism in the static wavelet-based MFCC fea-
tures, the differential features of WMFCC have been proposed in this work. The
procedure to extract WDMFCC features is enumerated below.
i. Extract the WMFCC features as per the procedure explained earlier.
ii. Apply the derivative algorithm to the WMFCC features extracted as in step (i) as
per Eq. (1).
Similar to the 16WMFCC feature extracted per utterance of an emotion, there are
16 numbers of WDMFCC features obtained per utterance. For fifty utterances per
emotion, the size of the WDMFCC feature matrix thus becomes 50 16.
2.4 Proposed Feature Combination

An attempt is made to combine the static and differential MFCC features for possible
enhancement of accuracy with the GMM classifier. For an N number of frames per
utterance, and 50 utterances per emotion, the feature matrix size thus becomes 50 N
(16MFCC + 16DMFCC) = 50 N 32. Similarly, the wavelet-based static MFCC
and differential MFCC are combined to form another hybrid feature set as indicated by
WMFCC + WDMFCC. In this case, the feature dimension using WMFCC and
WDMFCC feature becomes 50 32. The involvement of both static and dynamic
components of the baseline features has proved more beneficial for the recognition task
as evidenced in the result section.
3 Results and Discussion
The classification accuracy and the computation time concerning the feature dimen-
sions have been used for comparison. The utterances of emotion are divided into
training and testing sets. The utterances used for the training the classifiers are not used
for testing purposes. The classification accuracy of individual emotions with the
extracted feature sets with the RBFN classifier has been shown in Fig. 2, and that of
PNN is shown in Fig. 3.
Fig. 2. Recognition accuracy of RBFN with different variants of MFCC features
Fig. 3. Recognition accuracy of PNN for different variants of MFCC features

The classification performance of the individual speech emotion with the GMM
classifier is shown in Fig. 4. A GMM model is made for each emotion using 16 mean,
16 variances, and one mixture component.
Fig. 4. Recognition accuracy of GMM with different variants of MFCC features
The comparisons of classifiers in terms of lowest classification time, highest

average accuracy, and feature reliability have been provided in Table 1.
Table 1. Comparison of classifiers based on recognition accuracy, feature dimension, and

computation time with the proposed feature set
Classifiers Feature Average Classification Feature
accuracy (%) time (s) dimension
RBFN WMFCC + WDMFCC 80.51 28.41 Low
PNN WMFCC + WDMFCC 80.79 10.56 Low
GMM WAVELETS 71.82 07.52 High
The following important conclusions can be drawn from Figs. 2, 3, and 4. The
classification accuracy of high arousal emotions such as angry, fear, and happy is found
to be better due to their well-defined higher energy levels The sad state has the lowest
arousal level and found to provide the least classification accuracy. The PNN proves
better than RBFN as observed from Figs. 2 and 3. Unlike the RBF network where it is
needed to adjust the weights, it is possible to set all the weights of PNN to one. This has
made the PNN more reliable to low-dimensional feature sets without any over-fitting as
compared to the RBFN when Fig. 3 is compared with that of Fig. 2. However, the
difference in accuracy is not significant. The WMFCC and WDMFCC techniques have
provided the lowest feature size as compared to other feature extraction techniques as
explained in Sects. 2.1 and 2.3. Analysis of Figs. 2 and 3 shows enhanced classifi-
cation accuracy with these feature extraction methods. The use of the multiresolution
capability of wavelet analysis makes these features more reliable. Further, these fea-
tures provide both time and frequency localized information due to their extraction in
the wavelet domain besides the energy information provided by MFCC. These factors
have increased the available information of the signal under consideration that has
resulted in enhanced SER. The NN-based classifiers outperform the GMM classifier in
terms of classification accuracy for small feature sets such as WMFCC and WDMFCC.
On the contrary, better emotional modeling has been observed using a GMM classifier
than NNs for large feature dimensions. This has been observed when Figs. 2, 3, and 4
have been compared with Table 1. For large training features, the GMM classifier can
capture meager differences among distributions. As a result, the classifier is well suited
for frame-level spectral features extracted particularly from emotional speech signals.
The results in Table 1 also indicate the reliability of the chosen feature sets based on
the classification time and highest average recognition accuracy. An attempt is made
further to enhance the recognition accuracy of the emotions using feature hybridization
in Table 2.
Table 2. Effect of feature combination on the computation time, and classification accuracy
concerning the feature dimension using GMM
Features Feature Feature Classification Average
dimension extraction time time (s) accuracy
per emotions per utterance (s) (%)
MFCC + DMFCC 50 N 32 2.38 19.32 80.1
WMFCC + WDMFCC 50 32 2.65 09.42 60.2
The combination approaches have been tested to obtain better GMM modeling as
the classifier has proved better for large feature dimensions in our work. The use of
multiresolution capability in wavelet analysis has resulted in enhanced accuracy for the
WMFCC + WDMFCC as compared to the MFCC + DMFCC combination.
4 Conclusions
The use of both static MFCC and its dynamic (DMFCC) information in the extraction
of the features has improved the classification accuracy because of the availability of
temporal emotional information. The MFCC features extracted using the multiresolu-
tion wavelet analysis have resulted in a more reliable feature than either the baseline
MFCC or the wavelets. Further, the involvement of differential algorithm with wavelet-
based MFCC features is a novel effort in this direction. The combination of MFCC
variants having a larger dimension has resulted in enhanced accuracy in the case of
GMM. However, the NNs tend to be better performing with a small set of features. It
indicates the feature dimension plays an important role in deciding the classification
accuracy in the proposed techniques. Although, it has increased the computational

complexity of feature extraction, at the same time the classification time has been
reduced. Hence, a trade-off has to be made depending upon the task domain. Similarly,
the exploration of other reduction, modification, and combination mechanisms may
provide new research directions in this area. Further, the effectiveness of other con-
ventional classifiers to different feature size is an area needs further investigation.
References
1. Mehrabian A (1971) Silent messages, vol 8. Wadsworth, Belmont, CA
2. Deng J, Frühholz S, Zhang Z, Schuller B (2017) Recognizing emotions from whispered
speech based on acoustic feature transfer learning. IEEE Access 5:5235–5246
3. Joe CV, Sugi SSS (2016) Optimal feature for emotion recognition from speech. Afr J Basic
Appl Sc 8(3):136–144
4. Sato N, Obuchi Y (2007) Emotion recognition using Mel-frequency Cepstral coefficients. Inf
Media Technol 2(3):835–848
5. Palo HK, Chandra M, Mohanty MN (2018) Recognition of human speech emotion using
variants of Mel-Frequency Cepstral coefficients. In: Advances in systems, control and
automation. Springer, Singapore, pp 491–498
6. Palo HK, Mohanty MN (2018) Wavelet-based feature combination for recognition of
emotions. Ain Shams Engg J 9:1799–1806
7. Bishop CM (1995) Neural networks for pattern recognition. Oxford University Press
8. Javidi MM, Roshan EF (2013) Speech emotion recognition by using combinations of C5. 0,
neural network (NN), and support vector machines (SVM) classification methods. J Math
Comput Sc 6:191–200
9. Specht DF, Missiles L, Inc S, Alto P (1990) Probabilistic neural networks and the
polynomial Adaline as complementary techniques for classification. IEEE Tran Neural Net 1
(1):111–121
10. Palo HK, Sagar S (2018) Comparison of neural network models for speech emotion
recognition. In: 2nd international conference of data science and business analytics, IEEE,
pp 127–131
11. Ayadi E, Kamal MS, Karray F (2011) Survey on speech emotion recognition: features,
classification schemes, and databases. Pattern Recogn 44(3):572–587 (Elsevier)
12. Palo HK, Behera D, Analysis of speaker’s age using clustering approaches with emotionally
dependent speech features. In: Critical approaches to information retrieval research, IGI
Global, 2020, pp 172–197
13. Haq S, Jackson PJB (2010) Multimodal emotion recognition. In: Wang W (ed) Machine
audition: principles, algorithms, and systems. IGI Global Press, chapter 17, pp 398–423
14. Palo HK, Chandra M, Mohanty MN (2017) Emotion recognition using MLP and GMM for
Oriya language. Int J Comput Vision Robot 7(4):426–442
Handwritten Odia Numerals Recognition:
A Supervised Learning Perspective
Suchismita Behera(&) and Niva Das
Department of ECE, Siksha ‚‘O’ Anusandhan (Deemed to be University),

{suchismitabehera,nivadas}@soa.ac.in
1 Introduction
Handwritten Recognition (HR) systems measure the ability of a computer or device to

efficiently recognize the handwritten documents. HR systems may be developed to
operate in offline mode or in online mode. Many devices like smart phones, PDAs and
touch screens can take handwriting as input. Online handwriting verification may be
employed for password protected systems as it takes less time to write down numbers
and text to the screen [1, 2]. Recognition of handwritten documents is considered as
one of the challenging areas of research before the image processing community. Some
of the popular application areas of the numeral recognition problem include postal
automation, bank cheque processing, automatic data entry, etc. Though research has
attained a saturation in printed character recognition, it is limited in case of handwritten
numerals and characters basically due to large variation observed in shape, scale and
format arising out of individual writing styles. Already many pieces of work are
available in the area of HR systems for non-Indian language scripts such as Roman,
Japanese, Chinese and have been implemented for automatic sorting in postal systems.
Though India is a multilingual country and multiple scripts are used, not much work
has been done on character recognition. With respect to the Indian scripts, few pieces of
work are available in the area of numeral recognition. The major challenges in rec-
ognizing the handwritten numerals are the variations in size, style and the shape. These
problems can be best resolved by selecting good feature extraction techniques along
with an efficient classifier. Many feature extraction and classifier techniques have been
reported in the literature for this purpose.
In this paper, we have made an attempt to recognize the handwritten Odia numerals
with the help of the Support Vector Machine (SVM) classifier and single decision tree
classifier for pixel intensity-based features and Histogram of Oriented Gradient features
(HOG) separately. Experimental results on a newly developed Odia handwritten digit
database collected from IIT, Bhubaneswar are obtained through a simulation study and
presented. The paper has been organized as follows: Sect. 2 provides the literature
review of the work done on handwritten Odia numerals, Sect. 3 gives an overview of
handwritten Odia numerals. Feature extraction techniques used are highlighted in
Sect. 4, experimental results are analyzed in Sect. 5, followed by a Conclusion in
Sect. 6.

https://doi.org/10.1007/978-981-15-2774-6_11
Handwritten Odia Numerals Recognition: A Supervised Learning … 87
2 Literature
Various approaches reported in the literature for handwritten numeral recognition are
based on structural [3] and statistical [1] features. Some popular approaches used for
this purpose are Back-Propagation Neural Network (BPNN), support vector machines,
k-Nearest Neighbor(k-NN), etc. Pal et al. [4] used curvature features to achieve a
recognition rate of 94% in Odia numerals. A. Desai [5] used Neural Network for
recognition of Gujarati numerals. Bhattacharya et al. [6] used Adaptive Self Organizing
Neural Network and MLP network for feature extraction and classification, respec-
tively, for efficient recognition of Bangla handwritten numerals. Dash et al. [7] have
reported very high recognition efficiency for Odia handwritten numerals with Stockwell
and Slantlet transformation-based features. Roy et al. [8] have used a direction chain
code histogram of contour points of the numerals for efficient recognition of Odia
digits. In [9], Mishra et al. have reported a combination of Discrete Cosine Transform
(DCT) and Discrete Wavelet Transform (DWT)-based feature extraction techniques for
Odia handwritten digit recognition. In [10], Pujari et al. used curvature and gradient
features to get a recognition rate of 90% with SVM classifier. In [11], Impedevo et al.
have reported a classification accuracy of 98% with a k-NN classifier with MNIST
numeric dataset.
3 Odia Handwritten Numerals
Odia is the regional language of the Indian state Odisha, spoken by more than 40
million people, belonging to Odisha and the neighboring states like West Bengal,
Jharkhand, Chhattisgarh, and Andhra Pradesh. Like other language scripts, Odia script
consists of 10 numerals from 0 to 9. One set of handwritten Odia numerals is shown in
Fig. 1. Numerals such as one, two, six, seven are seen in two shapes. Nine is also
written in a variety of ways by different people. One numeral may also look similar to
another. These similarities of shape and structure in different numerals pose challenges
for the recognition task.
The handwritten Odia numeral database of IIT, Bhubaneswar is considered for
implementing this paper. The database contains more than 250 images for each
numeral collected from different sources. As the images are in.tif and.jpeg format, they
are converted to a uniform image format for further processing.
88 S. Behera and N. Das
Fig. 1. Sample handwritten Odia numerals from IIT Bhubaneswar database
4 Feature Extraction and Classification
4.1 Feature Extraction

Due to variations in size and presence of noise in the scanned images, some prepro-
cessing is required. The scanned images are first converted to grayscale format and then
filtered to remove noise. Then images are normalized in size so as to have a uniform
size of (24 24) for both training and testing phase. We have used two sets of features
for numeral recognition.
The first set is based on intensity information of each pixel. Here, each numeral
image is vectorized to a form of a column vector of size 576 1. We have considered
400 images of each numeral for training and 100 images for testing. Hence for 10
numerals, the training set is of size 4000 1 and testing set is of size 1000 1.
The other set is formed by extracting the HOG features of each image. To calculate
the HOG features of each image the following steps are adopted:
1. Gradient vector of each pixel is calculated along with magnitude and direction.
2. Each image is divided into cells of size 4 4.
3. In each cell, the magnitude and direction of the 16 pixels are represented by a 9-bin
histogram as in [12] which is known as HOG descriptors.
4. A block of size 8 8 is formed consisting of 4 cells which is represented by a
vector of (36 1) HOG descriptors. The block is slided horizontally and vertically
to cover the whole image which ultimately gives rise to five horizontal and five
vertical block positions. Hence for the entire image, the final HOG vector is
obtained by concatenation of HOG vectors of all blocks, i.e., 25 36 1.
4.2 Classification
We have used SVM classifier and single decision tree classifier for classification of the
handwritten numerals. In one experiment the pixel intensity information forms the
features whereas the HOG descriptors act as features for another experiment for both
training and testing set.
For implementation, we apply the intensity values and HOG descriptor-based feature
extraction techniques on handwritten Odia numeral database collected from IIT Bhu-
baneswar. Two experiments have been carried out on the database. For one experiment
pixel intensities are used as the feature set whereas HOG descriptors are used as
features for the second experiment. We randomly selected 400 images for training and
100 images for testing for each numeral class. SVM and single decision tree classifiers
are applied separately to the features extracted by two different methods in both the
experiments. The results are presented in the form of Confusion Matrix which specifies
the number of correctly identified numerals along the diagonal. Tables 1, 2, 3 and 4,
respectively, show the Confusion Matrix. Table 5 shows the recognition accuracy for
each numeral class for four different feature extraction cum classifier combinations.
Table 6 shows a comparison of the computational time for different feature extraction
cum classifier techniques. It is evident from the results that the decision tree classifier
requires less computational time and hence may be recommended for online applica-
tions and large database.
Table 1. Confusion matrix of pixel intensity-based single decision tree classifier

Table 2. Confusion matrix of pixel intensity-based SVM classifier
Table 3. Confusion HOG feature-based single decision tree classifier

Table 4. Confusion matrix of HOG feature-based SVM classifier
Table 5. Recognition accuracy matrix of individual numerals in %

Numeral Features + classifier
class Pixel Pixel HOG + single HOG + SVM
intensity + single intensity + SVM decision tree
decision tree
0 80 94 96 100
1 60 90 98 100
2 78 100 96 100
3 72 94 90 100
4 90 98 94 100
5 70 96 94 100
6 74 96 92 100
7 86 92 92 100
8 88 100 92 100
9 74 98 88 100
Overall 77.2 96 93.2 100
accuracy
Table 6. Computational time for different recognition techniques

Feature extraction + classifiers Computational time (s)
Pixel intensity-based single tree classifier 7.5001
Pixel intensity-based SVM 134.2926
HOG-based Single tree classifier 4.1538
HOG-based SVM 4.4532
6 Conclusion
This paper presents a comparison of performances of four different feature extraction

and classifier schemes for off-line handwritten Odia numeral recognition task. The
schemes are tested on a database that contains more than 400 numerals for each class
collected from various sources by IIT, Bhubaneswar. The overall recognition accuracy
obtained from SVM classifier is more compared to single decision tree classifier.
A further single decision tree requires less computational time compared to SVM.
Work on this may be extended towards designing a more robust classifier that can
perform well in the presence of irregularity and missing parts.
References
1. Mahadevan U, Srihari SN (1999) Parsing and recognition of city, state, and ZIP codes in
handwritten addresses. In: Proceedings of the fifth international conference on document
analysis and recognition, pp 325–328
2. Kimura F, Sridhar M (1991) Handwritten numeral recognition based on multiple algorithms.
Pattern Recogn 24:969–983
3. Plamondon R (2000) Srihari SN (2000) Online and off-line handwriting recognition: a
comprehensive survey. IEEE Trans. Pattern Anal. Mach. Intell. 22(1):63–84
4. Pal U, Wakabayashi T, Kimura F (2007) A system for off-line oriya handwritten character
recognition using curvature feature. In: 10th international conference on information
technology, pp 227–229
5. Desai A (2010) Gujrati handwritten numeral optical character recognition through neural
network. Pattern Recogn 33:2582–2589
6. Bhattacharya U, Das TK, Datta A (2002) A hybrid scheme for handprinted numeral
recognition based on a self-organizing network and MLP classifiers. Int J Pattern Recogn
Artif Intell 16:845–864
7. Dash KS, Puhan NB, Panda G (2015) Handwritten numeral recognitionusing non-redundant
stockwell transform and bio-inspired optimal zoning. IET Image Proc 9(10):874–882
8. Roy K, Pal T, Pal U, Kimura F (2005) Oriya handwritten numeral recognition system. In:
Proceedings of eighth international conference on document analysis and recognition, vol 2,
pp 770–774
9. Mishra TK, Majhi B, Panda S (2013) A comparative analysis of image transformation for
handwritten Odia numeral recognition. In: International Conference on Advances in
Computing, Communications and Informatics (ICACCI)
10. Pujari P, Majhi B (2015) Recognition of offline handwritten Odia numerals using support
vector machine. In: International conference on computational intelligence and networks
11. Impedevo S, Mangini FM, Barbuzzi D (2014) A novel prototype generation technique for
handwriting digit recognition. Pattern Recogn 4(3):1002–1010
12. Dalal N, Triggs B (2005) Histogram of oriented gradients for human detection. In:
Proceedings of CVPR, vol 1, pp 886–893
An Implementation of Neural Network
Approach for Recognition of Handwritten
Odia Text
Sachikanta Dash1(&) and Rajendra Kumar Das2

1
Department of CSE, DRIEMS Autonomous Engineering College, Tangi,
Cuttack, Odisha 754022, India
sachikanta_dash@rediffmail.com
2
Department of ENTC, DRIEMS Autonomous Engineering College, Tangi,
Cuttack, Odisha 754022, India
1 Introduction
In a practical point of view, all the ongoing innovations based on Computerization, for
example, word processors, fax machines and email are having their effect on hand-
writing. These in-varieties have prompted the adjusting and reinterpreting of the job of
handwriting and handwritten messages. HCR is the skill of a machine to get and
translate manually written contribution from numerous sources like paper records,
photos, and touch screen gadgets, and so on. HCR method is the process of recognizing
the writer from a set of handwritings obtained from a lot of writers. Identification and
verification are forms that resolve the unique idea of the work of a particular writer,
while handwriting interpretation and translation are forms whose goals are to sift
through the varieties of processes to determine the message. Recognition of manually
written and machine characters is a rising territory of research and finds broad appli-
cations in banks, workplaces, and businesses. In the HCR process, an image containing
text must be provided adequately and pre-processed. Next, the message text should
either undergo segmentation or feature extraction. Small piece of writing will be the
outcome, and these must go through various recognition models. Lastly, content-based
information should be applied to the recognized data images to verify the result.
Machines that process handwriting styles will be dealing with many styles and lan-
guages of arbitrary users. Numerous types of analysis, recognition, and translation can
link with the handwriting recognition process. The Table 1 encapsulates a few research
works proposed by different authors on handwritten characters.
Presently, there is no single methodology that involves it both effectively and
efficiently in all contexts. Artificial neural network systems, associated with the
handwriting recognition process, don’t require profound background knowledge to be
able to solve the handwritten recognition issues. The primary goal of this work is to
plan a master framework for, HOCR utilizing Neural Network that can viably perceive
Odia character kind of arrangement using the Artificial Neural Network approach.

https://doi.org/10.1007/978-981-15-2774-6_12
An Implementation of Neural Network Approach … 95
Table 1. Comparision study on HCR

Author name Year Proposal for type of Methodology
character
Vats et al. [1] 2014 Offline handwritten digits Correlation
Singh et al. [2] 2012 Handwritten Hindi Back propagation neural network
character
Sayyad et al. [3] 2013 Devnagiri script Back propagation and feed-forward
neural network
Mehfuz et al. [4] 2012 Review on offline HCR Neural network approach
Borde et al. [5] 2012 Handwritten OCR Fuzzy based approach
Kala et al. [6] 2001 Review on offline HCR Advanced ANN
Dongare et al. [7] 2014 Devnagiri HCR Neural network approach
Patil et al. [8] 2012 Devnagiri HCR Neural network approach
Kaur et al. [9] 2012 Handwritten Gurumukhi Back propagation approach
character
Nohaj et al. [10] 2011 Printed English text Feed-forward network
Sharma et al. [11] 2013 English letter Neural network approach
Pradeep et al. [12] 2011 Handwritten alphabets Multilayer feedforward network
Bhattacharya 2012 Handwritten Bangla MLP, MQDF
et al. [13] character
Pal et al. [14] 2015 Handwritten Bangala Convolutional neural network
character
2 A Brief Review on Odia Script
India is a multi-lingual and multi-script nation, including diverse official languages and
has more than a hundred-local script.
Odia is, the sixth most prominent language in India derived from Indo-Aryans
language. Odia is one of the most generally spoken languages in the world. More than
35 million individuals in the eastern part of Indian talk in this language. The Odia
language derived from the ancient Brahmi script through different alterations. The
writing style of Odia language is unique in comparison with some other regional
languages in India. The modern Odia script comprises of 13 vowels, 36 consonants and
five additional primary characters and ten digits.
Following are some properties of Odia Script:
• The Writing style of Odia is from left to right.
• The idea of upper and lower case (as in English) is missing here.
• Each word in Odia script made out of a few characters isolated from one another.
• In a compound Odia character, the ‘Maatra’ is available at the top, base, left, right
or combination of any of these four places of a primary Odia character.
96 S. Dash and R. K. Das
3 Artificial Neural Network
Pattern Recognition is amazingly hard to automate. Animals are familiar with different
items and make sense of the enormous measure of visual data with minimal effort. An
Artificial Neural Network as the backend is utilized for performing characterization and
Recognition assignments. In offline CR frameworks, the Neural Network (NN) has
risen as a fast and useful tool for classification towards accomplishing high recognition.
NN is broadly categories into two types: feed-forward(FF) and feedback(FB) network.
In an FFNN, hubs organized as the input layer, one or more Hidden Layer and an
output layer stacked on one above the other. Back Propagation is a learning rule for the
preparation of multilayer FFNN. Back Propagation(BP) gets its name from the pro-
cedure of propagating the errors in the network in backward from the output layer.
For BPNN, it must be accessible to a collection of training data set and the appropriate
responses or correct explanation of the set.
In Pattern recognition frameworks, typically, there are uses two passes. The first
pass is the feature extractor that includes the extraction of specific features in a char-
acter set (for example, discovering bars of pixels inside a picture for CR). The sub-
sequent pass is the classifier that is more general purpose and can be trained using a NN
process with sample data sets. Thus, without focusing on building complicated feature
extraction and detection algorithms, this paper focuses on implementing a standard
BPNN. It also encapsulates and deals with various pre-processing stages for effective
recognition.
Following are some benefits of NN over CR:
• The application of NN in HCR brings the idea of reading multiple combinations of
handwritings
• HCR will be an effective method as evidence collection in forensic applications.
• Noise can be reduced to a large extent from the original image with the help of NN.
• Here the main aim is to recognize character with more accuracy irrespective of font
and size.
4 Design and Implementation of Model
Initially, the Algorithm was designed for the extraction of Character. Here MATLAB is
used as a tool for implementing the Algorithm. Then the NN model is developed that
would give the optimum results. There was no specific way of finding the accurate
Model of NN. It created upon trial and error basis. We have focused on two main
phases; Pre-processing and CR. In the initial phase, the pre-processing is done for a
scanned document for the extraction of individual Characters from it and normalizing
each Character. Initially, we specify an input image file, which opened for reading and
pre-processing. The image in jpeg file would be in RGB format (usually), so we
convert it into a binary format that is in grayscale format. Then use the threshold to
save and convert this grayscale image to binary, i.e., to denote all the pixels below a
certain threshold as 0 and above it as 1.
In the implementation process of recognizing handwritten Odia character, firstly the

extracted characters are fit into a window without white spaces on all the four sides then
feature vector created for each character. A database designed for the characters after
normalizing into 32 32 pixels. Window to viewport transformation used in the
normalization process. Pixel-wise mapping from the original image to the pixel of
corresponding normalized image done for every character. After the normalization of
the extracted character of the input test image, the 2-D Normalized correlation coef-
ficients approach used to identify similar patterns between a test image and the standard
database images.
Following are the steps for the proposed method:
• Input or read the scanned handwritten image document (RGB).
• Convert it to a grey scaled image.
• Pre-processing of the grey-scaled image.
• Character-wise extraction process.
• Mapping of isolated character for recognition using NN.
• Conversion of character images to text after recognition.
5 Experimental Result
For the proposed method various experiments have been performed and tested. The
experiments performed on various text images using the software Matlab (R2015a) for
Odia Characters recognition [11, 12].
Fig. 1. Scanned image for recognition

98 S. Dash and R. K. Das
Fig. 2. Recognized characters
The result obtained in generated text file (Fig. 1 and 2) is correct owing to the line
and character-wise segmentation method, which starts to read the input from the first
line of the texts to the bottom line and then from the left side to the right side. The
advantages of using this method are owing to its simplicity but can apply for recog-
nizing [13] various sets of handwriting provided through templates.
The recognition process of Handwritten Odia Character using NN has completed

effectively and successfully. The significance of the pre-processing stage in this work is
enormous and incorporates steps like Skew correction, binarization, noise elimination.
The procedure of the NN approach consists of classification along with preparing a set
of the database of character for recognizing the Odia character from the handwritten
input text. Through this strategy, we can get our results with 99% accuracy. The
percentage of accuracy of this work can improvise by including a wide variety of
handwritings that enables us to encompass a broad set of variations with the inclusion
of more massive data sets and handwriting variation.
References
1. Vats I, Singh S (2014) Offline handwritten english numerals recognition using correlation
method. Int J Eng Res Technol (IJERT) 3(6). ISSN: 2278-0181
2. Singh G, Lehri S (2012) Recognition of handwritten hindi characters using back propagation
neural network. Int J Comput Sci Inf Technol 3(4):4892–4895. ISSN 0975-9646
3. Sayyad SS, Jadhav A, Jadhav M, Miraje S, Bele P, Pandhare A (2013) Devnagiri character
recognition using neural networks. Int J Eng Innovative Technol (IJEIT) 3(1)
4. Mehfuz S, Katiyar G (2012) Intelligent systems for off-line handwritten character
recognition: a review. Int J Emerg Technol Adv Eng 2(4)
5. Borde S, Shah E, Rawat P, Patil V (2012) Fuzzy based handwritten character recognition
system. Int J Eng Res Appl (IJERA) VNCET 30 Mar’12. ISSN: 2248-9622
6. Kala R, Vazirani H, Shukla A, Tiwari R (2001) An overview of character recognition
focused on off-line handwriting. IEEE Trans Syst Man Cybern Part Appl Rev 31(2)
7. MSA Dongare, Dhananjay B. Kshirsagar, Ms. Snehal V. Waghchaure (2014) Handwritten
devanagari character recognition using neural network. IOSR J Comput Eng (IOSR-JCE)
Ume 16(2):74–79. E-Issn: 2278-0661, P-Issn: 2278-8727 (Ver. X)
8. Patil MB, Narawade V (2012) Recognition of handwritten Devnagari characters through
segmentation and artificial neural networks. Int J Eng Res Technol (IJERT) 1(6). ISSN:
2278-0181
9. Kaur M, Kumar S (2012) A recognition system for handwritten gurmukhi characters. Int J
Eng Res Technol (IJERT) 1(6). ISSN: 2278-0181
10. Nohaj M, Jaka R (2011) Image preprocessing for optical character recognition using neural
networks. Prediction Multidimension Econ Data Civilization Comput Game 30
11. Sharma N et al (2013) Recognition for handwritten english letters: a review. Int J Eng
Innovative Technol (IJEIT) 2(7)
12. Pradeep J et al (2011) Diagonal based feature extraction for handwritten alphabets
recognition system using neural network. Int J Comput Sci Inf Technol (IJCSIT) 3(1)
13. Bhattacharya U, Chaudhuri BB (2009) handwritten numeral databases of indian scripts and
multistage recognition of mixed numerals. IEEE Trans Pattern Anal Mach Intell 31(3):444–
457
14. Pal U, Sharma N, Wakabayashi T, Kimura F (2007) Handwritten numeral recognition of six
popular Indian scripts. In: IEEE Ninth international conference on document analysis and
recognition, vol 2, pp 749–753
Recognition and Extraction of Rain Drops
in an Image Using Hough Transform
A. Susmitha1(&), Lipsa Dash1, and Sunanda Alamuru2

1
New Horizon College of Engineering, Bengaluru, India
2
SreeNidhi Institute of Science and Technology, Hyderabad, India
1 Introduction
In a dynamic weather condition such as rain, it produces sharp intensity changes in

images that can considerably destroy the visibility of scene. The raindrops can refract
and reflect the scene radiance and atmospheric illumination produced a complex visual
of raindrops. Due to the refraction and reflection, the raindrops are brighter than its
background. Raindrops are randomly distributed and moving at high velocity. The
movement of raindrops with very high velocity known as rain streaks will produce
spatial and temporal intensity variation in image and video. Therefore, detecting and
removing raindrops will improve the quality of the image. To detect and remove
raindrops and then restore the image back ground K-means clustering, Hough Trans-
form and Gaussian filter are used. The k-means clustering gives maximum correct
clustering rate and Hough transform is used to extract and remove raindrops from the
image.
2 Related Work
In past few years, many methods have been proposed for the removal of the rain. The
different papers and the researchers who contributed to it are discussed below.
• Henderson [1] proposed that weather is defined as the current atmospheric condition
such as temperature, rainfall, wind, and humidity. Bad weather conditions can be
classified into two which are the static or steady weather conditions and dynamic
weather conditions. The static weather condition including fog, mist and haze
whereas the dynamic weather conditions are rain, hail and snow. Figure 1 shows the
classification of bad weather conditions.
• Narasimhan and Nayar [2] developed a model to detect bad weather conditions in
images. The detection algorithm can be classified into four; no explicit detection,
per-pixel detection, patch-based detection and frequency-based detection. Figure 2
summarizes the detection algorithm that has been used by researchers.

https://doi.org/10.1007/978-981-15-2774-6_13
Recognition and Extraction of Rain Drops in an Image … 101
Fig. 1. Bad weather classification
Fig. 2. Taxonomy of rain detection in image [3]
(a) No Explicit Detection

Some researchers did not perform the detection algorithm to remove rain from images
Narasimhan and Nayar [2]. This detection dealt with single image of mild rain. Since
the faraway image of rain was taken, the image looked like in a foggy condition. They
used the same algorithm to restore the contrast of the rainy image. Some researchers
simply applied filtering to remove the rain from the images [4, 5]. Coincidentally, this
technique will perform poorly when it is applied to the moving objects scene.
(b) Per-Pixel Detection
The pixels of an image can be divided into rain and non-rain parts. Zhang et al. implement
k-means to the grayscale intensity of each pixel to divide the pixels into the rain part and
non-rain part. Since the rains have temporal properties, some pixels are misclassified.
(c) Patch-Based Detection.
This detection detects the region of rain in an image. By applying the constraints of
photometric model in each frame in the video, Garg and Nayar [6] manage to detect the
candidate rain pixels. For a three frames static background scene, the intensities of
previous frame (In − 1) must be equal to the intensities of the next frame (In + 1) and
the change of intensities in the current frame (nth) must satisfy the constraint, c
(threshold: minimum change in intensity produced by a raindrop that detected in the
presence of noise). When this constraint of photometric model applied to a scene with
object motions, this algorithm will detect some false positives due to the movement of
102 A. Susmitha et al.
the object 6. Barnum et al. [3] found that this algorithm fails to detect individual streaks
in an image [3]. Figure 4 shows the detection of rain by Garg and Nayar [6].
(d) Frequency-Based Detection
Barnum et al. [7] used spatio-temporal frequency method to detect rain in an image.
The inverse transform of the estimated proportion of energy of rain is used to detect the
rain streak in a single frame.
• Tripathi and Mukhopadhyay [8] proposed rain removal techniques can be divided
into two main approaches which are time domain-based approach and frequency
domain-based approach. In the time domain-based approach basically considered
the properties of rain itself including the chromatic, temporal or both. Figure 3
shows the taxonomy of rain removal techniques.
Fig. 3. Taxonomy of rain removal techniques [8]
(a) Temporal Properties-Based Approach

Due to the temporal properties, rains are not remaining fixed at one time. For a static
scene, Starik and Werman [4] applied temporal median filter for each pixel to reduced
rain effects from images. Even though this technique is one of the easiest techniques to
remove rain from images, but this technique will produce blurring when applied in a
dynamic scene. By taking the advantage of temporal properties of rain, Park and Lee
[9] proposed a new algorithm which is the intensity estimation by using Kalman filter
in a real-time situation. Since this algorithm estimated that the video captured by using
a static camera with static background, therefore it fails to estimate the pixels’ intensity
in a dynamic background.
(b) Chromatic Properties-Based Approach

The chromatic properties-based approach proposed by Liu et al. (2008) deals with
colored pixel images. By using the chromatic properties, two relations between the
same pixels in two consecutive frames can be developed (Liu et al. 2008). As the
raindrop falling to the ground, it will affect the pixels’ intensity. Therefore, the first
relation is between the background pixel and the rain-affected pixel where the rain-
affected pixels are brighter than the background and there are two related rain-affected
pixels in the consecutive frames.
To restore the rain-affected pixels, the intensity of neighborhood is used to estimate
the background value. The disadvantage of this technique as stated by Tripathi and
Mukhopadhyay [8], this technique fails to differentiate between the rain pixels and the
moving object pixels.
(c) Hybrid Properties-Based Approach
Basically, a hybrid properties-based approach is the composite of both temporal and
Chromatic based approach. Due to the temporal properties of rain, the raindrops are not
fixed at a particular pixel in all frames and the chromatic properties of raindrops stated
by Zhang et al. [7] show that the intensity changes of R, G and B channels due to the
raindrops are approximately the same. The raindrop pixels are detected by applying the
K-means clustering to the intensity histogram of the video. In the process of removing
the raindrops, Zhang et al. (2006) applied dilation with Gaussian blurring to the
detected pixels and then used them as the channel. The raindrops are removed from the
videos by using-blending technique.
(d) Frequency Domain-Based Approach
This approach basically analyze the frequency information of each frame. Proposed by
Barnum et al. [3] can be applied in either single frame or multiple consecutive frames.
After the rain streaks detected by using rain detection, the removing of rain pixels can
be done by replacing them with their temporal neighbors. This method rarely removes
the rain streak completely but it will modify the brightness of rain streak by reducing
the brightness.
• Garg and Nayar [10] proposed a method by adjusting the camera parameters. Here
exposure time is increased or the depth of field is reduced. However, this method
fails to handle heavy rain and fast-moving objects which are close to the camera.
• Zhang et al. [11] proposed a method based on the chromatic and temporal prop-
erties. This method uses k-means clustering to estimate the non-rain-affected pixel
value to inpaint the rain-affected pixels. This clustering method is effective only in
removing rain from static background when there is no moving object. This method
uses all the frames available in a sequence for the removal of the rain.
• Barnum et al. [7], Barnum and Narasimhan [3] proposed a method for the detection
and removal of rain streaks by using frequency information of each frame. Here a
blurred Gaussian model is used to approximate the blurring produced by the
raindrops. This model is suitable when the rain streaks are prominent, but this
blurred Gaussian model fails to detect the rain streak when it is not sharp enough.
• Liu et al. [12] proposed a method for the removal of rain by using chromatic based
properties in rain-affected videos. It fails to detect all possible rain streaks. The
reason could be that chromatic property is not satisfied in practice as described in
the previous discussion. This method requires at least three consecutive frames for
the removal of rain.
3 Proposed Method
The proposed hybrid method is shown in Fig. 4.
Fig. 4. Block diagram of rain drop recognition and extraction
3.1 Color Image to Grayscale Image Conversion

The input image is a rainy image which is in color. Grayscale image is convenient for
processing and to make clusters. Hence the color image must be converted into
grayscale image.
3.2 Gaussian Filter

After converting the color image into grayscale image, it needs to be processed that is
the noise present in the image has to be removed. For noise removal, Gaussian filter is
used. Filtering is an operation of removing frequency components from an image
selectively. The Gaussian Filter is smoothing filter that is used to ‘blur’ images and
remove detail and noise. It is similar to the mean filter, but it uses a different kernel that
represents the shape of a Gaussian (‘bell-shaped’) hump.
3.3 K-Means Clustering Algorithm

Clustering is performed after noise removal. cluster is a collection of objects which are
“similar” between them and are “dissimilar” to the objects belonging to other clusters.
To achieve this, the k-means algorithm is used. It is the most important unsupervised
learning algorithm. It determines the best separation of observations, based on the
minimizing function from each input parameter to the cluster centroid. In our proposed
method the image is divided into clusters of 2, 3…10 to find the correct cluster number
K. In this method, the first step of k-means clustering is to randomly choose 2 (In this
case where k = 2) arbitrary means. Second, all observations are assigned to 1 of the two
clusters, based on their distance to each mean. Third, the mean of each cluster is
updated based on associated observations. Steps 2 and 3 are iteratively repeated until a
stopping criteria is hit (i.e., no observations change membership, the sum of the dis-
tances is minimized or the maximum number of iterations is reached).
Algorithm
1. First, initialize k points, called means, randomly.
2. Categorize each item to its closest mean and update the mean’s coordinates, which
are the average of.
3. the items categorized in that mean so far.
4. Repeat the process for a given number of iterations until clusters are formed.
3.4 Hough Transform

The Hough transform is a technique that is used to isolate features of a particular shape
within an image. Because it requires that the desired features be specified in some
parametric form, the classical Hough transform is most commonly used for the
detection of regular curves such as lines, circles, ellipses, etc. A generalized Hough
transform can be employed in applications where a simple analytic description of a
feature(s) is not possible. Due to the computational complexity of the generalized
Hough algorithm, we restrict the main focus of this discussion to the classical Hough
transform. Despite its domain restrictions, the classical Hough transform retains many
applications, as most manufactured parts contain feature boundaries which can be
described by regular curves. The main advantage of the Hough transform technique is
that it is tolerant of gaps in feature boundary descriptions and it is relatively unaffected
by image noise. The Hough technique is particularly useful for computing a global
description of a feature(s), given local measurements. We can use Hough transform to
detect other features with analytical descriptions. For instance, in the case of circles, the
parametric equation is (x − a)2 + (y − b)2 = r2 where a and b are the coordinates of
the center of the circle and r is the radius. In this case, the computational complication
of the algorithm initiates to increase. We research with the Hough circle detector on
raindrop image. One way of reducing the computation required to perform the Hough
transform is to make use of gradient.
4 Experimental Result
The input image is a color image. To perform clustering, input image should be in
grayscale form. So the YCBCR color space is used to convert image into grayscale
form. YCBCR gives an accurate gray component. It is also useful to reconstruct the
original RGB color image once raindrop detection and removal process are over. Once
the conversion is performed we will move towards noise removal. After noise removal,
we perform clustering using k-means clustering algorithm. Then after we use Hough
transform to detect and remove the raindrops from the single image. Once we detect
and remove the raindrop from the image, we will get the output image which is devoid
of the rain drops. The output image which we obtain through various techniques such
as RGB to YCbCr conversion, K-means clustering, Gaussian filter, Hough Transform
is in the form of a GUI (Graphical User Interface) which is as shown Figs. 5, 6 and 7.
Fig. 5. OUTPUT1
Fig. 6. OUTPUT2
Fig. 7. OUTPUT3
5 Conclusion
In this paper, we have offered a new approach of recognition and extraction of raindrop
using a hybrid algorithm. The algorithm is framed in order to recognize the rain
droplets using clustering and shape modeling of raindrops. The proposed framework is
based on K-means clustering and Gaussian filter for the efficient retrieval of rain
droplets from the rainy image. The k-means clustering results in highest correct
clustering rate and Hough transform is used to detect and remove the raindrops from
single image. The proposed system showed better results than available approaches for
raindrop recognition and deletion.
References
1. Abd Wahab MH, Su CH, Zakaria N, Abdul Salam R (2013) Review on raindrop detection
and removal in weather degraded images. In: CSIT 2013: The 5th International Conference
on Computer Science and Information Technology, 27–28 Mar 2013. Amman, Jordan
2. Narasimhan and Nayar (2003) IEEE International Conference on Computer Vision, vol 2,
pp 1067–1074
3. Barnum P, Narasimhan SG, Kanade T (2010) Analysis of rain and snow in frequency space.
Int J Comput Vis 86(2–3):256–274
4. Starik S, Werman M (2003) Simulation of rain in videos. In: Proceedings of Texture: The
3rd International Workshop on Texture Analysis and Synthesis, 17 Oct 2003. Nice, France,
pp 95–100
5. Shariah S, Marhaban MM, Ramli AR, Mohd Noor SBB (2011) Rain detection and removal
using a new algorithm. In: Proceedings of the National Geoscience Conference, 11–12 Jun
2011. The Puteri Pacific Johor Bahru, Johor, Malaysia
6. Garg K, Nayar SK (2004) Detection and removal of rain from videos. IEEE Comput Soc
Conf Comput Vis Pattern Recognit 1:528–535
7. Barnum P, Kanade T, Narasimhan SG (2007) Spatio temporal frequency analysis for
removing rain and snow from videos. In: Workshop on Photometric Analysis For Computer
Vision (PACV), in conjunction with ICCV
8. Tripathi AK, Mukhopadhyay S (2014) Removal of rain from videos: a review. SIViP
8:1421–1430
9. Park WJ, Lee KH (2008) Rain removal using Kalman filter in video. In: International
Conference on Smart Manufacturing Application, pp 494–497
10. Garg K, Nayar SK (2005) IEEE International Conference on Computer Vision, vol 2,
pp 1067–1074
11. Zhang X, Li H, Qi Y, Leow WK, Ng TK (2006) Rain removal in video by combining
temporal and chromatic properties. In: IEEE International Conference on Multimedia and
Expo
12. Liu P, Xu J, Liu J, Tang X (2009) Pixel based temporal analysis using chromatic property for
removing rain from videos. Comput Inf Sci 2(1):53–60
13. Fattal R (2008) Single image dehazing. In: SIGGRAPH
14. Halimeh JC, Roser M (2009) Raindrop detection on car windshields using geometric-
photometric environment construction and intensity—based correlation. In: IEEE Intelligent
Vehicles Symposium, 2009, pp 610–615. 978-1-4244-3504-3/09/$25.00 ©2009
15 Hassim R, Bade A (2015) Taxonomy of rain detection and rain removal techniques. Transact
Sci Technol, 28–35
16. He K, Sun J, Tang X (2008) Single image haze removal using dark channel prior. In: IEEE
Computer Society Conference on Computer Vision and Pattern recognition (CVPR)
18. Ramesh Kanthan M, Naganandini Sujatha S (2013) Automatic grayscale classification using
histogram clustering for geometric active contour models. Int J Curr Eng Technol 3(2). ISSN
2277-4106
19. Ramesh Kanthan M, Naganandini Sujatha S (2015) Rain drop detection and removal using
K-means clustering. In: 2015 IEEE International Conference on Computational Intelligence
and Computing Research. 978-1-4799-7849-6/15
20. Roser M, Geiger A (2009) Video-based raindrop detection for improved image registration.
In: 2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV
Workshops
21. Tan RT (2008) Visibility in bad weather from a single image. In: IEEE Conference on
Computer Vision and Pattern Recognition (CVPR 08)
22. You S, Tan RT, Kawakami R, Ikeuchi K (2013) Adherent raindrop detection and removal in
video. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
(CVPR)
23. You S, Tan RT, Kawakami R, Mukaigawa Y, Ikeuchi K (2014) Raindrop detection and
removal from long range trajectories. In: IEEE Conference on Computer Vision and Pattern
Recognition (CVPR)
24. You S, Tan RT, Kawakami R, Mukaigawa Y, Ikeuchi K (2016) Adherent raindrop
modelling, detection and removal in video. IEEE Trans Pattern Anal Mach Intell 38(9)
25. Webster DD (2014) Automatic rain drop detection for improved sensing in automotive
computer vision applications. M.Sc. Thesis, Cranfield University
A Novel Technique for Target Recognition
Using Multiresolution Technique
Tamanna Sahoo and Bibhuprasad Mohanty(&)
Department of Electronics and Communication Engineering, ITER, Siksha ‘O’

Anusandhan University (Deemed to be University), Bhubaneswar, India
{tamannasahoo,bibhumohanty}@soa.ac.in
1 Introduction
Long ago, computer vision area has seen a rapid development in automation which has
enabled researchers to develop models or algorithms [1] using human visual system
and its basic components as important part of the models, providing better detection or
tracking results. The main objective is the identification of specified targets in a scene
under different categories (stationary or moving) and environment (atmospheric tur-
bulence) [2]. Identification of specific targets is one of the difficult tasks with these
restriction which led to many researches and approaches such as pattern recognition,
traditional thresholding methods, wavelets [3], texture-based approaches, descriptor-
based methods, and so on [4]. This paper mainly concentrates on approaches based on
wavelet transform and its wavelet coefficient features(WCFs).
One of the major advantages of multiresolution technique (wavelet transform) is
that it can be implemented both in pyramidal or tree structure. It also provides similar
subband decomposition in consecutive scales. Further properties like rapid processing,
adaptation to changing local image statistics, exact position information and adaptation
to high background noise, are certain advantages of Discrete Wavelet Transform.
In this paper, the main objective is to identify the target object in a scene, which is
to be achieved by calculating features using gray level co-occurrence matrix(GLCM),
of detail coefficients that is subband of transformed sub-block images(non-overlapping
and adjacent blocks) of same sizes of the input image. The approach for computing
GLCM of detail subband is given by:
(i) In the first approach, GLCM of each single detail subband image is computed and
summed to get the GLCM of whole sub-block image [1].
(ii) Second, the dynamics [5] of the sub-block image are first created by applying
inverse DWT to detail coefficient and then calculating GLCM from the dynamic
image of sub-block image.
The calculation of GLCM has led to computation of combination of wavelet
coefficient features (WCFs) (cluster prominence, cluster shade, and contrast), which are
used for this paper. The highest value obtained by combining the feature values is
recognized as seed block. Further seed point is extracted from seed block which is used
in region growing process to obtain the specified target object.

https://doi.org/10.1007/978-981-15-2774-6_14
110 T. Sahoo and B. Mohanty
The remainder of the paper is prepared accordingly. In Sect. 2, the scope of

approach has been discussed. Further in Sect. 3, the proposed system via Discrete
Wavelet Transform [6] and wavelet coefficient features are described. The experimental
results are shown and the details are discussed in Sect. 4, followed by a conclusion and
forthcoming work on the proposed work has been explained in Sect. 5.
2 Scope of Approach
The core approach of this work is to develop a constructive system which will be able
to recognize the target object under diverse types of atmospheric turbulence and var-
ious kind of object. And taking the advantage of multiresolution analysis of Discrete
Wavelet Transform, it has been used for this paper.
3 Proposed Method
The steps of the proposed system are shown in figure above and it is described below.
• Here, for the division of input image of size N M into sub-blocks of 32 32 or
16 16, the input image is first resized into 512 512. Further, the input image is
converted to gray-scale image(if it is a color one) for better processing (Fig. 1).
Fig. 1. Basic building block of the proposed detected target system

A Novel Technique for Target Recognition Using Multiresolution … 111
• The division of sub-block image is carried out in such a way that it is non-
overlapping and adjacent blocks. The distinct sub-block images, starting from the
top-left corner, are decomposed into two-level two dimensional discrete wavelet
transform [7].

ðx; yÞ ¼ h ðxÞ h ðyÞFIA ðx; yÞ # 2y # 2x
ðj þ 1Þ ðjÞ
FIA ð1Þ

ðx; yÞ ¼ h ðxÞ g ðyÞ FIH ðx; yÞ # 2y 2x
ðj þ 1Þ ðjÞ
FIH ð2Þ

ðx; yÞ ¼ g ðxÞ h ðyÞ FID ðx; yÞ # 2y 2x
ðj þ 1Þ ðjÞ
FID ð3Þ

ðj þ 1Þ ðjÞ
FIV ðx; yÞ ¼ g ðxÞ g ðyÞ FIV ðx; yÞ # 2y 2x ð4Þ
where FIA, FID, FIV, and FIH are the wavelet decomposition subbands (approximation,
diagonal, vertical and horizontal coefficient).
• From the transformed coefficient, gray level co-occurrence matrix (GLCM) is
calculated. GLCM [8] is the gray level occurrence in the configuration which is
represented as matrix, consisting of relative frequencies Co (p, q, d, Ɵ). The p,
q represents the two gray level pixels and how frequently these two pixels appear in
the window in the direction Ɵ and separation between the pixels is given by a
displacement vector d. For instance, if the displacement vector is (1, 0) then in the
direction of 90 it is represented as one pixel below and zero pixel to the left [1].
Fig. 2. Co-occurrence matrix orientation. a Ɵ = 0, d = (0, 1), b Ɵ = 45, d = (1, 1), c Ɵ = 90,
d = (1, 0), d Ɵ = 135, d = (1, −1)
Fig. 3. Dynamics in wavelet domain

The figure below demonstrates the co-occurrence matrix in different orientations. In

the present work, it is derived from Ɵ = 0 and d = (0, 1) (i.e., zero pixel below and one
pixel to the right) for detail sub-bands (Figs. 2 and 3).
• The calculation of GLCM [9] is one of the major tasks of the proposed system and
is performed only on the detail sub-bands of both the levels of the sub-block image
in two ways:
(a) Method 1: It is done by taking computing GLCM of each detail coefficient
image and then combining them to provide the GLCM of sub-block image.
(b) Method 2: By creating the dynamics which is given by the formula below [5]
for each sub-block image using the formulae given below and then computing
the GLCM.
j
W 1 ¼ ½W FIA ðx; yÞ ð5Þ
• From the obtained GLCM, the evaluation of wavelet coefficient features (cluster
prominence, cluster shade, and contrast) using the formula presented below are
calculated and this is performed for each sub-block.
X
N
Contrast ¼ ðp qÞ2 Co ðp; qÞ ð6Þ
p;q¼1
X
N 3
Cluster shade ¼ p Mx þ q My Co ðp; qÞ ð7Þ
p;q¼1
X
N 4
Cluster prominence ¼ p Mx þ q My Co ðp; qÞ ð8Þ
p;q¼1
where,
X
N X
N
Mx ¼ pCoðp; qÞ; My ¼ qCo ðp; qÞ ð9Þ
p;q¼1 p;q¼1
Contrast feature values are acquired by normalizing linearly whereas cluster

prominence and cluster shade are acquired by normalizing logarithmically, depending
upon the dynamic ranges of the features.
• The next step is the selection of seed block which is the sub-block image having
maximum of combined normalized feature value (cluster prominence, cluster shade,
and contrast). The concept behind the selection of seed block is that the high value
of wavelet coefficient feature represents that it is a part of the target.
The seed block is used along with region growing algorithm for the identification of
target object. Region growing is one of the segmentation process (based on region),
usually depends on some predefined criteria (adjacency and threshold) sub-regions are
grown into larger regions. In the present work, seed point is chosen as the criteria
where the center point of the seed block [2]. Further, in the process, a comparison of all
the unallocated neighboring pixels is created by differencing the pixel’s intensity value
with the region’s mean(seed point) to grow the region. This can be used as a measure of
similarity. The smallest pixel difference measured during the process is assigned to the
respective region. The method’s continuity depends on intensity comparison between
region mean and new pixel and thresholding.
3.1 Algorithm of Proposed System
Input: Input image(size: M N)

Output: Detected target image
(a) Read the input image and resize it to 512 512.
(b) Obtain 32 32 sub-block image, initiating from the top-left corner.
(c) Decompose the sub-block images using 2D-DWT.
(d) Derive the Gray level co-occurrence matrix (GLCM) for detail subbands of
transformed sub-block images using both the methods.
(e) Calculate and combine features (contrast, cluster prominence, and cluster shade)
in wavelet domain, from GLCM.
(f) Reiterate the steps (c) to (e) for each sub-block images.
(g) Select a sub-block image as seed block (having maximum of combined wavelet
coefficient feature value).
(h) Apply seed point (center of seed block) in region growing algorithm to obtain the
target object.
In the previous section, the proposed algorithm has been discussed and applied on
natural single-target image and the image chosen is having a clear background. Further,
the DWT decomposition level used for the purpose is two-level and filter used is haar
filter. In the region growing technique, the mean distance method using seed point is
used which is obtained from the seed block. The detailed information of single-target
image (a xylophone image) has been presented below.
Original image size = 360 240 3
Resized image size = 512 512
Method 1
Seed block = 105th sub-image block,
Seed points = (208,272),
Sum of WCF = 7.4292e + 03.
Method 2
Seed block = 89th sub-image block,
Seed points = (176,272),
Sum of WCF = 2.2956e + 03.
Fig. 4. a Input image. b Seed blocks of methods: 1(left) and 2(right). c Region grown image of
methods: 1(left) and 2(right). d Highlighted image of methods: 1(left) and 2(right)
The results of proposed methods are shown in Fig. 4 where (a) shows the original
resized image of 512 512 while (b), (c) shows the seed blocks obtained from two
methods of wavelet coefficient feature values, (d) shows the image after region growing
process using the seed points value and (e) represents the highlighted image.
5 Conclusion
Wavelet-based approach using dynamics for finding the coefficient features and seed
block in the proposed algorithm is found to be convincing with respect to visual
inspection in a single target image [10]. The proposed method can be extended for
multiple targets which can be either stationary or moving and with cluttered back-
grounds. In future, this proposed technique in wavelet domain can be extended at
different steps of visual attention based moving object detection which is presently our
research area [11].
References
1. Arivazhagan S, Ganesan L (2004) Automatic target detection using wavelet transform.
EURASIP J Appl Sig Process 17:2663–2674
2. Pulla Rao C, Guruva Reddy A., Rama Rao CB (2016) Target detection using multi
resolution analysis for camouflaged images. Int J Cybern Inf 5(4)
3. Espinal F, Huntsberger TL, Jawerth BD, Kubota T (1998) Wavelet-based fractal signature
analysis for automatic target recognition. Opt Eng 37(1):166–174
4. Mahmoodabadi SZ, Ahmadian A, Abolhasani MD, Eslami M, Bidgoli JH (2005) Feature
Extraction based on multiresolution wavelet transform. In: IEEE Proceedings, pp 3902–3905
5. Sahoo T, Mohanty B (2018, September) Moving object detection using background
subtraction in wavelet domain. In: 2nd International Conference on Data Science and
Business Analytics (ICDSBA)
6. Gomes JPP, Brancalion JFB, Fernandes D (2008) Automatic target recognition in synthetic
aperture radar image using multiresolution analysis and classifiers combination. IEEE
7. Arivazhagan S, Ganesan L (2003) Texture classification using wavelet transform. Pattern
Recogn Lett 24(9–10):1513–1521
8. Hazra D (2011) Texture recognition with combined GLCM, wavelet and rotated wavelet
features. Int J Comput Electr Eng 3(1):17938163
9. Shan Z, Aviyente S (2005) Image denoising based on the wavelet co-occurrence matrix.
ICASSP2005, IEEE, pp 645–648
10. Pan Y, Chen Y, Fu Q, Zhang P, Xu X (2011) Study on the Camouflaged target detection
method based on 3D convexity. Proc Modern Appl Sci 5:152–157
11. Singh SK, Dhawale CA, Misra S (2013) Survey of object detection methods in camouflaged
image. In: Proceedings of IERI, Elsevier, pp 1–6
A Novel Technique of Shadow Detection Using
Color Invariant Technique
Leeza Panda and Bibhuprasad Mohanty(&)

Anusandhan University (Deemed to be University),
Bhubaneswar, India
{leezapanda,bibhumohanty}@soa.ac.in
1 Introduction
When object occludes light from a source of light which appears as surface edges,
shadows occur in the image. Shadow detection and removal over the past decades
cover many specific applications such as traffic surveillance [1, 2], face recognition [3–
5] and image segmentation. Object shadow detection has been an active field of
research for several decades. A shadow in an image has become a serious issue as it
causes image distortion, object merging, object loss, etc. Based on the intensity of light,
shadows are classified into two types: cast shadow and self-shadow. Cast shadows
occur when an object is projected and a dark area forms in the direction of light
whereas self-shadows are a certain example of cast shadow where the shadow of the
body is extrapolated onto itself. Self-shadows are indefinite which do not have a
prominent edge whereas cast shadows have sub-region like shading and interreflection
(Fig. 1).
Images with shadows can decrease the performance of certain computer vision
applications such as object detection, tracking, surveillance, etc. Hence it is necessary
to find out the methods for shadow detection and removal, and moreover, it is also
important to keep the original information intact while removal.
2 Brief Literature Review
Finlayson et al. [6] came up with a method where the original image with illumination
invariant image recognizes the shadow features. Firstly the features are set to zero and
then the representation of edges is done to get a shadowless image. During reinte-
gration, error propagation can be reduced by closing the shadow features before the
reintegration process takes place. This process is proved by Fredembach and Finlayson
[7]. Lalonde et al. [8] proposed a method where geometrical features are being matched
which detects shadow with good accuracy. One of the challenging problems here is
when the shadow isn’t on the ground. It appears in larger variation when the shadow

https://doi.org/10.1007/978-981-15-2774-6_15
A Novel Technique of Shadow Detection … 117
Fig. 1. Shadow occurrence illustration
isn’t on the ground. Wu and Tang [9] proposed a method to detect vague shadows in an
image using derivatives of the input image. A method to remove the shadows from
curved areas retaining the background texture is proposed in [9]. The removal of
shadows is achieved by calculating different scale factors for shadow regions and
penumbra regions to cancel the effect of shadows.
Shadows detection can be categorized into five algorithms:
1. Invariant based shadow detection: 2D chromaticity projection was used along a line
upright in the change in direction of illumination. This was used to obtain a 1D
grayscale illumination invariant representation of an image which was the most
prominent contribution of Finlayson et al. [7]. Edge detection was performed both
on illumination invariant and original images. The shadow edges were the result of
a difference in edge-maps.
2. Feature-based shadow detection: An algorithm that started with a hypothesis testing
over a shadowed area on which they are formed was modeled by Salvador et al.
[10]. Detection of Shadow pixels in HIS, YCbCr color spaces and hue difference of
background and foreground regions were confirmed by the algorithm. Edge
information was used for the detection using Sobel operator and subsequently
Boolean AND was used to combine the detection results.
3. Region-based shadow detection: A region-based technique to detect shadows from
natural scenes was modeled by Guo et al. [11]. A single region classifier incor-
porating SVM with a kernel and a pairwise classifier incorporating SVM with RBF
kernel was trained by them. An energy function which amalgamated the single
region classifier and pairwise classifier predictions were then minimized by using
graph cut. Regions of the same material with the same illumination and different
illumination were used as their pair classifiers.
4. Color model-based shadow detection: Salvador et al.’s [12] contribution to this
category is one of the earliest ones. The generation of object contours was achieved
using an edge map of the c1c2c3 color invariant of the actual image. As the
luminance image is affected by shadows, dark regions that are shadow candidates
are extracted using the edge map of a luminance image.
118 L. Panda and B. Mohanty
5. Interactive-based shadow detection: A shadow extraction technique based on quad

map input by the user was proposed by Wu and Tang [9]. The quad map neces-
sitated a user to mark four different regions in an image: shadow, non-shadow,
uncertain and excluded regions. The input was subsequently used in a Bayesian
framework to extract the shadow.
Shor and Lischinski proposed one of the simplest interactive methods in which a
bare mouse click by the user marked the shadow region [13]. The shadow mask was
then derived from the clicked area by iterative region growing and an illumination
invariant distant measure. The technique worked only for a surface that had both
shadowed and non-shadowed areas.
The process of Shadow detection encompasses two significant affirmations—de-
tecting the shadowed region and re-establishing the illumination in the shadowed area.
Some degree of image understanding is involved in the detection task which is used to
determine whether a pixel is dark due to a shadow or caused due to the reflection at the
concerned scene point. Accomplishing this task mandates some assumptions about the
shadowed surfaces in the scene and expecting some hints from the user. The illumi-
nation restoration task is also appreciably challenging, as it attempts to remove any
noticeable differences between the originally lit and the restored parts of the images. To
be more particular, it is difficult to avoid the differences in local contrast and the
amount of noise between the original and restored areas.
3 Methodology
The concept of invariant image is based on the theory that the shadows appear in lower
light condition areas. This illumination invariant technique finds a depiction image
where shadows don’t exist. But two conditions need to be satisfied; the original image
must be noiseless and uncompressed. The edge map which is created from the original
image by the edge detection process can also hamper in the shadow detection process.
The illumination invariant algorithm for shadow detection which consists of five parts:
Image preprocessing which includes conversion of RGB image with the application of
median filter followed by the calculation of shadow ratio and the threshold value. After
the threshold value is calculated, a shadow mask is created. After the mask is been
created the area size threshold is adjusted or tuned followed by dilation of the grayscale
scale image which results in shadow boundaries. The RGB to greyscale conversion is
given by the formula

max½FR FG FB
Y ¼ log ð1Þ
min½FR FG FB þ 1
4 Results and Analysis
The above technique has been tested in Matlab R2016b. Figure 2 shows original test
images with the presence of shadows. Figure 3 shows the converted binary image.
Figure 4 shows us the binary image with shadow detected followed by Fig. 5 RGB
images with shadow detected.
Fig. 2. Flowchart representation for shadow detection
The above method has been implemented on Matlab R2016b and the illumination
invariant method has been used to detect the shadows in the images.
120 L. Panda and B. Mohanty
Fig. 3. Original images with the presence of shadows
Fig. 4. Binary images with shadows detected
Fig. 5. RGB images with shadows detected
5 Conclusion and Future Work
This paper carries forward the various works done in the field of shadow detection.
Various methodologies can be adopted for the determination of the contours of the
shadow and differentiating the shadows with the nearby objects. The experiment
illustrated the flow algorithms and illustrations for shadow detection. The illumination
invariant method was used for shadow detection. The method takes into account the
shadow ratios, calculates the shadow masks and finally dilates the grayscale binary
image of the shadow which in turn results in shadow boundaries. The illustrations and
algorithms can be further enhanced and utilized to work on automatic driver assistance.
The functionality of shadow detection being a very significant part of lane detection in
automatic driver assistance.
References
1. Wang JM, Chung YC, Chang CL, Chen SW (2004) Shadow detection and removal for traffic
images. In: 2004 IEEE international conference on networking, sensing and control, vol 1,
21–23 Mar 2004, pp 649–654
2. Bevilacqua A (2003) Effective shadow detection in traffic monitoring applications. WSCG
3. Chen T, Yin W, Zhou XS, Comaniciu D, Huang TS (2005) Illumination normalization for
face recognition and uneven background correction using total variation based image
models. CVPR 2(2005):532–539
4. Adini Y, Moses Y, Ullman S (1997) Face recognition: the problem of compensating for
changes in illumination direction. IEEE Trans Pattern Anal Mach Intell 19(7):721–732
5. Zhao W, Chellappa R (1999) Robust face recognition using symmetric shape-from-shading.
Technical report. Center for Automation Research, University of Maryland
6. Finlayson GD, Hordley SD, Drew MS (2002) Removing shadows from images. In:
Proceedings of the 7th European conference on computer vision, pp 823–836
7. Fredembach C, Finlayson, GD (2004) Fast re-integration of shadow free images. In:
Proceeding color imaging conference
8. Lalonde J-F, Efros AA, Narasimhan SG (2010) Detecting ground shadows in outdoor
consumer photographs. In: Proceeding 11th European conference computer vision
9. Wu TP, Tang CK (2005) A Bayesian approach for shadow extraction from a single image.
In: Proceedings of the IEEE international conference on computer vision, Beijing, China, vol
1, 17–21 Oct 2005, pp 480–487
10. Salvador E, Cavallaro A, Ebrahimi T (2004) Cast shadow segmentation using invariant color
features. Comput Vis Image Underst 95(2):238–259
11. Guo R, Dai Q, Hoiem D (2013) Paired regions for shadow detection and removal. IEEE
Trans Pattern Anal Mach Intell 35(12):2956–2967
12. Salvador E, Cavallaro A, Ebrahimi T (2001) Shadow identification and classification using
invariant color models. In: Proceedings of IEEE international conference on acoustics,
speech, and signal processing, Salt Lake City, USA, vol 3, 7–11 May 2001, pp 1545–1548
13. Shor Y, Lischinski D (2008) The shadow meets the mask: pyramid-based shadow removal.
Comput Graph Forum 27(2):577–586
Comparison of Performance Metrics of Star
Topology and Ring Topology in Wireless
Sensor Network
Laxmipriya Moharana(&), Bedanta Kumar Biswal, Raman Raj,

and Suraj Naik
Electronics and Communication Engineering, Faculty of Engineering, Siksha

‘O’Anusandhan (Deemed to be University), Bhubaneswar, India
{laxmimoharana,surajnaik}@soa.ac.in
1 Introduction
In today’s world, wireless sensor networks (WSNs) are an emerging area of wireless
computing, gathering information, and networking. A WSN comprises of a collection of
nodes organized into a cooperative network. The recent advancement in WSN is
observed in the field of railways for safety and security purposes. [1–4]. A sensor node
consists of an ADC, Microcontroller, Memory, power device and a Transceiver as in
Fig. 1 [5]. Energy conservation to be maintained within a wireless sensor network to
improve the lifetime of sensor nodes [6]. To decrease the energy consumption and to
improve the lifetime of a sensor network, the sensor deployment in a certain geometrical
area to be taken care of by implementing different topology criteria. The well-suited
topology is to be used for sensor deployment. In this paper, we have proposed two
different topology models i.e. star topology and ring topology and calculated and plotted
results of throughput of sensor nodes, average throughput, delay and jitter. The result
shows a detailed comparison between two topologies which comprises of six sensor
nodes in WSN connected in wireless medium based upon the above-mentioned matrices.
Fig. 1. Basic block diagram of a sensor node

https://doi.org/10.1007/978-981-15-2774-6_16
Comparison of Performance Metrics of Star Topology … 123
2 Literature Survey
A model is to be initiated to observe the in-network computation in a sensing field and

to find the optimum achievable throughput. Single path routing technique is helpful to
achieve such an optimal throughput. Another important aspect of sensor networks is to
maximize the lifetime of a sensor node [7]. By using matrices such as overall network
throughput and total energy consumption, the performance of topology control algo-
rithm can be frame worked [8]. An efficient topology may be designed such that the
energy consumed by each and every sensor nodes of a sensor network is to be econ-
omized; a possible way to achieve that is by coverage maintenance and graph con-
nectivity [9]. If we analyze the energy aware routing schemes by adopting different
topologies then we can say that Routing Information Protocol (RIP) is more energy
efficient compared to Intrazone Routing Protocol (IARP) and STAR [10].when we
compare the quality of service of star topology with multi-hop topology outdoor
propagation we can find that multi hope topology is superior than star topology in
certain constraints [11].
3 Testing, Analysis, and Evaluation
Each node is represented with green color as shown in Fig. 2, which is placed as per
the configuration provided in the NS-2 tcl file. The nodes are placed with a wireless
connection set up having some specific node parameters and are connected in space
with specific co-ordinates.
Fig. 2. Scenario of ring and star topology in NS 2 visual trace analyzer
In ring topology data packets are sent around the circular path until the packets will
be their final destination. 1st part of Fig. 3 represents a ring network formed by ns-2.35.
Here the source is node-2 and the destination is node-4. So, packets are transmitted
from node-2 to node-4. In either direction, the packets can travel to its destination.
When a node will be out of order during data transmission, data packets will be traveled
124 L. Moharana et al.
in another route to reach their destination for that this topology is called self-healing
network. In the Star Topology, a central connection point is there called a hub which is
a computer hub or sometimes just a switch. 2nd part of Fig. 3 represents a star network.
Here the hub is node-0 and from node-2 packets are transmitted to node-4 through
node-0. In a Star Network, a failure occurs in the cable one computer be affected but the
entire network will be safe. But when any problem arises to hub, the whole system
will fail.
Fig. 3. Scenario of ring and star topology of NAM file in NS 2.35 with simulation time of 1.4 s
(approx.)
As data transmission took place from node 2 to node 4, the node 2 is considered as
source node and node 4 as destination node, the trace file of ns2 code is analyzed to
calculate the throughput, packet delay and packet jitter at node-4. As we know
throughput is defined as a measure of how fast we actually send data through a network
without any loss. It is the calculation of no. of bits received without any error per
sec. Here it is calculated as (no. of bits received + hold rate)/(data rate sampling
time). It can also be calculated as Size of the packets/Transmission time, so unit of
throughput is expressed as bits/s. Delay can be termed as the total time data packets
need to transfer from the source node to the destination node. The sources of delay in
WSN are source and destination processing delay, propagation delay and network
delay. The given formula can be used to calculate the average Delay, where Arrival
time is the time when packet “i” reaches the destination and Start time is the time when
packet “i” leaves the source. “n” is the total number of packets. Average delay =
(Arrival Time − Start Time)/n. Average jitter for data transmission in a WSN can be
defined as the fluctuation in the delay introduced by the different devices along the
communication path. It is the variation in the time between packets arriving. Jitter is
generally associated with network consistency and stability. Average Jitter = ((Arrival
Time + 1) − (Start Time + 1)) − ((Arrival Time) − (Start Time))/n − 1 [12].
3.1 Analysis for Star and Ring Topology

The topology is formed to transform data from node 2 to node 4. In ns2 visual trace
analyzer, the star topology and ring topology was formed by taking 5 nodes with
certain co-ordination that is shown in Fig. 2, also in ns2 the NAM (network animation)
file was generated to show the formation of Star and ring topology and the corre-
sponding data transmission shown in Fig. 3. Here we were assuming data transmission
is taken place from node 2 to node 4 through node 0, since node 0 is considered as the
central hub for star topology and for Ring topology data packets are transmitted from
node 2 to node 4 through node 1 and node 0.
3.2 Analysis of Star Topology
Node 0 Statistics
By referring to the second part of Fig. 3 i.e. a star topology throughput, packet delay and
jitter are calculated at node 0. Total number of packets are transmitted from node 2 are
177 with a data rate of 93 Kbps. Through the transmission path, the total packets dropped
are 23 with a data rate of 12 Kbps. By referring Fig. 4 we got minimum packet delay is
0.055531097 ms and maximum packet delay is 0.40498733 ms. So, the average delay at
node 0 is calculated as 0.28157563 ms. From Fig. 5 we got the packet jitter values as
follows: minimum jitter is 0.00002 ms, maximum packet jitter is 0.22034536 ms. So,
the average packet jitter is 0.11018268 ms. Figure 6 shows the received throughput
at node 0 the statistic is as follows: minimum throughput is 6000 bytes/s, maximum
throughput is 42,000 bytes/s. So, the average throughput is 24,000 bytes/s.
Fig. 4. Delay over time at Node 0-star

Fig. 5. Jitter over time at Node 0-star
Fig. 6. Throughput received at Node 0-star

Node 4 Statistics
From node 0 the transmitted packets are 101 with 53 Kbps data rate, in between
the transmission the total packets dropped are 9 with 5 Kbps data rate. By referring
Figs. 7, 8 and 9.
Fig. 7. Delay over time at Node 4-star
Fig. 8. Jitter over time at Node 4-star

Fig. 9. Throughput received at Node 4-satr
The destination node i.e. the node 4 having the packet delay, packet jitter and
throughput statistics as follows: minimum packet delay is 0.039644157 ms, maximum
packet delay is 0.4034061 ms and the average packet delay is 0.28857745 ms. Mini-
mum packets jitter is 0.00018 ms, maximum jitter is 0.288336 ms. and the average
jitter is 0.144258 ms. Minimum received throughput at node 4 is 6000 bytes/s, max-
imum throughput 52,000 bytes/s and average throughput is 29,000 bytes/s.
3.3 Analysis of Ring Topology
Node 1 Statistics
By following part 1 of Fig. 3 the ring topology different performance metrics are
calculated at nodes 1, 0 and 4. which are given below. From node 2 the packets are
transmitted are 58 with a data rate 86 Kbps, then the total dropped packets at node 1 are
7 with data rate 9 Kbps. At node 1 Minimum packet delay is 0.12554 ms, maximum
packet delay is 1.1548 ms and the average delay is 0.64017 ms, minimum jitter is
0.00008 ms, maximum jitter is 0.62123 ms and average jitter is 0.3106 ms, minimum
throughput is 500 bytes/s, maximum throughput is 8000 bytes/s and average through is
4250 bytes/s as shown in Figs. 10, 11 and 12.
Fig. 10. Delay over time at Node 1-ring
Fig. 11. Jitter over time at Node 1-ring

Fig. 12. Throughput received at Node 1-ring
Node 0 Statistics
By analyzing the plots from Figs. 13, 14 and 15. From node 1 the packets are trans-
mitted are 102 with a data rate 152 Kbps. then the total dropped packet at node 0 is 1
with data rate 2 Kbps. At node 0 Minimum packet delay is 0.01388597 ms, maximum


packet delay is 1.1409631 ms and the average delay is 0.577424 ms, minimum jitter is
0.00006 ms, maximum jitter is 0.19961004 ms and average jitter is 0.099835 ms,
minimum throughput is 2000 bytes/s, maximum throughput is 15,000 bytes/s and
average through is 8500 bytes/s.
Node 4 Statistics
By analyzing the plots from Figs. 16, 17 and 18. From node 0 the packets are trans-
mitted are 11 with a data rate 15 Kbps. then the total dropped packets at node 4 are 2
with data rate 3 Kbps. At node 4 Minimum packet delay is 0.735275 ms, maximum
packet delay is 1.544114 ms and the average delay is 1.139694 ms, minimum jitter
0.016051866 ms, maximum jitter is 0.60370114 ms and average jitter is
0.3098765 ms, minimum throughput is 100 bytes/s, maximum throughput is 3200
bytes/s and average through is 1650 bytes/s.


4 Conclusion
Wireless Sensor Networks (WSN) promises a new domain on the way computers and
humans interact with our environment and play a vital role in area of new research and
development, invention and investigations. From the above topology design and
analysis, we concluded that the performance metrics of star topology is much better
than ring topology. The packets sent, received, throughput, end to end delay, jitter was
much higher in star topology. In real life scenario, we can deploy the sensor nodes by
forming ring topology because here each node can be treated as the normal node as
well as a hub so it is lowering the cost of implementation.
References
1. Hodge VJ, O’Keefe S, Weeks M, Moulds A (2015) Wireless sensor networks for condition
monitoring in the railway industry: a survey. IEEE Trans Intell Transp Syst 16(3):1088–1106
2. Nuffer J, Bein T (2006) Application of piezoelectric materials in transportation industry. In:
Proceedings of Global symposium on innovative solutions for the advancement of the
transport industry, San Sebastian, Spain, pp 1–11
3. Pourghodrat A (2011) Energy harvesting systems design for railroad safety. M.S. thesis,
Department of Mechanical and Materials and Engineering, University of Nebraska–Lincoln,
Lincoln, NE, USA
4. Lynch JP, Loh KJ (2006) A summary review of wireless sensors and sensor networks for
structural health monitoring. Shock Vibrat Dig 38(2):91–130
5. Tolani M, Singh SRK, Shubham K, Kumar R (2017) Two-layer optimized railway
monitoring system using Wi-Fi and zigbee interfaced wireless sensor network. IEEE Sens J
17(7):2241–2248
6. Nguyen T-T, Shieh C-S, Dao T-K, Wu J-S, Hu Robot W-C (2013) Prolonging of the
network lifetime of WSN using fuzzy clustering topology. In: Second international
conference on vision and signal processing (RVSP)
7. Sapidi R, Girard A (2013) Maximum achievable throughput in a wireless sensor network using
in network computation for statistical functions. IEEE/ACM Trans Netw 21(5):1581–1594
8. Sinha P (2003) Throughput and energy efficiency in topology-controlled multi-hop wireless
sensor networks. In: 2nd Second ACM international workshop on wireless sensor networks
and applications (WSNA), San Diego, California, USA
9. Santha Meena S, Manikandan J (2017) Study and evaluation of different topologies in
wireless sensor network. In: International conference on wireless communications, signal
processing and networking (WiSPNET), Chennai, India, March 2017
10. Alam S, De D, Ray A (2015) Analysis of energy consumption for IARP, RIP and STAR
routing protocols in wireless sensor networks. In: Second international conference on
advances in computing and communication engineering, Dehradun, India, May 2015
11. Pramono S, Putri AO, Warsito E, Budi Basuki S (2017) Comparative analysis of star
topology and multihop topology outdoor propagation based on quality of service (QoS) of
wireless sensor network (WSN). In: IEEE international conference on communication,
networks and satellite (Comnetsat), Semarang, Indonesia, Oct 2017
12. Birla J, Sah B (2012) Performance metrics in ad-hoc network. IJLTET 1(1)
A Framework for Real-Time Lane Detection
Using Spatial Modelling of Road Surfaces
Pankaj Prusty and Bibhuprasad Mohanty(&)
Electronics and Communication Engineering, Faculty of Engineering, ITER,

Siksha ‘O’ Anusandhan (Deemed to be University), Bhubaneswar, India
{pankajprusty,bibhumohanty}@soa.ac.in
1 Introduction
In this paper, a robust and novel method using spatial neighbourhood properties of non-
lane pixels and comparing the testing pixels with these pixels for detection of lane fol-
lowed by morphological operations has been proposed. This method works perfect for
various types of lane markings like straight, curved, continuous and discontinuous, etc.
Lane detection basically is a method of identifying the lane markers on a road. For
automatic driver assistance systems, lane mark is an important and essential part. The
objective of lane mark detection is to give a warning to the driver if the vehicle is
approaching towards the lane mark or near the lane mark. Nowadays because of highly
increase in traffic many accidents are caused due to lane departure and many of these
lane departures are due to drivers fault. So therefore, there is a high need for research to
develop robust systems for automated detection of lane marks. Lanes are of various
types which include road surface markings, zebra crossings, arrow markings, etc. Our
system is robust in capturing all these types of lane markings. The importance of lane
detection is it makes the system intelligent and can be a good part of driverless system.
But the basic requirement for this is that it should be accurate, robust and capable of all
types of lane detection.
Our spatial modelling of background algorithm works well for continuous and
discontinuous lane marking in real-time. Here we do not require any prior information
about the road surfaces and also no data base is required here. All works are done
intensity plane. Less memory requirement and no need for data base make the system
robust. Our algorithm is capable of detecting the lane immediately when the frame
arrives.
Previous work has been discussed in Sect. 2. Section 3 contains our proposed
method followed by experimental results in Sect. 4.

https://doi.org/10.1007/978-981-15-2774-6_17
136 P. Prusty and B. Mohanty
2 Literature Survey
Basic image features like histogram, edges, and gradients play a measure role in the
lane detection process which is based on vision. Supervised learning methods have
been used in papers [1–5]. In [3] a system is proposed which calculates the departure
degree from lane in correspondence with the boundaries angular relationship. They
have used the shadow of the front vehicle to locate the position of the vehicle in front of
them. Then they have used the neuro-fuzzy network to estimate the actual distance. The
authors of [6] have been proposed a spatiotemporal method with curve fitting prop-
erties. Authors of [4] have been proposed a method where they have used a base
classifier based on some properties of the environment. The system does not work in
dark shadow region, reflector region etc. Authors of [5] have combined the self-
clustering algorithm, fuzzy C-mean and fuzzy rules to improve the performance in
various light conditions. Author [7], in their paper, used Hough transformation along
with Kalman filtering method for 3° Bezier splines.
3 Proposed Work
In this section, we discuss our method for detection of lane mark from road surfaces.
A camera is mounted on the bonnet of a vehicle and it is assumed that this camera is
capable of sending the captured video frames to a computer which is on board. We
apply our spatial road surface modelling algorithm and detection method. First, the
road surface background is modelled by considering the non-lane pixels and then the
lane is detected by comparing the lane frame with the background.
3.1 Region of Interest Selection

We first select the region of interest in order to reduce the miss and false classification
rate. This makes our system faster. We select that part of the image which contains the
road surfaces only not the sky, pedestrian, etc., since these parts do not provide any
meaningful information to us. The region of selection can be better if the placement of
the camera is fixed and known.
3.2 Gray-Scaling of the Image

Then the region of interest is converted to grayscale where each pixel represents the
intensity of the original image. Working on gray scale will help us in analyzing the
variation in intensity which makes the system robust.
3.3 Spatial Modelling of Road Surfaces

As the statistical correlation between neighbouring pixels of an image is nearly equal,
so we have to consider the spatial neighbours of a pixel where no lane mark is there to
model the road surfaces. Initially, we have to assume that the first frame does not
A Framework for Real-Time Lane Detection … 137
contain any lane mark to model the road surfaces. Also, we can model the road surface
by considering the non-lane areas. The process of modelling the road surfaces is done
as per the procedure explained below.
P ¼ fq1 ; q2 ; q3 . . .qN g ð1Þ
where qi is the randomly selected N samples from the spatial n n neighbourhood of

pixel with location (x, y). We will apply this concept in the entire image so that we have
a set of pixels Q.

S ¼ P1 ; P2 ; P3 ; . . .; PS ð2Þ
where S represents the total number of pixels in the image. Choosing the value of
n more than five may degrade the performance of the algorithm due to the fact that the
more we go away from the pixel the resemblance between pixels decreases. N is taken
to be nearly 80% samples from n n. The model gets updated or new model is created
when the detected frame has no lane marks at all and discards all the previously stored
data values which make the system robust.
3.4 Lane Mark Identification

As explained in Eq. (2), we have now a spatial model of the road surfaces. We want to
classify a new pixel value in respect to its neighbouring value that whether this is the
background or lane mark. A new pixel is now compared to the background samples
and if it is a non-lane pixel then its value must be close to some of the values of the
corresponding spatial model of Pi. The classification of the pixel is done as per the
following equations.
X ¼ jtestðx; yÞ smðx; yÞj [ Th1; for all N ð3Þ
where ‘test’ is the frame for testing with pixel location (x, y) and ‘sm’ is the spatial
modelled intensities of location (x, y). N represents the number of intensity values
present for the location (x, y) in ‘sm’ and ‘Th1’ is empirically decided. Experimentally
it has been found that the value of ‘Th1’ is 150. Once all the value of X for location
(x, y) is calculated, then its cardinality is calculated.
The decision is taken as per the following equation. If the following equation is
satisfied then the pixel with location (x, y) is classified as non-lane pixel else it is
classified as lane pixel.
j X j Th2 ð4Þ
where |.| represents the cardinality of a set and Th2 is to be empirically calculated. It has
been found that 4 or more value of Th2 shows good results with least variation.
4 Experimental Results
Our proposed algorithm has been implemented in MATLAB 2016a on Windows 10

operating system. The implemented algorithm’s performance has been evaluated on
two videos shoot in a Nikon camera mounted on the hood of a vehicle and one video
from YouTube [3]. Our algorithm was run on a laptop with Intel Corei3 3110 M Cpu
with 2.4 GHz clock speed and 4 GB RAM. Figure 1 shows the videos used for
simulation.
Fig. 1. Videos used for simulation
First, two videos are taken from a camera mounted on the hood of a vehicle. Here
the camera is mounted in such a way that it focuses the road surfaces and lane only.
Whereas the third video is a YouTube video where the scene contains road surfaces,
sky, other vehicles, etc. The region of interest needs to be identified before proceeding
towards the algorithm. Figure 2 shows one frame of the YouTube video and its rep-
resentation after ROI selection.
Fig. 2. (Left) Frame 6 from original video (right) Same frame after ROI selection and grayscale
conversion
First row of Fig. 3 is a video taken by a camera mounted on the hood of a vehicle.
For simulation purpose, each frame is converted to 256 256 pixels. The result
obtained is compared with STIC method [6]. The morphological operations are then
applied on the detected lane pixels to give a clear boundary. Second row represents
frame 6 of a video downloaded from YouTube. In this video, we have no prior
information about the road surface and camera mounting condition. We simply
downloaded it as it is from a moving video and simulated it. The results obtained are
shown in Fig. 3.
A Framework for Real-Time Lane Detection … 139
Fig. 3. Results of Lane detection of two videos (first row: video taken by us, second row: video
from YouTube), From left to right of each row: grayscale of the frame, result after lane detection,
result after morphological operation
To compare the performance of the algorithm, we have taken the help of four
parameters: true positive (TP) and false positive (FP). Where true positive is the
number of correctly classified pixels and false positive is the number of pixels falsely
classified as lane. Table 1 represents the performance of the proposed method with the
STIC method.
Table 1. Comparison between the STIC method and proposed method with the following
performance metrics
Videos STIC method Proposed method
TP (%) FP (%) TP (%) FP (%)
Video 1 (taken from a camera) 93 5 97 2.3
Video 2 (taken from a camera) 94 4.23 98.2 0.7
Video 3 (YouTube video) 92 3.04 91.89 4.56
5 Conclusion
We have proposed a real-time lane detection method based on the concept of spatial
modelling method. The system is robust because it does not require any preexisting
data base and also less memory requirement is there. All the works have been done in
intensity plane. The method is capable of detecting various types of lane marks like
dotted discontinuous, etc. The results show that the proposed method is robust and
efficient in comparison to other methods. The method does not consider the poor light
condition, which can be the future scope of the work.
References
1. Zhou S, Lagnemma K (2010) Self-supervised learning method for unstructured road detection
using fuzzy support vector machine. In: IEEE international conference on Intelligent robots
and systems (IROS)
2. Zhou S, Gong J, Xiong G, Lagnemma L (2010) Road detection using support vector machine
based on online learning and evaluation. In: Intelligent vehicle symposium (IV)
3. Wu C, Lin C, Lee C (2012) Applying a functional neuro fuzzy network to realtime lane
detection and front vehicle distance measurement. IEEE Trans Syst Man Cybern 42(4)
4. Kuhnl T, Kummert F, Fritsch J (2012) Spatial ray features for real time ego-lane extraction’.
IEEE Trans Intell Transp Syst USA, 16–19 Sept 16–19
5. Wang J, Lin C, Chen S (2010) Applying fuzzy method to vision based lane detection and
departure warning system. Expert Syst. Appl. 37(2010):113126
6. Gupta A, Choudhary A (2017) Real-time lane detection using spatio-temporal incremental
clustering. In: Proceedings of IEEE 20th international conference on intelligent transportation
system, pp 1–6
7. Wang Y, Teoh EK, Shen D (2004) Lane detection and tracking using b-snake. Image Vis
Comput 22(4):269–280
Mobile Robot Path-Planning Using
Oppositional-Based Improved Firefly
Algorithm Under Cluttered Environment
Mohit Ranjan Panda1, Susmita Panda2(&), Rojalina Priyadarshini1,

and Pradipta Das3
1
School of Computer Engineering, KIIT University, Bhubaneswar,
Odisha, India
2
Institute of Technical Education and Research, SOA Deemed to be University,
susmitapanda@soa.ac.in
3
Veer Surendra Sai University of Technology, Burla, Odisha, India
1 Introduction
Robotics is one of the most important areas of research not only for its industrial
applications but also for its rapid involvement in many real-world aspects. And in
Robotics path planning is still a challenge for the mobile robots when the environment
is uncertain and unpredictable. Path-planning is a NP-complete problem which deals
with finding out the optimal and collision-free path from the source to the destination
either in a static environment or dynamic environment. Environment can be expressed
as the surrounding of the robot. If the environment is unchanged while a robot is
deliberating, then it is a static environment or else it is a dynamic environment. Pre-
viously traditional algorithms such as visibility graph-based, potential field method,
Voronoi-diagram, etc. are being used to find out the optimum path, but due to complex
calculations and having no guarantee that we can get the optimum solution, researchers
started to use various nature-inspired algorithms which showed better results.
There are many local and global path-planning techniques proposed by the
researchers for the navigation of the robots. When the environment is known, global
path-planning techniques are used. But the drawback of this is the inability to cope up
with the uncertainty of the environment. So to overcome this problem, various local
path-planning techniques are used. When the environment is not known or partially
known to the robot then it is called local path-planning. Various local path-planning
approaches are there like Genetic Algorithm (GA), Ant Colony Optimization (ACO),
Particle Swarm Optimization (PSO) and many more.
GA is a class of metaheuristic algorithm which is encouraged by the natural
selection procedure that belongs to the group of evolutionary algorithm and is used in
optimization problems to generate high-quality solutions. ACO is used to solve com-
putational problem and it uses the probabilistic technique. PSO is a method that is used
in optimization problems. It uses an iterative approach to improvise the candidate
solution with regard to given quality parameters. Like this, many more optimization

https://doi.org/10.1007/978-981-15-2774-6_18
142 M. R. Panda et al.
algorithms are there. In this paper, we are concentrating on the Firefly Algorithm
(FA) which is presented by Dr. Xin-She Yang in the year 2007 at Cambridge
University. This is again a type of metaheuristic algorithm which is motivated by the
flashing nature of the fireflies. FA is having simple and efficient features along with
firefly characteristics which makes it more feasible to be used in various optimization
problems and it has achieved good results. Because of its high convergence rate and
robustness, it is self-adoptable to the changing environment and unfavorable situation
in a complex system [1].
Although FA shows better performance, it has been observed from the experiments
that sometimes it suffers from premature convergence, i.e., it gets trapped in local
optima and this makes the optimization precession low or it may lead to the failure. The
loss of population diversity in the early stage is the reason behind this. To overcome
this problem, FA is enhanced to get higher convergence speed and greater accuracy. To
overcome that, we added a mutation to the FA-based approach, which involves
oppositional and dimensional learning.
2 Basic Firefly Algorithm
Firefly algorithm is a simple and effective optimization algorithm that is inspired by the
bioluminescence process of the fireflies by which they emit the light and by this light
they attract other fireflies towards them. Attraction between the fireflies depends on the
intensity of the light they emit. The main advantage of the FA is, it uses the random
movement of the individuals, and this movement depends on the mutual influence
among the fireflies at the same time. One firefly will be attracted towards the other
firefly if and only if the other firefly has greater light intensity. And this concept acts as
the basics of the firefly algorithm. There are three constraints that act as a base for the
firefly algorithm. They are given as follows:
(1) Fireflies will move towards the brighter firefly irrespective of the gender, i.e., they
are not gender-specific.
(2) The power of attraction of the fireflies is dependent upon the intensity of the same,
i.e., more the brightness, more eye-catching the fireflies will be. And the firefly
having less brightness will be attracted towards the brighter firefly.
(3) As the objective function is dependent on the brightness of the fireflies, it is
estimated by calculating the brightness of the firefly.
There are some definitions which we need to understand before getting into the
firefly algorithm, such as
Mobile Robot Path-Planning Using Oppositional … 143
Attractiveness function
It is represented by the symbol b. This is the objective function that plays a very vital
role in this algorithm. This attractiveness function is directly dependent on the
brightness of the fireflies. And with the increase in distance between the fireflies, the
brightness decreases. So the attractiveness function of the firefly is represented by an
unmodulating decreasing function, i.e.
bðr Þ ¼ b0 ecr ; ðm 1Þ
m
ð1Þ
where bðr Þ is the power of attraction which is at a distance r, b0 is the power of

attraction of the firefly at distance 0, c is one coefficient called as the light absorption
coefficient.
Here in exponential function ecr r is present in the exponent and it shows when
m
the distance will increase the brightness will decrease.

The distance between the fireflies can be calculated by using the formula,
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u d
u X 2
rij ¼ xi xj ¼ t
xi;k xk;j ð2Þ
k¼1
where rij is the distance between ith firefly and jth firefly, xi and xj are the position of
the ith and jth firefly.
Movement of the Firefly
A firefly will be attracted to the other firefly if the other firefly has greater brightness.
So this movement can be expressed as

crij2
1
x i ¼ x i þ b0 e xi xj þ a rand ð3Þ
2
There are different parameters that define the Firefly Algorithm and those parameters
are Absorption parameter, denoted by the symbol c and Random parameter, identified
by the symbol a.
Working of Firefly Algorithm
Sticking to local minima is a problem which original FA suffers from mainly while
handling dimensional rich problems. The reason is that, as it doesn’t take into account
each dimension from every firefly in isolation with others. By which some moves
towards the global optimal solution are missed. The proposed improved FA helps to get
rid of the problem of getting stuck in the local best solution, by considering the global
best solution in every generation. During the position updation process, the best
solutions are taken into account.
The limitation of the basic Firefly Algorithm which sometimes shows low con-
vergence rate and stuck at local optima, i.e., it is not so efficient in case of local search.
So to improvise the basic FA a new strategy is being applied, i.e., oppositional-based
learning which is described below.
3 Oppositional-Based Learning
Oppositional-based learning is a new way within field of robotics. It is being used to

increase the speed of convergence of various optimization algorithms. This technique is
developed by Tizhoosh which is applied successfully to various algorithms and which
showed better results. The concept behind this technique is to consider the currently
generated candidate solution with the corresponding opposite solution to get better
optimization results. Here opposite solution is also taken into consideration because it
has been proven by the researchers that sometimes opposite solution is a better solu-
tion. Let x is a real number and x′ is its opposite number and x ½a; b then x′ will be
defined as
x0 ¼ a þ b x ð4Þ

One may extend this description to higher dimensions, i.e., let P x1 ; x2; x3 ; . . .; xn
be the n-dimension vector, where xi ½ai ; bi and i ¼ 1; 2; 3; . . .; n. So the opposite
vector of p will be represented as P0 x01 ; x02 ; x03 ; . . .; x0n and x0i ¼ ai þ bi xi .
4 The Hybridized Algorithm
At the starting of the algorithm, the firefly population is set and fireflies are generated
randomly, so the diversity is highest at that time. Gradually diversity in the population
decreases with the increase in iteration number due to selection pressure or sample
generating bias. So when this occurs, the FA’s exploration capability will be reduced
and this will be called premature convergence. So to increase the searching ability of
the improved firefly algorithm, hereby combining the IFA with the oppositional-based
search technique, a new hybridized algorithm is being proposed so that the perfor-
mance will be increased. To overcome this problem, a way is to increase the diversity
in the population [13]. For this, we can replace the worst firefly i.e. the firefly having
the least brightness will be replaced by a newly constructed firefly. By this, the
diversity will be increased and this will eliminate a firefly from the local optima. As a
new firefly is added, the previous search routine is broken and this leads to the
enhancement of the diversity. This new approach can be formulated as,

xbest ðtÞ if rndðÞ\p
xworst ðt þ 1Þ ¼ ð5Þ
xmax þ xmin xworst ðtÞ otherwise
where worst and best is the representation of the worst and the best firefly respectively.
xworst ðtÞ and xbest ðtÞ is the position of the best and the worst firefly respectively at tth
iteration. rndðÞ is a uniform random number which ranges from 0 to 1. p whose value
stands between 0 and 1 is the probability that the worst firefly will move towards the
brightest firefly. After that, the elimination of the worst firefly from the normal path will
be done. And the new firefly will be constructed by taking some elements from
brightest firefly.
Proposed Algorithm
After combining the basic firefly algorithm with the oppositional-based learning
approach the new algorithm is described below.
OIFA is formed by using both the above modifications, i.e. inclusion of

dimensional-based generation of Gbest and including oppositional-based optimization.
The above Algorithm explains the pseudo-code of OIFA. Lines 1–5 explain the ini-
tialization and building of the input space of fireflies and The used fitness function f(xi)
is giving a value which is the level of light intensity of a particular firefly xi. Lines 7–16
explain the way for the updation of global best firefly by applying the context vector
Y. lines 18–24 shows the application of oppositional number for optimization. In lines
25–30, the Euclidean distance is computed among each firefly and best fire-fly. Then
each firefly is going to follow the Gbest.
5 Computer Simulation
A simulated environment has been created to test the multi-robot path-planning

algorithm. C Programming is used to satisfy the purpose of a Pentium processor with
14 spherical robots. The first round of code is executed in a group of size of 50 for
simulating four different algorithms such as OIFA, IFA and FA, each algorithm runs 30
times to get the simulated output. The input space is taken as a graphical area, where
each robot’s start and goal points are fixed priory. As robots are of a circular shape,
their radius is defined by taking 6 pixels. The obstacles are also marked and are taken in
varieties of seven unique shapes, where velocities of each robot for the experiment are
kept as fixed for a single run of the program, but again were modified for different runs
within the same program. In the experimental work, the very first world map config-
uration is shown in Fig. 1. Figure 2 presents the intermediary stages of the robot
movements. Finally, Fig. 3 is shows all the robots are reaching towards their predefined
goal position in the world map.
Fig. 1. Initial world map with 7 obstacles and 5 robots
For evaluation and validation of the proposed model’s performance, (ATTD)

Average total trajectory deviation and (AUTD) Average Untraveled trajectory Distance
are considered. From Fig. 4 it is observed that there is a change in AUTD with a
change in velocity and the number of steps. Here it could be observed more resources
Fig. 2. Intermediate state of the world map using OIFA after 4 steps
Fig. 3. Five robots reached their respective predefined goal after 22 steps
are necessary to measure AUTD to progress towards the convergence with number of
iterations. It can be observed from Fig. 5 that if the number of robots increases,
convergence rate decreases. Hence, this leads to a fall in AUTD. The exact path is
drawn for all three techniques e.g. OIFA, IFA, and FA. Figure 6 gives an overview of
calculated ATTD with varying the number of robots mostly between 1 and 8. From the
below graph, it can be concluded that OIFA performs better compared to IFA and FA.
Fig. 4. Average untraveled trajectory distance versus no of stages with variable velocity and
fixed number of obstacles
Fig. 5. Average untraveled trajectory distance versus no of stages with fixed number of
obstacles and varying the number of robot
Fig. 6. Average total trajectory path deviation versus no of obstacles with varying the number of
obstacles and velocity
Fig. 7. Average untraveled trajectory target distance versus no of steps required under different
algorithms
The minimum count of ATTD is sufficient to identify OIFA as better in comparison to

IFA and FA with varying the robot numbers. The behavior of the proposed model is
analyzed to establish the relation between the number of hops and AUTD, thus a graph
is drawn which is given in Fig. 7. This graph is generated by considering all three of
the mechanisms with 7 obstacles and 5 robots.
In this work, an oppositional-based improved firefly algorithm was proposed for getting
an optimal path for mobile robots which will also be collision-free, with a well-defined
start and end position for each of the robot present in the environment. The simulation
result is showing that the proposed approach provides a better result than the con-
ventional firefly algorithm. In the considered environment, the environment is a static
one where the landscape and obstacles, both are static in comparison to the robots; but
the robots are dynamic. Research will be carried out in the future using dynamic
obstacles other than robots during multi-robot route preparation, such as moving cars,
animals and onboard camera.
References
1. Yang XS (2009) Firefly algorithms for multimodal optimization. In: Stochastic algorithms:
foundations and applications, lecture notes in computer sciences, vol 5792, pp 169–178
2. Farahani SM, Abshouri AA, Nasiri B, Meybodi M (2012) Some hybrid models to improve
firefly algorithm performance. Int J Artif Intell 8(S12):97–117
3. El-Sawy AA, Zaki EM, Rizk-Allah R (2012) A novel hybrid ant colony optimization and
firefly algorithm for solving constrained engineering design problems. J Nat Sci Math 6(1)
4. Kasdirin HA, Yahya NM, Tokhi MO (2015) Hybridizing firefly algorithm with invasive
weed optimization for engineering design problems. In: 2015 IEEE international conference
on evolving and adaptive intelligent systems (EAIS). IEEE, pp 1–6
5. Hidalgo-Paniagua A, Vega-Rodríguez MA, Ferruz J, Pavón N (2017) Solving the multi-
objective path planning problem in mobile robotics with a firefly-based approach. Soft
Comput 21(4): 949–964
6. Cen Y, Song C, Xie N, Wang L (2008) Path planning method for mobile robot based on ant
colony optimization algorithm. In: 3rd IEEE conference on industrial electronics and
applications (ICIEA), Singapore, pp 289–301
7. Rahnamayan S, Tizhoosh H, Salama M (2008) Opposition-based differential evolution.
IEEE Trans Evol Comput 12(1):64–79
8. Rashid M, Baig A (2010) Improved opposition-based PSO for feed forward neural network
training. In: Proceedings of international conference on information science and application,
pp 1–6
9. Shaw B, Mukherjee V, Ghoshal SP (2012) A novel opposition-based gravitational search
algorithm for combined economic and emission dispatch problems of power systems. Int J
Electr Power Energy Syst 35(1):21–33
10. Chatterjee A, Ghoshal SP, Mukherjee V (2012) Solution of combined economic and
emission dispatch problems of power systems by an opposition-based harmony search
algorithm. Int J Electr Power Energy Syst 39(1):9–20
11. Panda MR, Das P, Dutta S, Pradhan S (2018) Optimal path planning for mobile robots using
oppositional invasive weed optimization. Int J Comput Intell 1–29, Wiley
12. Panda MR, Dutta S, Pradhan S (2017) Hybridizing invasive weed optimization with firefly
algorithm for multi-robot motion planning. Arab J Sci Eng 1–11, Springer
13. Panda MR, Das P, Pradhan S (2017). Hybridization of IWO and IPSO for mobile robots
navigation in a dynamic environment. J. King Saud Univ-Comput Inf Sci Elsevier
Design and Modal Analysis of Few-Mode
Fibers for Spatial Division Multiplexing
Bhagyalaxmi Behera1, S. K. Varshney2,

and Mihir Narayan Mohanty1(&)
1
ITER, Siksha ‘O’ Anusandhan (Deemed to be University), Bhubaneswar, India
2
Department of Electronics and Electrical Communication Engineering,
IIT Kharagpur, Kharagpur, India
1 Introduction
Optical communication has been around since the ‘70s with the fiber links utilizing
multi-mode optical fibers with less bit rate. However, with the advent of single-mode
fibers, it has been possible to go for a high bit rate (>few Gbps per channel) as well as
long distance. With the help of wavelength division multiplexing (WDM) or dense
WDM (DWDM). It is possible to achieve a data rate of order of terabits per second.
There is a huge demand for high bandwidth and higher bit rate at the metro and access
networks where spatial-mode multiplexing or mode-division multiplexing may play a
pivotal role.
A theoretical limit on the information-carrying capacity of a fiber link, due to the
non-linear nature of light was investigated by approximating a non-linear channel to a
linear channel with multiplicative noise. It is also observed that the non-linear effects
happen due to intensity-dependent refractive index variation is one of the important
cause of the reduction in spectral efficiency of a single-mode fiber [1].
With an aim to overcome the upcoming capacity crisis of current optical fiber links
several approaches have been evolved, such as space-division multiplexing (SDM) [2],
and mode-division multiplexing (MDM) [3]. Use of few-mode fibers and multi-core
fibers (MCFs) are gaining special attention in optical transmission using MDM and
SDM, where the transmission capacity increases substantially as compare to SMFs [2,
4]. MDM has attracted the attention of the scientific community to increase the
transmission capacity of a fiber link using FMFs. Several MDM transmission systems
using few-mode fibers have been proposed [5, 6]. In this regard, A 2 10 Gbps MDM
transmission system was realized over a distance of 10 km with the help of a two-mode
(LP01, LP11) fiber [5]. Authors in [6] have reported the effective transmission of five
modes in a mode-division multiplexed transmission, where they have concluded that
the performance of the MDM transmission system is limited by the inter-modal
crosstalk between the LP modes. Transmission of 32Wavelength Division Multiplexed
(WDM) channels using four-LP modes or 12 spatial/polarization modes (LP01, LP02,
LP11, LP21) over a distance of 177 km was demonstrated using few-mode fiber in [7],
where authors claimed a spectral efficiency of 32 bit/s/Hz per core for MDM trans-
mission. In their concluding remark, they have reported the strongly coupled trans-
mitted signals are recuperated using a 12 12 multi-input–multi-output (MIMO)

https://doi.org/10.1007/978-981-15-2774-6_19
Design and Modal Analysis of Few-Mode Fibers … 153
signal processing. A spectral efficiency-distance product of 31,500 bit/s/Hz km has

been obtained in [8] by using 33 WDM channels and three-mode graded-index fiber
over 3500-km with low differential group delay (DGD) low mode-dependent loss
(MDL)at a cost of 6 6 MIMO processing. A data rate of 280 Bits/s was reported in
[9], with the help of a mode multiplexed system using 3-mode graded-index few-mode
fibers. The transmission capacity of a MDM system increases with the increasing
number of LP modes in an FMF at cost of computational complexity and power
consumption of MIMO signal processing. The crosstalk due to strong mode coupling
between the transmitted modes and the large DGD dependent distortion in FMFs puts a
challenge for the researcher. The use of graded-index FMF is an approach to lower
down the group delay and/or the Differential Mode Delay (DMD) [10, 11]. From the
literature review, one can observe that FMF is the best solution to enhance the capacity
limit through MDM as long as the numbers of modes are few with lesser mode
coupling and low DGD.
FMF is a special type of multi-mode fiber that supports a few tens of guided modes,
where each individual mode is treated as a unique data channel. The FMF was designed
in the year 1978, with a graded-index profile to support LP01 and LP11 mode. Its modal
properties were demonstrated with a core diameter of 16.3 µm, and the normalized
frequency parameter was set with an approximation of (V) = 4.45 ± 0.11 under the
operating wavelength of 1.25 µm [12]. The initial motivation of this design is to
introduce a large core to compensate for the splicing loss present in SMFs. The
importance was lost through the invention of SMF with a splicing loss of 0.1 dB, and it
took more than 30 years to regain the importance of FMF in current research to
facilitate the high transmission capacity of optical fiber links.
In this article, the FMF design over-view of the modified-graded-index profile for a
few-mode operation is discussed. The fiber parameters are set to support four-LP
modes (LP01, LP02, LP11, and LP21) with low DGD, large mode area Aeff, and low
bending loss for MDM application C-band from 1530 to 1565 nm.
2 Design Principle
The few-mode fiber profile structures to be discussed in this article are based on the
following principles:
• The first three higher-order LP modes (LP02, LP11, and LP21) are allowed to
propagate along with the dominant LP01 mode.
• The modes are guided with less mode coupling by maintaining a large difference
between effective indices of the guided modes.
• The DGD between the above four guided modes is kept small.
• Veff is selected just below the cutoff V-value of LP31 mode.
The variation of refractive index with respect to the core radius of a graded-index
FMF (GI-FMF) and modified-graded-index FMF (MGI-FMF) is illustrated in Fig. 1a,
b, respectively.
154 B. Behera et al.
Fig. 1. Refractive index profile of GI-FMF and MGI-FMF
The GI-FMF consists of a graded-index core and a single cladding, whereas the
MGI-FMF is defined by a graded-index core with up doping (D+ = 0.3%), and fol-
lowed by a depression in the cladding which is down doped (D− = 0.7%). The graded-
index core helps to maintain equalized group delay between the guided modes, plus a
large effective index difference between the guided modes helps to reduce the mode
coupling. The relative refractive index difference (D+) between the core and cladding is
kept small to achieve a large mode area. The depression in the cladding is introduced to
reduce the bending loss. The parameters of the proposed fiber are core radius a, width
of the depression in the cladding d, the graded-index profile parameters a and q, and the
refractive index of the core and cladding nco and ncl, respectively. Where (D+) is
defined in Eq. (1). The core is assumed to be pure silica. The variation of graded-index
profile w.r.t core radius is expressed in Eq. (2) GI-FMF and in Eq. (3) MGI-FMF.

þ n2co n2cl =2n2co for GI FMF
D ¼ ð1Þ
nco ncl =nco for MGI FMF
(
nco 1 qD þ ðr=aÞa 0 r a
nðr ÞjGIFMF ¼ 1=2 ð2Þ
nco 1 2D þ r[a
8 a
< nco 1 qD þ ðr=aÞ ð0 r aÞ
nðr ÞjMGI FMF ¼ nco ð1 D Þða r d Þ ð3Þ
:
nco 1 2D þ ðr [ dÞ
The graded-index core of GI-FMF varies from a parabolic profile to a step-index

profile with a variation of a = 2 to a = ∞, respectively. Whereas q defines a sharper or
slower transition between nco and ncl. For our design simplification, we have assumed
q = 1. The effective mode indices and corresponding modal behavior of the proposed
fiber have been calculated using a finite difference method. In order to achieve low
bending loss, we have proposed MGI-FMF with a depression in the cladding just after
the core of width d. The depression is down doped by D− = 0.7%. The index profile of
MGI-FMF is expressed in Eq. (3).
3 Result and Analysis
It has been observed that the GI-FMF profile supports LP01 and LP11 modes at 1.55 µm
for any value of a 2. The effective index variation (neff) w.r.t a, and the mode field
distribution of LP01 and LP11 modes in GI-FMF is demonstrated in Fig. 2a–c,
respectively. With this result, we conclude that this fiber profile exhibits weak mode
coupling between LP01 and LP11 mode by maintaining a large index difference.
Fig. 2. a Variation of neff of LP01 and LP11 modes with respect to a, mode field distribution of
b LP01, c LP11 modes in GI-FMF with a core radius of 8 µm, D+ = 0.3% at k = 1.55 µm
The modal characteristic of GI-FMF at 1.55 µm is exhibited in Table 1. Here the

simulation results are obtained by the finite difference method. The graded-index
profile of GI-FMF, described in Eq. (2) supports only LP01 and LP11 modes with a core
radius of 8 µm, D+ = 0.3% for any value of a 2, where q is assumed to be 1. The
results are verified for the entire C-band center around 1.55 µm. From Table 1 we
conclude that the GI-FMF structure satisfies smaller mode coupling, large effective
area, and small DGD between the modes. Hence this fiber structure claims to support
mode multiplexing in C-band. But this design is suffered by large bending loss as well.
Table 1. Characteristics of GI-FMF profile with a = 5 in Fig. 1 at 1.55 µm

Property Value
Dneff of LP01 (neff – ncl) 3.8 10−3
Dneff LP01 − Dneff LP11 2.1 10−3
The cutoff wavelength of LP11 1843 nm
DGD between LP11 and LP01 −0.0007 ps/km
Bending loss of LP01 (at bending radius R = 5 mm) 9.244 101db/km
Bending loss of LP11 (at bending radius R = 5 mm) 2.319 102db/km
Aeff of LP01 135 µm2
Dispersion in LP01 13.5 ps/km nm
In this phase, we have discussed MGI-FMF. This fiber profile effectively guides
four-LP modes (LP01, LP11, LP21, and LP02). The effective indices of the guided modes
depend on both a and d. Hence, the neff versus normalized (d/2a) is plotted in Fig. 3.
With this, we conclude that the neff-variation is significant for three values of d (=1.6,
3.2, 4.8 µm). Hence, the rest of the simulations are performed with these d values.
Fig. 3. Variation of neff of LP01, LP11, LP02, and LP21 modes as a function of d/2a, in MGI-
FMF with a core radius of 8 µm, a = 5, D+ = 0.3%, D− = 0.7% at k = 1.55 µm
The MGI-FMF structure allows us to guide effectively, first three higher-order

modes (LP11, LP02, LP21) along with the dominant LP01 mode for any value of a
between 1 and ∞, and q = 1 at operating wavelength 1.55 µm. Figure 4 shows the
variation of neff as a function of a, for d = 1.6 µm. A large index difference (Dneff) is
aimed to maintain between the modes to achieve less mode coupling for MDM
applications in C-band communication.
The variation of normalized propagation parameter b w.r.t the normalized fre-
quency parameter V for different values of d (=1.6, 3.2, 4.8 µm) are depicted in
Fig. 5a–c, respectively. Where the V and b are calculated from Eq. (4) and (6),
respectively. With this result, we conclude that MGI-FMF allows guiding five LP
modes (LP01, LP02, LP11, LP21, and LP31). But we have set Veff just below the cutoff V-
value of LP31 to effectively guide the first four modes. We find that b of LP21 and LP02
moves closer with the increasing value of V. But a distinct gap is maintained between
the b of LP21 and LP02 by increasing the width d.
A large difference in the b of the guiding modes ensures less coupling between the
modes to avoid the crosstalk in MDM.
2pa
V¼NA ð4Þ
k
2
a 1=2 2 1=2
NA ¼ NAð0Þ½1 ðr=aÞ ffi nco ðr Þ ncl ra ð5Þ
0 r[a
Fig. 4. a Variation of neff of LP01, LP11, LP02, and LP21 modes with respect to a, mode field
distribution of b LP01, c LP11, d LP02, e LP21 modes in MGI-FMF with a core radius of 8 µm,
d = 1.6 µm, D+ = 0.3%, D− = 0.7% at k = 1.55 µm
Fig. 5. Variation of b of LP01, LP11, LP02, LP21, and LP31modes as a function of

V a d = 1.6 µm, b d = 3.2 µm c d = 4.8 µmin MGI-FMF with a core radius of 8 µm,
D+ = 0.3%, D− = 0.7% at k = 1.55 µm
n2eff n2cl
b¼ ð6Þ
n2co n2cl
Table 2 summarizes the modal characteristics of MGI-FMF at 1.55 µm. The results
have been calculated by using the finite difference method. The simulation results are
obtained with the following fiber parameters, core radius a = 8 µm, depression width
d = 4.8 µm, D+ = 0.3%, D− = 0.7%, a = 5. In comparison with GI-FMF, MGI-FMF
has lower bending loss at the cost of a fractional reduction in the mode area. MGI-FMF
guides four-LP modes effectively with low DGD and less mode coupling between the
modes.
Further, the bending loss of GI-FMF and MGI-FMF is compared in Fig. 6. The
bending loss is calculated at a bending radius of R = 5 mm and k = 1.55 µm. From
Fig. 6 we conclude, that the low index contrast of depression in the cladding of MGI-
FMF decreases the penetration of optical power into the cladding and thus, reduces the
bending loss excessively.
4 Conclusion
Though the graded-index fibers have been designed earlier, still there was a conflict
about the few-mode operation. In this work the modified-graded-index structure which
guides four modes have been simulated successfully. The performance shows the
Table 2. Characteristics of MGI-FMF profile with a = 5, d = 4.8 µm in Fig. 1 at 1.55 µm

Property Value
Bending loss of LP01 (at bending radius R = 5 mm) 1.743 10−1 dB/km
Bending loss of LP11 (at bending radius R = 5 mm) 5.687 101 dB/km
Bending loss of LP21 (at bending radius R = 5 mm) 4.088 104dB/km
Bending loss of LP02 (at bending radius R = 5 mm) 2.318 106dB/km
Fig. 6 Comparison of bending loss in GI-FMF and MGI-FMF as a function of wavelength at a

bending radius R = 5 mm
efficiency of our proposed method. It is well observed that the suitability condition for
mode-division multiplexing in C-band is possible. Whereas, the fiber parameters are
under study to support more than four modes with a V-value at which the DGD of the
guided modes are equalized. Simultaneously, the width of the depression in the clad-
ding, and its distance from the core must be taken care of for better accuracy; the work
is kept for future work.
References
1. Mitra PP, Stark JB (2001) Nonlinear limits to the information capacity of optical fibre
communications. Nature 411(6841):1027–1029
2. Richardson DJ, Fini JM, Nelson LE (2013) Space-division multiplexing in optical fibres. Nat
Photon 7(5):354
3. Berdagué S, Facq P (1982) Mode division multiplexing in optical fibers. Appl Opt 21
(11):1950–1955
4. Kitayama KI, Diamantopoulos NP (2017) Few-mode optical fibers: original motivation and
recent progress. IEEE Commun Mag 55(8):163–169
5. Hanzawa N, Saitoh K, Sakamoto T, Matsui T, Tomita S, Koshiba M (2011). Demonstration
of mode-division multiplexing transmission over 10 km two-mode fiber with mode coupler.
In: Optical fiber communication conference. Optical Society of America, p. OWA4
6. Salsi M, Koebele C, Charlet G, Bigo S (2012). Mode division multiplexed transmission with
a weakly-coupled few-mode fiber. In: OFC/NFOEC, IEEE, pp 1–3
7. Ryf R, Randel S, Fontaine NK, Montoliu M, Burrows E, Chandrasekhar S, Gnauck AH,
Xie C, Essiambre RJ, Winzer P, Delbue R (2013) 32-bit/s/Hz spectral efficiency WDM
transmission over 177-km few-mode fiber. In: Optical fiber communication conference.
Optical Society of America, pp PDP5A-1
8. Rademacher G, Ryf R, Fontaine NK, Chen H, Essiambre RJ, Puttnam BJ, Luis RS, Awaji Y,
Wada N, Gross S, Riesen N (2017). 3500-km mode-multiplexed transmission through a
three-mode graded-index few-mode fiber link. In: 2017 European conference on optical
communication (ECOC). IEEE, pp 1–3
9. Rademacher G, Luís RS, Puttnam BJ, Ryf R, Furukawa H, Maruyama R, Aikawa K,
Maruta A, Awaji Y, Wada N (2018). 93.34 Tbit/s/mode (280 Tbit/s) transmission in a 3-
mode graded-index few-mode fiber. In Optical fiber communication conference. Optical
Society of America, pp W4C-3
10. Ferreira F, Fonseca D, Silva H (2013) Design of few-mode fibers with arbitrary and flattened
differential mode delay. IEEE Photon Technol Lett 25(5):438–441
11. Gruner-Nielsen L, Sun Y, Nicholson JW, Jakobsen D, Jespersen KG, Lingle R, Palsdottir B
(2012) Few mode transmission fiber with low DGD, low mode coupling, and low loss.
J Lightwave Technol 30(23):3693–3698
12. Sakai J, Kitayama K, Ikeda M, Kato Y, Tatsuya K (1978) Design considerations of
broadband dual-mode optical fibers. IEEE Trans Microw Theory Tech 26(9):658–665
Speech Enhancement Using Wiener Filter
Based on Voiced Speech Probability
Rashmirekha Ram, Abhisek Das,

and Saumendra Kumar Mohapatra(&)
1 Introduction
Communication through speech is the fundamental and basic needs of human beings.
Due to clarity of communication, speech emerges as the most effective and advanta-
geous source. At the point, when the speaker and audience are near to one other,
communication is simple and with less noise. Nevertheless, in case of distance com-
munication, the listener faces difficulty to listen due to involvement of various noises
that degrades the original signal. The quality of speech needs to be prominent for
accuracy of information exchange. Background interfering noise and channel noise are
some of those which drastically degrade the performance of the system. However,
because of the intricacies of the speech signal, increasing the quality creates challenge
in speech processing research and communication system applications [1].
The characteristics of speech signal depend on time. Various kinds of noise present
in the background degrade the clarity and quality of the speech signal. To improve the
quality and to reduce the noise level, researchers have proposed numerous methods, for
example, adaptive techniques, filtering algorithms, statistical and subspace methods
[2–5]. Upadhyay et al. designed Wiener filter with recursive noise estimation for
enhancement. For performance measures, additive noise and pink noise were evaluated.
Nevertheless, speech signal was distorted [6]. To overcome this, a conventional fil-
tering was proposed based on voiced and unvoiced frames of the speech. The filter was
incorporated with empirical mode decomposition (EMD) and adaptive center weighted
average (ACWA) filter. The additive Gaussian noise was tested for the proposed
method. However, residual noise was present in the enhanced signal because of EMD.
The filtering can be improved by using proper classification methods to distinguish
voiced frames [7]. Authors used gamma tone filter bank for signal sub-band decom-
position. This phase-dependent low-delay filter bank analyzed the signal frame by
frame, and the fundamental period was estimated. For estimation, periodicity degree
(PD) was proposed to detect the period and to estimate the SNR in each frame. The
comb filter ratio and the normalized autocorrelation were multiplied to calculate the PD
in each sub-band frame signal unit. The noise energy level and the Wiener gain applied
in each frame formed the enhanced signal. The combination of the filter bank and the
Wiener filter provided better enhanced signal compared to other baseline statistical-
based methods [8].

https://doi.org/10.1007/978-981-15-2774-6_20
162 R. Ram et al.
Wavelet methods have occupied a special attention in speech enhancement area

because of their noise reduction characteristics. Depending on this, an enhancement
method was designed by utilizing voiced speech probability with Wavelet decompo-
sition. The speech probability was decided on the voiced and unvoiced frames
depending on VAD. The probability was computed between two Gaussian mixture
models (GMM) and the likelihood ratio test. The mean square error was minimized by
the gain estimator and the wavelet decomposition method. TIMIT database was used to
test the results with five different noisy conditions. The speech presence probability
(SPP) plays an important role in frequency domain speech enhancement techniques.
Most SPP methods reduced the noise variance by wavelet denoising methods. How-
ever, this method could not optimize the multitaper spectrum and denoise the speech
signals. To overcome this disadvantage, wavelet transform was utilized to the peri-
odogram, and then, an oracle was used to indicate the noise spectrum. The generalized
likelihood ratio in SPP estimated the noise spectrum and provided better improvement
in quality and intelligibility [9, 10]. Nevertheless, in stationary environment, single-
channel speech enhancement is quite difficult, and for this, Wiener filter is more
preferable. The filter transfer function used in frequency domain and the local mean and
variance were calculated sample to sample. In this method, noise was estimated by
using present and previous power values of spectrum. The smoothing parameter uti-
lized in the Wiener filter enhanced the signal preferably better. Both male and female
speakers from the NOIZEUS database were taken to evaluate the results, and it pro-
vided significantly better results than other conventional methods [6, 11].
The paper is organized as follows. Speech enhancement and some related literatures
of enhancement techniques are discussed in Sect. 1. The proposed method is
enlightened in Sect. 2. Different noisy signals from NOIZEUS database are tested, and
the results are depicted in Sect. 3. Lastly, Sect. 4 concludes the work with some future
scope.
2 Methodology
Voice activity detection is used in this work to find the voiced speech probability. For
this, decision directed approach is chosen, and the decision rule is achieved from
likelihood ratio test [12]. The noise is estimated from unvoiced region of the speech by
minimum mean square error estimation (MMSE) technique. For filtering the signal,
Wiener filter is used. As the novelty, the output of the Wiener filter is obtained from the
a priori SNR of the signal which is estimated from decision directed approach. It
provides the smoother estimates of the signal. As a result, better enhanced signal is
obtained. This section is subdivided into two subsections: (A) voiced speech proba-
bility (VSP) using voice activity detection, (B) Wiener filtering using VSP for speech
enhancement.
Speech Enhancement Using Wiener Filter … 163
2.1 Voiced Speech Probability (VSP) Using VAD

Consider the clean speech signal sðtÞ is degraded by uncorrelated additive noise nðtÞ.
For kth frequency bin, the discrete Fourier transform (DFT) coefficient vectors of noisy
speech yðtÞ can be represented as Yk . Similarly, Sk and Nk are the DFT coefficients of
clean speech and noise. For VAD, two hypotheses may arise in each frame are:
H0 : speech absent : Yk ¼ Nk
ð1Þ
H1 : speech absent : Yk ¼ Nk þ Sk
The probability density functions (pdfs) for DFT coefficients are asymptotically and
independent Gaussian random variables. The pdfs on H0 and H1 can be expressed as
[12]
( )
Y
M1
1 jYk j2
pðY jH0 Þ ¼ exp ð2Þ
k¼0
paN ðkÞ aN ðkÞ
( )
Y
M1
1 j Sk j 2
pðSjH1 Þ ¼ : exp ð3Þ
k¼0
p½aN ðkÞ þ aS ðkÞ aN ðkÞ þ aS ðkÞ
where aS ðkÞ and aN ðkÞ represent the variances of Sk and Nk , respectively. M denotes
the signal dimension of DFT coefficient vectors. The decision rule is derived from the
mean of likelihood ratio test for each frequency band. The likelihood ratio for the kth
frequency band is

pðYk jH1 Þ 1 kk bk
Lk , ¼ exp ð4Þ
pðYk jH0 Þ 1 þ bk 1 þ bk
where bk is the a priori SNR and kk is the a posteriori SNR. These two parameters can
be expressed as
bk , PS ðkÞ=PN ðkÞ ð5Þ
and
2.
kk , jYk j aN ðkÞ ð6Þ
where PS ðkÞ is the power spectra of clean speech and PN ðkÞ is the power spectra of
noise The block diagram of VAD is shown in Fig. 1.
164 R. Ram et al.
Fig. 1. Block diagram of voice activity detection
We assume that aN ðkÞ coefficients are known by noise estimation, and the a priori
SNR bk needs to be estimated. To estimate bk , decision directed (DD) method is
chosen and is expressed as [13]
^
^ ðmÞðDDÞ ¼ l Ak ðm 1Þ þ ð1 lÞP½kk ðmÞ 1
b ð7Þ
k
aN ðk; m 1Þ
where m is the frame index and A^ k ðm 1Þ ¼ signal amplitude of previous frame that is
estimated by MMSE. Equation (7) provides the DD estimates of a priori SNR. It
reduces the fluctuations arise in the estimated likelihood ratio of noise periods. An
example of probability of speech presence by VAD is shown in Fig. 2.
Fig. 2. Speech presence probability for a noisy input signal

2.2 Wiener Filtering Using VSP for Speech Enhancement

After finding the speech from the noisy speech signals, the noise spectrum is updated.
The MSE estimator is used to estimate the signal as it is based on conditional prob-
ability. To estimate the SNR, the gain function is used, and it depends on a priori and a
posteriori estimates. Again, both these SNR values depend on the estimated conditional
speech and noise spectral variance. To reduce the MSE, the spectral component of
noisy signals and the spectral gain function is multiplied. The speech variance and the
noise spectral variance are estimated from the voiced frames and the unvoiced frames,
respectively. The spectral gain function used in MMSE estimation is considered from
[14, 15]. To estimate the a priori SNR, the gain function predicts the spectral variance
of clean signal from the noisy signals. The block diagram of the proposed speech
enhancement technique is shown in Fig. 3.
Fig. 3. Block diagram of proposed speech enhancement technique
Further to minimize the estimation error optimally, Wiener filter is used. The
objective is to generate an estimate of the clean signal spectrum Sk from the desired
signal Dk . The Wiener filter output in frequency domain can be written as,
PS ðkÞ
HðkÞ ¼ ð8Þ
PS ðkÞ þ PN ðkÞ
166 R. Ram et al.
The filter coefficients are real, even and non-negative because both the power
spectra are greater than zero and are even symmetry [2]. This signifies that the impulse
response hðkÞ is even and indicates that hðkÞ is not causal. Consequently, Eq. (8) can
be expressed as
bk
HðkÞ ¼ ð9Þ
bk þ 1
The values of bk are estimated using Eq. (7) and can be expressed as
^ ðmÞDD
b
HðkÞ ¼ k
ð10Þ
^ ðmÞDD þ 1
b k
The above equation provides the enhanced signal by Wiener filter based on voiced
speech probability. This proposed technique emphasizes the portion of the speech
spectrum wherever the SNR value is high and attenuates the portion of the signal
spectrum wherever the SNR value is low. Different signals are tested to prove the
efficacy of the proposed technique, and the results are shown in the next section.
For experimentation, clean speech and noisy speech signals are taken from NOIZEUS
database [2]. It has been generally used for benchmark experimentation. This database
consists of 30 IEEE phonetically balanced sentences. These sentences are produced by
both male and female speaker. Eight different noises with four different SNR levels are
there. In our experimentation, Babble noise of SNR 0–15 dB is considered with an
increment of 5 dB. For simulation, MATLAB is used. The sentence ‘He knew the skill
of the great young actress’ is uttered by a male speaker and is taken as the clean signal.
The spectrogram of the clean speech signal is shown in Fig. 4. Its corresponding noisy
signal corrupted by Babble noise of SNR 5 dB is shown in Fig. 5.
To prove the effectiveness, we compare the enhanced signal of proposed algorithm
and the output of Wiener filter. For pre-processing, 16 ms of frame size with 10 ms
skip rate is used. Hamming window of length 256 is considered. The windowed frame
is then analyzed using DFT. For speech probability, VAD is used with decision
directed approach. The proper separation of voiced and unvoiced frames provides
better estimation of the gain function. Unlike the MMSE gain function, Wiener gain
function does not depend on kk . Therefore, estimation of bk is very much important to
estimate the noise from the unvoiced region of the speech. The output of the voiced
speech probability is shown in Fig. 2.
The Wiener filter output and the proposed technique are shown in Figs. 6 and 7,
respectively. It is observed from the figures that the frequency distribution is less in the
proposed technique as compared to Fig. 6. Because when HðkÞ ¼ 0, bk ! 0. This
signifies very low SNR regions. Similarly, HðkÞ ¼ 1 when bk ! 1 and signifies very
high SNR regions.
Fig. 4. Spectrogram of clean speech signal (‘He knew the skill of the great young actress’)
Fig. 5. Spectrogram of noisy speech (Babble noise of SNR 5 dB)
To examine the performance of the enhancement techniques, SNR and PESQ are
used [16]. The maximum SNR improvement is achieved 2.54 dB (17.54–15 dB).
Similarly, the maximum PESQ score is obtained as 2.78. In case of Wiener filter, the
maximum PESQ score is 2.25 (Table 1).
168 R. Ram et al.
Fig. 6. Spectrogram of enhanced signal using Wiener filter
Fig. 7. Enhanced signal using proposed method
Table 1. Comparison between Wiener filter and proposed method in terms of SNR and PESQ
score
Babble noise Wiener filter Proposed method
SNR
0 dB SNR (dB) 1.23 2.32
PESQ 1.65 1.92
5 dB SNR (dB) 6.58 7.23
PESQ 1.99 2.38
10 dB SNR (dB) 11.67 12.29
PESQ 2.23 2.45
15 dB SNR (dB) 16.88 17.54
PESQ 2.25 2.78
4 Conclusion
In this work, Wiener filter with VSP is used for enhancement. The decision directed
approach with likelihood ratio determines the voiced and unvoiced region of the
speech. The noise is estimated from the unvoiced region and for this MMSE is used.
Both a priori and a posteriori SNR are estimated. However, in both MMSE and Wiener
estimator does not depend on a posteriori estimation. Therefore, the a priori SNR
estimated from DD approach used in Wiener filter provides better enhanced signal. The
proposed method is compared with Wiener filter. The proposed method provides a
maximum SNR of 17.54 dB and a PESQ score of 2.78. Results show that the proposed
Wiener filter outperforms better than Wiener filter. In the future, different VAD method
can be applied. Other filtering methods may be tested for quality improvement.
References
1. Quatieri TF (2002) Discrete-time speech signal processing: principle and practice. Prentice
Hall, New York
2. Loizou PC (2013) Speech enhancement: theory and practice. CRC Press
3. Ram R, Mohanty MN (2016) Performance analysis of adaptive algorithms for speech
enhancement applications. Indian J Sci Technol 9(44):1–9
4. Ram R, Mohanty MN (2017) Design of fractional fourier transform based filter for speech
enhancement. Int J Control Theory Appl 10(7):235–243
5. Ke Y, Hu Y, Li J, Zheng C, Li X (2019) A generalized subspace approach for multichannel
speech enhancement using machine learning-based speech presence probability estimation,
vol 146. Audio Engineering Society Convention, Audio Engineering Society
6. Upadhyay N, Jaiswal RK (2016) Single channel speech enhancement: using wiener filtering
with recursive noise estimation. Procedia Comput Sci 84:22–30
7. Khaldi K, Boudraa AO, Turki M (2016) Voiced/unvoiced speech classification-based
adaptive filtering of decomposed empirical modes for speech enhancement. IET Signal
Process 10:169–180
8. Chen Z, Hohmann V (2015) Online monaural speech enhancement based on periodicity
analysis and a Priori SNR estimation. EEE/ACM Trans Audio Speech Lang Process
(TASLP) 23(11):1904–1916
9. Bhowmick A, Chandra M (2017) Speech enhancement using voiced speech probability
based wavelet decomposition. Comput and Electr Eng 62:706–718
10. Lun DPK, Shen TW, Hsung TC, Ho DK (2016) Wavelet based speech presence probability
estimator for speech enhancement. Digit Signal Process 22(6):1161–1173
11. El-Fattah MAA, Dessouky MI, Abbas AM, Diab SM, El-Rabaie ESM, Al-Nuaimy W, … El-
Samie FEA (2014) Speech Enhancement with an adaptive wiener filter. Int J Speech
Technol 17(1):53–64
12. Sohn J, Kim NS, Sung W (1999) A statistical model-based voice activity detection. IEEE
Signal Process Lett 6(1):1–3
13. Ephraim Y, Malah D (1984) Speech enhancement using a minimum-mean square error
short-time spectral amplitude estimator. IEEE Trans Acoust, Speech, Signal Process 32
(6):1109–1121
14. Mowlaee P, Scheran D, Stahl J, Wood SU, Kleijn WB (2019) Maximum a posteriori speech
enhancement based on double spectrum. Proc. Interspeech 2019:2738–2742
170 R. Ram et al.
15. Ram R, Mohanty MN (2019) Use of radial basis function network with discrete wavelet
transform for speech enhancement. Int J Comput Vis Robot 9(2):207–223
16. Ram R, Mohanty MN (2018) Performance analysis of adaptive variational mode
decomposition approach for speech enhancement. Int J Speech Technol 21(23):69–381
Analysis of Different Parameter of FSO
Communication System by Using Robust
Model Design
Shubham Mahajan and Amit Kant Pandit(&)
School of Electronics & Communication, Shri Mata Vaishno Devi University,

Katra, Reasi, J&K, India
1 Introduction
“FSO is an observable pathway innovation which uses lasers and photo-identifiers to

give optical associations between two points—without the fiber. FSO can transmit
information, voice, or video at velocities equipped for arriving at 2.5 Gbps. Items
equipped for paces up to 10 Gbps are relied upon to hit the business sectors inside one
year. FSO units comprise an optical handset with a laser (transmitter) and a photo-
identifier (beneficiary) to give full duplex (bidirectional) capacity [1]. FSO systems
utilize imperceptible infrared laser light wavelengths in the 750–1550 nm territory.
FSO gives point-to-point transmission of communication data through the climate
utilizing the optical flag as the transporter frequencies. It has attracted consideration
media transmission industry, because of its cost adequacy—simple establishment,
snappy foundation of communication interface particularly in the calamity the execu-
tive’s situation, high data transfer capacity provisioning, and wide scope of uses. The
scope of frequencies where it works makes FSO communication free from authorizing
[2]. With FSO communication, most extreme information moves rates up to 2.5 Gbps
is conceivable, not normal for the greatest information move paces of 622 Mbps
offered by RF communication systems. FSO includes the optical transmission of voice,
video, and information utilizing air as the mode of transmission (Fig. 1).
Transmission utilizing FSO innovation is generally straightforward [3, 4]. It
includes two systems each comprising of an optical handset which comprises a laser
transmitter and a collector to give full duplex (bidirectional) capacity. Each FSO system
utilizes a high-control optical source (e.g., laser) in addition to a telescope that transmits
light through the climate to another telescope that gets the data. By then, the accepting
telescope interfaces with a high affectability recipient through an optical fiber.
2 Literature Review
“Throughout the most recent two decades free-space optical communication (FSO) has
turned out to be increasingly intriguing as an aide or choice to radio recurrence cor-
respondence. This chapter gives an outline of the difficulties a system fashioner needs

https://doi.org/10.1007/978-981-15-2774-6_21
172 S. Mahajan and A. K. Pandit
FSO Link
Transmitter Receiver
Fig. 1. Basic FSO communication system link
to consider while executing an FSO system. Commonplace increases and misfortunes

along the way from the transmitter through the medium to the collector are presented in
this chapter [5]. Nitty-gritty discourses of these points can be found in this uncommon
issue of the Radio building Journal. Its high transfer speed abilities and low lessening
attributes make it perfect for gigabit transmission. Fiber-optic communication has
altered the broadcast communications industry. It has additionally made its essence
broadly felt inside the information organizing network too. Utilizing fiber-optic link,
optical communication shave empowered broadcast communications connects to be
made over a lot of more noteworthy separations and with much lower levels of losing
the transmission medium and conceivably most imperative to fall, fiber-optical inter-
changes have empowered a lot of higher information rates to be obliged. Fiber-optic
systems are significant media transmission foundation for overall broadband systems.
Wide data transmission signal transmission with low postponement is a key prereq-
uisite in the present-day applications. Optical filaments furnish gigantic and fantastic
transmission transfer speed with unimportant inactivity and are presently the trans-
mission mode of decision for long separation and high information rate transmission in
media transmission systems. Free-space optics is a medium with high transfer speed
having most extreme information rates and security issues supporting its advancement
for the present period. Fierce air influences the presentation of the connection. Stick-
iness, water vapor, signals ingestion, bar shine, spreading, and meandering are a
portion of the elements which cause laser pillar debasement [6]. Keeping up a free-
space optical connection between two intersections is an extreme test and needs
improvement in its highlights.
3 Methodology
FSO Link: The transmitter balances information on to the prompt intensity of an optical
pillar. We consider power modulated direct recognition channels utilizing on/off keying
(OOK) balance, which is broadly utilized in functional systems [7, 8]. The got pho-
tograph current sign is identified with the occurrence optical power by the finder
responsivity r.
Gotten Signal Y ¼ Hr X þ N
X Transmitted intensity
H Channel state
Y Resulting electrical signal
N Signal-independent AWGN.
Analysis of Different Parameter of FSO … 173
Fig. 2. FSO communication system
Parameters: We assessed the bit error rate (BER) execution of the proposed system
changing with various system parameters, for example, distance transmitter power path
loss factor, atmospheric turbulence received signal power. BER is profoundly cor-
rupted on extreme environmental turbulence condition ever for a short separation of
free-space channel. The impact of way misfortune factor because of thick haze is
additionally extreme on the BER despite the fact that the disturbance impact and free-
space separation are short. The ideal transmitted info power and collector reception
apparatus span at which the BER is least is emphatically subject to free-space sepa-
ration and climatic parameters. In FSO, there are different parameters which we need to
dissect, and for it, we need to utilize different squares from reenactment library with the
goal that goal can be accomplished. During this the need to overcome the errors,
dispersion, scattering and other atmospheric conditions (Figs. 2 and 3).
I/p Convolution
(Conv) O/p
Hadamard
Code Generator
Fig. 3. Block diagram of Hadamard code generator
Free-space optics (FSO) innovation-based remote systems are not without diffi-
culties. The key restriction of free-space optical interchanges emerges from the earth
through which it engenders. Albeit moderately unaffected by downpour and day off,
space optical communication systems can be seriously influenced by haze and envi-
ronmental disturbance. The primary plan difficulties in free-space optical interchanges
are delineated in figure.
Atmospheric Turbulence: For weak turbulences, the probability of changing like-

lihood density function is demonstrated as a log-normal dispersion the log-amplitude of
the optical intensity has a Gaussian PDF with log-amplitude difference given by
r2r
r2x
4
Haze and free-space optics: The primary test is mist. Mist is vapor made out of
water droplets, which are just a couple of hundred microns in distance across however
can adjust light attributes or totally impede the section of light through a blend of
retention, dissipating, and reflection. This can prompt a decline in the power thickness
of the transmitted pillar, diminishing the successful separation of a free-space optical
connection [9, 10].
Scintillation and free-space optics: Scintillation is the spatial variety in light force
brought about by environmental turbulences. Such turbulences are brought about by
wind and temperature slopes that make pockets of air with quickly fluctuating densities
and, accordingly, quick changing records of optical reflection. These air pockets act
like focal points with time-changing properties and can prompt sharp increments in the
bit-mistake paces of free-space optical communication systems, especially within the
sight of direct daylight (Fig. 4).
Fig. 4. Block diagram of atmospheric effects on FSO communication system
Beam wander and free-space optics: Beam meander emerges when tempestuous
breeze current (whirlpools) bigger than the distance across of the transmitted optical bar
cause a moderate, however critical, relocation of the transmitted bar. Shaft meander
may likewise be the aftereffect of seismic movement that causes a relative relocation
between the situation of the transmitting laser and the getting photograph finder
(Fig. 5).
Fig. 5. Block diagram of transmitter and receiver FSO communication system
FSO CHALLENGES
A. Absorption and scattering loss
In process of propagation through earth’s atmosphere, the laser beam has to interact
with particles and gas molecules present in atmosphere. Absorption and scattering loss
are major reasons of loss in atmosphere channel. Attenuation in atmosphere is repre-
sented as below:
a ¼ erd
distance of measurement is denoted by d; coefficient attenuation particular to length per

unit is denoted by r.

3:912 k q
rffi
V 550
Visibility is denoted by V and is measured in km; wavelength is denoted by k and

measured in nm; diffusing particle distribution is of size q attained by Kim and Kruse.
B. Fog
It results in atmospheric attenuation leading to absorption and scattering losses. It will
result in high attenuation 350 dB/km in dense fog conditions.

10 log V% k q
aFog ¼
V ðkmÞ ko
Visibility is denoted by V, measured in km; V% is percentage of air drop; wave-

length is denoted by k in nm; range of visibility is denoted by ko (550 nm); and
diffusing particle distribution is of size q. Now, by Kruse model parameter of q is
defined as
8
>
> 1:6 if V [ 50 km;
<
1:3 if 6 km\V\50 km;
q¼
>
> 0:585V if V\6 km:
:
Kim [9] has defined q parameter as below:

8
>
> 1:6 V [ 50 kms
>
>
< 1:3 6 kms\V\50 kms
q ¼ 0:16 V þ 0:34 1 kms\V\6 kms
>
>
> V 0:5
> 0:5 kms\V\1 kms
:
0 V\0:5 kms
C. Rain
Its brunt is not more than fog. Wavelength of FSO is smaller than drops of rain. It has
been suggested that 2.5 mm/hr. corresponds to light rain and 25 mm/hr corresponds to
heavy rain [10], [15]. According to Carbonneau.
Attenuation of rain ¼ 1:076 R23
D. Snow
Attenuation because of snow is more than rain, and not as much as fog as size of
snow particles is in the middle of fog and rain. Attenuation because of snow is
30–350 dB/km.
Software Used MATLAB 2015: It is amazing programming that gives a domain to
numerical calculation just as graphical showcase of yields. In MATLAB, the infor-
mation is in the ASCII design just as paired configuration. It is elite language for
specialized registering coordinate calculation, representation, and programming in a
straightforward manner where issues and arrangements are communicated in com-
monplace scientific documentation.
• Acquisition, Data Exploration, Analyzing, and Visualization
• Engineering complex drawing and logical illustrations
• Analyzing of algorithmic structuring
• Mathematical and Computational capacities
• Modeling and mimicking issues prototyping
• GUI (graphical UI) building condition.
Utilizing MATLAB, you can take care of specialized figuring issues all around
effectively and efficient when contrasted with customary programming dialects, for
example, C, C++, and FORTRAN. The name MATLAB represents network research
center.
Features of MATLAB:-
• MATLAB is an abnormal state language utilized for numerical calculation, per-
ception, and application advancement.
• It makes all around well-disposed condition for iterative investigation, plan, and
critical thinking.
• Mathematical capacities for understanding conventional differential conditions,
Fourier examination, direct variable-based math, insights, separating, enhancement,
numerical mix.
• Development instruments for upgrading code quality and augmenting execution.
• Tools for structure applications with custom graphical interfaces (GUI).
• Functions for incorporating MATLAB-based calculations with outside applications
and we can ready to create code in hex document, c, implanted and so on.
4 Conclusion
In the wake of examining different research papers, we will dissect the effect of different
parameters like air disturbance, way misfortune factor, and pointing blunder on the
exhibition of free-space optical communication system. To acquire our objective, we will
recreate a model of free-space communication system in MATLAB utilizing Simulink.
From transmitter to recipient, we will utilize different square which can give us better
outcome intended to state when we transmit information from transmitter; then, it is
proliferated viva remote channel and certainly different parameters like haze, lessening
component, downpour, and environmental turbulences influence our information so that
at collector side we got twisted information it implies BER will be there and to decrease it
is our fundamental target. Other than this another significant factor is control that must be
least. At that point, the BER is assessed by shifting the few parameters, for example,
separation, transmitter control, way misfortune factor, got signal power, and climatic
disturbance. Way misfortune factor is extreme in nearness of thick haze. The impact of
climatic turbulences was likewise seen at most extreme disturbance conditions; the BER
is additionally roughly 12–28. In research, our fundamental center will be to dissect the
effect of different parameters on information transmission.”
Acknowledgements. This work in itself is an acknowledgment to the inspiration, drive, and the
technical assistance contributed to it by many people. It would have never been possible without
the help and guidance that it received from my teachers. Many thanks go to my colleagues
around me for providing a friendly and entertaining working environment. Special thanks to my
friends for her personal support.
On a personal level, I am forever indebted to my parents for the love they provided me through
my entire life.
References
1. Fidler F, Knapek M, Horwath J, Leeb WR (2010) Optical communications for high-altitude
platforms. IEEE J Sel Top Quantum Electron 16(5)
2. Al-Habash MA, Andrews LC, Philips RL (2001) Mathematical model for the irradiance
probability density function of a laser propagating through turbulent media. Opt Eng 40(8):
1554–1562
3. Noshada M, Rostami A (2012) FWM minimization in WDM optical communication systems
using the asymmetrical dispersion managed fibers. Int J Light Electron Opt 123(9):758–760
4. Wang X, Kitayama K (2004) Analysis of beat noise in coherent and incoherent time-
spreading OCDMA. IEEE/OSA J Light Technol 22(10):2226–2235
5. Zhu X, Kahn JM (2003) Performance bounds for coded free-space optical communications
through atmospheric turbulence. IEEE Trans Commun 51(8):1233–1239
6. Ahmed A, Hranilovic S (2007) Outage capacity optimization for free-space optical links
with pointing errors. IEEE J Light Wave Technol 25(7)
7. Shake TH (2005) Confident performance of encoded optical CDMA. IEEE/OSA J Light
Technol 23:1652–1663
8. Sharma Prachi et al (2013) A review of the development in the field of fiber optic
communication systems. Int J Emerg Technol Adv Eng 3(5):113–119
9. Takaaki O (2010) Recent status and trends in optical submarine cable systems. NEC Tech J
5(1):4–7
10. Kaushal H, Kaddoum G, Optical communication in space: challenges and mitigation
techniques. IEEE Commun Surveys Tutorials. https://doi.org/10.1109/comst.2016.2603518
11. Zhu X, Kahn JM (2002) Free space optical communication through atmospheric turbulence
channels. IEEE Trans Commun 50(8):1293–1300
12. Naboulsi MA, Sizun H, de Fornel F (2004) Fog attenuation prediction for optical and
infrared waves. Opt Eng 43(2):319–329
13. Milev A, Minchev C (2007) A simulation model of an optical communication CDMA
system. Int Conf Comput Syst Technol Compo Sys Tech
14. Theodore S, Rappapo R (2007) Wireless communications: principles and practice, 2nd edn.
Prentice Hall
An Improved Grasshopper Optimization
Algorithm for Solving Numerical Optimization
Problems
Puneet Mishra1(&), Vishal Goyal2, and Aasheesh Shukla2

1
Department of Electrical and Electronics Engineering,
Birla Institute of Technology and Science, Pilani, Pilani Campus, Jhunjhunu,
Rajasthan 333031, India
2
Department of Electronics and Communication Engineering, GLA University,
Mathura, India
aasheesh.shukla@gla.ac.in
1 Introduction
To avoid the entrapment of optimization techniques in local optimal points, a class of

optimization techniques known as stochastic optimization has arisen [1]. The stochastic
optimization problems rely on random generation of candidate solution and are free
from gradient-based techniques to solve the optimization problem. A subcategory of
stochastic optimization algorithms is bio-inspired optimization algorithm. The bio-
inspired optimization algorithms have proven their efficacy to solve real-world chal-
lenging optimization problems in recent past. Some of the bio-inspired optimization
algorithms are genetic algorithm (GA) [2, 3], particle swarm optimization (PSO) al-
gorithm [4, 5], bat algorithm (BA) [6, 7], flower pollination algorithm (FPA) [8, 9],
grey wolf optimizer (GWO) [10, 11], cuckoo search algorithm (CSA) [12, 13], ant-lion
optimizer (ALO) [14–16], Krill Herd optimization algorithm (KHOA) [17, 18], moth
flame optimization algorithm (MFOA) [14, 15], ant colony optimization (ACO) [19],
honey bee mating optimization [20], and many more.
It may be noted that the search algorithms relying on natural physical laws are
adequately efficient in searching the global optimum points; however, bio-inspired
algorithm development and their application have become more popular. The appli-
cation areas of these algorithms range in a variety of fields such as route planning [21],
controller parameter optimization [22], structure design problems [23], micro-
electronic mechanical system design [24], UAV path planning [25], and system
identification [26]. It may be noted that, though optimization algorithms based on the
salient features of different species have evolved in literature, the quest for a more
efficient algorithm is always continued, as it is very hard to say, that which algorithm
will be applicable and suitable to a wide variety of optimization problems. In the same
trend, a bio-inspired optimization algorithm which mimics the swarm and food search
behaviour of grasshoppers has been recently proposed in the literature [23].

https://doi.org/10.1007/978-981-15-2774-6_22
180 P. Mishra et al.
This grasshopper optimization algorithm (GOA) is claimed to provide a balance

exploration and exploitation in the search space of optimization problem. However,
GOA quickly moves towards the current optimal point in the search space and thereby
increases the risk of finding the local optimum point. The current work deals with the
same problem of GOA, and an attempt has been made by providing two unique
modifications in the approach of conventional GOA. The improved GOA (IGOA)
proposed in this work has been tested for 19 benchmark test functions. Various per-
formance measures have been considered to verify the superiority of IGOA over GOA
and have been discussed in detail in this paper.
Further, this paper is divided in five sections. A brief introduction and the need for
the improvement of GOA are presented in Sect. 1. Implementation and brief details of
GOA are presented in Sect. 2. Section 3 presents the improvements made in GOA and
implementation aspects of IGOA in detail. Section 4 presents and compiles the
experimental results for considered benchmark test functions. Finally, Sect. 5 con-
cludes the paper with inference drawn from the work carried out in the paper.
2 Grasshopper Optimization Algorithm (GOA)
Saremi et al. [23] related the movements of different grasshoppers to solve optimization
problem by giving equivalence to these movements to exploitation and exploration in a
search space.
A mathematical model as proposed in [23] is presented here
Xi ¼ r1 Si þ r2 Gi þ r3 Ai ð1Þ
where Xi is the position of the ith grasshopper, Si is the social interaction, Gi is gravity
force on the ith grasshopper, and Ai shows the wind advection. r1, r2, and r3 are random
numbers in [0, 1] to provide random weight to different factors in (1). Equation 2
defines the social interaction [23], obtained by social forces as given by s(r) function
given by Eq. (3). dij is the distance between ith and jth grasshopper.
X
N
Si ¼ s dij c
dij ð2Þ
j¼1
j6¼i
sðr Þ ¼ f e l er
r
ð3Þ
where f is intensity of attraction and l is attractive length scale. The set of equations
from Eqs. (1) to (3) are then further used to mathematically model some aspects of
grasshopper’s movement in order to solve optimization problems. It may be noted that
the second and third terms in Eq. 1 are less significant in the developed mathematical
model in [23], and is represented as,
An Improved Grasshopper Optimization … 181
2 3
X
N
ðubd lbd Þ xj xi 7 c
6
Xid ¼ c4 c s xdj xdi 5 þ Td ð4Þ
j¼1
2 dij
j6¼i
where ubd and lbd are upper bound and lower bound in dth dimension. c is a coefficient
to shrink the comfort, repulsion, and attraction zone. A detailed representation of these
zones is explained in [23]. The factor c plays a very important role in mathematical
modelling of the two important aspects of the optimization procedure, i.e., exploration
and exploitation. The c factor is multiplied two times in update Eq. (4), which provides
a balance between the above said two procedures. It may be noted that the c factor is
not a constant value and is varied as the optimization procedure progresses further. The
original version of GOA decreases the c factor linearly from unity to a minimum value
of 10−5 with iterations. However, GOA has a strong tendency to move very quickly
towards the current optimal point and the choice of factor like r and l becomes very
crucial in order to determine an optimal balance between random walks and targeted
walks of search agents. Though provisions and care have been taken while proposing
GOA, many times, it gets stuck in local optimal points. In order to avoid this issue, a
modified and improvised version of GOA has been proposed in this paper, which is
discussed in subsequent section.
3 Improved Grasshopper Optimization Algorithm (IGOA)
GOA is prone to quickly converge and confines the search space to move towards the
current optimal solution. This behaviour of GOA increases the risk of entrapment in
local optimal solution which should be avoided. In order to address this issue, two
modifications are proposed in this work. These modifications enable the IGOA to
converge faster and avoid local optimal solutions in contrast to conventional GOA.
These modifications are in the form of,
(a) update equation for c factor
(b) random walk application in grasshopper swarm.
The improved version of GOA, i.e., IGOA makes use of an interesting update
equation for c factor to be used in update of search agents. Since c factor is supposed to
provide a balance between exploration and exploitation, it is essential that c factor
remains close to unity for initial iterations and then decreases and then maintains at a
minimum level to support the exploitation procedure. This paper adopts the above-
mentioned methodology using an interesting c factor update equation as given in Eq. 6.
It is worth mentioning here that the value of c factor will always start from unity and
may reach to very low values as the iteration values progress. The value of r plays an
important role in variation of c factor as it tells the rate at which the c factor will
decrease. A lower value of r will decrease the decrement rate and the value of c factor
will remain high for a high number of iterations. Similarly, a very high value of r
(greater than 6) will provide a steeper decrement in c factor and will largely support the
exploitation procedure. A typical value of three to four can be taken for r. This paper
adopts the value of r as 4 for all the optimization procedures. A typical variation of
c factor for 200 iterations is shown in Fig. 1 for GOA (linear c factor variation) and for
IGOA (exponential variation). It may be seen that for IGOA, the values of c for iter-
ations lesser than 50 support more exploration than GOA and later support a balanced
approach and then emphasize more on exploitation. This increases the chances of
IGOA to explore the search space in a more judicial manner. The simulation results
support this claim and are presented in Sect. 4.
8 0 ( )2 1
>
>
< exp@0:5 l A if c [ cmin
c¼ L=
r ð5Þ
>
>
:
cmin otherwise
Further, another step is also added to IGOA to avoid premature convergence of

optimization procedure. A biased random walk nature, inspired by the cuckoo search
algorithm [12], is added to conventional GOA to increase the exploration possibility.
A randomly selected grasshopper in the swarm is moved towards another randomly
selected grasshopper with a random step size. This random walk of grasshoppers in a
swarm will increase the exploration possibility and will avoid entrapment of procedure
in local minima. It is worth mentioning here that this random walk applied to a
grasshopper is associated with a probability p lying in range of [0, 1]. A random
number is generated in each iteration which is compared with p, and if the value of
random number is lesser than p, then the random walk is initiated.
1
GOA
0.9 Exploration feasible scenario IGOA
0.8
0.7
0.6
C factor
0.5
0.4
Exploitation feasible
0.3 scenario
0.2
0.1
0
0 20 40 60 80 100 120 140 160 180 200
Iterations
Fig. 1. Variation of c factor for GOA and IGOA

In order to test the efficacy of the proposed IGOA over GOA, nineteen benchmark test
functions were optimized using both algorithms. The results were compared on the
basis of achieved best values, mean, and standard deviation from 20 independent trial
runs. These benchmark test functions range from unimodal to multimodal function with
higher dimensions as well. Since it was essential to maintain the same platform for
GOA and IGOA, both algorithms are implemented in LabVIEW environment.
Benchmark test functions [27] listed in Table 1 are optimized by GOA and IGOA for
200 iterations with a swarm size of 20 for all the functions except F11 and F18, whose
dimensions are considered as 10 in contrast of 2, which is for all other function. For
these function (F11 and F18), the swarm size is taken as 30 and maximum number of
iterations as 500. Further, for the optimization procedures the values of all other
parameters of GOA is taken same as in [23], and also for IGOA, wherever relevant.
Figure 2 shows the convergence graph for F1 obtained by GOA and IGOA. It may be
clearly seen from this graph that the IGOA performance is far better than GOA, as it
converges very fast in comparison with GOA and also obtains the minimum value
better than GOA. It should be noted here that the convergence graphs plotted for
different functions are for the best result obtained from the 20 independent trial runs.
The faster convergence of IGOA in comparison with GOA is due to the modified
variation in c factor in optimization procedure. It enabled the IGOA to search more
randomly in initial phase of optimization, and then for later parts, it searched with a
smaller value of c factor to facilitate the exploitation procedure. This resulted in faster
and more steep convergence of IGOA in contrast to GOA which only had a steeper
convergence at very late stage as shown in Fig. 2. Further, due to space constraint,
convergence curves are not shown for other functions; however, quantitative com-
parison is shown between IGOA and GOA for considered functions.
102
Absolute Deviation from optimum value
GOA
IGOA
100
10-2
10-4
10-6
10-8
0 20 40 60 80 100 120 140 160 180 200
Iterations
Fig. 2. Convergence graphs for absolute deviation from optimum values for F1 function
To evaluate and support the supremacy of IGOA over GOA, a detailed comparative
analysis was also done using 20 independent trial runs for each test function. Different
factors such as mean value, standard deviation, and percentage improvement in mean
values (PIMV). The expressions to calculate PIMV, which is used to rank the algo-
rithms, are given as follows:
jGMV MeanGOA j jGMV MeanIGOA j

PIMVð%Þ ¼ 100 ð6Þ
jGMV MeanGOA j
where GMV is the global minimum value for the specified test function, and Mean is
the average value obtained by GOA or IGOA, for the specified test function. This
performance measure is an interesting criterion to look for. This performance criterion
looks for two important things: First it incorporates the IGOA performance to locate the
global minimum value, and second it also incorporates the relative improvement of the
IGOA over GOA. Similarly, performance criterion for logarithmic improvement in best
value obtained by GOA and IGOA has also been taken and is calculated as,

jGMV BestGOA j
LIBV ¼ log10 ð7Þ
jGMV BestIGOA j
These logarithmic performance criteria are helpful to determine the IGOA’s

improvement over GOA in terms of mean value and best value. These values must be
positive which indicate that the IGOA is superior in terms of deviation of mean or best
value from the global minimum of a specified test function. Details of best values
obtained by GOA and IGOA for different considered test functions are shown in
Table 1. It may be noted that the best values obtained from 20 independent trial runs by
IGOA is always superior or equivalent to conventional GOA. It may be noted from the
data provided in Table 1 that none of the values obtained by IGOA is worse than GOA
and all the LIBV performance measures are positive. To further investigate the per-
formance of considered GOA and IGOA, the statistical analysis using the 20 inde-
pendent runs test data has to be taken into consideration. The PIMV seems to be a
better performance criterion, since it incorporates the mean values information from a
set of independent test runs. It may also be possible that the optimized values are
dispersed in a large range and the mean value gets close to the global minimum values.
To make this into consideration, the standard deviation of the optimized values has also
been taken into consideration, to find whether the algorithm has the necessary
repeatability feature in it or not. To showcase the effectiveness of IGOA over GOA in
this manner, the values of mean and standard deviation are listed in Table 2 for
different test functions. It may be noted from data provided in Table 2 that the standard
deviation values for IGOA are very close to zero, which tells that IGOA performance is
repeatable in nature. While GOA has a tendency to stick in local minima for some of
the test functions and thereby has a very high value of standard deviations for some test
functions. The high value of standard deviations for GOA can be observed particularly
for F2, F8-10, F13, F15-16 and F19.
Table 1. Best values as obtained by GOA and IGOA for considered test functions
S. Function Best optimized values from 20 trial runs Logarithmic
No. name GOA IGOA improvement
(LIBV)
F1 Ackley 1.3805E−06 8.6150E−08 1.204788
F2 Beale 4.92E−13 0 NA
F3 Bohachevsky 6.4208E−10 1.00E−14 4.807594
F4 Bohachevsky 5.9789E−10 4.30E−14 4.143159
1
F5 Bohachevsky 1.5328E−10 1.30E−14 4.071542
2
F6 Booth 1.047E−12 0 NA
F7 Branin 0.397887357 0.397887358 0.431408
F8 Easom −0.99999999994437 −0.99999999999999 4.444563
F9 Eggholder −959.640662720651 −959.64066272085 2.32E−06
F10 Goldstein 3.00000000000375 2.99999999999991 1.619052
Price
F11 Griewank 0.109267741 0.02546913 0.632478
F12 Matyas 5.00E−15 0 NA
F13 Rosenbrock 1.10129E−07 0 NA
F14 Schaffer 2.00E−15 0 NA
F15 Schwefel 2.545556E−05 2.545513E−05 7.40E−06
F16 Shubert −186.730908828 −186.730908831 NA
F17 Six-Hump −1.031628453 −1.031628453 NA
camelback
F18 Sumsquares 4.13675E−07 6.5711E−11 3.798625
F19 Zakharov 1.75422E−04 3.69E−13 8.677059
This data shows that the repeatability of the IGOA is higher than GOA. Further,
PIMV can indicate more about the efficacy of IGOA while optimizing different test
functions. It may be noted that the maximum value for PIMV can be 100 and the values
which are closer to the maximum PIMV support the excellent performance shown by
IGOA while optimizing different test functions. Positive PIMV values indicate that the
IGOA performance is better than the GOA; however, a value close to 100 indicates that
the IGOA performance is not only better than GOA but also the mean value obtained is
either very close or equal to the global optimum value for a particular benchmark test
function. The PIMV presented in Table 2 also indicates the supremacy of IGOA over
GOA for all considered benchmark test functions.
Table 2. Mean and standard deviation values for different functions

S. No. Mean values Standard deviations PIMV (%)
GOA IGOA GOA IGOA
F1 2.25E−05 3.37E−07 1.17E−05 1.27E−07 98.50
F2 0.330795669 7.95E−15 0.417640307 3.41E−14 100
F3 6.28E−09 6.92E−13 6.48E−09 6.13E−13 99.99
F4 2.36E−08 2.69E−12 2.60E−08 2.52E−12 99.99
F5 3.99E−08 1.01E−12 7.43E−08 7.37E−13 99.99
F6 4.59E−11 4.00E−15 8.32E−11 3.73E−15 99.99
F7 0.397887358 0.397887358 3.48E−11 1.70E−15 99.97
F8 −0.45 −1 0.510417785 2.79E−13 100
F9 −753.5810622 −954.674431 151.5400571 5.69E−02 97.58
F10 8.4 3 18.7880141 3.17E−14 100
F11 0.2065592 0.11519567 0.088591911 6.42E−02 44.23
F12 3.03E−12 0 6.94E−12 0 100
F13 0.240502042 2.01E−07 0.737820273 7.83E−07 99.99
F14 4.82E−13 0 3.81E−13 0 100
F15 169.76381516 2.55E−05 141.1030576 8.16E−12 99.99
F16 −172.9084996 −186.730908 42.54447783 7.01E−12 100
F17 −0.786779065 −1.03162845 0.383729877 7.73E−15 100
F18 4.04E−04 8.37E−06 8.57E−04 2.62E−05 97.92
F19 1.0128319 1.68E−07 3.965683405 6.09E−07 99.99
5 Conclusion
This work proposed an improved version of grasshopper optimization algorithm

(GOA). The improved GOA (IGOA) proposed in this work eliminates a few short-
comings of GOA. The variation of c factor throughout the optimization procedure in a
linear fashion is an important functionality of GOA which is claimed to provide a
balanced exploration and exploitation of the search space. This paper improves this
balance by proposing a new method for the variation of c factor which helps in
determining the optimum point more quickly and accurately. To further increase the
effectiveness of IGOA, another step which provides the random walks to grasshopper
in its swarm has also been proposed. This random walk further reduces the risk of
IGOA to get trapped in local minimum points. These two modifications have signifi-
cantly improved the efficacy of IGOA over GOA, and the same has been verified by
performing optimization procedure on 19 benchmark test functions of different nature.
To avoid any biasing towards a particular algorithm, 20 independent trial runs were
made and parameters like best, mean, and standard deviation were calculated. To
showcase the effectiveness of IGOA over GOA, different performance measures were
also considered, namely percentage improvement in mean value, logarithmic
improvement in best values. Based on rigorous experimentation and obtained perfor-
mance measures, it was observed that the IGOA was able to provide a significant
improvement over GOA and the mean values for all functions were very close to global
optimum values. Further, repeatability of IGOA was also substantially better than GOA
for all the considered test functions. Based on the obtained results, it may be concluded
that IGOA provides more accurate and repeatable results than GOA and can be applied
to real-world optimization problems.
References
1. Spall JC (2005) Introduction to stochastic search and optimization: estimation, simulation,
and control, vol. 65. John Wiley & Sons
2. Harik GR, Lobo FG, Goldberg DE (1999) The compact genetic algorithm. IEEE Trans Evol
Comput 3(4):287–297
3. Holland JH (1992) Genetic algorithms. Sci Am 267(1):66–73
4. Kennedy J (2011) Particle swarm optimization. In: Encyclopedia of machine learning.
Springer USA, pp 760–766
5. Poli R, Kennedy J, Blackwell T (2007) Particle swarm optimization. Swarm Intell 1(1):33–
57
6. Yang XS (2010) A new metaheuristic bat-inspired algorithm. In: Nature inspired cooperative
strategies for optimization (NICSO 2010). Springer, Berlin, Heidelberg, pp 65–74
7. Yang XS, Hossein Gandomi A (2012) Bat algorithm: a novel approach for global
engineering optimization. Eng Comput 29(5):464–483
8. Yang XS (2012) Flower pollination algorithm for global optimization. In: International
conference on unconventional computing and natural computation. Springer, Berlin,
Heidelberg, pp 240–249
9. Yang XS, Karamanoglu M, He X (2014) Flower pollination algorithm: a novel approach for
multiobjective optimization. Eng Optim 46(9):1222–1237
10. Mirjalili S, Mirjalili SM, Lewis A (2014) Grey wolf optimizer. Adv Eng Softw 69:46–61
11. Mirjalili S, Saremi S, Mirjalili SM, Coelho LDS (2016) Multi-objective grey wolf optimizer:
a novel algorithm for multi-criterion optimization. Expert Syst Appl 47:106–119
12. Yang XS, Deb S (2009) Cuckoo search via Lévy flights. In: Nature and biologically inspired
computing, 2009. NaBIC 2009. World Congress on. IEEE, pp 210–214
13. Yang XS, Deb S (2010) Engineering optimisation by cuckoo search. Int J Math Model
Numer Optim 1(4):330–343
14. Mirjalili S (2015) Moth-flame optimization algorithm: a novel nature-inspired heuristic
paradigm. Knowl-Based Syst 89:228–249
15. Mirjalili S (2015) The ant lion optimizer. Adv Eng Softw 83:80–98
16. Mirjalili S, Jangir P, Saremi S (2017) Multi-objective ant lion optimizer: a multi-objective
optimization algorithm for solving engineering problems. Appl Intell 46(1):79–95
17. Gandomi AH, Alavi AH (2012) Krill herd: a new bio-inspired optimization algorithm.
Commun Nonlinear Sci Numer Simul 17(12):4831–4845
18. Wang GG, Guo L, Gandomi AH, Hao GS, Wang H (2014) Chaotic krill herd algorithm. Inf
Sci 274:17–34
19. Dorigo M, Birattari M (2011) Ant colony optimization. In: Encyclopedia of machine
learning. Springer, Boston, MA, pp 36–39
20. Afshar A, Haddad OB, Marino MA, Adams BJ (2007) Honey-bee mating optimization
(HBMO) algorithm for optimal reservoir operation. J Franklin Inst 344(5):452–462
21. Fu Y, Ding M, Zhou C (2012) Phase angle-encoded and quantum-behaved particle swarm
optimization applied to three-dimensional route planning for UAV. IEEE Trans Syst, Man,
Cybern-Part A: Syst Humans 42(2):511–526
22. Mishra P, Kumar V, Rana KPS (2015) A fractional order fuzzy PID controller for binary
distillation column control. Expert Syst Appl 42(22):8533–8549
23. Saremi S, Mirjalili S, Lewis A (2017) Grasshopper optimisation algorithm: theory and
application. Adv Eng Softw 105:30–47
24. Chattaraj N, Ganguli R (2017) Multi-objective optimization of a triple layer piezoelectric
bender with a flexible extension using genetic algorithm. Mech Adv Mater Struct 1–9
25. Roberge V, Tarbouchi M, Labonté G (2013) Comparison of parallel genetic algorithm and
particle swarm optimization for real-time UAV path planning. IEEE Trans Industr Inf 9
(1):132–141
26. Mishra P, Kumar V, Rana KPS (2018) An efficient method for parameter estimation of a
nonlinear system using backtracking search algorithm. Eng Sci Technol Int J 21(3):338–350
27. Surjanovic S, Bingham D (2013) Virtual library of simulation experiments. https://www.sfu.
ca/*ssurjano/optimization.html. Last Accessed on 29 May 2018
Design of Array Antenna for Body Area
Network
Sarmistha Satrusallya, Shaktijeet Mahapatra,

and Mihir Narayan Mohanty(&)
Department of ECE, ITER, Siksha ‘O’ Anusandhan (Deemed to be University),

Bhubaneswar, India
{sarmisthasatrusalya,shaktijeetmahapatra,
mihirmohanty}@soa.ac.in
1 Introduction
Antennas with smaller profile and lightweight are developed nowadays to be used in
various applications. One of the applications is the development of wearable antenna.
These antennas are widely used in the field of medicine, military, and different sci-
entific community. The wearable antenna in the UWB range (3–10.6 GHz) provides
better performance in terms of bandwidth and gain.
Researchers have developed different antennas in the desirable band such as
coplanar waveguide fed disc antenna and microstrip fed annular slot antenna [1].
Authors also presented conical dielectric resonator antenna for use in body area net-
work operating in the range of (3.4–5 GHz). The antenna provides a good performance
in frequency and time domains [2]. A cavity slot antenna was proposed at 2.45 GHz
resulted in high efficiency of more than 50% in the vicinity of the human body [3].
Similarly, the radiation of three types of UWB antenna was experimentally charac-
terized. Among omnidirectional, monopole antenna and printed diversity antenna, the
latter exhibits improved quality of radio frequency link for on-body communication [4].
A CPW-fed ultra-wideband antenna and button-shaped antenna based on a microstrip
Yagi array were also presented for the same purpose. The antennas were suitable for
operation in the range of 2.45 GHz and had application in the fields of telemedicine [5,
6]. Kang et al. designed a folded antenna with reduced backward radiation and
proximity effect in the human bodies. A switchable beam textile antenna capable of
generating beam steering in four directions was also proposed. The antenna is attached
in safety jackets and rain coats to be used for search and rescue operation [7, 8].
Different array antennas were also proposed for on-body communication. Millimetre
wave spectrum has application in health monitoring and live streaming. A frequency
scanning array antenna was designed to provide reconfigurability in the direction of
main beam pattern [9–11].
In this paper, we present an array antenna of five elements with circular slots cut on
the centre of the patch. The antenna is small and is easy to fabricate. A defective
ground plane has been used for better performance.

https://doi.org/10.1007/978-981-15-2774-6_23
190 S. Satrusallya et al.
The paper is organized into the following sections: Sect. 2 presents the design of
the antenna; simulation results are described in Sect. 3, and in the last section, con-
clusions are drawn.
2 Design Considerations
The proposed antenna is an array having five elements. The substrate is chosen to be
FR4-epoxy having a relative permittivity of 4.4 and dielectric loss tangent of 0.02. The
rectangular patches are placed on a substrate having dimension 30 mm 180 mm
1.6 mm. Each rectangular patch has a circular slot of radius 2 mm. The ground plane
dimensions are 15 mm 180 mm. The array is fed by a microstrip fed with 50 X
impedance. Figure 1 shows the proposed antenna array (Table 1).
Fig. 1. Design of the proposed antenna array
Table 1. Design parameters

Parameters Values (in mm)
Length of substrate 180
Width of substrate 30
Height of substrate 1.6
Length of patch 16
Width of patch 18.7
Radius of the slot 2
Figure 2 shows the impedance bandwidth of the proposed design. The antenna res-
onates at two frequencies, 3.1 and 3.6 GHz. The impedance bandwidths are 0.2,
0.8 GHz with a return loss of −13 dB and −31 dB, respectively.
Design of Array Antenna for Body Area Network 191
Fig. 2. Impedance bandwidth of the proposed antenna
Figure 3 shows the VSWR of the design. It can be clearly seen that the VSWR at
the resonating frequencies lies between 1 and 2.
Fig. 3. VSWR of the proposed antenna
Figure 4 represents the gain of the antenna. The gain is found to be 3.85 db which
is suitable for on-body communication.
192 S. Satrusallya et al.
Fig. 4. 3D polar plot of the proposed design
4 Conclusions
An array antenna has been designed for wearable application in the range of (3.1–
5 GHz). The antenna resonates at 3.1 and 3.6 GHz. The defective ground structure
provides a better performance of the proposed design. The antenna can be used for on-
body and off-body communication.
References
1. Klemm M, Troester G (2006) Textile UWB antennas for wireless body area networks. IEEE
Trans Antennas Propag 54(11):3192–3197
2. Almpanis G, Fumeaux C, Frohlich J, Vahldieck R (2008) A truncated conical dielectric
resonator antenna for body-area network applications. IEEE Antennas Wirel Propag Lett
8:279–282
3. Haga N, Saito K, Takahashi M, Ito K (2009) Characteristics of cavity slot antenna for body-
area networks. IEEE Trans Antennas Propag 57(4):837–843
4. See TS, Chen ZN (2009) Experimental characterization of UWB antennas for on-body
communications. IEEE Trans Antennas Propag 57(4):866–874
5. Shin H, Kim J, Choi J (2009) A stair-shaped CPW-fed printed UWB antenna for wireless
body area network. In: 2009 Asia pacific microwave conference, pp 1965–1968
6. Khaleel HR, Al-Rizzo HM, Rucker, DG, Elwi TA (2010) Wearable Yagi microstrip antenna
for telemedicine applications.In: 2010 IEEE Radio and Wireless Symposium (RWS),
pp 280–283
7. Kang CH, Wu SJ, Tarng JH (2011) A novel folded UWB antenna for wireless body area
network. IEEE Trans Antennas Propag 60(2):1139–1142
Design of Array Antenna for Body Area Network 193
8. Jais MI, Jamlos MFB, Jusoh M, Sabapathy T, Kamarudin MR, Ahmad RB, Ishak NLK
(2013) A novel 2.45 GHz switchable beam textile antenna (SBTA) for outdoor wireless
body area network (WBAN) applications. Prog Electromagn Res 138:613–627
9. Park J, Tak J, Choi J (2014) A low-profile dipole array antenna for on-body communi-
cations. In: 2014 IEEE international workshop on electromagnetics (iWEM), pp 4–5
10. ElSalamouny MY, Shubair RM (2015) Novel design of compact low-profile multi-band
microstrip antennas for medical applications. In: 2015 loughborough antennas & propaga-
tion conference (LAPC), pp 1–4
11. Chauhan SS, Abegaonkar MP, Basu A (2017) Millimeter wave frequency scanning antenna
for body centric communication. In: 2017 international symposium on antennas and
propagation (ISAP), pp 1–2
Optimal Load Condition for a Resonant
Wireless Charging System
Jagadish C. Padhi, Durga P. Kar, Praveen P. Nayk(&),

Biswaranjan Swain, Dipti Patnaik, and Satyanarayan Bhuyan

{praveennayak,satyanarayanbhuyan}@soa.ac.in
1 Introduction
Recent years have witnessed the vast improvement in the wireless charging system for
low-power electrical/electronic appliances of few mW to high-power application such
as electric vehicle (EV) of kW level. This is due to the convenience, safety, and utmost
suitable in all weather conditions [1–8]. Charging can be carried out through con-
ventional inductive coupling but not so far achieved its height due to the short-range
energy transfer efficiency and limitation in coil size. Providentially, the alternative in
the form resonant inductive coupling-based wireless charging system came into the
picture to enhance the power transfer efficiency as well as the transfer range with
adequate dimension that can be embedded into the vehicle [9–13]. But to make the
wireless charging system more practical, it is important to the issues dealt with the
operating frequency, load, and the power transfer range; there is a need of experimental
investigation. So, in order to mitigate the issues related to transfer load power and
transfer efficiency, the optimal load condition is very much essential. Hence, load
dependencies of power transfer efficiency have been experimentally investigated in this
paper. The solution will provide a noteworthy advancement in the field of wireless
charging under non-ideal scenario.
2 Experimental Setup
The experimental setup consists of a plastic rod graded with a scale, mounted on a base
along which circular acrylic plates which can slide up or down the plastic rod. On these
plates, required coils of various dimensions can be mounted. This setup makes it
convenient to swap coils and take experimental results with accuracy. The material
used in making this apparatus is non-magnetic and does not affect the experiment in
any way. Copper conductor of gauge SWG12 (1.3208 mm radius) is wound into
circular coils, and capacitors are connected to the ends of the transmitting and receiving

https://doi.org/10.1007/978-981-15-2774-6_24
Optimal Load Condition for a Resonant Wireless Charging System 195
coils. A high-frequency function generator is used as a power source to drive the source
coil. The dimensions of the coil are chosen such that they are easy to handle and
capable of demonstrating wireless power, fixed as two cylindrical coils of 7.5 cm
radius made of copper wire SWG12 (conductor radius 0.13208 cm), with 5 turns each
(Fig. 1).
Fig. 1. Experiment setup of wireless power transfer system
The value of inductance is 6.8µH which gives a resonant frequency of 1.93 MHz
with a capacitance of 1000pF, but experimentally, it is later shown that resonant
frequency is 2.05 MHz. This small discrepancy is due to the structures not being
perfect cylinders. With this setup experiments are conducted to determine the effect of
size, shape, distance, and frequency on efficiency of power transfer.
196 J. C. Padhi et al.
Experiments are conducted to observe the variation of output power with regard to
three parameters, namely frequency, load, and distance. The frequency dependence on
the output power is shown in Fig. 2.
output power vs frequency

0.7
0.6
Output power (watts)
0.5
0.4
0.3
0.2
0.1
0
1.5 2 2.5
frequency (MHz)
Fig. 2. Frequency characteristics of output power at a fixed load
In this case, the frequency is varied with the load set at 220 X and distance at
10 cm. It is expected that at resonance frequency, output power is maximum since
evanescent coupling would take place between the resonators. The graph proves that
the resonant frequency for this system is 2.05 MHz.
output power vs load

1.2
1
output power (watts)
0.8
0.6
0.4
0.2
0
0 2000 4000 6000 8000 10000
Load resistance (in ohms)
Fig. 3. Electric load characteristics of output power
The electric load characteristic is shown in Fig. 3. The load is varied from 10 to
10,000 X with the frequency set at 2.05 MHz, and distance between the transmitter and
receiver is set at 10 cms. It is observed that maximum output power is achieved at
220 X (Fig. 4).
output power vs distance

0.9
0.8
0.7
output power (watts)

0.6
0.5
0.4
0.3
0.2
0.1
0
1 6 11 16 21
distance (cms)
Fig. 4. Dependency of output power with distance
The distance between the transmitter and receiver is varied from 1 to 20 cm with
the frequency of power supply set at 2.05 MHz and the load resistance set at 220 X.
From the result, it is clearly observed that output power initially increases and attains its
maximum at a separation distance of 14 cm and then gradually decreases afterward.
This confirms the counter-EMF effect in the experimental setup. This also shows a key
difference from simple electromagnetic induction where maximum power output would
be achieved when distance is minimum. From the graph, it is noted that at 14 cm
distance the output power is maximum at 0.8 W.
The power transfer efficiency corresponding to frequency, separation distance, and
load is discussed and given below. The frequency characteristic of power transfer
efficiency is illustrated in Fig. 5. The high-frequency signal from the signal generator is
varied with the load set at 220 X and distance at 10 cm. It is found that at resonance
frequency, efficiency attains its maximum since evanescent coupling would take place
between the resonators and which confirms that the resonant frequency for this system
is 2.05 MHz.
Efficiency vs frequency
80
70
60
50
Efficiency
40
30
20
10
0
1.5 1.7 1.9 2.1 2.3 2.5 2.7
frequency (MHz)
Fig. 5. Frequency characteristics of wireless power transfer efficiency

Efficiency vs load
80.00%
70.00%
60.00%
Efficiency 50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
0 2000 4000 6000 8000 10000
Load resistance (ohms)
Fig. 6. Dependency of efficiency on the load resistance
The load variation corresponding to the power transfer efficiency is shown in

Fig. 6. The load is varied from 10 to 10,000 X with the frequency set at 2.05 MHz and
distance between the transmitter and receiver set at 10 cms. It is observed from the
result that the maximum efficiency is achieved at 220 X. This gives the indication that
power levels do not change the resonance peak. This value indicates the maximum
output power point.
Efficiency vs distance
80
70
60
efficiency (%)
50
40
30
20
10
0
1 6 11 16 21
distance (cms)
Fig. 7. Effect of separation distance on efficiency
The effect of separation distance on the power transfer efficiency is shown in Fig. 7.
The distance between the transmitter and receiver is varied from 1 to 20 cm with the
frequency of power supply set at 2.05 MHz and the load resistance set at 220 X. It is
observed that efficiency initially increases and starts decreasing then. This confirms the
counter-EMF effect in the experimental setup. In this case efficiency of 81% is achieved
at about 10 cm distance.
From the above three experiments, it is concluded that a maximum efficiency of
73% is achieved at the resonance frequency of 2.05 MHz with load resistance 220 X
and 10 cm separation.
Efficiency vs load
80.00%
70.00%
60.00%
Efficiency
50.00%
40.00% 1.44 amplitude
2.2 amplitude
30.00%
2.8 amplitude
20.00%
10.00%
0.00%
0 5000 10000
Load resistance (ohms)
Fig. 8. Effect of load on efficiency for different signal amplitude
The effects of load on efficiency for different signal amplitude are shown in Fig. 8.
Comparing to the previous result, the difference is the input power is increased by
increasing the amplitude in the signal generator which is then amplified 10 times by
the power amplifier. It was observed that increasing power levels has minimal effect on
efficiency and the small reduction in efficiency is due to the heating of the ceramic
capacitor and wires which increases the resistance of the circuit.
4 Conclusion
The power transfer to load and the power transfer efficiency corresponding to fre-
quency, load, and separation air gap have been experimentally investigated. The power
transfer efficiency is measured for different load values. It is found that both the power
delivered to the load and the power transfer efficiency are very much sensitive to
electrical load connected to the receiver coil. Moreover, the resonance frequency is
seen to be independent of the power level. From the result, it is also confirmed that
there exists an optimal load for which maximum power transfer is achieved. From the
experimental result, the maximum efficiency 73% is achieved for an optimal load
220 X at the resonant frequency 2.05 MHz for a separation distance of 10 cm.
References
1. Pedder DAG, Brown AD, Skinner JA (1999) A contactless electrical energy transmission
system. IEEE Trans Ind Electron 46:23–30
2. Chang-Gyun K, Dong-Hyun S, Jung-Sik Y, Jong-Hu P, Cho BH (2001) Design of a
contactless battery charger for cellular phone. IEEE Trans Industr Electron 48:1238–1247
3. Covic GA, Boys JT (2013) Modern trends in inductive power transfer for transport
applications. IEEE J Emerg Sel Top Power Electron 1(1):28–41
4. Li S, Mi CC (2015) Wireless power transfer for electric vehicle applications. IEEE J Emerg
Sel Top Power Electron 3(1):4–17
5. Kurs A, Karalis A, Moffatt R, Joannopoulos JD, Fisher P, Soljacic M (2007) Wireless power
transfer via strongly coupled magnetic resonances. Science 317:83–86
6. Ho SL, Wang J, Fu W, Sun M (2011) A comparative study between novel witricity and
traditional inductive magnetic coupling in wireless charging. IEEE Trans Magn 47:1522–
1525
7. Kuipers J, Bruning H, Bakker S, Rijna H (2012) Near field resonant inductive coupling to
power electronic devices dispersed in water. Sens Actuators, A 178:217–222
8. Sample AP, Meyer DA, Smith JR (2011) Analysis, experimental results, and range
adaptation of magnetically coupled resonators for wireless power transfer. IEEE Trans
Industr Electron 58(2):544–554
9. Hui SYR, Ho WWC (2005) A new generation of universal contactless Battery Charging
platform for portable Consumer Electronic equipment. IEEE Trans Power Electron 20:620–
627
10. Mon G, Arcuti P, Tarricone L (2015) Resonant inductive link for remote powering of
pacemakers. IEEE Trans Microw Theory Tech 63(11):3814–3822
11. Orekan T, Zhang P, Shih C (2018) Analysis, design, and maximum power-efficiency
tracking for undersea wireless power transfer. IEEE J Emerg Sel Top Power Electron 6
(2):843–854
12. RamRakhyani AK, Mirabbasi S, Chiao M (2011) Design and optimization of resonance-
based efficient wireless power delivery systems for biomedical implants. IEEE Trans
Biomed Circuits Syst 5:48–63
13. Kar DP, Nayak PP, Bhuyan S (2015) Bi-directional magnetic resonance based wireless
power transfer for electronic devices. Appl Phys Lett 107(13):3901
Design of Microstrip Antenna
with Corner Truncation for UWB
Application and Its Analysis
Swarnaprava Sahoo and Laxmi Prasad Mishra(&)

Institute of Technical Education and Research, Siksha ‘O’ Anusandhan
(Deemed to be University), Bhubaneswar, Odisha, India
laxmimishra@soa.ac.in
1 Introduction
Nowadays, the microstrip patch antenna is very much important in wide range of
multifunctional wireless communication systems like UWB, WLAN, spacecraft, radar,
and applications of telemetry in biology. These systems have motivated the invention
of miniaturizing UWB radiators with high gain, good radiation efficiency, and patterns.
The possibilities of abilities of these applications are because of popular advantages
like low profile, lightweight, compact in size, planar configuration, easy fabrication,
inexpensive manufacturing, and integration with other microwave circuitries. The main
disadvantage of patch antenna is its narrow bandwidth. So, many researchers have
developed different familiar methods for improvement of microstrip antenna impe-
dance bandwidth by sustaining the patch radiator size as compact as feasible, by
applying thicker substrate and low permittivity value [1], inserting slots in radiating
patch [2], using different feeding methods and impedance matching [3], by ground
plane modification [4]. The impedance bandwidth and radiating patch radiator size are
generally complementary contrary to each other. So, the characteristics development of
one normally results in deterioration of the other. Under these conditions, the research
workers developed many different methods to enhance such limitations, basically
improving the impedance bandwidth. So, different antenna designs are proposed. They
are inserting slot and analysis of U-slot loading patch [5], by loading stacked patch [6].
The impedance bandwidth of above antennas is up to 40%.
The integration of all the important characteristics like high gain, wide bandwidth,
and good radiation pattern into a single antenna design is very complicated task. Some
limitations are overcome by this innovative work which is based on ultra-wideband U-
slot insertion and incorporation of two identical corner truncations in metallic patch. In
this research article, the proposed antenna is analyzed for improvement of impedance
bandwidth. The antenna characteristics are analyzed on the basis of corresponding
circuit approximation, cavity model analysis, and circuit theory concept. The coaxial
feed is placed nearby center of metallic patch for better excitation of the intended
radiator over wide bandwidth. The parameters like gain and radiation patterns of the
intended radiator are also deliberated.

https://doi.org/10.1007/978-981-15-2774-6_25
202 S. Sahoo and L. P. Mishra
2 Antenna Geometry
Figure 1 presents the geometrical consideration of the intended radiator with optimized
dimensions. The intended radiator consists of an inserted U-slot and a pair of truncated
corners rectangular microstrip patch. The dimension of the rectangular patch is L
W ¼ 35:5 26 mm2 with center frequency 7.2 GHz. The truncated length is dl =
9.5 mm. The rectangular patch is designed on Rogers R04350 (tm) dielectric substrate
(size: 71 52 mm2 Þ with dielectric constant ðer Þ 3.6 of height (h) 6 mm, loss tangent
ðtan dÞ 0.004, and full ground plane. The U-slot of dimensions ðlvs whs t ¼ 17:3
12 2:1 mm3 Þ is inserted in the center of the patch. The three arms of the U-slot are
symmetrically placed w.r.t the feed point yf = 1.5 mm.
3 Theoretical Consideration
Basedon cavity expansion model [7], the equivalent circuit of the radiating element is
Rpatch Lpatch Cpatch as shown in Fig. 2.

Rpatch ¼ Qpatch = xpatch Cpatch ; Lpatch ¼ 1= x2patch Cpatch ;
ð1Þ
Cpatch ¼ ½e0 eeff LW=2hcos2 ðpyf =LÞ
pffiffiffiffiffiffi
where xpatch = angular design frequency, Qpatch ¼ c eeff =4fh, and f = design
frequency.
Fig. 1. Proposed antenna geometry for UWB application

Design of Microstrip Antenna with Corner Truncation … 203
Fig. 2. Circuit of rectangular microstrip antenna
The impedance of the rectangular microstrip radiator is:

Zpatch ¼ 1= 1=Rpatch þ 1=xLpatch þ jxCpatch ð2Þ
In rectangular patch shown in Fig. 1,
DSTC ¼ Dl2TC ð3Þ
The effect of capacitance due to truncating corner is:
CTC ¼ e0 eeff DSTC =h ð4Þ
The impedance of the truncated rectangular microstrip radiator is:

Zpatch ¼ Rpatch þ jXpatch ¼ 1= 1=Rpach þ 1=jxLpatch þ jxCT ð5Þ
where CT ¼ Cpatch þ CTC .

The corresponding circuit of the inserted U-slot presented in Fig. 3 is based on
duality relationship of slot and dipole. The resonant frequency is changed due to the
insertion of U-slot in the radiating element that changes the radiator electrical
dimension. The path flow length increases due to the current flow throughout the
inserted slot and on the boundaries. This consequence can be modeled as series con-
nection of radiation resistance (R) and reactive component (X) of the two perpendicular
and one parallel arms of the inserted U-slot. Then, the individual series combination
will be the parallel connection of radiation resistance and reactive components
ðRvs ; Xvs and Rhs ; Xhs Þ. So, the corresponding circuit of the inserted U-slot can be
evaluated [8] which is presented in Fig. 3.
Fig. 3. Corresponding circuit of inserted U-slot
In this research work, for the analysis of inserted U-slot rectangular microstrip
patch radiator, the capacitive reactance is taken parallel to the reactance of the radiating
element. So, the impedance of inserted U-slot can be evaluated [7] as:
ZUslot ¼ zvs zhs =ðzvs þ 2zhs Þ ð6Þ
where zvs and zhs are impedances of the two vertical arm slots and one horizontal arm
slot, respectively.
The corresponding circuit of the intended radiator can be represented as the parallel
connection of truncated radiating element and inserted U-slot as given in Fig. 4. So, the
input impedance of the intended radiator is given as:

Zpatch ZUslot ¼ Zpatch ZUslot = Zpatch þ ZUslot ð7Þ
The radiation characteristics of the intended radiator can be determined [9] from
Eqs. (8) and (9). The gain of the proposed antenna can be obtained [7, 10] from Eq. (10).
jK0 WV0 ejk0 r sinððk0 W=2Þ sin h sin uÞ

Eh ¼ cosðkh cos hÞ
pr ðk0 W=2Þ sin h sin u

ð8Þ
k0 L
cos sin h sin u cos u for 0 h p=2
2
jk0 wv0 ejk0 r sinððk0 W=2Þ sin h sin uÞ

Eu ¼ cosðkh cos hÞ
pr ðk0 W=2Þ sin h sin u

ð9Þ
k0 L
cos sin h sin u cos u sin u for 0 h p=2
2
pffiffiffiffi
where v0 = radiating edge voltage, r = distance of an arbitrary point, k0 ¼ K er ,
K ¼ 2p=k,
G¼gD ð10Þ
Fig. 4. Equivalent circuit of the intended antenna
where,
g ¼ efficiency
of antenna
¼ Radiated power; Pr =Input power; Pt ,
Pr ¼ 1= 2xpatch Cpatch and Cpatch is defined in Eq. (1),

Pt ¼ 1= xpatch Cpatch ,
D = directivity of the proposed antenna = 4W 22 P=I1 k20 ,
Rp
I1 ¼ 0 sin2 ððk0 W cos hÞ=2Þ tan2 h sin hdh:
The return loss ðS11 Þ value less than −10 dB of microstrip patch radiator without U-slot
varies from 2.5 to 2.73 GHz (simulated B.W.% = 8%) and 2.54 to 2.98 GHz (theo-
retical B.W.% = 15.94%) with center frequencies at 2.6 and 2.71 GHz which is
ðfl þ fh Þ=2, where fl and fh are lower and upper limit frequencies, respectively, pre-
sented in Fig. 5a, whereas the corresponding impedance bandwidths of the proposed
antenna determined from -10 dB are 10 GHz: 120.48% from 3.3 to 13.3 GHz (theo-
retical), 7.6 GHz: 105.5% from 3.4 to 11 GHz (simulated), respectively, presented in
Fig. 5b. The resonant frequencies are spaced nearly to each other which provide UWB
characteristic. This bandwidth is also applicable for S-, C-, X-, and Ku-band modern
wireless communication systems. Also, the impedance bandwidth of proposed antenna

has been improved from 15.94% to 120.48%. Some discriminates occur in between
theoretical and simulated results. This is due to as the suggested assumption is suitable
for thin material but in this research work large material thickness is used which
influences the radiation resistance and feed inductance. The larger impedance band-
width is due to the insertion of U-slot in the radiating element.
Figure 6 shows that gain of the proposed antenna varies from 1.5 to 5.5 dBi
(theoretical), 1.6 to 5.9 dBi (simulated) over the whole impedance bandwidth.
Fig. 5. Simulated and theoretical return loss plot against frequency a Rectangular microstrip
patch antenna, b intended antenna
Fig. 6. Gain plot against frequency of intended antenna
Figure 7 presents the radiation patterns of the intended radiator which are quite
stable throughout the operating band. The E-plane cross-polarization level is higher
than that of H-plane due to feed location on the Y-axis. In H-plane, the slight increase
of cross-polarization level is observed at 3.8 GHz. The corner truncation in rectangular

radiating element may create the small variance among the suppositional and simula-
tion radiation characteristics. Again, the small broadside nature of the radiation patterns
is also changed due to the assumptions proposed in the theory.
(a) (b)
Fig. 7. Simulated and measured radiation characteristics of the intended radiator at resonant
frequencies 9.9 GHz
5 Conclusion
This research article demonstrates bandwidth enhancement analysis of microstrip

antenna with equivalent circuit approximate using cavity expansion model for ultra-
wideband application. The intended radiator is successfully designed and simulated
with bandwidth 105.5% (simulated: from 3.4 to 11 GHz), 120.48% (theoretical: from
3.3 to 13.3 GHz). The proposed antenna with large bandwidth is suitable for multi-
functional wireless S-band (from 2 to 4 GHz), C-band (from 4 to 8 GHz), X-band
(from 8 to 12 GHz), and Ku-band (from 12 to 18 GHz) applications.
References
1. Schaubert DH, Pozar DM, Adrian A (1989) Effect of microstrip antenna substrate thickness
and permittivity: comparison of theories and experiment. IEEE Trans Antennas Wirel
Propag Lett 37:677–682
2. Sahoo S, Mohanty MN, Mishra LP (2018) Bandwidth improvement of compact planar
antenna for uwb application with dual notch band performance using parasitic resonant
structure. Prog Electromagnet Res M 66:29–39
3. Shakib MN, Moghavvemi M, Mahadi WNL (2015) A low profile patch antenna for ultra
wideband application. IEEE Antennas Wirel Propag Lett 14:1790–1793
4. Sahoo S, Mishra LP, Mohanty MN, Mishra RK (2018) Design of compact uwb monopole
planar antenna with modified partial ground plane. Microwave Opt Technol Lett 60:575–578
5. Khidre A, Lee KF, Elsherbeni AZ, Yang F (2013) Wide band dual beam U-slot microstrip
antenna. IEEE Trans Antennas Propag 61(3):1415–1418
6. Yang W, Zhou J, Yu Z, Li L (2014) Single fed low profile broadband circularly polarized
stacked patch antenna. IEEE Trans Antennas Propag 62(10):5406–5410
7. Wolf EA (1988) Antenna analysis. Artech House, Norwood (USA)
8. Narayan S, Sharma S, Vishvakarma BR (2005) Analysis of slot loaded rectangular
microstrip patch antenna. Indian J Radio Space Phys 34:424–430
9. Derneryd AG, Lind AG (1979) Extended analysis of rectangular microstrip resonator
antennas. IEEE Trans Antennas Propag 27(6):846–849
10. Kumar G, Ray KP (2003) Broadband microstrip antennas. Artech House, Norwood, MA
Analysis of Vector-Controlled Onshore DFIG
with STATCOM for Power Compensation
Sovit Kumar Pradhan1(&), Radhakrishna Das2, Debswarup Rath3,

and S. S. P. B. K. Prasad3
1
Department of Electrical Engineering, NIT, Silchar, India
2
Department of Electrical Engineering, GIET, Bhubaneswar, India
3
Department of Electrical & Electronics Engineering, GIET, Bhubaneswar,
India
1 Introduction
With ever-increasing desire for variable speed operation, the doubly fed induction
generator (DFIG) is used along with wind turbines. This system is in great demand.
Voltage control and reactive power management are the concerns for power system,
while grid-connected DFIG is also affected due to continuous voltage fluctuations.
All equipment terminal voltages in the system are within suitable and operatable
boundary conditions. Both the equipment from utilities and customer are designed to
function at a predefined standard operatable voltage level. Overuse or operation of the
equipment at voltages outside the permissible or boundary limit array should unfa-
vorably upset their performance and cause them harm [1–3].
System steadiness is improved to make the most use of transmission system
arrangement. control have an important role For system stability analysis, voltage and
reactive power are major concern for researcher [4].
The reactive power flow is reduced to decrease secondary resistance and secondary
reactance to applied lowest. This safeguards the transmission system operations pro-
ficiently for active power transfer.
Because of nonlinear quality of transmission line due to loading, it consumes or
absorbs reactive power [5]. This creates problematical situations for transmission lines.
2 Background
2.1 Doubly Fed Induction Generator (DFIG)
A doubly fed induction generator (DFIG) exhibits more advantages in comparison with
other generators that applied in wind turbines [6]. In DFIG, as its name suggests rotor
and stator both are connected with grid for supply of power irrespective of synchronous
or subsynchronous operations [7]. The rotor currents create the magnetic field in the
rotor, and the rotor magnetic field interacts with the stator magnetic field to develop
torque in the machine. The torque is produced due to above interaction of magnetic
fields and can be considered as vector multiplication of the stator field and rotor field.
https://doi.org/10.1007/978-981-15-2774-6_26
210 S. K. Pradhan et al.
2.2 DFIG Operation

The AC–DC–AC converters are the voltage-source converters used on the connection
from rotor to grid named as rotor-side converter (RSC) and grid-side converter (GSC).
This is also known as back-to-back converter. For the reduction of ripple voltage
between converters, a DC-link capacitor is placed [7]. VSC with IGBT application are
used for this purpose [8]. Simulations have been performed with considering converter
on both sides (Fig. 1).
Fig. 1. Wind turbine and doubly fed induction generator system [7]
2.3 Static Synchronous Compensator

The static synchronous compensator (STATCOM) is a shunt-connected reactive
compensation equipment preferred in DFIG as it provides better result than shunt
capacitors. In STATCOM, power electronic switching devices are used to provide
controllability in magnitude and phase angle of the voltages. It also gives support to
bus voltages with modulation results improved transient characteristics such as better
transient stability limit and reduced oscillations in response.
Usually, a STATCOM is implemented in the networks that have reduced power
factor and less voltage stability. The active power capability of the network can be
increased with placement of energy storage device across the DC capacitor.
The STATCOM also provides better reactive power support in comparison with SVC
in case of low voltage.
2.4 Reasons for Selection of STATCOM

Shunt capacitors in wind firm can provide required reactive power but not responsive to
voltage change. At higher than the voltage limit of 1.05 pu, the required reactive power
compensation cannot be provided through fixed capacitor [9].
In case of imbalances in reactive power, which critically affects the entire power
systems along with consumers should be normalized. The STATCOM shows better
result regarding compensation to grid [10]. In this paper, PWM technique applied to
VSC STATCOM is planned to soothe grid-connected DFIG-based wind turbines.
Analysis of Vector-Controlled Onshore DFIG with STATCOM … 211
2.5 Voltage-Source Converter in DFIG

For improvement, most DFIG converters use a two-level six-switch converter as shown
in Fig. 2. A two-level converter output voltage has two levels of voltage. The insulated
gate bipolar transistors (IGBTs) are considered as most appropriate switch for this
operation. The output voltage of six-switch converter should have magnitude and
frequency within the determined value that the max value of voltage is less than the
DC-link voltage. The switching frequency of PWM converter is considered in such a
way that control of output voltage is possible.
Fig. 2. Vector structure of rotor-side converter [7]
2.6 Vector Control of Both Side Converter Circuit

The Rotor-Side Converter (RSC): The role of the rotor-side converter is to administer
the rotor currents effectively to produce necessary torque due to interaction between
fluxes produced at stator and rotor. The torque is normalized the wind turbine output
power. In order to achieve maximum power or track the maximum power point of
DFIG, the controller of power is required. A proportional–Integral (PI) regulator or
controller is implemented to minimize the error to zero at current loop. The controller
output is considered as reference for rotor current that added in the winding current of
rotor through rotor-side converter [7]. This quadrature part controls the electro- mag-
netic torque. With comparison between irq and irqref, the error can be found out and
the fault is reduced to zero.
The Grid-Side Converter (GSC): The grid-side converter targets to achieve the
proper control of the voltage. Furthermore, it is allowable to take up reactive power for
voltage holdup requirements with two control closed paths. Outer closed path consists
of path through the voltage regulator. The icdref is the reference current for regulator
[7]. The inner closed path is through the current regulator, and current regulation can be
performed with reference to both direct and quadrature axis dc current (Fig. 3).
Fig. 3. Vector structure of grid-side converter [7]
3 Simulation and Results
3.1 Simulation Model

Vector control of DFIG has been performed for analysis purpose of output voltage. The
value of parameters used in simulation analysis is considered with real values of
99 MW capacity of Damanjodi windmill, India (Fig. 4).
Proportional–integral controller is used in DFIG controller for better result as PI
controller improves dynamic response. Here, vector control is implemented so that
active power and reactive power can be controlled by current components. Phase and
the magnitude of the three phase are calculated and fed to the PI controller. Harmonics
can be compensated by using this model. This stator flux field-oriented control
improvises the output by diminishing the fluctuations. Simulations have been per-
formed using MATLAB software.
3.2 Result Analysis

Figure 5a shows simulated result of compensated graph of STATCOM applied to
DFIG, and it indicates the maximum value of 400 occurs at time 0. 04 s. The minimum
value is around 264 at 0.15 s (approx). The setting time is 0.02 s. In Fig. 5b, rotor
angle characteristics can be analyzed and steady-state condition is achieved at 0.04 s.
The DC output of grid-side converter is shown in the next figure, and the difference in
between maximum voltage and minimum voltage decreases as the time increases. From
that, it can be assumed that the total harmonic distortion is also decreasing. The average
voltage value is near about 1150 V.
From the above discussion, stator flux-oriented control of PI controller is working
better in DFIG with STATCOM for power compensations. The DC output voltage
fluctuation is approximately 22 V after 1 s with an average output voltage of 1150 V.
The frequency of rotor angle is remaining stable in this vector-controlled model.
Fig. 4. a Simulation model of DFIG with STATCOM, b proportional–integral controller used in

DFIG, c subsystem model of STATCOM
Fig. 5. a Compensate graph of the STATCOM, b load angle characteristics with respect to time,
c DC output voltage of converter, d rotor angle characteristics
4 Conclusion
In this paper, vector control strategy to PI controller with reactive and active power
compensation is performed. The control of DC bus voltage and reactive power is
achieved through controlling the rotor current with help of back-to-back converter. The
model includes the variation in wind speed for the simulation of characteristics of the
wind turbine. In relation with the results, this model is capable of delivering stable
output voltage with maintaining the power quality and this is achieved by eliminating
harmonics of the utility grid by vector control. For future work, various experiments
will be performed to verify the other uses of PI regulator with vector control schemes.
References
1. Yasa Y, Sozer Y, Mese E (2013) Harmonic analysis of doubly fed induction generator based
utility interactive wind turbine systems during fault conditions. In: IEEE energy conversion
congress and exposition, pp 2270–2276
2. Fan L, Yuvarajan S, Kavasseri R (2010) Harmonic analysis of a DFIG for a wind energy
conversion system. IEEE Trans Energy Conver 25(1):181–190, March 2010
3. Lindholm M, Rasmussen T (2003) Harmonic analysis of doubly fed induction generators. In:
Proceedings of the 5th international conference on power electronics and drive systems, Nov
2003, vol 2, pp 837–841
4. Xu L, Wang Y (2007) Dynamic modeling and control of DFIG-based wind turbines under
unbalanced network conditions. IEEE Trans Power Syst 22(1):314–323
5. Das RK, Pradhan SK, Pati SK (2018) An investigation of harmonics in grid connected
DFIG-WT experiencing transmission faults. In: International IEEE conference on technolo-
gies for smart city energy security and power (ICSESP-2018), Bhubaneswar, India. https://
doi.org/10.1109/icsesp.2018.8376689, Mar 2018
6. Muller S, Deicke M, Doncker RWD (2002) Doubly fed induction generator systems for
wind turbine. IEEE Ind Appl Mag 8(3):26–33
7. Fletcher J, Yang J (2010) Introduction to doubly-fed induction generator for wind power
applications. University of Strathclyde, Glasgow, UK
8. Abo-Khalil AG, Lee D-C, Jang J-I (2007) Control of back-to-back PWM converters for
DFIG wind turbine systems under unbalanced grid voltage. In: Proceedings of IEEE
international symposium on industrial electronics, pp 2637–2642
9. Petersen L, Kryezi F, Iov F (2017) Design and tuning of wind power plant voltage controller
with embedded application of wind turbines and STATCOMs. IET Renew Power Gener 11
(3):216–225
10. Izadpanahi N et al (2017) DFIG-based wind turbine using STATCOM for improvement
performance. In: 2017 14th International Conference on Electrical Engineering/Electronics,
Computer, Telecommunications and Information Technology (ECTI-CON). IEEE
11. Pradhan SK, Srivastava RK (2013) Analysis of PMSG under stator turn to turn fault using
winding function theory. In: International conference on condition assessment techniques
organised by IEEE DEIS Kolkata, India, pp 116–121. https://doi.org/10.1109/catcon,
6737483, 6–8 Dec 2013
12. Liao Y, Ran L, Putrus GA, Smith K (2003) Evaluation of the effects of rotor harmonics in a
doubly-fed induction generator with harmonic induced speed ripple. IEEE Trans Energy
Convers 18(4):508–515
13. Kheshti M, Kang X, Song G, Jiao Z (2015) Modeling and fault analysis of doubly fed
induction generators for Gansu wind farm application. IEEE J Mag 38:52–64
Dictionary Design for Block-Based
Intra-image Compression
Arabinda Sahoo1(&) and Pranati Das2

1
Department of ECE, ITER, Siksha ‘O’ Anusandhan Deemed to be University,
2
Department of Electrical Engineering, Indira Gandhi Institute of Technology,
Sarang, Odisha, India
1 Introduction
An effective dictionary can lead to efficient sparse approximation. The dictionary can
be fixed, generated from pre-defined sets of functions or designed from the learning
sets. The fixed dictionary is efficient for sparse approximation of certain classes of
signals, but quite disappointing for others. The shortcomings of the fixed dictionary
have paved the way for another versatile dictionary called learned dictionary. The
performance of block-based intra-image compression scheme depends on how well the
residuals obtained from intra-prediction are encoded. Transform-based intra-
compression method uses fixed DCT dictionary for encoding of prediction residuals.
However, DCT dictionary is not suited for efficient encoding of prediction residuals
with complex and non-periodic characteristics.
In recent years, various learned dictionaries have been developed by researchers for
image compression. Bryt and Elad [1] designed a dictionary based on K-SVD algo-
rithm for facial image compression. Zepeda et al. [2] proposed an iteration-tuned
dictionary (ITD). Zhan et al. [3] proposed a dictionary for compression of complex
synthetic aperture radar (SAR) images. Shao et al. [4] proposed a dictionary for fin-
gerprint compression. Zhu et al. [5] proposed a dictionary for surveillance image
compression. Most of the above methods are related to some specific class of images.
These dictionaries are not suitable for compression of prediction residuals.
Recently, Sahoo and Das [6] developed a dictionary-based image compression for
general arbitrary images. Experimental result shows the compression efficiency using
this scheme is better than other standard methods but it is computationally complex.
This motivated us to design a dictionary for block-based intra-image compression to
reduce the computational complexity.
This paper presents an algorithm to design an over-complete residual dictionary,
denominated as residual dictionary, suitable for encoding prediction residuals. The
proposed over-complete residual dictionary is embedded in block-based intra-image
compression framework to encode the residuals [6]. Comparative analysis demon-
strates that the proposed dictionary design method yields superior performance in terms
of compression efficiency and computational complexity as compared to other standard
methods.

https://doi.org/10.1007/978-981-15-2774-6_27
218 A. Sahoo and P. Das
The rest of the paper is organized as follows. Section 2 introduces the basics of
dictionary learning. Section 3 presents the proposed algorithm to design an over-
complete residual dictionary. Results and comparative analysis are discussed in Sect. 4.
Finally, Sect. 5 presents the conclusion.
2 Dictionary Design
The central idea in dictionary learning is to train a dictionary based on set of training
signals until a well-suited specialized final dictionary is obtained that can be used in
fixed form to perform efficient sparse representation of other signal sets of the same
class. The learned dictionary has been shown to be very well suited for sparse
approximation.
2.1 Background
Let y represent a signal set given by:
y ¼ ½y1 ; y2 ; y3 ; . . .; yK 2 RMK ð1Þ
Dictionary learning aims to find a dictionary D ¼ ½d1 ; d2 ; d3 ; . . .; dN 2 RMN with

M < N, such that each yK can be sparsely represented by atoms of D. The sparse
representation of each yK is summarized in the matrix X ¼ ½x1 ; x2 ; x3 ; . . .; xN 2 RNK ,
where xK is sparse representation vector of yK . The minimization problem with sparsity
constraint T0 can be represented as:
minkY DX k2 subject to 8i; kxi k0 T0 ð2Þ

D;X
In dictionary learning, the first step finds the coefficients with constant D. Then, in
second step the atoms of dictionary D are updated by keeping fixed coefficients. These
two steps continue for a certain number of iterations or until a desired approximation
error is reached. The first step of the problem is usually solved with sparse decompo-
sition algorithm such as MP [7] and OMP [8]. The second step of the problem, dic-
tionary update, can be solved by various update techniques. The K-SVD algorithm [9] is
very efficient and often used as an offline preprocessing step to design the dictionary.
3 Proposed Residual Dictionary Design
In block-based intra-compression method [10], the prediction residual is encoded. In

intra-compression method, the prediction residuals are transformed, quantized and
entropy-encoded. It uses fixed DCT dictionary. The fixed DCT dictionary is particu-
larly well suited for representing prediction residuals with periodic texture patterns.
Dictionary Design for Block-Based Intra-image Compression 219
However, DCT dictionary is not suited for efficient encoding of prediction residuals
with complex and non-periodic characteristics. In this section, we will describe the
detailed algorithm to design an over-complete residual dictionary.
The proposed residual dictionary design consists of three main steps: intra-
prediction, generation of the training samples and dictionary learning.
3.1 Intra-prediction
Let I denote the original image block, and IP is its predicted block for best mode N. The
prediction residual image block is represented as:
IR ¼ IIP ð3Þ
After intra-prediction, the best mode and the corresponding prediction residues of
each block are generated for encoding.
3.2 Generation of the Training Samples

In this step, prediction residuals IR generated from a wide variety of images are used as
training samples to train the dictionary. Prediction residuals from each mode are ran-
domly chosen as training set to train the residual dictionary.
3.3 Dictionary Learning
Let IT ¼ ½IR1 ; IR2 ; IR3 ; . . .; IRK 2 Rn K , where IR1 ; IR2 ; IR3 ; . . .; IRK represents K train-
2
ing vectors of training set IT . Residual dictionary DR ¼ ½dR1 ; dR2 ; dR3 ; . . .; dRN 2 Rn N
2
2
with N > n aims to sparsely represent each IRK by linear combination of few atoms of
DR . The minimization problem to find the best dictionary DR with sparsity constraint T0
is represented by:
min k IT DR X k2 subject to 8i; k xi k0 T0 ð4Þ

DR ;X
The overall problem can be represented as a joint optimization:
k
X
DR ¼ arg min kIRi DR xi k22 þ kkxi k0 ð5Þ
DR ;x i¼1
This optimization problem can be solved by splitting the problems into two parts:
1. With fixed DR, compute sparse representation X (sparse coding step).
2. With fixed X, update atoms of D (dictionary update step).
Sparse coding:
With fixed DR , it can be represented in k distinct problem of the form
minkIRi DR xi k22 subject to 8i; kxi k0 T0 ð6Þ

Xi
This problem is adequately addressed by the OMP sparse decomposition algorithm.

Dictionary update step:
K-SVD algorithm is used to update the dictionary atoms. Let dRk be the one column of
the residual dictionary that needs to be updated, then the error matrix without atom dRk
is represented as:
X
Ek ¼ IT dRj xTj ð7Þ
j6¼k
SVD decomposes Ek ¼ UDV T . The updated atom for dRk is the first column of U,
i.e., dRk = U1. This step repeats for all atoms or columns of DR . The above two steps
are repeated until the sparsity constraint is reached, i.e., 8i; k xi k0 T0 . Algorithm 1
summarizes the steps of residual dictionary design.
Algorithm 1: Residual dictionary design

2
K
Input: Training prediction residual IT ¼ ½IR1 ; IR2 ; IR3 ; . . .; IRK 2 Rn
2
N
Output: DR ¼ ½dR1 ; dR2 ; dR3 ; . . .; dRN 2 Rn
1: initialization: randomly initialize DR 2 Rn N with normalized columns
2
2: repeat sparse coding stage and codebook update stage until halting criteria is satisfied
Sparse coding stage:
For i = 1 to k do
3: solve min k IRi DR xi k22 subject to 8i; k xi k0 T0 using OMP sparse decomposition
xi
algorithm
Codebook update stage:
For k = 1 to N do
4: Compute the overall representation error matrix
P
Ek ¼ IT dRj xTj
j6¼k
5: Apply SVD which decomposes Ek ¼ UDV T

6: Update atom dRk by first column of U, i.e., dRk = U1
7: Until halting criterion k IRi DR xi k22 d is true 8i
3.4 The Compression Scheme

The block diagram of the residual dictionary-based compression process is shown in
Fig. 1 [6].
Fig. 1. Detailed block diagram of the compression process
In this scheme, after intra-prediction, the prediction mode number N of each block
is transmitted for encoding and the corresponding prediction residue IR is transmitted
for sparse representation. The coefficients and position of nonzero elements of solution
vector obtained from sparse representation are encoded.
4 Experimental Results
In order to evaluate the performance of proposed dictionary, experiments are conducted

on several test images taken from the Berkeley segmentation database [11, 12]. Test
images with different features are selected for experiment. The experiments concentrate
on compression performance and computational complexity. The implementation is
done in MATLAB.
4.1 Experimental Setup

In the experiment, a set of 45,000 prediction residual blocks, 5000 from each mode, are
randomly chosen as training set to train the residual dictionary. Residual dictionary
D of size 64 512 is selected. Our proposed residual dictionary design algorithm is
employed for dictionary training. Sparsity constraint T0 = 4 and 100 iterations are
selected as stopping condition of the algorithm.
Our proposed residual dictionary D and OMP algorithm are used for sparse rep-
resentation of prediction residual IR. Error d = 20 and T0 = 4 are selected for OMP.
After sparse representation, fixed-length codes are used to encode mode number
and position. Four bits are used to encode mode number. Positions of nonzero elements
are encoded with log2k bits, where k is the number of atoms in dictionary. The coef-
ficient is uniformly quantized and entropy-coded similar to JPEG to form the final bit
stream.
4.2 Compression Performance Comparison

We compared the image compression performance of our dictionary with intra-image
compression scheme using fixed DCT dictionary and JPEG 2000. The objective
evaluation of our proposed method is accomplished by assessment of PSNR [13].
Table 1 shows comparison of PSNR for test image at low bit rate as well as high bit
rate. It is evident from Table 1 that the proposed method gives rise to a PSNR
improvement up to 0.35 dB in comparison with JPEG 2000 and 0.6 dB in comparison
with intra-image compression scheme using DCT.
Table 1. Compression performance comparison of proposed dictionary with intra-image

compression scheme using fixed DCT dictionary and JPEG 2000
Images PSNR (dB) at bit rate 0.2 bpp PSNR (dB) at bit rate 1 bpp
DCT JPEG Proposed DCT JPEG Proposed
2000 dictionary 2000 dictionary
Barbara 26.92 27.30 27.61 36.94 37.16 37.53
Sailboat 28.88 29.13 29.54 36.38 36.82 37.02
Baboon 24.74 25.25 25.46 30.22 30.66 30.82
Couple 28.12 28.49 28.81 36.47 36.80 37.11
Hill 29.44 29.89 30.22 36.25 36.43 36.85
Jet plane 31.56 31.93 32.29 41.61 41.89 42.21
Lena 32.87 33.03 33.47 40.23 40.41 40.83
Lighthouse 28.11 28.40 28.83 38.23 38.43 38.84
Peppers 32.13 32.50 32.89 38.12 38.39 38.73
Average 29.19 29.55 29.90 37.16 37.44 37.77
4.3 Computational Complexity Comparison

In order to evaluate the computational complexity of our dictionary, we compared our
method with the dictionary-based image compression scheme proposed by Sahoo and
Das [6]. Both the methods differ due to sparse representation of image block. Both the
methods are applied over 100 residual image blocks of size 8 8. It is observed that
for error d = 20 around 12 dictionary elements are required using the method proposed
by Sahoo and Das [6] for sparse representation of residual image blocks, whereas using
our method around 4 dictionary elements are required for sparse representation of
residual image blocks. Note that sparse representation using our dictionary is around 3
times faster compared to other one.
5 Conclusion
In this paper, we presented an algorithm to design an over-complete residual dictionary

suitable for block-based intra-image compression. The performance of proposed image
compression method is compared with intra-image compression scheme using fixed
DCT dictionary and state-of-the-art JPEG 2000 compression standard. The computa-
tional complexity is also compared with other methods. Simulation results demonstrate
that the proposed image compression method outperforms other standard methods in
terms of compression efficiency and computational complexity.
References
1. Bryt O, Elad M (2008) Compression of facial images using the K-SVD algorithm. J Vis
Commun Image Representation 19(4):270–282
2. Zepeda J, Guillemot C, Kijak E (2011) Image compression using sparse representations and
the iteration-tuned and aligned dictionary. IEEE J Sel Topics Signal Process 5(5):1061–1073
3. Zhan X, Zhang Y, Huo C (2013) SAR image compression using multiscale dictionary
learning and sparse representation. IEEE Geosci Remote Sens Lett 10(5):1090–1094
4. Shao G, Wu Y, Yong A, Liu X, Guo T (2014) Fingerprint compression based on sparse
representation. IEEE Trans Image Process 23(2):489–501
5. Zhu J-Y, Wang Z-Y, Zhong R (2015) Dictionary based surveillance image compression.
J Vis Commun Image R 31(7):225–230
6. Sahoo A, Das P (2017) Dictionary based image compression via sparse representation. Int J
Electr Comput Eng 7(4):1964–1972
7. Mallat S, Zhang Z (1993) Matching pursuit with time-frequency dictionaries. IEEE Trans
Signal Process 41(12):3397–3415
8. Pati YC, Rezaiifar R, Krishnaprasad PS (1993) Orthogonal matching pursuit: recursive
function approximation with applications to wavelet decomposition. In: Proceedings of the
twenty-seventh Asilomar conference on signals, systems and computers, pp 40–44
9. Aharon M, Elad M, Bruckstein A (2006) K-SVD: an algorithm for designing overcomplete
dictionaries for sparse representation. IEEE Trans Signal Process 54(11):4311–4322
10. Turkan M (2012) Novel texture synthesis methods and their application to image prediction
and image inpainting, Ph.D. Thesis
11. (Online). Available:www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/segbench/
12. (Online). Available: http://sipi.usc.edu/services/database/
13. Wang Z, Bovik AC, Sheikh HR, Simoncelli EP (2004) Image quality assessment: from error
visibility to structural similarity. IEEE Trans Image Process 13(4):600–612
Performance Analysis of Fractional Order
Low-pass Filter
Kumar Biswal1(&), Madhab Chandra Tripathy2, and Sanjeeb Kar3

1
School of Electronics Engineering, Kalinga Institute of Industrial Technology
(DU), Bhubaneswar, India
kumar.biswalfet@kiit.ac.in
2
Department of Instrumentation & Electronics Engineering, CET, BPUT,
Bhubaneswar, India
mctripathy@cet.edu.in
3
ITER, SoA, Deemed to be University, Bhubaneswar, Odisha, India
sanjeebkar@soa.ac.in
1 Introduction
Fractional order differential and integral equations are the generalization of conven-
tional integral and differential equations that can be used for modeling the real world
we live in. The integer-order models were used due to the absence of solution methods
for fractional differential equations. At present, many methods have been developed for
realizing the applications based on fractional derivative and integration [1–3]. Frac-
tional calculus covers all non-integer numbers including fractions and irrational
numbers. The integer-order filters have integer coefficients in its mathematical repre-
sentation which leads to some design criterion that makes it difficult to obtain sharp
cutoff frequency [4–7]. The exact design requirements can be satisfied by fractional
order filters, but in conventional integer-order filters this was not possible. This frac-
tional concept adds a little bit of safety in the design specifications in comparison with
the integer-order filters [8]. In fractional order filters, the circuit design specifications
with tuning parameters can be easily achieved. For example, the roll-off frequency can
be varied and set to any desired slope which is only possible in case of fractional order
filters. Hence, fractional order filters provide better design flexibility as compared to
integer-order filters [9–14]. This paper is organized as follows. Section 2 deals with
model, transfer function and response of fractional order low-pass filter. Section 3
illustrates the time response of single-stage fractional low-pass filter. Similarly, Sect. 4
deals with frequency response of single-stage fractional order low-pass filter. Finally,
conclusion is presented in Sect. 5.

https://doi.org/10.1007/978-981-15-2774-6_28
Performance Analysis of Fractional Order Low-pass Filter 225
2 Fractional Order Low-Pass RC Filter
Low-pass filters allow low-frequency signals to pass through and attenuate high-frequency
signals through it. The cutoff frequency (xc) locates the passband (x < xc) and the stop-
band (x > xc) in a complete frequency response. A low-pass filter allows signal of low
frequency to pass and gradually attenuates high signals having frequencies higher than the
cutoff frequency.
Low-pass filters have a wide range of applications in analog and digital filters. The
filter is characterized by its cutoff frequency and rate of roll-off frequency. So, the value
of attenuation frequencies which are generally higher than cutoff frequency is deter-
mined by the order of filter [13, 14]. The circuit model for the single-stage fractional
order low-pass filter is presented in Fig. 1.
R cα
Vin VO
Fig. 1. Circuit model of a single-stage fractional order low-pass filter
The characteristic equation for the circuit is given by an expression:
VO d/ VO ðtÞ Vin
þ C/ ¼
R dt/ R
Taking Laplace transform, the transfer function of the circuit is
V O ð SÞ s
H ðsÞ ¼ ¼ a ;
Vi ðSÞ s þ s
Here, s ¼ 1 a is a constant (i.e., time constant).

RC
H(s) represents the transfer function expression of the (integer) first-order low-pass
filter. From the above transfer function, the magnitude H(w) and phase ;ðwÞ of the
(fractional) first-order low-pass filter [15–18] are given by
s
jH ðwÞj ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi ð1Þ
w þ s þ 2wa cos ap
2a 2
2
" #
1 wa sin ap
2
;ðwÞ ¼ tan ð2Þ
s þ wa sin ap
2
226 K. Biswal et al.
The step response is expressed as
vo ðtÞ ¼ 1 Ea;1 ðsta Þ ð3Þ
The impulse response is expressed
hðtÞ ¼ sa1 Ea;a ðsta Þ ð4Þ
1
jH ðwÞj ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð5Þ
1 þ ðwRCÞ2
;ðwÞ ¼ tan1 ðwRCÞ ð6Þ
Equations (5) and (6) represent the transfer function expressions for magnitude and
phase of (integer) first-order low-pass filter.
3 Time Response of a Single-Stage Low-Pass Filter
In fractional order filter, the time domain response defines the different measuring
parameters such as the transfer function, magnitude, phase, stability, peak overshoot
(Mp), settling time (ts) and rise time (tr) [6, 14]. The step response and impulse response
of single-stage fractional low-pass filter are plotted for the above parameters with
different orders (a) of a fractional capacitor.
Settling time: It is the time required for the response of a system to be within the
fraction of the steady-state value and to stay there.
Rise time: It is the time taken for the response of a system to go from 10% to 90%
of the steady-state value at its first instant.
Peak time: It is the time taken for the response of a system to reach the peak or
maximum value of overshoot.
The time at which the system’s output attains its maximum value is expressed as
[19],
0:131ð/ 0:255Þ2
tp ¼ ð7Þ
ð/ 0:921Þwc
The unit step response of single-stage fractional order low-pass filter is shown in
Fig. 2, and its response is plotted against order a. This figure can be used for comparing
the responses of fractional order and integer-order filters. A standardized value of time
constant, i.e., RC = 1, is assumed for all simulations.
An impulse is a short duration signal that starts from zero to a maximum value and
returns to zero again in a short time. The filter’s impulse response is obtained by
applying an impulse input and studied with order a of filter as shown in Fig. 3.
Step Response
1.4
G1
G2
1.2 G3
G4
1
Amplitude
0.8
0.6
0.4
0.2
0
0 0.5 1 1.5 2 2.5 3
Time (sec)
Fig. 2. Unit step response of single-stage low-pass filter where order a G1 = 0.1, b G2 = 0.4,
c G3 = 0.6, d G4 = 0.8
Impulse Response
60
G1
G2
50 G3
G4
40
Amplitude
30
20
10
-10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (sec)
Fig. 3. Impulse response of a single-stage low-pass filter where order a G1 = 0.1, b G2 = 0.4,
c G3 = 0.6, d G4 = 0.8
Furthermore when a varies from 0 to 1, the circuit’s behavior approaches toward an

equivalent second-order system as observed in Fig. 3. The different measuring speci-
fications such as rise time, settling time, peak overshoot and peak time are given in
Table 1 by changing order a of fractional capacitor.
Table 1. Time domain response specification of a fractional low-pass filter

a Rise time Settling time Peak overshoot Peak time
0.1 0.0331 0.0566 0.475 0.141
0.2 0.0483 0.0810 1.377 0.177
0.3 0.0709 0.4751 2.929 0.238
0.4 0.1021 0.7999 5.381 0.010
0.5 0.1361 1.0019 9.101 0.367
0.6 0.1780 1.1159 14.210 0.468
0.7 0.2257 1.2067 20.771 0.594
0.8 0.2781 2.1160 28.789 0.728
0.9 0.3351 3.4006 38.277 0.486
1 0.3968 5.1004 48.353 1.063
From Figs. 4, 5 and 6, the circuit’s response like percent overshoot, rise time and
settling time is studied for different values of a. So by using the results of figures shown
below, an approximate fractional capacitor can be chosen to meet the desired param-
eters given in Table 1.
Over shoot
50
45
40
35
30
%po
25
20
15
10
5
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
alpha
Fig. 4. Percent overshoot versus alpha (a) for single-stage fractional low-pass filter
Figure 4 shows the percentage overshoot versus order of the fractional capacitor a
used in the low-pass filter. From Fig. 4, it is observed that overshoot decreases when
order of the fractional capacitor a decreases. It is maximum in classical capacitor.
Figure 5 indicates the variation of settling time by varying the order of the frac-
tional capacitor. The settling time is maximum at a = 1, i.e., when normal capacitor is
used. This shows the circuit output stabilizes quickly when fractional capacitor is used.
Settling time
6
4
Ts
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
alpha
Fig. 5. Settling time versus alpha (a) for single-stage fractional low-pass filter
rise time
8
5
rise time
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
alpha
Fig. 6. Rise time versus alpha for single-stage fractional low-pass filter
Figure 6 shows rise time versus order of the fractional capacitor a. It is clear from
the figure that rise time decreases when fractional capacitor replaces the classical
capacitor. This indicates that fractional circuits have less delay and high speed, and thus
the speed of response has low delay time.
Bode Diagram
20
Magnitude (dB)
-20
-40
-60
0 G1
G2
Phase (deg)
G3
-45 G4
-90
-135
-3 -2 -1 0 1 2 3 4
10 10 10 10 10 10 10 10
Frequency (rad/sec)
Fig. 7. Bode plot of single-stage fractional low-pass filter where order a G1 = 0.1, b G2 = 0.4,
c G3 = 0.6, d G4 = 0.8
4 Frequency Response of a Single-Stage Low-Pass Filter
The Bode plots for magnitude and phase of a fractional order filter with order a are
shown in Fig. 7; it is observed that the slope of magnitude plot in the stopband region
varies proportionally to a. As observed from Fig. 7, the asymptotic approximation
becomes inaccurate for very small values of a, whereas the asymptotic part in the high-
frequency region has greater slope which increases with increase of a. Hence, the cutoff
frequency becomes sharper, which leads to an increase in speed of response obtained in
frequency domain response as shown in Fig. 7.
The Bode plot for phase is also shown in Fig. 7 which indicates for small value of
a, the phase variation is almost linear over a narrow frequency range. From the figure, it
is also observed that as a increases, the phase plot moves toward saturation state having
asymptotic values of 0° and 90° for low and high frequencies, respectively.
5 Conclusion
In this article, the time domain and frequency domain responses of a RC low-pass
fractional order filter are discussed. First, the fundamentals of magnitude and phase
characteristics have been worked out as fundamental for the fractional order filter.
Secondly, from the filter’s response the amplitude–frequency characteristics, phase–
frequency characteristics and cutoff frequency are analyzed in detail with respect to
time constant s and order a that shows better flexibility of the fractional order filter.
The filter behavior is studied and represented in figures by changing the order a.
Finally, it is concluded that the speed of response increases with decrease in order a of
a filter. It is also observed that the magnitude and phase change almost linearly with the
fractional order in the region of stopband.
References
1. Debnath L (2003) Recent applications of fractional calculus to science and engineering. Int J
Math Math Sci 54:3413–3442
2. Ortigueira MD (2008) An introduction to the fractional continuous-time linear systems, the
21st century systems. IEEE Circ Syst Mag 8(3):19–26
3. Elwakil AS (2010) Fractional-order circuits and systems: an emerging interdisciplinary
research area. IEEE Circ Syst Mag 10(4):40–50
4. Machado JT (2013) Fractional generalization of memristor and higher order elements.
Commun Nonlinear Sci Numer Simul 18:264–275
5. Moreles MA, Lainez R (2017) Mathematical modelling of fractional order circuit elements
and bio-impedance applications. Commun Nonlinear Sci Numer Simul 46:81–88
6. Zahra WK, Hikal MM, Bahnasy TA (2017) Solutions of fractional order electrical circuits via
Laplace transform and nonstandard finite difference method. J Egypt Math Soc 25:252–261
7. Sierociuk D, Podlubny I, Petráš I (2013) Experimental evidence of variable-order behavior
of ladders and nested ladders. IEEE Trans Control Syst Technol 21:459–466
8. Walczak J, Jakubowska A (2015) Analysis of resonance phenomena in series RLC circuit
with super capacitor. In: Analysis and simulation of electrical and computer systems, Lecture
Notes in Electrical Engineering, vol 324, Springer, Heidelberg, pp 27–34
9. Rousan AA, Ayoub NY, Alzoubi FY, Khateeb H, Al-Qadi M, Hasan (Qaseer) MK,
Albiss BA (2006) A fractional LC-RC circuit, Fractional calculus and applied analysis. Int J
Theor Appl 9(1):33–41
10. Obeidad A, Gharaibeh M, Al-Ali M, Rousan A (2011) Evaluation of a current in resistor,
Fractional calculus and applied analysis. Int. J Theor Appl 14(2):247–250
11. Gómez-Aguilar JF, Escobar-Jiménez RF, Olivares-Peregrino VH, Taneco-Hernández MA,
Guerrero-Ramírez GV (2017) Electrical circuits RC and RL involving fractional operators
with bi-order. Adv Mech Eng 9:1–10
12. El-Sayed MA, Nour HM, Raslan WE, El-Shazly ES (2012) Fractional parallel RLC circuit.
Alexandria J Math 3:11–23
13. Radwan AG, Salama KN (2012) Fractional-order RC and RL circuits. Circ Syst Signal
Process 31(6):1901–1915
14. Jakubowska A, Walczak J (2016) Analysis of the transient state in a series circuit of the class
R-Lb-Ca. Circ Syst Signal Process 35:1831–1853
15. Hu Y, Luo Y, Lu Z (2008) Analytical solution of the linear fractional differential equation by
Adomian decomposition method. J Comput Appl Math 215:220–229
16. Gorenflo RR, Kilbas AA, Mainardi F, Rogosin SV (2014) Mittag-Leffler functions, related
topics and applications. Theory and applications. Springer, Heidelberg, New York,
Dordrecht, London
17. Garrappa R (2015) Numerical evaluation of two and three parameter Mittag-Leffler
functions. SIAM J Numer Anal 53(3):1350–1369
18. Garrappa R (2014) The Mittag–Leffler function, MATLAB Central File Exchange. File ID:
48154
19. Podlubny I (1999) Fractional differential equations. Academic Press, London
Chronic Disease Risk (CDR) Prediction
in Biomedical Data Using Machine
Learning Approach
Lambodar Jena1(&), Soumen Nayak1,2, and Ramakrushna Swain3

1
Department of Computer Science and Engineering, Siksha O Anusandhan
(Deemed to Be University), Bhubaneswar, India
lambodarjena@soa.ac.in
2
Department of Computer Science and Engineering, Indian Institute of
Technology (ISM), Dhanbad, India
3
Department of Computer Science and Engineering, Silicon Institute of
Technology, Bhubaneswar, India
1 Introduction
Day by day, large amount of data are collected due to the rapid growth in technology
and science. Particularly in medical science, huge amount of medical databases are
created to assist the healthcare system [1]. Now, a major field of research is man-
agement of large scale of data and efficiently extraction of knowledge from large
databases [2]. In the recent era, machine learning has become an emerging field to be
used on bioinformatics and complex studies of medical science. Large volumes of data
are yet to be explored and mined to gather knowledge. The aim of this work is to
process the chronic kidney disease dataset for prediction of chronic kidney disease of a
large population. CKD is a major health issue in many societies because it gradually
decreases the function of kidney. Based on the symptoms identified in patients, it
predicts the sick ones [5]. In this work, genetic search algorithm is used for feature
selection and three machine learning classifiers are used for classification. The clas-
sifiers are namely naive Bayes, J48, and decision table. For better prediction and
diagnosis of chronic kidney disease, first the classifiers are used and compared with
each other. Then, subsequently the features are reduced from the dataset by using
genetic search algorithm and again classification results of each classifier are compared
with each other.
2 Related Work
Many authors have implemented different machine learning techniques for classifica-
tion and future prediction.
Dhamodharan [1] predicted risk in liver disease by using a number of machine
learning classifiers, and he found that Naïve Bayes becomes suitable algorithm with
high prediction accuracy. Solanki [2] has used J48 and random tree algorithms for the

https://doi.org/10.1007/978-981-15-2774-6_29
Chronic Disease Risk (CDR) Prediction in Biomedical Data … 233
classification of sickle cell disease and found that random tree performs better. Breast
cancer diagnosis and prognosis have been done by Joshi et al. [3] using some rules of
classification. They found that logistic model tree (LMT) algorithm identify more
correctly (76%) the healthy and sick patients. However, the disease leukemia risk has
been predicted by David et al. [4]. They have implemented a number of classifiers like
KNN, J48, random tree, and Bayesian n/w and found that Bayesian performs well.
Heart disease risk prediction has been done by Vijayarani and Sudha [5] by using
different types of classifiers. Kumar [6] has used alternating decision trees for dengue
fever diagnosis. In healthcare sector, Durairaj and Ranjani [7] have used and compared
the efficiency of various types of data mining algorithms. They have worked on the
prediction and diagnose of life killing diseases. Sugandhi et al. [8] used weka to
analyze cataract patient’s database. Yasodha and Kannan [9] analyzed diabetic patient
database using different methods namely network, Bayes tree, J48, and random tree.
Different classification techniques such as radial basis function, Bayes network, deci-
sion tree, and pruning algorithms are compared by Bin and Yau [10] for breast cancer.
Jena et al. [11] worked on chronic kidney disease dataset by comparing different
classification techniques such as SVM, naive, conjunctive rule, MLP, J48, and decision
table. Also, Jena et al. [12] further worked on the prediction of kidney disease using
different algorithms and feature selection techniques. Jena et al. [13] have also worked
on the prediction of human depression using apriori algorithm.
3 Methodology
3.1 Preliminary
Classification comes under supervised learning technique. Large scales of population of
records are classified by a classification model. There is a mapping of the data into
predefined groups by the model. Based on the attribute value, new classes will be
defined by classification model. A measure in terms of classification accuracy is
considered. To carry out the experiments, we have obtained the biomedical dataset on
chronic kidney disease with 400 instances and 24 attributes.
3.2 Classifiers Used

Classifiers are compared with each other to note the similarity and dissimilarity present
in each parameter. In machine learning, classifiers are divided into various types:
• Bayes
• Functions
• Lazy
• Rules
• Tree-based classifiers, etc.
A balance mixture of classifiers, namely naive Bayes, J48, and decision table, from
different groups has been chosen.
234 L. Jena et al.
3.3 Performance Indicators

The five performance indicators namely accuracy, error rates such as mean absolute
error (MAE) and root mean square error (RMSE), kappa statistics and ROC value are
considered to measure the performance of all the classifiers.
The experiments have been conducted on the dataset using all mentioned classifiers.
The result before applying feature selection procedure and the result after applying
feature selection are recorded in Table 1. The result of feature selection based on
chronic kidney disease dataset using genetic search algorithm is shown below. The
attributes selected by genetic search after 100 iterations are 2, 3, 4, 6, 10, 12, 13, 14, 15,
16, 17, 19, 20, 22, 23, 24 = 16. That is, the selected attributes are {bp, sg, al, rbc, bgr,
sc, sod, pot, hemo, pcv, wbcc, htn, dm, appet, pe, ane} from the dataset. The attributes
which are reported by genetic search as irrelevant are neglected here. Taking relevant
attributes selected by genetic search into consideration is the objective of our further
study. That is, after attribute selection we have also applied all the three classification
algorithms to see their behavior for various parameters shown in Table 1.
Table 1. Result of accuracy, kappa statistics, MAE, RMSE, and ROC

Experiment 1: Classification result before feature reduction
Algorithm Accuracy (%) Kappa statistics MAE RMSE ROC
Naïve Bayes 95.5 0.8962 0.0478 0.2047 1
J48 99 0.9786 0.0225 0.0807 0.999
Decision table 99 0.9786 0.1815 0.2507 0.992
Experiment 2: Classification result after feature reduction using genetic search
Algorithm Accuracy (%) Kappa statistics MAE RMSE ROC
Naïve Bayes 97 0.937 0.0334 0.1608 1
J48 99 0.9786 0.0225 0.0805 0.999
Decision table 98.75 0.9732 0.163 0.2331 0.99
5 Analysis of the Performance Indicators (Parameters)

5.1 Classification Accuracy
The accuracy tells that the estimated result is how much away from the actual value.
TP þ TN
Accuracy ¼
PþN
where
TP = True positive
TN = True negative
P = TP + FN (total positive)
N = FP + TN (total negative).
Any classifier with high accuracy will be more suitably chosen for prediction task.
Accuracy result of the algorithm varies from dataset to dataset. The accuracy of an
algorithm changes if we change the dataset. From experiment 1 in Table 1, it is found
that J48 and decision table algorithm give the same accuracy result, i.e., 99%, whereas
naïve Bayes classifier gives classification accuracy of 95%. All algorithms produce
more than 90% of classification accuracy. Hence, it is concluded that J48 and decision
table performs well for CKD dataset by taking all features of the dataset into consid-
eration. However after feature selection using genetic search algorithm, it is found that
J48 gives high accuracy result (= 99%) comparing to all other algorithms. Hence, it is
concluded that J48 performs well in both the cases for the biomedical dataset to
determine chronic kidney disease, which is shown in Fig. 1.
99%
98%
97%
Classifica on Accuracy with
96% all features
Classifica on Accuracy a er
95% feature reduc on
94%
93%
Naïve Bayes J48 Decision
Table
Fig. 1. Classification accuracy analysis
But from Table 1, it is observed that the accuracy of naive Bayes classifier algo-
rithm is increased by 2%, whereas the accuracy of decision table is getting reduced
mildly. Other algorithms, i.e., J48 have the same classification accuracy in both the
experiments.
5.2 Kappa Statistics

In both the experiments, classification with J48 has better kappa statistics value with
0.9786 than other classifiers shown in Fig. 2. From Table 1, it is interesting to note that
the values of kappa statistics, MAE, and RMSE for the classifiers are also changed. J48
and decision table has same kappa statistics value in comparison to naïve Bayes in the
first experiment.
236 L. Jena et al.
0.98
0.96
0.94
0.92 Kappa stascs with all

features
0.9 Kappa stascs aer feature
reducon
0.88
0.86
0.84
Table
Fig. 2. Kappa statistics analysis
5.3 Error Rate in Terms of Mean Absolute Error (MAE)

The classifier J48 achieves minimum MAE value = 0.0225 that remains the same in
both the experiments shown in Fig. 3. Thus in case of J48 classifier, the calculated
result is closer to the true value of the chronic kidney disease criteria. But it is also
observed from Table 1 that the MAE value varies slightly from experiment 1 to
experiment 2 for naïve Bayes and decision tree algorithms.
0.2
0.18
0.16
0.14
0.12
0.1 MAE with all features
0.08 MAE after feature reduction

0.06
0.04
0.02
0
Table
Fig. 3. Mean absolute error (MAE) analysis
5.4 Error Rate in Terms of Root Mean Square Error (RMSE)

From Table 1, it is found that J48 classifier results in minimum error rate (i.e., 0.0807
with all features and 0.0805 after feature reduction) compared to other algorithms,
naïve Bayes and decision tree shown in Fig. 4. Here, also it is observed that in Table 1,
the error rate (RMSE) value gets reduced in case of both naïve Bayes and decision table
as the number of features is reduced.
0.3
0.25
0.2
0.15 RMSE with all features

RMSE after feature reduction
0.1
0.05
0
Table
Fig. 4. RMSE analysis
5.5 ROC Value

From Fig. 5 and Table 1, it is seen that the ROC value of algorithms is approximately
the same in both experiments except decision table algorithm. In decision table algo-
rithm, the ROC value is slightly reduced as the numbers of features are reduced.
0.998
0.996
0.994
0.992 ROC using all features

ROC after feature reduction
0.99
0.988
0.986
0.984
Table
Fig. 5. Value of ROC

238 L. Jena et al.
From Figs. 1, 2, 3, 4, and 5, it is observed that in all the cases, J48 classifier’s
performance is quite better than all other classifiers being used for our experiment.
6 Conclusion
The major focus of this work is to analyze the complex usage of the classifiers to
predict the risk in chronic disease accurately. Here, three classifiers have been used to
determine the target class accurately. They have been used in two typical scenarios; one
with all attributes of the dataset and second after reducing the number of attributes. In
the experiments, we observe that J48 algorithm performs better with high accuracy of
99% and hence gives better performance in prediction. But the prediction accuracy
level of naïve Bayes and decision table is fluctuating with the increase of 2% in naïve
Bayes and decrease of 0.25% in case of the later one.
References
1. Dhamodharan S (2014) Liver disease prediction using Bayesian classification, Special
Issues. In: 4th National conference on advance computing, application technologies, May
2014
2. Solanki AV (2014) Data mining techniques using WEKA classification for sickle cell
disease. Int J Comput Sci Inf Technol 5(4):5857–5860
3. Joshi J, Rinal D, Patel J (2014) Diagnosis and prognosis of breast cancer using classification
rules. Int J Eng Res General Sci 2(6):315–323
4. David SK, Saeb AT, Al Rubeaan K (2013) Comparative analysis of data mining tools and
classification techniques using WEKA in medical bioinformatics. Comput Eng Intell Syst 4
(13):28–38
5. Vijayarani S, Sudha S (2013) Comparative analysis of classification function techniques for
heart disease prediction. Int J Innov Res Comput Commun Eng 1(3):735–741
6. Kumar MN (2013) Alternating decision trees for early diagnosis of dengue fever. arXiv
preprint arXiv:1305.7331
7. Durairaj M, Ranjani V (2013) Data mining applications in healthcare sector a study. Int J Sci
Technol Res IJSTR 2(10)
8. Sugandhi C, Ysodha P, Kannan M (2011) Analysis of a population of cataract patient
database in WEKA tool. Int J Sci Eng Res 2(10), October 2011
9. Yasodha P, Kannan M (2011) Analysis of population of diabetic patient database in WEKA
tool. Int J Sci Eng Res 2(5), May 2011
10. Bin Othman MF, Yau TMS (2007) Comparison of different classification techniques using
WEKA for breast cancer. In: 3rd Kuala Lumpur international conference on biomedical
engineering 2006. Springer, Heidelberg, pp 520–523, Jan 2007
11. Jena L, Kamila NK (2015) Distributed data mining classification algorithms for prediction of
chronic- kidney-disease. Int J Emerg Res Manage Technol 4(11):110–118. ISSN 2278-9359
12. Jena L, Swain R (2017) Chronic disease risk prediction using distributed machine learning
classifiers. In: Proceedings of IEEE International Conference on Information Technology
(ICIT)-2017. doi:https://doi.org/10.1109/icit.2017.46, 978-1-5386-2924-6/17 © 2017 IEEE
13. Jena L, Kamila NK (2014) A model for prediction of human depression using apriori
algorithm. In: IEEE Xplore Digital Library proceedings, pp 240–244. ISBN 978-1-4799-
4152-0
A Novel Approach to Detection
of and Protection from Sybil Attack in VANET
Binod Kumar Pattanayak1(&), Omkar Pattnaik2, and Sasmita Pani2

1
Department of Computer Science and Engineering, Institute of Technical
Education and Research, Siksha ‘O’ Anusandhan Deemed to be University,
binodpattanayak@soa.ac.in
2
Department of Computer Science and Engineering, Government College of
Engineering, Keonjhar, Odisha 758002, India
1 Introduction
The research community nowadays focus on the concept of mobile ad hoc network
(MANET) which is implemented in road and transport communication where vehicles
act as nodes, and it forms a vehicular ad hoc network (VANET). It is also regarded as a
subset of MANET. Vehicular ad hoc network (VANET) was created in October 2002
by the Federal Communications Commission (FCC). The main objective of this con-
cept is to provide safety on roads and secure driving. To create a VANET environment,
some parameters are necessarily required like onboard unit (OBU) and a roadside unit
(RSU) that must be installed along the roads. Dedicated short-range communications
(DSRC) protocol is used to communicate between OBU and RSU. Again this protocol
is modified later, and a new protocol is formed called as wireless access in vehicular
environment (WAVE). It is based on IEEE 802.11p group. The architecture of VANET
supports vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) interchanges
associated with RSU and OBU.
Security is necessarily required for VANET due to its wireless communications.
So, there are different types of attacks possible in VANET. Sybil attack is one of the
harmful attacks among them. In this attack, multiple identities are created by the
attacker, thereby trying to forge or steal new identities from the neighboring vehicles.
The attacker performs various malicious operations with the help of Sybil attack. An
illusion is created by the attacker by sending messages to different vehicles from
different identities. So, there occurs a jamming in the network which enforces the driver
of vehicles to divert from the original route.
Generally, it affects two major areas like routing and voting and reputation systems
as well. It may appear as a malicious node at more than one locations at a time in
geographical routing protocols [1]. Also, it may disrupt the cluster head and disrupt the
valuable information in cluster-based routing protocols [2]. But in the second case, a
voting process is performed and the vehicle’s exact location is monitored and the
misbehaving node is identified. When a malicious vehicle performs Sybil attack, all
messages come from a particular location that circulates the false information in the
entire network. So, the network may be unable to communicate with neighbor nodes.

https://doi.org/10.1007/978-981-15-2774-6_30
A Novel Approach to Detection of and Protection from Sybil … 241
The rest of the paper is organized as follows. Section 2 describes the attack model
along with the assumptions. Section 3 covers the way Sybil attack works. Section 4
includes the methods of protection from Sybil attack in VANET. Section 5 concludes
the paper.
2 Attack Model and Assumptions
In this part, the Sybil attack model is defined, thereby presenting some assumptions on
the system and the attacker’s probable actions.
Attack Model:
In wireless networks, the new neighboring nodes are discovered by the mobile nodes to
whom the beacon packets are periodically broadcast in the network. In VANET, the
vehicle is to be considered as a node. Sybil attack is based on the idea of multiple
identities being produced by a malicious node [3]. Also, the identities and positions are
determined by the beacon packets. A malicious node has the potential to create multiple
identities without the knowledge of legitimate nodes. A malicious driver may have
additional identity information because either he has borrowed identity from another
driver or might be stealing identity from neighbors. The main objective of Sybil attack
detection is that only one identity is assigned to each physical node [4].
In this paper, we assume a malicious node to be a physical node which produces
multiple identities, and correspondingly, the Sybil nodes are the fabricated identities
produced by the malicious node.
Assumptions:
The assumptions are focused here on the basis of Sybil attack detection and attacker’s
actions.
Attack Detection:
In this paper, we have assumed the following assumptions about detection of Sybil
attack in VANET. (i) The Sybil attack is launched by the individual drivers (vehicles),
which is the basic threat to the system. We also assume that the Sybil attack may be
created by multiple malicious vehicles. (ii) The RSUs are deployed separately along the
road, and to maintain the authentication of infrastructure, an electronic license plate
(ELP) is issued to whole network [5]. (iii) We focus on the same radio module called
received signal strength indicator (RSSI) such as DSRC which is based on radio-
frequency (RF) communication strategy [6]. (iv) Finally, we assume that all the
vehicles must be equipped with GPS devices along with digital maps since the accurate
position is determined by GPS technology. But in real-world scenario, we cannot rely
on roadside base stations present in RSU for the whole network.
Attackers’ Actions:
The following are the possible actions conducted by the attacker.
242 B. K. Pattanayak et al.
(i) Compromised RSBs: The attacker may be successful in compromising the RSBs
(part of RSU) which are a semi-trusted parties. The detection of compromised
RSBs can be canceled by the DMV (Department of Motor Vehicle). But, there is
still a chance for the attacker to retrieve the information stored in the RSBs. So,
the attacker uses its technique to recover the amount of information quickly from
the RSBs.
(ii) Announce false messages: It is also regarded as false data injection. The false
message can be signed by a vehicle before it is broadcast in the network. The
message can also be signed by a pseudonym that is certified by CA. So, it is
impossible to detect the message. When the number of vehicles is more than the
number of attackers in a network, the voting scheme is used. But, if the false
identities are generated sufficiently by the attacker, then the Sybil attack occurs.
(iii) Sybil attack: It refers to a situation wherein the attacker may pretend to be
multiple vehicles by virtue of using multiple pseudonyms. When multiple
pseudonyms are used to sign the message by a vehicle, the Sybil attack is formed
in VANET. The privacy of a vehicle is compromised when the vehicle is
identified from a group of pseudonyms it uses. Therefore, RSBs and vehicles
tend to sign multiple messages in attacker’s pseudonyms [7].
3 How Sybil Attack Works
The concept of Sybil attack was first introduced by Douseur [8]. This is a threat to
functionality of VANET. The attacker uses multiple identities in order for sending
messages to other nodes that belong to the network. The identities of other nodes may
be spoofed by a node called as malicious node that is also regarded as the Sybil
attacker. Sybil nodes are those nodes whose identities are spoofed by the Sybil attacker
[9]. An illusion is created in the network, wherein message is sent to the vehicles that
an accident or a traffic jam has already taken place, and thus, other vehicles are forced
to change their routes and tend to move along the attacker’s route. Also, the false
information may be injected by the Sybil attacker into the network. The attacker also
tries to disrupt the warning message sent by the actual vehicle in the network and fulfill
its target. So, this may put the life of passengers in danger [10].
The different types of possible Sybil attacks are given in Fig. 1.
Fig. 1. Different forms of Sybil attack

In Fig. 1, it is clear that Sybil attack classification is carried out on the basis of the
type of communication, identity, and their participation in the network [11]. These
features are explained clearly in the following paragraphs.
(i) Communication category: When a Sybil node receives a radio message from
an honest node, the message may be listened by a malicious node present in the
network. Along the same path, the malicious devices actually send messages but
not the Sybil nodes. In direct communication, the legitimate nodes communicate
with all the Sybil nodes which are created by malicious devices. But in indirect
communication, the malicious node helps the legitimate nodes to enter the area
of Sybil nodes.
(ii) Identity category: In this category, a new Sybil identity is created by an attacker
and that is generated randomly with a 32-bit integer which is a fabricated
identity or the identities of neighboring nodes are stolen by the attacker.
(iii) Participation category: The malicious nodes create multiple Sybil identities
which can participate simultaneously in the attack. In this category at a time, one
identity is used for each node, and thus, an observation takes place in the
network that many times a particular node may join or leave the network.
4 Protection Methods for Sybil Attack in VANET
In VANET, the protection mechanism can be classified into three different types that
are: (i) public key cryptography method, (ii) methods based on position verification,
and (iii) resource testing method.
Public Key Cryptography Method:
Public key cryptography is a standard method to detect Sybil attack, and also, it
provides authentication mechanism [12]. The concept of a symmetric cryptography is
used to develop the security solution which is based on the combination of digital
certificates with signatures.
In this technique, a CA (central authority) is present that is responsible for issuing
the certificates, thereby maintaining the hierarchy. A secure channel is used to com-
municate between CAs and monitor the issued certificates which are used by every
signed message. In VANET, the signed information sent by vehicles may be verified by
the receiver as demonstrated in Fig. 2. In this technique, it is verified if a message is
valid or invalid. The messages which are attached with certificates are to be considered
as valid ones, and the messages having no certificates are ignored. So, the Sybil attacks
can be prevented. At a time only one certificate is assigned to each node, and to
maintain better privacy, it is necessary to change the certificates from time to time. But
in VANET, due to rapid changes in mobility, it is difficult to implement PKI. But to
maintain privacy, the key pair (public/private) must be authenticated by the CA. But, it
may not be useful to check the identity of a vehicle [13].
Fig. 2. Hierarchy of central authority
Position Verification Method:

In vehicular environment, the concept of sensors, GPS has also important role in
creating a node (vehicle). It may also be possible to tamper the sensors on the part of an
attacker and send false information to other vehicles present in the network. To achieve
effective moving of vehicles in the network accurate positions of vehicle are essentially
required. So it is an important task for all the nodes to maintain accurate position in the
network. This is very much important for geographic routing. In this routing scheme,
accurate positions of all nodes are essential because intermediate nodes are responsible
for transmitting messages to the destination based on their locations.
The method of verification of a secure position was first addressed by the authors in
[14], called as verifiable multilateration approach. To fix the position of the vehicle, a
trustworthy network is built by the base stations which measures the sending and
receiving information sequentially in time. In this approach, a node neither delays the
reply time nor advances the time. So, the nearer distance is fixed for communication.
The second proposal is given by authors in [15] where the concept of threshold
values is used to determine the correct position of the node by receiving information
from other nodes. In this technique, nodes rely on GPS rather than base station scheme
to determine their positions. Beacon messages are used for sending and receiving
location information of nodes and also for verifying it through specified checks.
There are 4 types of check strategies based on the above concept, which are given
below:
• Acceptance range threshold;
• Mobility grade threshold;
• Maximum density threshold;
• Map-based technique.
In Acceptance Range Threshold, the verified node obtains a maximum threshold
value which determines their communication ranges in the network. So, it helps to
detect wrong positions of nodes.
In Mobility Grade Threshold, the last node position is known, and thus, if any
false information about changing of position is detected, then an extreme speed is
defined for a hub. In order to communicate among the nodes, signal messages are used.
The idea about Maximum Density Threshold is to accommodate a specific region
where maximum number of vehicles can be physically present.
In map-based technique, the correct location of the vehicle is to be determined.
In order to detect Sybil attack based on the position, the parameters like collision
avoidance, resource availability, emergency alert, and traffic conditions can be useful.
A VANET may be affected by inserted bogus packets, dropping packets, replying
packets, and modifying existing packets which are created by the malicious attacker
[15, 16].
There are two efficient schemes for localization like spectrum-based and spectrum-
free techniques. In spectrum-based technique, in order to locate vehicle’s position, the
estimated distance between a receiver and transmitter must be calculated. Some of the
techniques are: time of arrival (TOA), angle of arrival (AOA), and received signal
strength indicator (RSSI). But distance-free localization technique helps in gathering
other position information. In order to achieve more accuracy in localization, distance-
based method must be adopted [17].
According to Yan et al., a new method has been developed for Sybil attack in
which vehicles can listen GPS coordinates of other vehicles and also visualize the
neighbor’s vehicle with the help of radar transmission range. It is also called as a virtual
vehicle eye. Therefore, it is easy to track the actual location of neighboring vehicle and
malicious vehicles are separated from others [18].
The limitation of this method is that when an existing vehicle position is acquired
by a target vehicle, then they are in the radar network of established vehicle [19, 20].
Such a scenario is depicted in Fig. 3.
Fig. 3. A potential Sybil attack
As shown in Fig. 3, A gets C’s position to be Lc. A claims to have victimized node
B that its position is Lc, and its ID is IDa. B recognizes a vehicle that is located at Lc
and then concludes that it is the location of A [15].
Resource Testing Method:

In this technique, assumption is generated in user’s mind for limited resources of each
physical entity. The concept of radio resource testing specifies that a message is
broadcast by a node to all its neighbor nodes and tries to listen the response message by
channel which is selected randomly.
The nodes which do not response the broadcast message positively in the same
channel are assumed as Sybil entities, and the Sybil attack is detected. Otherwise the
neighbor node response in the same channel [21].
The resources which can be shared between malicious vehicles and Sybil entities
are memory, IP, and computational recourses. So it is possible to generate for Sybil
attack conditions.
According to researcher opinion, this approach may not be suitable for VANET.
The malignant vehicle can produce numerous integrity which does not match with the
identity of the vehicle present in the network. So they are registered separately in other
list [17].
The main objective of resource testing method is to detect the Sybil attack and
check the fake identities from the attacker. But in practical scenario to obtain sufficient
fake identities for a attacker is not a difficult task. So this method may not be imple-
mented in VANET [20, 22].
In this paper, we propose a method to detect and a protection mechanism for Sybil
attack in vehicular communication. Also, we have assumed the attacker’s action in
different scenarios. In this case, we prefer position verification method as an efficient
technique to protect from Sybil attack in real-world environment. Extensive work is
still required in the future. Also, we focus on the acceptance range threshold technique
of position verification method to detect the wrong position of more number of Sybil
nodes in VANET. In future, the performance of our method can be examined using
NS2 and traffic simulators called SUMO and MOVE for real-world environment.
References
1. Karlof C, Wagner D (2003) Secure routing in wireless sensor networks: attacks and
countermeasures. Ad hoc Netw J 1:293–315
2. Sood M, Vasudeva A (2013) Perspectives of Sybil attack in routing protocols of mobile ad
hoc network. Comput Netw Commun (NetCom) 131:3–13
3. Newsome J, Shi E, Song D, Perrig A (2004) The Sybil attack in sensor networks: analysis &
defenses. In: Third international symposium on information processing in sensor networks,
IPSN 2004, pp 259–268
4. Pouyan AA, Alimohammadi M (2014) Sybil attack detection in vehicular networks. Comput
Sci Inf Technol 2(4):197–202
5. Hubaux JP, Capkun S, Luo J (2004) The security and privacy of smart vehicles. IEEE Secur
Priv Mag 2(3):49–55
6. 5.9 GHz DSRC. http://grouper.ieee.org/groups/scc32/dsrc/index.html
7. Yu B, Xu C-Z, Xiao B (2013) Detecting Sybil attacks in VANETs. J Parallel Distrib
Comput. Available online 9 Feb 2013
8. Douceur JR (2002) The Sybil attack. In Proceedings of the international workshop on peer to
peer systems, March 2002, pp 251–260
9. Saini M, Kumar K, Bhatnagar KV (2013) Efficient and feasible methods to detect Sybil
attack in VANET. Int J Eng Res Technol 6:431–440
10. Grover J, Gaur MS, Laxmi V (2011) Sybil attack in VANETs detection and prevention.
Published in security of self-organizing networks. Printed in the United States of America on
acid-free paper-10 in 2011
11. Newsome J, Shi E, Song D, Perrig A (2004) The Sybil attack in sensor networks: analysis
and defenses. In: Proceedings of international symposium on information processing in
sensor networks, April 2004, pp 259–268
12. Raya M, Hubaux JP (2007) Securing vehicular ad hoc networks. J Comput Secur 39–68
13. Khalili A, Katz J, Arbaugh WA (2003) Towards secure key distribution in truly ad-hoc
networks. In: Proceedings of IEEE workshop on security and assurance in ad-hoc networks
14. Martucci LA, Kohlweiss M, Anderson C, Panchenko A (2008) Self-certified Sybil-free
pseudonyms. In: WiSec’08: proceedings of the first ACM conference on wireless network
security, New York, NY, USA, ACM Press, pp 154–159
15. Yan G, Olariu S, Weigle MC (2008) Providing VANET security through active position
detection. Comput Commun 31(12):2883–2897
16. Pathre A, Agrawal C, Jain A (2013) Identification of malicious vehicle in vanet environment.
J Glob Res Comput Sci 4:30–34
17. Isaac T, Zeadally S, Camara JS (2010) Security attacks and solutions for vehicular ad hoc
networks. Commun IET 4(7)
18. Samara G, A-Salihy WAH, Sures R (2010) Security Analysis of Vehicular Ad Hoc
Networks (VANET). In: IEEE second international conference on network applications,
protocols and services
19. Ibrahim K (2011) Data aggregation and dissemination in vehicular ad-hoc networks.
Doctoral dissertation, Old Dominion University, Norfolk, Virginia
20. Rahbari M, Jamali MAJ (2011) Efficient detection of Sybil attack based on cryptography in
VANET. Int J Netw Secur Its Appl (IJNSA) 3:185–195
21. Zhou T, Choudhury RR, Ning P, Chakrabarty K (2011) P2 DAP-Sybil attacks detection in
vehicular ad hoc networks. IEEE J. Sel Areas Commun 29(3)
22. Levine BN, Shields C, Margolin NB (2006) A survey of solutions to the Sybil attack.
University of Massachusetts, Amherst
23. Hao Y, Tang J, Cheng Y (2011) Cooperative Sybil attack detection for position based
applications in privacy preserved VANETs. In: IEEE communications society in the IEEE
Globecom 2011 proceedings
Design of Monochromatic Photonic Filter
in Near-Infrared Region Using Plane Wave
Expansion Method
K. P. Swain1(&), Sangram Kishore Mohanty2, P. S. Das1,

and G. Palai1
1
Department of Electronics and Communication Engineering, GITA,
Bhubaneswar, India
2
Department of ECE, ICE, Bhubaneswar, India
1 Introduction
PWE is customarily used technique in the photonics fraternity to solve Maxwell

equation for finding the dispersion relation. Generally, it uses eigenvalue problem to
solve the Maxwell equation. Many literatures [1–10] implement this method, especially
in the field of photonics. In [1], PWE is used to compute the band structure by
considering material dispersion, whereas paraxial beam propagation is demonstrated in
[2] using PWE. Photonic band gap (PBG) is calculated using PWE method in [3–5] to
design a 2-D band-pass filter for C-band WDM application, band-pass filter using dual-
mode square lattice and band-pass filter using dual square ring, respectively.
Using PWE, a detail calculation of defect mode in PCF is proposed in [6] and in [7], a
detail investigation is proposed using PWE to design add/drop filter. A PWE-based
model for laser beam propagation is also proposed in [8] in anisotropic medium to
calculate the optical field distribution and in [9], PWE is reported to study the electric
field distribution in MSOF, whereas doping concentration is investigated in [10] in an
optical thyristor.
Photonics is nowadays found many applications to design the different kinds of
filters as described in [11–13]. An agile RF photonic notch filter is demonstrated
experimentally in [11], where silicon Bragg grating waveguide is implemented.
A tunable photonic single notch filter in microwave region is proposed and demon-
strated experimentally in [12] which is based on phase modulation technique and in
[13], tunable photonic band-pass filter based on Brillouin is presented.
Monochromatic photonic filters with some real-life application are designed and
implemented in [14–17] where silicon grating structure is used as background material.
PWE simulation method is carried out to study the reflectance curve of the filter. In the
present work, monochromatic filters are designed in near-infrared region using PWE
simulation method, where the width of SiO and Si decide the allowable signal
wavelength. Simulations are made for the three commonly used ranges like 800–900,
1300–1400 and 1500–1600 nm and a relation is also established between the width of
the grating materials and the allowable signal wavelength.

https://doi.org/10.1007/978-981-15-2774-6_31
Design of Monochromatic Photonic Filter in Near-Infrared Region … 249
2 Proposed Grating Structure
The design of monochromatic photonic filter using silicon grating structure is shown in
the figure where SiO and Si materials are used alternatively to form total 68 layers with
a defect at 18th and 52nd position. Defect or imperfection in material science plays a
vital role to change the physical and mechanical property in the crystalline materials.
Here, defect is introduced deliberately as a line defect to obtain the characteristics of an
ideal filter. Here, air replaces Si material at the 18th and 52nd layer whose refractive
index is 1. In the proposed structure, the combination of refractive index of the
materials (SiO and Si), the width of the materials (taken as tʹ for SiO and tʹʹ for Si) and
overall grating length decide the allowable signal wavelength in the range. Helmholtz
equation which is the representation of Maxwell equation is used to calculate the
reflectance of grating structure. The final equation can be obtained by solving the
Helmholtz equation with help of Crammer’s rule as given below (Fig. 1).
Fig. 1. Grating structure of proposed structure
k ¼ 2 ðn0 t0 þ n00 t00 Þ
where
nʹ Refractive index of SiO (taken as 1.913)
nʹʹ Refractive index of Si (taken as 3.6726)
tʹ Width of SiO (varies depending upon the allowable wavelength)
tʹʹ Width of Si (varies depending upon the allowable wavelength).
3 Simulation Result and Discussion
PWE simulations are carried out to get the reflectance characteristics of monochromatic
filters for NIR. Basically, simulations are implemented maximum times for three ranges
like 800–900, 1300–1400 and 1500–1600 nm but only two reflectance curve are dis-
played in Figs. 2, 3 and 4 for each range.
Fig. 2. Reflectance characteristics curve for the allowable signal wavelength 850.7 and
875.1 nm
Fig. 3. Reflectance characteristics curve for the allowable signal wavelength 1322 and 1370 nm
Fig. 4. Reflectance characteristics curve for the allowable signal wavelength 1514 and 1591 nm
Considering the reflectance characteristics curve (Fig. 2) for the allowable wave-
length of 850.7 nm and 875.1 nm in the 800–900 nm range, the reflectance curve is
approached to zero for the allowable signal 850.7 nm and 875.1 nm, whereas one for
rest part of the signal in the range. So, it clear that this particular design is used to allow
only 850.7 nm for the width of SiO (tʹ) is equal to 0.1075 µm and Si (tʹʹ) is equal to
0.055 µm while rejecting completely other signals within the range. Similarly, Figs. 3
and 4 are representing the reflectance characteristics of 1322 nm and 1370 nm in the
range 1300–1400 nm and 1514 nm and 1591 nm in the range 1500–1600 nm range,
respectively. In the entire reflectance curves, which are obtained from PWE simula-
tions, X-axis represents the range of the signals in nanometer (nm), whereas reflectance
is represented in Arb. Units by Y-axis.
After getting all the simulation results for each range, one interesting facts is
revealed that the width of the SiO and Si is varied almost linearly with the increasing
allowable signal wavelength in the range as shown in Fig. 5, 6 and 7. So, any
researcher can easily find the width parameters of SiO and Si by applying simple
mathematics or any machine learning algorithm.
Width of SiO (t ) and Si (t ) in µm
Width of SiO and Si w.r.t allowable signals of filters

within the range of 800- 900 nm
0.12
0.1
R≤ = 0.979
0.08
0.06
0.04 R≤ = 0.9794 Thickness of Si in

nm(t1)
0.02
Thickness of SiO in nm
(t2)
0
800 820 840 860 880 900
Signal wavelength for the range of 800- 900 nm
Fig. 5. Width of SiO and Si w.r.t. allowable signal of range 800–900 nm

Width of SiO and Si w.r.t allowable signals of filters

Width of SiO (t ) and Si (t ) in µm 0.2 within the range of 1300- 1400 nm
0.16 R≤ = 0.9918
0.12
0.08
R≤ = 0.9861
0.04
0
1300 1320 1340 1360 1380 1400
Signal wavelength for the range of 1300- 1400 nm
0.25 Width of SiO and Si w.r.t allowable signals of

Width of SiO (t ) and Si (t ) in µm
filters within the range of 1500- 1600 nm
0.2
R≤ = 0.9897
0.15
0.1
R≤ = 0.9896
0.05
0
1500 1520 1540 1560 1580 1600
Signal wavelength for the range of 1500 - 1600 nm
4 Conclusion
Monochromatic photonic filters in the near-infrared region are successfully designed

using PWE simulation method, where the width of waveguide structure (SiO and Si) plays
a vital role to select the allowable signal. Aside this, the allowed or disallowed by the
waveguide is lucidly investigated through the analysis of the reflectance characteristics
pertaining to the wavelength of the signal. It is also divulged that the increasing allowable
signal wavelength is varied almost linearly with the width of the waveguide structures.
References
1. Guryev IV, Sukhoivanov IA (2007) Plane wave expansion method with considered material
dispersion. In: 2007 9th international conference-The Experience of Designing and
Applications of CAD Systems in Microelectronics (CADSM). IEEE, pp 23–24
2. Shuklal BK, Patel RH (2008) Simulation of paraxial beam propagation using plane wave
expansion method. In: 2008 International Conference on Recent Advances in Microwave
Theory and Applications (ICM). IEEE, pp 652–656
3. Robinson S, Nakkeeran R (2011) Two dimensional photonic crystal ring resonator based
bandpass filter for C-band of CWDM applications. In: 2011 National Conference on
Communications (NCC). IEEE
4. Kamiji Y, Nagaoka N, Chun-Ping C, Anada T, Jui-Pang H (2012) A novel photonic crystal
bandpass filter using degenerate modes of point-defect microcavity for terahertz commu-
nication system. In: 2012 Asia Pacific Microwave Conference Proceedings (APMC). IEEE,
pp 583–585
5. Sharma A, Inaniya PK (2015) Dual square ring with 3 3 dielectric rods structure based
band pass filter using two dimensional photonic crystal. In: International Conference on
Computer Science and Network Technology (ICCSNT). IEEE pp 45–49
6. Zhu Y, Zhuang Y, Shi X (2014) An improved algorithm of photonic crystal fibers’ defect
mode based on the plane-wave expansion and supercell method. In: 2014 International
Symposium on Computer, Consumer and Control (ISCCC). IEEE, pp 322–324
7. Kuzma A, Uherek F, Skriniarova J, Kuzma A, Uherek F (2016) Photonic crystal based
add/drop filters for sensing. In: 2016 Photonics North (PN). IEEE
8. Pilgun Y, Smirnov E (2016) Plane-wave expansion based modelling of laser beam
propagation in anisotropic medium. In: 2016 IEEE 7th International Conference on
Advanced Optoelectronics and Lasers (CAOL). IEEE, pp 139–141
9. Das PS, Jena D, Palai G (2016) PWE approach to MSOF for beam splitting application.
Optik 127:10228–10231
10. Panda A, Mishra CS, Palai G (2016) PWE approach to optical thyristor for investigation of
doping concentration. Optik 127:4831–4833
11. Burla M, Bazargani HP, St-Yves J, Shi W, Chrostowski L, Azana J (2014) Frequency agile
microwave photonics notch filter based on a waveguide Bragg grating on silicon. In: 2014
international topical meeting on Microwave Photonics (MWP) and the 2014 9th Asia-Pacific
Microwave Photonics Conference (APMP). IEEE, pp 392–394
12. Xu E, Yao J (2015) Frequency- and Notch-depth-tunable single-notch microwave photonic
filter. IEEE Photo Technol Lett 1–4
13. Mahendra A, Magi E, Choudhary A, Liu Y, Marpaung D, Eggleton BJ (2018) High
performance, low noise figure Brillouin-based tunable microwave photonic bandpass filter.
In: 2018 international topical Meeting on Microwave Photonics (MWP). IEEE
14. Swain KP, Palai G, Moharana JK (2017) Analysis for ‘101’ channels of MUX/DEMUX
using grating SOI Structure at sub nanometer scale. Optik 129:78–82
15. Swain KP, Palai G, Moharana JK (2019) Optical filter based electrical control vis-à-vis cloud
data: a new hybrid optoelectronics device for embedded application. Optik Int J Light
Electron Opt 178:964–969
16. Swain KP, Palai G, Moharana JK (2019) Design and implementation of opto-electro decoder
using photonic structure: a new application of Li-fi vis-a-vis optical embedded system. Optik
Int J Light Electron Opt 178:658–663
17. Swain KP, Palai G (2018) Photonic structure for embedded application: realization of optical
filter based locking system. Optik Int J Light Electron Opt 169:344–349
Modeling, Analysis, and Control of Vehicle
Suspension System Based on Linear Quadratic
Regulator
Akshaya Kumar Patra1(&), Alok Kumar Mishra1, Anuja Nanda1,

Narendra Kumar Jena1, and Bidyadhar Rout2
1
Department of EEE, ITER, S‘O’A University, Bhubaneswar 751030, India
narendrajena@soa.ac.in
2
Department of EEE, Veer Surendra Sai University of Technology, Burla
768017, India
1 Introduction
In recent times, many researchers in the field of vehicle suspension (VS) system
dynamics have devoted to arrive at an optimal solution with a compromise between
vehicle handling, ride comfort, and stability. There is a need for better approach today
as the above problems are very much evident in the modern vehicle cases. Specific to
the large sedan and luxury cars, even if excellent ride qualities are achieved, it is
limited to acquiring adequate handling behavior. With reference to sports vehicles,
although it is provided with very good handling capability, but fails to provide
desirable ride quality. There are many options in between for variations in the
designing stage of the vehicle manufacturers to meet the customer needs. Designing
point of view, passenger comfort and vehicle control are the two primary objectives to
be considered. Road disturbances such as bumps or potholes are to be handled not to
sacrifice the passengers comfort. At the control stage, these factors are generally
considered either through keeping the vehicle body from rolling and pitching exces-
sively, or by maintaining good contact between the tire and the road.
Nowaday’s hydraulic dampers (shock absorbers) and springs are extensively used
for vehicle suspensions. As a principle, these are charged with the job of absorbing
bumps, minimizing the vehicle’s body motions during acceleration, braking and
turning of the vehicle, and keeping the tires in contact with the road surface. However,
from the design point view, these objectives are contradictory to each other to achieve
all simultaneously at their optimum level.
The spring and the damper are the two essential components in VS design. The
spring design mostly depends on the weight of the vehicle. The damper design is based
on the suspensions placement on the compromise curve and so, it is essential to be
perfectly chosen to make the optimal vehicle performance for any type of vehicle. For
ideal performance, the damper should act such that passengers isolate from low-
frequency road disturbances and absorb high-frequency road disturbances. High
damping is essential to achieve for best isolation of passengers from low-frequency
disturbances.

https://doi.org/10.1007/978-981-15-2774-6_32
Modeling, Analysis, and Control of Vehicle Suspension System … 255
However, it is desirable to design a high-damping system, on the other hand it

degrades the high-frequency absorption rate. In other way, by providing low damping
the damper offers adequate high-frequency absorption in terms of scarifying low-
frequency isolation. To meet these contradictory objectives, it is essential to design and
focus on automotive suspensions without compromising any of the factors mentioned
above. As a solution to the above, three types of suspensions can be improved. These
are passive, fully active, and semi-active-type of suspensions. The spring and damper
are the two basic components of the conventional passive suspension. Both the com-
ponents are considered and fixed at the design stage. The suspension stores energy in
the spring. Later it dissipates energy through the damper.
It is very much needed for further research to develop robust control algorithms to
enhance the performance of VS system. This in turn enhances the vehicle capabilities
to handle the aforementioned issues. An ideal design of VS system needs to achieve
many performance characteristics such as: (1) control of body movement, (2) control of
suspension movement and force distribution. Performance point of view, the VS should
able to isolate the body for comfort against the road impact and inertial disturbances.
These are generally associated with cornering and braking or acceleration of the vehicle
system [1]. As discussed above, many performance objectives are conflicting in nature,
that to enhance one the other factor degrades. Considering all the objectives, as a goal
for the designing of a suspension system is difficult to meet [2]. Minimization of
vertical force to the passengers can be achieved by minimizing the vertical vehicles
body acceleration of the suspension. Another factor which plays a vital role in pas-
sengers comfort is the optimal contact between wheel and road surface and this is
essential in various driving conditions in order to maximize safety factor [3]. Among
few design in the past, the system presented in [4] based on unconstrained optimiza-
tions for passive suspension (PS) system case is widely accepted and used. This suc-
cessfully performs to achieve the desirability of low suspension stiffness, reduced
unsprung mass (UM), and an optimum damping ratio for the better controllability. As
the PS system performs satisfactorily some extent, it is considered in many applications
for the VS system. However, both the suspension spring and damper do not supply
energy to the suspension system. They only control the motion of the vehicles body
and wheel by limiting the suspension velocity. This is computed according to the rate
specified by the designer. To overcome this issue, the active suspension (AS) system is
considered as an efficient option for this application. The AS systems have the ability to
dynamically respond to changes in the road profile. It is due to this fact that it can
supply energy to produce relative motion between the body and the wheel. Sensors are
provided in suspension system to measure the parameters dynamically. The parameters
such as body velocity, suspension displacement, wheel velocity, and wheel and body
acceleration are sensed and computed for the controller as input parameters [5]. An AS
can be thought as the integration of the passive components to actuators that supply
additional forces. These additional forces are computed by a feedback control law
based on the input data from the sensors fixed to the vehicle. For modeling the real-time
dynamic conditions, the uncertainties due to system design and other external distur-
bances are needed to be considered for the controller. This motivates for a better control
design to increase the robustness and controllability under uncertainties and
disturbances.
256 A. K. Patra et al.
During the past three decades, so many control strategy techniques are suggested
and tested by absorbing the shocks due to a rough and bumpy road in case of VS
system. Time-discrete and switching PID control strategy is implemented in VS
problems with variable control gains based on the measured suspension variables [6,
7]. However, the optimal gain parameter setting, a lesser range of robust control, and
need of change of gain setting with varying conditions are the major limitations to limit
the real-time application of these controllers. Among other projected robust control
algorithms applied for limiting the oscillation and velocity of the VS system are fuzzy-
logic control [8], fuzzy-PID control [9], genetic algorithm [10], neural network [11],
neuro-fuzzy (NF) control [12], linear quadratic regulator (LQR) [13], H-infinity control
[14], and sliding mode (SM) control [15]. However, even if these control techniques
are implemented effectively by absorbing the shocks due to the rough and bumpy road
in case of VS system with enhanced accuracy and damping of oscillation, still fail to
handle various constraints and random change found in a suspension environment.
These approaches are not fully insensitive to the disturbances and the uncertainties of
the model in spite of the improved performance. Hence, optimal control parameters
setting for still better performance and for avoiding slow response following road
disturbance (road impact), the current work suggests an alternative novel technique
based on the LQR concept. Application of the suggested approach to control the
oscillation and velocity in vehicle suspension system results to ensure a better robust
controller in comparison to other contemporary well-established approaches under both
harmonized and incompatible uncertainties.
2 Problem Formulation and Modeling
2.1 System Overview

The overall model of VS system is depicted in Fig. 1. The road disturbance wðtÞ and
vðtÞ are reflected as the process disturbance and the sensor noise, respectively, in this
study. The controller receives information about the oscillation of the VS system as
input to provide the optimal control force uðtÞ, and it is applied in between the UM and
the sprung mass (SM) to reduce the relative motion between them.
2.2 Modeling of VS System

In the year 1958, the concept of VS system was emerged into limelight. It has been
found in the later period that the vibration suppression capability is limited in case of
traditional PS and semi-active suspension systems. This leads to focus on an AS system
with additional control forces as a better alternative to the above drawbacks to suppress
the oscillations and used in modern vehicle industry. The major factor that differen-
tiates the active suspension system performance wise is the ability to inject energy into
vehicle dynamic system via actuators unlike dissipates energy in case of conventional
suspension system (SS). To produce a desirable control force to handle the variety of
road disturbances (road impact) in real-time applications, actuators are placed in
between the UM and the SM in AS system. The most challenging task of designing an
AS system is to enhance the ride comfort by absorbing the shocks due to the rough and
bumpy road. The force actuator in the case of the AS system is capable to add and also
dissipates energy from the system. This results an increase in the ride comfort and
vehicle handing due to the capability of the suspension system to regulate the vehicle
altitude and to eliminate the adverse effects of braking and vehicle roll during cornering
and braking. The schematic model of the VS system is illustrated in Fig. 1.
Fig. 1 Schematic model of the VS system
All the physical activities of the VS system are mathematically expressed and
specified by Eqs. (1) and (2). The SIMULINK diagram of the VS model is established
with respect to Eqs. (1) and (2) as displayed in Fig. 1 [13].
2
d x1 ðtÞ dx1 ðtÞ dx2 ðtÞ
m1 ¼ b1 k1 ðx1 ðtÞ x2 ðtÞÞ þ uðtÞ ð1Þ
dt2 dt dt
2
d x2 ðtÞ dx1 ðtÞ dx2 ðtÞ
m2 ¼ b 1 þ k1 ðx1 ðtÞ x2 ðtÞÞ
dt2 dt dt
ð2Þ
dvðtÞ dx2 ðtÞ
þ b2 þ k2 ðvðtÞ x2 ðtÞÞ uðtÞ
dt dt
2.3 Linearization of VS System

Figure 1 shows the non-linear model of VS system, and its simplified form is shown in
Fig. 1. The state-space equation of the VS model with wðtÞ and vðtÞ can be expressed
as [13]:

x_ m ðtÞ ¼ Am xm ðtÞ þ Bm uðtÞ þ Bd wðtÞ
ð3Þ
yðtÞ ¼ Cm xm ðtÞ þ Dm uðtÞ þ vðtÞ
where, wðtÞ, xm ðtÞ, vðtÞ, and yðtÞ are the road disturbance, state variable, sensor noise,
and measured output of the VS system, respectively. The state variable xm ðtÞ consists
of x1 ðtÞ, x2 ðtÞ, x_ 1 ðtÞ, and x_ 2 ðtÞ. The Am , Bm , Cm , Dm , and Bd represent the matrices of
the state-space model of VS system. The state-space matrices are found out from the
linearization of the system dynamics nearby the reference point. The detailed derivation
is described in the literature [13]. In the current study, the initial reference such as the
oscillation and the velocity of the suspension system are taken as zero. The VS model
is linearized surrounding the reference point for the design of robust controllers to
control the oscillation and the velocity of the suspension system within the stable range.
2.4 Performance Analysis of VS System

There are four poles in VS system. One of them lies in right-hand side of the complex
plane. As a result, the system becomes unstable. This needs the design of an adaptive
controller for improving the stability of the system by means of shifting the poles into
the left-hand side of the complex plane. The VS system in the open loop form is
depicted in Fig. 1.
The VS system has two inputs and two outputs. The control force uðtÞ and road
disturbance wðtÞ are the two inputs of the VS system. The oscillation and velocity of
the SS are the two outputs of the VS system. An uncontrolled oscillation and velocity
of the suspension system are being observed owing to the application of 0.1 m
impulsive road disturbance on the unsprung mass at the simulation time of 1.0 s. The
irregular oscillation and velocity of the suspension system are illustrated in Fig. 2a, b.
Figure 2a, b illustrates the unstable dynamics under various model uncertainties and
disturbances. These unstable dynamics can be reduced by applying the suitable control
algorithms. In this case, the oscillation of the suspension system is the most essential
outcome needs to be controlled within a stable range through suitable control tech-
niques, and velocity of the suspension system is analyzed in order to view the motion
trajectory.
Fig. 2 a Response of oscillation of the SS with road impact; b response of velocity of the SS
with road impact
3 Control Algorithm
The control specifications such as settling time ts , steady-state error ess , maximum
overshoot OMax , and maximum undershoot UMax are evaluated and examined with
proper validation of the controller action.
3.1 LQR Design

The linearized model of the VS system as discussed in Sect. 2.3 has been taken for the
formulation of the suggested control algorithm to regulate the oscillation and the
velocity of the suspension system. For accomplishing an upgraded performance and the
adjustment of controller parameters of the suggested LQR, it is integrated to the lin-
earized model of the VS system as illustrated in Fig. 3. The linearized model of the VS
system with wðtÞ and vðtÞ is formulated as [16–19]:
Fig. 3 VS system with LQR in the state-space representation

dxm ðtÞ
¼ Am xm ðtÞ þ Bm uðtÞ þ Bd wðtÞ ð4Þ
dt
yðtÞ ¼ Cm xm ðtÞ þ vðtÞ ð5Þ
uðtÞ ¼ Kc eðtÞ ð6Þ
Figure 3 shows the linearized model of the diabetes patient with the LQR gain Kc .
The evaluation procedures of Kc is mentioned in detail in Sect. 3.1.1. The evaluation of
the TF of the LQR is described in Sect. 3.1.2.
3.1.1 LQR Gain Kc

The Kc is evaluated with respect to the system dynamics for the minimum value of the
quadratic performance index j as specified in Eq. (7).
Z1

j¼ xm ðtÞT Q1 xm ðtÞ þ uðtÞT R1 uðtÞ dt ð7Þ
0
where, Q1 ¼ CmT q1 Cm . The Q1 , q1 , and R1 are denoted as the positive semi-definite

weighted matrix of state, the intensity of a weighted matrix of state, and the positive
definite weighted matrix of input, respectively. The Kc is estimated for the minimum
value of the quadratic performance index as specified in Eq. (8).
Kc ¼ R1
1 Bm Pk
T
ð8Þ
Pk is the solution of the Controller Algebraic Reccati Equation (CARE).

The CARE is specified as follows:
ATm Pk þ Pk Am Pk Bm R1
1 Bm Pk þ Cm q1 Cm ¼ 0
T T
ð9Þ
3.1.2 TF of LQR
The TF of suggested controller KðsÞ is estimated with the use of Kc . The KðSÞ of LQR
is specified as follows:
KðsÞ ¼ Kc ðsIn Am þ Bm Kc Þ1 Bm ð10Þ
The packed matrix notation of KðSÞ is specified as follows:

Am Bm Kc Bm
KðsÞ ¼ ð11Þ
Kc 0
For the design of suggested LQR, the optimal values of control parameters are
evaluated with help of MATLAB and represented as in the Table 1.
Table 1 Optimal values of control parameters

Q1 R1
1000 CmT Cm 1
4 Result and Discussions
The performance of the closed loop VS system with LQR is described in detail in this
section. The suggested control approach is compared with other popular control
algorithms to justify its enhanced performance.
4.1 Performance Analysis of VS System with LQR

In this section, all physical activities of the closed loop VS model with suggested LQR
are examined under different conditions and the huge deviation of road disturbance.
The time domain response of the oscillation and the velocity of the suspension system
with 0.1 m impulsive road disturbance at the simulation time of 1.0 s are displayed in
Fig. 4a, b. The output results clearly specify the suspension system attains the zero
oscillation and zero velocity with less settling time and also attains the balance position
where the system is absolutely steady. To achieve the enhanced system response, the
required control force uðtÞ is generated by the suggested LQR.
Fig. 4 a Response of oscillation of the suspension system with road impact based on LQR;
b response of velocity of the suspension system with road impact based on LQR
4.2 Comparative Study

The proposed LQR control approach is compared with other popular control approa-
ches such as PID, fuzzy, NF, LQR, H1 , and SM to justify its enhanced performance as
a controller. Figure 4a illustrates the effect of road impact in the oscillation of the
suspension system with the suggested LQR approach. Table 2 presents a comparative
analysis with respect to ts (sec), OMax ðmÞ, UMax ðmÞ, noise (%), and ess ð%Þ . The effect
of road disturbance in the oscillation of the suspension system applying different
control approaches such as PID, fuzzy, NF, LQR, H1 , and SM is also documented in
Table 2 based on the references, and [6, 8, 12–14] and [15], respectively. Similar
working conditions are followed with the same level of sensor noise in all control
techniques application for comparison.
Table 2 Comparative result analysis related to oscillation of the suspension system

Controller PID Fuzzy NF LQR H1 SM LQR
[6] [8] [12] [13] [14] [15] (Proposed)
Road disturbance 0.1 0.1 0.1 0.1 0.1 0.1 0.1
(m)
ts (sec) 4.2 4.1 3.7 3.4 3.1 3.0 2.3
OMax ðmÞ 0.0050 0.0049 0.0047 0.0045 0.0043 0.0040 0.0022
UMax ðmÞ 0.0042 0.0045 0.0046 0.0042 0.0041 0.0048 0.0026
Noise (%) 10 10 5 5 5 5 1
ess (%) 0 0 0 0 0 0 0
The oscillations of the suspension system under 0.1 m impulsive road disturbance
are tested. The corresponding results are presented for the various control approaches
along with the proposed LQR with respect to time domain specifications such as
OMax ðmÞ, UMax ðmÞ, and ts (sec). The results signify the better controllability of the
LQR. The simulation results also demonstrate the high noise and chattering elimination
capability with high robustness for the proposed approach. Overall, by looking to the
above comparative analysis, the findings of suggested approach advantages are the
higher accuracy and stability, more robustness, high noise and chattering elimination
capability, and better capability to handle uncertainty under various conditions, and
huge deviation of road disturbance.
5 Conclusions
In this manuscript, a novel control strategy LQR is proposed based on self-tuned

approach. To justify its enhanced performance, it has been applied and tested to control
the oscillation and the velocity of the suspension system in the vehicle. In suggested
LQR, self-tuned algorithm is utilized to enhance the control performance. The com-
parative results clearly reflect the suggested LQR is arrived at better performance than
the other control approaches such as PID, fuzzy, NF, LQR, H1 , and SM with respect to
stability, reliability, and robustness under various abnormal conditions and distur-
bances. The better performance of the suggested approach in terms of improved
accuracy and stability, enhanced robustness, high noise elimination capability, and
better ability to control uncertainty justifies its real-time application.
References
1. Wright PG (1984) The application of active suspension to high performance road vehicles.
In: Microprocessors in fluid engineering IMechE conference publications
2. Fateh MM, Alavi SS (2009) Impedance control of an active suspension system. J Mechatron
19:134–140
3. Lin JS, Kanellakopoulos I (1995) Nonlinear design of active suspension, In: 34th IEEE
Conference 1995, vol 17, pp 45–59
4. Alleyen A, Hedrick JK (1995) Nonlinear adaptive control of active suspensions. IEEE Trans
Control Syst Technol 3(1):845–860
5. Esmailzadeh E, Taghirad HD (1997) Active vehicle suspensions with optimal state feedback
control. J Mech Sci 200(4):1–18
6. Kumar MS (2008) Development of active suspension system for automobiles using PID
controller
7. Talib MHA, Darns IZM (2013) Self-tuning PID controller for active suspension system with
hydraulic actuator. In: 2013 IEEE Symposium on Computers & Informatics (ISCI), April
2013, pp 86–91
8. Pekgökgöz RK, Kısa M (2010) Actıve suspension of cars using fuzzy logic controller
optimized by genetic algorithm. Int J Eng Appl Sci 2(4):27–37
9. Yang Q, Yuan C (2008) Fuzzy-PID control on sem-i active air suspension. Trans Chin Soc
Agric Mach 39(9):24–29
10. Du H, Lam J, Sze KY (2003) Non-fragile output feedback H1 vehicle suspension control
using genetic algorithm. Eng Appl Artif Intell 16:667–680
11. Wang YJ (2007) Analysis of vehicle suspension control using neural networks. Northeastern
university, Shenyang
12. Aldair AA, Wang WJ (2012) A neurofuzzy controller for full vehicle active suspension
systems. J Vib Control 18(12):1837–1854
13. ElMadany MM (1999) Linear quadratic Gaussian control of a quarter-car suspension. Veh
Syst Dyn 32(6):479–497
14. Kaleemullah, M., Hasbullah, F.: Design of robust H∞, fuzzy and LQR controller for active
suspension of a quarter car model, In 2011 4th International Conference on Mechatronics
(ICOM), (2011, May) 1–6, IEEE
15. Li H, Liu H (2013) Adaptive sliding-mode control for nonlinear active suspension vehicle
systems using T-S fuzzy approach. IEEE Trans Indus Electron 60(8):3328–3338
16. Patra AK, Rout PK (2018) Backstepping sliding mode Gaussian insulin injection control for
blood glucose regulation in TIDM patient. J Dyn Sys Meas Control 140(9):091006
17. Patra AK (2018) Backstepping model predictive controller for blood glucose regulation in
type-I diabetes patient. IETE J Res. https://doi.org/10.1080/03772063.2018.1493404
18. Petra AK (2017) Adaptive continuous-time model predictive controller for implantable
insulin delivery system in type I diabetic patient. Optimal Control Appl Meth 38(2):184–204
19. Patra AK, Rout PK (2017) Adaptive sliding mode Gaussian controller for artificial pancreas
in TIDM patient. J Process Control 59:13–27
Performance and Evaluation of Different
Kernels in Support Vector Machine for Text
Mining
Ashish Kumar Mourya, ShafqatUlAhsaan, and Harleen Kaur(&)
Department of Computer Science and Engineering, School of Engineering

Sciences and Technology, JamiaHamdard, New Delhi, India
1 Introduction
The significance of utilization of an electronic data has increased in recent years. In

recent years, approximately 75% of digital data is in the form of text or multimedia
form [1]. This tremendous amount of data is not only unstructured and unclassified but
also contains unless data, useful data, heterogeneous data, and scientific data. Cur-
rently, more than 35% of corporation is working with this huge data. Text mining and
analysis is similar to data mining, excluding data mining techniques applied on rela-
tional data, whereas text pre-processing works on unstructured or semi-structured data
[2]. In text mining or analysis, we extracted the useful information from heterogeneous
or multimedia data by separating the plain text and heterogeneity and analyzing the
classification frequency or correlation. Machine learning technique such as SVM or
DT, the possibility of the high frequency result even when the amount of data is
heterogeneous or enormous, is confirmed. Data mining is liable for pulling out of
implied, potential data and liable for explicitly data [3]. Machine learning techniques
and algorithms designed to analyzed these explicitly data know the configuration of
numbers, and they are pre-programmed to process it with precision. We have used
SVM for predicting the performance and analyzing the multimedia in text mining.
SVM algorithm is also known to be very powerful in terms of accomplishing
performance on the numerical data. SVM technique was introduced to solve the linear
and binary classification problem of two classes [4]. Now, it is currently implemented
in any kind of problem [9]. We have now used them in classification as well as
regression problem. It is broadly used in lots of applications like gene classification,
time series prediction, image segmentation, and face recognition. SVM is successfully
applied on different classification tasks that take statistical data or heterogeneous data.
On the other hand, artificial neural network techniques are biologically inspired
machine learning methods designed like a human brain process the information [5]. Its
critical structure allows knowledge practically with any function beneficially with
adequate hidden layers of processing nodes with training data. SVM and neural net-
work are normally O (n3) or at best quadratic ally complex with more examples.
SVM falls under the category of supervised machine learning. SVM classifier
performs classification by example within multiple data classes. SVM algorithms are
used in the field of computational biology on a large number of problems, like

https://doi.org/10.1007/978-981-15-2774-6_33
Performance and Evaluation of Different Kernels in Support … 265
detection of homology in protein, analyses of gene expressions and tandem mass, etc.
SVMs work on the principle of structural risk minimization [6]. SVMs are universal
learning systems [7]. SVM, within a multi-dimensional feature space, works on training
samples of multiple classes in order to get a best possible hyper-plane that separates the
two classes. The training samples that fall on hyper-plane that is closest to each other
are called support vectors, and the minimum distance between the support vectors is
called margin. The result with the maximum margin is considered as optimal, and the
error rate for this result is very low. SVM, by means of a kernel function, projects
vectors into a higher-dimensional input space and obtains the best possible hyper-plane
that performs the distinction of data classes via optimization function. Let consider
pattern x; the corresponding estimated label Z ^ is given by equation (1).
!
X
n
Z^ ¼ sign½f ð xÞ ¼ sign ai y i K ð x i ; x Þ þ c ð1Þ
i¼1
Kernel functions or tricks give a principled means to characterize the data and thus
depict assumptions from different variety of data. The kernel methods while deploying
multiple kernel functions correspond to a collection of varied data objects and char-
acterize the relationship among any pair of data objects. For example, that defines a set
of multiple data objects that can be adequately and sufficiently represented by means of
‘m * n’ matrix of pair-wise kernel standards.
While dealing with heterogeneous data, SVM makes use of different types of kernel
functions that perform efficiently on the data objects and must be combined numeri-
cally. In [8], the authors used the SVM approach in their study to combine gene data
and phylogenetic profiles in an un-weighted mode, by deploying numerical operators to
axis the SVM within dataset correlations surrounded by features, whereas do not take
into account the correlations among various datasets [9]. Un-weighted sums of kernels
have also been used successfully in the forecasting of protein relations.
Researchers proposed a new learning method known as multi-kernel learning
(MKL), which is an ensemble of different kernels contained by the SVM classifier itself
[10]. These methods combine a single optimization procedure, which concurrently
finds the solution for classification followed by SVM and weights on the individual
facts types in the heterogeneous dataset.
2 Background
In [11] this paper, a comparison is made among multilayer perceptron and convolu-
tional network and concluded that convolutional networks perform better as compared
to multilayer perceptron. The experimental results are performed on Czech newspaper
text dataset. Jadon and Sharma [12] deployed the SVM algorithm for document
classification. In their approach, they used SVM to categorize the text documents
automatically. They worked on Reuters-corpus-text document dataset and obtained
86% accuracy while applying SVM algorithm. Devika et al. [13] applied the sentiment-
266 A. K. Mourya et al.
based classifier to analyze and determine the relevant news content. Their dataset
consists of 250 news documents in the form of text labeled with multiple sentiments on
which classification is done. Arundthati [14] explained that Naïve Bayes classifier is
used for text classification. They first applied the approach in linear way and then
hierarchically. The hierarchical method results to be more effective and improved the
efficiency and accuracy of the classifier.
Purohit [15] has compared multiple types of sentiment analysis methods like rule-
based approach, learning classifier approach, and lexicon-based approach. In their
work, the parameters that are being highlighted are accuracy, performance, and effi-
ciency, irrespective of advantages and disadvantages of these techniques. Tilve and
Jain [16] used the decision tree for the assessment of dataset related to educational
background. The dataset constitutes the name, marks, enrollment number, and grades
of students. A number of algorithms have been deployed like random forest, REPTree,
Hoeffding, and J48. It has been evaluated that the decision tree algorithms show a
descent increment in performance while dealing with student’s dataset.
Kumar and Verma [17] went through different research works and used different
datasets for sentiment analysis. In their work, a number of algorithms such as asso-
ciation rule-based decision tree and Naïve Bayes classifiers have been used. A com-
parison is performed among these algorithms. The algorithm proposed by Purohit A.
et al. makes use of Porter stemmer, association rule, a priori algorithm, Naïve Bayes
classification which showed a nice performance and attained a classification accuracy
of 75%. Vaghela et al. [18] compared the proposed algorithms, Naive Bayes and vector
space model, while deploying them on newsgroup datasets. The dataset constitutes 50
categories of news text data. The experimental results show that Naive Bayes classifier
has best accuracy. Kharde and Sonawane [19], Kalarikkal and Remya [20] explored
feature selection and submitted that if it is done reasonably, the classification algo-
rithms would perform better specially in terms of accuracy. In their study, a comparison
among lexicon and machine learning algorithms has been performed and came to the
point that machine learning algorithms are domain dependent. Supervised learning
classifiers have been used in the comparative study, and the authors explored that
supervised learning models are easy to train and need much less maintenance.
Supervised machine learning models acquire high accuracy in comparison to lexicon-
based models.
Das et al. [21] performed a detailed review and study of supervised and untrained
methods of learning by means of some estimation metrics. They focus on challenges
and applications of sentiment analysis. A comparative study has been carried out on
multiple classification algorithms such as Naive Bayes, SVM, logistic regression, and
random forest. Shrivatava et al. [22] mainly focused on processing the datasets using
machine learning techniques like support vector network, principal component anal-
ysis, and maximum entropy. They used qualitative training set for better implemen-
tation and outcome. Chavan et al. [23] developed an application that allows users to
gather data from social media Twitter handle, analyze the data, and visualize it by
means of pie charts and tables. Medhat et al. [24] deployed an efficient method to sort
out the structure of tweets with the help of SVM and attained a precision of 70.5%.
Senthilkumar and Paulraj [25] proposed sentiment classification using multiple
techniques of machine learning, particularly Naïve Bayes, SVM, and decision trees.
They used these classifiers to carry out the text categorization. In their study, the
authors came to know that SVM classifier performs better in terms of accuracy as
compared to other classifiers as Naïve Bayes and decision trees. In addition to accu-
racy, SVM is the best candidate to adjust the parameter settings automatically. Moraes
et al. [26] studied and came up with a detailed summary regarding utilization of
sentiment classification via feature selection.
The survey highlighted those techniques of sentiment classification that are very
effective and depicted how to choose features that play a key role in improving the
performance of sentiment classification. Kang et al. [27] presented a quality of clas-
sification techniques that can be deployed in data mining. In this paper, a couple of
classification techniques like SVM, k-nearest neighbor classifier, Bayesian networks,
neural network, and decision tree are reviewed. Abbasi et al. [28] classified the reviews
in text form for positive and negative sentiments. The authors evaluated the support
vector network and artificial neural networks on the bases of document-level sentiment
analysis.
The empirical results presented that the performance of artificial neural networks
generates superior results as compared to SVM classifiers. Terachi et al. [29] presented
an innovative improved Naïve Bayes algorithm. The proposed algorithm performed
better and showed a descent improvement in comparison to basic Naïve Bayes algo-
rithm when trained on bigram and unigram features. The proposed algorithm expected
very powerful than SVM and Naïve Bayes. [30] evaluated the Naïve Bayes and
maximum entropy classification models for Twitter data, and the main objective of
their study is the performance evaluation. The exploratory results that are being carried
out in their study clearly proved that Naïve Bayes models are very proficient as
compared to maximum entropy model.
3 Experiments and Results
In our experiments, we have compared the performance of SVMs using polynomial and
RBF kernels using text mining. For performing experiment, we have taken dataset from
UCI data repository. The name of dataset is the Reuters-21578 R8 dataset. The data has
various documents whose outcome classes belong to eight of the ten most frequent
document classes of Reuters’ articles. We have split the entire data into two parts. The
first part is training dataset, which contains 5485 documents, and the second part
contains 2189 documents for testing purpose.
The problem of text classification consists of text example, training example, and
newly derived example. Newly derived examples consist of predictor feature only, and
testing examples are used to weigh up the models; therefore, they have both predictor
and target attributes. The objective of text classification is to construct a model as a
function of the values of the attributes using training examples or documents that
would then forecast the target values of the new documents as accurately as possible.
The proposed study has applied different corpus tools on the raw data as shown in
Fig. 1. The raw data contains various stop words such as the, and, is, and for. This
process is known as pre-processing. There are many libraries available in R package for
performing various tasks on text classification. Caret and tm are used in this study.
Before evaluating the model performance, we have tested the cross-validation using
tenfold methods. This method uses 10 time repeated values for evaluating the model.
Thereafter, we have applied three different kernel tricks with SVM. The linear SVM
has achieved 0.91 training accuracy, whereas 0.89 testing accuracy. The radial basis
function has tremendous performance over any kind of data as shown in Fig. 2. These
kernel tricks perform good on text as well as numeric data. The outcomes of all three
kernel tricks have been shown in Table 1.
Fig. 1. Word cloud of the raw data

Fig. 2. Accuracy plot with RBF kernel
Table 1. Performance table

Kernels Training accuracy Testing accuracy
Linear kernel 0.91 0.89
Polynomial kernel 0.86 0.90
Radial basis function 0.94 0.91
4 Conclusion
SVM is a subfield of machine learning. SVM can be applied in both classification and
regression problem. This paper discusses how different kernel functions that can be
implemented and applied in conjunction with the SVMs in the automatic text classi-
fication problem. Several models for text analysis are represented; in most of them, text
is determined by a deposit of vectors in a higher-dimensional space. The use of SVM
classifier in combination with some text pre-processing techniques and kernel function
assortment is illustrated by some experiments, conducted in a limited datasets obtained
from the different frequent classes of the datasets. In this experiment, SVM repeatedly
achieves good performance on the text analysis and categorization task. SVM elimi-
nates the need for feature space due to its ability to generalize in high-dimensional
vector space. The proposed study shows that radial basis function has the highest
tendency to classify data.
References
1. Kaushik A, Naithani S (2015) A study on sentiment analysis: methods and tools. Int J Sci
Res 4(12). ISSN (Online): 2319-7064
2. Gupta V, Lehal GS A survey of text mining techniques and applications. J Emerg
3. Duda RO, Hart PE, Stork DG (2001) Pattern classification. Wiley, New York
4. Bishop CM (2006) Pattern recognition and machine learning. Springer, New York
5. Marsland S (2009) Machine learning. An algorithmic perspective. Chapman and Hall/CRC
Press, Boca Raton
6. Mohri A, Rostamizadeh A, Talwalker A (2012) Foundations of machine learning. The MIT
Press, Cambridge
7. Tapia E, Bulacio P, Angelone L (2012) Sparse and stable gene selection with consensus
SVM-RFE. Pattern Recogn Lett 33(2):164–172
8. Ruiz R, Riquelme JC, Aguilar-Ruiz JS (2006) Incremental wrapper-based gene selection
from microarray data for cancer classification. Pattern Recogn 39(12):2383–2392
9. Huerta R, Vembu S, Muezzinoglu MK, Vergara A (2012) Dynamical svm for time series
classification. Pattern recognition. Springer, Berlin, Germany, pp 216–225
10. Fatima S, Srinivasu B (2017) Text document categorization using support vector machine.
Int Res J Eng Technol (IRJET) 4(2):141–147
11. Hui JLO, Hoon GK, Zainon WMNW (2017) Effects of word class and text position in
sentiment-based news classification. Procedia Comput Sci 124:77–85
12. Jadon E, Sharma R (2017) Data mining: document classification using Naive Bayes
classifier. Int J Comput Appl 167(6):13–16
13. Devika MD, Sunitha C, Ganesh A (2016) Sentiment analysis: a comparative study on
different approaches. Procedia Comput Sci 87:44–49
14. Arundthati A (2017) Assessment of decision tree algorithm on student’s recital. Int Res J
Eng Technol (IRJET) 4
15. Purohit A, Atre D, Jaswani P, Asawara P (2015) Text classification in data mining. Int J Sci
Res Publ 5(6):1–7
16. Tilve AKS, Jain SN (2017) A survey on machine learning techniques for text classification.
Int J Eng Sci Res, Technol
17. Kumar R, Verma R (2012) Classification algorithms for data mining: a survey. Int J
Innovations Eng Technol (IJIET) 1(2):7–14
18. Vaghela VB, Jadav BM, Scholar ME (2016) Analysis of various sentiment classification
techniques. Int J Comput Appl 140(3):0975–8887
19. Kharde V, Sonawane P (2016) Sentiment analysis of twitter data: a survey of techniques.
arXiv preprint arXiv:1601.06971
20. Kalarikkal S, Remya PC (2015, June) Sentiment analysis and dataset collection: a
comparative study. In: 2015 IEEE International Advance Computing Conference (IACC).
IEEE, pp 519–524
21. Das TK, Acharjya DP, Patra MR (2014, January) Opinion mining about a product by
analyzing public tweets in Twitter. In: 2014 international conference on computer
communication and informatics. IEEE, pp 1–4
22. Shrivatava A, Mayor S, Pant B (2014) Opinion mining of real time twitter tweets. Int J
Comput Appl 100(19)
23. Chavan GS, Manjare S, Hegde P, Sankhe A (2014) A survey of various machine learning
techniques for text classification. Int J Eng Trends Technol 15(6)
24. Medhat W, Hassan A, Korashy H (2014) Sentiment analysis algorithms and applications: a
survey. Ain Shams Eng J 5(4):1093–1113
25. Senthilkumar D, Paulraj S (2015, March) Prediction of low birth weight infants and its risk
factors using data mining techniques. In: Proceedings of the 2015 international conference
on industrial engineering and operations management, pp 186–194
26. Moraes R, Valiati JF, Neto WPG (2013) Document-level sentiment classification: an
empirical comparison between SVM and ANN. Expert Syst Appl 40(2):621–633
27. Kang H, Yoo SJ, Han D (2012) Senti-lexicon and improved Naïve Bayes algorithms for
sentiment analysis of restaurant reviews. Expert Syst Appl 39(5):6000–6010
28. Abbasi A, Chen H, Salem A (2008) Sentiment analysis in multiple languages: feature
selection for opinion classification in web forums. ACM Trans Inf Syst (TOIS) 26(3):12
29. Terachi M, Saga R, Tsuji H (2006, October) Trends recognition in journal papers by text
mining. In 2006 IEEE international conference on systems, man and cybernetics, vol 6,
pp 4784–4789
High Gain Slot Antenna by Using Artificial
Magnetic Conductor
Prakash Kumar Panda1(&) and Debalina Ghosh2

1
Department of ECE, ITER, SOA Deemed to be University, Bhubaneswar,
India
prakashpanda@soa.ac.in
2
School of Electrical Sciences, IIT Bhubaneswar, Bhubaneswar, India
deghosh@iitbbs.ac.in
1 Introduction
With the increase demand in communication technology, requirement of efficient

antenna system has also been increased. The need of high gain and high directive
antennas is highly essential in satellite communication and personal communication
system. Besides these features, antennas also need to maintain low profile behaviour
[1]. In the evolution of research on increasing gain and directivity of planar monopole
or dipole antenna, metamaterial plays an important role in transforming the bidirec-
tional radiation pattern into directional radiation pattern.
Metamaterial, an artificial periodic structure, is formed by placing metallic patches
above a grounded dielectric slab. When this structure reflects an incident wave with a
0° refection phase, then it is referred to as artificial magnetic conductor (AMC), and
basically the phase reflected by an AMC surface varies with frequency. This surface
behaves like a reflector when placed behind a planar monopole or patch antenna [2].
The radiation pattern of a monopole or dipole antenna is omni-directional in nature. If
such antennas are supported by a PEC ground plane or a reflector, then it radiates in a
single direction [3, 4]. But the antenna must be k/4 distance apart from the ground
plane for avoiding antenna impedance mismatch. To overcome this problem, AMCs are
used in such antennas for increasing bandwidth, gain and directivity instead of PEC
ground plane [5–7].
In this work, a planar slot antenna having omni-directional radiation pattern has
been placed above the AMC surface for obtaining directional pattern. The gain of the
slot antenna is restricted to only 2–3 dBi due to omni-directional pattern. By trans-
forming this radiation pattern to directional, the gain is increased to 8.38 dBi. Also, the
impedance bandwidth of the antenna does not change due to the AMC surface.
The AMC surface benefits to yield an in-phase image current in the distinct locality of
the antenna. But conversely, PEC creates a current which is reversed with the antenna
current. Section 2 describes the design techniques of AMC surface. Section 3 suggests
the slot antenna and designs its combination with AMC surface. Section 4 describes the
simulation results, and Sect. 5 concludes the overall research work.

https://doi.org/10.1007/978-981-15-2774-6_34
High Gain Slot Antenna by Using Artificial Magnetic Conductor 273
2 Design Methodology of Artificial Magnetic Conductor
The reflection phase characteristics and the boundary condition applied to the unit cell
are depicted in Fig. 1a, b, respectively. FR4 material having relative permittivity (ɛr) of
4.4 and tangent loss (tan d) of 0.02 is chosen as the substrate for the unit cell. Square-
type unit cell is considered for wide bandwidth and easy manufacturing [8]. The
substrate’s length and width are 19.3 mm, and the length and width of the patch are
18.76 mm. The ground plane size is similar to the unit cell substrate. The unit cell
substrate height is considered as 3.2 mm. Periodic boundary condition is given on the
unit cell structure, and a normal plane wave is incident on it to find the phase char-
acteristics. The reflecting phase bandwidth of the unit cell is found to be 2.3–2.54 GHz
(bandwidth in the range of ±90°) [8, 9], and it crosses 0° at 2.39 GHz.
Fig. 1. a Phase diagram. b Boundary condition applied in unit cell
3 Design of Slot Antenna and Its Integration

with the AMC Surface
Figure 2 shows the geometrical configuration of a coplanar folded slot antenna. It is

based on RO4003C having a 0.8 mm thick and occupies an area of 44.3 55 mm2.
The relative permittivity and the loss tangent of RO4003C are 3.38 and 0.0027,
respectively. Coplanar waveguide with 50Ω characteristic impedance is fed to the
antenna. The resonating frequency of the antenna is adjusted by the length of the stub
added in the slot or aperture created on the radiating patch. After a simulation study on
varying the stub length and finally choosing its length b = 35.6 mm, the antenna
resonates at a centre frequency of 2.4 GHz (Fig. 3).
The CPW-fed slot antenna is employed on an AMC consisting of 4 4 unit cells.
The separation distance among the slot antenna and the artificial magnetic conductor is
maintained to be 10 mm for proper impedance matching of the antenna. If the antenna
will be placed very close to the AMC surface, then due to the high electromagnetic
interference among antenna and AMC surface mismatch the antenna input impedance.
So after a parametric analysis, this separation distance is found to be 10 mm for proper
impedance matching and a suitable gain pattern. The simulated return loss parameter of
the slot antenna and the integration with artificial magnetic conductor is shown in
274 P. K. Panda and D. Ghosh
Fig. 2. Geometrical dimensions of slot antenna. [LS = 44.3, WS = 55, a = 47, b = 35.6,
c = 0.35, d = 2.06, e = 0.21, f = 3.1, Z = 0.8] (all dimensions in mm)
Fig. 3. Slot antenna over AMC surface: a top view; b side view with h1 = 3.2 mm and
h2 = 0.8 mm
Fig. 4. Return loss parameter of slot antenna and the integrated structure
Fig. 4. The S parameter results between the slot antenna and its integration with the
AMC surface look almost similar. The slight change in the result is due to the mutual
coupling effect between the antenna and AMC surface.
The simulated results for gain and directivity of the slot antenna along with the
composite antenna are displayed in Figs. 5 and 6 correspondingly. The following
results are taken at the resonating frequency of 2.4 GHz. After integrating the antenna
with the artificial magnetic conductor, the gain and directivity of the composite antenna
have been increased by 4.25 dB and 4.99 dBi, respectively. Also, the simulated results
verify that the bidirectional radiation is converted into directional after adding the
antenna with the AMC surface. Additionally, the integrated antenna is characterized in
the form of radiation pattern in elevation and horizontal plane. The elevation and
horizontal plane of the slot antenna and the composite antenna at the resonating fre-
quency of 2.4 GHz are presented in Fig. 7.
Fig. 5. Gain of the antenna at 2.4 GHz: a slot antenna, b slot antenna with AMC surfaces
276 P. K. Panda and D. Ghosh
Fig. 6. Directivity of the antenna at 2.4 GHz: a slot antenna, b slot antenna with AMC surfaces
Fig. 7. a E-plane radiation pattern. b H-plane radiation pattern
5 Conclusions
This work proposes a slot antenna combined with artificial magnetic conductor surface
by which bidirectional radiation behaviour of the slot antenna has been transformed
into directional radiation pattern. This phenomenon also provides very low profile
behaviour of the antenna and confirms an increased gain and directivity of 8.38 dB and
9.5 dBi at the resonant frequency, i.e. 2.4 GHz. Overall, this analysis concludes that
slot antenna with AMC surface is good for wireless system (at ISM band) for its high
gain and directivity. Further, the antenna possesses low profile characteristics.
References
1. Zaker R, Abdipour A (2010) A very compact ultrawide band printed omnidirectional
monopole antenna. IEEE Antennas Wirel Propag Lett 9:471–473
2. Sievenpiper D, Zhang L, Broas RFJ, Alexopolous NG, Yablonovitch E (1999) High
impedance electromagnetic surfaces with a forbidden frequency band. IEEE Trans Microwave
Theor Tech 47(11):2059–2074
3. Kraus JD (1940) The corner—reflector antenna. Proc IRE 28(11):513–519
4. Uda S, Mushiake Y (1954) Yagi-Uda Antenna. The Research Institute of Electrical
Communication, Japan, Tohoki University
5. Zhang Y, Hagen JV, Younis M, Fischer C, Wiesbeck W (2003) Planar artificial magnetic
conductors and patch antennas. IEEE Trans Antennas Propag 51(10):581–583
6. Panda PK, Ghosh D (2018) Isolation and gain enhancement of patch antennas using EMNZ
superstrate. AEU-Int J Electron Commun 86:164–170
7. Panda PK, Ghosh D (2019) Wideband bow-tie antenna with increased gain and directivity by
using high impedance surface. Int J RF Microwave Comput Aided Eng 29(3):1–9
8. Foroozesh A, Shafai L (2011) Investigation into the application of artificial magnetic
conductors to bandwidth broadening, gain enhancement and beam shaping of low profile and
conventional monopole antennas. IEEE Trans Antennas Propag 59(1):04–19
9. Aminian A, Yang F, Rahmat-Samii Y (2003) In-phase reflection and EM wave suppression
characteristics of electromagnetic band gap ground planes. In: Proceedings of the IEEE
antennas and propagation society international symposium, vol 4, pp 430–433, June 2003
Electrical Energy Storage: A Great Business
Ahead
Renu Sharma(&), Satish Choudhury, Bibekananda Jena,

and Sonali Goel
Department of Electrical Engineering, ITER Siksha ‘O’ Anusandhan (Deemed to

{renusharma,satishchoudhury,
bibekanandajena}@soa.ac.in
Abbreviations
GW Giga watt
MW Mega watt
kW Kilo watt
UPS Uninterruptible power supply
LCOS Levelized cost of storage
ISOs Independent system operators
BESS Battery energy storage system
VRF Vanadium redox flow
PCS Power control system
BMS Battery management system
PER Primary frequency response
AGC Automatic generation control
FFR Fast frequency response
PFR Primary frequency response
BTW Behind-the-meter
PV Photovoltaic
MISO Midcontinent independent system operator
CAISO California Independent System Operator Corporation
BESS Battery energy storage systems
RegD Dynamic regulation signal
RegA Traditional regulation signal
FERC Federal Energy Regulatory Commission
PHS Pumped hydro storage
DOD Depth of discharge
1 Introduction
Energy storage is one of the fastest growing areas in the Indian utility industry and
rapidly becoming the mainstream. About 202 GW of grid-connected storage systems
are available globally of which 135 GW is pumped hydro and 65 GW is UPS systems

https://doi.org/10.1007/978-981-15-2774-6_35
Electrical Energy Storage: A Great Business Ahead 279
and rest about 2677 MW being new storage technologies [1]. Policymakers are using
mandates to encourage storage investments. National Mission on Transformative
Mobility and Battery Storage approved by Cabinet Mission have approved for setting
up a National Mission on Transformative Mobility and Battery Storage for driving
sustainable and holistic mobility initiatives. The Phased Manufacturing Program valid
for 5 years till 2024 to support setting up of a few large-scale, export-competitive
integrated batteries and cell-manufacturing giga plants in India for confine production
of the entire electric vehicle value chain. Several new application areas for batteries
have opened up that require large-scale batteries for storage and grid balancing
applications. Flow batteries are suitable for energy storage applications requiring long
hours of backup, ranging from a few hours to several days. As power (in kW) and
energy capacity (in kWh) are separated, the battery capacity (or discharge time) can
simply be enlarged by filling more liquid into the tanks and the power capacity can be
increased by increasing the battery stacks. Flow batteries usually offer long life and
stable capacity, but low energy density makes them more suitable for use in stationary
applications.
A large number of battery technologies are available in the market. The conven-
tional battery technology like the lead–acid battery is still good for residential and
industrial applications. The advantages in this type of battery is the low cost ($100–
200/kWh), it has no major safety issues and can deep discharge. The drawbacks, on the
other hand, are a short life (300–1000 cycles, i.e., 2–3 years of life), low specific
energy (35–50 Wh/kg), limited depth of discharge (up to 50%), thermal runway
(performance and life degrades as battery heats up and as temperature changes), and
disposal issues, as lead is toxic. The selection of batteries depends entirely upon the
application and one should consider various factors such as capital cost, LCOS, life-
time, efficiency, depth of discharge, footprint, etc., for selecting the right battery
technology. The Li-ion battery is a market leader for its portable applications and the
electric vehicle. Lead–carbon batteries are an improved version of lead–acid batteries
and provide long life and fast charging/discharging Li-ion batteries provide higher
energy density (80–240 Wh/kg) are compact in size, fast charging, and hence are best
for portable applications and for electric vehicles. However, Li-ion and lead–acid
batteries are designed to discharge for 3–4 h only. For storage applications requiring
long hours of backup, Li-ion and lead–acid batteries are not suitable. Further, safety
issues like fire hazard associated with the Li-ion batteries are still a concern over their
use in large-scale storage. Flow batteries are suitable for large-scale energy storage
applications requiring long hours of backup and heavy-duty cycles. Among them,
vanadium redox battery is one of the suitable candidates for renewable integration.
Vanadium redox flow (VRF) battery is one of the proven and commercially
available flow battery technologies ideals for applications in micro-grids, grid bal-
ancing, electric vehicle charging stations, etc., worldwide, about 1 GWh of VRF
batteries has been installed and is in operation. Since vanadium at different oxidation
states is stored in two separate tanks, there is no detrimental effect of cross mixing of
the electrolyte. The battery stacks are made of a graphic electrode and polymer
membranes, and no components are consumed or degraded during the operation.
Therefore, the lifetime of the battery is more than 20 years (>10,000 cycles) and
provides a stable performance (no capacity decay) over its lifetime price wise, VRF
280 R. Sharma et al.
batteries are available $1000/kWh for small to medium systems and $500/kWh for
MWh level systems. Due to the long life of VRF batteries, the leveled cost of energy
storage is the cheapest compared with other battery types. The major supply of
vanadium currently comes from China, and due to the recent rebar policy in China the
price of vanadium is a commodity and can be recovered 100% from the used battery, a
leading model is being developed to eliminate the upfront cost burden to developers
and users.
Zinc bromine hybrid flow batteries are emerging as a suitable technology for res-
idential, telecommunication, and other energy storage markets. This battery is good for
deep discharging and provides stable performance for 10 years. However, the use of
halide chemistry and the need for routine maintenance of this battery are the draw-
backs. In large-scale energy storage applications, technology requiring frequent and
long duration maintenance also adds extra cost and necessitates the redundancy plan. In
terms of price, zinc–bromine batteries are available at $1000/kWh.
Iron chromium and all-iron flow batteries are other types of flow batteries available
at low cost, but they have limited installations and their long-term performance in the
field has to be evaluated. Sodium sulfur batteries are also available for large-scale
energy storage applications. They have a high energy density, high roundtrip efficiency,
and long life, but they need high temperature (>300 °C) for operation and have a safety
risk associated with sodium. Rechargeable zinc–air batteries are also emerging as a low
cost option at 4100/kWh battery for energy storage with long hours of discharge
capability, high energy density (450 Wh/l), and good performance of over 10,000
discharge hours. The drawbacks of this battery are low-power density, relatively short
life, low roundtrip efficiency, and the requirement of preventive maintenance. The
overall analysis about different type of batteries is given in Fig. 1.
Fig. 1. Performance analysis of different type of batteries in terms of cost, life span, and DOD
In terms of cost of energy storage, one should consider not only the battery or
module cost but also the cost for power control system (PCS), battery management
system (BMS), balance of the plant, fire suppression system, air-conditioning, etc.
2 Storage System
Storage can provide more value to a system or to infrastructure that is highly stressed.
For example, systems that are running up against minimum generation constraints on
thermal units and curtailing wind and solar, or distribution feeders that are hitting
thermal overloads, or island systems where the rate of change of frequency can be very
high, may all get more value from storage than systems that are less stressed. Storage
can provide system flexibility, fast reserves, etc., but the value of these is greater in
very constrained systems. Other infrastructure and/or modifications to planning and
operations may also provide the same benefits that storage provides, so it is also a
matter of least-cost mitigation options. Basic application of storage system is shown in
Fig. 2. It shows the interconnection between different sources of electricity and the
existing utility grid [2]. Out of many sources shown in the figure, energy storage
device, i.e., battery is one of the vital components to provide bridge between source and
the sink [3]. It also plays a key role at the time of power outage from different existing
sources and in the remote places where electricity is unreachable.
Fig. 2. Interconnection of different sources of electricity with utility grid

2.1 Solar and Wind Integration

The variability and uncertainty of solar and wind create challenges for utilities and
ISOs that need to maintain system reliability. A large number of potential options have
been discussed in literature for mitigating the impacts of solar and wind on the grid [4,
5]. Storage has been considered an option, but cost recovery for storage to integrate
wind is that the high forecast errors (over—or under—forecasts) [6]. To take advantage
of that price spread, storage needed to better forecasts than the wind forecasts.
Storage can provide a better match for solar integration than for wind integration
[7]. First, solar is much more predictable and storage owners easily know when to
charge and discharge, unlike with wind. Second, solar integration is much more about
time-shifting of large amounts of energy because the midday solar peak is so pro-
nounced, whereas there tends to be some wind at all hours of the day, on average.
2.2 Where Is Storage Most Effective?

Storage can be implemented at various levels on the power system:
• Bulk power system/transmission level—providing energy arbitrage, avoiding
transmission congestion or losses
• Distribution feeder level—in addition to the above, storage can also provide more
location-specific services such as deferring distribution upgrades.
• Behind-the-meter (BTM)—some of the above benefits may be accessible to BTM
storage. Storage can also provide services such as reducing demand. Charges or
managing energy bills with time-of-use rates.
The further downstream storage is, the more benefits it can theoretically provide
[8]. However, the further downstream storage is, the less accurately the
markets/incentives/rates reflect the physics of power flows. For example, the volu-
metric and bundled rates for most residential customers reflect very little about the
actual cost of service at any given time. In another example, the rates for commercial
customers may include a demand charge or a time-of-use charge but still are likely to
ignore the impacts of that commercial customer on ancillary services.
Additionally, the further downstream storage is, the less it may be able to monetize
the various value streams. BTM customers are unlikely to have a large enough system,
and the telemetry and control sophistication to participate in energy, capacity or
ancillary service markets. For example, while a BTM system may defer upgrades of the
distribution system, it cannot monetize this value.
2.3 Ownership
Ownership of the storage asset constrains the applications that the storage unit pro-
vides. On the bulk power system level, there may be issues as to whether storage is
treated as a generator or as a load. As a load, it fits into demand-side management
markets but perhaps not when it acts as a generator.
BTM systems can bring qualitative benefits that are not economic. For example,
consumers that already own distributed PV systems may wish to add storage to
increase their usage of self-generated electricity, thus increasing their self-sufficiency.

While these benefits may not make economic sense, they can be important drivers in
decision-making.
BTM systems are most likely to be under-utilized. Part of this is the lack of
sophistication of the end-users; part is the lack of access to the markets that would
allow the owners to better utilize the systems. For example, in providing demand
charge reduction, there will be many hours when the unit could be providing regulation
and still have time to charge to prepare for offsetting the demand charge. This means
there are opportunities for third parties to step in, aggregate these systems and install
the telemetry and controls needed to bid services into the markets.
2.4 Market Rules

Some market rules limit the ability of storage to fully monetize its benefits. For
example, in MISO, storage resources that have one hour or less duration can provide
regulation, but if it does provide regulation, it cannot simultaneously provide spin,
supplemental reserve or ramping reserve. That limits the value stream that storage can
access. In CAISO, storage is treated as a generator that produces both positive and
negative energy.
2.5 Co-location with Solar and Wind

Many utility-scale battery projects are co-located with new or existing wind and solar
plants. This allows the battery to leverage significant interconnection infrastructure.
Additionally, storage paired with solar or wind plants can be eligible for federal
investment tax credits, state incentives, and/or subsidized loans. Despite co-location
with a wind plant and capability of a BESS may be operated independently of the wind
plant. In PJM, this is frequently done to prioritize BESS use the highly valued fast
regulation service.
3 Energy Arbitrage
Even though energy arbitrage (changing when prices are low and discharging when
prices are high) is the main application that most people consider for energy storage, it
is difficult to make a business case for new storage with arbitrage as the sole appli-
cation. First, the current spate of low natural gas prices does not drive frequently high
price spreads that drive arbitrage opportunities. Additionally, price forecast errors tend
to be worse when arbitrage opportunities are best (prices are volatile), making bidding
less certain.
As mentioned earlier, storage in energy arbitrage applications should be able to
provide high value to grids with high solar penetrations. Unlike wind, the predictability
of solar means that high solar penetrations will drive much more predictable price
signals. The negative prices in the day-ahead market due to the very high levels of solar
are an indication of this predictability.
A final issue for storage is determining when to bid on as a generator and when to
bid on as a load. Because storage owners do not have the full information that the ISO
has, e.g., transmission congestion, it is difficult for storage owners to know how to bid
the storage and storage owners may attempt to schedule the storage in a way that is not
beneficial to the grid. Incorporating storage scheduling into the ISO’s security-
constrained unit commitment in ISOs that allow this, such as PJM, can essentially give
the ISO control over the resource to optimize operation.
Storage is one of many options for flexibility on the grid. Other option also holds
promise for rapid adoption that provides time-shifting of the load. Dynamic pricing is
being considered as rate restructuring is debated in many states, partly driven by
shortcomings of net metering and partly driven by advanced metering capabilities. To
some extent, dynamic pricing will compete with storage to provide the same arbitrage
value to the system. Essentially, dynamic pricing could drive arbitrage storage out of
the bulk power system and into the BTM market segment.
Another emerging technology option is transportation electrification. Electric
vehicles (EVs) with vehicle to grid (V2G) technology essentially look like storage with
limited runtimes. But even without V2G technology, EVs provide one part of the
storage arbitrage value, that of providing load, and if EV charging is controlled, this
may reduce curtailment of excess wind or solar.
4 Frequency Regulation
One of the most promising applications for storage technologies is frequency regula-
tion. Frequency regulation (also known as secondary reserves or automatic generation
control) tends to be the highest value reserve service. While conventional storage
technologies such as pumped hydro storage, fit well in this application, there is a “fast”
frequency regulation application for which new storage technologies (such as batteries
and flywheels) are ideal.
4.1 Tuning the Fast Regulation Signal

In 2009, PJM had collaborated with AES on the use of lithium battery storage to
provide higher performance regulation. At first, they tried to use a fast ramping
hydrogenation-type signal. However, lithium-ion batteries tend to have a limited
energy capacity, typically 4:1 power to energy ratio. The limited energy capacity was
not compatible with this signal, which had tended to push the battery to its top or
bottom limit and pin it there. PJM designed the dynamic control signal (RegD) with
this type of limited energy capacity in mind. The target time for the RegD signal to
converge to zero was 5 min and 95% of the time, it converges to zero in less than
15 min [9].
4.2 Performance-Based Regulation

In 2011, The Federal Energy Regulatory Commission (FERC) Order 755 directed
system operators to “pay for performance” for providing frequency regulation and grid
stabilization services. Under this order, resources that can more accurately follow an
operator’s regulation signal should be rewarded. PJM was the first ISO in the USA to
implement performance-based regulation. Regulation resources are compensated for
capability (reserve MWs), performance (mileage), and lost opportunity cost ($/MW of
lost revenue in the 5-min interval). Performance of resources determines whether a
resource can bid on the traditional regulation signal (RegA) or the faster RegD signal.
Additionally, fast regulation provides greater benefits: PJM has found that the first MW
of RegD offsets 3 MW of RegA, but that this benefit declines as more resources
provide RegD, and that above 60% fast resources, there is no additional benefit [10].
Performance-based regulation has enabled PJM to reduce regulatory requirements.
Previously, PJM required 1% of the on or off-peak load forecast, depending on the time
of day. After introducing RegD, the total regulation requirement ratcheted down over
subsequent months to its current level of 0.7%. PJM’s control performance standard
metrics demonstrate a continued performance in regulating area control error.
4.3 Impact on Prices

Regulation prices increased from $13.75/MWh prior to performance-based regulation
to $38.75/MWh when RegD was introduced partly because the old framework did not
include lost opportunity cost as part of the compensation for regulation. If this lost
opportunity cost is included, then PJM’s overall regulation cost remains similar to its
cost prior to performance-based regulation.
4.4 Storage Market Impacts

PJM’s performance-based regulation, including the fine-tuning of the RegD signal for
the limitations of battery storage, has made PJM the leading battery storage market in
the USA. Battery storage projects in PJM’s interconnection queue went from 9 MW in
2012 to 534 MW in 2015. The share of regulation capability provided by batteries
increased from 16% in 2014 to 28% in 2015. At the same time, the share of regulation
credits earned by batteries increased from 12 to 21%. There is a concern; however, that
the RegD market is saturated because RegD is over-procured due to an inconsistency in
how RegD substitutes for RegA in the market design [11].
5 Demand Charge and Bill Management
The fastest growing sector for energy storage is the BTM sector which grew from
6 MW in 2014 to 35 MW in 2015. The ability of BTM storage, especially BESSs, to
reduce demand charges and manage energy bills, has driven the commercial sector to
be the fastest growing storage segment in the USA over the last two years [12].
This application is heavily dependent on the characteristics of the particular load

and the rate structure. Finding storage technologies that fit the needs of the load and
rates are also critical.
Energy storage systems for demand charge reduction and bill management may
have relatively low utilization factors. For example, if the system’s primary application
is to help meet energy use during expensive afternoon and evening hours, the system
may have all night and morning (far more hours than necessary) to ensure a full charge.
There is thus an opportunity for this system to provide other grid benefits during that
downtime. However, these systems tend to be too small to qualify to be in the market
and the technical and financial hurdles to participate in the market tend to be too great.
6 Fast Frequency Response
New storage technologies such as batteries, flywheels, and supercapacitors, may be

capable of providing a very fast response to frequency deviations. With increased
penetrations of non-synchronous generators and decreased levels of synchronous
generators that provide inertia, the system may have more challenges maintaining
frequency stability. If a large generator trip offline, a power system with lower inertia
will have a faster drop in frequency. To recover the system, frequency responsive
generation is needed to increase output. Conventional generators providing primary
frequency response are relatively slow to respond (several seconds), and may be too
late to help a system that has very low inertia. The FFR is a new service in some
regions (required or likely to be required in Ireland, UK, ERCOT) that can provide a
very fast response (less than one second) to help recover frequency. Additionally,
simultaneous have shown that in some system conditions (low inertia), 1 MW of FFR
can provide the same level of support to the system as twice the amount of conven-
tional primary frequency response (PFR) [10]. This means that if some of the frequency
response is FFR, the total requirement for frequency response could be lower.
While it is too soon to know what prices will be for this service and how large of a
market there will be, there are relatively few resources that can provide this service,
depending on requirements for response speed, and so storage will likely have a large
role to play.
7 Life Cycle Costing
A cost analysis [13] is made to explain how the storage device is economical as
compared to the existing sources and how it leads to great business ahead. A small
community has taken into consideration for the life cycle costing analysis. Let 100
numbers of people are staying in the community and they are accessing electricity from
a hydel plant. Let the capital cost of consumption of electricity in hydel plant is about
`50,000 + 5000 * l, where it includes installation cost, transmission cost, cost of the
line, and losses. ‘l’ denotes the length of the line in km. For lifetime of 25 years and
taking a present worth factor of 7.84, the annual life cycle cost (ALCC) of hydel system
is calculated in (1).
50000 þ 5000 l
ALCChydel ¼ 6378 þ 636 l ð1Þ
7:86
Similarly, if the same community wants to install a solar PV system with energy
storage device it costs approximately 200 per watt, where the availability of solar
insolation is expected to be 5 kWh/m2/day. For ‘x’ kW of consumption, the capital cost
of solar PV system will be 200x. So, taking a life span of 25 years and present worth of
7.84 the ALCC of PV system is calculated in (2).
200x
ALCCpv ¼ 25:5x ð2Þ
7:84
Annual energy generation in hydel plant is calculated in (3).
Ehydel ¼ 1 kW 24 h 365 days g ¼ 1 24 365 0:3 ¼ 2628 kWh ð3Þ
ALCChydel 6378 þ 636 l

Cost per unit consumption ¼ ¼
Ehydel 2628 ð4Þ
¼ 2:42 þ 0:242 l
Annual energy consumption in PV system is calculated in (5).

x
Epv ¼ kW 5 kWh=m2 =day 365 days
1000 ð5Þ
x
¼ 5 365 ¼ 1:825x
1000
ALCCpv 25.5 x
Cost per unit consumption = = = 14 ð6Þ
E pv 1.825 x
To compare between two systems, breakeven point can be calculated as in (7).
2:42 þ 0:242 l ¼ 14
ð7Þ
) l 48 km
It can be concluded from the above analysis that if your community is about 48 km
distance from the hydel system then it is more beneficial to use such sources of
electricity, whereas if your location is more than 48 km from hydel station then it is
more economical to use PV system with battery storage facility. In this paper, the cases
of remote area where electricity is unreachable and the area, where the power outage is
more frequent are taken into consideration. The analysis shows the storage system (PV
along with battery) can be better replacement for the sources of electricity for such type
of community and the business can be more propagated in such commodity.
8 Conclusion
Storage has the potential to provide a large number of values to the power system.
However, different actors can access and monetize different portions of that overall
value stream, and even that varies depending on the ISO or jurisdiction. Existing PHS
that was installed in the past years to integrate the nuclear fleet has earned revenue by
stacking these benefits for years, but it is difficult to recover costs on new utility-scale
storage by stacking benefits, at least in the current low natural gas price paradigm.
The main commercial markets for storage are utility-scale BESSs offering fast
regulation reserves and BTM BESSs for demand charge/bill management. FFR is
likely to be a new niche market for utility-scale BESSs. As rules evolve, to allow
aggregators into the market and to allow different market participants to access more of
the value stream, storage growth should accelerate. And as rates are restructured to
send more granular signals to end-users, dynamic pricing is likely to drive much more
BTM storage down to the residential level.
Acknowledgements. The assistance provided by the authorities of Siksha ‘O’ Anusandhan

(Deemed to be University), Bhubaneswar are highly acknowledged.
References
1. Smart grid handbook for regulators and policy makers. Nov 2017
2. Vazquez S, Lukic SM, Galvan E, Franquelo LG, Carrasco JM (2010) Energy storage
systems for transport and grid applications. IEEE Trans Ind Electron 57(12):3881–3895
3. Kalaiselvam S, Parameshwaran R (2014) Thermal energy storage technologies for
sustainability: systems design, assessment and applications. Elsevier
4. GE Energy (2010) Western wind and solar integration study. NREL/SR
5. GE Energy Consulting (2012) Hawaii solar integration study
6. Gharehpetian GB, Agah SM (eds) (2017) Distributed generation systems: design, operation
and grid integration. Butterworth-Heinemann
7. Mills A, Wiser R (2014) Strategies for mitigating the reduction in economic value of variable
generation with increasing penetration levels. LBNL-6590E
8. Fitzgerald G, Mandel J, Morris J, Touati H (2015) The economics of battery energy storage.
RMI, Boulder, CO
9. Benner S, PJM (2015) A brief history of regulation signals at PJM
10. PJM (2013) Performance based regulation: year one analysis
11. Monitoring Analytics (2016) LIC, state of the market report for PJM January through June,
Aug. 11, 2016
12. GTM Research/Energy Storage Association (2016) U.S. energy storage monitor 2015 year
in review
13. Zakeri B, Syri S (2015) Electrical energy storage systems: a comparative life cycle cost
analysis. Renew Sust Energy Rev 1(42):569–596
Embedded Entropy-Based Registration of CT
and MR Images
Sunita Samant, Subhaluxmi Sahoo(&), and Pradipta Kumar Nanda
Image & Video Analysis Laboratory, Department of ECE,

S O A, Deemed to be University, Bhubaneswar, India
1 Introduction
Multimodal image registration is a methodology of spatial mapping between different

images acquired by different diagnostics processes. Multimodal image registration is
among the most challenging problems in medical imaging. In recent years, different
registration methods have been proposed. One of the methodologies is the measure
similarity in image registration is maximizing the MI. Registration of multimodal
images using MI is proposed by Maes et al. [1] and for registration, where no
assumptions were made and no boundary constraints were imposed. Subsequently,
Chen and Varshney [2] and Juni et al. [3] have addressed the artifacts using the
interpolation method. Woo et al. [4] have proposed a new registration technique
incorporating spatial and geometric information to high resolution images. Mohanalin
et al. [5] have proposed an innovative idea to increase the fidelity of multimodal image
registration using Tsallis entropy. Khadar and Hamza [6] have proposed a Jensen–
Tsallis divergence along with histogram-based modality transformation, and Li et al.
[7] have proposed Jensen–Arimoto divergence. But, to take care of number of bins
used for the computation of mutual information, a minimum MSE histogram-based
estimation method has been proposed [8]. Han et al. [9] have proposed a non-rigid
image registration technique that is based on a weighted Horn-type regularizer. Another
registration approach that has been used in [10] involves finding structural information
of the images to register. In our work, we have proposed a new approach for registering
noisy images. Our registration scheme for noisy images has been developed on the
notion of mutual information to find out the optimal registration parameter. The 2D
joint histogram of the noisy MR and CT image is assumed to be degraded which would
result uncertainty in finding mutual information and optimal registration parameter. So
instead of registering the noisy images, we have created embedded images for regis-
tration process to take care of noisy situation. We have considered noisy MR-T1 image
and CT images from retrospective image registration evaluation (RIRE) database and
noisy T2 and T1 from brain web database. MR image is considered to be noisy one, so
Gaussian kernel filter is applied to smooth out the noise component. We have obtained
structural representations of the images by calculating embedded entropy of the images
and then input images are replaced by the embedded entropy images for the registra-
tion. Then, maximization of mutual information determines the registration parameter.
The proposed method has been tested for 5, 10, 20 dB SNRs of MR-T1 (RIRE) and
MR-T2 (brain web) data set.
https://doi.org/10.1007/978-981-15-2774-6_36
290 S. Samant et al.
2 Background on Entropy-Based Image Registration
Viola [11] has introduced a notion of image registration based on mutual information
(MI). The motive of registration is to compute a transformation that will align two or
more images captured with different techniques at different time. Consider two images
If and Ir as the floating and reference images, respectively. The registration of these
images is then formulated using mutual information. The estimation of the transfor-
mation (T) which will align the floating image F with reference image R is obtained by
maximizing the mutual information over transformation T,
T ¼ arg max MIðF; RÞ ð1Þ
MI of two random variables F and R with marginal probability and joint probability
mass function is defined by
pFR ðf ; r Þ
MIðF; RÞ ¼ pFR ðf ; r Þlog2 ð2Þ
pF ð f ÞpR ðr Þ
MI measures the degree of interdependence between F and R in terms of the

distance measured between joint distribution PF,R (f, r) and distribution associated with
PF(f) and PR(r). Entropy can be interpreted as a measure of uncertainty, variability or
complexity. The MI related to entropies is stated as
MIðF; RÞ ¼ H ðF Þ þ H ðRÞ H ðF; RÞ ð3Þ
where H(F), H(R) and H(F, R) are the marginal and joint entropies of the random
variable F and R, respectively.
3 Embedded Entropy Images
The similarity between images can be interpreted as whether intensity changes

occurring at similar locations. So, the intensity changes or the bound for a lossless
compression in an image can be quantified by the entropy measures.
Here, we have introduced a new structural representation of the image by
embedding Arimoto entropy and Tsallis entropy. In the following, Arimoto and Tsallis
entropies are defined.
3.1 Arimoto Entropy

For any discrete random variable X(x1, x2,…, xN) with its distribution function
P(p1, p2, …, pN), i.e., Arimoto entropy is given by
Embedded Entropy-Based Registration of CT and MR Images 291
2 !1=q 3
q 4 XX
HA ð X Þ ¼ 1 pq 5 ð4Þ
q1 i2X j2X
ij
where ‘q’ is the Arimoto’s parameter. pi,j is the joint probability of the current pixel with
the corresponding neighborhood. Arimoto entropy can be viewed as generalization of
Shannon’s entropy. Arimoto entropy is pseudo-additivity and non-negative in nature.
3.2 Tsallis Entropy

For any discrete random variable X(x1, x2,…, xN) with its probability distribution P(p1,
p2,…, pN), i.e., Tsallis entropy is given by
" #
q XX q
HT ð X Þ ¼ 1 pij ð5Þ
q1 i2X j2X
It is extended version of Shannon’s entropy and generalization of Boltzmann–

Gibb’s entropy. It contains a ‘q’ parameter which is a real and represents the system’s
non-extensivity.
3.3 Embedded Entropy

The Arimoto and Tsallis parameters play a vital role for obtaining the structural
information of the image. The different values of the parameters result in sub-extensive
or super-extensive property. Therefore, we have introduced the function to embed
Arimoto and Tsallis entropy.
EEð X Þ ¼ ð1 þ qÞHT ð X Þ þ ð1 qÞHA ð X Þ ð6Þ
3.4 Embedded Entropy Image Generation

In order to generate entropy image for image registration, we have enforced a second-
order entropic measure to image. To avoid the spatial ambiguities, each region is
processed through Gaussian filter. For the original image with a size of MN, the
neighborhood of size b is extracted for each image pixel i in l. In each region, Gaussian
kernel filter is applied. Then, we have calculated the embedded entropy. This
embedded entropy gives the statistical dependency between current pixel and its
neighborhood pixels, and calculation of entropy value depends upon the number of
bins used. Higher the bin size, more finer the histogram is, which gives more infor-
mation about the image and which tends to more accurate result. But, at the same time
higher bin size also requires more storage space, which leads to increase in redundancy
in the joint histogram. So, we have limited the bin size to 10 by using global nor-
malization. The normalized image is used for the estimation of the second-order
entropy for all the pixels.
Due to local intensity changes, there is a similarity between current pixel intensity
and its neighborhood intensity. And second-order entropy technique addresses changes
in intensity in an image region locally when the second-order entropy values of these
regions are minimized. In this proposed work, we have considered MR-T1 image as
shown in Fig. 1a. The corresponding entropy images have been computed using Ari-
moto entropy, Tsallis entropy and embedded entropy as shown in Fig. 1b–d.
Fig. 1. a MR-T1 image, b–d entropy images using Arimoto entropy, Tsallis entropy and
embedded entropy, respectively
4 Embedded Entropy Images-Based Registration
Registration of images has been achieved from the mutual information of the embedded
entropy images.
Our registration process has been represented in Fig. 2. In our work, we first used
the embedded entropy (EE) operator which indicates the structural changes in the
images on floating image If and reference image Ir. Then, joint histogram of the entropy
images is computed, and then the marginal and joint probability distributions have been
calculated.
Fig. 2. Block diagram of proposed registration algorithm

Therefore, the entropy image registration is formulated as follows:
t ¼ arg max MIðEEðIR ðr ÞÞ; EEðIF ðtð xÞÞÞÞ ð7Þ
where MI is the similarity measure criteria given as
pFR ðf ; r Þ
MIðF; RÞ ¼ pFR ðf ; r Þlog2 ð8Þ
pF ð f ÞpR ðr Þ
and transformation function ‘t’ maps the correspondence of the pixels x of the floating
image IF in the reference image IR.
5 Pseudo Codes of the Algorithm
a) Calculation of embedded entropy images:

Input: If(Floating image) and Ir(Reference image), b=size of
neighborhood, standard deviation = σ = 1.0, mean = μ = 0.0
Output: Ie_f, Ie_r, IG_f = gaussian(If, sigma, mu), IG_r = gaussian(Ir , sigma,
mu), I_f = normalization (IG_f), I_r = normalization (IG_r)
for i= 1 to MN in I (both floating and reference image) do
P = neighborhood region( i,b)
e_Arimoto = second-orderArimotoEntropy(P)
e_Tsallis = second orderTsallisEntropy(P)
Ie (i) = (1+alpha) * e_Tsallis + (1-alpha) *e_Arimoto
end for
b) Calculation of MI of Ie_r & Ie_f
c) Determination of registration (α) from the computed MI
6 Results and Discussions
For the registration process, we have considered CT and MR-T1 images of RIRE
database and MR-T1 and MR-T2 images of brain web database. In order to reveal the
efficiency of our proposed method, Gaussian noise has been added to MR images at
different SNR label. SNR can be defined as SNRdB = 10log10SP/NP, where SP is
signal power and NP is noise power.
In our paper, we have considered MR-T1 as the floating image and CT image as
reference image from RIRE database for registration. MR-T1 image has been degraded
by 2, 5, 10 and 20 dB SNRs as shown in Fig. 3b–e, respectively. We have extracted the
structural images corresponding to CT and noisy MR-T1 images with 2, 5, 10 and 20
dB SNRs as shown in Fig. 3f–i, respectively. Then, the MI curve of the structural CT
and MR images with 20 dB SNR has been computed from the joint histogram, and the
MI plot is shown in Fig. 4a. Maximum value of MI is found to be −6°, and the
registration parameter alpha (rotational angle) that is estimated has been found to be at
−6°. Subsequently, noise content has been increased in MR-T1 image to have of 10
and 2 dB SNRs. The mutual information for 10 dB and 2 dB SNR has been shown in
Fig. 4b, c, respectively. The estimated registration parameter (alpha) using proposed
method is found to be −4°. For the noise-free case, this alpha (registration) parameter
has been found to be −4°. We can observe that the efficiency of our proposed method
increases with increase in noise content. We have compared our proposed method with
two existing methods as shown in Fig. 5. As we can see from the figure, the estimated
errors do not change up to 10 dB SNR. MI, SSD method diverged beyond 5 dB SNR;
whereas, the estimated error showed a converging trend beyond 5 dB SNR. It has been
found that the performance of existing methods decreases as increase in noise. But, the
proposed method gives better result with increase in noise content. We have also tested
our proposed methods with MR-T1 and MR-T2 from brain web database. We have
obtained the noisy MR-T2 images for 20, 10, 5 and 2 dB as shown in Fig. 6a–e. The
corresponding structural images of MR-T1 and noisy MR-T2 are shown in Fig. 7a–e.
The performance of our method has been found to be better than the other exiting
method for 2 and 10 dB as shown in Fig. 8a–c for 10 dB, 5 dB and 2 dB SNR,
respectively. The estimate errors of proposed method and existing methods have been
found out as shown in Fig. 9.
Fig. 3. a–e CT, noisy MR-T1 images with 2 dB, 5 dB, 10 dB and 20 dB SNR, respectively,
f–j embedded entropy image of a–e, respectively
Fig. 4. Estimation of similarity metric for CT and noisy MR-T1 with a 20 dB and b 10 dB SNR
c 2 dB SNR, respectively
Fig. 5. Error plot for registration parameter (rotational angle)

Fig. 6. a MR-T1, b–e MR-T2 with 2 dB, 5 dB, 10 dB and 20 dB SNR, respectively (brain web
database)
Fig. 7. a–e Embedded entropy images for MR-T1 and MR-T2 with 2 dB, 5 dB, 10 dB and
20 dB SNR, respectively (brain web database)
Fig. 8. Estimation of similarity metric for MR-T1 and noisy MR-T2 with a 10 dB SNR, b 5 dB
SNR and c 2 dB SNR, respectively
Fig. 9. Error plot for registration parameter (rotational angle)
7 Conclusions
In this work, we have presented an embedded entropy image-based multimodal image

registration. The proposed method is based on the statistical relation between a pixel
and its neighborhood in an image region. The embedded entropy, which is computed
from this joint histogram, gets affected with increase in noise content. However, the
proposed method achieved better results as compared to other existing methods with
high noise levels in terms of estimation error.
References
1. Maes F, Vandermeulen D, Suetens P (1997) Medical image registration using mutual
information. IEEE Trans Med Imaging 16(2):187–198
2. Chen H, Varshney PK (2003) Mutual information-based CT-MR brain image registration
using generalized partial volume joint histogram estimation. IEEE Trans Med Imaging
22(9):1111–1119
3. Li J, Cong R, Jin L, Wei P (2008) A medical image registration method based on weighted
mutual information. In: 2nd International conference on bioinformatics and biomedical
engineering, Shanghai, pp 2549–2552
4. Woo J, Stone M, Prince JL (2015) Multimodal registration via mutual information
incorporating geometric and spatial context. IEEE Trans Image Process 24(2):757–769
5. MohanalinB, Kalra PK, Kumar N (2010) An automatic image registration scheme using
Tsallis entropy. Biom Sig Process Control 5(4):328–335
6. Mohammed Khader A, Hamza B (2012) An information-theoretic method for multimodality
medical image registration. Expert Syst Appl 39(5):5548–5556
7. Li B, Yang G, Shu H, Coatrieux JL (2014) A new divergence measure based on Arimoto
entropy for medical image registration. In: 22nd International conference on pattern
recognition, Stockholm, pp 3197–3202
8. Hacine-Gharbi A, Deriche M, Ravier P, Harba R, Mohamadi T (2013) A new histogram-

based estimation technique of entropy and mutual information using mean squared error
minimization. Comput. Electr. Eng. 39(3):918–933
9. Han Y, Feng X-C, Baciu G (2013) Local joint entropy based non-rigid multimodality image
registration/ Pattern Recogn. Lett. 34(12):1405–1415
10. Wachinger Christian, Navab Nassir (2012) Entropy and Laplacian images: structural
representations for multi-modal registration. Med Image Anal 16(1):1–17
11. Viola P, Wells III WM (1997) Alignment by maximization of mutual information. Int j
comput vis 24(2):137–154
Power Factor Corrected SEPIC and Cuk
Converter—A Comparison
Alok Kumar Mishra1(&), Akshaya Kumar Patra1,

Ramachandra Agrawal1, Lalit Mohan Satapathy1, Shekharesh Barik2,
Samarjeet Satapathy1, and Jnana Ranjan Swain1
1
EEE Department, ITER, S‘O’A (Deemed to be University),
lalitsatapathy@soa.ac.in
2
CSE Department, DRIEMS (Autonomous), Cuttack 754022, India
1 Introduction
According to the international standards and for the best power transfer and utilization,
power factor correction has become a necessity. To get a DC output voltage we use a
rectifier and a parallel capacitor [1, 2] as shown in Fig. 1a. The input voltage and
current are shown in Fig. 1b. Figure 1b depicts a pulse-shaped input current with much
harmonics content causing a very poor power factor. The IEEE and IEC are some
international entities to define or standardize the permissible limits of the harmonic
content in-line current such as IEEE 519 and IEC 61000-3-2 [3]. The design of an AC
to DC power converter to overcome these power quality issues like obtaining unity
power factor (UPF) at the AC input mains and a close regulation of the DC output
voltage has been discussed in [4–7]. So, for small power uses in single-phase supplies,
DC-DC converters are used which are switch-based and controlled accordingly to
ensure high power factor (HPF) at the mains side. The main motive is to emulate a
resistive circuit when seen from mains side and improve power factor. Theoretically,
there exist three families of non-isolated power factor correction topologies which are
buck, boost and buck-boost topology [8, 9]. The buck topology is generally used when
low output voltage is required but has high-frequency commuted current at input. Due
to commuted nature, it exhibits a discontinuous nature, and hence should be connected
with high-speed recovery circuit. The major drawback faced by buck topology is the
inclusion of a filter of high-frequency range within the diode bridge and the power
source [10]. The filtering inductor causes a soft variable current in the input end in the
topology of a boost converter. This topology provides high output voltage and leads to
over-voltage stress at the switches [11]. The third family is buck-boost topology and
finds its applications in power factor correction circuits. It includes Cuk and SEPIC
converter [12, 13]. The output voltage polarity makes these two converters different
from each other. In case of Cuk converter, the position of free-wheeling diode and the
inductor is reversed from that of the SEPIC converter to obtain reversed output voltage
polarity [14, 15]. This paper presents two different topologies (namely SEPIC and Cuk)

https://doi.org/10.1007/978-981-15-2774-6_37
300 A. K. Mishra et al.
type power factor correction (PFC) circuit to improve the input-side power factor of the
diode bridge rectifier, lower the THD factor with decreased ripple contained voltage at
output [16]. The entire modeling and the design scheme of input power factor
(IPF) rectified AC to DC power converters are carried out in the MATLAB/Simulink
environment.
(a) (b)
is
i s ,Vs
L
AC O
Vs A
Source Co
D
Diode Bridge
Fig. 1. a AC-DC rectifier, b source voltage and current waveform
2 Circuit Configuration and Design Equation
To improve the power factor at the input and to get a regulated voltage at the output, we
have to interface a DC to DC (SEPIC or Cuk) converter in between the bridge rectifier
and the load as shown in Fig. 2 (SEPIC) Fig. 3 (Cuk), respectively. By adopting these
topologies, a low current ripple can found at input for a low-level DC-bus voltage,
which is rare in case of conventional buck/boost converter. With adequate coupling
coefficient design in a Cuk or SEPIC converter, the same magnetic core can accom-
modate the output and input inductors. In both the converter, when switch is on
inductor L1 and L2 stores energy, inductor current increases linearly and output
capacitor supply power to the load and diode is reverse biased. When switch is off
diode gets forward biased and stored energy of the inductor is supplied to the load.
is L1 C1 D
Io
SW L2 Vo
Vs Co RL
SEPIC Converter
Diode Bridge
Fig. 2. SEPIC converter topology for power factor correction

Power Factor Corrected SEPIC and Cuk Converter … 301
is L1 C1
Io
L2
Vs SW D Co RL Vo
Cuk Converter
Diode Bridge
Fig. 3. Cuk converter topology for power factor correction
Table 1. Designed equation for SEPIC and Cuk converter

Parameters SEPIC Cuk
L1 vs d vs d
DiL1 fs DiL1 fs
L2 vs d vs d
2DiL2 fs DiL2 fs
C1 iL2 d I0 ð1d Þ
Dvc1 fs Dvc1 fs
Co Po vs d
4pfs Vo Dvo 8fs2 L2 Dvo
The different parts of all the converter system are modeled using basic equations
and all equations are summarized in Table 1. Where vs is the rms source voltage, d is
the duty cycle, DiL represents inductor current ripple, Dvc , Dvo are the voltage ripple of
the output voltage and capacitor voltage, respectively, fs is the switching frequency. Po
is the output power, Vo is the output voltage, Io is the output current.
3 Different Methods of Power Factor Correction
There are two objectives of power factor correction: 1. To get a regulated voltage at the
output. 2. The wave shape of current at the input must be a sine wave. To achieve the
first objective, we have to use a feedback loop at the output. There are two methods to
achieve the second objective. First method is called “multiplier approach” and the
second is called “voltage follower approach”. In multiplier approach, a feedback loop
of input current is used to govern the DC to DC power converter to function as an input
voltage programmed current sink as shown in Fig. 4. In this paper, multiplier approach
is used, where multiplier approach control is further sub-divided into four different
methods for generating the gate pulse for the SEPIC and Cuk converter, such as:
is I actual Io
DC-DC
Converter
Vs Vin RL Vo
SW
Diode Bridge
1
K +
Gate
Sinusoidal -
Pulse
Reference
PWM
- Modulator
+
Current error Amplifier
X I ref PI Controller
-
Multiplier +
V o,ref
Voltage error Amplifier

PI or Fuzzy Logic Controller
Fig. 4. Multiplier approach with average current control technique
1. Hysteresis current control

2. Peak current control
3. Average current control
4. Borderline current control.
In this paper, the average current control method is considered which allows a
better input current waveform, which is depicted in Fig. 4. In this technique, the sensed
inductor current is to be filtered with the help of a current error amplifier and the output
of the same drives the PWM modulator. Hence, the error between input current Ig and
its reference gets minimized with the help of inner current loop. The reference current
is generated by a voltage error amplifier (PI or FLC). Due to the PWM modulator,
average current control method gives a constant switching frequency because of current
filtering commutation noises gets eliminated. This method does not need any com-
pensation ramp. Inductor current must have to be sensed which is the demerits of this
method.
4 Controller Used for Output Voltage Regulation
In this paper, two different controllers are used to regulate the output voltage namely PI
controller and fuzzy logic controller and the performance of the same is compared.
4.1 PI Controller
The name itself proportional-integral controller is a type of control loop mechanism
adopted in continuous modulated operation of control. A PI controller repeatedly
calculates the error in between a set point (SP) value and measured process value(PV)
and get a correction accordingly in PI base. In real-time operation, it accordingly
provides an accurate and respective change of correction to a control function. For
example, furnace temperature control that it applies a derivative term to effectively
correct the error despite a huge change.
Vo - LPF
+
Rule Base
Integrator Imax,ref
Vo,ref
e(n) ce(n)
Fuzzification Rule Evaluator Defuzzification
Data Base
Fig. 5. Basic block diagram of FLC
4.2 Fuzzy Logic Controller

The basic block diagram of the proposed FLC for output voltage regulation of SEPIC
and Cuk converter is shown in Fig. 5.
There are two separate meanings of fuzzy logic. Generally, fuzzy logic is the further
application of multivalued logic and is also known as logic system. We can also say,
fuzzy logic is same as fuzzy sets theory that which relates to collection of objects with
unsharp boundaries and the membership is a point of degree. There is a specific object
which degree of membership in a given set which can vary between the ranges 0–1 in
fuzzy set theory. Fuzzy logic is based on sound quantitative and also deals with
imprecise information and data. In mathematical theory, the values of fuzzy variable are
expressed by proper English language. Error in voltage of a power converter can be
defined in linguistic variables like negative big (NB), negative medium (NM), negative
small (NS), zero (ZE), positive small (PS), positive medium (PM), positive big (PB),
and each variable can be defined by varying triangular membership function. Seven
fuzzy levels were chosen and were defined by fuzzy set library which value of the error
is e and change in error is de. The larger the number of fuzzy levels, the higher the input
resolution. We know that a rule is n dimensional and n is the number of variable
included in the rule. The sum of rules is known as rule R. FIS editor edits the input and
output variables, which are e, de and output. After editing, we have to design the
membership function for each variable. The final step involves of writing rules in rule
editor using the rule given in Table 2.
Table 2. Fuzzy rules for closed loop control of SEPIC and Cuk converter
e\de NB NM NS ZE PS PM PB
NB NB NB NB NB NM NS ZE
NM NB NB NB NM NS ZE PS
NS NB NB NM NS ZE PS PM
ZE NB NM NS ZE PS PM PB
PS NM NS ZE PS PM PB PB
PM NS ZE PS PM PB PB PB
PB ZE PS PM PM PB PB PB
To investigate the performance of the proposed system, a Simulink model of a Cuk and
SEPIC converter used for power factor correction is developed as shown in Figs. 6 and
7, respectively. The designed parameter used in simulation for both SEPIC and Cuk
converter is given in Table 3. The system performance is investigated in six steps.
Fig. 6. Simulink model of Cuk converter

Fig. 7. Simulink model of SEPIC converter
Table 3. Designed parameter of the system

Parameters Values
Input voltage RMS value 120 V
Supply frequency 50 Hz
L1 , L2 6 mH, 10 mH
C1 , Co 10, 10 mF
Load power 1 kW
Reference output voltage 100 V
Switching frequency 40 kHz
Voltage and current ripple 0.05
Duty cycle 0.45
Step: 1 when none of these converters is connected to the system. Step: 2 when Cuk
and SEPIC converters are connected to the system with inner current loop only and
without the output voltage feedback loop. Step: 3 when Cuk converter is connected to
the system with PI controller in the feedback loop. Step: 4 when Cuk converter is
connected to the system with FLC in the feedback loop. Step: 5 when SEPIC converter
is connected to the system with PI controller in the feedback loop. Step: 6 when SEPIC
converter is connected to the system with FLC in the feedback loop. To convert AC to
DC, we generally connect a bridge rectifier and filter capacitor across the load as shown
in Fig. 1, the simulation result of the same along with its harmonic spectrum is shown
in Fig. 8a–d.
To improve the performance of the input current and output voltage, Cuk or SEPIC
converter is used as discussed in Figs. 3 and 4. Various waveforms at steady state in
open loop are shown in Fig. 9 (Fig. 9a–c for SEPIC and Fig. 9d–f for Cuk). From
Fig. 8. Simulation results of the proposed system without any converter
(Fig. 9b or e), it is clear that when load is increased at t = 1 s, load voltage is not a
regulated one. To get a regulated output voltage and improved input current wave, PI or
FLC is used in the output feedback loop. The dynamic response of the system is shown
in Figs. 10a–f and 11a–f for Cuk converter, Figs. 12a–f and 13a–f for SEPIC converter,
respectively, for PI and fuzzy logic controller in the outer loop for both the converter.
To test the validity of the PI and FLC, a load change and reference voltage change are
created at t = 1 s and at t = 2 s, respectively. With this dynamic change, the regulated
output voltage is obtained which can be clearly seen from Figs. 10e, 12e and 11e, 13e,
respectively, for PI and FLC. Various performance parameters obtained in simulation
are given in Table 4 for comparison.
Fig. 9. Simulation results of the system in open loop (a–c for SEPIC and d–f for CUK)
Fig. 10. Simulation results of PI Controlled Cuk converter
6 Conclusion
The design, modeling and simulation of Cuk and SEPIC converter for power factor
correction purposes considered in MATLAB/Simulink environment. The outcomes of
the simulation work provided low THD of the supply current with upgraded AC mains
power factor and reduced output voltage ripple. Comparing with PI and FLC used in
the output feedback loop, FLC gives better result in terms of THD of supply current,
IPF and output voltage ripple. From the obtained result, it is clear that fuzzy logic
Fig. 11. Simulation results of fuzzy logic controlled Cuk converter
controlled Cuk converter gives better result when load disturbance and reference
change occur. The prototype of the proposed Cuk converter can be developed that
would be attempted as future work.
Fig. 12. Simulation results of PI controlled SEPIC converter

Fig. 13. Simulation results of fuzzy logic controlled SEPIC converter
Table 4. Performance parameters of the system under different condition

Performance parameter THD (%) Power factor Output voltage
Different condition ripple (%)
Without any converter 191.36 0.4479 9.2
Open loop SEPIC converter 4.37 0.9986 2.0
Open loop Cuk converter 4.70 0.9986 2.2
PI controlled SEPIC converter 4.17 0.9983 2.1
PI controlled Cuk converter 3.13 0.9990 2.1
Fuzzy logic controlled SEPIC converter 4.16 0.9997 1.5
Fuzzy logic controlled Cuk converter 2.97 0.9999 1.8
References
1. Mishra AK, Pathak MK, Das S (2011) Isolated converter topologies for power factor
correction—a comparison. In: IEEE International conference on energy, automation and
signal 28 Dec 2011, pp 1–6
2. Kuiyuan W (2006) The comparison and choice of several power factor correction methods.
In: IEEE vehicle power and propulsion conference, Sep 2006. IEEE, pp 1–5
3. 519–2014 IEEE recommended practice and requirements for harmonic control in electric
power systems, 11 June 2014, pp 1–29
4. Pressman I (1991) Switching power supply design. McGraw-Hill, New York
5. Singh B, Singh BN, Chandra A, Al-Haddad K, Pandey A, Kothari DP (2003) A review of
single-phase improved power quality AC-DC converters. IEEE Trans Ind Electron 50(5):
962–981
6. Batarseh I Power electronics circuits, 3rd edn. Wiley
7. Richard R (1994) The fundamentals of power factor correction. Int. J. Elect. Eng. Educ.
31:213–229
8. Dah D, Ki SK (2013) Light-load efficiency improvement in buck-derived single-stage single-
switch PFC converters. IEEE Trans Power Electron 28(5):2105–2110
9. Liu X, Xu J, Chen Z et al (2015) Single-inductor dual-output buck–boost power factor
correction converter. IEEE Trans Ind Electron 62(2):943–952
10. Kanaan HY, Al-Haddad K, Fnaiech F (2004) Switching-function-based modeling and
control of a SEPIC power factor correction circuit operating in continuous and discontinuous
current modes. In: IEEE international conference on industrial technology, vol 1, no 8, Dec
2004, pp 431-437
11. Umamaheswari MG, Uma G (2013) Analysis and design of reduced order linear quadratic
regulator control for three phase power factor correction using Cuk rectifiers. Electr Power
Syst Res 1(96):1–8
12. Umamaheswari MG, Uma G, Vijitha SR (2012) Comparison of hysteresis control and
reduced order linear quadratic regulator control for power factor correction using DC–DC
Cuk converters. J Circ Syst Comput 21(1):1250002
13. Umamaheswari MG, Uma G, Isabella LA (2014) Analysis and design of digital predictive
controller for PFC Cuk converter. J Comput Electron 13(1):142–154
14. Poorali B, Adib E (2016) Analysis of the integrated SEPIC-flyback converter as a single-
stage single- switch power-factor-correction LED driver. IEEE Trans Ind Electron 63(6):
3562–3569
15. Melo PF, Gules R, Ribeiro EF et al (2010) A modified SEPIC converter for high-power-
factor rectifier and universal input voltage applications. IEEE Trans Power Electron 25(2):
310–321
16. Sudhakarababu C, Veerachary M (2004) Zeta converter for power factor correction and
voltage regulation. In: 2004 IEEE region 10 conference TENCON 2004, vol 500. IEEE,
pp 61–64
An Efficient Classifier-Based Approach
for Early Arrhythmia Detection with Feature
Reduction Using Ranker Search Algorithm
Monalisa Mohanty1, Asit Kumar Subudhi1(&),

Pradyut Kumar Biswal2, and Sukanta Sabut3
1
Department of Electronics and Communication, Institute of Technical
Education & Research, SOA Deemed to be University, Bhubaneswar, India
asitsubudhi@soa.ac.in
2
Department of Electronics Engineering, IIIT Bhubaneswar, Bhubaneswar,
India
3
School of Electronics Engineering, KIIT Deemed to be University,
1 Introduction
The VF and the VT are hazardous cardiac chaos, which requires quick response and
request of high-energy shock in the heart area [1]. Due to VF, there occurs an improper
electrical movement in the ventricles, and hence, the heart is no longer an effective
pump.
The authors have used the temporal and spectral domain approach to discriminate
ventricular arrhythmias [2]. The wavelet transforms (WT), support vector machines
(SVMs), and neural network approaches are used for differentiating several ventricular
arrhythmias [3–5]. The spectral and temporal features were combined with the mor-
phological characteristics of ECG signal for the detection of VF and VT conditions [6].
The SPRT, i.e., sequential probability ratio test, and TCI, i.e., threshold crossing
interval methods, have been used in time domain for precise classification of the
VF/VT condition. Zhang et al. [7] have described a complexity measure approach for
the detection of VT/VF conditions by changing the ECG signal into 0–1 string.
Various algorithms and methods have been proposed for recognition of the cardiac
arrhythmias which are required for efficient classification of VT and VF arrhythmia
rhythms. The detection performance of VF conditions may be improved using com-
bined features extracted from ECG signal. The current work aims at recognition and
classification of VT/VF arrhythmia conditions with the help of spectral, statistical, and
temporal domain features. For a 5 s window length, the de-noised signals were sub-
jected for feature extraction. The extracted features have been ordered and ranked with
the help of correlation attribute evaluation along with a ranker search method. The
arrhythmia conditions were classified with the help of ordered features using J48
decision tree algorithm, SVM, MLP, and by CVR which has been discussed thoroughly
in Sect. 2. The evaluation of results with different parameters and comparison with
other time–frequency domain methods has been done and is explained in Sect. 3.

https://doi.org/10.1007/978-981-15-2774-6_38
314 M. Mohanty et al.
Section 4 corresponds to the conclusion of the entire work that has been used for
classification the ventricular arrhythmia effectively.
2 Methodologies
The proposed block diagram of ventricular arrhythmia detection process is as shown in

Fig. 1.
Fig. 1. Proposed block diagram of ventricular arrhythmia detection process
2.1 Database
Two different databases of PhysioNet repository, i.e., CUDB and VFDB, were con-
sidered. Data of 35 recordings with eight-minute duration of single-channel ECG signal
from CUDB database along with data of 22 recordings with thirty-five-minute duration
of two-channel ECG signal from VFDB database were taken for feature extraction of
ECG signal. A total 57 records and 78 complete duration annotated ECG channels were
used for the study. A suitable window length of 5 s is considered for feature extraction
methods.
2.2 Preprocessing of ECG Signal

With a sampling frequency of 250 Hz, the preprocessing of acquired ECG signals was
done to reduce high-frequency noise, to suppress baseline wandering effect, and to
eliminate power line interference. Thus, the steps of preprocessing are:
(1) High-pass filtering with a fc = 1 Hz to reduce baseline wander;
(2) Second-order Butterworth low-pass filtering with a cutoff frequency of fc = 30 Hz
to reduce high-frequency noise; and
(3) Notch filtering with a fc = 60 Hz to eliminate power line interference.

A total nineteen number of features were extracted in statistical, spectral, and temporal
domains for precise classification of the ventricular arrhythmia. The weighted and
ranked features are given as LKG, TCSC, A3, AB, C1, CO, C3, FSMN, FB, VAL,
MEA, CROSS, TCI, A2, A1, KUR, HURST, SKW, and PERM.
An Efficient Classifier-Based Approach … 315
2.4 Feature Selection

The feasibility of extracted parameters was measured using a feature selection tech-
nique called “correlation attribute evaluation method.” The selected features are further
ranked according to their respective weightage value using a ranker search method.
Correlation Attribute Evaluation (CAE) [8].
This algorithm estimates the significance of a feature by computing the correlation
between the class and the feature and is conferred as:
nCzi
Czc ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n þ nðn 1ÞCii
where
Czc indicates the correlation between the class variable and the summed feature
subsets.
n indicates the no. of subset features.
Czi indicates the average correlations of the subset features and the class variable,
and Cii indicates the average inter-correlation between subset features.
2.5 Classifiers
Four different classifiers have been used in this section. The ranked and ordered fea-
tures are fed as the input to the J48 decision tree, SVM, multilayer perceptrons, i.e.,
MLP and classification via regression classifiers, i.e., CVR, for precise classification of
NSR, VT, and VF arrhythmia rhythms.
2.5.1 J48 Decision Tree Algorithm

J48 decision tree algorithm is a classifier which classifies a specified data point using a
top-down approach [9]. The node having the maximum information gain has been
carefully chosen as the test attribute. For a given attribute A, the information gain
(IG) at any node is calculated as:
X
j SV j
IGðS; AÞ ¼ EntropyðSÞ Entropy(SÞ ð1Þ
vValuesð AÞ
jSj
where S denotes the set of instances; |S| denotes its cardinality; Sv denotes the subset of
S for which attribute A has value v.
2.5.2 Support Vector Machine

SVM is usually a supervised machine learning algorithm which has been used for both
classification
and regression
tasks [5]. Let us consider a training data set with N pairs,
i.e., x1; y1 ; x2; y2 ; . . .; xN; yN with xi 2 <p and yi 2 f1; 1g.
Let us assume the hyper-plane is given as fx : f ð xÞ ¼ xT b þ b0 ¼ 0g, where b is a

unit vector,

GðxÞ ¼ sign xT b þ b0 C ð2Þ
2.5.3 Multilayer Perceptron

The multilayer perceptron calculates a single output from number of real-valued inputs
[10]. It can be given mathematically as:
!
X
n
Y ¼u v i ai þ b ¼ u VTA þ b ð3Þ
i¼1
where
V represents the weight vector.
A represents the input vector.
b represents the bias.
u represents the activation function.
2.5.4 Classification via Regression

It is used for performing classification using regression methods [11]. It generates a
binary class along with one regression model corresponding to each class.
2.6 Performance Measures

The typical statistical indices used for performance evaluations such as sensitivity, i.e.,
Se, specificity, i.e., Sp, precision, i.e., PPV, and accuracy, i.e., Acc, are given as:
TP
Se ¼
TP þ FN
TN
Sp ¼
TN þ FP
TP þ TN
PPV ¼
TP þ FP
TP þ TN
Acc ¼
TP þ FP þ TN þ FN
where TP is true positive, TN is true negative, FN is false negative, and FP is false

positive.
3 Results
A total 19 number of features were extracted from the de-noised ECG signals. The
ordered and ranked features are given as LKG, C1, C3, CROSS, FSMN, A1, A2, A3,
VAL, MEA, SKW, KUR, CO, FB, AB, HURST, PERM, TCI, and TCSC. The eval-
uation was done by correlation attribute evaluator by computing the correlation
between the attribute and the class. Table 1 demonstrates the ordered ranking of
weighted features, and the selected features are given as LKG, TCSC, A3, AB, C1, CO,
C3, FSMN, FB, VAL, MEA, CROSS, TCI, A2, A1, KUR, HURST, SKW, and PERM.
Table 1. Ranking of the selected features using correlation attribute evaluation technique along
with ranker search method
Rank Feature name Weightage
1 LKG 0.6809
2 TCSC 0.5642
3 A3 0.5261
4 AB 0.4564
5 C1 0.4250
6 CO 0.4111
7 C3 0.3962
8 FSMN 0.3640
9 FB 0.2208
10 VAL 0.2206
11 MEA 0.2206
12 CROSS 0.2206
13 TCI 0.2056
14 A2 0.1827
15 A1 0.1754
16 KUR 0.1604
17 HURST 0.1072
18 SKW 0.0441
19 PERM 0.0390
The performance valuation of J48 algorithm with different features combinations at

a confidence factor, i.e., CF = 1, was done, and the finest classification was obtained
for all 19 features together with a CF = 1 as shown in Table 2. The maximum accuracy
rate was found to be 97.7317% with Se = 92.7822%, Sp = 98.1123%, and PPV =
96.0063%, respectively. Figures 2 and 3 correspond to the accuracy with different
feature combinations and the comparative analysis of evaluated parameters,
respectively.
Table 2. Performance valuation of J48 algorithm with different feature combinations at CF = 1

Feature TP FP FN TN ACC SE SP PPV
combinations
1 4214 902 902 9330 88.25 62.67 89.65 79.17
1, 2 4233 881 881 9347 88.52 64.51 90.01 72.95
1, 2, 3 4402 712 712 9516 90.72 71.13 92.40 81.61
1, 2, 3, 4 4402 712 712 9516 90.72 71.13 92.40 81.61
1, 2, 3, 4, 5 4457 657 657 9571 91.44 73.21 93.02 82.16
1, 2, 3, 4, 5, 6 4449 665 665 9563 91.33 73.07 92.96 81.42
1, 2, 3, 4, 5, 6, 7 4607 507 507 9721 93.39 79.25 94.62 86.75
1, 2, 3, 4, 5, 6, 7, 8 4662 452 452 9776 94.11 81.39 95.17 89.20
1, 2, 3, 4, 5, 6, 7, 8, 4801 313 313 9915 95.92 87.73 96.72 91.93
9
1, 2, 3, 4, 5, 6, 7, 8, 4850 264 264 9964 96.56 89.22 97.19 93.65
9, 10
1, 2, 3, 4, 5, 6, 7, 8, 4812 302 302 9926 96.06 87.44 96.78 92.77
9, 10, 11
1, 2, 3, 4, 5, 6, 7, 8, 4812 302 302 9926 96.06 87.44 96.78 92.77
9, 10, 11, 12
1, 2, 3, 4, 5, 6, 7, 8, 4854 260 260 9968 96.61 89.11 97.22 93.87
9, 10, 11, 12, 13
1, 2, 3, 4, 5, 6, 7, 8, 4873 241 241 9987 96.86 89.72 97.39 94.54
9, 10, 11, 12, 13,
14
1, 2, 3, 4, 5, 6, 7, 8, 4844 270 270 9958 96.48 88.27 97.08 94.02
9, 10, 11, 12, 13,
14, 15
1, 2, 3, 4, 5, 6, 7, 8, 4847 267 267 9961 96.52 88.39 97.13 94.01
9, 10, 11, 12, 13,
14, 15, 16
1, 2, 3, 4, 5, 6, 7, 8, 4847 267 267 9961 96.52 88.89 97.16 93.35
9, 10, 11, 12, 13,
14, 15, 16, 17
1, 2, 3, 4, 5, 6, 7, 8, 4934 234 234 10,102 96.98 90.48 97.51 94.33
9, 10, 11, 12, 13,
14, 15, 16, 17, 18
All 19 features 4940 174 174 10,054 97.73 92.78 98.11 96.01
Bold indicates the best results as compared to the other feature combinations
Fig. 2. Results of accuracies for different feature combinations using four classifiers
Fig. 3. Comparison of evaluated parameters of four different classifiers
4 Conclusion
Precise detection and accurate classification of arrhythmia rhythms are extremely vital
in order to provide external defibrillation. A group of 19 features in statistical, spectral,
and temporal domain such as LKG, TCSC, A3, AB, C1, CO, C3, FSMN, FB, VAL,
MEA, CROSS, TCI, A2, A1, KUR, HURST, SKW, and PERM was considered for a
5 s window length. The extracted features were subjected to a feature selection pro-
cedure called correlation attribute evaluation to evaluate the prejudiced properties of
the extracted ECG features. The weighted and ranked features are combined together
for the J48 decision tree algorithm to provide a robust classifier for ventricular
arrhythmia. The results in this study show that the use of J48 decision tree algorithm
significantly improves the capability for the detection of life-threatening arrhythmias as
compared to the SVM, MLP, and CVR classifiers.
References
1. Jekova I (2007) Shock advisory tool: detection of life-threatening cardiac arrhythmias and
shock success prediction by means of a common parameter set. Biomed Sig Process Control
2:25–33
2. Li Q, Rajagopalan C, Clifford G (2014) Ventricular fibrillation and tachycardia classification
using a machine learning approach. IEEE Trans Biomed Eng 61:1607–1613
3. Clayton RH, Murray A, Campbell RWF (1994) Recognition of ventricular fibrillation using
neural networks. Med Biol Eng Comput 32:217–220
4. Khadra L, Al-Fahoum AS, Al-Nashash H (1997) Detection of life-threatening cardiac
arrhythmias using the wavelet transformation. Med Biol Eng Comput 35:626–632
5. Alonso-Atienza F, Morgado E, Fernandez-Martınez L, Alberola AG, Rojo-Alvarez JL
(2012) Combination of ECG parameters with support vector machines for the detection of
life-threatening arrhythmias. Comput Cardiol 39:385–388
6. Krasteva V, Jekova I (2005) Assessment of ECG frequency and morphology parameters for
automatic classification of life-threatening cardiac arrhythmias. Physiol Meas 26:707–723
7. Zhang XS, Zhu YS, Thakor NV, Wang ZZ (1999) Detecting ventricular tachycardia and
fibrillation by complexity measure IEEE Trans Biomed Eng 46:548–555
8. Karegowda AG, Manjunath AS, Jayaram MA (2010) Comparative study of attribute
selection using gain ratio and correlation based feature selection. Int J Inf Tech Know Man
2:271–277
9. Panigrahi R, Borah S (2018) Rank allocation to J48 group of decision tree classifiers using
binary and multiclass intrusion detection datasets. Procedia Comput. Sci. 132:323–332
10. Raji G, Chandra SCSV (2017) Long-term forecasting the survival in liver transplantation
using multilayer perceptron networks. IEEE Trans Syst Man. Cybern 47:2318–2329
11. Hou C, Nie F, Yi D, Wu Y (2013) Efficient image classification via multiple rank regression.
IEEE Trans Image Proc 22:340–352
The Fractional Order PID Controller Design
for BG Control in Type-I Diabetes Patient
Akshaya Kumar Patra1(&), Anuja Nanda1, Santisudha Panigrahi2,

and Alok Kumar Mishra1
1
Department of EEE, ITER, S‘O’A University, Bhubaneswar 751030, India
2
Department of CSE, ITER, S‘O’A University, Bhubaneswar 751030, India
1 Introduction
As per the World Health Organization report, one among the widespread diseases is
diabetes mellitus and is resulted due to the malfunctioning of the pancreas. This
reduces insulin sensitivity affecting the normoglycaemic range of BG concentration
(70–120 mg/dl) in a healthy human being. At present, numerous research projects are
undertaken by several researchers to get rid of this problem by devising advanced
medical equipment like automated MID. Till date, BG concentration is manually
controlled to adopt the open-loop control strategy. The hypoglycemic or hyperglycemic
conditions may evolve due to the difficulties in handling the internal system changes
and external disturbances by implementing the control loop technique. The develop-
ment of implanted artificial pancreases (AP) enabling the adequate dose of insulin
delivery proportionate to the sensor measurement in the patient’s body may provide the
means to incorporate the closed loop control strategy. Figure 1 describes a closed loop
patient model with an AP. The AP consists of glucose sensor, MID, and controller. The
sensor measures the BG concentration of the human body continuously and sends
signal to the controller for generating the desired control actions. The control signal
generated by this controller also depicts the association of the additional model
uncertainties and disturbances. Thereafter, according to the control signal u(t), the
optimal insulin dose is infused into the patient’s venous blood by MID to achieve the
normoglycaemic range of BG concentration [1, 2].
To determine an optimal solution for the AP like building an appropriate model of
the complex BG regulatory system, a number of obstacles and challenges such as the
effects of non-linear behavior, time-dependent dynamics, presence of several sources of
disturbance, uncertainty, and lack of glucose sensing are to be faced. Additionally, the
challenges and constraints related to control, specifically for BG regulation controller
design are considerable glucose measurement delay, insulin absorption delay after
being injected, irreversible action of insulin, meal detection and estimation, model
parameter variation, asymmetric risk of extreme BG concentration variations and time-
based control needs, etc. [1]. Despite of technical progress and considerable devel-
opment on aforesaid issues, substantial improvement is still required in the control
algorithm. Assessment of the glucose excursions following the insulin dose adjustment
needs a controller for BG regulation in AP and many authors suggested the PID
controller as a viable solution [3, 4]. However, due to glucose sensing time delay,
https://doi.org/10.1007/978-981-15-2774-6_39
Fig. 1. Compartmental model of TIDM patient with AP
insulin action and non-variable gain parameters, the desired performance, assuring high
accuracy, reliability, and robustness could not be achieved. The fuzzy control [5, 6],
LQG control [7], H∞ control [8–10], Sliding Mode (SM) control [11–14], and model
predictive (MP) control [15, 16] are some of the well proven controllers to deal with the
BG regulation issues. The BG control in the diabetic patients within the normogly-
caemia range inculcating the above controllers enhanced the accuracy and robustness
to some extent compared to the PID controllers. However, these control algorithms are
not fully insensitive to the disturbances and the uncertainties of the model in spite of
the improved performance. Hence, optimal control parameters setting for better per-
formance and for avoiding slow response following meal disturbance, the current work
suggests an alternative novel technique implementing the fractional order PID
approach.
The FOPIDC approach concept leads to an enhanced control performance with
respect to robustness and delay of time compensation characteristics to counteract the
negative impact of associated errors, disturbances, and uncertainties. The stabilizing
control law uðtÞ is developed based on the fractional order approach. Application of the
FOPIDC to control the BG concentration in TIDM patient results to ensure a better
robust controller in comparison to other contemporary well-accepted methods under
both harmonized and incompatible uncertainties.
This manuscript is organized as follows. Section 2 concisely illustrates the TIDM
patient model with mathematical details reflecting to its dynamic characteristics of its
GM process. A detailed presentation on how the control technique is formulated and
how it is implemented for this problem is presented in Sect. 3. Comparative results of
the proposed approach are presented in Sect. 4. The concluding comments are sum-
marized in the Sect. 5.
The Fractional Order PID Controller Design … 323

2.1 System Overview
The compartmental model schematic diagram as depicted in Fig. 1 reveals the GM
dynamics of the human body. The proposed controller is tested using the above model.
The human body has liver, kidney, periphery, heart/lungs, gut, and brain as different
functional parts and is shown as six compartments separately. The blood flow direc-
tions or circulation are indicated by the arrow marks. The controller receives the
measured arterial glucose value as input to provide the optimal insulin dose to be
injected as an output with the help of MID.
2.2 Clinical Background

A human body with a prolonged high arterial BG level that exceeds 144 mg/dl is
known to suffer from hyperglycemia disease. This type of the clinical disorder is also
termed as diabetes mellitus. The insulin deficiency along with its less resistive or
insensitiveness independently or combinedly cause hyperglycaemia and are some of
the major reasons as discussed in [17, 18]. Diabetes mellitus may occur if the blood
glucose remains unutilized effectively under any normal human life cycle. Diabetes is
classified as type-I and type-II as per the rate of insulin creation in the patient’s body. In
the type-I diabetic patients, the insulin creation by the pancreatic cells is completely
impossible, but in case of the type-II is at a bit smaller rate, which also gets inhibited in
due course. Hence, in the present situation, an immediate insulin injection to restore the
normal BG level is essential that also prevents allied adverse impacts on the normal
operation of any human body. Subject to diverse fooding, fasting, and exercise
behaviors, the regulation of the BG levels are not very easy in the real human life cycle.
The external sources like the carbohydrate foods are digested down into glucose at
different parts of the digestive system (gut), and then filtered. The filtered glucose is
added to the venous blood. Later on, the liver stores this glucose after being transferred
from the venous blood as glycogen. When the BG concentration in the venous blood is
less than normoglycaemic range of glucose concentration, glycogen is re-injected as
glucose by the liver into venous blood and is known as the internal glucose source. The
glucose utilization and production by the liver is known as the net hepatic glucose
balance (NHGB). which is the general phenomena in the glucose metabolism process.
The total energy needed by human body organs irrespective of insulin dependency
derives from the degree of glucose utilization.
The creation of insulin is due to the b-cells present in the pancreas and thus the BG
level is regulated. A normal human with high BG level has two major functions of
insulin, firstly to allow the liver for the glucose absorption along with the storing it in
the form of glycogen, specifically under the meal intake conditions. As a result, the
production of excess ‘internal’ glucose by the liver and muscles is stopped. Secondly,
insulin speeds up the glucose absorption in the muscles and fulfills the peripheral
energy needs of the body. But, in case of diabetes patients, the abnormality in the
metabolic process dynamics, both the aforementioned functions are found to impair
partially or completely. In case of diabetes patient, the cells stop the glucose utilization
and internal glucose is produced by the liver, then an uncontrolled BG level is noticed.
Later, as the BG level is more than the renal threshold glucose (RTG) value of
162 mg/dl, venous BG in excess is extracted through the kidney.
2.3 Modeling of a TIDM Patient

In present scenario, with due consideration of the dynamics of glucose metabolism
process, several simulation models are proposed to control the BG concentration in the
TIDM patients [19, 20]. Owing to the simplified structure and desired approximation to
the dynamics of human metabolism with reduced error, the model suggested by
Lehmann [19, 20] are among the widely accepted ones. The testing and verification of
the proposed technique using Lehmann [19, 20] model for BG control are considered in
the present study. Figure 1 describes the compartmental set up of diabetic patient with
the implanted AP. The controller computes the insulin injection amount to the venous
blood at 5 min interval using an integrated implanted insulin pump. Depending on the
supporting device and sensing technology, the sampling rate varies when is being
applied in the real time domain [21, 22]. The gut compartment input is the meal and the
peripheral input is the exercise for integrating and executing the process disturbances.
Compartmental modeling method considering the equations of the fast principle has
been used to develop the GI interaction process with a MID model of the patient.
Figure 1 demonstrates the simulated model diagram of the patient with MID.
2.4 Analysis of Patient Dynamics

The patient model dynamic and characteristic operations are verified under several
operating constraints like the actuator and sensor noises, intake amount of carbohy-
drates, varying exercises, etc. The BG concentration and the insulin dose of the con-
sidered patient model with 60 gm meal at 600 min and the exercise for half-an-hour at
1300 min are illustrated in Fig. 2a. The glucose production and consumption rate of
liver, and glucose consumption rate by the BG profile organs such as gut compartment,
CNS, and peripheral cells are shown in Fig. 2b. The overall BG regulation in the
human body is carried out naturally through glucose utilization by the organs, such as
peripheral cells (adipose tissues, muscle cells) and the liver that exclusively depend on
insulin. The peripheral cells and liver consume the least amount of glucose under the
condition of lack of insulin. Under this situation, BG level rises abnormally and go
beyond the glucose level of 144 mg/dl. This results in the hyperglycaemia problem.
The kidney removes part of the glucose from venous blood under the condition when
the BG concentration rises up to RTG value. The rate of glucose excretion of the
kidney proportionate to the BG level is represented by Fig. 2b. Figure 2a, b illustrate
the unstable dynamics under various model uncertainties and disturbances. These
abnormal dynamics can be reduced by applying the AP based on suitable control
algorithm.
Fig. 2. a BG concentration with respect to insulin dose; b glucose production and consumption
rate by GM organs
3 Control Algorithm
The FOPIDC control algorithm is demonstrated in this section. The closed loop system
response with respect to stability, accuracy, and robustness is analyzed. The control
specifications such as settling time (min), steady-state error ess (%), overshoot (mg/dl),
and undershoot (mg/dl) are also evaluated and examined with proper validation of the
controller actions.
3.1 FOPIDC Design

The TIDM patient model with fractional order PIa Dc controller is shown in Fig. 1a. In
fractional order PIa Db control, the error signal e(t) is used to generate control signal u
(t). The transfer function (TF) of the proposed controller can be formulated as [23]:
Ki
TF ¼ Kp þ þ K d sb ð1Þ
sa
where, a and b are the fractional orders of the integrator and differentiator of the
proposed controller, respectively. Kp, Ki, and Kd are denoted for the proportional,
integral, and derivative gains of the proposed controller, respectively. The optimal
values of the FOPIDC parameters are considered based on the MATLAB/SIMULINK
Table 1. Optimal values of control parameters

Kp Kd Ki a b
10 8 0.05 0.7 0.9
Fig. 3. Simulink diagram of FOPIDC
environment and FOMCON toolbox as represented in Table 1. The structure of

FOPIDC is shown in Fig. 3, which is designed based on the Eq. (1).
4 Result and Discussions
Time-domain response of glucose profiles, stability, and robustness of the closed loop
model with proposed FOPIDC are described in detail in this section. The proposed
control approach is compared with other popular control algorithms to justify its
enhanced performance.
4.1 Analysis of Patient Dynamics with FOPIDC

In this section, all glucose profiles of the non-linear patient model with FOPIDC is
examined under different operating conditions such as variation of exercise, intake
amount of carbohydrate, and noise of sensors and actuators. All glucose profiles like
BG level, insulin dose, NHGB rate, gut rate, CNS and peripheral glucose utilization
rate in TIDM patient model with proposed controller are illustrated in Fig. 4. Com-
pared to the condition of the uncontrolled process, the result obtained evidences a
higher utilization of the plasma glucose by the peripheral cells and liver, those are
dependent on the insulin availability and sensitivity. These bring back the BG level to
81 mg/dl with the quick settling time that result in the reduction or prevention of the
hyperglycaemia occurrence probability as shown in Fig. 4a. The insulin-independent
organ like the CNS, consume the plasma glucose at a constant rate. Figure 4b illustrates
the constant rate of glucose utilization by CNS at 84 mg/min. As the BG level is below
Fig. 4. a BG concentration with respect to insulin dose; b glucose production and consumption
by GM organs
the RTG level, kidney does not extract any glucose and is depicted in Fig. 4b by the
constant glucose excretion rate at zero level. By implementing the proposed control
techniques to the MID, different operational parameters are evidenced and have the
enhanced performances.

The suggested FOPIDC approach is compared with other popular control approaches
such as PID, fuzzy, H∞, and SM to justify its enhanced performance as the controllers.
Figure 4 illustrates the effect of meal disturbance in the BG concentration of the patient
with the FOPIDC approaches. Table 2 presents a comparative analysis with respect to
settling time (min), peak overshoot (mg/dl), peak undershoot (mg/dl), noise (%), and
steady-state error ess (%). The effect of meal disturbance in the BG concentration of the
patient model applying different control approaches such as PID, fuzzy, H∞, and SM
is also presented in Table 2 based on the references [4, 5, 10, 14], respectively. Similar
working conditions are followed with the same level of actuator and sensor noise in all
control techniques application for comparison. The BG concentration in TIDM patient
model ingested to 60 gm meal under different controllers is tested. The corresponding
results are presented for the various control approaches along with the proposed
FOPIDC with respect to control specifications such as overshoots (mg/dl), undershoots
(mg/dl), and settling time (min). The results signify the better controllability of the
FOPIDC. The simulation results also demonstrate the high noise elimination capability
with high robustness for the proposed approach. Overall, by looking to the above
comparative analysis, the findings of suggested approach advantages are the higher
Table 2. Comparative result analysis related to the BG concentration

Controller PID Fuzzy H∞ SM FOPIDC
[4] [5] [10] [14] (Proposed)
Meal intake (mg) 60 60 60 60 60
Insulin infusion rate 59.6 59.2 59.1 59 59.1
(mU/min)
Settling time (min) 290 260 255 250 255
Peak overshoot (mg/dl) 5.2 5.3 6.6 6.5 4.4
Peak undershoot (mg/dl) 3.1 2.1 1.2 1.5 1.1
Noise (%) 10 10 5 5 5
ess (%) 0 0 0 0 0
accuracy and stability, more robustness, high noise elimination capability, and better
capability to handle uncertainty under various abnormal conditions and huge variation
meal disturbance.
5 Conclusions
In this paper, a novel control strategy FOPIDC is proposed for BG control in TIDM
patient model. To justify its enhanced performance, it has been applied and tested to
control the BG concentration in patient within normoglycaemic range. Initially, the
patient is modeled as a ninth-order state-space representation. Then the proposed
control approach FOPIDC is designed. The comparative results clearly reflect that the
suggested FOPIDC is arrived at better performance than the other control approaches
such as PID, fuzzy, H∞, and SM, with respect to stability, reliability, and robustness
under various abnormal conditions and disturbances. The related better performance of
the suggested approach (FOPIDC) in terms of improved accuracy and stability,
enhanced robustness, high noise elimination capability, and better capability to handle
uncertainty justify its real-time application.
References
1. Chee F, Fernando T (2003) Closed-loop glucose control in critically ill patients using
continuous glucose monitoring system (CGMS) in real time. IEEE Trans Inf Technol
Biomed 7(4):43–53
2. Kamath S, Vidyasagar S (2009) Simulation study on closed loop control algorithm of type 1
diabetes mellitus patients. IETE J Res 55:230–235
3. Frederick C, Tyrone L (2003) Expert PID control system for blood glucose control in
critically ill patients. Information technology in biomedicine. IEEE Trans 7(4):419–425
4. Sutradhar A, Chaudhuri S (2002) Analysis and design of an optimal PID controller for
insulin dispenser system. J Inst Eng (India) 82(2):304–313
5. Ibbini M (2006) The PI-fuzzy logic controller for the regulation of blood glucose level in
diabetic patients. J Med Eng Technol 30(2):83–92
6. Singh M, Gupta JRP (2007) A new algorithm-based type-2 fuzzy controller for diabetic
patient. Int J Biomed Eng Technol 1(1):18–40
7. Patra AK, Rout PK (2015) An automatic insulin infusion system based on LQG control
technique. Int J Biomed Eng Technol 17(3):252–275
8. Chee F, Andrey V (2005) Optimal H∞ insulin injection control for blood glucose regulation
in diabetic patients. IEEE Trans Biomed Eng 52(10):1625–1631
9. Yasini S, Sistani MBN (2012) Knowledge-based Closed-loop control of blood glucose
concentration in diabetic patients and comparison with H∞ control technique. IETE J Res
58:328–336
10. Patra AK, Rout PK (2014) Optimal H-infinity insulin injection control for blood glucose
regulation in IDDM patient using physiological model. Int J Autom Control 8(4):309–322
11. Gallardo H, Ana G (2013) High-order sliding-mode control for blood glucose: practical
relative degree approach. Control Eng Pract 21(5):747–758
12. Rmileh A, Gabin W (2012) Wiener sliding-mode control for artificial pancreas: a new
nonlinear approach to glucose regulation. Comput Methods Programs Biomed 107:327–340
13. Patra AK, Rout PK (2018) ‘Backstepping sliding mode Gaussian insulin injection control for
blood glucose regulation in TIDM patient. J Dyn Sys Meas Control 140(9):091006–091006-15
15. Patra AK, Mishra AK, Rout PK (2018) Backstepping model predictive controller for blood
glucose regulation in type-I diabetes patient. IETE J Res. https://doi.org/10.1080/03772063.
2018.1493404
16. Patra AK, Rout PK (2017) Adaptive continuous-time model predictive controller for
implantable insulin delivery system in Type I diabetic patient. Opt Control Appl Methods
38(2):184–204
17. Guyton JR, Foster RO (1978) A model of glucose-insulin homeostasis in man that
incorporates the heterogenous fast pool theory of pancreatic insulin release. Diabetes Care
27(1):1027–1042
18. Barger M, Rodbard D (1989) Computer simulation of plasma insulin and glucose dynamics
after subcutaneous insulin injection. Diabetes Care 12(1):725–736
19. Lehmann ED, Deutsch T (1992) Physiological model of glucose–insulin interaction in Type-
1 diabetes mellitus. J Biomed Eng 14(3):235–242
20. Lehmann ED, Deutsch T (1998) Compartmental models for glycaemic prediction and
decision support in clinical diabetes care: promise and reality. Comput Methods Programs
Biomed 56(1):193–204
21. Sperr G (1997) Biosensor research targets medical diagnostics. Medical device and
diagnostic industry magazine, Nov 1997
22. Cochin L, Cadwallender W (1997) Analysis and design of dynamic systems, 3rd edn.
Addison-Wesley, New York
23. Bingul Z, Karahan O (2018) Comparison of PID and FOPID controllers tuned by PSO and
ABC algorithms for unstable and integrating systems with time delay. Opt Control Appl
Methods 39(4):1431–1450
Mutual Fund Investment Method Using
Recurrent Back Propagation Neural Network
Smruti Rekha Das1(&), Debahuti Mishra2, Pournamasi Parhi2,

and Prajna Paramita Debata3
1
Department of Computer Science and Engineering, Gandhi Institute for
Education and Technology, Baniatangi, Bhubaneswar, Odisha, India
smrutirekhadas@gietbbsr.com
2
Department of Computer Science and Engineering, Siksha ‘O’Anusandhan
debahutimishra@soa.ac.in
3
International Institute of Information Technology, Bhubaneswar, Odisha, India
1 Introduction
Mutual fund is managed professionally. The basic advantages of investing money in

mutual fund are the highest level of diversification. Through diversification, it is able to
hold many types of securities and diversification also reduces the risk of investing
money in securities. In the recent years, a number of methods have been introduced by
the researchers and it is found that neural network-based methods have proved its
efficiency over a different area of financial market prediction. W-C. Chiang has iden-
tified and explored [1] on back propagation neural network (BPNN) to predict net asset
value (NAV) of mutual funds. In addition to this, D. C. Indro et al. has applied artificial
neural network (ANN) for equity mutual [2] fund prediction. Exploring about the
recurrent methods, Manjunath Patel GC et al. proved that in neural network methods,
reverse mapping gives [3] effective prediction performance. Various research articles
have published, exploring on the efficiency of neural network-based methods, from
which, very few are introduced here. By going through the above research, this study
was motivated to explore the efficiency of neural network-based methods for predicting
mutual fund, as the proper prediction can yield a highest benefit to the investor. Three
neural network-based methods such as BPNN, RBPNN and RRBFNN are employed in
this paper, as these are the widely accepted models. An empirical comparison based on
the mean square error (MSE) results, during training, is established among these pre-
diction models. NAV along with some statistical measures calculated for of SBI
Magnum Equity and UTI Equity mutual fund dataset are given as the input to the
prediction model, to validate its predicted performance. The prediction methods have to
be employed for 1 day ahead, 5 days ahead and 7 days ahead. These three different
days ahead are to be predicted by the prediction methods. In addition to this, the
prediction models are also validated through various performance measures at the time
of testing.

https://doi.org/10.1007/978-981-15-2774-6_40
Mutual Fund Investment Method Using Recurrent … 331
The rest of the works are organized in the following manner, Sect. 2 focuses on the
details of methodologies that have been used in this study. The dataset description and
experimental analysis are found in Sects. 3 and 4, respectively. Finally, Sect. 4 con-
cludes the work with future scope.
2 Methodologies
The neural network is developed on the concept of the working principle of brain of
human and its nervous system. It is described by consisting an interconnection among
each layer, each layer having the number of neurons, the learning algorithm that
required for network training and the activation functions. The network is unknown
about the prior assumption of the generating process with approximate of any nonlinear
continuous function. The training [4–6] of neural network based on a simple concept
that whenever the network gives the wrong answer, the weight between the layers is
updated with the expectation of getting a minimum error, because of this network’s
future responses possibly to be correct. Modifying the concept of BPNN, RBPNN is
designed, where the network contains a cyclic connection from the output to the hidden
layer input, for making a more powerful model than [7, 8] the feed-forward neural
network. RBPNN technique has been applied by the resources in many of the research
areas. Among different types of feed-forward neural network, RBFNN is able to solve
the nonlinear problems. However, RRBFNN is having a major drawback that basically
their input to output map is static and to represent the nonlinear data, their efficiency is
very [9, 10] limited. It has been observed that both RBFNN and RRBFNN can
approximately be used for any nonlinear functions if the structures of any nonlinear
functions are constructed properly.
3 Dataset Description
Two mutual fund datasets SBI Magnum Equity and UTI Equity mutual fund are
collected. This study considered these two mutual funds for experimentation as UTI
Equity is the oldest mutual fund and SBI Magnum is the first non-UTI mutual fund in
India. Eleven years of data have selected, with the period of March 1, 2006 to March 1,
2017. In mutual fund only the NAV is available, hence using the NAV the dataset is
regenerated by calculating the [11] mean, standard deviation, kurtosis and skewness.
These statistical measures have generated using the proper mathematical formula. For
calculating the statistical measure, 12 is considered as the window size. After regen-
erating the data from the available NAV, now the dataset is ready for input to the
model, in the ratio of 70% data is for training and 30% for is for testing. Table 1 shows
the total number of available data and the number of data patterns that are generated
during the windowing process. The training patterns and testing patterns are generated
accordingly in the ratio of 7:3.
332 S. R. Das et al.
Table 1. Details of mutual fund datasets

Name of Data Total Total Training Testing
mutual fund available generated patterns patterns
data patterns
SBI magnum 1-03-2006 to 3153 3140 2198 942
equity 1-03-2017
UTI equity 1-03-2006 to 2676 2663 1864 799
mutual fund 1-03-2017
4 Experimental Analysis
The schematic layout of the proposed model is depicted in Fig. 1. At first, two mutual
fund datasets such as SBI Magnum Equity and UTI Equity have been collected. Then,
in the second step, the dimensionality of the dataset has increased using the statical
measures such as mean, standard deviation, kurtosis and skewness with 12 window
size. Further, the dataset is divided into 70% is for training and 30% is for testing.
These training data are set input for all the prediction models such as BPNN, RBPNN
and RRBFNN. An empirical comparison is carried out among these predictions used
for experimental purpose to get the best predicted outcome.
Regeneration
of datasets using
Mean, Standard BPNN
Deviation,
Kurtosis and RBPNN
Skewness in the Predicted
Mutual fund Outcome
ratio of 70% is
datasets RRBFN
for training and
30% is for
testing.
Fig. 1. Abstract view of the proposed model
The simulation result presented in actual versus predicted form for BPNN, RBPNN
and RRBFNN is shown in Figs. 2, 3 and 4, respectively. From the figures, it is
visualized that in RBPNN, the distance between actual graph and predicted graph is
very less.
From the regenerated dataset, five elements such as NAV, mean, standard devia-
tion, kurtosis and skewness for every experimented day are employed to BPNN,
RBPNN and RRBFNN. For all the three prediction models, different parameter values
are set according to the specification of the algorithm. 100 number of iterations is well
thought out for all the models to make the comparison standardize. Apart from itera-
tion, BPNN needs the number of hidden layers and the number of nodes present in each
hidden layer. This study has chosen three numbers of hidden layers and each hidden
layer is having five number of nodes. Similarly, for RBPNN, along with the above
NAV of SBI Magnum Equity Mutual fund NAV of SBI Magnum Equity Mutual fund NAV of SBI Magnum Equity Mutual fund
90 90 90
85 80 80
80 70 70
NAV
NAV
NAV
75 60 60
70 50 50
65 40 Actual 40 Actual
Actual
Predicted Predicted Predicted
60 30 30
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time Series in Day Time Series in Day Time Series in Day
(a) 1 day(SBI) (b) 5 days(SBI) (c) 7 days(SBI)
NAV of UTI Equity Mutual fund
NAV of UTI Equity Mutual fund 50 NAV of UTI Equity Mutual fund
50 50
Actual Actual Actual

Predicted 45 Predicted Predicted
45 45
40 40 40
NAV
NAV
NAV
35 35 35
30 30 30
25 25 25
20 20 20
0 100 200 300 400 500 600 0 100 200 300 400 500 600 0 100 200 300 400 500 600
(a) 1 day(UTI) (b) 5 days(UTI) (c) 7 days(UTI)
Fig. 2. NAV of SBI and UTI mutual fund using BPNN
100 90 90
90 80 80
80
70 70
NAV
70
NAV
NAV
60 60
60
50 50
50
Actual 40 Actual 40 Actual
40 Predicted Predicted Predicted
30 30 30
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400
NAV of UTI Equity Mutual fund NAV of UTI Equity Mutual fund NAV of UTI Equity Mutual fund
65 65 65
Actual
60 Predicted 60 Actual 60 Actual
Predicted Predicted
55 55 55
50 50 50
NAV
NAV
NAV
45 45 45
40 40 40
35 35 35
30 30 30
25 25 25
20 20 20
0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
Fig. 3. NAV of SBI and UTI mutual fund using RBPNN

160 160 160
Actual Actual Actual
140 Predicted 140 140 Predicted
Predicted
120 120 120
100 100 100
NAV
NAV
NAV
80 80 80
60 60 60
40 40 40
20 20 20
0 0 0
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
NAV of UTI Equity Mutual fund NAV of UTI Equity Mutual fund NAV of UTI Equity Mutual fund
250 250 250
Actual Actual
Predicted Actual
Predicted Predicted
200 200 200
150 150 150

NAV
NAV
NAV
100 100 100
50 50 50
0 0 0
0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000
Fig. 4. NAV of SBI and UTI mutual fund using RRBFNN
parameters of BPNN, three other controlling parameters are defined with its suitable
value. The value of the threshold is set to 0.001 and for both the learning rate parameter
and momentum coefficient 0.01 are set for the experimentation. Further, in RRBFNN,
the radial function is the Gaussian function, and the Gaussian function needs two
values: One is the value of center c and the other is the value of radius r. This
experimental work fixed 0 for center c and 1 for radius r. In the training phase, mean
square error (MSE) is calculated to train the network. In Table 2, the MSE results of all
the models during training are given. From the above table, it can be observed that, for
both SBI magnum equity mutual fund and UTI equity mutual fund, RBPNN is per-
forming better than the rest two prediction models. Comparing among the rest two
models, it is found that the performance of BPNN is better than RRBFNN. In addition
to this, the performance of different days ahead has also examined. The prediction
efficiency is very high for the prediction of 1 day ahead than 5 and 7 days. It proves
that when the day increases the model performance is decreases. The prediction
Table 2. MSE of BPNN, RBPNN and RRBFNN during training phase

Day(s) BPNN RBPNN RRBFNN
9
1>= SBI 0.8229 0.41571 1.2769e+03
5 1.2225 0.45536 1.2810e+03
>
;
7 1.4436 0.75775 1.2853e+03
9
1>= UTI 1.1427 0.091836 1.8423e+03
5 1.1675 0.18402 1.8526e+03
>
;
7 2.2503 0.19421 1.9752e+03
methods are validated during the testing phase considering the mean absolute per-
centage error (MAPE) and root mean square error (RMSE) result. For both the
experimented dataset, the validation work is carried out and referring to Table 3, it is
noticed that RBPNN is giving less value in the form of error compared to BPNN and
RRBFNN.
Table 3. Performance evaluation during the testing phase for 1 day ahead prediction
Methods MAPE RMSE
9
BPNN >
= SBI 1.56432 1.56432
RBPNN 0.94251 0.98242
>
;
RRBFNN 2.45231 2.18721
9
BPNN >
= UTI 1.89522 1.98356
RBPNN 0.98882 0.99231
>
;
RRBFNN 2.01432 2.18164
The bold letter indicates of minimum error. That means RBPNN is giving minimum error than
other two models (BPNN and RRBFNN). Minimum error means again maximum accuracy
From the convergence graph depicted in Figs. 5 and 6, it is noticed that RBPNN is
converging faster than BPNN and RRBFNN. RBPNN requires only five number of
iterations to converge, which is very less compared to the total number of experiments
considered here, in this study. By analyzing these convergence graphs, it can be
examined that the convergence speed of BPNN is faster than RRBFNN. This con-
vergence graph is plotted for the MSE error during training, considering the data of
1 day ahead to be predicted for both SBI and UTI Mutual fund.
Error Convergence graph Error Convergence graph Error Convergence graph

60 250 1300
50 1290
200
40 1280
150
MSE
MSE
MSE
30 1270
100
20 1260
50 1250
10
0 0 1240
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
No.of Experiments No.of Experiments No.of Experiments
(a) BPNN (b) RBPNN (c) RRBFNN
Fig. 5. Error convergence graph of SBI during training for BPNN, RBPNN and RRBFNN
Error Convergence graph Error Convergence graph Error Convergence graph

11 1200 1970
10 1960
9 1000
1950
8
800 1940
7
MSE
MSE
MSE
1930
6 600
1920
5
400 1910
4
1900
3
200 1890
2
1 1880
0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
No.of Experiments No.of Experiments No.of Experiments
(a) BPNN (b) RBPNN (c) RRBFNN
Fig. 6. Error convergence graph of UTI during training for BPNN, RBPNN and RRBFNN
5 Conclusion
This study has explored the predictive analytics for different time horizon using the
base neural network models. So, for these three prediction models such as BPNN,
RBPNN and RRBFNN are considered for this experimental work and an empirical
comparison has established among these prediction models to assess their performance.
Observing the simulated graph and MSE result in training as well as RMSE and MAPE
during testing, it is noticed that RBPNN is outperformed over BPNN and RRBFNN. In
future, the work can be extended using different optimization techniques.
References
1. Chiang W-C, Urban TL, Baldridge GW (1996) A neural network approach to mutual fund
net asset value forecasting. Omega 24(2):205–215
2. Indro DC, Jiang CX, Patuwo BE, Zhang GP (1999) Predicting mutual fund performance
using artificial neural networks. Omega 27(3):373–380
3. GC MP, Shettigar AK, Krishna P, Parappagoudar MB (2017) Back propagation genetic and
recurrent neural network applications in modelling and analysis of squeeze casting process.
Appl Soft Comput 59:418–437
4. Sutheebanjard P, Premchaiswadi W (2010) Stock exchange of Thailand index prediction
using back propagation neural networks. In 2010 second international conference on
computer and network technology (ICCNT), IEEE, 2010, pp 377–380
5. Khoa NLD, Sakakibara K, Nishikawa I (2006) Stock price forecasting using back
propagation neural networks with time and profit based adjusted weight factors. In:
International joint conference SICE-ICASE, 2006, IEEE, 2006, pp 5484–5488
6. Lee C-T, Chen Y-P (2007) The efficacy of neural networks and simple technical indicators in
predicting stock markets. In: International conference on convergence information
technology, 2007, IEEE, pp. 2292–2297
7. Sak H, Senior A, Beaufays F (2014) Long short-term memory recurrent neural network
architectures for large scale acoustic modeling. In: Fifteenth annual conference of the
international speech communication association, 2014
8. Huang Z, Huang D, Lyu MR, Lok T (2006) Classification based on Gabor filter using
RBPNN classification. In: 2006 International conference on computational intelligence and
security, IEEE, 2006, vol 1, pp 759–762
9. Han HG, Guo YN, Qiao JF (2017) Self-organization of a recurrent RBF neural network
using an information-oriented algorithm. Neurocomputing 225:80–91
10. Han Hong-Gui, Qiao Jun-Fei (2012) Adaptive computation algorithm for RBF neural
network. IEEE Trans Neural Netw Learn Syst 23(2):342–347
11. Das SR, Mishra D, Rout M (2019) An optimized feature reduction based currency
forecasting model exploring the online sequential extreme learning machine and krill herd
strategies. Phys A Statist Mech Appl 513:339–370
Stabilizing and Trajectory Tracking
of Inverted Pendulum Based on Fractional
Order PID Control
Akshaya Kumar Patra1(&), Alok Kumar Mishra1, Anuja Nanda1,

Dillip Kumar Subudhi2, Ramachandra Agrawal2, and Abhishek Patra2
1
Department of EEE, ITER, S‘O’A University, Deemed to be University,
2
Department of CSIT, ITER, S‘O’A University, Deemed to be University,
1 Introduction
The control of IP is a standard problem among all other problems in the control system
engineering field owing to non-minimum phase, nonlinear, and under-actuated char-
acteristics as reported in the references [1–3]. Additionally, the IP system exhibits a
significant number of industrial applications like self-balancing two-wheeled vehicles
or a kind of sag-way, guided missiles, rockets, intelligent robots, and other crane
models [4]. In this present study, an adaptive law of control approach is being tested
and analyzed to choose the IP system as it possesses a significant relevance with the
control dynamics. During the past three decades, so many control strategy techniques
are suggested and tested for control of AP of pendulum within the stable range. Time
discrete and switching PID control strategy is implemented in IP problems with vari-
able control gains based on the measured AP of the pendulum [5, 6]. However, the
optimal gain parameter setting, a lesser range of robust control and need of change of
gain setting with varying conditions are the major limitations to limit the real-time
application of these controllers. Among other projected robust control algorithms
implemented for limiting the AP of the pendulum are fuzzy [7], linear quadratic reg-
ulator (LQR) [8–10], neurocontrol [11], backstepping control [12], passivity control
[13], state feedback control [14], H-infinity ðH1 Þ control [15], sliding mode
(SM) control [16], fuzzy-sliding mode (FSM) control [17, 18], and BLQG control [19].
However, even if these control techniques are implemented effectively to control the
AP, the pendulum with better accuracy still fails to handle various constraints and
random change found in a trajectory of motion in the pendulum. These control tech-
niques are not completely insensitive to the disturbances and the uncertainties of the
model in spite of the improved performance. Hence, optimal control parameters setting
for better performance and avoiding slow response, the current work suggests an
alternative novel technique implementing the fractional order PID approach.
The FOPIDC approach concept leads to an enhanced control performance with
respect to robustness and delay of time compensation characteristics to counteract the
negative impact of associated errors, disturbances, and uncertainties. The stabilizing

https://doi.org/10.1007/978-981-15-2774-6_41
Stabilizing and Trajectory Tracking of Inverted Pendulum … 339
control law uðtÞ is developed based on the FOPIDC approach [20–22]. Application of
the FOPIDC to control the AP of pendulum results to ensure a better robust controller
in comparison with other contemporary well-accepted methods under both harmonized
and incompatible uncertainties.
This manuscript is structured as follows. Section 2 concisely illustrates the IP
system with mathematical details reflecting to its dynamic characteristics. Also, it
clearly demonstrates the simulation execution of the system on MATLAB environ-
ment. A detailed presentation on how the control technique is formulated and how it is
implemented for this problem is presented in Sect. 3. Comparative results of the
proposed approach with other published control techniques and the related analysis are
provided in Sect. 4. The concluding comments are summarized in the Sect. 5.
2.1 System Overview

The closed-loop model of the IP is depicted in Fig. 1a. The applied horizontal force
FðtÞ and vðtÞ is reflected as the process disturbance and the sensor noise, respectively,
in this study. The controller receives information about the AP of the pendulum as an
input to provide the optimal control force uðtÞ, and it balances the pendulum.
2.2 System Modeling

Figure 1b reflects the cart-pendulum model connected to a flexible cart rail with a free
swinging pole. The CP is being controlled by a DC motor. The nonlinear IP modeling
through the Newton’s law-based mathematical equations has been carried out. It is
presumed that the vertical force does not affect the CP, and the CP is disturbed by the
horizontal force FðtÞ based on the operation of the DC motor [23–25]. All the physical
activities of the IP system are expressed mathematically and specified by Eqs. (1) and
(2). All nomenclature and specifications for IP are shown in Table 1 and Table 2,
respectively. The Simulink model of the IP is established with respect to the Eqs. (1)
and (2) as displayed in Fig. 1c.

d2 xðtÞ d2 hðtÞ dhðtÞ 2 dxðtÞ
ðM þ mÞ 2
ml 2
cos hðtÞ þ ml sin hðtÞ þ b ¼ FðtÞ ð1Þ
dt dt dt dt
d2 ht d2 xðtÞ
ði þ Ml2 Þ mgl sin hðtÞ ¼ ml cos hðtÞ ð2Þ
dt2 dt2
2.3 Response of IP Model

There are four roots in IP system. One of them lies in right-hand side of the complex
plane. As a result, the system becomes unstable.
Fig. 1. a IP model with LQR; b schematic model of the IP system; c simulation model of the
nonlinear IP system
Table 1. IP model states and parameters

Symbol Description
FðtÞ Horizontal force acting on the cart
b ddxðtÞ
t
Frictional force acting on the cart
xðtÞ; ddxðtÞ ; d2 xðtÞ Cart position, cart velocity, and cart acceleration, respectively
t dt2
2
dhðtÞ d hðtÞ AP, AV, and angular acceleration of pendulum, respectively
hðtÞ; dt ; dt2
Table 2. IP model specification values

Symbol Physical meaning Value
M Cart mass 0.5 kg
m Pendulum mass 0.2 kg
i Inertia 0.3 kg.m2
g Gravitational acceleration 9.8 m/s2
b Frictional coefficient 0.1 N.s/m
l Pendulum length 0.3 m
This needs the design of an adaptive controller for improving the stability of the
system by means of shifting the roots into the left-hand side of the complex plane.
The IP system Simulinkmodel in the open-loop form is depicted in Fig. 1c. The IP
system consists of two inputs and four outputs. The control force uðtÞ and applied
horizontal force FðtÞ are the two inputs of the IP system. The CP and AP of the
pendulum are two outputs of the IP system. An uncontrolled system dynamics such as
AP of the pendulum and CP are being observed owing to the application of 1 N
impulsive horizontal force FðtÞ on the cart at the time t ¼ 1:0 sec. The uncontrolled
system dynamics are illustrated in Fig. 2a, b. Figure 2a, b illustrates the unstable
dynamics under various model uncertainties and disturbances. The unstable dynamics
can be reduced by applying the suitable control techniques. In this case, the AP of the
pendulum is the most essential outcome needs to be controlled within a stable range
through suitable control techniques, and CP is analyzed in order to view the motion
trajectory.
Fig. 2. a AP of pendulum with application of FðtÞ; b CP with application of FðtÞ
3 Control Algorithm
The FOPIDC algorithm is demonstrated in this section. The closed-loop system

response with respect to robustness, accuracy, and stability is analyzed. The control
specifications such as settling time ts , steady-state error ess , maximum overshoot OMax ,
and maximum undershoot UMax are also evaluated and examined with proper valida-
tion of the controller action.
3.1 FOPIDC Design

The IP model with fractional order PI a Dc controller is shown in Fig. 1a. In fractional
order PI a Db control, the error signal eðtÞ is used to generate control signal uðtÞ. The
transfer function (TF) of the proposed controller can be formulated as [26]:
Ki
TF ¼ Kp þ þ Kd sb ð3Þ
sa
where a and b are the fractional orders of the integrator and differentiator of the
proposed controller, respectively.Kp , Ki , and Kd are denoted for the proportional,
integral, and derivative gains of the proposed controller, respectively. The optimal
values of the FOPIDC parameters are considered based on the MATLAB/Simulink
environment and FOMCON toolbox as represented in Table 3. The structure of
FOPIDC is shown in Fig. 3, which is designed based on the Eq. (3).
Table 3. The optimal values of control parameters

Kp Kd Ki a b
18 6 1.5 0.7 0.9
4 Outcomes and Discussions
The response of the closed-loop IP with proposed FOPIDC is described in detail in this
section. The proposed control approach (FOPIDC) is compared with other popular
control algorithms to justify its enhanced performance.
4.1 Performance Analysis of IP System with FOPIDC

In this section, all physical activities of the closed-loop IP model with suggested
FOPIDC are examined under different conditions and the huge deviation of applied
horizontal force. The closed-loop system dynamics with 1 N impulsive horizontal force
FðtÞ at the time of 1.0 s are displayed in Fig. 4a, b. The outcomes clearly specify the
pendulum achieves the zero AP with less ts , and cart also attains the balance position
where the system is absolutely stable. To achieve the enhanced system response, the
required control force uðtÞ is generated by the suggested FOPIDC and demonstrated in
Fig. 5a.
Fig. 3. Simulink diagram of FOPIDC
Fig. 4. a AP of the pendulum with the application of FðtÞ based on FOPIDC; b CP with the
application of FðtÞ based on FOPIDC
Fig. 5. a Controlled signal uðtÞ of the FOPIDC; b AP of the pendulum with a deviation of
40% applied FðtÞ based on FOPIDC
4.2 Robustness of the FOPIDC

Figure 5b illustrates the AP of the pendulum with suggested FOPIDC under the huge
deviation of applied horizontal force FðtÞ. The time-domain outcomes under huge
deviation of applied forces show the enhanced performance of the closed-loop system
with FPIDC. Overall in each case, the pendulum achieves finally zero AP with less ts .
As indicated by the outcomes, the suggested FOPIDC robust performance under the
huge deviation of applied horizontal forces is much better.

The suggested FOPIDC approach is compared with other popular control approaches
such as PID, fuzzy, LQR, H1 , FSM, and BLQG justify its enhanced performance as a
controller. Figure 4(a) illustrates the effect of applied force in the AP of the pendulum
with the proposed FOPIDC approach. Table 4 presents a comparative analysis with
respect to ts ðsecÞ, OMax ðradÞ, UMax ðradÞ, noise (%), and ess ð%Þ. The effect of applied
force in the AP of pendulum applying different control approaches as PID, fuzzy, H1 ,
FSM, and BLQG is also demonstrated in Table 4 based on the references [5, 7, 9, 15,
18, 19], respectively. Similar working conditions are followed with the same level of
sensor noise in all control techniques application for comparison.
Table 4. Comparative result analysis related to AP of pendulum

Controller PID Fuzzy H1 FSM BLQG FOPIDC
[5] [7] [15] [18] [19] (Proposed)
Applied force (N) 1 1 1 1 1 1
ts ðsecÞ 2.8 3.0 2.1 1.8 1.9 0.37
OMax ðradÞ 0.1 0.226 0.107 0.042 0.041 0.037
UMaz ðradÞ 0.01 0.087 0.045 0.151 0.01 0.008
Noise (%) 10 10 5 5 5 5
ess ð%Þ 0 0 0 0 0 0
The AP of pendulum under 1 N impulsive horizontal force is tested. The corre-

sponding results are presented for the various control approaches along with the pro-
posed FOPIDC with respect to time-domain specifications such as OMax ðradÞ,
UMax ðradÞ, and ts ðsecÞ. The results signify the better controllability of the FOPIDC.
The simulation results also demonstrate the high noise and chattering elimination
capability with high robustness for the proposed approach. Overall, by looking to the
above comparative analysis, the findings of suggested approach advantages are the
higher accuracy and stability, more robustness, high noise and chattering elimination
capability, and better capability to handle uncertainty under various conditions and
huge deviation of road disturbance.
5 Conclusions
The manuscript proposed a novel control strategy (FOPIDC) to balance the pendulum.
To justify its enhanced performance, it has been applied and tested to control the
system dynamics of IP system within the stable range. Initially, the IP system is
modeled as a fourth-order state-space representation. Then, the proposed control
approach (FOPIDC) is designed and implemented. The comparative results clearly
reflect that the suggested FOPIDC arrives at better performance than the other control
approaches such as PID, fuzzy, H1 , FSM, and BLQG with respect to stability, relia-
bility, and robustness under various abnormal conditions and disturbances. The better
performance of the suggested approach in terms of improved accuracy and stability,
enhanced robustness, high noise and chattering elimination capability, and better ability
to control uncertainty justifies its real-time application.
References
1. Iqbal J, Ajwad SA (2016) Automating industrial tasks through mechatronic systems – a
review of robotics in industrial perspective. Tehnic ki vjesnik – Technical Gazette 23:917–
924
2. Ajwad SA, Iqbal J (2017) Role and review of educational robotic platforms in preparing
engineers for industry. Maejo Int J Sci Technol 11:17–34
3. Bettayeb M, Al-Saggaf U (2014) Stabilization of an inverted pendulum-cart system by
fractional PI-state feedback. ISA Trans 53:508–516
4. Iqbal J, Khan SG, Ukovic SC (2017) Nonlinear control systems -a brief overview of
historical and recent advances. Nonlinear Eng 6:301–312
5. Ghosh A, Subudhi B (2012) Brief paper – robust proportional-integral-derivative
compensation of an inverted Cart-Pendulum System: an experimental study. IET Control
Theory Appl 6(8):1145–1152
6. Wang C, Liu C, Fu W (2016) Design and simulation of IP system based on the fractional
PID controller. In: IEEE 11th Conference on Industrial Electronics and Applications
(ICIEA), pp 1760–1764
7. Magana ME, Holzapfel F (1998) Fuzzy-logic control of an inverted pendulum with vision
feedback. IEEE Trans Educ 41(2):165–170
8. Ozana S, Hajovsky R (2012) Design and implementation of LQR controller for inverted
pendulum by use of REX control system. In: IEEE International Conference on Circuits and
Systems. vol 1, pp 343–347
9. Kumar EV, Jerome J (2013) Robust LQR controller design for stabilizing and trajectory
tracking of inverted pendulum. Procedia Eng 64:169–178
10. Prasad LB, Gupta HO (2012) Modelling and simulation for optimal control of nonlinear
inverted pendulum dynamical system using PID controller and LQR. In: Sixth Asia
Modelling Symposium (AMS), pp 138–143
11. Pasemann F (1998) Evolving neurocontrollers for balancing an inverted pendulum. Netw
Comput Neural Syst 9:1–4
12. Deng L, Gao S (2011) The design for the controller of the linear IP based on backstepping.
In: International conference on electronic and mechanical engineering and information
technology (EMEIT). vol 6, pp 2892–2895
13. Jörgl M, Gattringer H (2013) Passivity based control of a cart with inverted pendulum. Appl
Mech Mater 332:339–344
14. Žilić T, Pavković D (2009) Modeling and control of a pneumatically actuated inverted
pendulum. ISA Trans 48:327–335
15. Lambrecht P, Vander G (1988) H-infinity control of an experimental inverted pendulum with
dry friction. IEEE Control Syst Mag 13(4):44–50
16. Wai RJ, Chang LJ (2006) Adaptive stabilizing and tracking control for a nonlinear inverted-
pendulum system via sliding-mode technique. IEEE Trans Ind Electron 53:674–692
17. Tao CW, Taur J, Chang J (2010) Adaptive fuzzy switched swing-up and sliding control
for the double-pendulum-and-cart system. IEEE Trans Syst Man Cybern B Cybern 40(1):
241–252
18. Chen CS, Chen WL (1998) Robust adaptive sliding-mode control using fuzzy modelling for
an inverted-pendulum system. IEEE Trans Ind Electron 45(2):297–306
19. Patra AK, Rout PK (2019) Backstepping linear quadratic gaussian controller design for
balancing an inverted pendulum. IETE J Res 1:1–15
20. Khalil HK (2002) Nonlinear systems, 3rd edn. Prentice Hall, Upper Saddle River, New
Jersey
22. Patra AK, Rout PK (2018) Backstepping sliding mode Gaussian insulin injection control for
blood glucose regulation in TIDM patient. J Dyn Sys Meas Control 140(9):091006
23. Irfan S, Razzaq MT, Iqbal J (2018) Advanced sliding mode control techniques for IP:
modelling and simulation. Eng Sci Tech Int J. https://doi.org/10.1016/j.jestch.2018.06.010
24. Ronquillo-Lomeli G, Ríos-Moreno GJ (2016) Nonlinear identification of inverted pendulum
system using Volterra polynomials. Mech Based Des Struct Mach 44(1):5–15
25. Kajita S, Fujiwara K (2003) Biped walking pattern generation by a simple three-dimensional
inverted pendulum model. Adv Robot 17(2):131–147
26. Bingul Z, Karahan O (2018) Comparison of PID and FOPID controllers tuned by PSO and
ABC algorithms for unstable and integrating systems with time delay. Optimal Control Appl
Meth 39(4):1431–1450
Inter-area and Intra-area Oscillation Damping
of Power System Stabilizer Design Using
Modified Invasive Weed Optimization
Mohammad Salik1, Pravat Kumar Rout1(&),

and Mihir Narayan Mohanty2
1
Department EEE, ITER, Siksha ‘O’ Anusandhan
2
Department of ECE, ITER, Siksha ‘O’ Anusandhan
1 Introduction
In power system operation, fault disturbances create low-frequency oscillations that

leads to many abnormal situations related to stability and protection. For acquiring a
smooth operation, it needs to provide adequate damping to avoid sustained oscillation
and may cause later on system outage. From the generator side, PSSs are responsible to
provide an adequate damping to these possible oscillations by providing supplementary
feedback stabilizing effects to the excitation system. At present times, the conventional
PI controller based on lead–lag compensator concept with constant gain parameters is
extensively used in the power system utilities [1]. However, this controller fails to offer
substantial stability to the system with frequently changing power system environment.
Secondly, these controllers are required to be tuned frequently by an optimal tuning
technique to set the right gain parameter values. In past few decades, although many
methods are suggested and tested, till this issue is an open research with possibility of
better methodology to cope up this problem. Among many successful ideas, the
approaches based on modern control theory are prominent and well accepted such as
optimal control, adaptive control, variable structure control, and intelligent control for
the PSS design [2–5].
Heuristic optimization techniques are successfully applied in many engineering
applications to optimally finding the gain parameters and system parameters. The major
advantages like simplicity in design, easy to apply, less possibility to be trapped in
local minima, non-derivative solution approach, and high possibility to arrive at global
minima are the major factors of attraction than the classical techniques. Classical
techniques are complex and have high chance of arriving at local minima. With respect
to PSS design, recently many heuristic search algorithms are successfully applied such
as GA [6], tabu search algorithm [7], and simulated annealing [8]. It has been found
that the results are promising and arrive at a better solution. But premature conver-
gence, degradation in efficiency, and reduction in search capability in difficult situations
force to apply new evolutionary computation techniques to solve complex nonlinear
power system problems.

https://doi.org/10.1007/978-981-15-2774-6_42
348 M. Salik et al.
A novel evolutionary optimization-based computation technique, known as inva-

sive weed optimization (IWO), has been suggested recently [9]. Basically, it is a new
numerical stochastic optimization algorithm. The mathematical formulation and
searching strategy are inspired from colonizing weeds. Weeds are plants characterized
with its inherent habits like: (1) Vigorous, (2) habitual practices and desirable natural
threats, and (3) harvesting of the full grown seeds are challenges in the cultivation. On
the other hand, weeds respond according to the climatic changes. Taking this adap-
tiveness of the weeds into account and efficient search algorithm can be formulated.
These optimization techniques titled as invasive weed optimization (IWO) mimic the
robustness, adaptation, and randomness of colonizing weeds. Considering the above
advantages, the researchers are able to formulate a simple in structure and robust in
searching strategy to arrive at global optimum value in many applications. The test
results reflect that the solutions from IWO are better in comparison with the other
methods.
In this study, the basic problem related to damping of oscillation with PSS design is
formulated in terms of an objective function. This objective function is considered to
decide the optimum parameter setting through a novel optimization technique based on
IWO. The operating constraints are taken care of during the simulation process to arrive
at a real-time feasible result. A single-machine infinite bus (SMIB) system has been
considered to test the efficacy of the proposed approach. Different test cases with
different disturbances and loading conditions are simulated, and the corresponding
analysis and nonlinear simulated results are presented to assess the effectiveness of the
proposed PSSs. In addition to that, comparative results of the proposed approach are
presented to that of GA-based PSS (GAPSS) and conventional approach-based PSS
(CPSS).
2 Problem Statement
2.1 System Modeling and Structure of PSS

The small-signal stability model of the PSS-integrated power system can be formulated
by a set of nonlinear differential equations and in a generalized form can be expressed as:
dX=dt ¼ f ðX; UÞ ð1Þ
where X denotes the vector of the related state variables defining the characteristic of
the system. U denotes the vector of input variables. A fourth-order machine model is
considered in this study. The differential equations used for the nonlinear time domain
simulation are mathematically represented as follows.
dX=dt ¼ x x0 ð2Þ
dd=dt ¼ ðPm Pe Þ=M ð3Þ

Inter-area and Intra-area Oscillation Damping of Power … 349

dEq =dt ¼ Efd ðXd Xdd ÞId Eq =Tdo ð4Þ

dEfd =dt ¼ Ka ðVref Vt VPSS Þ Efd =Ta ð5Þ
where d is the rotor angle; x is the rotor speed; Pm is the mechanical power; Pe is the
electrical power; Efd is the equivalent excitation voltage; d and q are the direct and
quadrature axes, respectively; Eq is the internal voltages behind the reactance Xd and
Xq; Ta is the regulator time constant; Tdo is the Time constant of excitation circuit;
and Ka is the regulator gain.
The proposed method is applied for designing a PSS in a single-machine infinite
bus (SMIB) system and a three-machine power system. Figures 1 and 2 show below
the schematic diagram of the SMIB and the three-machine power system, respectively.
Fig. 1. Single-machine power system
Fig. 2. Three-machine power system

350 M. Salik et al.
A conventional lead–lag PSS is considered in this test system looking to its wider
acceptance in recent times for similar types of study. It can be represented as:
sTw ð1 þ sT1 Þ ð1 þ sT3 Þ

VPSS ðiÞ ¼ Ki Dwi ð6Þ
1 þ sTw ð1 þ sT2 Þ ð1 þ sT4 Þ
where Dwi is the machine speed deviation from synchronous speed = (x − xo); VPSS
(i) is the PSS output signal at the ith machine; Tw is the washout time constant; T1, T3 are
the lead time constants; and T2, T4 are the lag fine constants. The time constants Tw, T2,
and T4 are generally taken from the standard prescribed values for the PSS. However, for
the better performance the stabilizer gain, Ki, and time constants T1i and T3i are to be
computed and set at their optimized values within the allowed prescribed limits.
2.2 Objective Functions

The PSS is designed with an objective to minimize the power system oscillations
through effective damping under a system disturbance to ensure an enhanced power
system stability and control. These oscillations are reflected in the system responses as
deviations from its normal conditions. The parameters such as power angle, rotor
speed, and line power are generally investigated for the analysis and system design.
The objective function is formulated according to the deviations occurred for these
parameters. In the present study, the two performance indices are considered such as
the settling time and overshoots. The objective function in terms of these indices is
formulated and considered for the optimum value setting by the searching techniques.
These indices can be mathematically represented as:
n Z sim
tt
X
PI1 ¼ ðtDwÞ dt ð7Þ
i¼1
t¼0
n Z sim
tt
X
PI2 ¼ ðDwÞ dt ð8Þ
i¼1
t¼0
where n is the number of machines and tsim is the simulation time.

Optimize F(x). Subject to
K min K K max
T1min T1 T1max
T3min T3 T3max
The standard limits considered in this study are as follows: The optimized
parameters are [0.001 − 50] for K and [0.06 − 1.6] for T1i and T3i. The time constants
Tw1, T2, and T4 are set as 5, 0.05, and 0.055, respectively. The objective function as
represented in Eq. (7) is taken to evaluate the fitness value to compute the optimal set
of solution vectors in the IWO and GA algorithm. The optimal set of parameters are
found out by tuning these parameters within their limits as PSS parameters like {Ki, T1i,
T3i, i = 1, 2, 3, … nPSS}.
3 Genetic Algorithm
3.1 Overview
The genetic algorithm (GA) is a stochastic-based searching scheme and applied suc-
cessfully in many real-world applications due to its capability to result in approximate
solutions. However, it is found in many cases, fails to arrive at global minima, is
trapped at local minima and is stagnant due to saturation effect. The idea of searching in
GA algorithm is formulated by the biological principles like survival of fittest and
evolution processes like inheritance, crossover, mutation, and selection.
Genetic algorithms are applied in computer simulation for many engineering
problems. Like other evolutionary optimization techniques, a set of random solution is
formulated named as population. Each member of the population represents one
specific solution. This mimics the chromosome structure in biology with genes as
variables. Generally, the mathematical representation of solutions is framed in a binary
string of 0 and 1 s. The encoding and decoding can be followed in various ways. With
every generation represented in simulation as iterations, the solutions vary optimally
and better results are obtained by generation after generation by following the major
steps like crossover, mutation, and selection. The objective function based on the
desirable result to be obtained is formulated, and this function is utilized to find out the
fitness value of all the solutions. The fitness vector helps to select the best value
considered as an optimal value for the problem it is used for. The stepwise procedure
followed in GA algorithm is explained in the next section.
3.2 Details About the Algorithm

3.2.1 Initialization
Initial population is randomly formulated by choosing many solutions as a member of
the population. The values are selected within the prescribed range. The size of the
population should be chosen suitably according to the complexity and type of the
problem. In this study as followed by many authors for similar GA application, here it
is taken as 20. Conventionally, even if the size of the population is randomly taken,
however it should cover the entire range of the searching apace.
3.2.2 Procedure of Selection

Throughout each succeeding epoch, a proportion of the existing population is preferred
by a selection method to breed a new generation. Individual solutions are selected
through a fitness-based process by finding the fitness vector by the objective function.
For a better selection, in this study the most popular and well-studied selection method
based on roulette wheel selection is taken into consideration.
352 M. Salik et al.
3.2.3 Reproduction
In this step, crossover and mutation operation are followed to generate new population
from the existing population. The crossover and mutation process are regulated by
crossover probability factor (Cpf) and mutation probability factor (Mpf). In this study,
the Cpf and Mpf are considered as 0.8 and 0.25, respectively, according to the standard
values taken in many papers. For producing new solution, two parent solutions are
chosen from the pool of selected solutions followed by a selection approach based on
the survival of fittest concept. The characteristics of parents as found in the child due to
inheritance, similarly here also the typical knowledge or factors are shared by the new
child solutions from their parent solutions expecting a better generation of solution in
comparison. The new solutions are added with the old solutions, and by the selection
process new better solutions are selected in proper population size for the next
generation.
3.2.4 Termination
In this study, the number of iterations has been taken as a key parameter for the
termination of the computation loop. The generation process is allowed to repeat
continuously until the termination criteria have been reached. Here, 100 is taken as the
maximum number of iterations to be executed.
4 IWO Algorithm
4.1 Overview
Among the stochastic search method, IWO algorithm is one of the novel approaches
ever proposed in the literature. This algorithm is formulated based on the natural
behavior of weed colonizing and used in the search domain for achieving global
optimal point in a complex system. As weeds adopt the environmental changes for
proper growth, IWO algorithm is capable to adapt the system complexity for ensuring a
complete search in the global search domain. This scheme reproduces seeds for
increasing the rate of growth of the population. Implementing the concept of fittest, the
colony tends to become dense, as it eliminates the one with lesser fitness and lesser
opportunity of life.
4.2 Outline of the Algorithm

4.2.1 Initialization
A population of initial solutions is being spread over the D-dimensional and NP
number of population problem space with random positions.
4.2.2 Process of Reproducing the Population Index

An element indexed under the given population is permitted to reproduce elements.
The rate of reproduction is indirectly regulated by the maximum and the minimum
fitness of the element present in the population. The rate of reproduction of the element
through a plant is linearly functioned in this algorithm. Therefore, the rate of
reproduction is directly dependent on the fitness factor of an individual plant. Hence, it

can be observed in Fig. 3 that the rate of reproduction is always linear and further
bounded by the maximum and minimum fitness factor. Eventually, the algorithm
ensures the growth of fittest element in the search domain. This forces to increase
convergence toward the group’s best value. As like other evolutionary algorithms, the
fittest element remains in the domain for the reproduction and thus weeds survive to
produce in the next generation for better generation. On the other hand, the worst ones
are eliminated due to less fitness value from the growth process.
Fig. 3. Reproduction strategy
4.2.3 Spatial Dispersal

This particular part of the manuscript describes the process of randomization and
adaptation in the IWO. Reproduced seeds are spread in the search domain, i.e., NP
D randomly, maintaining the mean value at zero. As a result, distributed seeds in the
search domain maintain an equal density factor. The parent plant is encircled by its own
reproduced seeds. Therefore, an uninterrupted search throughout the space is ensured.
In contrast, to maintain the speed and accuracy simultaneously, standard deviation
(SD) is allowed to vary between the initial and final values rinitial and rfinal, respec-
tively. In generation to generation, the procedure converges to the better fitness value
and at last arrives at an optimal value considered as the best possible value.
A nonlinear alteration has found to be arriving at better performance. Looking at
this idea, the r is allowed to vary in different iterations as in Eq. (9).
ðitermax iterÞn
riter ¼ ðrinitial rfinal Þ þ ðrfinal Þ ð9Þ
ðitermaxÞn
where itermax denotes the maximum number of iterations. The riter denotes the
standard deviation for the current iteration. Superscript n denotes the weighted average
modulation factor.
354 M. Salik et al.
4.2.4 The Process of Selecting the Fittest Zone of Element

According to the algorithm, if the parent of reproduced seeds is defined as offspring,
then it will be eliminated from the space in the next generation of iteration.
This is done by eliminating the low fitness value members. For limiting maximum
number of plants in a colony, a competition between the plants is followed based on
their fitness value. After consecutive iterations, the number of plants in a colony will
reach its maximum due to a combined and fast reproduction. There is a need, and it is
expected that the fitter plants have to reproduce more number of new solutions in
comparison with the undesirable plants. Once the maximum number of plants (Pmax)
reached the pre-specified number, a mechanism for elimination is followed in terms of
comparing their individual fitness values. When the maximum number of weeds in a
colony is reached, each weed is allowed to generate seed as specified in reproduction.
The produced seeds are then added with the population and allowed to spread over the
NP D-dimensional search area. However, they ranked together with their parents as
a colony of weeds. To avoid the allowed population to be crossed, the lower fitness
value members are eliminated. This is necessary to satisfy the constraint that the
population should be within the maximum allowable population in a colony. By this
process, the plants and offspring are ranked together, and subsequently the ones with
better fitness survive as the iteration moves forward. Later, they are only allowed to
replicate for producing new and better solutions or members of the population. With a
probability to result in a better solution even from the weeds with lower fitness value,
all the weeds are allowed to produce new offspring according to their fitness. If they are
comparatively better fitness value, they may survive in the colony and be a part of the
process in later iterations. However, the population control mechanism is always fol-
lowed even to their offspring at the end of a given run. This leads to better selection by
executing competitive exclusion.
5 Result and Performance Evaluation
For illustration and comparison purposes, a single-machine infinite bus system and a
three-machine system have been taken as two test systems as shown in Figs. 1 and 2,
respectively. A three-phase ground fault with duration of 0.1 s is considered as a fault
to study about the dynamic stability of the system incorporating the effect of PSS on
this situation. The test results demonstrate clearly that the proposed IWO-PSS out-
performs the GAPSS and CPSS.
Figure 4 shows how quickly the IWO technique optimizes the parameters of PSS
with less function value compared to the GA for same type of fault, loading condition,
and iterations. Figures 5, 6, and 7 show the control on restoration of bus voltage, the
active and reactive power flow in a line in pre- and post-fault condition. Also, the load
angle and speed variations of machine are shown in Figs. 8 and 9 in both the tech-
niques. Figures 10, 11, 12, and 13 show the intra-area and inter-area oscillations after
the fault and how the damping occurs very quickly due to correct design settings of
0.058
0.056
0.054
0.052 GAPSS
Iteration
IWOPSS
0.05
0.048
0.046
0.044
0.042
0 20 40 60 80 100 120 140 160 180 200
Error
Fig. 4. Convergence characteristics of the IWO and GA
1.2
CPSS
1.1 GAPSS
IWOPSS
1
Vt (p.u.)
0.9
0.8
0.7
0.6
0.5
0 1 2 3 4 5 6 7 8 9 10
t (sec)
Fig. 5. Terminal bus voltage response at node one
PSS parameters by GA and IWO for a three-multi-machine system. Only selected

figures are shown here due to lack of space. Simulation results in different conditions
show that the proposed IWO technique is better than the GA.
356 M. Salik et al.
2.5
CPSO
GAPSO
2 IWOPSS
Pe (p.u.)
1.5
0.5
0
0 1 2 3 4 5 6 7 8 9 10
t (sec)
Fig. 6. Active power flow in the transmission line
1.4
CPSS
1.2 GAPSS
IWOPSS
0.8
Qe (p.u.)
0.6
0.4
0.2
-0.2
0 1 2 3 4 5 6 7 8 9 10
t (sec)
Fig. 7. Reactive power flow in the transmission line
2.2
CPSS
2 GAPSS
IW0PSS
1.8
1.6
Delta (p.u)
1.4
1.2
0.8
0 1 2 3 4 5 6 7 8 9 10
t (sec)
Fig. 8. Rotor angle variation for single-machine case

6
CPSS
GAPSS
4 IWOPSS
Omega (p.u.)
-2
-4
-6
0 1 2 3 4 5 6 7 8 9 10
t (sec)
Fig. 9. Speed variations for single machine
Fig. 10. Speed deviations between generators one and two
Fig. 11. Speed deviations between generators one and three

358 M. Salik et al.
Fig. 12. Rotor angle variations between generators one and two
Fig. 13. Rotor angle variations between machines one and three
6 Conclusion
This paper investigates to find out a novel approach IWO for system stability
enhancement by quick damping through PSS design with optimal parameter setting.
The tuning of power system stabilizers using GA and IWO techniques is investigated
under different disturbances and loading conditions in a single-machine and a multi-
machine power system. Extensive time domain simulation results have been presented
to establish the efficiency and improved dynamic performance of the proposed stabi-
lization techniques. The IWO-PSS technique provides good damping characteristics
and outperforms GAPSS. So, the proposed stabilizer bears much potential for practical
implementation.
Appendix
For single-machine infinite bus system:

Xd = 1.5, Xdd = 0.2, Xq = 1, Tdo = 6, H = 6, Xe = 0.3, f = 50, Ke = 60,
Te = 0.4, P = 1, Q = 0.4, Vt = 1, Pref = 1
Load data: P11 = 0.8, Q11 = 0.2
References
1. Anderson PM, Fouad AA (1977) Power system control and stability. Lowa state university
press
2. Sauer PW, Pai MA (1998) Power system dynamics and stability. Prentice Hall
3. Larson RV, Swann DA (1981) Applying power system stabilizers, I, II and III. IEEE
Trans PAS 100(6):3017–3046
4. IEEE Working Group (1996) Annotated bibliography on power system stability controls
1986–1994. IEEE Trans Power Syst 88(4): 794–804
5. Abido MA (1999) A novel approach to conventional power system stabilizer design using
tabu search. Int J Electr Power Energy Syst 21(6): 443–454
6. Kennedy J (1997) The particle swarm: social adaptation of knowledge. In: Proceedings of 1997
IEEE international conference on evolutionary computation. In diaganpolis, IN, pp 303–308
7. Shi Y, Eberhant R (1998) Parameter selection in particle swarm optimization. In: Proceedings
7th annual conference evolutionary program, Mar 1998, pp 591–600
8. Kennedy JE, Shi R (2001) Swarm intelligence. Morgan Raufmann
9. Mehrabian AR, Locus C (2006) A novel numerical optimization algorithm inspired from
weed colonization. Ecol Inf, pp 355–366, Elsevier Science
IoT-Based Baby Swing Monitoring,
Alerting and Security System
Shuvendra Kumar Tripathy1(&) and Pradyut Kumar Biswal2

1
Trident Academy of Technology, Bhubaneswar, India
2
Department of Electronics Engineering, IIIT Bhubaneswar,
Bhubaneswar, India
pradyut@iiit-bh.ac.in
1 Introduction
This paper presents the smart cradle baby swing prototype model. There is immense
issue the working parents are confronting in their day-to-day life. They are not able to
pay appropriate time for their infants. Babies and infants need 24 7 attention, which
is highly difficult for working parents. It is awkward for them to always look out for
their infant while doing their work or errands. So, we have structured a basic frame-
work which helps the parents in dealing with their infant. This framework proposes a
basic voice identification system which can be connected essentially for planning a
gadget with capacity to recognize an infant’s cry and naturally turning on baby rest
music. This model takes care of the issue of time and vitality of such guardians. The
model of the cradle contains humidity and temperature sensor, moisture sensor, load
cell, voice recognizer and other electronic gadgets associated with Arduino Uno and
Raspberry Pi. The cradle machines are working with coordination of sensors and the
processors.
2 Literature Survey
Many child care baby swing frameworks are available; however, dominant part of the
framework uncommonly fails due to the absence of proper technology. This paper
presents a style of a child observance framework supported by the GSM network.
A model is proposed and created which supplies a solid and affordable child recog-
nition framework that may assume a significant job in giving higher child care [1],
individuals nowadays are by all accounts consistent in inclination with their work, and
they will most likely be unable to screen the tyke constantly. So as to come over this
trouble, the youngster checking framework is built up that chips away at the guideline
of image processing. The framework comprises of high-goal camera to identify the
development of the youngster. Here, the picture preparing is utilized to examine the
skin tone of the kid. Arduino controller can store the caught movements. This con-
troller transmits the data to move the robot by utilizing a servo engine. The robot moves
alongside the tyke because of the movement caught. Be that as it may, the frameworks’

https://doi.org/10.1007/978-981-15-2774-6_43
IoT-Based Baby Swing Monitoring, Alerting and Security System 361
reaction time is extremely high. The fundamental intention of this paper is to keep the
youngster safe and move the live data by a message to that specific parent cell phone
through GSM [2]. The author presents a structure of a baby checking system reliant on
the GSM composed. A model is made which gives a reliable and profitable baby
watching system that can expect a critical activity in giving better infant youngster care.
This system screens fundamental parameters, for instance, body temperature, beat rate,
moistness condition, advancement of an infant youngster [3]. The baby’s health con-
dition is continuously checked and also communicated with live video in mobile
application using IoT [4] as well as wearable technology [5]. The author also proposed
a system where PIR sensor, temperature sensor, pressure sensor, voice sensor, light
sensor are being used inside the smart room to monitor baby using WSN technology
[6]. In 2018, “IoT-based child and women security” paper was recommended to light
upon the dislodge youngster away from home. The model is made up of GSM and GPS
self-ruling utilizing cell phone [7].
3 Proposed System
From the above study, it is found that the drawback of the existing system is the
absence of real-time monitoring and controlling of the baby cradle. The system sends
the message signal if baby is crying remotely but not the real-time control. Also, all the
systems have used GSM module which needs SIM card, proper network for operation.
The major advantage of the proposed system is that it works on wireless sensor net-
work as well as on real-time platform. The block diagram of the proposed system is
shown in Fig. 1.
Fig. 1. Block diagram of IoT-based system

362 S. K. Tripathy and P. K. Biswal
4 Working of the Proposed System
The working of the proposed system is explained through the flowchart shown in
Fig. 2. In the proposed system, temperature and crying sound of the baby are always
monitored through two modules, i.e., temperature sensor and voice module. If baby
crying is detected by the voice module, then the cradle automatically starts to swing
continuously for 5 min with a recorded audio playing in their parent’s tone and
one SMS as well as voice alert of “baby crying” will be sent to the parent’s mobile.
Fig. 2. Flowchart of IoT-based system

Parents also can remotely monitor as well as control the cradle by sliding the slider
which is available in the mobile application. Also, when the temperature increases
beyond the threshold level, then the fan or the air conditioner automatically turns on
without involving any human interference.
Moisture sensor checks the wetness of the bed. If the bed is found to be wet, then a
voice alert will be sent to the parent’s mobile. Sometimes, it may happen due to busy
work inside the home without parent knowledge, and unauthorized person may enter
and try to steal the baby. So, to prevent the theft of the baby by the unauthorized
person, there is a load cell below the cradle bed so that if the weight decreases below a
certain level then a buzzer starts working.
All the sensor and input modules are being connected to Raspberry Pi through
Arduino Uno using Firmata software. Arduino will be the interface between all sensor
and Raspberry Pi. Camera is directly connected to the Raspberry Pi. All analog data are
collected by Arduino Uno, then transferred to the Raspberry Pi through USB port and
stored in a database using WSN technology.
A Web application as well as an android-based application is developed through
which users (parents) can screen and control the temperature of the room by making fan
on or off according to the room temperature where the child is kept and also can control
the swing of the cradle through a slider.
5 Implementation of the System
In this proposed framework, Python programming language is used. Python is a

translated state programming language for broadly useful programming. Apache server
with MySQL database is utilized in this task. The client’s parameters gained by
Raspberry Pi are put away on a server persistently every 2 min, and the equivalent is
transferred to the Web site page which is made utilizing PHP script which makes it
comprehensively open. Furthermore, security is given by giving the login id and secret
phrase using which the approved individual can login to the page and access client’s
parameters from anyplace.
The hardware requirements are: Raspberry Pi, Arduino Uno, sound sensor (voice
module), sound recorder and play system, NPK sensor, a USB camera and a SD card.
The various softwares that are used to develop the proposed system are: New Out
Of Box Software (NOOBS), mobile application, HTML and PHP, MQTT protocol,
Firmata protocol, Python and Node-RED.
6 Outcome of the System
The software NOOBS can be installed in the SD card of the Raspberry Pi. Using VNC
server, we can access Raspberry Pi window in our desktop. PuTTY is an exceptionally
valuable application that is utilized to interface with sequential ports and Secure Shell
364 S. K. Tripathy and P. K. Biswal
(SSH) to Raspberry Pi. A static IP is assigned to the Raspberry Pi through which we

can remotely open the Web page.
A Web page as well as mobile app is created to get the real-time information of the
child using HTML language and PHP method. Data are stored in the database using
MySQL database system. In our case, the Node-RED dashboard window will be
opened on the IP address of 192.168.1.102 with the port number 1880. The mobile app
view of baby swing monitoring, alerting and security system is shown in Fig. 3.
Fig. 3. Mobile app view of the system
Through the slider, the parents can control the cradle rotation by sliding the finger
on the mobile screen. As it can be seen from the front screen of the mobile app, the
electronic devices in the room can be controlled remotely by the parents depending on
the suitability of the situation.
7 Conclusion
In this paper, a framework is proposed and a real-time system is implemented for real-
time monitoring, controlling and automation of the baby swing cradle from the remote
location. The major advantage of the proposed system is that instead of using a GSM
module, a SMS-based notification is designed using IoT technology. The whole system
works on WSN technology. The parents can monitor and control various electronic
devices that are placed in the room for well-being and comfort of the child. Parents can
control the swing of baby cradle by controlling slider which is available in the mobile
application; also, a theft alarm system is proposed using a simple and efficient tech-
nology. This system can be very much useful for the working parents; even if they are
in office, they can take care of their baby.
References
1. Naiknaware S, Gajare D, Vibhute S, Kenjale S, Kulkarni S (2017) Review on baby
monitoring system. IJCSN Int J Comput Sci Netw 6(2), Apr 2017
2. Clenisha M, Sandra A, Sajin P, Sam BB (2018) Survey on automated child monitoring system
using image processing. Int J Pure Appl Math 118(24)
3. Patil SP, Mhetre MR (2014) Intelligent baby monitoring system. ITSI Trans Electr Electron
Eng (ITSI-TEEE), 2014
4. Rajesh1 G, ArunLakshman R, Hari Prasad L, ChandiraMouli R (2014) Baby monitoring
system using wireless sensor networks. ICTACT J Commun Technol, Sept 2014
5. Sharma O, Lewis SN, Telang U, Almeida LD, Lewis LDS (2017) Design of a bluetooth
enabled health monitoring system for infants using wearable technology. J Adv Res Dyn
Control Syst, vol 15
6. Kaur T, Mittal M, Singh H (2018) The baby monitoring room prototype model using IoT.
Int J Adv Res Sci Eng 07(6), Apr 2018
7. Budebhai M (2018) IoT based child and woman safety. Int J Comput Sci Mob
Comput IJCSMC 7(8), Aug 2018
Effective ECG Beat Classification and Decision
Support System Using Dual-Tree Complex
Wavelet Transform
Nihar Kar1, Badrinarayan Sahu1, Sukanta Sabut2,

and Santanu Sahoo1(&)
1
Department of ECE, Faculty of Engineering, Siksha ‘O’ Anusandhan,
santanusahoo@soa.ac.in
2
School of Electronics Engineering, KIIT, Bhubaneswar, Odisha 751024, India
1 Introduction
In worldwide, the main cause of death is due to cardiovascular diseases. In a recent

report by WHO, 17.1 million deaths are claimed due to CVD every year [1]. The
irregular electrical activities of the heart are called cardiac arrhythmia. Due to blockage
in coronary, artery heart attack occurs. The lack of blood and oxygen supply affects the
heart muscle which may die if the blockage is not treated immediately. Change in
lifestyle, smoking habits, unhealthy diet, obesity, diabetes, hypertension, use of alcohol
may increase the death cause of CVDs in worldwide [2]. Early detection of cardiac
arrhythmia may improve the death rate due to cardiac arrest by appropriate analysis of
ECG signals. The ECG signal gets corrupted due to the interference such as power line
interference, baseline wander, and many other biological artefacts [3]. Several methods
such as empirical mode decomposition [4], adaptive filtering [5], mathematical mor-
phology [6], and weighted averaging filter [7] are reported for eliminating these noises.
L. Smital et al. have reported wavelet wiener filtering techniques for eliminating noise
related to broadband myopotentials (EMG) in ECG signals [8]. In preprocessing stage,
the P and T attenuate from noises by smoothing techniques [9]. Proper analysis of ECG
signal requires accurate detection of QRS complex which is defined as the depolar-
ization of the right and left ventricles of the human heart and contraction of the large
ventricular muscles. Several methods have been reported for the detection of QRS
complexes. Discrete wavelet transform (DWT) mostly used to detect the accurate
features from ECG signal, where the signal decomposes into different coarseness and
coefficients which contains sufficient information of the processed signal in the sub-
bands [10, 11]. Multiscale feature of WT [12] can detect the P-, T-wave, and QRS
complex with a detection error rate 99.8% in the presence of baseline wander. To
evaluate the performance of QRS complexes, P-wave and T-wave, multiresolution
wavelet transform with optimum coefficients is used where the author declares an
accuracy of 97% for detecting heart rate, 96% for detecting P-wave, 95% for QRS
complex, and 98% for T-wave [13]. Martinez et al. proposed a technique to find
positive maximum and negative minimum of peaks using dyadic wavelet [14].

https://doi.org/10.1007/978-981-15-2774-6_44
Effective ECG Beat Classification and Decision Support System … 367
Extracting the good features is the main key point of signal classification. In a recent
work for the detection of QRS complex, Joao et al. [15] present an advanced approach
using first derivative, Hilbert transform, wavelet transforms, and adaptive thresholding
methods. Author claims 99.5% error rate in the PhysioNet arrhythmia database using
moving average computing method for the detection of QRS complexes [16]. In a
recent work, using adaptive thresholding-based multiresolution wavelet transform is
used to detect the peaks and valleys of a signal, S. Banerjee et al. claimed 99.8% of
sensitivity and 99.6% of positive predictivity using MIT-BIH arrhythmia database;
however, using PTB ECG database claimed 99.84% and 99.98%, respectively [17].
Based on the power spectrum of decomposition, multiresolution wavelet analysis is
used for selecting detail coefficient for the detection of QRS complex with 99.87% of
Se, 99.79% of Pp, and detection error rate of 0.34% [18]. Junior et al. in 2016 [19]
worked on real-time data to perceive QRS complex using redundant discrete wavelet
transform (RDWT) which achieves a detection rate of 99.32% in detecting QRS
complexes. Using statistical, geometric, and nonlinear HRV features, Jovic et al.,
proposed an efficient and accurate classification method for detecting cardiac disorders
[20]. Dong et al. [21] proposed a method based on deterministic learning model for
efficient ECG beat classification using radial basis function (RBF). Author claims an
average accuracy of 97.78 and 97.21% for global and patient-adapting beat classifi-
cation. Acharya et al. [22] proposed an effective method for automatic detection of
myocardial infarction in recorded ECG signals by implementing a convolutional neural
network (CNN) algorithm. Using feature selection algorithm, the author claimed an
accuracy of 93.53 and 95.22% with and without noise, respectively. Dokur et al. [23]
proposed a method which consists of Fourier and wavelet features and classified for
ECG beat classification. The features are searched with divergence values and classi-
fied 10 types of ECG beat with success of 97% using genetic algorithm-based neural
network. Over the last decades, many researchers are working in analysis of ECG
signal for automatic detection of cardiac arrhythmias with the use of computer tech-
nology. The computer-aided diagnosis system is a mimic to clinician and a reliable tool
in the application of detecting cardiac diseases in routine and long-term monitoring of
ECG signals. The major steps include noise elimination, detection of R-peaks,
extraction and formulation of important feature and Classification of cardiac arrhyth-
mias are given in Fig. 1.
Fig. 1. Overview of ECG beat detection system

368 N. Kar et al.
In this paper, we have used an application of dual-tree complex wavelet transform

to get the accurate features from the processed signal. Signal is first preprocessed by
DT-CWT up to four levels and using informative features, and five types of ECG beats
were classified from 48 recorded ECG signals taken from MIT-BIH arrhythmia data-
base. Temporal, morphological and statistical are the most important features contain
various clinical information are used as an input to the K-NN classifier to classify the
arrhythmia beats. The performance of the classifier was calibrated in terms of sensi-
tivity (Se), specificity (Sp), and accuracy (Acc).
2 Methodology
2.1 MIT-BIH Arrhythmias Database

In this work, MIT/BIH arrhythmia database is used for classification of arrhythmia
beats. The database comprises of 47 records of 2 channel ECG signals each of thirty
minutes period obtained from 47 subjects between 1975 and 1979 at the BIH
Arrhythmia Laboratory [24]. The database has 25 men of age 32–89 years, and 22
women aged 23–89 years which consists of 116, 137 number of QRS complexes.
The ECG signals were sampled at 360 samples per second with 11-bit resolution over a
10 mV range and band pass-filtered at 0.1–100 Hz. For timing and beat class infor-
mation, the data were verified by independent cardiologist for annotation. Normal (N),
left bundle branch block (LBBB), right bundle branch block (RBBB), atrial premature
beats (APBs) and premature ventricular contractions (PVCs) beats of data were
selected from MIT-BIH arrhythmia database for analysis. The first channel signals
taken from database were processed using the algorithm implemented and tested using
MATLAB version 2016.
2.2 Dual-Tree Complex Wavelet Transform

The concept of dual-tree complex wavelet transform is introduced in the year 1998 by
Kingsbury. Two real DWTs were employed in dual-tree CWT, where the real and
imaginary parts of the transform are given by first and second DWT. The poor
directional selectivity and shift variance drawbacks of DWT are eliminated by dual-tree
CWT. Figure 2 shows that the signal x(n) is implemented using DT-CWT framework
where two DWTs are in parallel to the same data. The implementation is much
expensive because for an N point signal, it gives 2N DWT coefficients [25]. A DT-
CWT can be demarcated as:
Wðl; mÞ ¼ WðlÞWðmÞ ð1Þ

Fig. 2. Dual-tree CWT consistent to three levels
WðlÞ and WðmÞ are complex wavelets.
WðlÞ ¼ Wa ðlÞ þ jWb ðlÞ ð2Þ
ðmÞ ¼ Wa ðmÞ þ jWb ðmÞ ð3Þ
where Wh ðxlÞ and Wh ðmÞ are real coefficient of upper and lower filter banks. The
overall equation becomes:
Wðl; mÞ ¼ ½Wa ðlÞ þ jWb ðlÞ½Wa ðmÞ þ jWb ðmÞ ð4Þ
The sets of coefficients used in DT-CWT in level 1 and level 2 are given in Table 1.
Table 1. Filter coefficients used in first and second levels

Level 1 Level 2
Tree A Tree B Tree C Tree D
Coeff1 Coeff2 Coeff1 Coeff2 Coeff1 Coeff2 Coeff1 Coeff2
0.000 0.000 0.000 0.000 0.011 0.000 0.011 0.000
0.088 0.011 −0.088 −0.011 0.011 0.000 0.011 0.000
0.088 0.011 0.088 0.011 −0.088 −0.088 −0.088 −0.088
0.696 0.088 0.696 0.088 0.088 −0.088 0.088 −0.088
0.696 0.088 0.696 0.088 0.696 0.696 0.696 0.696
0.088 −0.696 0.088 −0.696 0.696 −0.696 0.696 −0.696
0.088 0.696 −0.088 0.696 0.088 0.088 0.088 0.088
0.011 0.088 0.011 −0.088 −0.088 0.088 −0.088 0.088
0.011 0.088 0.011 −0.088 0.000 0.011 0.000 0.011
0.000 0.000 0.000 0.000 0.000 −0.011 0.000 −0.011
370 N. Kar et al.
2.3 K-Nearest Neighbor (K-NN) Classifier

The K-NN classifier is the supervised learning process that classifies the signal of
unknown instances relating to the known according to the distance/similarity function.
The K-NN is a robust versatile and classifier that is often used in complex classification
problems. In the classification setting, by the appropriate distance function two similar
instances are assigned to the same class. Its similarity function is defined according to a
distance metric between two data points; the Euclidean distance of the given metric is
given by
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uX
u n 2
dpq ¼ xpj yqj ð6Þ
j¼1
The classifier performance depends on the distance metric and K number of

neighbors. Training and classification phase of the classifier consist of multidimen-
sional feature space and class labels, respectively. In classification phase, K is a con-
stant that assigns a label which is nearest to the query point.

The classification accuracy of the cardiac abnormalities is highly depending upon the
information retains in the features. For ECG beat classification, different features were
obtained from one cardiac cycle. Accurate detection of Rloc could help to determine the
features required for the classifier. In this work, we used the features such as temporal,
R-R interval, and morphological features. The statistical features like mean, median,
and variance are also considered for classification. Different features such as R-R in-
tervals, RR-pre, and RR-post are used as temporal features, and QRS duration is the
heartbeat features, minimum, maximum, mean, and standard deviation of real and
complex wavelet coefficients of DT-CWT which were considered as statistical features
used as the input to the classifier for the detection of five types of arrhythmia beats.
All the recorded signal is initially de-noised with bandpass filters and then processed
for signal analysis at preprocessing stage. Due to the irregularity in heartbeats and
noises, it is quite a tough work to diagnose the correct arrhythmia conditions of the
heart. A dual-tree complex wavelet transform-based method has been proposed in this
work to extract discriminant features where the signal is decomposed up to four levels
to collect some temporal and morphological features to classify ECG arrhythmia beats.
The evaluated result shows the better accuracy and sensitivity in K-NN classifier.
Figure 3 shows the acquired signal and its preprocessing stages. In Fig. 4, different
arrhythmia beats are presented. The measure of performance of the classifier in terms of
Se, Sp, Pp, and accuracy is evaluated and presented in Fig. 5. The combined feature set
with K-NN classifier is tested over tenfold cross validation, and it is observed higher
performance is achieved at tenfold with Se (%) of 96.45, Pp (%) of 99.34, Sp (%) of
96.16, and Acc (%) of 98.92 for detecting the arrhythmia beats. The details of result are
shown in Table 2.
1.05 1
Amplitude (mV)
Amplitude (mV)
0.95 0.5
0.9
0.85 0
0.8
0.75
-0.5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Samples Samples
(a) Acquired ECG signal of rec. no. 111 (b) Denoised signal
1
Amplitude (mV)
0.5
-0.5
0 100 200 300 400 500 600 700 800 900 1000
Samples
(c) Detection of R-peaks
Fig. 3. Preprocessing stages of the signal
1 1
0.8
0.8
Amplitude (mV)
Amplitude (mV)
0.6
0.6
0.4
0.4 0.2
0.2 0
-0.2
0
-0.4
-0.2 -0.6
-0.4 -0.8
0 100 200 300 400 500 600 700 800 900 1000 1400 1450 1500 1550 1600 1650 1700 1750 1800 1850 1900
Samples Samples
Normal LBBB
1
0.6
0.8
Amplitude (mV)
0.6 0.4
0.4
0.2
0.2
0 0
-0.2
-0.4 -0.2
-0.6 -0.4
-0.8
-1 -0.6
1400 1500 1600 1700 1800 1900 2000 3000 3500 4000 4500 5000 5500 6000 6500 7000
Samples APB
RBBB
Fig. 4. Types of arrhythmia beats

372 N. Kar et al.
(a) Performance of Se and Pp of the classifier (b) Performance of Sp and Acc of the classifier
Fig. 5. Experimental result of the classifier over tenfold cross validation
Table 2. Classifier performance for ECG beat detection over tenfold cross validation
Fold No. Se (%) Pp (%) Sp (%) Acc (%)
2 84.76 96.83 83.28 95.09
3 84.80 96.84 83.37 95.42
4 84.79 96.84 83.47 96.64
5 92.95 98.67 91.67 97.75
6 93.17 98.75 92.38 97.91
7 93.17 98.79 92.68 97.98
8 93.58 98.79 92.55 97.99
9 93.94 98.86 92.94 98.09
10 96.45 99.34 96.16 98.92
4 Conclusion
This work presents a method for processing and extracting of features from ECG signal
using DT-CWT and classifying different heartbeat abnormal conditions using K-NN
classifiers. With various informative features the projected method confirmed the
superiority in terms of higher accuracy for classifying normal, LBBB, RBBB, PVC,
and APB beats. The algorithm reached 98.92% of accuracy using K-NN classifier. The
results suggest that the dual-tree CWT-based features with K-NN classifier can be used
effectively in the decision-making process of cardiac abnormalities. The machine
learning techniques given in this work can be effectively used in diagnosis and
decision-making process of accurate and timely detection of abnormal events in the
ECG signals so such patients can be treated early in critical care to save life. For real-
time ECG signal analysis, a deep learning-based classifier may be used in the future to
improve the accuracy and reduce computational time.
References
1. WHO (2015) Cardiovascular diseases (CVDs) (Internet). (Cited 25 Dec 2015). Available
from http://www.who.int/mediacentre/factsheets/fs317/en/
2. Aje TO, Miller M (2009) Cardiovascular disease: a global problem extending into the
developing world. World J Cardiol 1:3–10
3. Rai HM, Trivedi A, Shukla S (2013) ECG signal processing for abnormalities detection
using multi-resolution wavelet transform and artificial neural network classifier. Measure-
ment 46:3238–3246
4. Xing H, Huang M (2008) A new QRS detection algorithm based on empirical mode
decomposition. Conf Proc Bioinf Biomed Eng
5. Almenar V, Albiol A (1999) A new adaptive scheme for ECG enhancement. Sig Process
3:253–263
6. Zhang F, Lian Y (2009) QRS detection based on multiscale mathematical morphology for
wearable ECG devices in body area networks. IEEE Trans Biomed Circuits Syst 4:220–228
7. Momot A (2009) Methods of weighted averaging of ECG signals using Bayesian inference
and criterion function minimization. Biomed Signal Process Control 2:162–169
8. Smital L, Vítek M, Kozumplík J, Provazník I (2013) Adaptive wavelet wiener filtering of
ECG signals. IEEE Trans Biomed Eng 60:437–445
9. Kohler BU, Hennig C, Orglmeister R (2002) The principles of software QRS detection.
IEEE Eng Med Biol Mag 21:42–57
10. Addison PS, Watson JN, Clegg GR, Holzer M, Sterz F, Robertson CE (2000) Evaluating
arrhythmias in ECG signals using wavelet transforms. IEEE Eng Med Biol 19: 104–109
11. Fahoum AS, Howitt I (1999) Combined wavelet transformation and radial basis neural
networks for classifying life-threatening cardiac arrhythmias. Med Biol Eng Comput
37:566–573
12. Li C, Zheng C, Tai C (1995) Detection of ECG characteristic points using wavelet
transforms. IEEE Trans Biomed Eng 42:21–28
13. Pal S, Mitra M (2010) Detection of ECG characteristic points using multi resolution wavelet
analysis based selective coefficient method. Measurement 43:255–261
14. Martinez JP, Almeida R, Olmos S, Rocha AP, Laguna P (2004) A wavelet-based ECG
delineator: evaluation on standard database. IEEE Trans Biomed Eng 51:570–581
15. Madeiro JP, Cortez PC, Marques JA, Seisdedos CR, Sobrinho CR (2012) An innovative
approach of QRS segmentation based on first-derivative, Hilbert and wavelet transforms.
Med Eng Phys 34:1236–1246
16. Chen SW, Chen HC, Chan HL (2006) A real-time QRS method based on moving averaging
incorporating with wavelet denoising. Comput Method Programs Biomed 82:187–195
17. Banerjee S, Gupta R, Mitra M (2012) Delineation of ECG characteristic features using
multiresolution wavelet analysis method. Measurement 45:474–487
18. Bouaziz F, Boutana D, Benidir M (2014) Multiresolution wavelet-based QRS complex
detection algorithm suited to several abnormal morphologies. IET Signal Proc 8:774–782
19. Junior EA, Valentim RADM, Brandao GB (2016) Real-time QRS detection based on
redundant discrete wavelet transform. IEEE Latin Am Trans 14:1662–1668
20. Jovic A, Bogunovic N (2011) Electrocardiogram analysis using a combination of statistical,
geometric, and nonlinear heart rate variability features. Artif Intell Med 51:175–186
21. Dong X, Wang C, Si W (2017) ECG beat classification via deterministic learning.
Neurocomputing 240:1–12
374 N. Kar et al.
22. Acharya UR, Fujita H, Oh SL, Hagiwara Y, Tan JH, Adam M (2017) Application of deep
convolutional neural network for automated detection of myocardial infarction using ECG
signals. Inf Sci, pp 190–198
23. Dokur Z, Olmez T, Yazgan E (1999) Comparison of discrete wavelet and Fourier transforms
for ECG beat classification. Electron Lett, vol 35
24. Moody GB, Mark RG (2001) The impact of the MIT-BIH Arrhythmia database. IEEE Eng
Med Biol Mag 20:45–50
25. Kingsbury N (2010) Complex wavelets for shift invariant analysis and filtering of signal.
Appl Comput Harmon Anal 10: 234–253
Classical 2D Face Recognition: A Survey
on Methods, Face Databases, and Performance
Evaluation
Manoj Kumar Naik and Aneesh Wunnava(&)
Department of ECE, ITER, Siksha O Anusandhan, Bhubaneswar,

Odisha 751030, India
{manojnaik,aneeshwunnava}@soa.ac.in
1 Introduction
The face recognition is a special category of pattern recognition where the input is a
face image. The face recognition has started in the early and mid- 1970s, which is used
to measure the features between the faces or the face profile. During the 1980s to
1990s, face recognition has treated as not active or not passive. From the early 1990s,
face recognition is gaining pace. The reason for the gaining pace in the face recog-
nition: more research project funding for security and surveillance; availability of new
generation real-time hardware; the ability of neural networks as a classification and
adaptation tolls, etc. So, to get an optimal feature vector in the linear or nonlinear
approach is a challenge in the face recognition problem. There is no concrete evidence
of the best method to find the optimal feature vectors in the linear face recognition
approach.
This section briefly introduces the face recognition. The survey on face recognition
methods has stated in Sect. 2. In Sects. 3 and 4, a brief survey on face databases and
statistical performance evaluation techniques have been discussed. At last, the con-
clusion has been drawn in Sect. 5.
2 Face Recognition Methods
The face recognition is one of growing interest due to security reason. The face
recognition method based upon the 2D or 3D or combination of 2D and 3D face image.
For our limitation, we only discussed the 2D face recognition. To understand the face
recognition or face recognition system, we need to investigate various aspects such as
classification, classical 2D face recognition methods, face databases, and statistical
performance evaluation.
The backbone of any face recognition approach is to find the optimal feature vector
using different classification methods or matching the template between the training
and testing images. The research on face recognition stated somewhere 1970s, so we
are now standing in the fifth decade of FR research, so there are a huge number of

https://doi.org/10.1007/978-981-15-2774-6_45
376 M. K. Naik and A. Wunnava
journal, magazine, technical reports, books, face database, and face recognition systems
already developed by the commercial organization, government agencies, research
laboratories, and researcher. According to the face recognition review described in [1–
6], the methods can broadly classify into two groups. One is a holistic or appearance-
based or template-based approach, and the other one analytic or feature-based or
geometric-based approach [7].
• Holistic/Appearance-based/Template-based approach: In this approach, the feature
was extracted by considering the whole face image. These approaches directly
operate on the intensity value of the pixel of the facial image. This approach mostly
depends on the transform that reduces the high-dimensional feature space to a low-
dimensional feature vector because, in high-dimensional space, the probing a face
image for recognition is very difficult [8, 9]. The dimensionality reduction makes
the face recognition a computationally efficient and traceable.
• Analytic/Feature-based/Geometric-based approach: In this approach, the set of
geometric features is computed, such as eyes, nose, forehead, and mouth. The face
outline and the position of the geometric feature form a feature vector. The location
is used to find the geometrical relationship like area, distance, and angle, among the
different face images used for the identification or recognition [7, 10, 11].
Mostly all face recognition methods fall on the holistic or analytical approach. The
feature extracted using a holistic approach is known as the global feature, and the
feature extracted using the analytic approach has referred to as the local feature.
The well-known classical face recognition classification methods are principal
components analysis (PCA), linear discriminant analysis (LDA), independent com-
ponent analysis (ICA), hidden Markov model (HMM)-based, elastic bunch graph
matching (EBGM), support vector machine (SVM), and intrinsic discriminant analysis
(IDA), etc. The evolution and development of these classification methods have dis-
cussed below.
2.1 Principal Component Analysis (PCA)

The feature vector extraction for classification was the backbone of face recognition,
where PCA has its significance. The PCA [12] used the well-known Karhunen–Loéve
(KL) expression [13–15]. The idea of PCA was first introduced in [16], and then by
[17]. Sometimes, the PCA is also known as Hotelling transform [18]. For the pattern
recognition problem, the PCA was first time introduced in [19] to represent face images
in terms of optimal coordinates. For the first time in [20], try to find the eigenfaces from
the face image using Karhunen–Loéve expansion. Then in 1991, Turk and Pentland
[21, 22] developed a face recognition system using eigenfaces that can locate and
identify face images. A more detailed study on computational and performance aspects
of the PCA algorithm has been described in [23, 24].
2.2 Linear Discriminant Analysis (LDA)

The LDA method is a class-specific method proposed by Etemad and Chellappa in [25]
in the year 1997 by considering the discriminant analysis [26, 27]. The LDA is also
Classical 2D Face Recognition: A Survey on Methods, Face … 377
known as Fisher’s linear discriminant (FLD) [26]. Then, in the same year 1997, [28]
proposed a new face classification image known as Fisherfaces by implementing
dimension reduction technique PCA in FLD. One common problem in face recognition
is the “small sample size” (SSS) problem, which can be overcome by perturbation or
transform modification. To overcome the SSS problem in LDA-based face recognition,
a method new LDA (NLDA) has proposed in [29]. Most of the researchers believe that
LDA outperforms PCA, but according to [30], PCA can outperform the LDA when the
training data set is small. A new direct LDA (DLDA) algorithm has proposed in [31] to
overcome the singularity problem in LDA; the diagonalization procedure has modified
to make computationally efficient. A new direct fractional-step LDA (DF-LDA)
method was proposed in [32], which improves the LDA performance considering the
removal of the SSS problem and improving the classification ability. A new algorithm
PCA plus LDA has been proposed in [33] that performs LDA in PCA transformed
space to overcome the singular problem in LDA. The PCA generally gives linear
features, and kernel PCA (KPCA) [34] gives a nonlinear feature. Then, the researcher
tries to develop complex LDA-based combine Fisherfaces algorithm known as com-
plex Fisherfaces [35] by fusing the feature extracted from complex LDA and integrated
feature of PCA and KPCA. Some more variant LDA algorithm has improved LDA
(ILDA) [36], regularized LDA (R-LDA) [37], soft label-based linear discriminant
analysis (SL-LDA) [38], uncorrelated linear discriminant analysis (ULDA) [39],
orthogonal linear discriminant analysis (OLDA) [39], and perturbation LDA (PLDA)
[40] in related extraction of the face feature vector.
2.3 Independent Component Analysis (ICA)

The independent component analysis (ICA) [41, 42] is an extension of the PCA [12],
which imposes independence up to the second-order and defines orthogonal directions.
The two different architectures have discussed in it: one face images have considered as
random variables, and in another one pixel of the face image are treated as random
variables. An enhanced independent component analysis coined as Fisher independent
component analysis (FIDA) was proposed in [43] by augmenting Fisher linear dis-
criminant (FLD) analysis [26–28] in the generic ICA [42].
2.4 Hidden Markov Model (HMM)

The hidden Markov model (HMM) [44] has also referred to as Markov sources or
probabilistic functions of Markov chains [45–48]. The HMM model has successfully
applied in the field of speech processing [44]. For the first time, HMM is used for face
identification in [49] to extract features and segment the face image automatically.
Then in [50], a new face recognition approach based on HMM was discussed, which
extracts a two-dimensional discrete cosine transform (2D-DCT)-based feature for face
recognition. Again, to reduce the complexity of HMM-based face recognition, the
researcher in [51] proposed an HMM-based face recognition, in which the observation
vector to characterize the different state of HMM is obtained using Karhunen–Loéve
transform (KLT) [15, 19–22, 27].
2.5 Elastic Bunch Graph Matching (EBGM)

The elastic bunch graph matching (ENGM) based face recognition algorithm [52] that
implement bunch graph a data structure that serves the general representation of a set of
complex Gabor wavelet coefficients [53] known as jets of a small set of individual
faces, and also employed the object adapted graphs to find the correspondence between
face images. Some more researchers proposed face recognition uses EBGM in [54–56].
2.6 Intrinsic Discriminant Analysis (IDA)

The face image is classified into three components: common facial differences, indi-
viduality difference, and interpersonal difference. By considering all these, a new
dimensionality reduction technique has proposed as intrinsic discriminant analysis
(IDA) in [57]. The face recognition using the IDA algorithm was known as intrinsic
faces.
The researcher also presents many more face recognition methods. Some of them
are Gabor-Fisher classifier (GFC) [58], kernel eigenfaces [59], kernel Fisherfaces [59],
kernel direct discriminant analysis (KDDA) [60], locality preserving projections
(LPP) [61, 62], local binary patterns (LBP) [63], locally linear embedding (LLE)
[64–66], isometric feature mapping (Isomap) [67], and locality sensitive discriminant
analysis (LSDA) [68] algorithms.
3 Face Databases
The human faces are three-dimensional (3D) structures. So, when they record it on the
digital camera, it will transform into two-dimensional (2D) coordinates. So, the
appearance of the facial image affected by some factors like pose, illumination, facial
expression, occlusion, age, and facial hair. To validate a face recognition algorithm,
one should consider all these factors. For this reason, the researcher required some
standard databases that will be recognized by the researcher community for compar-
atively evaluate algorithms. Some of these database are Olivetti Research Ltd
(ORL) database of faces [69, 70], AR face database [71, 72], face recognition tech-
nology (FERET) [73–76], Yale face database [77], extended Yale face database B
(B +) [78], the Carnegie Mellon University (CMU) pose, illumination, and expression
(CMU-PIE) database [79–81], labeled faces in the wild (LFW) database [82, 83],
plastic surgery face database [84], University of Manchester Institute of Science and
Technology (UMIST) face database [85], labeled Wikipedia faces (LWF) [86, 87],
labeled faces in the wild-a (LFW-a) [88], YouTube makeup (YMU) face image [89],
virtual makeup (VMU) face image [89], US Adult Faces Database [90], makeup in the
“wild” (MIW) face database [91], Indian movie face database (IMFDB) [92], facescrub
database [93], and sibling database [94].
4 Statistical Performance Evaluation
The face recognition algorithm had better been evaluated to study the statistical per-
formance. In the statistical evaluation, for matching of the probe image to the gallery
image, one requires a similarity measure. So, the most common algorithm uses a
similarity measure as a nearest-neighbor classifier to validate the identity of faces. The
similarity measures are the distance function between the feature vectors. Some of
these are baseline (L1), Euclidean (L2), angle, Mahalanobis, L1 + Mahalanobis,
L2 + Mahalanobis, and angle + Mahalanobis distance [23]. For the performance
evaluation, the researcher specifies a CMC curve [75, 76, 95]. Sometimes, the
researcher also used the ROC curve for system performance evaluation, which is a
special case of the CMC curve [75, 76, 95].
5 Conclusion
The paper briefly discussed the development of the classical 2D-based face recognition
method existing in the literature, available databases, and various metrics related to
performance evaluation. From the above, it clearly stated that the outperformer clas-
sification methods are PCA and LDA for robustness and simplicity. The above
methods have been proposed based on the still face databases, so more challenging in
the future is how they can be applied to real-time simulations.
References
1. Chellappa R, Wilson CL, Sirohey S (1995) Human and machine recognition of faces: a
survey. Proc IEEE 83:705–740
2. Perronnin F, Dugelay J-L (2003) An introduction to biometrics and face recognition. In:
IMAGE’2003: learning, understanding, information retrieval
3. Zhao W, Chellappa R, Phillips PJ, Rosenfeld A (2003) Face recognition: a literature survey.
ACM Comput Surv 35:399–458
4. Kong SG, Heo J, Abidi BR, Paik J, Abidi MA (2005) Recent advances in visual and infrared
face recognition - a review. Comput Vis Image Underst 97:103–135
5. Abate AF, Nappi M, Riccio D, Sabatino G (2007) 2D and 3D face recognition: A survey.
Pattern Recogn Lett 28:1885–1906
6. Beham MP, Roomi SMM (2013) A review of face recognition methods. Int J Pattern Recogn
Artif Intell 27:1356005
7. Brunelli R, Poggio T (1993) Face recognition: features versus templates. IEEE Trans Pattern
Anal Mach Intell 15:1042–1052
8. Beyer K, Goldstein J, Ramakrishnan R, Shaft U (1999) When Is “Nearest Neighbor”
Meaningful? In: Beeri C, Buneman P (eds) Database Theory — ICDT’99, vol 1540.
Springer, Berlin, Heidelberg, pp 217–235
9. Sirovich L, Kirby M (1987) Low-dimensional procedure for the characterization of human
faces. J Opt Soc Am A 4:519–524
10. Bruce V (1999) Identification of human faces. In: 1999 Seventh international conference
on (Conference Publication No. 465) image processing and its applications, vol 612,
pp 615–619
11. Goldstein AJ, Harmon LD, Lesk AB (1971) Identification of human faces. Proc IEEE
59:748–760
12. Jolliffe IT (1986) Principal cornponent analysis. Springer, New York
13. Karhunen K (1946) Uber lineare methoden in der wahrscheinlichkeits-rechnun. Ann Acad
Sri Fennicae ser A1 Math Phys 37
14. Fukunaga K (1972) lntroduction to statistical pattern recognition. Academic, New York
15. Fukunaga K, Koontz WLZ (1970) Application of the Karhunen Loeve expansion to feature
selection and ordering. IEEE Trans Comput C-19:311–318
16. Pearson K (1901) On lines and planes of closest fit to systems of points in space.
Philosophical Magazine 2\
17. Hotelling H (1933) Analysis of a complex of statistical variables into principal component.
J Educ Psychol 24 (1933)
18. Gonzalez RC, Wintz PA (1987) Digital image processing. Addison-Wesley, Reading, MA
19. Watanabe S (1965) Karhunen-Loeve expansion and factor analysis theoretical remarks and
applications. In: 4th Prague Conference Information Theory
20. Kirby M, Sirovich L (1990) Application of the Karhunen-Loeve procedure for the
characterization of human faces. IEEE Pattern Anal Mach Intell 12:103–108
21. Turk MA, Pentland AP (1991) Face recognition using eigenfaces. In: Proceedings CVPR’91,
IEEE computer society conference on computer vision and pattern recognition, pp 586–591
22. Turk MA, Pentland AP (1991) Eigenfaces for recognition. J Cogn Neurosci 3:71–86
23. Moon H, Phillips PJ (2001) Computational and performance aspects of PCA-based face
recognition algorithms. Perception 30:303–321
24. Yambor WS, Draper BA, Beveridge JR (2000) Analyzing PCA-based face recognition
algorithms: eigenvector selection and distance measures
25. Etemad K, Chellappa R (1997) Discriminant analysis for recognition of human face images.
J Opt Soc Am A 14:1724–1733
26. Fisher RA (1936) The use of multiple measurements in taxonomic problems. Ann Eugenics
7:179–188
27. Duchene J, Leclercq S (1988) An optimal transformation for discriminant and principal
component analysis. IEEE Trans Pattern Anal Mach Intell 10:978–983
28. Belhumeur PN, Hespanha JP, Kriegman D (1997) Eigenfaces vs. Fisherfaces: recognition
using class specific linear projection. IEEE Trans Pattern Anal Mach Intell 19:711–720
29. Chen L-F, Liao H-YM, Ko M-T, Lin J-C, Yu G-J (2000) A new LDA-based face recognition
system which can solve the small sample size problem. Pattern Recogn 33:1713–1726
30. Martinez AM, Kak AC (2001) PCA versus LDA. IEEE Trans Pattern Anal Mach Intell
23:228–233
31. Yu H, Yang J (2001) A direct LDA algorithm for high-dimensional data with application to
face recognition. Patttern Recogn 34:2067–2070
32. Kostantinos JL, Plataniotis KN, Venetsanopoulos AN (2003) Face recognition using
LDA-based algorithms. IEEE Trans Neural Networks 14:195–200
33. Yang J, Yang J-Y (2003) Why can LDA be performed in PCA transformed space? Pattern
Recogn 36:563–566
34. Schlkopf B, Smola A, Muller K-R (1998) Nonlinear component analysis as a kernel
eigenvalue problem. Neural Comput 10:1299–1319
35. Yang J, Yang J-Y, Frangi AF (2003) Combined fisherfaces framework. Image Vis Comput
21:1037–1044
36. Xiao-Yuan J, Zhang D, Yuan-Yan T (2004) An improved LDA approach. IEEE Trans Syst
Man Cybern Part B Cybern 34:1942–1951
37. Lu J, Plataniotis KN, Venetsanopoulos AN (2005) Regularization studies of linear
discriminant analysis in small sample size scenarios with application to face recognition.
Pattern Recogn Lett 26:181–191
38. Zhao M, Zhang Z, Chow TWS, Li B (2014) Soft label based linear discriminant analysis for
image recognition and retrieval. Comput Vis Image Underst 121:86–99
39. Ye J (2005) Characterization of a family of algorithms for generalized discriminant analysis
on undersampled problems. J Mach Learn Res 6:483–502
40. Zheng W-S, Lai JH, Yuen PC, Li SZ (2009) Perturbation LDA: learning the difference
between the class empirical mean and its expectation. Pattern Recogn 42:764–779
41. Jutten C, Herault J (1991) Blind separation of sources, part 1: an adaptive algorithm based on
neuromimetic architecture. Sign Proces 24:1–10
42. Comon P (1994) Independent component analysis, a new concept? Sign Proces 36:287–314
43. Keun-Chang K, Pedrycz W (2007) Face recognition using an enhanced independent
component analysis approach. IEEE Trans Neural Networks 18:530–541
44. Rabiner L, Huang B (1993) Fundamentals of speech recognition. Prentice-Hal, Englewood
Cliffs, NJ
45. Baum LE, Petrie T (1966) Statistical inference for probabilistic functions of finite state
Markov chains. Ann Math Stat 37:1554–1563
46. Baum LE, Eagon JA (1967) An inequality with applications to statistical estimation for
probabilistic functions of Markov processes and to a model for ecology, pp 360–363
47. Baum LE, Petrie T, Soules G, Weiss N (1970) A maximization technique occurring in the
statistical analysis of probabilistic functions of Markov chains. Ann Math Stat 41:164–171
48. Baum LE (1972) An inequality and associated maximization technique in statistical
estimation for probabilistic functions of Markov processes. In: Inequalities III, Academic
Press pp 1–8
49. Samaria F, Young S (1994) HMM-based architecture for face identification. Image Vis
Comput 12:537–543
50. Nefian AV, Hayes MH, III (1998) Hidden Markov models for face recognition. In:
Proceedings of the 1998 IEEE international conference on acoustics, speech and signal
processing, vol 2725, pp 2721–2724
51. Nefian AV, Hayes MH, III (1998) Face detection and recognition using hidden Markov
models. In: Proceedings 1998 International Conference on Image Processing, ICIP 98, vol
141, pp 141–145
52. Wiskott L, Fellous JM, Kuiger N, von der Malsburg C (1997) Face recognition by elastic
bunch graph matching. IEEE Trans Pattern Anal Mach Intell 19:775–779
53. Daugman JG (1988) Complete discrete 2D gabor transform by neural networks for image
analysis and compression. IEEE Trans Acoustics Speech Signal Proces 36:1169–1179
54. Kela N, Rattani A, Gupta P (2006) Illumination invariant elastic bunch graph matching for
efficient face recognition. In: CVPRW’06 Conference on Computer Vision and Pattern
Recognition Workshop, pp 42–42
55. Pervaiz AZ (2010) Real time face recognition system based on EBGM framework. In: 2010
12th International conference on computer modelling and simulation (UKSim), pp 262–266
56. Hanmandlu M, Gupta D, Vasikarla S (2013) Face recognition using elastic bunch graph
matching. In: 2013 IEEE applied imagery pattern recognition workshop: sensing for control
and augmentation, pp 1–7
57. Wang Y, Wu Y (2010) Face recognition using Intrinsicfaces. Pattern Recogn 43:3580–3590
58. Chengjun L, Wechsler H (2002) Gabor feature based classification using the enhanced fisher
linear discriminant model for face recognition. IEEE Trans Image Proc 11:467–476
59. Yang M-H (2002) Kernel Eigenfaces vs. kernel fisherfaces: face recognition using kernel
methods. In Proceedings of the fifth IEEE international conference on automatic face and
gesture recognition, IEEE computer society, p 215
60. Lu J, Plataniotis KN, Venetsanopoulos AN (2003) Face recognition using kernel direct
discriminant analysis algorithms. IEEE Trans on Neural Networks 14:117–126
61. He X, Shuicheng Y, Yuxiao H, Niyogi P, Hong-Jiang Z (2005) Face recognition using
Laplacianfaces. IEEE Trans Pattern Anal Mach Intell 27:328–340
62. He X, Niyogi P (2003) Locality preserving projections. In: advances in neural information
processing systems
63. Ahonen T, Hadid A, Pietikainen M (2006) Face description with local binary patterns:
application to face recognition. IEEE Trans Pattern Anal Mach Intell 28:2037–2041
64. Saul LK, Roweis ST (2003) Think globally, fit locally: unsupervised learning of low
dimensional manifolds. J Mach Learn Res 4:119–155
65. Roweis ST, Saul LK (2000) Nonlinear dimensionality reduction by locally linear
embedding. Science 290:2323–2326
66. Ridder DD, Duin RPW (2002) Locally linear embedding for classification. Delft University
of Technology
67. Samko O, Marshall AD, Rosin PL (2006) Selection of the optimal parameter value for the
Isomap algorithm. Pattern Recogn Lett 27:968–979
68. Cai D, He X, Zhou C, Han J, Bao H (2007) Locality sensitive discriminant analysis. In:
Proceedings of the 20th international joint conference on Artifical intelligence, Morgan
Kaufmann Publishers Inc., Hyderabad, India, pp 708–713
69. http://www.cam-orl.co.uk/facedatabase.html
70. Samaria FS, Harter AC (1994) Parameterisation of a stochastic model for human face
identification. In: Proceedings of the second IEEE workshop on applications of computer
vision, pp 138–142
71. Martinez AM, Benavente R (1998) The ar face database
72. Li B, Li J, Tang K, Yao X (2015) Many-objective evolutionary algorithms: a survey. ACM
Comput Surv 48:1–35
73. Phillips PJ, Hyeonjoon M, Rizvi SA, Rauss PJ (2000) The FERET evaluation methodology
for face-recognition algorithms. IEEE Trans Pattern Anal Mach Intell 22:1090–1104
74. Vázquez D, Fernández-Torres MJ, Ruiz-Femenia R, Jiménez L, Caballero JA (2018) MILP
method for objective reduction in multi-objective optimization. Comput Chem Eng
108:382–394
75. Phillips PJ, Wechsler H, Huang J, Rauss PJ (1998) The FERET database and evaluation
procedure for face-recognition algorithms. Image Vis Comput 16:295–306
76. Rizvi SA, Phillips PJ, Hyeonjoon M (1998) The FERET verification testing protocol for face
recognition algorithms. In: Proceedings third IEEE international conference on automatic
face and gesture recognition, pp 48–53
77. http://cvc.yale.edu/projects/yalefaces/yalefaces.html
78. Wang H, Jin Y, Yao X (2017) Diversity assessment in many-objective optimization. IEEE
Trans Cybern 47:1510–1522
79. Sim T, Baker S, Bsat M (2002) The CMU pose, illumination, and expression (PIE) database.
In: Proceedings of fifth IEEE international conference on automatic face and gesture
recognition, pp 46–51
80. Sim T, Baker S, Bsat M (2003) The CMU pose, illumination, and expression database. IEEE
Trans Pattern Anal Mach Intell 25:1615–1618
81. Sedarous S, El-Gokhy SM, Sallam E (2017) Multi-swarm multi-objective optimization
based on a hybrid strategy. Alexandria Eng J 57(3):1619–1629
82. Huang GB, Ramesh M, Berg T, Learned-Miller E (2007) Labeled faces in the wild: a
database for studying face recognition in unconstrained environments
83. Meza J, Espitia H, Montenegro C, Giménez E, González-Crespo R (2017) MOVPSO: vortex
multi-objective particle swarm optimization. Appl Soft Comput 52:1042–1057
84. Singh R, Vatsa M, Bhatt HS, Bharadwaj S, Noore A, Nooreyezdan SS (2010) Plastic
surgery: a new dimension to face recognition. IEEE Trans Inf Forensics Secur 5:441–448
85. http://images.ee.umist.ac.uk/danny/database.html
86. Hasan MK, Pal CJ (2011) Improving alignment of faces for recognition. In: 2011 IEEE
international symposium on Robotic and Sensors Environments (ROSE), pp 249–254
87. Hasan MK, Pal C (2014) Experiments on visual information extraction with the faces of
wikipedia. In: AAAI 2014
88. Wolf L, Hassner T, Taigman Y (2011) Effective unconstrained face recognition by
combining multiple descriptors and learned background statistics. IEEE Trans Pattern Anal
Mach Intell 33:1978–1990
89. Dantcheva A, Cunjian C, Ross A (2012) Can facial cosmetics affect the matching accuracy
of face recognition systems? In: 2012 IEEE fifth international conference on Biometrics:
Theory, Applications and Systems (BTAS), pp 391–398
90. Bainbridge WA, Isola P, Oliva A (2013) The intrinsic memorability of face photographs.
J Exp Psychol Gen 142:1323–1334
91. Chen C, Dantcheva A, Ross A (2013) Automatic facial makeup detection with application in
face recognition. In: 2013 International Conference on IEEE Biometrics (ICB), pp 1–8
(2013)
92. Setty S, Husain M, Beham, P, Gudavalli J, Kandasamy M, Vaddi R, Hemadri V, Karure JC,
Raju R, Rajan B, Kumar V, Jawahar CV (2013) Indian movie face database: a benchmark for
face recognition under wide variations. In: 2013 Fourth National Conference on Computer
Vision, Pattern Recognition, Image Processing and Graphics (NCVPRIPG), pp 1–5
93. Hong-Wei N, Winkler S (2014) A data-driven approach to cleaning large face datasets. In:
2014 IEEE International Conference on Image Processing (ICIP), pp 343–347
94. Vieira T, Bottino A, Laurentini A, De Simone M (2014) Detecting siblings in image pairs.
Vis Comput 30:1333–1345
95. Grother P, Micheals R, Phillips PJ (2002) Face recognition vendor test performance metrics.
In: Kittler J, Nixon M (eds) Audio- and video-based biometric person authentication, vol
2688. Springer, Berlin Heidelberg, pp 937–945
Multi-level Authentication-Based Secure
Aware Data Transaction on Cloud Using
Cyclic Shift Transposition Algorithm
Prasanta Kumar Bal(&) and Sateesh Kumar Pradhan
Department of Computer Science & Application, Utkal University,

Bhubaneswar, India
sateesh.cs@utkaluniversity.ac.in
1 Introduction
A data “transaction server or transaction processing system (TPS)” is a software sys-

tem, or combination of hardware and software, that supports transaction processing. In
data transaction, computing task is divided into small, indivisible operations, called
transactions. Data center approach is used in cloud computing that comprises of
thousands of servers to provide services to the end users [1]. Cloud computing has
desirable properties, such as elasticity, scalability, and pay-per-use and fault tolerance.
As a promising service platform, its applications demand huge data storage. Cloud
application also requires data reliability that to be in a highly cost-effective manner [2].
Data storage is most popular and important service. Generally, independent institutions
check and certify the security requirements imposed by the users to cloud providers.
The cloud provider and cloud end user make formal agreement that includes the
security and privacy of the data stored and processed in the cloud. Recently, cloud
storage has received huge attractions in personal and business organizations because it
is convenient and economical [3].
The main drawback in cloud storage is information integrity and confidentiality. In
this paper, these issues are handled by encryption techniques for correctness, confi-
dentiality, and integrity tamper resistance, [4]. Internet is used to connect cloud users
having mobile phones, laptops, or fashionable desktops with the cloud platform. Cloud
services are usually classified into software as a service (SaaS), platform as a service
(PaaS), and infrastructure as a service (IaaS) like raw computing power or cloud
storage. A decent cloud security supplier can provide a scalable answer that detects
anomalies before they reach the “data center,” serving to security [5]. “A broad set of
policies, technologies, applications, and controls utilized to protect virtualized IP, data,
applications, services, and the associated infrastructure of cloud computing” is referred

https://doi.org/10.1007/978-981-15-2774-6_46
Multi-level Authentication-Based Secure Aware Data Transaction … 385
to as cloud security. In recent years, many encryption techniques are used for the above
purpose. They have Triple DES, Data Encryption Standard (DES), Advanced
Encryption Standard (AES), Twofish encryption, and Rivest–Shamir–Adleman
Encrypion (RSA). In this paper, we have used cyclic shift transposition (CST) algo-
rithm. The CST algorithm is basically a transposition algorithm obtained by shifting
operations [6].
Lot of researcher has been developing a secure data transaction on cloud. Among
them, some of the works are analyzed here; In 2017, Luna et al. [7] developed
quantitative reasoning about cloud security using service-level agreement. The tech-
nique proposed by them was to help improve the security in cloud data transaction.
Adoption framework was used toward achieving data security in cloud by Chang et al.
[6] in 2016. They experimented to reduce the injection of Trojans and viruses with
success rate of 97.43%. In 2016, Gupta et al. [8] analyzed implementation of cloud
ERP with reduced effect of compliance, network, and security. This paper proposed a
link between the concerns of cloud vendor and the CSFs that are categorized people,
and technological factors. Experimental result showed that the implementation of cloud
ERP was moderated. Wei et al. [9] in 2016 developed Revocable Storage Identity
Based Encryption (RS-IBE) for secure data sharing in cloud computing. Finally,
demonstrated result has advantages in efficiency and functionality. It takes more time to
take the conversion of cipher text in cloud computing; its main limitation is security in
cloud computing. Awad et al. [10] in 2018 proposed a chaotic searchable encryption
for mobile cloud storage in fuzzy keyword. Experimental results showed that the
proposed scheme was efficient and suitable for a secure searchable cloud storage
system. In 2018, Hu et al. [11] had developed a secure and verifiable access control
based on the Nth degree truncated polynomial ring (NTRU) cryptosystem for big data
storage in cloud. The design of a secured and privacy preserving for big data storage in
a cloud is an extremely challenging problem. It indicates that eligible user can be
prevented from cheating and reduce various attacks such as the collusion attack.
2 Proposed Methodology
As the demand for cloud computing is growing across the globe, data security and
privacy protection are also two rising issues in cloud computing. To avoid this prob-
lem, multi-level authentication-based security aware data transaction on cloud is pro-
posed. The proposed system consists of three modules, namely multi-level
authentication, data encryption, and data retrieval. The multi-level authentication
process is used to avoid the unauthorized person login process. The proposed
methodology is depicted in Fig. 1.
386 P. K. Bal and S. K. Pradhan
User Registration Stage
Enter the user data Select the Crop Stored on

(id, password) image Image server
Registered user Security stage
Transaction Encrypt the data Encrypted Data

data using CSTA
Request
Verification success and get the

authentication key from the user Data on cloud
User “B”
Fig. 1. Overall concept of the proposed methodology
2.1 Multi-level Authentication Process

To avoid the malicious attack and information loss, the authentication process is
mandatory. “It becomes easy for unauthorized users to modify the data without the
knowledge of the owner in a centralized environment. Hence, it is unavoidable for
security breach, thereby making it mandatory to secure the data using multi-level
authentication.” The multi-level authentication process consists of two stages, namely
registration and login.
Registration process:-
During the registration process, initially, the user opens the registration page. Then, the
user creates the user id Uid and password UPa . After Uid creation, user has to enter all
the information about the user. Then, user selects one image SI from the N number of
images available in the database Db . After the selection process, the selected image SI
is cropped. The cropping is done based on pixel position values. After the cropping
process, user information along with cropped image is stored on the cloud server. The
registration process is given in Fig. 2.
Users Server
User id
Password
Set of
images
Selected image
User id, password,

Cropped image
selected image and
cropped image
Fig. 2. Registration process
Login process:-
After the registration process, the user is able to upload the data on cloud. A user wants
to upload the data means user must login to the server. Using this process, we can avoid
the unauthorized person to access the data. In login process, initially, user enters their
id Uid and password Upa . The information is correct means the server displays n
number of images along with registered image. Among the displayed images, the user
selects one image SI which should be same as the registered image. If the selected
image is correct, then user crops the image and sends to the server. Finally, server
checks the cropped image with registered cropped image. If the information given by
the user is correct for the respective user name, the server will allow the user to access
the data otherwise, neglect the user request.
2.2 Data Security Using Cyclic Shift Transposition Algorithm (CSTA)

The main concern of cloud data owner is to store the data securely on cloud as data
confidentiality is an unavoidable task. Even though different algorithms are used for
sensitive data protection from the malicious users, they can easily access the cloud
storage framework. To avoid the problem in this paper CSTA method is proposed that
does not require any centralized authority such as key management system. In this
method, initially, the file is separated into N N matrix and so it consists of “n” blocks
size and its key size is 22N þ 1 . The CSTA algorithm consists of two stages, namely
encryption and decryption.
Encryption process:-
The encryption process is used to hide the original information. The proposed CSTA
performs partition and shifting operation, namely column shift, row shift, primary
diagonal shift, and secondary diagonal shift. The step-by-step process of encryption
process is explained below;
Step 1 Consider the input file Fi . Initially, we split the file Fi into N N matrix
format.
Step 2 Then, perform shift column (SC) operation. The SC is performed based on
Eq. (1).
0
Fr;c ¼ Fr þ shiftðr;Nb Þ mod Nb;c ð1Þ
Here, shiftðr; Nb Þ only depends upon key value. The key values vary from 0 to 9.
The key values represent the cyclic shift of the number of elements, and the modular
arithmetic is represented by mod operation.
Step 3 Perform shifting the row (SR) in a particular order. SR transportation can be
expressed in Eq. (2).
0
Fr;c ¼ Fr;c þ shiftðr;Nb Þ mod Nb ; ð2Þ
Step 4 Perform diagonal shift operation, i.e., shift the diagonal elements from top to
right bottom. This can be presented as follows:
0
Fr;c ¼ Fr þ shiftðr;Nb Þ mod Nb;c þ shiftðr;Nb ÞmodNb ð3Þ
Step 5 Perform secondary diagonal shift operation in particular order. This can be
expressed as follows;
0
Fr;c ¼ Fðr1Þ mod Nb;c:
Step 6 Then, find the output in a particular order.
0
Fr;c ¼ Fðc þ ðNb1 ÞÞ;c:
Step 7 Obtain the encrypted text by converting the output into ASCII format.
Step 8 Finally, we calculate the hash value along with a timestamp and store it in
the cloud.
Decryption process The decryption process is a reverse of the encryption process. If

any cloud user tries to retrieve the information, the cloud server decrypts the data into
original file and sends to the cloud user. The step-by-step process of decryption process
is explained below:
Step 1 Evaluate the hash value along with a timestamp and transfer it to the
receiver.
Step 2 Consider the output and convert the output into ASCII format to get the
encrypted file.
Step 3 After that, shift the row into a specific order.
Step 4 Shifting the column into a specific order
Step 5 Then, shift the matrix diagonally into the corresponding order.
Step 6 Perform secondary diagonal shift operation in corresponding order.
Step 7 Find out the decrypted file.
2.3 Query-Based File Retrieval

After the encryption process, the encrypted data are stored in a cloud. To avoid
malicious attack and information loss, the data are securely stored in the cloud. After
uploading, any user wants to access the data means they send a request to CSP. After
receiving request, the CSP checks the authentication process. If the user is already
registered, CSP sends the data; otherwise, it neglects the request.
The experimental results obtained from the proposed methodology are analyzed in this
section. Here, the implementation is done in JAVA with cloud simulator. The effec-
tiveness of the proposed methodology is analyzed in terms of different metrics, namely
memory usage, execution time, encryption time, and decryption time.
3.1 Experimental Results

The main objective of proposed methodology is to securely send the data transaction on
cloud. Here, two level securities are developed. The first level is in authentication stage,
and another one is data security stage. For authentication process, multi-level
authentication is developed which is used to avoid the unauthorized user login. For data
security, CSTA is developed which is securely transact the data. To prove the effec-
tiveness of the proposed methodology, here, we compare our proposed methodology
with different algorithms. The experimental results of proposed against existing is
given in this section.
Figure 3 represents the running time of the proposed method. The data size is
10,000 kb, the AES is 5637 s, and the DES is 4632 s. But the proposed CSTA (2) is
3905 is better than the existing methods. Because the “proposed model follows the
decentralized architecture, it does not depend on any third-party system.”
6000 AES DES CSTA
5000
4000
Running time (s)
3000
2000
1000
0
500 2500 5000 10000
File size (kb)
Fig. 3. Running time
To prove the effectiveness of the proposed authentication process, we compare our

method with a different authentication process. The performance of the proposed
methodology is analyzed in terms of successful authentication is given in Fig. 4. When
analyzing Fig. 4, our proposed method attains a better authentication rate. This is
because our proposed method has five mandatory fields, namely user id, password,
image selection, cropping, and encryption.
98%
Successful authentication (%)
96%
94%
92%
90%
88%
86%
84%
Password Pin Cropping Proposed
Methods
Fig. 4. Performance based on successful authentication
Figure 5 represents the encryption time of the proposed method. The data size is
5000, the AES is 1,911,324s, and the DES is 1,893,721s. But the proposed CSTA
1,857,651s is better than the existing methods. Because the proposed model follows the
decentralized architecture, it does not depend on any third-party system.
2500000
AES DES CSTA
Encryption time (s) 2000000
1500000
1000000
500000
0
500 2500 5000 10000
File size (KB)
Fig. 5. Encryption time
Figure 6 represents the decryption time of the proposed method. The data size is
10,000, the AES is 16,856,331s, and the DES is 1,620,511s. But the proposed CSTA
1,595,432 is better than the existing methods. Because the proposed model follows the
decentralized architecture, it does not depend on any third-party system.
1800000
AES DES CSTA
1600000
1400000
Decryption time (s)
1200000
1000000
800000
600000
400000
200000
0
500 2500 5000 10000
File size (kb)
Fig. 6. Decryption time
Figure 7 represents the memory size of the proposed method. The data size is
10,000, the AES is 27,632,481, and the DES is 26,947,263. But the proposed CSTA
26,446,232 is better than the existing methods. Because the proposed model follows
the decentralized architecture, it does not depend on any third-party system.
30000000
AES DES CSTA
25000000
Memory usage (bits)
20000000
15000000
10000000
5000000
0
500 2500 5000 10000
File size (kb)
Fig. 7. Memory size
4 Conclusion
In this paper, multi-level authentication-based secure data transaction has been pro-
posed. To avoid the unauthorized user login, multi-level authentication has been
developed. To enhance the security of the data, CSTA algorithm has been developed.
This method was fully secured; therefore, the third party cannot able to access the data.
To prove the performance of the proposed methodology, the proposed algorithm has
been compared with different algorithms. The security level of authentication process is
also analyzed. The result has clearly shown the effectiveness of the proposed
methodology.
References
1. Yan Z, Ding W, Yu X, Zhu H, Deng RH (2016) Deduplication on encrypted big data in
cloud. IEEE Trans Big Data 2(2):138–150
2. Li W, Yang Y, Yuan D (2016) Ensuring cloud data reliability with minimum replication by
proactive replica checking. IEEE Trans Comput 65(5):1494–1506
3. Jakóbik A, Palmieri F, Kołodziej J (2018) Stackelberg games for modeling defense scenarios
against cloud security threats. J Netw Comput Appl 110:99–107
4. Kandil A, El-Deeb H (2016) Exploration of application migration to cloud environment. In
2016 6th international conference-cloud system and big data engineering (Confluence) IEEE,
pp 109–114
5. Basu S, Pattnaik PK (2017) A consistency preservation based approach for data-intensive
cloud computing environment. In 2017 8th International Conference on Computing,
Communication and Networking Technologies (ICCCNT), IEEE, pp 1–5
6. Chang V, Ramachandran M (2016) Towards achieving data security with the cloud
computing adoption framework. IEEE Trans Serv Comput 9(1):138–151
7. Luna J, Taha A, Trapero R, Suri N (2017) Quantitative reasoning about cloud security using
service level agreements. IEEE Trans Cloud Comput 5(3):457–471
8. Gupta S, Misra SC (2016) Moderating effect of compliance, network, and security on the
critical success factors in the implementation of cloud ERP. IEEE Trans Cloud Comput 4
(4):440–451
9. Wei J, Liu W, Hu X (2016) Secure data sharing in cloud computing using revocable-storage
identity-based encryption. IEEE Trans Cloud Comput 1–1
10. Awad A, Matthews A, Qiao Y, Lee B (2018) Chaotic searchable encryption for mobile cloud
storage. IEEE Trans Cloud Comput 6(2):440–452 (2015), (Apr 2018)
11. Hu C, Li W, Cheng X, Yu J, Wang S, Bie R (2018) A secure and verifiable access control
scheme for big data storage in clouds. IEEE Trans Big Data 4(3):341–355
A Novel Graphical User Interface-Based
Toolbox for Optimization and Design
of Linear Antenna Array
Guru Prasad Mishra, Shibanee Dash,

and Saumendra Kumar Mohanty(&)
Department of ECE, FET, ITER, Siksha ‘O’ Anusandhan

saumendramohanty@soa.ac.in
1 Introduction
The performance of single-element antenna is limited. They cannot meet the gain and
radiation pattern requirements for any specific application. In many applications, it is
also required to design antennas that are suitable for long-distance broadband com-
munication [1]. This demanded antenna should be highly directive and also wideband.
A possible solution to obtain a high-gain, high bandwidth antenna is by using the array
antenna, which is essentially an assembly of several radiating antennas to form a single
antenna system in an electrical and geometrical configuration [2]. The performance of
the array increases with the number of elements in the array, but the cost and com-
plexity increase with higher number of elements. Nowadays, antenna arrays are more
prominent in wireless terminals, smart antenna, and they are widely used in applica-
tions like radar, sonar, and communications [3–5]. Before design, a perfect analysis is
required. In this regard, computational electromagnetics is essential for numerical
analysis and subsequent computer simulation and physical modeling.
Bossavit [6] proposed computational electromagnetics to model the interaction of
electromagnetic fields with the physical structure and environment. In [7], the rela-
tionship between field and circuit theory is examined. An extensive study has been
carried out on different analytic methods, which are used to analyze antenna structures
using computational electromagnetics. Such methods are the method of moments
(MoM) and finite-difference time-domain (FDTD). Abd-Alhameed et al. [8] investi-
gated and discussed the design of microstrip patch antennas using MoM. In [9, 10], the
multiple sweep method of moments (MSMM) are applied to the three-dimensional (3-
D) problems of a long dipole antenna with a reactive load in the center of each arm and
a TEM horn antenna with a bend.
Stochastic and metaheuristic optimization methods are now very common and have
many advantages over deterministic methods. These methods include genetic algorithm
(GA) [11–15], simulated annealing, differential evolution (DE), Tabu search (TS), PSO
[16–18], ant colony optimization (ACO), CS optimization, and many others. Further, as
global optimizations do not depend on initial conditions, and they handle discontinuous
and non-differentiable functions, so they are preferred over local optimization

https://doi.org/10.1007/978-981-15-2774-6_47
A Novel Graphical User Interface-Based Toolbox for Optimization … 395
techniques. Hence, we have used particle swarm optimization (PSO) and cuckoo search
(CS) in optimization of LAA. In [19], the study presents a multi-level optimization
(MLO) method applied in particle swarm optimization (PSO) to solve a 3-objective
optimization problem. PSO algorithm is introduced in [20] to optimize the element
spacing and lengths of a mutually coupled planar array antenna. Baskar has described a
method of using particle swarm optimization (PSO) algorithms to optimize the element
spacing and lengths of Yagi–Uda antennas in [21]. The radiation pattern and rSLL of a
42 elements linear array, with 20 optimization variables, have been compared using CS
and differential evolution [22]. So a comparative study is made between PSO and CS in
PAA design using MOM technique.
It is very useful as the actions in a GUI are usually performed through direct
manipulation of the graphical elements. Hence, we have gone through the literature or
process that involves designing the toolbox. In [23], it gives detail on a quick start with
guide, design process, and principles, GUI layout with guide, and some GUI examples.
2 Design Representation of LAA
As MoM considers mutual coupling among the elements of the antenna array, so it is
the most prominent method for our purpose. Here, the frequency of operation is in the
VHF band, i.e., 30–300 MHz. The antenna array consists of several linear dipole
elements and all are energized directly by a discrete feed as shown in Fig. 1.
This array is exclusively designed to operate as an end-fire array. The lengths and
diameters of elements, as well as their respective spacing, determine the optimum
characteristics. These types of antenna arrays are quite practicable to use because they
are lightweight, simple to build, low cost and provide moderately desirable charac-
teristics for many applications.
For this method, Pocklington’s integral equation was used to analyze the array
structure. So in this chapter, the sections describe the derivation of E and H fields from
integral equation. Also, the detailed steps considering in MATLAB to design LAA
using MoM will discuss in the next section. As the results from MATLAB need to be
validated in a globally accepted simulation software, here LAA is being designed in
CST microwave studio. The steps included in design at CST platform will also discuss
here in detail. Finally, in the last section, the outcomes from both platforms are mat-
ched and analyzed.
Fig. 1. 16-element dipole LAA

396 G. P. Mishra et al.
2.1 Integral Equation Moment Method

Here, the design is based on the integral equation for the electric field radiated by the
elements in the array, and it will be used to describe the complex current distributions
on all the elements and the corresponding radiation patterns. This method uses
Pocklington’s integral equation to analyze any structure by point matching numerical
methods. The number of discrete points must be sufficient in number to allow com-
puted data to match well with experimental results. This equation states that the total
field radiating in an unbounded space is generated by an electric current source. Mutual
coupling is also considered here, and there is no restriction on the number of elements.
Pocklington’s integral equation,
Zþ 2 jkR
l
@2 2 e
I ðz0 Þ þ k dz0 ¼ j4pxe0 Ezt ð1Þ
@z2 R
2l
For small diameter wires, the current on each element can be approximate:
where
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
R¼ ðx x0 Þ2 þ ðy y0 Þ2 þ ðz z0 Þ2
X
0
M
pz0
In ðz Þ ¼ Inm cos ð2m 1Þ ð2Þ
m¼1
ln
where
Inm complex current coefficient of mode m on element n.
ln corresponding length of the element.
Once the current distribution is found, the total field of the exciting array is
obtained by summing the contribution of each element. The far-zone electric field
generated by the m modes of the nth element-oriented parallel to z-axis is given by:
Ehn jxAhn ð3Þ
where
2 3
Zþ 2
ln
jkR
le 6 0 7
Ahn sin h4ejkðxn sin h cos / þ yn sin h sin /Þ þ In ejkzn cos h dz0n5 ð4Þ
4pr
l2n
So the total field is,
X
N
Eh ¼ Ehn ð5Þ
n¼1
These above equations are used in MATLAB to get the total field and their
respective parameters of an antenna. MATLAB implementation of the design of LAA
using MoM involves the following process. The number of elements, length, and their
respective separations is the input parameters to the structure code. Here the frequency
of operation is in VHF band (i.e., 30–300 MHz). So here k is taken as 1 m, and all the
dimensions are taken in terms of k. The radius of dipoles is taken as 0.001 k, as here
they are considered as small wires. After assigning all the above parameters as input,
the structure code starts from here. First, the position of the elements in an array is
calculated in y-coordinate parallel to z-axis. After placing them, a supply of 1 V is
given to the center of each dipole. Then current on each element is found out after
placing Eq. (2) in MATLAB code. To operate functions like finite integration and
summation, the functions are declared separately to reduce the complexity of the
structure code and there call back is used when necessary. Then, electric field is
calculated by using Eq. (5), and their respective parameters, like 3 dB beamwidth,
front-to-back ratio, directivity (D) and SLL, are found out. Similar steps are also carried
out for magnetic field and their respective parameter calculation. Finally, far-field plots
are visualized in both polar and Cartesian plots.
3 GUI-Based Toolbox Implementation in MATLAB
In this paper, a GUI has been created where one can design the LAA with and without
optimization. This GUI contains six push buttons that perform differently. By clicking
the first push-button, one can design an LAA of his own specification of any number of
elements, lengths, and their respective spacing. Push-button two and three perform the
design of optimized LAA with desired directivity and FSLL for single-objective and
multi-objective fitness functions, respectively, using PSO, whereas the fourth and fifth
button does the same process as second and third, using CS optimization technique.
Finally, a quit button is added to close the GUI.
The input parameters to the different push-button for the above GUI are:
I. The number of elements, lengths of antennas, and their respective spacing.
II. Number of elements, desired directivity, and number of iteration for the PSO
optimization process.
III. The number of elements, desired directivity, desired FSLL, and number of
iteration for PSO optimization process.
IV. Number of elements, desired directivity, and number of iteration for the CS
V. The number of elements, desired directivity, desired FSLL, and number of
iteration for the CS optimization process (Fig. 2).
Fig. 2. GUI for optimized LAA design
3.1 Steps for LAA Design in CST Microwave Studio

Before going to start design in CST, one must have to set the units. Units of dimen-
sions, frequency, temperature, and time are the point of concern here. Here, the
dimensions, frequency, and temperature are taken in mm, MHz, and ns, respectively.
Materials used to design the antenna must be declared then. Here in this work, copper
and vacuum materials are used to design the antenna array. Then the structure of the
LAA has to be created according to the array. Here 32 numbers of cylinders and 16
numbers of gaps are being designed for a 16-element LAA. According to this design,
the frequency range is set in between 0 and 500 MHz as the operating frequency is
300 MHz. The operating frequency must be greater than half of the frequency range.
After declaring all the parameters and structure, excitations are set to the gap between
the dipoles. Here to match the result with MATLAB, voltage source of 1 V is given to
all the elements. Field monitors are an important part of the design process. The user
can set E-field, H-field, or far-field with the operating frequency, whichever is needed
to analyze the structure. Here only far-field at 300 MHz is set for this purpose. Before
going to start the solver, all excitation ports are selected, and simultaneous excitation is
set, as a linear array is being designed in CST. There are many solver options like
transient solver, frequency solver, eigenmode solver, and TLM solver. Here transient
solver is started to solve the structure, and far-field results are obtained. Then the results
are analyzed and matched with the MATLAB results by choosing polar, Cartesian plots
and 2-D or 3-D plots.
4 Optimization Procedure
4.1 Implementation of PSO
Here, a 16-element LAA is optimized using PSO in MATLAB. Since there are only 16
lengths and 15 spacing along y-axis, total columns of the matrix are taken as 31 and 25
rows for 25 particles in optimization. Here, k is 1 m and the spacing is fractional
multiple of the values of the matrix parameter are taken from 0.6 to 1 for better results.
Cases are considered here. Case-I is for a single-objective fitness function, which
enhances the directivity only, and case-II considers both directivity and sidelobe level
into account. For the time, we have taken only ten iterations (in total 250 times) in the
4.2 Implementation of CS
Here we have taken 25 numbers of nests. Then the discovery rate of alien eggs is taken
as 0.25. We have restricted the l and d between 0.6l and l. With these limits, 25 21
random matrix is generated, similar to that of PSO.
4.3 Parameters of PSO and CS

To compare both optimizations, we have to keep some variables the same for both the
techniques and those are:
• Ddes = 21 dBi
• FSLLdes = −17 dB
• No. of particles or nests = 25
• No. of iterations = 10
• Tolerance limit = 2.
Multi-parameter single-objective fitness function (case-I):
First, we have designed a fitness function, whose only aim is to optimize D without
considering the other output parameters of the antenna and is given as,
F = |Ddes−D(y)|, where ‘Ddes’ stands for the desired directivity and y = {l, d}.
Multi-parameter multi-objective fitness function (case-II):
Here, both D and FSLL are being optimized with antenna lengths and their
respective spacing as input parameters to the process. The fitness function is given
as,
F = c1|Ddes − D(y)| + c2|FSLL(y) – FSLLdes|, where the weighting parameters c1
and c2 are taken both as 0.5. ‘Ddes,’ ‘FSLLdes,’ stands for the desired directivity and
desired FSLL and y = {l, d}.
Two different cases, as we discussed earlier, are considered during optimization. case-I
directivity optimization of the linear array without taking FSLL into account and case-
II a trade-off between D and FSLL. The single-objective (for case-I) results and the
multi-objective (for case-II) results are shown in the following table. The lengths and
spacing are given in units of k as we are considering our antenna directivity concerning
an isotropic antenna, so the unit is taken in dBi. As the maximum radiation of any field
pattern is kept to zero dB, here the values of SLL are expressed in negative. For case-I,
the D is 20.673 dBi for PSO and 20.048 dBi for CS. The corresponding FSLL is
(−5.27 dB) and (−5.05 dB) for PSO and CS, respectively. The time taken for con-
vergence is found to be 4663.417872 s for PSO and 333.162295 s for CS. For case-II,
the D is 19.332 dBi for PSO and 18.138 dBi for CS. The corresponding FSLL is
(−17.65 dB) and (−18.71 dB) for PSO and CS, respectively. The time taken for
convergence is found to be 3296.936648 s for PSO and 322.260406 s for CS. So if a
comparison is made between PSO and CS in terms of convergence time, CS converges
approximately 10 times faster than the PSO for both cases having the same number of
iterations and same fitness function (Table 1).
6 Simulation Results Obtained from MATLAB and CST

Microwave Studio
All the simulations are carried out using Intel(R) Pentium(R) CPU B960@ 2.20 GHz
processor with 4 GB RAM. The optimized performance parameters are estimated, and
the radiation patterns are verified in CST microwave studio. It is observed from the
figures and the table that radiation pattern obtained in MATLAB best approaches with
the pattern obtained from CST microwave studio.
The simulation results of the PSO optimized linear array are as shown below
(Fig. 3) (Table 2).
The simulation results of CS optimized linear array are as shown below (Fig. 4).
Table 1. Optimized parameters of 16-element linear array (case-I and case-II)
Case-I Case-II
PSO CS PSO CS
Element Length (in k) Spacing (in k) Length (in k) Spacing (in k) Element Length (in k) Spacing (in k) Length (in k) Spacing (in k)
1 0.9314 – 0.9782 – 1 0.9961 – 0.9828 –
2 0.9311 0.9021 0.9242 0.7841 2 0.9958 0.8779 0.9881 0.9737
3 0.9404 0.8282 0.9584 0.7593 3 0.9614 0.8923 0.9070 0.7285
4 0.9752 0.9448 0.9202 0.8225 4 0.9461 0.8664 0.9478 0.9553
5 0.9766 0.9178 0.9127 0.8088 5 0.9805 0.8931 0.9739 0.8567
6 0.9506 0.9164 0.9686 0.7486 6 0.9354 0.8818 0.9734 0.7049
7 0.9679 0.8637 0.9911 0.9681 7 0.9691 0.7280 0.9816 0.7388
8 0.9522 0.9104 0.9751 0.8321 8 0.9040 0.7895 0.9533 0.8780
9 0.9560 0.8371 0.9823 0.9011 9 0.9788 0.8522 0.9561 0.9165
10 0.9698 0.9096 0.9225 0.7631 10 0.9186 0.9134 0.9347 0.7111
11 0.9558 0.8115 0.9387 0.8413 11 0.9308 0.9981 0.9551 0.7529
12 0.9793 0.8747 0.9419 0.7150 12 0.9522 0.7504 0.9021 0.7697
13 0.9582 0.8559 0.9936 0.8488 13 0.9125 0.9069 0.9737 0.9145
14 0.9682 0.9229 0.9106 0.7809 14 0.9047 0.9554 0.9834 0.9312
15 0.9894 0.9248 0.9531 0.9053 15 0.9617 0.7209 0.9836 0.8768
16 0.9881 0.8404 0.9621 0.9244 16 0.9494 0.7708 0.9033 0.9619
D (dBi) 20.673 20.048 D (dBi) 19.332 18.138
FSLL ( dB) −5.27 −5.05 FSLL (dB) −17.65 −18.71
Time (s) 4603.41 333.16 Time (s) 3296.93 322.26
A Novel Graphical User Interface-Based Toolbox for Optimization …
401
(a) Matlab (case-I) (b) CST (case-II) (c) Matlab (case-I) (d) CST (case-II)
Fig. 3. E-theta versus h (at U = 90°) for free-space environment using MATLAB and CST
microwave studio for case-I and case-II, respectively
Table 2. Comparison of results from MATLAB and CST microwave studio

Case-I Case-II
PSO CS PSO CS
D FSLL D FSLL D FSLL D FSLL
(dBi) (dB) (dBi) (dB) (dBi) (dB) (dBi) (dB)
MATLAB 20.67 −5.27 20.04 −5.05 19.33 −17.65 18.13 −18.71
CST 19.93 −4.92 19.15 −4.87 19.23 −16.52 18.62 −16.91
(a) Matlab (case-I) (b) CST (case-II) (c) Matlab (case-I) (d) CST (case-II)
Fig. 4. E-theta versus h (at U = 90°) for free-space environment using MATLAB and CST
microwave studio for case-I and case-II, respectively
7 Conclusion
For this work, the structure code is developed in MATLAB for LAA, and the structures
are optimized by PSO and CS optimization techniques. The performances of optimized
and un-optimized results of LAA are compared. The optimized proved to have better
performance than un-optimized in both design and performance parameters. The
comparative study shows that CS is better than PSO in terms of convergence time for
all cases. Then by taking these optimized parameters such as lengths and their
respective spacing of elements, the LAA is designed in CST simulation software. Both
outcomes are seemed to be fairly agreeing with each other for non-uniform spacing,
non-uniform LAA. The novelty of the above discussion is that a GUI-based toolbox is
developed here to design and to optimize LAA in MATLAB, which reduces the
complexity for an amateur. The GUI having both options to design LAA with or
without process optimization made it completely from all aspects.
References
1. Wang X, Gao F, Liu Q (2004) Design of antenna array used as smart antenna for
TD-SCDMA systems. Int Conf Comm Circuits Syst 1:176–180
2. Balanis CA (2005) Antenna theory: analysis and design, 3rd edn. Wiley, London
3. Kummer WH (1992) Basic array theory. Proc IEEE 9(3):127–139
4. Hejres JA (2004) Adaptive sidelobe using the positions of selected elements of the phased
antenna array. IEEE Antennas Propag Soc Symp 3:2655–2658
5. Lee A, Chen L, Wei J, Hwang HK (2005) Simulation study of wideband interference
rejection using adaptive array antenna. In: IEEE aerospace conference, pp 1–6
6. Bossavit A (1998) Computational electromagnetism. Eng Sci Educ J 7(6):275–281
7. Kettunen L (2001) Fields and circuits in computational electromagnetism. IEEE Trans
Electromagn 37(5):3393–3396
8. Abd-Alhameed RA, McEwan NJ, Excell PS, Ibrahim MM, Hejazi ZM, Musa M (1997) New
procedure for design of microstrip patch antennas using method of moments. In: Tenth
International Conference on Antennas and Propagation, vol 1, no. 436, pp 178–181
9. Colak D, Newman EH (1998) The multiple sweep method of moments (MSMM) design of
wide-band antennas. IEEE Trans Antenna Propag 46(9):1365–1371
10. Mishra GP, Jena MR, Mangaraj BB (2016) Investigation on design and performance of
linear cantor array using strip dipole and V-dipole for UHWF band application. In: IEEE
International Conference on Wireless Communication, Signal Processing and Networking,
pp 1810–1814
11. Wang W, Gong S, Wang X, Guan Y, Jiang W (2010) Differential evolution algorithm and
method of moments for the design of Low-RCS antenna. IEEE Antennas Wirel Propag Lett
9:295–298
12. Sivanandam SN, Deepa SN (2008) Introduction to genetic algorithms. Springer, Berlin,
Heidelberg
13. Johnson J, Rahmat-Samii Y (1997) Genetic algorithms in engineering electromagnetics.
IEEE Trans Antennas Propag Magazine 39(2):7–21
14. Yan KK, Lu Y (1997) Sidelobe reduction in array pattern synthesis using genetic algorithm.
IEEE Trans Antennas Propag 45:1117–1122
15. Robinson J, Rahmat-Samii Y (2004) Particle swarm optimization in electromagnetics. IEEE
16. Eberhart RC, Shi Y (2001) Particle swarm optimization: developments, application, and
resources. In: Proceedings of IEEE conference on evolutionary computation, pp 81–86
17. Khodier M, Al-Aqeel M (2009) Linear and circular array optimization: a study using particle
swarm intelligence. Prog Electromagn Res 15:347–373
18. Nguyen TH, Morishita H, Koyanagi Y, Izui K, Nishiwaki A (2013) A multi-level
optimization method using PSO for the optimal design of an L-Shaped folded monopole
antenna array. IEEE Trans Ant Propag 62(1):209–215
19. Mohanty SK, Sahoo AB, Pradhan H, Mangaraj BB (2017) Comparative study of PSO and
CS for optimization of 3 x 5 planar antenna array using MOM. Int J Inf Commun Technol
11(1):128–149
20. Baskar S, Alphones A, Suganthan PN, Liang JJ (2005) Design of Yagi-Uda antennas using
comprehensive learning particle swarm optimization. IEEE Proc Microwaves Antennas
Propag 152(5):340–346
21. Mohanty SK, Mishra GP, Mangaraj BB (2014) Implementing Taguchi and Cuckoo Search
to Optimize LAA. In: Annual IEEE India Conference (INDICON), 11–13 December 2014,
pp 1–5
22. Building GUIs with MATLAB (1997) The math works, 5th edn.
23. Moshe Y (2004) GUI with Matlab. In: Signal and Image Processing Laboratory (SIPL)
Wideband Patch Antenna Element
for Broadcast Applications in Lower 5G Bands
Sonagara Abhishek M(&) and Rahul Vishwakarma

Anusandhan (Deemed to be University), Bhubaneswar, Odisha 751030, India
1 Introduction
The telecom industries have currently been preparing themselves to embrace new 5G
cellular technologies. The commercial products supporting 5G are supposed to be
available by 2020 [1]. Increasing demands of 5G mobile technology has given a boost
to ongoing research in 5G supporting antennas. In accordance with demand, the fre-
quency bands between 3 GHz and 5 GHz have been allocated for 5G services in many
countries such as 3.1–3.55 GHz and 3.7–4.2 GHz in USA, 3.4–3.8 GHz in Europe,
and 3.3–3.6 GHz and 4.8–4.99 GHz in China [2]. To cover all 5G bands, base station
antenna having wide bandwidth with overall high efficiency and high gain is required.
For that, many new antenna designs are proposed in 5G bands [3]. In the evolution of
modern 5G services, compared to other types of antennas, patch antenna can play vital
role due to its characteristics like small size, low profile, easy integration, and low cost
[4]. But at the other side, patch antenna also has limitations in the form of low
bandwidth and lower efficiency, especially in wideband [5]. So microstrip patch
antenna can be a good candidate for 5G base station antenna design due to its men-
tioned advantages, if such design can overcome its limitations [6, 7]. Wideband
operation of simple microstrip patch antenna can be achieved by applying partial
ground and stacking technique [8, 9]. Considering demand on 5G, the proposed
antenna design provides good results for one single antenna element which can be
arrayed to produce good results for base station application [10].
2 Antenna Design
The proposed antenna consists of two grounds one of which is partial and another one
is having length and width same as of substrate as shown in Figs. 1 and 2.
Patch antenna with partial ground plane is suspended over full ground plane. The
substrate used is FR-4 (lossy) dielectric having thickness of 1.6 mm with loss tangent
0.025 and dielectric constant 4.3. The procedure to find length and width for

https://doi.org/10.1007/978-981-15-2774-6_48
406 S. A. M and R. Vishwakarma
(a) Top view (b) Back view
Fig. 1. Patch antenna with partial ground
(a) Side view (b) Perspective view
Fig. 2. Patch antenna with partial ground suspended over full ground plate
rectangular patch is carried out using the below-mentioned equations. The specific
parameters required are dielectric constant (2r ), height of substrate (h), and frequency
of resonance (fr) [5].
rffiffiffiffiffiffiffiffiffiffiffiffiffiffi
v0 2
W¼ ð1Þ
2fr 2r þ 1
2r þ 1 2r 1 1
2reff ¼ þ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð2Þ
2 2 1 þ 12 h W

ð2reff þ 0:3Þ Wh þ 0:264
DL ¼ 0:412 h W ð3Þ
ð2reff 0:258Þ h þ 0:8
v0
L¼ pffiffiffiffiffiffiffiffiffi 2DL ð4Þ
2fr 2reff
The metal patch over the substrate is extended at one end by loading rectangular
metal strips with successively reduced widths and lengths which is shown in Fig. 1a.
Partial ground can also be seen in Fig. 1b. Metal thickness used here is 0.020 mm.
Wideband Patch Antenna Element for Broadcast … 407
Inset feeding technique is used to achieve better impedance matching. The air gap
between patch antenna with partial ground and full ground plate is also shown in
Fig. 2a. The perspective view is also shown in Fig. 2b for better understanding.
The given lengths and widths of the metal strips and partial ground are obtained to
achieve higher required bandwidth of 2 GHz. For the simulation of the proposed
design CST studio software is used.
Results are obtained for two designs, (1) patch antenna with partial ground and
(2) patch antenna with partial ground suspended over full ground plate. S11 parameter
for patch antenna with partial ground remains less than −10 dB for 3.26–4.23 GHz
with bandwidth of 1.16 GHz (31% of center frequency 3.745 GHz) and minimum
return loss reaching at −14.10 dB at 3.68 GHz. These wideband characteristics are due
to partial ground and extended patch elements of different widths and lengths. S11
parameter and VSWR results for both designs are given in Figs. 3 and 4, respectively.
Fig. 3. S11 parameter comparison

Fig. 4. VSWR comparison
S11 parameter for patch antenna with full ground plate shows wideband charac-
teristics for frequency band from 3.1–5 GHz having 10 dB impedance bandwidth of
1.9 GHz (47% of center frequency 4.05 GHz), which covers entire lower 5G bands
with minimum return loss of −51.5 dB at 3.4 GHz and −24.5 dB at 4.6 GHz. VSWR
value for the proposed antenna is also below 2 for most of the covered bands. As
shown in figures, both return loss and VSWR is improved in patch antenna with partial
ground suspended over full ground plate.
From realized gain plot in Fig. 5, role of ground plate below patch antenna can be
understood. For most of the interested frequency bands, realized gain is more than
5 dBi with reaching maximum gain of 6.77 dBi at 3.4 GHz frequency. But the design
of antenna without ground plate has maximum gain of only 2.3 dBi, and it remains less
than the antenna with full ground plate for whole bandwidth.
Fig. 5. Realized gain comparison

Figure 6 shows that total efficiency is also improved in the antenna with full ground
plate compared to antenna without ground plate. Impedance smith chart for both the
designs has been plotted in Fig. 7a, b, and we can observe that smith chart is also in
good agreement with the return loss graph showing that more impedance matching
occurs when impedance curve moves nearer to the normalized 1 value of impedance.
Fig. 6. Total efficiency comparison
(a) Without ground plate (b) With ground plate
Fig. 7. Impedance Smith chart
In Figs. 8 and 9, E plane and H plane directivity polar plots indicate that after
applying ground plate, back radiation is eliminated and giving more radiation in desired
direction, which is desired characteristic for base station antenna (Table 1).
(a) E plane (b) H plane
Fig. 8. Directivity at 3.4 GHz and 3.7 GHz
(a) E plane (b) H plane
Fig. 9. Directivity at 4.6 GHz

Table 1. Summary of results

Parameter Proposed designs
Without ground plate With ground plate
S11 −14.1 dB @ 3.7 GHz −51.5 dB @ 3.4 GHz
−24.5 dB @ 4.6 GHz
VSWR 1.5 @ 3.7 GHz 1.005 @ 3.4 GHz
1.13 @ 4.6 GHz
10 dB Impedance 1.145 GHz (31% of center 1.9 GHz
bandwidth freq. 3.745 GHz) (47% of center
freq. 4.05 GHz)
Total efficiency 92.5% @ 3.7 GHz 95.35% @ 3.4 GHz
86% @ 4.6 GHz
Realized gain 2.23 dBi @ 3.7 GHz 6.77 dBi @ 3.4 GHz
5.36 dBi @ 4.6 GHz
4 Conclusion
In this simulation work, microstrip patch antenna with partial ground backed by full
ground plate is proposed for entire lower 5G bands for broadcast purpose. The sim-
ulated S11 parameter of the proposed structure shows that return loss remains under
−10 dB for all the required 5G frequency bands between 3 and 5 GHz. The proposed
design has overall good efficiency which is more than 80% for almost all required
frequencies. Good performance for gain and radiation patterns is obtained. Therefore,
the proposed antenna can be a good candidate for antenna element of antenna array for
broadcasting purpose in 5G communication technologies.
References
1. Shafi M, Molisch AF, Smith PJ, Haustein T, Zhu P, De Silva P, Tufvesson F, Benjebbour A,
Wunder G (2017) 5G: a tutorial overview of standards, trials, challenges, deployment, and
practice. IEEE J Sel Areas Commun 35(6):1201–1221
2. Ancans G, Bobrovs V, Ancans A, Kalibatiene D (2017) Spectrum considerations for 5G
mobile communication systems. Procedia Comput Sci 104:509–516
3. Ali MMM, Sebak AR (2016) Dual band (28/38 GHz) CPW slot directive antenna for future
5G cellular applications. In: 2016 IEEE International Symposium on Antennas and
Propagation (APSURSI), pp 399–400
4. Chandana RS, Sai Deepthi P, Sriram Teja D, Veera JayaKrishna N, Sujatha M (2018)
Design of a single band microstrip patch antenna for 5G applications. Int J Eng Technol
7:532
5. Balanis C (2005) Antenna theory analysis and design, 3rd edn. Wiley, Publication
6. Yassine J, Gharnati F, Said AO (2017) Design of a compact dual bands patch antenna for 5G
applications. In: 2017 International Conference on Wireless Technologies, Embedded and
Intelligent Systems (WITS), pp 1–4
7. Tarpara N, Rathwa R, Kotak N (2018) Design of slotted microstrip patch antenna for 5G
application. Int Res J Eng Technol (IRJET) 05(04)
8. Marzudi WNNW, Abidin ZZ, Dahlan SH, Ramli KN, Majid HA, Kamarudin MR (2015)
Rectangular patch with partial ground wearable antenna for 2.4 GHz applications. In: 2015
IEEE international RF and microwave conference (RFM), pp 104–9
9. Ooi B-L, Qin S, Leong M-S (2002) Novel design of broadband stacked patch antenna. IEEE
10. Khattak MI, Sohail A, Khan U, Barki Z, Witjaksono G (2019) Elliptical slot circular patch
antenna array with dual band behaviour for future 5G mobile communication networks. Prog
Electromagnet Res C 89:133–147
Anonymized Credit Card Transaction Using
Machine Learning Techniques
B. K. Padhi1(&), S. Chakravarty1, and B. N. Biswal2

1
Department of Computer Science & Engineering, Centurion University
of Technology & Management, Bhubaneswar, Odisha, India
2
Department of Computer Science & Engineering, Bhubaneswar Engineering
Colleges, Bhubaneswar, Odisha, India
1 Introduction
Today, due to the rapid growth of digitalization that has increased online transactions in
the mobile phone drastically created mobile transaction fraud. Mobile phones have
become the best devices for the attacker to gain lots of money within less time with less
effort. Due to the fraudulent transaction, the business and bank are facing massive
financial loss. For years, it has been considered that the information or data collected
from credit card transactions are anonymized or genuine.
1.1 Fraud
Fraud can be known as acquiring someone’s property illegally. Credit card fraud
(CAF) means the use of a credit card without authorization and even stealing the
sensitive data by the fraudsters such as card number, security pin, and one-time
password [1]. CAF can be classified into two categories, such as internal fraud and
external fraud [2], while other classification categories, such as, are conventional-
related card frauds, tradesman-related frauds, and network-related frauds [3].
1.2 Fraud Detection

Fraud detection (FD) is the most effective way to identify a successful transaction as a
fraudulent or non-fraudulent [4]. Credit card fraud detection is a challenging job due to
the following; firstly, information on normal and fraudulent behaviors changes from
time to time, and secondly, due to the skewed and imbalanced credit card transactional
datasets. FD is considered as a data mining classification problem, where the main
objective should be to accurately classify the credit card transactions as fraud or non-
fraud [3].
1.3 Machine Learning

ML techniques can be highly recommended for FD, because of its correctness rate and
detection rate [5]. In ML, we provide data to the particular generic algorithm without

https://doi.org/10.1007/978-981-15-2774-6_49
414 B. K. Padhi et al.
writing any programming code, and thus, it builds logic based on the data provided.
ML algorithms use computational methods to “learn” information directly from data
without relying on a predetermined equation as a model.
2 Literature Survey
In paper [1], a bagging classifier is applied to the credit card transaction dataset to
explain the benefits of the bagging ensemble techniques. In paper [2], support vector
machine (SVM)-based method with multiple kernels is used to analyze the user profile.
In paper [3], the analysis and judgment on the performance of naive Bayes, k-nearest
neighbor, and logistic regression on highly skewed and imbalanced credit card fraud
datasets are evaluated based on an undersampling and oversampling. In paper [4], ten
different machine learning algorithms in the field of credit card fraud detection have
been used with time features in the dataset which is considered as one of the important
features for the performance evaluation. In paper [5], the objective is to create secure
credit card transactions based on deep learning algorithms such as logistic regression,
naive Bayesian, support vector machine, neural network, artificial immune system, k-
nearest neighbor. In paper [6], artificial neural network techniques have been adopted
to detect fraud. In paper [7], different types of supervised machine learning algorithms
are applied in credit card transactions to detect fraud transactions. Undersampling is
used for handling the imbalanced dataset. In paper [8, 9], genetic algorithm has been
used to detect fraud transactions. In paper [10], support vector machine techniques are
used for credit card fraud detection. In paper [11], the mining techniques such as a
frequent itemset mining algorithm have been used for credit card fraud detection. In
paper [12], the decision tree algorithm has been used for detecting fraud. In paper [13],
migrating birds optimization techniques has been used for fraud detection in credit
cards. In paper [14], naive Bayes techniques are been used for fraud detection in credit
cards. In paper [15], for credit card fraud detection, good approaches such as logistic
regression and naive Bayes are carried out. In paper [16], the performance of Bayesian
and neural network has been calculated in credit card fraud detection. In paper [10],
decision tree, neural networks, and logistic regression have been performed and tested
for their applicability in fraud detection.
3 Methodology
In this phase, we have followed the working flow diagram for detecting fraud. In this
section, we have focused on two approaches:
• Classification algorithm
• Ensemble learning algorithm (Fig. 1).
Anonymized Credit Card Transaction Using Machine Learning … 415
Fig. 1. Working flow diagram
3.1 Classification Algorithms

3.1.1 SVM, QDA, LDA, and CART
In 1995, the SVM algorithm was introduced by Vapnik and his team members [4]. The
main idea behind the SVM algorithm is to find the optimal hyperplane that has the
maximum width between the vectors in any dimensional space. An optimal hyperplane
can be considered as one of the linear decision functions with maximum width between
the vectors [5, 7, 17].
Quadratic discriminant analysis (QDA) is similar to LDA, where QDA considers
that estimation from each class in the dataset is normally distributed. As comparing to
LDA, QDA does not consider the covariance of each of the classes is identical [18].
LDA is one most preferred supervised learning techniques where the input region is
divided into regions known as decision regions, and its boundaries are called decision
surfaces or decision boundaries [19]. It is used for solving and providing a better result
for the classification predictive modeling problems.
Classification and regression trees (CART) are implemented by using a binary
decision tree which was developed by Breiman et al. (1984). This technique is used for
continuous and non-parameter data. The condition for division of tree is depended on
attributes, quantity, and Gini index. Two subsets of the data are generated based on the
condition, and the same process continues again for each subset to find out the next
condition for the division.
3.2 Ensemble Learning Algorithm

Ensemble learning is one of the meta-classifier approaches which help to get better
results by the mixture of multiple machine learning classifiers to achieve well the
predictive results [20, 21]. This algorithm includes two types of learning technique:
• Sequential learning (boosting)
• Parallel learning (bagging).
Our main focuses will be on the boosting algorithm.
3.2.1 Boosting Algorithm

Boosting machine learning methods are used to solve complex and convoluted. In this
case, all the features are taken into consideration to make the classification and pre-
diction accuracy. Boosting is a process that combines many machine learning
algorithms to convert weak learners (WL) to strong learners so that the accuracy of the
model increases. In the above figure, the output result (Out) of all the WL are grouped
together in order to produce strong learner (SL) [21].
XGBoost, CatBoost, LGBoost, and Random Forest
The advanced version of gradient boosting is XGBoost, which improves the efficiency
of the models to get better output results. It extremely boosts the performance of the
models [22].
In CatBoost, the categorical columns are given with the indices so that it can be
encoded as one-hot encoding using one_hot_max_size for all the features with no of
different value less than or equal to the given parameter [22].
LGBoost (also known as light gradient boosting machine), from Microsoft and
published in 2017, is used to increase the performance as compared to XGBoost.
LGBoost uses XGBoost as a base and performs better than it in training dataset and
handling capacity sizes of the dataset [22].
RF method is commonly known as the random forest method used for solving
classification and regression problems. RF uses the collection of the decision tree
values; each one is slightly different from each other [7, 20].
4 Performance Evaluation
4.1 Measuring Algorithm Performance
To evaluate the performance of the model, we have considered the measuring factors
such as false positive (FP) and positive predictions that are considered as negative; true
positive (TP) and positive predictions that are considered as positive; false negative
(FN) and negative predictions that are considered as positive; true negative (TN) and
negative predictions that are considered as negative.
4.2 Confusion Matrix

This is a graphical table representation for evaluating the performance of any classi-
fication model [4, 23, 24]. “Y” is yes and “N” is no, whereas “P Class” is predicted
class and “A Class” is an actual class. Confusion matrix looks like (Fig. 2).
A Class= Y A Class =N
P Class =Y TP FP
P Class =N FN TN
Fig. 2. Confusion matrix

4.2.1 Precision, Recall/Sensitivity/True-Positive Rate (TPR),

Specificity/True-Negative Rate (TNR), and F1 score
TP
Precision ¼ ð1Þ
ðTP þ FPÞ
TP
TPR ¼ ð2Þ
ðTP þ FNÞ
TP
TNR ¼ ð3Þ
ðTP þ FPÞ
Rc
F1 Score ¼ 2 Pr ð4Þ
ðPr þ RcÞ
4.2.2 Area Under ROC Curve

A receiver operating characteristic curve (ROC curve) is a graph representation that
shows the performance of a classification algorithm as compared to all classification
thresholds [23, 24]. This curve plot is obtained from TPR versus FPR.
4.2.3 Matthews Correlation Coefficient Score (MCCS)

It is a performance score obtained from true and false positives, as well as true and
false negatives [24].
TP TN FP FN
MCCS ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð5Þ
ðTP þ FPÞ ðFN þ TNÞ ðTP þ FNÞ þ ðFP þ TNÞ
4.2.4 Area Under Precision–Recall Curve

A precision–recall curve is the graph representation curve that shows performance
between precision against recall [25].
4.2.5 Cross-validation Score Report

Cross-validation is one of the statistical methods used to estimate the accuracy of
machine learning models [25].
5 Result Analysis
5.1 Exploring the Dataset

We have taken a dataset that is available in the ULB Machine Learning Group [6]
collected in two days made through credit cards by European cardholders in September
2013. This dataset is transformed by the PCA transformation method, and it provides
28 principal components or features from V1 to V28, but 30 components are taken for
the evaluation process. It is found that there are 284,807 transactions, out of which 492
are fraud transactions and only 0.172% are of fraud transactions from all transactions;
thus, this dataset is highly unbalanced and skewed toward fraud transactions.
5.2 Pre-processing and Features Selection

Pre-processing and selection of features from the above dataset can be done in two
ways, such as:
• Data normalization
• Outlier removal and dimensional reduction.
5.2.1 Data Normalization

Due to the highly unbalanced dataset which is skewed toward the positive class [26],
we have used a StandardScaler method, which will transform the data such that its
distribution will have a mean value 0 and a standard deviation of 1. To create a
subsample dataset with balanced class distributions, we have taken 446 fraudulent
transactions in the training data and 446 non-fraudulent transactions randomly from
normal samples. We got 892 transactions as a subsample dataset for analyzing by
concatenation (Fig. 3).
Fig. 3. Before sampling and after resampling
Then, we shuffled the subsampled dataset. Now after resampling, we found out of
892, non-fraudulent transactions are 446 and fraudulent transactions are 446.
5.2.2 Outlier Removal and Dimensional Reduction

To remove extreme outliers, we have used the quantile technique which reduces our
data size from 892 transactions by 264 transactions to 628 transactions. For dimen-
sional reduction, we have t-Distributed Stochastic Neighbor Embedding (t-SNE) to
reduce the number of random variables.
5.3 Model Evaluation and Result Analysis

5.3.1 Cross-validation Score Report
In the cross-validation technique, we have used k = 10 and random state = 42. We got
the CatBoost model with better results (Tables 1 and 2).
Table 1 Comparison of cross-validation score

Classification algorithms Ensemble learning algorithms
Mean value Standard deviation Mean value Standard deviation
value value
SVM: 0.972,559 0.021100 CatBoost: 0.977419 0.016286
LDA: 0.975472 0.018522 XGBoost: 0.972844 0.017558
QDA: 0.967962 0.021107 LGBoost: 0.968263 0.018790
CART: 0.909202 0.044487 RF: 0.961604 0.026427
Table 2 Confusion matrix

Classification algorithms Ensemble learning algorithms
SVM: QDA: CatBoost XGBoost:
[[83 2] [[81 4] [[84 1] [[84 1]
[10 31]] [6 35]] [10 31]] [9 32]]
LDA: CART: LGBoost: RF:
[[85 0] [[76 2] [[85 2] [[85 0]
[12 29]] [8 33]] [8 33]] [9 32]]
5.3.2 Confusion Matrix Report

From the confusion matrix, we state that LGBM gives better outcomes. It predicts high
positive values and less negative values.
5.3.3 ROC Curve

From Fig. 4, we found that the LGBM algorithm gives better results (Figs. 4, 5 and 6).
Fig. 4. ROC curve

Fig. 5. Comparison of algorithms based on ROC-AUC curve
Fig. 6. Comparison of algorithms based on precision–recall curve
5.3.4 ROC-AUC Curve

From the comparison plot, we focus not only on achieving the highest accuracy but
also to extract business values and information. Therefore, we found that LGBM takes
over the CatBoost to achieve the highest degree of comprehensiveness.
5.3.5 Precision–Recall Curve

From the precision–recall curve, we found LGBM gives better result than other models
(Table 3).
Table 3 Comparison of precision, recall, F1 score, and MCC score

Algorithms
Selected model Precision value Recall value F1 score value MCC score value
SVM 0.94 0.76 0.84 0.78
LDA 1.00 0.71 0.83 0.79
QDA 0.90 0.85 0.88 0.82
CART 0.86 0.78 0.82 0.78
CatBoost 0.97 0.76 0.85 0.80
XGBoost 0.97 0.78 0.86 0.82
LGBoost 1.00 0.80 0.89 0.86
RF 0.94 0.81 0.87 0.80
To evaluate imbalanced datasets, it is better to consider the four-parameters, in

order to achieve high comprehensiveness and if the dataset is highly imbalanced, i.e.,
Matthew Coefficient Correlation Score value, F1 score value, and the area under the
precision–recall curve value. Therefore, we found that the LGBM algorithm gives
better outcomes among all models.
In order to detect fraud is a tremendous and complex issue that requires better active
planning before applying a machine learning algorithm on it. Fraud detection is an
application of data science, and machine learning can give a better result, to keep the
customer’s money safe without losing integrity. We propose our future work to include
a comprehensive tuning of the LGBoost algorithm and apply grid search techniques,
with a non-anonymized dataset in which we can find important features for detecting
fraudulent transactions.
References
1. Zareapoor M, Shamsolmoali P (2015) Application of credit card fraud detection: based on
bagging ensemble classifier. In: International Conference on Computer, Communication and
Convergence (The authors published by Elsevier B.V)
2. Shen A, Tong R, Deng Y (2007) Application of classification models on credit card fraud
detection. In: 2007 International conference on service systems and service management,
IEEE, pp 1–4
3. Awoyemi JO, Adetunmbi AO, Oluwadare SA (2017) Credit card fraud detection using
machine learning techniques: a comparative analysis. In: 2017 International Conference on
Computer Networking and Informatics (ICCNI), Lagos, pp 1–9
4. Rajora S, Li D-L, Jha C, Bharill N, Patel OP, Joshi S, Puthal D, Prasad M (2018) A
comparative study of machine learning techniques for credit card fraud detection based on
time variance, 978–1-5386-9276-9/18
5. Popat RR, Chaudhary J (2018) A survey on credit card fraud detection using machine
learning. In: Proceedings of the 2nd International Conference on Trends in Electronics and
Informatics (ICOEI 2018) IEEE Conference Record:# 42666; IEEE Xplore ISBN:978-1-
5386-3570-4
6. Mittal S, Tyagi S (2019) Performance evaluation of machine learning algorithms for credit
card fraud detection, 978-1-5386-5933-5/19/$31.00
7. Dhankhad S, Mohammed EA, Far B (2018) Supervised machine learning algorithms for
credit card fraudulent transaction detection: a comparative study. In: 2018 IEEE international
conference on information reuse and integration for data science
8. Rama Kalyani K, Uma Devi D (2012) Fraud detection of credit card payment system by
genetic algorithm. Int J Sci Eng Res 3(7):1–6 (ISSN 2229-5518)
9. Meshram PL, Bhanarkar P (2012) Credit and ATM card fraud detection using genetic
approach. Int J Eng Res Technol 1(10):1–5 (ISSN: 2278-0181)
10. Singh G, Gupta R, Rastogi A, Chandel MDS, Riyaz A (2012) A machine learning approach
for detection of fraud based on SVM. Int J Sci Eng Technol 1(3):194–198 (ISSN : 2277-
1581)
11. Seeja KR, Zareapoor M (2014) Fraud miner: a novel credit card fraud detection model based
on frequent itemset mining. Sci World J 2014:1–10. https://doi.org/10.1155/2014/252797
(Hindawi Publishing Corporation)
12. Patil S, Somavanshi H, Gaikwad J, Deshmane A, Badgujar R (2015) Credit card fraud
detection using decision tree induction algorithm. Int J Comput Sci Mobile Comput 4(4):
92–95 (ISSN: 2320-088X)
13. Duman E, Buyukkaya A, Elikucuk I (2013) A novel and successful credit card fraud
detection system implemented in a turkish bank. In: 2013 IEEE 13th International
Conference on Data Mining Workshops (ICDMW), pp 162–171
14. Bahnsen AC, Stojanovic A, Aouada D, Ottersten B (2014) Improving credit card fraud
detection with calibrated probabilities. In: Proceedings of the 2014 SIAM international
conference on data mining, Society for Industrial and Applied Mathematics, pp 677–685
15. Ng AY, Jordan MI (2002) On discriminative vs. generative classifiers: a comparison of
logistic regression and naive bayes. Adv Neural Inf Process Syst 2:841–848
16. Maes S, Tuyls K, Vanschoenwinkel B, Manderick B (2002) Credit card fraud detection
using Bayesian and neural networks. In: Proceedings of the 1st international naiso congress
on neuro fuzzy technologies, pp 261–270
17. Hsu C, Chang C, Lin C (2003) A practical guide to support vector classification. Technical
report, Department of Computer Science and Information Engineering, National Taiwan
University
18. Ogwueleka FN (2011) Data mining application in credit card fraud detection system. J Eng
Sci Technol 6(3):311–322
19. Mahmoudi N, Duman E (2015) Detecting credit card fraud by modified Fisher discriminant
analysis. Expert Syst Appl 42:2510–2516
20. Sorournejad S, Zojaji Z, Atani RE, Monadjemi AH (2016) A survey of credit card fraud
detection techniques: data and technique oriented perspective. arXiv:1611.06439
21. Herawan T et al (eds) An application of oversampling, undersampling, bagging and boosting
in handling imbalanced datasets. In: Proceedings of the first international conference on
advanced data and information engineering (DaEng-2013). Lecture notes in electrical
engineering vol 285. Springer Science+Business Media Singapore. https://doi.org/10.1007/
978-981-4585-18-7_2
22. CatBoost, LGBoost, XGBoost: https://towardsdatascience.com/catboost-vs-light-gbm-vs-
xgboost-5f93620723db
23. Syaripudin A, Khodra ML (2014) A comparison for handling imbalanced datsets. In: 2014
International Conference of Advanced Informatics: Concepts, Theory and Aplications
(ICAICTA), 978-1-4799-5100-0/14
24. Thennakoon A, Bhagyani C, Premadasa S, Mihiranga S, Kuruwitaarachchi N (2019) Real-
time credit card fraud detection using machine learning. 978-1-5386-5933-5/19
25. Nuno Abrunhosa Carneiro (2017) A data mining approach to fraud detection in e-tail.
Published in Decision Support Systems. https://doi.org/10.1016/j.dss.2017.01.002
26. Lee S (2000) Noisy replication in skewed binary classification. Comput Stat Data Anal
34:165–191
Performance Comparison of Adaptive
Algorithms for Smart Antenna of Dipole Array
Barsa Samantaray1(&), Kunal Kumar Das1, and Jibendu Sekhar Roy2

1
ECE Department, Institute of Technical Education and Research,
Siksha ‘O’ Anusandhan Deemed to be University, Bhubaneswar, Odisha, India
{barsasamantaray,kunaldas}@soa.ac.in
2
School of Electronics Engineering, Kalinga Institute of Industrial Technology
(KIIT), Deemed to be University, Bhubaneswar, Odisha, India
1 Introduction
Smart antenna is smart with reference to smart signal processing algorithm. In adaptive
smart antenna, by using signal processing algorithm the direction of arrival can esti-
mate and beam direction toward the desired user and null toward the undesired
direction can be generated [1, 2]. Various beam-forming Adaptive signal processing
algorithms with its advantages and disadvantages are discussed in [1–4]. Most of the
algorithms used for adaptive beamforming in smart antenna have fixed step-size
parameters. Some literatures are available where variable step-size parameters are used
[5–8], where mainly variable step-size least mean square (LMS) algorithm is used.
Here, sign LMS (SLMS) and variable step SLMS (VS-SLMS) algorithms are used
for the design of smart antenna. SLMS is a variant of LMS algorithm. Here, a com-
parison is done between the performances of these two algorithms. The VS-SLMS
algorithm has better performance than SLMS algorithm in adaptive beam generation of
smart antenna.
2 Adaptive Beamforming Algorithms
For the design of smart antenna, adaptive signal processing algorithms, sign LMS
(SLMS) algorithm and variable step sign LMS (VS-SLMS) algorithm, are used. Sign
(also called signed regressor) LMS (SLMS) is used for the adaptation process to be
very fast.
The equation used for updating the weight in sign LMS is [9].
xðn þ 1Þ ¼ xðnÞ þ lðnÞ e ðnÞ sgn½xðnÞ ð1Þ

https://doi.org/10.1007/978-981-15-2774-6_50
Performance Comparison of Adaptive Algorithms for Smart … 425
The sign function can be defined as

where
sgn½xðnÞ ¼ 1; for xðnÞ [ 0

¼ 0; for xðnÞ ¼ 0 ð2Þ
¼ 1; for xðnÞ\0
l(n) is the step size, and e*(n) is the error signal between input and output signals.
In VS-SLMS, the step size l is also variable [5, 6] with iteration number and in
every iteration, and during weight updating, step-size parameter varies according to the
formula [5]
ln þ 1 ¼ aln þ den ; if 0\ln þ 1 \lmax

ð3Þ
¼ lmax ; otherwise
where maximum value of convergence parameter, lmax, is defined as lmax \2=kmax and
is the largest kmax which is the maximum eigen value of the correlation matrix of the
signal. In (3), ‘a’ and ‘d’ are constant parameters, and in this paper, in simulation, the
values chosen for those are a = 0.95 and d = 0.0003.
The factor en in (3) is related to the weight vector as
en ¼ kwðn þ 1Þ wðnÞk=kwðn þ 1Þk ð4Þ
The weight updated VS-SLMS algorithm equation is
wðn þ 1Þ ¼ wðnÞ þ ln þ 1 e ðnÞsgn½xðnÞ ð5Þ
3 Simulated Design of Smart Antenna of Dipole Array
The diagram of a linear dipole array is shown in Fig. 1, where the dipole antennas are
separated by a distance of ‘d’.
Fig. 1. Linear dipole array

426 B. Samantaray et al.
For a dipole antenna of length ‘l’, the radiated electric field is expressed as [10]
2 3
bl
jbr
I0 e cos 2 cos h cos bl2
EðhÞ ¼ jg 4 5 ð6Þ
2pr sin h
where b ¼ 2p=k is the propagation constant, I0 is current amplitude fed to the dipole
antenna, η = 120pX, and r is the distance from the radiator array.
For a linear array of N number of dipole antennas, the total radiated field, with
mutual coupling, is given by
X
N
I0 EðhÞejðn1Þð k cos h þ aÞ
2pd
Etotal ðhÞ ¼ ð7Þ
n¼1
where k is the wavelength, a is the progressive phase shift of the array, and I0 is the
current fed to the antennas, and it is assumed that all the antennas are fed by equal
current I0. Array factor of (7) is used for dipole array using adaptive algorithms SLMS
and variable step-size VS-SLMS. Adaptive beam generation algorithm makes possible
to generate beam in the desired direction and null in the unwanted direction.
In simulated design of smart antenna of dipole array, a number of dipole antennas
in arrays are N = 16 and N = 20, and in all the cases, inter-element spacing is d = 0.5k.
The results for different beam directions (BD) and null directions (ND) for N = 16,
obtained using SLMS and VS-SLMS, are plotted in Figs. 2 and 3, respectively
(BD = 50°, ND = 60°).
-20
-40
Normalised Radiation Pattern (dB)
-60
-80
-100
-120
-140
-160
-180
0 20 40 60 80 100 120 140 160 180
Angles (Degree)
Fig. 2. Beamforming using SLMS algorithm

-10
Normalized Radiation Pattern (dB)
-20
-30
-40
-50
-60
-70
0 20 40 60 80 100 120 140 160 180
Angles (Degree)
Fig. 3. Beamforming using VS-SLMS algorithm
Similarly, the results for different beam directions (BD) and null directions (ND) for
N = 20, obtained using SLMS and VS-SLMS, are plotted in Figs. 4 and 5, respectively
(BD = 70°, ND = 75°).
The variation for variable step-size parameter for VS-SLMS algorithm for N = 16
is shown in Fig. 6.
Mean square error (MSE) plots for SLMS and VS-SLMS are plotted in Fig. 7 and
in 8, respectively, for N = 20.
From the graphs of Figs. 7 and Fig. 8, it is evident that the convergence of VS-
SLMS algorithm is better than SLMS algorithm. The results for beamforming are
tabulated in Table 1.
From Table 1, it is evident that the performance of VS-SLMS algorithm is better to
obtain lower side lobe level (SLL). Using SLMS for N = 16 and N = 20, maximum
SLLs (SLLmax) are −6.5 dB and −6 dB, respectively, whereas using VS-SLMS N = 16
and N = 20, maximum SLLs (SLLmax) are −11.5 dB and −8.2 dB, respectively. Also,
convergence of VS-SLMS is better than SLMS algorithm.
-20
-40
Normalised Radiation Pattern (dB)
-60
-80
-100
-120
-140
-160
0 20 40 60 80 100 120 140 160 180
Angles (Degree)
Fig. 4. Beamforming of SLMS algorithm
-10
Normalized Radiation Pattern (dB)
-20
-30
-40
-50
-60
-70
-80
0 20 40 60 80 100 120 140 160 180
Angles (Degree)
Fig. 5. Beamforming using VS-SLMS algorithm

-3
10
5.05
4.95
mu
4.9
4.85
4.8
4.75
0 100 200 300 400 500 600 700 800 900 1000
no of iteration
Fig. 6. Variation of step-size parameter of VS-SLMS algorithm
0.025
0.02
mean square error
0.015
0.01
0.005
0
0 100 200 300 400 500 600 700 800 900 1000
iteration
Fig. 7. MSE plot for SLMS (N = 20)

0.01
0.009
0.008
0.007
mean square error
0.006
0.005
0.004
0.003
0.002
0.001
0
0 100 200 300 400 500 600 700 800 900 1000
iteration
Fig. 8. MSE plot for VS-SLMS
Table 1 Comparison of beamforming performances of SLMS and VS-SLMS algorithms

Adaptive No. of Desired Desired Obtained Obtained SLLmax
algorithms dipoles angle (°) null (°) angle (°) null (°) (dB)
VS-SLMS N = 16 50 60 51 60 −11.5
N = 20 70 75 69.6 75 −8.2
SLMS N = 16 50 60 51.6 60 −6.5
N = 20 70 75 69 75.4 −6
4 Conclusion
Smart antennas of dipole arrays are generally used for cellular communication from
UHF to microwave frequencies, typically up to 6 GHz. A performance comparison
between SLMS and VS-SLMS algorithms is presented here which will help to design
adaptive smart antenna of dipole array. In both the algorithms, the number of iteration
is 1000. In SLMS, the step-size parameter, l = 0.05. Convergence of variable step-size
algorithm is better, and using variable step size algorithm, more accurately desired
beam direction and null direction can be achieved. Side lobe level reduction using VS-
SLMS algorithm is more than SLMS algorithm. For a smart antenna array of 16
dipoles, 5 dB lower side lobe level is achieved using VS-SLMS algorithm than SLMS
algorithm.
References
1. Godara LC (1997) Application of antenna arrays to mobile communications, part II: beam-
forming and direction-of-arrival considerations. Proc IEEE 85(8):1195–1245
2. Bellofiore S, Balanis CA, Foutz J, Spanias AS (2002) Smart antenna systems for mobile
communication network, part 1, overview and antenna design. IEEE Antennas Propag Mag
44(3):145–154
3. Halder A, Senapati A, Roy JS (2018) Smart antenna of microstrip array for sector
beamforming for 28 GHz millimeter wave mobile communication. Int J Microwave Opt
Technol (IJMOT) 13(3):203–208
4. Sharma R, Senapati A, Roy JS (2018) Beamforming of smart antenna in cellular network
using leaky LMS algorithm and its variants. Int J Microwave Opt Technol (IJMOT) 13(4):
263–268
5. Lau YS, Hussain ZM, Harris RJ (2004) A weight-vector LMS algorithm for adaptive
beamforming. In: IEEE TENCON Conference, New Jersey, USA, pp 94–498
6. Luo XD, Jia ZH, Wang Q (2006) A new variable step size LMS adaptive filtering algorithm.
Acta Electronica Sinica 34(6):1123–1126
7. Zhang Y, Li N, Chambers JA, Hao Y (2008) New gradient-based variable step size LMS
algorithms. EURASIP J Adv Signal Process 2008(7):1–9
8. Huang H, Lee J (2012) A new variable step size NLMS algorithm and its performance
analysis. IEEE Trans Signal Process 60(4):2055–2060
9. Senapati A, Ghatak K, Roy JS (2015) A comparative study of adaptive beamforming
techniques in smart antenna using LMS algorithm and its variants. In: International
Conference on Computational Intelligence & Networks (CINE-2015), KIIT University,
India, 12–13 Jan 2015. IEEE Xplore, pp 58–62
10. Balanis CA (2005) Antenna theory—analysis and design, 3rd edn. Wiley
Effect of Buffer Size on Performance
of Wireless Sensor Network
Manoranjan Das(&) and Benudhar Sahu
Institute of Technical Education and Research, SOA (Deemed to be University),

Bhubaneswar, India
1 Introduction
Wireless communication technology though an old concept for information transmis-

sion between a pair of transmitter and receiver is still treated as one of the advanced
communication technologies so far as the data sharing among users is concerned. More
specifically when the users belong to a communication network where the network is
limited to distance between users and cost of installation, the wireless is preferred as an
ultimate among various types of communication technology. Thus, wireless commu-
nication network has reserved a worthy place in the current research scenario. The
cellular network, WLAN, wireless ad hoc network are the examples of wireless
communication network. These wireless communication networks are further can be
categorized as either infrastructure-based or infrastructure-less networks. The
infrastructure-based wireless communication networks are those where the presence of
a fixed infrastructure (like a fixed base station) plays an important role for functioning
of the network. Example of such a kind wireless network is the cellular network. In the
other side, an important feature of an infrastructure-less wireless communication net-
work is that the users (here onwards named as nodes) in these networks communicate
among themselves without the help of any fixed base station (or access point). Wireless
ad hoc networks where the nodes are either static or mobile by nature come under this
category. The popularly known wireless sensor network (WSN) [1, 2] is the noted
example to describe the operation of an infrastructure-less wireless communication
network.
The performance of such wireless communication networks in compared to wired
communication networks suffers from various factors such as network size, traffic
density, and routing. To be more explainable, in a wireless sensor network the amount
of data collected at the end node (i.e., sink node) directly reflects the network per-
formance. Further, in order to have a good collection of data at the end node it is
required that the sensor nodes (which are the only constituents of a WSN) must have
good efficiency not only in data sensing but also data handling (i.e., both data storing
and data forwarding). It is worthy enough to mention here that in a wireless sensor
network the job of a sensor node is not limited to data sensing and transmitting its data
to the end node only. Rather a sensor node is also required to carry forward the data
sensed by other sensing node which is not in the transmission range of sink node.
Alternatively, the sensor nodes also take the responsible of handling the data traffic
efficiently. In order to handle this job, the sensor node uses its buffer which is of a fixed
https://doi.org/10.1007/978-981-15-2774-6_51
Effect of Buffer Size on Performance of Wireless Sensor … 433
size. However, many a times it may so happen that the network gets congested in route
when the traffic in the network is heavy enough. Further, when the sensor node is
loaded with heavy traffic, the buffer size is not enough to handle. In this condition, the
data packets may get dropped because of buffer overflow [3]. Thus, the buffer size of
the sensor node is also treated as one strong factor that affects the performance of a
wireless sensor network. Hence, it is important to study the effect of buffer size on the
performance of wireless sensor network [3, 4].
As explained above, it is important in a wireless sensor network to have a good
amount of collection of data packets from the sensing nodes that show a good network
performance. Thus in this paper, we study the effect of buffer size on the performance
of a wireless sensor network that assuming the nodes are static by nature [5, 6].
The rest of the paper is organized as follows. In Sect. 2, a brief overview about
wireless sensor network is given. Section 3 defines the performance metrics used for
analyzing the performance of the network. The list of parameters being chosen for
simulation is also given in Sect. 3. In Sect. 4, analysis toward the performance of
network under different environments with the help simulation plots is presented.
Conclusion of the paper is presented in Sect. 5.
2 Overview of Wireless Sensor Network
Among the different infrastructure-less wireless network, the WSN is quite popular so
far as the application is considered. The WSN is preferably suitable for application
environment such as battlefield, habitat monitoring, and mines area. The installation of
a wireless sensor network is done through the deployment of large number of tiny
sensor nodes in a given geographical area. The distribution of the nodes in the specified
geographical area may follow a random (mostly) or regular fashion depending on the
network design. Further, it is also noticed that in almost all WSN applications the
sensor nodes have two major responsibilities (1) data acquisition and (2) data trans-
mission to an end node (normally called as sink node) as and when desired. Thus, as
the name suggests the nodes belonging to a WSN are usually equipped with both
sensing devices and communication devices along with limited resources such as
power supply (i.e., battery-powered) and storage. Usually, all the nodes present in
WSN shares their information through wireless mode of communication. However,
since the nodes are small in size, and associatively, they are battery-powered the
communication range of some sensor node is not enough to directly communicate the
information with the sink node. This problem is solved by allowing a sensor node to
carry forward the data packet of another sensor node which is neighbor of each other.
Thus, the sensor nodes in WSN also bear the responsibility of routing the data packets
in a route between source and destination node pair.
434 M. Das and B. Sahu
3 Simulation Environment and Performance Metrics
In this work to study the effect of buffer size on the network performance, we preferred
a simulation methodology where the simulation environment replicates the specified
real-time scenario. The network considered for the simulation is a static one which
means the nodes present inside network area are static by nature. Also, the distribution
of nodes in the network follows random topology. The network simulator used for the
said purpose is Global Mobile Information System Simulator (GloMoSim) [7], and the
associated parameters assigned with the values are as given in Table 1.
Table 1. Parameters used for simulation

Parameters Value/Specification
Terrain area 1500 M 300 M
Number of nodes 100
Node mobility None
Number of sources 25
Simulation time 5M
Transmission power 20 dBm
Mac protocol 802.11
Routing protocol AODV
Packet size 512 bytes
Type of data traffic Constant bitrate (CBR)
Packet transmission rate 1 packet/s, 5 packet/s
Buffer size 2, 5, 10, 50, 100
Performance Metrics. In this study, the performance analysis is carried with the help
of the following performance metrics as defined below.
(a) Packet Delivery Fraction. This metric gives a value that defines the ratio of the
amount of data packet delivered at the destination node to the total data packets
transmitted from the source nodes during the simulation period.
(b) Packet Drop Fraction. This metric presents the fraction of amount of data packets
lost (dropped) at the routing nodes due to buffer overflow for a given of data
packets being transmitted by the source nodes in the network during the simulation
period.
4 Result Analysis
In order to get the plots for different performance metrics, the network simulator runs
for 300 s against five different seeds, where the seed differs from each other based on
spatial coordinates of nodes in the network area. This means for each seed, the spatial
coordinates of nodes are different. Accordingly, the average of five output values
corresponding to five different seeds is computed to give a point on the plot.
Figure 1 presents the variation in packet delivery fraction versus buffer size for
different packet transmission rate. It is observed from Fig. 1 that the amount of packets
delivered successfully at the destination increases with increase in buffer size. This is
due to the fact that, the increase in buffer capacity at the sensor node enables it to store
more incoming data packets which are to be forwarded later. It is also noticed from the
plots in Fig. 1 that with increase in packet transmission rates the packet delivery
fraction decreases for a fixed value of buffer size. This happens so because the sensor
nodes have a fixed data receiving rate. Thus, when packet transmission rate increases,
the receiving sensor node gets less time to receive all the data packets and to put it in
the buffer. However at the same time, it can be seen in Fig. 1 that for higher packet
transmission rate the percentage increase in packet delivery fraction is more in com-
pared to that for lower packet transmission rate. The specific reason behind this is
nothing but the number of data packets that get a successful route to be transmitted in-
between the source and destination nodes. Lastly, it is also depicted from Fig. 1 that for
increase in buffer size at the lower range (i.e., from 2 to 10) the percentage change in
packet delivery fraction is more in compared to the increase in buffer size in the higher
range (i.e., from 10 to 100). Also for low packet transmission rate, it seems that the
change in buffer size at a higher range has almost nil effect on packet delivery fraction.
However for high packet transmission rate, there is a small change in packet delivery
fraction with increase in buffer size at the higher range. This is because when packet
transmission rate is low, a précised portion of the buffer area remains empty most of the
time.
0.65
Packet delivery fraction
0.6
0.55
packet transmission rate = 1packet/sec

packet transmission rate = 5packet/sec
0.5
0.45
2 5 10 50 100
Buffer size
Fig. 1. Packet delivery fraction versus buffer size for different packet transmission rate in a
WSN of 100 nodes with 25 sources
436 M. Das and B. Sahu
Figure 2 given indicates the effect of buffer size on performance metric (i.e., packet
drop fraction) for different packet transmission rate. It is observed from Fig. 2 that with
increase in buffer size, the packet drop fraction due to buffer overflow decreases. It is
quite correct since the increase in buffer size at the sensor node allows the node to store
more incoming data packets. It is also noticed in Fig. 2 that there is no packet drop take
place for high range of buffer size (i.e., 10 and more than 10) when the packet
transmission rate is low. However for high packet transmission rate, the trend of
decrease in packet drop fraction continues even in high buffer size range. Further, it is
clear from Fig. 2 that with increase in packet transmission rate the amount of data
packet dropped due to buffer flow is more in compared to that for lower packet
transmission rate for a fixed buffer size. This is because when the packet transmission
rate is made to increase, the number of data packets that arrive at a sensor node also
increases. This increase in incoming data packets at the sensor node creates an overload
burden to the sensor node since the node has also some other responsibility (i.e., data
acquisition) at the same time and collected data also has to be stored in the same buffer.
Hence, congestion at the intermediate sensor node occurs and so the incoming data
packets cannot be accommodated in the buffer and dropped.
packet drop fraction due to buffer overflow
0.14
packet transmission rate = 1 packet/sec
0.12
packet transmission rate = 5 packet/sec
0.1
0.08
0.06
0.04
0.02
-0.02
2 5 10 50 100
Buffer size
Fig. 2. Packet drop fraction due to buffer overflow versus buffer size for different packet
transmission rate in a WSN of 100 nodes with 25 sources
5 Conclusions
The study on different performance metrics for different buffer sizes in a static WSN
shows that selection of suitable buffer size is important so as to maintain a good quality
performance. Further, it is also observed from the plots obtained from simulation
results that the effect of buffer size is more precious when packet transmission rate in
the network increases. It is also seen from the simulation plots that with increase in
buffer size to a higher range, the performance of the network improves. However, since
the sensor nodes are very small in size and the number of sensor nodes being deployed
in the geographical area is large, the increase in buffer size to support high data
transmission rate is not economical. Hence, there must be a tradeoff between selecting
the buffer size with respect to the rate of transmission of data packet in between nodes.
References
1. Sohraby K, Minoli D, Znati T (2010) Wireless sensor networks, 2nd edn. Wiley, India
2. Fahmy HMA (2016) Wireless sensor networks: concepts, applications, experimentation and
analysis. Springer
3. Kempa WM (2017) On buffer overflow duration in WSN with a vacation-type power saving
mechanism. In: Proceedings of the 24th international conference on systems, signals and
image processing (IWSSIP)
4. Shakir ZD, Yoshigoe K, Lenin RB (2012) Adaptive buffering scheme to reduce packet loss on
densely connected WSN with mobile sink. In: Proceedings of IEEE consumer communica-
tions and networking conference (CCNC)
5. Pathak AA, Deshpande VS (2015) Buffer management for improving QoS in WSN. In:
Proceedings of ICPC
6. Rault T, Bouabdallah A, Challal Y (2013) WSN lifetime optimization through controlled sink
mobility and packet buffering. In: Proceedings of the global information infrastructure
symposium (GIIS)
7. Zeng X, Bagrodia R, Gerla M (1998) GloMoSim: a library for the parallel simulation of large-
scale wireless networks. In: Proceedings of 12th workshop on parallel and distributed
simulation
Noise Cancellation Using a Novel Self-adaptive
Neuro-fuzzy Inference System (SANFIS)
Laxmipriya Samal, Debashisa Samal(&), and Badrinarayan Sahu
Department of ECE, ITER, S‘O’A Deemed to be University, Bhubaneswar, India

debashishamal@soa.ac.in
1 Introduction
In order to remove noise contained in signals, the noise cancellation is a unique method
[1, 2]. It is applied in systems like transformers, hands-free phones, machineries,
speech enhancement, echo cancellation, biomedical signal and image processing [3–7].
The concept of noise cancellation using adaptive filter was first proposed by Widrow
[8, 9]. Many methods are suggested in order to overcome this problem. More recently
least mean square algorithm (LMS), normalized least mean square algorithm (NLMS),
recursive least square (RLS) algorithms are used [10–12]. In this work, self-adaptive
neuro-fuzzy inference system (SANFIS) is proposed for noise cancellation. It is being
extended to the nonlinear area by using nonlinear adaptive system which uses the
concept of neural network and fuzzy logic. The neural network is used to adapt the
parameter of the fuzzy system of the noise source into an interfering component for the
detection of signal. The performance of SANFIS technique gives better results with
previously proposed LMS and NLMS algorithm.
2 Adaptive Noise Cancellation (ANC)
In the above block diagram [13], the information-bearing signal s(k) was mixed with
noise which was originated from another noise by a dynamic procedure. The objective
of this is to produce a signal which is exactly the same as that of source signal (Fig. 1).
3 Adaptive Algorithms for Noise Cancelling
In the past, many algorithms have been widely studied, and some are discussed below.
3.1 The Least Mean Square (LMS) Adaptive Algorithm

This algorithm suggested by Widrow and Hoff [14] uses the method of gradient descent
approach. To reduce the error, it calculates minima adjusting the filter coefficients.
LMS algorithm is discussed below

https://doi.org/10.1007/978-981-15-2774-6_52
Noise Cancellation Using a Novel Self-adaptive Neuro-fuzzy … 439
Fig. 1. Generalized block diagram of adaptive noise cancelling
wl ðp þ 1Þ ¼ wl ðpÞ þ g eðpÞ sðp lÞ ð1Þ
X
k1
yðpÞ ¼ wl ðpÞ sðp lÞ ð2Þ
l¼0
eðpÞ ¼ dðpÞ yðpÞ ð3Þ
where
yðpÞ filter output,
eðpÞ signal error,
sðpÞ input signal,
g step size,
wðpÞ filter weights.
3.2 Normalized Least Mean Square (NLMS) Adaptive Algorithm

In order to overcome the drawback of LMS algorithm, NLMS algorithm [15, 16] was
suggested which uses variable step size. By doing so, it converges faster than the LMS.
The steps of NLMS algorithm are given below
b
gðpÞ ¼ ð4Þ
d þ jsðpÞj2
b
wðp þ 1Þ ¼ wðkÞ þ eðpÞ sðpÞ ð5Þ
d þ jsðpÞj2
where
b is the NLMS adaption parameter
d normalization constant.
440 L. Samal et al.
4 The Self-adaptive Neuro-fuzzy Inference (SANFIS) System
The self-adaptive neuro-fuzzy inference system creates its name from adaptive neural
fuzzy inference system where the parameters are updated itself using back-propagation
algorithm. The combination fuzzy systems with neural networks possess good the
learning ability in optimizing the coefficients (Fig. 2).
Fig. 2. SANFIS architecture
The architecture consist of [17–19] described as below
axiom 1:letfx is A1 and y is B1 g so f1 ¼ fp1 x þ q1 y þ r1 g ð6Þ
axiom 2:letfx is A2 and y is B2 g so f2 ¼ fp2 x þ q2 y þ r2 g ð7Þ
First layer: each node with activation function is given by
R1;k ¼ lAk ðxÞ; k ¼ 1; 2 ð8Þ
R2;k ¼ lBk2 ðxÞ; k ¼ 3; 4 ð9Þ

Second layer: every node in this level is given by
R2;k ¼ wj ¼ lAk ðxÞlBk ðyÞ; k ¼ 1; 2 ð10Þ
Third layer: every node in this is given by

wk
k ¼
O3;k ¼ w ; k ¼ 1; 2 ð11Þ
w1 þ w2
Fourth layer: every node in this level is given by
k fk ¼ w
O4;k ¼ w k ð pk x þ qk y þ r k Þ ð12Þ
Fifth layer: every node in this level is given by overall response

X P
wk fk
O5;k ¼ k fk ¼ Pk
w ð13Þ
k k wk
The suggested methodology of noise cancellation with SANFIS approach is

expressed below.
In Fig. 3, X(k) is the information-bearing signal which is to be reconstructed from
the primary noise signal n(k). This is passed through nonlinear dynamics (f) and
generates signal Z1(k), which summed to X(k) to form the measured output signal Y(k).
The purpose is to retrieve X1(k) from the measured signal Y(k) and delayed version of
primary noise Z2(k) which should be the same version of information signal X(k).
Fig. 3. Block diagram of the proposed SANFIS method

442 L. Samal et al.
The results of simulation are analysed using MATLAB 2018. Figure 4 shows the
information signal which is a sinusoidal signal. Figure 5 shows the noise signal gen-
erated from a primary noise source by a nonlinear process. Figure 6 is the measured
signal obtained by sum of information with noise signal. Figure 7 is the output of
SANFIS system, and Fig. 8 is the recovered signal which is the exact replica of
Fig. 4. Information signal
Fig. 5. Noise signal Z1

Fig. 6. Measured signal Y
Fig. 7. Output of SANFIS Z2
information signal. In order to find the performance measure, we plot mean squared
error (MSE) curves of LMS, NLMS and SANFIS algorithms shown in Fig. 9. We
observe that LMS algorithm converges around 40 iterations whereas NLMS algorithm
around 30 iterations. Similarly, for the same data, SANFIS system converges around 20
iterations. From these results, we conclude that NLMS algorithm converges faster
compared to LMS algorithm. Similarly, the converge rate of SANFIS system algorithm
is more faster compared to NLMS. Thus, we conclude SANFIS system is better than
LMS and NLMS algorithms in terms of quality performances.
444 L. Samal et al.
Fig. 8. Recovered signal X1
Fig. 9. MSE curve for comparisons of LMS, NLMS and SANFIS system
6 Conclusion
In this work, the idea of fuzzy logic and neural network are merged to develop
SANFIS. This SANFIS technique is used for noise cancellation. We extend the method
of linear adaptive noise cancellation into the nonlinear by using nonlinear adaptive
systems. The performance of SANFIS technique for adaptive noise cancellation,

compared with the LMS, NLMS algorithm. The future work may include the opti-
mization algorithms for different kinds of noises and to implement with DSP micro-
controller and LABVIEW.
References
1. Sambur M (1978) Adaptive noise canceling for speech signals. IEEE Trans Acoust Speech
Signal Process 26(5):419–423
2. Burgess JC (1986) Adaptive signal processing. In: Widrow B, Stearns SD (eds)
3. Wu JD, Lin SL (2010) Audio quality improvement of vehicular hands-free communication
using variable step-size affine-projection algorithm. Int J Wavelets Multiresolut Inf Process 8
(06):875–894
4. Sasaoka N, Shimada K, Sonobe S, Itoh Y, Fujii K (2009) Speech enhancement based on
adaptive filter with variable step size for wideband and periodic noise. In: IEEE international
midwest symposium on circuits and systems, August 2009, pp 648–652
5. Ahmad MS, Kukrer O, Hocanin A (2013) A 2-D recursive inverse adaptive algorithm. SIViP
7(2):221–226
6. Kim PU, Lee Y, Cho JH, Kim MN (2011) Modified adaptive noise canceller with an
electrocardiogram to enhance heart sounds in the auscultation sounds. Biomed Eng Lett 1
(3):194
7. Widrow B, Glover JR, McCool JM, Kaunitz J, Williams CS, Hearn RH, Goodlin RC (1975)
Adaptive noise cancelling: principles and applications. Proc IEEE 63(12):1692–1716
8. Haykin S (2008) Adaptive filter theory. In: 27th annual international conference of the
engineering in medicine and biology society. IEEE Press, Pearson Education India,
pp 1212–1215
9. Albert TR, Abusalem H, Juniper MD (1991) Experimental results: detection and tracking of
low SNR sinusoids using real-time LMS and RLS lattice adaptive line enhancers. Naval
Ocean Systems Center, San Diego, CA
10. Kazemi R, Farsi A, Ghaed MH, Karimi-Ghartemani M (2008) Detection and extraction of
periodic noises in audio and biomedical signals using Kalman filter. Sig Process 88(8):2114–
2121
11. Akingbade KF (2014) Separation of digital audio signals using least-mean-square
(LMS) adaptive algorithm. Int J Electr Comput Eng 4(4):2088–8708
12. Dhiman J, Ahmad S, Gulia K (2013) Comparison between adaptive filter algorithms (LMS,
NLMS and RLS). Int J Sci Eng Technol Res (IJSETR) 2(5):1100–1103
13. Dixit S, Nagaria D (2017) LMS adaptive filters for noise cancellation: a review. Int J Electr
Comput Eng 7(5):2088–8708
14. Haykin S (2008) Adaptive filter theory. In: 27th Annual international conference of the
engineering in medicine and biology society. IEEE Press, Pearson Education India,
pp 1212–1215
15. Paulo SD (2008) Adaptive filtering algorithms and practical implementation. Int Series Eng
Comput Sci 23–50
16. Mohammed J (2012) A study on the suitability of genetic algorithm for adaptive channel
equalization. Int J Electr comput Eng 2(3):285
17. Juang CF, Chiou CT, Lai CL (2007) Hierarchical singleton-type recurrent neural fuzzy
networks for noisy speech recognition. IEEE Trans Neural Netw 18(3):833–843
446 L. Samal et al.
18. Theocharis JB (2006) A high-order recurrent neuro-fuzzy system with internal dynamics:
Application to the adaptive noise cancellation. Fuzzy Sets Syst 157(4):471–500
19. Qin H, Yang SX (2007) Adaptive neuro-fuzzy inference systems based approach to
nonlinear noise cancellation for images. Fuzzy Sets Syst 158(10):1036–1063
20. Chandrakar C, Kowar MK (2012) Denoising ECG signals using adaptive filter algorithm.
Int J Soft Comput Eng (IJSCE) 2(1):120–123, 2088–8708
21. Dixit S, Nagaria D (2017) Design and analysis of cascaded LMS adaptive filters for noise
cancellation. Circuits Syst Signal Process 36(2):742–766
22. Prasetyowati SAD, Susanto A (2015) Multiple processes for least mean square adaptive
algorithm on roadway noise cancelling. Int J Electr Comput Eng 5(2):355
Frequency Regulation in an Islanded
Microgrid with Optimal Fractional
Order PID Controller
Narendra Kumar Jena1, Subhadra Sahoo1, Amar Bijaya Nanda1,

Binod Kumar Sahu1(&), and Kanungo B. Mohanty2
1
Department of Electrical Engineering, Siksha ‘O’ Anusandhan Deemed
to be University, Bhubaneswar, Odisha, India
{narendrajena,subhadrasahoo,amarnanda}@soa.ac.in
2
Department of Electrical Engineering, NIT, Rourkela, Rourkela, Odisha, India
kbmohanty@nitrkl.ac.in
1 Introduction
Microgrid (MG) paves a solution for electrification in an isolated area. Apart from this,
incessant rise in demand, limited conventional resources, and environment issues entail
to develop MGs. MGs in off-grid condition bring the voltage and frequency regulation
problem, which seldom faces such problem in the grid-connected mode. The problems
in the isolated MG are (a) low inertia due to the microturbine DGs, (b) sporadic nature
of renewable energy sources, and (c) no inertia with complex structure of energy-
storing elements. So, by the robust secondary controller, LFC problem tackled to bring
the frequency deviation within a predefined boundary [1, 2].
Different researchers delved with LFC study on MGs endorsing different kinds of
controllers such as PID controller [2], adaptive PI controller [3], fuzzy PI controller [4],
type-2 fuzzy PI controller [5–7], H-infinite controller [8], H∞, and l-synthesis-based
controller [9], tilted integral derivative (TID) controller [10], and virtual inertia control
of microgrid [11]. In fuzzy-type controller, rule formulation is a difficult task by which
with improper input–output scaling parameters leads the system into an unstable region
or the response carries high steady-state error. In H-infinite type controller, it is difficult
to know all the internal states, and in practice, it is too difficult to implement. But PID is
a conventional one which is used profusely in industries due to its simple structure and
can be easily implemented. From this angle in this paper, a fractional order-based PID
controller is adopted which is a maiden attempt to use as a secondary controller in the
LFC of a microgrid. This FOPID controller can reach out to produce an improved
dynamic response in comparison with PID controller because it carries two numbers of
extra adjusting knobs. Adjusting these extra two knobs, undershoot, overshoot, and
settling time can be improved than the result produced by PID controller. To get the
optimal gains of controllers, various researchers employed different computational

https://doi.org/10.1007/978-981-15-2774-6_53
448 N. K. Jena et al.
techniques like PSO [2], Big Bang–Big Crunch optimization [5], salp swarm opti-
mization (SSA) [7], and grasshopper optimization [12]. In this paper, to have a better
transient response of the microgrid, the gains of PID and FOPID controllers are
optimized by a SOS algorithm.
The objective of this paper is assimilated as follows:
a. A fractional order PID controller is designed by SOS algorithm to employ as a
secondary controller in the proposed isolated MG.
b. Performance of this proposed controller (FOPID) is validated by comparing with
PID controller.
c. Robustness of the controller is examined by subjecting power generation variation
as well as load variation.
2 Systems, Proposed Controller, and Algorithm

2.1 System Considered
The MG shown in Fig. 1 [9] is configured by its linearized model as depicted in Fig. 2.
The microgrid contains PV source, WTG, diesel generator, FC, MTG, FH storage unit,
and BES unit whose specifications are tabled in Appendix and are taken from [9]. The
total power required to meet the demand is the sum of power generated by renewable
sources, diesel generator, FC, MTG, and power exchange of storage elements. The
power balance equation is expressed in Eq. (1).
Micro
turbine Diesel
engine
Fuelcell
converter
ac bu s
a c lo a d
ac bu s
converter converter converter converter
Wind turbine Flywheel BES PV array
Fig. 1. Schematic diagram of an isolated MG

Frequency Regulation in an Islanded Microgrid with Optimal … 449
ΔPwind 1 Δ Pwtg
1 + sT wtg
ΔPinso 1 Δ PPV
1 + sT PV
Δ PL = SLP
1 Δ PMT
1 + sT MT
FOPID 1 Δ PDi ΔP 1 Δf MG
controller 1 + sT Di D + sM
1 ΔPFC
1 + sT FC
1 ΔPFH
1 + sTFH
1 Δ PBES
1 + sT BES
Fig. 2. Linearized model of MG
PL ¼ PPV þ Pwtg þ PMT þ PDi þ PFC PFH PBES ð1Þ
With an incremental change of load, the power balance equation becomes as given
in Eq. (2).
DPL þ DPPV þ DPwtg þ DPMT þ DPDi þ DPFC þ DPFH þ DPBES ¼ 0 ð2Þ
The frequency deviation from Fig. 2 is given in Eq. (3).
Dp
Df ¼ ð3Þ
D þ sM
where
Dp ¼ DPPV þ DPwtg þ DPMT þ DPDi þ DPFC þ DPFH þ DPBES DPL ;
D = damping coefficient, and M = inertia constant.

Mathematical modeling of different sources is elaborated below.
DPwind DPwtg
DPwtg ¼ ð4Þ
Twtg
DPinso DPPV
DPPV ¼ ð5Þ
TPV
u DPMT
DPMT ¼ ð6Þ
TMT
u DPDi
DPDi ¼ ð7Þ
TDi
u DPFC
DPFC ¼ ð8Þ
TFC
DfMG DPFH
DPFH ¼ ð9Þ
TFH
DfMG DPBES
DPBES ¼ ð10Þ
TBES
Dp DDf
Df ¼ ð11Þ
M
The state-space model of the microgrid as in [9] taking all the differential equations
from Eqs. 4–11 is expressed in Eq. 12.
0
1
DP 0 1 1
B wtg C 0 0 0 0 0 0 0 0 1
B C B Twtg C DPwtg
B DP C B
B PV C 0 1
0 0 0 0 0 0 C CB DPPV C
B C B TPV
CB C
B DP C B B 1
0 CB DPMT C
C B
B MT C 0 0 0 0 0 0 C
B C B TMT
CB C
B C B 1
0 CB DPDi C
B
B DPDi C B 0 0 0 0 0 0 C
B C¼B CB
TDi
B C B C DP C
B DPFC C B 0 0 0 0 1
0 0 0 CB C
CB C
FC
B C B
TFC
B B
1 CB DPFH C
C
B C 1
B DPFH C B 0 0 0 0 0 0 TFH CB C
B C B C
TFH
B C B 0 1 1 C@ DPBES A
0 0 0 0 0
B DPBES C @ TBES TBES A
Df
@ A 2D
0 0 0 0 0 0 0 M
Df ð12Þ
0 1 1 0 1
Twtg 0 0 0
B C B 0 C
B 0 1
0C B C
B TPV C B 1 C
B C
0 C0 1 B TMT C
B
B 0 0 C
B C DPwind B 1 C
B 0 0 C
0 CB C B TDi C
B
þB
B
C
C@ DPinso A þ B 1 Cu; and
B 0 0 0C B TFC C
B C DPL B C
B 0 0 0C B 0 C
B C B C
B C B C
@ 0 0 0A @ 0 A
2
0 0 M 0
y ¼ Df
To evaluate a prompt and smooth dynamic response of the proposed microgrid,

FOPID and PID controllers are optimized by SOS algorithm which is explained in the
next section.
2.2 Fractional Order PID Controller

Podlubny [13] explored non-integer-based controller PI / Dk in which / and k are
two fractional operators. The operational plane of fractional controller from Fig. 3a
shows that PID only works at a single point but PI / Dk controller works on the whole
quadrant. So, from this whole area suitable points are available for controlling the
system. The control signal of PI / Dk controller is given in Eq. (13).
(a) (b)
Fig. 3. a Operational quadrant of PID, and fractional PID. b Structure of FOPID controller
uFOPID ðtÞ ¼ kp eðtÞ þ ki D/t eðtÞ þ kd Dkt eðtÞ ð13Þ
where eðtÞ is the error signal injecting into the controller. In the Laplace domain, the
transfer function of PI / Dk is expressed in Eq. (14).
ki
GFOPID ðsÞ ¼ kp þ þ kd s k ð14Þ
s/
2.3 Symbiotic Organisms Search (SOS)

The SOS algorithm [14] imitates the symbiotic collaboration procedures for the most
part received by different organisms to endure and scatter in the biological system.
Organisms never isolate themselves to lead their life because of dependence on dif-
ferent species for sustenance and even endurance. Symbiosis epitomizes the bonding
between at least two kinds of organisms for their feeding association. Often, living
beings get advantage one another or one gets an advantage from the other without
doing any damage to the other living being. Therefore, according to their dependency
with each other, the SOS algorithm technique contains three stages in particular
mutualism stage, commensalism stage, and parasitism stage which are represented in
the flowchart in Fig. 4 eloquently. SOS has only few parameters which are fixed and
does not require any parameter tuning like other algorithms.
Start
Initialisation phase
Intialise no of population, Maximum no. of iteration and termination criteria
Evaluate the initial population and identify the best performing organism as Xbest
Mutualism phase
Randomly select one organism X j where Xj ≠ X i to calculate the mutual vector and benefit
factors
Update the population and evaluate their performance
Accept the modified Yes No

If the newly generated organism is Accept the old one and
organism and reject the
better than previous one reject the new one
old one
Commensalism phase
Randomly select one organism X j where Xj ≠ X i to modify organisms and evaluate their
performance
Accept the modified Yes If the newly generated organism is No Accept the old one and
organism and reject the
better than previous one reject the old one
old one
Parasitism phase
Randomly select one organism X j where X j ≠ X i and create a parasite vector from X i
Accept the parasite Yes No

If the parasite vector is better thanX i Accept X i
vector
Update the iteration count
No
If the maximum iteration reached
Yes
Store the best solution
End
Fig. 4. Flowchart for SOS computational technique

The proposed model in Fig. 1 is deployed in MATLAB/Simulink to study the fre-

quency control under five different cases which are based on different combinations
among the change of solar power, wind power, and load. The dynamic response is
assessed by employing PID and FOPID controller designed by SOS algorithm. The
enumerated gain parameters are given in Table 1. In this paper, with these same
designed gain values all the rest cases have been studied which is different from other
cited papers.
Table 1. Gain parameters of controllers

Controller kp ki kd / k
PID −14.3624 −2.7633 −16.1832 – –
FOPID −19.6294 −15.4307 −8.4892 0.9321 0.8007
Case-1 Dynamic behavior under variable load and constant solar, and wind
power
In this case, a load deviation of 0.01 is imposed injecting constant solar power and
wind power of each 0.2 pu. The frequency response is depicted in Fig. 5a. The
overshoots and settling times (2% band) of frequency excursions by PID and FOPID
controllers are 13.876 MHz, and 5.985 s, and 6.797 MHz and 3.875 s, respectively.
The FOPID controller gives a prompt response curtailing both the overshoot and the
settling time.
Case-2 Dynamic behavior under variable load and constant solar, and wind
power
After getting the success in the first case, frequency control is studied in the same
system by subjecting a wider load variation (−0.035 to 0.1 pu) endorsing FOPID
controller. The wind power and solar power are 0.2 pu each. The load variation curve
and frequency deviation curve are captured in Fig. 5b. The time curve in Fig. 5b shows
a minor deviation in frequency which shows the design of the controller and is optimal.
Case-3 Dynamic behavior under constant load, constant solar power, and variable
wind power.
Under this condition, frequency deviation dynamics is studied keeping solar power at
0.2 pu, load at 0.01, and variable wind power ranging from −0.1 to 0.4 pu. This excess
increment or decrement of power injected into the system is balanced by the energy-
storing element by which the frequency deviation dynamics are stabilized. The
excursion of frequency portrayed in Fig. 5c shows a skimpy overshoot than the
response observed in case 1.
Case-4 Dynamic behavior under constant load, constant wind power, and variable
solar power
In this context, a variable solar power varied from −0.1 to 0.3 pu along with a constant
wind power of 0.2 pu and a constant load of 0.01 pu. The time response of frequency
-3
10
(a) 14
PID
12 FOPID
10
8
f in Hz
-2
0 2 4 6 8 10
Time in sec
(b) 0.1
f in Hz 0.01
0.08 SLP in pu 0
-0.01
0.06 58 60 62 64 66 68
SLP
0.04
f/
0.02
-0.02
-0.04
0 10 20 30 40 50 60 70 80 90 100
Time in sec
(c) 0.4
10
-3
0.3
-2
P wind
-4
0.2
f in Hz 68 70 72 74
f/
P in pu
0.1 wind
-0.1
0 10 20 30 40 50 60 70 80 90 100
Time in sec
Fig. 5. a Frequency deviation curve. b Frequency deviation curve against variable load.
c Frequency deviation curve against variable wind power. d Frequency deviation curve against
variable solar power. e Frequency deviation curve against variable load, solar power, and wind
power
(d) 0.3
f in Hz
0.01 P PVinsolation in pu
0.2
PVinsolation
0
-0.01
0.1
P
60 65
f/
-0.1
0 20 40 60 80 100
Time in sec
(e) 0.4
PVinsolation
SLP in pu
P in pu
wind
0.3
P PVinsolation in pu
Pwind / P
f in Hz
0.2
0.1
SLP /
0
f/
-0.1
0 20 40 60 80 100
Time in sec
Fig. 5. (continued)
deviation depicted in Fig. 5d shows an improved transient response in comparison with

case 2 because of the power balance correction is carried out by the energy-storing
components.
Case-5 Dynamic behavior under variable load, wind power, and solar power The
last case deals with the perturbation of load in the range of −0.035 to 0.1 pu, wind
power variation from −0.1 to 04 pu, and solar power variation from −0.1 to 0.4 pu.
The transient behavior frequency along with variable load, wind power, and solar
power time diagram as shown in Fig. 5e shows the stability of the system and is
increased due to the power exchanged by the energy-storing element into the system.
The transient response is quite faster with exhibiting minimal overshoot.
4 Conclusion
The simulation of islanded MG enticing FOPID as a secondary controller optimized by

SOS algorithm is carried out successfully. The FOPID controller ameliorates the
transient response produced by the PID controller efficiently. The proposed controller
tackles the response proficiently and promptly in the wake of load variation, solar, and
wind power variation independently as well as combining. The appreciable excursion
of frequency deviation in all cases and in all forms (undershoot, overshoot, and settling
time) confirms the robustness of the proposed controller that is perceived.
Appendix
D ¼ Damping coefficient TMT ¼ Time constant of MT TFH ¼ Time constant of FH

¼ 0:012 ¼ 2s ¼ 0:1 s
M ¼ Inertia constant TDi ¼ Time constant of diesel engine TBES ¼ Time constant of BES
¼ 0:2 ¼ 2s ¼ 0:1 s
TPV ¼ Time constant of PV TFC ¼ Time constant of FC Twtg ¼ Time constant of WTG
¼ 1:8 s ¼ 4s ¼ 1:5 s
References
1. Yang J, Zeng Z, Tang Y, Yan J, He H, Wu Y (2015) Load frequency control in isolated
micro-grids with electrical vehicles based on multivariable generalized predictive theory.
Energies 8(3):2145–2164
2. Bevrani H, Habibi F, Babahajyani P, Watanabe M, Mitani Y (2012) Intelligent frequency
control in an ac microgrid: online PSO based fuzzy tuning approach. IEEE Trans Smart Grid
3(4):1935–1944
3. Zheng S, Tang X, Song B, Lu S, Ye B (2013) Stable adaptive PI control for permanent
magnet synchronous motor drive based on improved JITL technique. ISA Trans 52(4):
539–549
4. Li X, Song YJ, Han SB (2008) Frequency control in micro-grid power system combined
with electrolyzer system and fuzzy PI controller. J Power Sources 180(1):468–475
5. Engin Y (2014) Interval type-2 fuzzy PID load frequency controller using big bang-big
crunch optimization. Appl Soft Comput 15:100–112
6. Khooban MH, Niknam T, Blaabjerg F, Davari P, Dragicevic T (2016) A robust adaptive
load frequency control for micro-grids. ISA Trans 65:220–229
7. Sahu PC, Mishra S, Prusty RC, Panda S (2018) Improved-salp swarm optimized type-II
fuzzy controller in load frequency control of multi area islanded AC microgrid. Sustain
Energy Grids Netw 16:380–392
8. Singh VP, Mohanty SR, Kishor N, Ray PK (2013) Robust H-infinity load frequency control
in hybrid distributed generation system. Int J Elect Power Energy Syst 46:294–305
9. Bevrani H, Feizi MR, Ataee S (2015) Robust frequency control in an islanded microgrid:
H∞, and l-synthesis approaches. IEEE Trans Smart Grid 7(2):706–717
10. Khadanga RK, Padhy S, Panda S, Kumar A (2018) Design and analysis of tilt integral
derivative controller for frequency control in an islanded microgrid: a novel hybrid dragonfly
and pattern search algorithm approach. Arab J Sci Eng 43(6):3103–3114
11. Kerdphol T, Rahman FS, Watanabe M, Mitani Y, Turschner D, Beck HP (2019) Enhanced
virtual inertia control based on derivative technique to emulate simultaneous inertia and
damping properties for microgrid frequency regulation. IEEE Access 7:14422–14433
12. Barik AK, Das DC (2018) Expeditious frequency control of solar photovoltaic/biogas/
biodiesel generator based isolated renewable microgrid using grasshopper optimisation
algorithm. IET Renew Power Gener 12(14):1659–1667
13. Podlubny I (1999) Fractional-order systems and PI/sup/spl/lambda//D/sup/spl/mu//-con-
trollers. IEEE Trans Autom Control 44(1):208–214
14. Min-Yuan C, Prayogo D (2014) Symbiotic organisms search: a new metaheuristic
optimization algorithm. Comput Struct 139:98–112
Biometric Detection Using Stroke Dynamics
Abhishek Das, Saumendra Kumar Mohapatra(&),

and Laxmi Prasad Mishra

laxmimishra@soa.ac.in
1 Introduction
Alphonse Bertillon, a famous French Police Officer, was the first person who devel-
oped an identification system based on body measurements for criminals in the 1870s.
‘Anthropometry’ developed by him was the first scientific biometric system used by
police to identify criminals. The reason behind this development was impersonation.
Impersonation is a crime associated with pretending to be someone else. Earlier
criminals changed their appearances to evade capture, so he used body measures to
identify and capture criminals. That means life measures can be used to authenticate a
person.
Biometrics means life measures is a method to identify persons based on their
characteristics. It depends on unique biological characteristics to verify an individual.
These biometric authentication systems compare currently extracted data with the
stored data. If both data match, then authentication is confirmed; otherwise, the person
is blocked to use the system. The applications for biometrics can be found in gov-
ernment centers, military applications, commercial applications, banking, etc. Bio-
metrics can be divided into two kinds: physiological and behavioral.
The system requires no dedicated action from the user, so it avoids the problems of
need for physical contact. And also, for capturing biometric data, no external hardware
is required. The system discriminates the users by using the combination of human
motion analysis and typing rhythm biometrics. Continuous authentication (CA) is the
method in which we have to check the identity of the user continuously, based on each
user’s action performed. This method is to authenticate persons both statistically and
continuously to provide continuous security without much interrupting the user.
Our work is ordered as follows: Section 1 gives the brief introduction of this project
with its objective and motivation. Section 2 describes the existing system. Section 3
gives the design of the proposed system and the description of the method. Imple-
mentation is organized in Sect. 4. Result analysis is included in Sect. 5. Section 6
concludes the work.

https://doi.org/10.1007/978-981-15-2774-6_54
Biometric Detection Using Stroke Dynamics 459
2 Related Works
(A) Face Recognition
The principal completely mechanized personification-based face acknowledgment
approach fit for working with information procured in the wild is presented. 3D face
structure is produced from a solitary 2D picture and contrasts it and a reference model
for acquiring a minimal portrayal of face highlights deviations. But it requires
expensive camera, and construction of the 3D model is difficult and takes more time
[1]. Face recognition is done using linear discriminant analysis method. SVD-LDA
method is used to reduce the time complexity. LDA is taking mean as the discrimi-
nating factor not taking the covariance, and so it does not contain as much varied data.
Also, this method is supervised so cannot work efficiently with new features [2].
Then, user’s face detection is done using the principal component analysis method.
By using PCA, the researchers have reduced the amount of variables required. This
identification was carried out by providing a new image in the eigenface subspace and
by comparing the spot in eigenface subspace with the spot of already stored images.
But the method has worked with the still images and is not useful for real-time face
recognition. And also, the method did not work for different size face image recog-
nition [3]. The fiducial points were determined by the system in one matching process,
and thus, it eliminates the requirement for analyzing the equality of all the models with
corresponding graphs individually. But the method is responsive to illuminating con-
ditions. A lot of graphs have to be stored manually, and it will use huge storage [4].
(B) Keystroke dynamics
In literature, we can find a number of works were done related to key stroke estimation.
They used amount of disarray of the array to differentiate among different typing data.
They computed the mean distance of the provided sample by using all the data of
different users. If the corresponding mean distance is small to the mean distance of the
samples, then the user is considered as a legal user. So authentication depends on all
legal users. That is, for the development of a user profile, it relies on other users [5].
In the approach by Dowland, calculating the mean and also the standard deviation
of digraphs were used for determining a user profile. Typing example was collected by
monitoring the users during their regular behavior. But it showed high error rates [6].
In authentication via keystroke, dynamics clustered the data with the help of a
variation in the maxi-min distance algorithm. Another variable is the typing speed of a
user which is used as the criteria for clustering. But there is a need to recluster the data
as the system is used [7].
(C) Mouse Dynamics
A power transformation method is used to analyze the holistic attributes of mouse
trajectories. A schematic representation is made using these holistic attributes. But the
method listening carefully on the user verification in a static scenario. And also, it is
using fixed sequence of mouse behavior [8].
The users have to perform a precise job, and the job consisted of moving the mouse
between two lines, i.e., from a point where it was started and the ending point. The
460 A. Das et al.
participants are instructed to perform this task 5 times per session, 6 sessions in total
resulting in total 30 runs. But the problem is that, instruction to do a specific task did
not represent the natural user’s behavior [9].
Active authentication by mouse movements developed a random look like envi-
ronment for revealing user’s heuristic mouse movements than just a fixed path for
every user. User’s one sample movement of all procedure is recorded as a reference
data. Training and verification phases are there. Features from 9 paths between 7
squares are displayed consequently [10].
User’s behaviors are extracted from interaction with a Web page in Web biomet-
rics. It uses the normal login process of a conventional online security system. User’s
verification is by Web interaction [11].
(D) Combining Keystroke and Mouse Dynamics
A work uses MD and KD in multi-model decision fusion for continuous authentication.
Architecture with 11 number of sensors are used. A dataset having data of 67 users
while working in an office for a week is used. A mixture of naive Bayes and SVM is
used to analyze the performance [12].
A new reporting technique was used to evaluate the system performance. They
have obtained a good result, i.e., 50 out of 53 unblocked users out by their system while
3 users were occasionally wrongly locked out by the system [13].
3 Methodology
3.1 Problem Statement
The problem is to develop a mechanism for providing static and continuous authen-
tication for humans. Static authentication is based on face recognition, and continuous
authentication is provided by mouse and keyboard interaction behavior. On the off
chance that the present conduct of a client digresses altogether from the ordinary
conduct, the framework hails this conduct as irregular.
3.2 Design
Biometric authentication systems compare confirmed authentic data in a database with
captured biometric data. Authentication is confirmed, if the two examples of biometric
information coordinate. Ceaseless confirmation is a type of dynamic validation which
changes the point of view of verification from an occasion to a procedure. Design of the
proposed system is displayed in Fig. 1.
The proposed method starts with static authentication, and then, the continuous
analysis is performed by the combination of mouse and keystroke dynamics. Both the
static and continuous authentication methods consist of two different phases:
1. Training state:
During which the user’s data is collected and a final model is produced.
Fig. 1. System architecture
2. Authentication phase:
The model is then used to constantly screen they believed clients account. In the event
that the present conduct of a client strays altogether from typical conduct, at that point
the framework hails this conduct as atypical.
3.3 Overall System Design

Building of face recognition is by principal element study. Eigenface approach trans-
forms faces into a small set of essential characteristics, eigenfaces. Recognition is
performed by extracting the face features and comparing them with the measurements
stored in the database. Face recognition allows the user to continue to use the system, if
both face features are matched. It can use the existing hardware infrastructure and
existing camera. Then, the user is continually monitored by the combination of key-
stroke and mouse dynamics to avoid the session hijacking. Keystroke dynamics is a
form of behavioral biometrics. The aim behind this keystroke dynamics is to identify
humans and it is based on the analysis of their typing rhythms on the keyboard. By
estimating the stay time (the period of time a key is held down) and fly time (the time
term from a key discharged to the following key squeezed) for every console activity,
elements can be removed.
In mouse dynamics, for each user, a schematic characteristic representation is made
by using the holistic attributes (travelled distance, time, etc.). Then, procedural char-
acteristic representation is made using the dynamic process of mouse movements by
performing behavior segmentation.
462 A. Das et al.
4 Implementation
4.1 Face Recognition
Face identification is a biometric method of identifying an individual by comparing live
capture with the stored record for that person. It is an “Easy-to-collect” biometrics. The
whole recognition process involves two steps:
A. Initialization process.
B. Recognition process.
When a user starts using a system, then his face image is captured. After prepro-
cessing the face images, the features are extracted.
PCA is used to overcome the expensive computation and the need for great
amounts of storage of older face recognition technique such as correlation methods.
The features extracted are stored.
In authentication phase also, the features are extracted in the same way as in the
training phase and classification method of nearest neighbor is used for classification.
When the authorized user uses the system after his training phase, then the user
identified as genuine and if a user, whose features are not stored starts using the system
he is identified as an unauthorized user and system is blocked to prevent further
movements of the unauthorized user in the system.
If an authorized user is using the system, he is continued to use the system and
keystroke, and mouse dynamics is used to avoid session hijacking. So that the user can
be continually monitored (Fig. 2).
Fig. 2. Face recognition
PCA is the method for compressing a lot of data into something that captures the
essence of original data. It is done by the decomposition of covariance matrix. PCA
finds the principal components of data, and it does what it says on the name. The
eigenvector with the highest eigenvalue is the principal component. Every image in the
training set is represented as eigenfaces. Eigenface maximizes the scatter between all
face images in the training set.
Keystroke dynamics aims to identify humans based on the analysis of their typing
rhythms on the keyboard. And it is a behavioral biometric. The time and value of all
key press and key release events are recorded. The method has two phases: enrollment
and verification. Learning phase extracts the dynamics, and the verification phase
compares the current user with the dataset. K-nearest neighbor is used for the classi-
fication at authentication phase. By using KNN, the genuine user and malicious user
can be found out. Each user exhibits distinctive hand motion gestures and habits, and
the mouse interaction behavior appears to be unique. Here, users are discriminated by
the way they interact with the mouse.
4.2 Procedural Feature Construction

The dynamic process of a mouse movement is captured by partitioning each mouse
movement into several segments. By partitioning the mouse movement, movement
segments are obtained. Dividing each mouse movement into four movement segments
(four phases).
1. First phase: Speed increases from 0 and user starts moving the mouse from starting
position.
2. Second phase: Speed reaches the maximum user accelerates.
3. Third phase: Speed reduces and user decelerates the mouse moving in the target
direction.
4. Fourth phase: Speed reduces to 0. User adjusts mouse to approach the target.
Features are extracted from segments, and the extracted features then combined to
form a vector. It represents the mouse segments of each mouse operation.
The verification system mainly consists of four main components:
1. Recorder
2. Feature construction
3. User verification model
4. Fusion and decision maker.
Fsch = Schematic feature vector classification score.
Fpro = Procedural feature vector classification score.
The combination rules used are:
1. Product rule: Fsch * Fpro.
2. Sum rule: Fsch + Fpro.
3. Max rule: max (Fsch, Fpro).
4. Min rule: min (Fsch, Fpro).
5. Weighted sum rule: Wsch*Fsch + Wpro*Fpro.
Recognize the legitimate user by calculating the scores. The classification scores
from both the schematic features and procedural features are combined using the
combination rules. The weighted sum rule gives the best performance while the product
rule showed the worst performance.
464 A. Das et al.
The result from both the verification models (keyboard and mouse) is compared to
make a decision. The result from the keyboard authentication and mouse authentication
is compared to decide whether the user is genuine or not. So a continuous authenti-
cation of the user is performed by comparing both the results (Fig. 3).
Fig. 3. Combination of keyboard and mouse
5 Results and Discussions
In order to evaluate the proposed recognition system, experiments were conducted.

These total 25 users were tested using the method. Out of 25, 13 are genuine users and
12 are unauthorized users. We do test in initial learning and also in behavioral learning.
The results of test done after behavioral learning are shown in Fig. 4. And the result of
test done after initial learning is shown in Fig. 6. The performance measures found out
during the test are shown in Figs. 5 and 7. From the figures, it is clear that the results
obtained by the proposed method are better than doing single biometric authentication.
BEHAVIORAL LEARNING
NO. OF TESTS TP TN FP FN
FACE 25 9 11 2 1
KEYBOARD 25 10 9 2 2
MOUSE 25 11 10 1 4
COMBINATION 25 10 10 3 2
Fig. 4. Behavioral learning test result
METHOD ACCURACY PRECISION RECALL F SCORE

FACE 81% 0.81 0.75 0.78
KEYBOARD 83% 0.32 0.85 0.85
MOUSE 76% 0.74 0.67 0.70
COMBINATION 91% 0.95 0.85 0.86
Fig. 5. Performance measures in behavioral learning

INITIAL LEARNING
No. of Tests TP TN FP FN
FACE 25 9 9 1 4
KEYBOARD 25 10 8 2 4
MOUSE 25 9 9 2 3
COMBINATION 25 11 11 1 3
Fig. 6. Initial learning test result
METHOD ACCURACY PRECISION RECALL F SCORE

FACE 79% 0.88 0.68 0.77
KEYBOARD 80% 0.78 0.79 0.76
MOUSE 78% 0.68 0.69 0.74
COMBINATION 86% 0.91 0.79 0.83
Fig. 7. Performance measure in initial learning
Both in initial learning and behavioral learning, the combination of face, keyboard,
and mouse is showing more accuracy, precision, etc., than the single biometric system.
On comparing both the performance measures results, it is shown that the results
got from the behavioral learning test are best than the initial learning test.
6 Conclusion
Biometric-based recognition of humans based on keystroke and mouse dynamics with

face recognition is a simple and efficient static and continuous authentication system
based on human–device interaction behavior. Continuous authentication is a form of
dynamic and risk-based authentication. It changes the perspective of an event to a
process. Face recognition is appropriate for static authentication because it cannot be
fooled by any other methods. After the successful face authentication, the user is
continually monitored by the keyboard (by using the typing pattern) and mouse (by
using the user’s mouse behavior). By behavioral learning, more accuracy can be
achieved. Without right security measures, critical data would get into the hands of
cybercriminals and also impersonation is a huge challenge. So utilizing biometrics for
personal authentication is convenient and considerably more accurate than the existing
methods.
466 A. Das et al.
References
1. “A Leopard Cannot Change Its Spots”: Improving Face Recognition Using 3D-Based
Caricatures. Joao Neves. IEEE Trans Inform Forensics Secur 14(1), Jan 2019
2. Deswal M, Kumar N, Rathi N (2014) Face recognition based on singular value
decomposition linear discriminant analysis method. Int J Eng Innov Technol 3(12), June
2014
3. Chandrapaul L, Al Sumam A (2012) Face recognition using principal component analysis.
Int J Adv Res Comput Eng Technol 1(9), Nov 2012
4. Wiskott L, Fellous JM, Kruger N, Von Der Malsburg C (1999) Face recognition by elastic
bunch graph matching. IEEE Trans Pattern Anal Mach Intell 19(07), July 1999
5. Gunetti D, Picardi C (2005) Keystroke analysis of free text. ACM Trans Inform Syst Secur 8(3)
6. Dowland PS, Fvrnell SM, Papadaki M (2002) Keystroke analysis as a method of advanced
user authentication and response. In: Proceedings of the IFIP TC11 17th information security
7. Monrose F, Rubin A (1997) Authentication via keystroke dynamics. In: Fourth ACM
conference on computer and communication security, April 1997
8. Shen C, Cai Z, Liu X, Guan X, Maxion RA (2016) Mouse identity: modelling mouse
interaction behavior for a user verification system. IEEE Trans Human Mach Syst 46,
October 2016
9. Bours P, Fullu CJ (2009) A login system using mouse dynamics. In: Proceedings of the 5th
international conference on intelligent information hiding and multimedia signal processing,
pp 1072–1077
10. Aksari Y, Artuner H (2009) Active authentication by mouse movements. In: Proceedings of
the international symposium on computer and information sciences
11. Gamboa H, Fred ALN, Jain AK (2007) Web biometrics: user verification via web
interaction. In: Proceedings of the biometrics symposium, p 16
12. Fridman L, Stolerman A, Acharya S, Brennan P, Juola P, Kam M (2015) Multi model
decision fusion for continuous authentication. Comput Electr Eng
13. Bours P, Mondal S (2015) Performance evaluation of continuous authentication system. In:
IET Biometrics, pp 1–7
Hardware Partitioning Using Parallel Genetic
Algorithm to Improve the Performance
of Multi-core CPU
Suryakanta Nayak(&) and Mrutyunjaya Panda
Department of Computer Science and Application, Utkal University,

Bhubaneswar, India
mrutyunjaya74@outlook.com
1 Introduction
Hardware platform comprises the outstanding digital accessories that perform software
application programming packages. Owing to the need of limiting the development
time of digital products, advancement in CAD tools for hardware architecture is a
major challenging matter [1–3]. VLSI circuit partitioning is a key technique owing to
fact that it regulates the components which to be applied on hardware circuit, which
works on the software program. Over the last decades, several researches had been
implemented for HW partitioning [4, 5]. Further, HW partitioning can confirm that
faster throughput, less cost and efficient performance. At the commencement and
procedure stage, HW partitioning was taken as a graph partitioning representation. The
partitioning procedure is demonstrated as an undirected graph partitioning [6]. More-
over, specified model is marked into two unique kinds of HW partitioning issue, out of
which first one may be determined ideally by polygon time complex in nature, whereas
the second one is basically used to NP-hard problem [7]. Moreover, it is pertaining to
mathematical aspects of partitioning, and two types of classifications for calculations
are used, specific accurate techniques and approximate approaches which is applied
with heuristic calculations. The predetermined calculation is utilized to obtain excep-
tional goals for the small size. The common exceptional ways are involved in dynamic
programming branch and bound [8, 9] and linear programming [10]. In the partitioning
process, the deficiency estimation and the arrangement region of HW partitioning are
incremented exponentially. Heuristic calculations are used in mode simulation choices
because of their modern ability to procure the best choices inside the restricted pro-
cessing time. Aside this, the commencement stage of hardware circuit partitioning
where space is required in heuristics, including tools arranged for heuristic calculation,
had been proposed [11, 12]. The very beginnings complete challenges in product goals
and continuously moves from one to other the technique, while the last arrangement
begins with a product application converting segments to equipment. The equipment
placed in the system is treated as a presentation requirement process. The software
application technique is treated as a period limitation process. After that point, many
general heuristics and meta-heuristics methods had been additionally evolved, like
hereditary calculation [7, 13], simulated annealing [14, 15], particle swarm

https://doi.org/10.1007/978-981-15-2774-6_55
468 S. Nayak and M. Panda
optimization (PSO) [16]. In the present work, Sect. 2 discusses related work, Sect. 3
discusses proposed methodology, Sect. 3.1 discusses multi-core correction by genetic
algorithm, Sect. 3.2 discusses operations of hardware partitioning, and Sect. 5 narrates
conclusion.
2 Related Work
The application of segmentation to an undirected graph is nicely demonstrated in [7],

where it is described the segmentation with the guide of measurements are, commu-
nication price (CP), hardware partitioning (HP), which might be developed. So, it deals
with the insufficiencies of the above procedure, an inclining framework fixated based
upon the two-dimensional search was suggested to take care of the arrangement space
of problem (P). Further, the strategy for searching of P was governed with the guide of
the goals of boundary P0 [7]. Summing up the advantages of direct method and indirect
method, at first, the solution of the problem is investigated by genetic algorithm, and
next, the development idea is adopted from the second indirect ways for problem-
specific constraint matter. More often, our improvement strategy is unlike that of
existing methods [12, 13], and simply one impracticable solution is corrected. Here, a
number of impracticable solutions are corrected. In this way, plan can improve the
search concentration and ultimately improve the solution quality and address the
constraint concern. Moreover, techniques are projected to equal the architecture of CPU
hardware and reduce the run time, otherwise it would be time taking. In our framework,
various infeasible arrangements are adjusted. In this strategy, our procedure can expand
this pursue. A novel parallel example for our revision methods is proposed to analyze
great the structure of CPU and GPU equipment. The analyses confirm the adequacy of
our framework. In multi-core CPU and GPU has emerged as a well-liked parallel
computing platform with low energy absorbs and a high ratio of performance to cost.
This is a vital role in the science and engineering domains [17, 18]. For that reason, we
determine the process of use of how to use a parallel multi-core CPU and GPU on
individual computers for hardware partitioning is an attractive matter. The present
research proposes a brand name—new CPU and GPU by employing parallel genetic
algorithm scattering revision for HW apportioning. The difference in structure among
CPU and GPU information said that switch among CPU and GPU is through PCI-E
bus. The design of the CPU is limpness arranged. At the point when there is complex
decision-making ability in programming applications, the structure of CPU shows
enormous information over that of GPU. The significant matters are that CPU is
throughput oriented, and it measures CPU performance, the primary part supervises the
nodes and keeps a large portion of the region, to minimize the circuit even as in a
introverted chip of CPU, number reduction of time and space majority of the circuit.
There are more than a couple of parallel programming dialects helping multi-core CPU
and GPU. This paper represents that how to compute unified device architecture.
Hardware Partitioning Using Parallel Genetic Algorithm … 469
3 The Proposed Method
In a circuit partitioning, meta-heuristic technique is utilized for calculation of

throughput by using genetic selection. The basic components inside the genetic cal-
culation integrate replacement, amalgam and change. In our framework, these center
components are adjusted for HW partitioning. We stop our calculation when the same
type of result will be done, in minimization techniques. In this proposed technique,
calculation is done with scattering rectification for the HW partitioning issue. It
introduces a basic and simple to utilize a component of multi-stringing on more than
one CPU when parallelizing a sequential application. Hence, a feasible calculation
must rationalize the exchange of fixed cost among GPU and CPU. In this original copy,
we planned an innovative CPU and GPU parallel example to assist genetic calculation
through dispersion revision. The interior elements in the genetic search contain
crossover, reproduction and mutation (Fig. 1).
3.1 Multi-Core Correction by Genetic Algorithm

In normal generic calculation, an underlying factor is the initial element to the sub-
sequent stages. Let us consider for hardware partitioning, x represents (x1, x2, xn). Here,
x means an HW partitioning of concern P. It proposes that the node Vi is allocated to
hardware partitioning. Subsequently, the length of communication is equivalent to the
number of nodes in the task graph. It recommends an outline of randomized members.
The rest of the graph are produced to utilize arbitrarily arrangement of their properties,
two qualities are inside the circuit first, and second is distinct and arbitrary calculation
said that population cost is high. The full-adder circuit will take as our information
structure. Figure 2 demonstrates the 17 hubs, and the underlying allocation is finished
by the hereditary calculation depends on the guideline of typical inherited qualities and
predictable assurance (Fig. 3).
3.2 Genetic Operations for Hardware Partitioning

At every single cycle, common genetic calculation chooses two groups in the present.
For registering the product expense and the equipment cost, they consent to the
identical technique because of the calculation of the equipment charge. In Fig. 4
illustrate the computing cost, generally position of hardware partitioning initialization
are required for computation. For the processing, the discussion rate relies upon the
representation of the endeavor diagram on CPU. The weight on every single region
must indicate the discussion rate between couples of hubs. The consecutive method of
correspondence rate is O (n2). In view that two hubs of each characteristic are taken by
methods for their strings, the information situation of the response is the vector set on
mutual memory. In this manner, the summary of every single feature is completed
strings.
Algorithm 1
1. residual 1 = R – (S(y) + C(y))
2. at the same time residual 0 three.
3. for each and every yi = 0
Fig. 1. Proposed model of hardware partitioning
4. do
5. assume yi = 1
6. calculate Oci and hello/(si + Oci)
The goal is our to minimizing the hardware cost, for HW partitioning, the smaller
the hardware fee, the more the saved hardware cost.
Fig. 2. A full-adder circuit
Individual 1 1 0 1 0 1 0 1
Individual 2 1 0 1 0 1 0 1 0
Individual n 1 0 1 0 1 1 1 0
Fig. 3. An initialized population for HW partitioning
Fig. 4. Computing cost on CPU
Algorithm 2
1. residual1 = R – (S(y) + C(y))
2. while residual 0
3. for each yi = 0
4. do
5. assume yi = 1
6. Compute Oci and hi/(si + Oci)
7. decide node k with maximum residual
8. end for
9. if exists node k then
10. if exists node k then

11. update H (y),S(y), C(y)
12. residual
13.update H (y), S(y),C(y)
14.else
15.else
16.break
17.end if
18.end while.
The proposed methodology is consolidated the advantages of both the immediate

techniques and roundabout strategies. The current deviating strategies are outperformed
by the listed methods [6, 7], which is a realistic method. In the present research
calculating data transfer to the CPU and GPU, minimization of the equipment cost is
our objective. It is calculated in the first output to represent data transfer send to GPU.
The technique is the high caliber and straightforward. The figure four represents the
performance analysis in between array size and speed. It recommends that the
appropriate response quality by utilizing our system is the high caliber in general cases.
In the figure five represents the data transfer to CPU and GPU simultaneously, x-axis
represents the array size in byte, y-axis represents transfer speed GB/s, and in the read
write data transfer, CPU is better than GPU. Aside from, taking the data transfer
technique, the current ways are better than the proposed hereditary calculation in [7].
4.1 Physical Realization of the System

See Figs. 5 and 6.
5 Conclusion
The present investigation accomplishes a CPU and GPU parallel hereditary calculation
with inclusion of scattering rectification for the hardware dividing issue. The work is
portrayed as per the following: at first, a more grounded inherited calculation with
distribution adjustment is exhibited. Further, including preparing costs calculation and
dispersion revision are kept running in parallel. At last, we have displayed a novel
parallel technique by method for utilizing the corresponding intensity of multi-center
CPU as well as many centers GPU. Finally, the suggested method reveals that at the
appointed time an innovative substitute design for CPU and GPU.
Fig. 5. Relationship between array size and data transfer speed
Fig. 6. Relationship between read write bandwidth and data transfer speed
References
1. Wolf WH (2011) Hardware-software co-design of embedded systems. Proc IEEE 82:968–
989
2. Wolf WH (2017) A decade of hardware/software co-design. Computer 36(4):38–43
3. Teich J (2018) Hardware/software co-design: the past, the present, and predicting the future.
Proc IEEE 100:1411–1430
4. Mhadhbi I, Ben Othman S (2016) A comprehensive survey on hardware/software
partitioning process in co-design. Int J Comput Sci Inf Secur 14(3):263
5. Wang R, Hung WN (2016) Uncertainty model for configurable hardware/software and
resource partitioning. IEEE Trans Comput 65(10):3217–3223
6. Arató P, Juhász S (2003) Hardware-software partitioning in embedded system design. In:
Presented at the IEEE international symposium intelligent signal process., Budapest,
Hungary, Sept 2003, pp 197–202
7. Arató P, Mann ZA, Orbán A (2015) Algorithmic aspects of hard-ware/software partitioning.
ACM Trans Des Autom Electron Syst 10:136–156
8. Shi W, Wu J (2016) Algorithms for bi-objective multiple-choice hardware/software
partitioning. Comput Electr Eng 50:127–142
9. Trindade AB, Cordeiro LC (2016) Applying SMT-based verification to hardware/software
partitioning in embedded systems. Des Autom Embedded Syst 20(1):1–19
10. Mann ZA, Orbán A (2012) Finding optimal hardware/software partitions. Formal Methods
Syst Des 31(3):241–263
11. Gupta RK, Micheli GD (1993) Hardware-software co synthesis for digital systems. IEEE
Des Test Comput 10(3):29–41
12. Ernst R, Henkel J, Benner T (1993) Hardware-software co synthesis for microcontroller.
IEEE Des Test Comput 10(4):64–75
13. Janakiraman N, Kumar PN (2014) Multi-objective module partitioning design for dynamic
and partial reconfigurable system-on-chip using genetic algorithm. J Syst Archit 60(1):119–
139
14. Abdelhalim MB, Habib SED (2011) An integrated high-level hard-ware/software partition-
ing methodology. Des Autom Embedded Syst 15: 19–50
15. Yan X, He F, Hou N, Ai H (2017) An efficient particle swarm optimization for large-scale
hardware/co-design system. Int J Cooperat Inf Syst 9:59–64
16. Henkel J, Ernst R (2009) An approach to automated hardware/software partitioning using a
flexible granularity that is driven by high-level estimation techniques. IEEE Trans Very
Large Scale Int (VLSI) Syst 9(2): 273–28
17. Wang G, Gong W, Kastner R (2006) Application partitioning on programmable platforms
using the ant colony optimization. J Embedded Comput 2(1):119–136
18. Ferrandi F, Lanzi PL, Pilato C, Sciuto D (2013) Ant colony optimization for mapping,
scheduling and placing in reconfigurable systems. In: Presented at the IEEE NASA/ESA
conference adaptive hardware system, Turin, Italy, (2013), pp 47–54
19. Koudil M, Benatchba K, Tarabet A, Sahraoui EIB (2007) Using artificial bees to solve
partitioning and scheduling problems in codesign. Appl Math Comput 186:1710–1722
20. Wang N, Liu C, Yuhong L, Shen J (2017) A visible light communication (VLC) based
intelligent transportation system for lorry fleet. IEEE, ICOCN, pp 1–3
An Internet of Healthcare Things (IoHT)-
Based Healthcare Monitoring System
Harleen Kaur1(&), Mohd. Atif1, and Ritu Chauhan2

1
2
Centre for Computational Biology and Bioinformatics, Amity University,
Noida, U.P, India
1 Introduction
Traditional strategies of supplying protection cannot be directly carried out in IoT

because of extraordinary requirements and verbal exchange stacks involved. Infor-
mation and communication technologies (ICTs) deployed as part of medical records
systems must assure various sizable safety requirements together with integrity, con-
fidentiality, availability, authorization, non-repudiation, and authentication duty with a
purpose to comfortable medical facts without impacting the efficiency of services and
privacy of sufferers’ records.
The primary problem that each affected person mainly residing in faraway places
observed turned into unavailability of medical doctors and remedy on crucial situations.
This had very dreadful effects on human’s mind about the hospitals and medical
doctors’ offerings. Nowadays, with the implementations of new technology by using
making use of IoT devices for healthcare monitoring machine, these problems have
been taken care of to massive quantity. IoT is now not able to keep patients safe and
healthy most efficiently but to enhance the proper way doctors provide care. Healthcare
IoT can also enhance patient engagement and pride by allowing patients to spend more
time with their doctors. The Internet of things (IoT) is a widespread environment in
health care. More incorporated blessings and tactics with a position for the so-called
Internet of healthcare things (IoHT) or the Internet of medical things (IoMT) are sought
within the general context of medical care and eHealth.
Kevin Ashton coined the time net of factors for the first time in 1999. The RFID
organization, based upon fashionable conversation protocols, defines the Internet of
factors as the worldwide network of interconnected devices. It contains conventional
areas such as embedded systems, controlling and automation systems, and wireless
sensor networks to facilitate network conversation between devices (D2D). The idea
was first used at MIT’s Auto-ID center. RFID has become visible as a precondition for
implementing structures classified as IoTs. RFID is an IoT. Today, it has both private
and corporate customer programs.
The development of miniature devices has been brought about by improvements
and the convergence of micro-electromechanical systems (MEMS), wireless

https://doi.org/10.1007/978-981-15-2774-6_56
476 H. Kaur et al.
Fig. 1. Internet of things
communications, and digital electronics. They are capable of feeling, calculating, and
talking wirelessly. These small devices are referred to as wireless sensor networks
(WSNs) nodes [1].
In the Internet for things (IoT), it is easy to collect, file, and analyze new streams of
information faster and more precisely by making gadgets quickly collect and share
statistics with each other and the cloud. Numerous application domains take place on
the Internet of things. The packages can be classified mainly based on Web access,
insurance, scale, diversity, lack of originality, participation with individuals, and
implications [2].
2 Related Works
Firstly, in 1999 Kevin Ashton suggested an IoT idea and described IoT as a unique
identification of the RFID generation, loosely coupled linked accessories. Luigi et al.
addressed on the IoT and the combination of several technology and communication
solutions that make this promising paradigm possible. Identification and tracking
technology, wired and wireless networks for sensors and actuators, improved com-
munication protocols, and intelligence for smart products maximum applicability. The
fundamentals of IoT as the aggregate of net and the rising technology have been
mentioned [3].
Shen has studied that the eHealthcare gadget contains three domain names in
particular: frame region, conversation and connectivity, and provider. A number of
wireless placement (WAAN) networks, each similar to a consumer, describe the
framework location domain. The key function of the communication and networking
An Internet of Healthcare Things (IoHT) … 477
domain is to connect frame locations and provider domains. WAAN gateways link to
the Web and allow for the efficient communication of mutual data between WAANs by
advanced Wi-fi technologies. In the service field, an authority reliance maintains an
online server receiving personal health and fitness information and records [4].
The IoT platform’s structure and problems in the design of IoT’s hardware and
software additives have been addressed [5]. The IoT platform’s structure and problems
in the design of IoT’s hardware and software additives have been addressed. They have
worked out the various IoT utilities, together with intelligent cities, healthcare, farming,
and nanoscale programs. In their paper, the country of IoT artwork was studied and
numerous key drivers of technology were provided. The possibility of lowering the
DTLS overhead by means of compression of 6LoWPAN headers was investigated, and
the first DTLS compression specifier for 6LoWPAN was presented. The updated
hardware, communication and computing requirements for uHealth’s next technology
systems have been fully assessed. Instead, they talked about new technological and
technological trends and how to deal with health predictions.
Current strategies have been studied to design efficient and comfortable eHealth
monitoring. They provided a comprehensive framework for superior eHealth moni-
toring by describing the entire life cycle in detail. They also highlighted the critical
supplier additives, accurately recognizing the data series in the aspect of the individual.
In order to make the proposed framework more efficient, we have offered and discussed
the major challenges to solve, in order to develop a green and relaxed person-oriented
tracking machine [6].
First, there is a description of the safety and privacy issues of the use of the frame
sensor community (ASN) in health applications. Later, although most of the popular
ASN studies have fully recognized the security challenges, they found that they still fail
to incorporate robust safety offers that could preserve the privacy of individuals
affected. Finally, it was proposed that the ASN, called ASNCare, could efficiently
fulfill many safety needs of the ASN-based healthcare system, using mainly aligned
IoT device [7].
The first vulnerabilities, including the latest TMIS protocol proposed in the liter-
ature, were explored. The attacks proposed mainly on the basis of stamping weak-
nesses, reader request calculation, and tag answer messages were not monitored in
detail using a one-way hash function. Second, they proposed an efficient protocol to
protect health and high performance for cell authentication dual RFID-TMIS. Its
inspiration was to develop and expand the previous protocol in which TMIS was
designed to use both promising technologies with the same authentication device to
support the RFID era. The evaluation showed that the progressive protocol should
address the weaknesses in protocol protection studied and provide mobility and effi-
ciency [8].
In order to eliminate these vulnerabilities, a new authentication protocol on radio-
frequency identification has been proposed primarily on the basis of elliptic curve
encryption (ECE). In order to generate a shared short key for encrypting subsequent
messages, they have also used an elliptical curve Diffie–Hellman (ECDH) key settlement
protocol. Their protocol, carried out in a set of safety homes, likes mutual authentication,
anonymousness, confidentiality, forward safety, privacy, middle class attack resistance,
resistance to replay and impersonation attack resistance. The consequences of
478 H. Kaur et al.
implementation suggest that our proposed protocol exceeds complexity over time
compared with other similar protocols and calls for significantly less operational
volatility [9].
There is a comfortable IoT framework that ensures end-to-end protection against
IoT devices from an IoT utility is proposed and is composed of IoT, IoT and IoT
devices. The IoT broker manages its personal devices and collects sensing information.
The IoT app offers IoT offers to customers. It wants access to sensing information to
utilize IoT services. In particular, intermediate protection problems should be taken into
account when offering real-time healthcare services since patient’s medical facts are
surely a very sensitive data on privacy. The maximum IoT protocols in terms of end-to-
end security, including CoAP and MQTT, however, were of greatest importance to
DTLS safety. Consequently, we proposed a new IoT framework which would comply
with the safety feature of the end-to-end CoAP communication. A symmetrical
encryption and mainly based encoding of attributes are encoded for oral exchange and
computing efficiencies in this context. In addition, every IoT tool has as one of its
attributes a completely unique identity. Thus, while the IoT broker is a medium node,
only decrypt and display facts when all the attributes are met [10].
A secure user profiling structure [11] was provided which includes personal
information on the individuals involved. At the same time, an affected person and a
clinic are keeping the information up-to-date. Data can be leaked while they are being
percented and communicated. To solve the problems of security, the safe communi-
cations channel must be mounted and an OTP input price generated using a two-way
hash feature, with a single password between a boss and health center. This work
provides a dual hash functional method for creating a one-time password to ensure that
the secured key is relaxed on the channel. As a result, attackers cannot decrypt the
information leaked due to the secured key and the proposed method exceeds existing
strategies in calculated value phrases. The Internet of things is being used; a concept
within the health field is not now without further data and therefore overhead transfers
of facts. In an attempt to maximize capabilities and broaden reputation/use of things
that have been provided by the Internet, new metrics and methods are added to cope
with these overheads. The above technique is evaluated experimentally within the
paradigm of health without losing its generality. The focus is on an overview that
contains statistical codecs and transmission methods and the choice of the highest
combination, which can result in cost reduction/minimization. An analytical method is
submitted and experimentally evaluated using theoretical metrics.
Various famous ICT paradigms, including cloud, IoT, and big data, have been
mentioned [12]. It provided a thorough evaluation of the works of art and their con-
vergence. Then, they proposed a M2M gadget for electronic health programs, based on
the decentralized cloud structure, fashionable systems, and RTUs. The device has been
designed in such a way that data can be aggregated to create virtual sensors for big data
processing of sensor records, and some dimension implications were supplied.
The data protection problems of the patient have been diagnosed [13], and their
corresponding mechanisms have been found in the literature selected. The evaluation
found that IAE, AAE, and its variations are the most used way to deal with the privacy
concerns of the affected person in health clouds. Other strategies based on theoretical
modes and frameworks which do not use any encryption strategy are not therefore
implemented in the current world situation.
A light smash glass has been suggested to manipulate (LiAAC), which will help to
provide access to encrypted medical files by entering and accessing damage glass
attribution-based files [14]. In daily situations, a scientist with an attribute set can
decipher and get admission to the information, which please the medical file policy.
Under emergency conditions, the access mechanism for break-glasses bypasses the
right of entry to clinical report coverage to allow a timely transfer of data via emer-
gency hospital therapy or rescue. LiAAC is lightweight because the overheads for
garage and transmission are low and only a few calculations are made through devices
within the healthcare network. The LiAAC version is officially proven comfortable,
and extensive experiments in the famous version are carried out to show its
performance.
A light smash glass has been suggested to manipulate (LiAAC), which will help to
provide access to encrypted medical files by entering and accessing damage glass
attribution-based files [15]. In daily situations, a scientist with an attribute set can
decipher and get admission to the information, which please the medical file policy.
Under emergency conditions, the access mechanism for break-glasses bypasses the
right of entry to clinical report coverage to allow a timely transfer of data via emer-
gency hospital therapy or rescue. LiAAC is lightweight because the overheads for
garage and transmission are low and only a few calculations are made through devices
within the healthcare network. The LiAAC version is officially proven comfortable and
extensive experiments in the famous version are carried out to show its performance.
A review of and analysis of the weakness of an authentication protocol for WMSNs
is mainly based on the medical services [6]. They found that the protocol within the
authentication and session key agreements phase was not always correct so that Ui and
Sn could not authenticate each other correctly and could not agree on a consultation
key. The biometrics are provided as the third aspect of authentication [11] and in
comparison with preceding protocols the new consumer authentication protocol, which
is based primarily on WMSNs is designed in an attempt to eliminate drawbacks of (He
2015) protocol.
The proposed [10] scheme was analyzed, and its scheme was indicated to suffer
from a repeat attack and to be deficient. Crypto et al. proposed a reliable power-efficient
control system to overcome the weaknesses of the system for wireless sensor networks.
In addition, their new scheme has proven itself easy to cope with many kinds of attacks.
The framework for coalesce storing personal and oral exchange has been proposed
instead of developing uneven safety answers for garage and verbal trade, [9]. Fusion
has been suggested. Fusion utilizes existing IoT relaxed conversation protocols con-
sisting of Internet Protection Protocol (IPsec) and DTLS, and re-uses defined storage
conversation protection mechanisms. Thus, advanced security frameworks for com-
munication extend to the storage area. With this mechanism, the records requested can
be transmitted from the recording device directly without decrypting statistical blocks.
As a consequence, Fusion offers speed and performance advantages that are essential
for useful IoT devices with limited resources.
480 H. Kaur et al.
An upheaval [16] method was derived called GSD technique for the detection of
malicious attacks in IoTs which include intrusion detection, based entirely on mathe-
matical morphology. They have successfully produced GSD clusters to show the
number of energy nodes in a wireless sensor community at once because the GSD
curves are comparable while the amount of power nodes in the wireless sensor network
is consistent. The facts of each node are used because the network parameters are used
in this method.
Various aspects of IoT technology have been examined, and diverse architectures
and structures have provided for the healthcare network to facilitate the access of the
IoT backbone and the transmission and reception of scientific records. In the IoT-
driven fitness services and applications, substantial R&D efforts were made. The paper
also provides research on how the IoT can address pediatric and old people’s treatment,
persistent health surveillance, personal health, and health management. The paper
presents a broad overview of the way that recent and ongoing progress has inspired
cheap healthcare gadgets and linked health services to unlimitedly increase the capacity
of IoT-based mainly healthcare services to further developments for deeper insights
into business developments and technological enabling solutions. The paper considers
various safety requirements and challenges to increase the understanding of IoT
healthcare safety and reveals extraordinary research problems in the vicinity that could
suggest a version that mixes the related safety dangers. The dialog on many critical
issues, including standardization, the community type, business model, the quality of
the carrier, and protection of health facts, will help to provide the basis for similar IoT-
based studies on entirely healthcare services. This paper provides rules and rules on
eHealth and IoT for the advantage of various stakeholders interested in evaluating IoT
healthcare technology.
3 Need and Challenges of IoHT in Healthcare

A. Need of IoT in Healthcare
(i) In order to make better patient decisions, IoT ensures that all information is
taken into account. Networked health alternatives and fully interactive
technology strengthen and improve treatment effectiveness.
(ii) Sophisticated IoT technology enriches emergency treatment and makes
emergency assistance more efficient.
(iii) Sophisticated medical devices call on patients to use prescribed medication
on time and other essential health initiatives. IoT tends to improve patient
comfort.
(iv) IoT helps physicians to discuss urgent cases with specialists across the
world.
(v) Various sophisticated sensors and different equipments help doctors to
examine the patient’s health and his/her internal feelings with ease.
B. Challenges of IoT in Healthcare

(i) As the daily data generated is huge, therefore to maintain security and
privacy of data is the biggest challenge.
(ii) Number of medical devices can create network congestion due to overload
and can lower the speed of data transfer. This is due to the lack of
sophisticated protocols.
(iii) The cost of affording the IoT devices is yet a big challenge.
4 Conclusion
The long projected IoHT health revolution is already under way in this paper. As new
applications increase, the urgent demand for cheap, hand care continues to be
addressed. In the meantime, the automatic IoHT building blocks and verbal exchange
of device to machine are still hooked up are added bureaucracy to the service layer
throughout the IoT infrastructure. This revolution is characterized by providing IoT-
driven healthcare solutions for end-to-end treatment and connectivity. However, a
machine would like to be designed according to this assessment, where doctors can
analyze the critical parameters of the affected person in real time using comfortable
mechanisms elsewhere. A device must be designed to provide a whole security against
a certain attack, including access to the IoT-based healthcare control and authentication
protocol. IoHT demands are enormous and can massively support the healthcare sector.
Use of IOHT technology can reach every patient globally and communicate physicians
with patients. In future, it will keep growing and serve the purpose of treating patient at
ease. In coming year, hospitals will acquire the IoHT technology to strengthen their
medical facility and patient will get treated in less amount of time. IoHT helps to
enhance the interaction between patients and doctors.
References
1. Bagci IE, Raza S, Roedig U, Voigt T (2016) Fusion: coalesced confidential storage and
communication framework for the IoT. Sec Commun Netw, pp 2656–2673
2. Bandyopadhyay D, Sen J (2011) Internet of things: applications and challenges in
technology and standardization. Wireless Personal Commun, pp 49–69
3. Korteum G, Kawsar F, Fitton D, Sundramoorthy V (2010) Smart objects as building blocks
for the internet of things, pp 44–51
4. Liang X, Xu i, Shen Q, Lu R, Lin X, Shen XS, Zhuang W (2012) Exploiting prediction to
enable secure and reliable routing in wireless body area networks, pp 388–396
5. Benssalah M, Djeddou M, Drouiche K (2016) Dual cooperative RFID-telecare medicine
information system authentication protocol for healthcare environments. Sec Commun Netw,
pp 4924–4948
6. He D, Kumar N, Chen J, Lee C, Chilamkurti N, Yeo SS (2015) Robust anonymous
authentication protocol for health-care applications using wireless medical sensor networks.
Multimedia Syst, pp 49–60
482 H. Kaur et al.
7. Gope P, Hwang T (2016) BSN-care: a secure IoT-based modernhealthcare system using

body sensor network, pp 1368–1376
8. Li X, Niu J, Kumari S, Liao J, Liang W, Khan MK (2016) Anew authentication protocol for
healthcare applications using wirelessmedical sensor networks with user anonymity. Sec
Commun Netw, pp 2643–2655
9. Alamr AA, Kausar F, Kim J, Seo C (2016) A secure ECC based RFID mutual authentication
protocol for internet of things. J Supercomput, pp 1–14
10. Garkoti G, Peddoju SK, Balasubramanian R (2014) Detection of insider attacks in cloud
based e-healthcare environment. Int Conf Inf Technol, pp 195–200
11. Ko H, Song MB (2015) A study on the secure user profiling structure and procedure for
home healthcare systems. J Med Syst, pp 1–9
12. Cvitic I, Vujic M, Husnjak S (2016) Classification of security risks in the IoT environment.
In: Proceedings of the 26th DAAAM international symposium on intelligent manufacturing
and automation, pp 0731–0740
13. Choi J, In Y, Park C, Seok S, Seo H, Kim H (2016) Secure IoT framework and 2D
architecture for End-To-End security, pp 1–15
14. Bruce N, Sain M, Lee HJ (2014) A support middleware solutionfor e-healthcare system
security. In: 16th international conference on advanced communication technology
15. Chi L, Hu L, Li H, Sun Y, Yuan W, Chu J (2013) Improvedenergy-efficient access control
scheme for wireless sensor networks basedon elliptic curve cryptography. Sensor Lett,
pp 953–957
16. Islam SMR, Kwak D, Kabir MH, Hossain M, Kyung-Sup Kwak, KS (2015) The internet of
things for health care: a comprehensive survey. IEEE Access, pp 678–708
Botnet Detection Technology Based
on DNS-Based Approach
Bhavya Alankar(&)

1 Introduction
All the businesses and services are dependent on the Internet and the servers that store
the data in the cloud and provide easy access across the globe. It gives several services
that are extremely helpful to use. The Internet may be helpful positively, but the cyber
crimes are prevailing across it as well. The confidentiality and integrity of the data can
be harmed by doing breaches in the information security, theft of identity, and several
other attacks. The person who attacks also known as botmaster spread Trojan or
malwares or both which increases the bots number present in the network. Bot in the
botnet is referred to robot and net to the Internet which states that botnet is a robot
network or computer/servers where attacker controls and gains the access of the sys-
tems of the network, while the end is not even known to it.
The bots are used remotely through command and controlling server which is
controlled by the attacker. The attacker uses one of the bots of the botnet which is the
command and control server to control and communicate with the other bots through
instructions. The bots can be singular as well as in a group at the same time through
commands. Botnet’s size can be increased compromising the devices or servers in the
network. Botnet sustains the property of propagation which helps it in spreading all
over the Internet [1]. The attacks that can be done through botnets are the fraud of
phishing, fraud of clicks, stealing of password, spamming, fraud of bitcoin, theft of
mass identity, traffic sniffing, new malware spreading, and logging of the key.
Internet Relay Chat (IRC) and Hypertext Transfer Protocol (HTTP) are the com-
munication protocols for botnets [1]. The widely used botnet protocol is IRC as they
are extremely popular and it can be easily found for use by a botmaster. IRC server is
easily found and taken down if the botnet is detected, which makes a disadvantage of
IRC. A zombie computer is a name given to the compromised systems; hence, botnet is
also known as a zombie network. The most crucial time in the making of the botnet was
the first five years. The first botnet is created by the attackers which was called as
“Eggdrop” in 1993. More advanced botnets are created after that which have new
features and functions until 2002. During these years, the use of botnets was started by
most of the attackers which resulted in a rapid increase of cyber attacks. The world has
not got sufficient data and a convincing procedure for detecting a botnet. The hazard of
botnet is not strange to anyone as they are existing from a long time; there is a lot of

https://doi.org/10.1007/978-981-15-2774-6_57
484 B. Alankar
research that needs to be done to put a check on them [1]. Nowadays, the things are
adding on to check the botnet in the server.
A lot of research work has been done on botnet detecting ways. However,
researchers have advised ways that are distant from each other on botnets. It has been
examined that to notice and audit botnets, two major ways are used [3, 6]. The major
first way is using honeynets by installing them into the servers. This honeynet will be
set up to focus on collecting reports of botnets and know the ways they behave. They
may not be able to find botnet, but it collects the information which is further used to
create a protection for botnets [7]. A setup with intentional vulnerabilities is known as
honeynet. It main motive is to get attacked by the botnets; this enables us to study the
activities of the attacker and several methods can be developed to secure the network.
The second approach for botnet detection is established on quietly checking the net-
work movement and analysing it. The approaches which follow the ways for detecting
a botnet are the anomaly-based and Domain Name System (DNS)-based detection
techniques which we have studied further in the paper. The movement of the DNS
server and IDS server is been tracked, and any abrupt behaviour is saved to further
study the attacker and develop methods to secure the network. This approach is more
efficient as compared to other approaches.
2 Background
A lot of research has been done in botnet detection. We focus on different techniques of
botnet detection and botnet suppression in this section [1–3] gave the mechanism of
various botnet detection techniques. A study also gave network-based botnet detection
techniques in which network traffic is inspected related to IRC protocol which may
sense the presence of a botnet. The method of detection of a botnet which related to
alarms when an intruder from a different network.
IRC botnet deployment is done all around in past years. The use of IRC protocol as
a C&C method spreads because of its versatility, redundancy, and scalability. There are
a large knowledge and code base for IRC-based bots, which enables botnet creators to
reuse the source code and create a new botnet, for example, Agobot variants. A well-
designed and modular Agobot’s code is online available, which makes it easy to create
their own botnet for botnet author.
The fundamental weaknesses of an IRC botnet are that stem from its C&C servers
which are centralized. The source of command is easily identified and a single point is
possessed by these servers, which make it very easy to disrupt a botnet which is based
on IRC and get the IP addresses of bots present in a botnet. Botnets which use IRC as
communication means for C&C are easy to detect relatively to others. The hostility is
adapted by the botmasters towards IRC bots by somehow optimizing the IRC botnet’s
C&C architecture or by creating a new botnet. The shift currently is to decentralized
P2P C&C architectures from centralized client–server. The IRC-based bots are being
changed by the authors to make them resilient because of which IRC bots are still
seeing widespread use P2P botnets are documented and discovered in the wild, e.g.,
Phatbot, Peacomm, Sinit, Slapper, and Nugache. The shift is already made to a new
Botnet Detection Technology Based on DNS-Based Approach 485
botnet C&C control methodology, i.e., P2P C&C. Following are the approaches which
follow the ways for detecting a botnet:
1. Detection based on signature
For detecting botnets, this approach tries to find a signature. It gives us the important
instruction around the form of botnet intrusion, and it is easy to apply. This approach
detects the botnet at very huge per cent, and most of the detection detected by this is
real. No new botnet can be detected by this approach as it has signature saved for only
already familiar botnet which can be the only defect in this approach. The correlation
produces signature for IDS system. The strategy driven by IDS to detect botnet has
problem. The signature produced by the IDS system sometimes causes problem in
detection as they can be in encrypted form and it can be difficult to decrypt the
signature. The signature or sign provided by a botnet after detection is saved into the
database on this technique, and the resistance is created for that particular botnet which
is the main drawback as it is related to a particular botnet only [19, 20]. If another
botnet of same type attacks the system again, then it is not able to detect it as it has
signature saved of only one botnet of that type, the botnet can attack again and again,
and this system will not be able to detect the botnet.
2. Detection based on anomaly
The botnet is detected in this approach by movement irregularity, also known as
anomaly on the net [6]. This approach works by analysing the whole grid and finds
each irregular act of the movement all over the grid. The defect in detection based on
signs is also removed by this approach as it has the ability to find botnets that are never
used in any intrusion. It costs more as compared to detection based on signature and
also superior to it in detecting a botnet. The main two proposed anomaly detection
methods are based on resulting large deviations for packet level and flow data. Ano-
maly represents a set of records of interaction in both types of anomaly detecting
method.
3. Detection based on DNS
The approach is based on DNS, where botnet is detected by getting the data of the DNS
and checking if a botnet connects to the DNS or not which is not like above-mentioned
approaches. Detection-based DNS and anomaly are similar as procedure practiced in
detection through anomaly is practiced in DNS movement too. The group of hosts
which behave abruptly are taken into account by this method which is like botnets.
DNS TTL is not honoured in these group of hosts, and DNS queries are carried out to
servers which are not local. This method looks into large number of response from
DNS with NXDOMAIN as code of error. This technique is able to detect botnet with
high efficiency.
486 B. Alankar
3 Life Cycle of Botnet Technology
The botnet life cycle is a procedure for adding a new bot in a botnet. The different
phases for making a botnet or creating a new bot in a botnet are contained in a life
cycle. There are different researchers who state differently about the life cycle of botnet.
The distinct steps in a lifecycle of botnet are infection and injection, control and
command and application of botnet [1, 2]. Figure 1 describes about a botnet life cycle.
Fig. 1. Life cycle of a botnet
Other researchers state that the life cycle of the botnet is in five phases: infecting the
server, injecting to the server, connecting to the server, sending virus to the server and
maintaining and updating the botnet [3]. When the botnet infection is inserted into a
new device which is connected to the Internet, then the injection of virus into device by
HTTP protocol and P2P protocol connects to the C&C server of botnet. It starts
working as a zombie computer for botnet after malicious code injection into the system.
Now, the new victim device is controlled by the attacker through command and control
server. The masters keep the updating and maintenance work [1, 4].
Network Topologies in Botnets
When talking about botnets network topologies needed to be discussed because it
creates a huge difference in botnet performing its functions. Different topologies and
architecture have been used by attackers [5]. The main topologies botnets uses are:
A. Star Topology
A very good bot managing and communicating system is provided between the bots by
this topology. The big disadvantage of this topology is that it has only C&C server,
which creates the problem of failing at one place; i.e., the whole system breaks if C&C
gets blocked. The connection from botnet can be blocked by even a legitimate user
themselves [5]. The Main C&C contains all the information that is needed to run the
botnet. If the main C&C fails, then the botnet is of no harm. One does not need to
remove that if the main C&C is not working for a botnet in star topology. A number of
systems that are connected to this type of topology may vary; it can be as few as one or
can be in large numbers as hundreds. The risk of connecting several systems to this
botnet is high as it becomes dead if the main C&C server breaks.
B. Multi-server Topology
In this topology, multiple servers or C&C control and manage the network as a whole
which makes the communication system better between all C&Cs and bots in the
network. For some reasons, if a C&C fails, all other servers remain working and also
make a decision about removing the C&C which has failed. The risk failing of the
network at a point is removed in this topology [5]. This topology is highly efficient as
compared to the star topology. The C&C servers in this topology can be many which
can create a big network as each C&C server can connect to several other systems. If
the C&C server which has many systems connected to it fails, then it is replace by other
C&C servers which also connect to the other systems which were previously connected
to the failed C&C server and the operation continues without smoothly. Figure 2
describes above topology (Fig. 3).
Fig. 2 Star topology

488 B. Alankar
Fig. 3 Multi-server topology
C. Hierarchical Topology
In this topology, botnet contains multiple C&C servers are set in a way in cluster for
providing a reliable network. It is easier for a botmaster in renting a bot due to
hierarchical topology. Using hierarchical topology, botnets have many benefits as they
are not easily detectable. In botnets based on hierarchical topology, the location of the
bot agent is not known to the other bot agent which makes it tough to find a botnet and
it also does not reveal the botnet size [5]. The botnets using hierarchical topology are
very hard to detect as the C&C server location is not known if one C&C server gets
caught, then another C&C server cannot be detected. Each server has many systems
connected to it in a hierarchical way so that each level contains a C&C server and
makes the work faster and easier. The same C&C server can be rented out to many
clients and one server can be used to carry out functions for different clients at the same
time. It reduces the number of machines used in the botnet which makes it more
efficient than the star topology and multi-server topology. The cost factor also rises in
this type of topology as the server increases so it takes space to put data of several
clients into a C&C server and use it according to the functions and commands.
Hierarchical topology can be very useful and efficient as its servers are distributed all
over the network at different levels which reduce the risk of failure in the system
(Fig. 4).
D. Random Topology
Botnet agents spread the malware through the same procedures although this topology
does not contain any centralized C&C infrastructure. As per the condition, any zombie
computer can act as a C&C server. Bots can communicate through many paths and
ways, and in this topology, botnets are very difficult to detect and control as C&C
infrastructure is not centralized; i.e., a new C&C server will be replaced into the
network if a working C&C server is hijacked and is removed by new server. Bots
respond to the command given by C&C in more time as compared to other
Fig. 4 Hierarchical topology
topologies. It is still better to use as it provides many better functions and features than
other topologies of botnet [5]. This can also be known as dynamic master–slave or
peer-to-peer relationship topology. In the dynamic master–slave, master refers to the
C&C servers, while the slave is the system which is dynamic means the master–slave
keeps on changing. Any system can be a master or a slave at a particular instance of
time. The peer-to-peer relationship is that the system and server both are treated as
peers which can be related to each other in the same way at any given time. The peers
remain same as it can act as server and the system as well (Fig. 5).
Fig. 5 Random topology

490 B. Alankar
4 Problem Statement
The vital elements of our daily lives are the Internet and World Wide Web which are
critical to business and personal activities of the society, but they have also became a
home to cyber crimes. Several means have been developed which protect the network
against the attacks which include computer worms, viruses, scanning, and denial of
service, but more stealthy techniques have been developed by the attackers. The
attackers are motivated for cyber crimes by the inclusion of money in network.
A significant security threat on the network is botnet, and its operator is growing for
their own profit. One can attack a country or enterprise for political gain through
botnets by getting the secret data. Botnets exist not only in old well-known network but
also in a new created application states [15–17]. Signature-based techniques on selected
network or honeypots are used to detect botnet in the current state.
5 Proposed Solution
The proposed solution mainly consists of:

(i) C&C servers: The command and control servers are the servers that control the
botnet. All the instructions of the botnets like attacking a server, infecting a
server, and overtaking a server are passed through these servers only. These are
the headquarters of the botnets, and the botnet can be destroyed if its C&C
servers stop working.
(ii) Positive Bot: The bot that will be introduced by us that will detect the botnet if
any of it is present in the system. It then intrudes in the system and kills the main
C&C servers that are working with the botnet to infect the system.
(iii) Killed Servers: The C&C servers that are infected by our bot and will stop
working are known as killed servers. Our bot attacks those servers and kills them
by infecting virus inside them which make stop its working.
Procedure: Currently, the solution to remove the botnet is to reset the whole server
which will remove all the data from the server. This process may result in the loss of
some important data; to overcome this problem, the proposed solution is that we can
create a bot which will first detect the presence of botnets in the network and then
intrude into the network if it is present. It will then kill and replace a C&C server by
first infecting it with anti-bot code.
After intruding in the network, the positive will detect and kill the C&C servers in
the network of the botnet. In the botnets using star topology, it will be easy to remove
the whole network by detecting botnet, intruding into a that network, and killing a
server as there is only one server which will be removed by us, while in the botnets
using multi-server topology and hierarchical topology it will detect the botnet first,
intrude into the network by killing and replacing a C&C server, and then kill the
remaining servers one by one.
This will be a lot harder in the botnets using random topology as each and every
computer can act as a C&C server; for this topology, it can be a long task as it will have
to check and kill each and every system connected to the network for a botnet (Fig. 6).
Fig. 6 Detection of botnet
In this figure, we can see that our bot is detecting the presence of bot in the server.
This is mainly done through three techniques detection based on signature, detection
based on anomaly, and detection based on DNS [17] (Fig. 7).
Fig. 7 Intrusion of positive bot into botnet

492 B. Alankar
In this figure, we can see that out bot it trying to intrude into the network of the
bots. The positive bot intrudes into the system and infects the main C&C servers by the
virus so that they stop working and the botnet is totally malfunctioned (Fig. 8).
Fig. 8 Positive bot killing servers in the botnet
The positive bot finally kills the botnet by injecting the virus into the system of the
botnet and disabling the C&C server of the botnet which stops the functioning of the
botnet and hence make our system bot-free.
6 Conclusion
In this paper, we have analysed that what are the botnets and what threats they can
cause tous. This study discussed some techniques through which the botnets can be
detected and monitored and the network topologies used by the botnets in attacking a
server. In the end, we can say that the Internet has become a basic need to everyone
around the world nowadays and there are several works running across it. The cyber
thieves use new techniques every time to introduce and steal the data through botnets.
We need create new techniques to detect and remove the botnet from the server without
affecting our data and to safeguard our server.
Millions of users use the Internet globally as it is becoming more common in the
upcoming days. New threats are coming on Internet for attacking users. Botnets are into
Internet for quite a long time; still we did not see emergence of best detection tech-
niques. We need a technique to detect and kill the botnet in the network.
References
1. Rossow C, Andriesse D, Werner T, Stone-Gross B, Plohmann D, Dietrich CJ, Bos H (2013)
SoK: P2PWNED—modeling and evaluating the resilience of peer-to-peer botnets. In:
Proceedings of the 2013 IEEE symposium on Security and Privacy (SP), San Francisco, CA,
USA, 19–22 May 2013, pp 97–111
2. Hu X, Knysz A, Shin KG (2011) Measurement and analysis of global IP-usage patterns of
fast-flux botnets. In: Proceedings of the IEEE INFOCOM, Shanghai, China, 10–15 Apr 2011
3. Ji Y, He Y, Jiang X, Cao J, Li Q (2016) Combating the evasion mechanisms of social bots.
Comput Secur 58:230–249
4. Jerkins JA (2017) Motivating a market or regulatory solution to IoT insecurity with the Mirai
botnet code. In: Proceedings of the 2017 IEEE 7th annual Computing And Communication
Workshop and Conference(CCWC), Las Vegas, NV, USA, 9–11 Jan 2017, pp 1–5
5. Zhao G, Xu K, Xu L, Wu B (2015) Detecting APT malware infections based on malicious
DNS and traffic analysis. IEEE Access 3:1132–1142
6. Asha S, Harsha T, Soniya B (2005) Analysis on botnet detection techniques. In: Proceedings
of the international conference on Research Advances in Integrated Navigation Systems
(RAINS), Karnataka, India, 6–7 May 2016, pp 1–4. Saha B, A Gairola (2005) Botnet: an
overview. CERT-In White Paper CIWP-2005-05, 2005
7. Alomari E (2012) Botnet-based Distributed Denial of Service (DDoS) attacks on web
servers: classification and art. 49(7): 24–32
8. Barford P, Yegneswaran V (2006) An inside look at botnets. Adv Inf Sec, Springer
9. Plohmann D, Gerhards-Padilla E, Leder F (2011) Botnets: detection, measurement,
disinfection and defence. Eur Netw Inf Secur Agency Tech Rep
10. Gu G, Perdisci R, Zhang J, Lee W (2011) Botminer: clustering analysis of network traffic for
protocol and structure-independent botnet detection. Usenix Security Symposium, 2008.
BotHunter, Bothunter: a network-based botnet diagnostic system
11. http://www.bothunter.net/. (Online). Accessed 12 Dec 2011
12. Mody N, O’Reirdan M, Masiello S, Zebek J (2009) Common best practices for mitigating
large scale bot infections in residential networks. MAAWG, July 2009
13. Sakib MN, Huang C (2016) Using anomaly detection based techniques to detect HTTP-
based botnet C&C traffic. In: Proceedings of the IEEE international Conference on
Communications (ICC), Kuala Lumpur, Malaysia, 23–27 May 2016, pp 1–6
14. Stone-Gross B, Cova M, Gilbert B, Kemmerer R, Kruegel C, Vigna G (2011) Analysis of a
botnet takeover. IEEE Secur Priv 9:64–72
15. The Honeynet Project (2016) Know your enemy: fast-flux service networks. Available
online http://www.honeynet.org/papers/ff/. Accessed on 19 May 2016
16. Passerini E, Paleari R, Martignoni L, Bruschi D (2008) FluXOR: detecting and monitoring
fast flux service networks. In: Proceedings of the 5th international conference on Detection
of Intrusions and Malware, and Vulnerability Assessment (DIMVA ’08), Paris, France, 10–
11 July 2008, pp 186–206
494 B. Alankar
17. Wang T, Lin H, Cheng W, Chen C (2017) DBod: clustering and detecting DGA-based
botnets using DNS trafficanalysis. Comput Secur 64:1–15
18. Kolias C, Kambourakis G, Stavrou A, Voas J (2017) DDoS in the IoT: mirai and other
Botnets. Computer 50:80–84
19. Kwon J, Lee J, Lee H, Perrig A (2016) PsyBoG: A scalable botnet detection method for
large-scale DNS traffic. Comput Netw 97:48–73
Development of an IoT-Based Tourism Guide
System
Vedanta Prusty1(&), Abhisek Rath1, Kshirod Kumar Rout2(&),

and Sivkumar Mishra3
1
Department of ETC, IIIT, Bhubaneswar, Odisha, India
{b214059,b214003}@iiit-bh.ac.in
2
Department of EE, IIIT, Bhubaneswar, Odisha, India
kshirod@iiit-bh.ac.in
3
Department of EE, CAPGS, BPUT, Rourkela, Odisha, India
sivkumar@iiit-bh.ac.in
1 Introduction
Tourism is one of the promising areas over which the fulcrum of economic prosperity
of many countries, including India, depends. With an annual average GDP of 6–7%
since 2014, India’s tourism story is no alien when it comes to contribution in the
finances of the country. According to the World Travel and Tourism Council, tourism
not only produced US$ 210 billion, i.e., contributing 9.4% of country’s GDP, but also
created 41.622 million jobs thereby contributing 8% of the total employment. In real-
time scenario, people are investing more on traveling places, seeking adventure and
aspiring to bridge or at least understand the mysticism and achievements of the past and
present-day utility-based life, thereby making India as the most viable option. India not
only provides charm of heavenly, picturesque and pristine hill stations, adrenaline
pumping travel destination but is also considered as the bedrock of culture and tradi-
tions and legacy of physical artifacts, inherited from the past generations.
The concept of tourism is not a new phenomenon in India. Since 1948, with the
setting up of a Traffic Tourism Committee (TTC), different agencies like India Tourism
Development Corporation (1966) and campaigns, namely “Incredible India” (2002)
and followed by “Incredible India-2.0” (2017), were launched. These agencies and
campaigns have played a significant role in boosting tourism prospects of the country.
Different types of variations have been bought into tourism as a concept, and it has
resulted into diversification of the idea into the following forms, namely—Leisure
Tourism, Business Tourism, Ecological Tourism, Pilgrimage Tourism, Historical
Tourism, Medical Tourism, Ayurveda and Yoga Tourism, Adventure Tourism, Sports
Tourism and Wildlife Tourism, hereby making India as the one-stop destination for
vast number of tourists.
Figure 1 demonstrates how Tourism impacts other dominant sectors of the Indian
Economy. It not only can boost investment in Transportation Sector and Hospitality
Sector but also can significantly impact progress of small and medium enterprises
(SMEs). Following are some factors contributing to the prosperity of the Tourism
Sector in India:

https://doi.org/10.1007/978-981-15-2774-6_58
496 V. Prusty et al.
Fig. 1. Impact of Indian Tourism sector
(a) Price competitiveness of Tourism.

(b) Good freight/air transport facilities.
(c) Ground transport infrastructure.
(d) Availability of vast number of natural and cultural resources.
With the above-mentioned positives, there are certain oddities in terms of low
penetration of technological advancements which if achieved, can act in favor of
India’s growth story. The most versatile and one of the futuristic technical advancement
which can harness tourism sector to its fullest is Internet of things, abbreviated as IoT
[1]. The concept of IoT has shrunk the world into a smartphone, which facilitates the
interaction between the physical entities via Internet. The main application of IoT
consists of controlling and monitoring functions. This innovative and simple to
understand and easy-to-use technology has penetrated into both social, economic,
academic and energy sectors, where it facilitates efficient and optimal usage of
resources. In the present-day scenario, the multi-faceted nature of technology raises
various possible transient and random failures which are vulnerabilities that can be
exploited for illicit means by cyber attackers. The various possible attacks on IoT
applications include—denial-of-service (DoS), controlling, eavesdropping, physical
damage and node capture. These threats can be handled by different features of IoT,
namely: reliability, responsibility, privacy which include data sharing and management,
data collection and data security, safety and identification and authentication, being the
crucial features [2].
Research in the field of IoT has paved way to application oriented domains like
home automation [2], smart irrigation system [3], smart tourism and many other fields
which form a part of our day-to-day life. Many previous pioneering works focused on
development results of IoT studies and discuss ideas on how to apply them to business
[4], IoT platform for supporting a real-time recommender system [5] and signature-
based authenticated key establishment scheme for IoT environment [6], which stressed
upon the security aspect of the IoT. The area of research was extended to the big data
analytics where many works on behavioral pattern analysis [7] using the existing
database which made further research possible in the study of many e-commerce
models [8]. Big data analytics study enabled synthesis of millions of data points into a
meaningful pattern, thereby resulting in automation of a physical entity for which the
Development of an IoT-Based Tourism Guide System 497
analytics was performed. The IoT and big data together has resulted in some of
research opportunities like design of the intelligent monitoring system for scenic spot
[9], providing an IoT information technology public platform for regional revival [10],
concept of smart and connected communities (SCC) [11], innovation approach of
mobile dynamic trip planner [12], application of Geo2Tag LBS Platform [13],
implementation of GPS and Electronic Map as a multi-mode [14] and numerous other
techniques and technologies. The ideation for the above-mentioned researches is the
result of various case studies in the field of tourism conducted world-wide, namely on
Liaoning Tourism Resources Management Information System Based on GIS [15],
Guided Tour [16], IoT application of South African Tourism [17] and i-Tour model
[18]. The sustainability of the project work depends on the flexibility of the idea and
focusing on this various research opportunities in the field of rural folk tourism [19]
and eco-tourism has gained momentum [20–23].
In this paper, we have tried to bring in an entirely new concept of transforming
Indian Tourism sector digitally by automating the tourism sector with the right use of
technology such as embedded systems and Internet of things. And by this we think we
can really capitalize our place on world’s tourism fraternity and also help in improving
tourist’s satisfaction level, meet to the expectations of tourists and hence safeguard our
country’s image and precious amazing history.
1.1 Problem Statement
(A) Problems in Present Scenario:

The present scenario presents us with many opportunities when the below-
mentioned problems are catered with essential and efficient solutions:
1. Indian history, culture and heritage at stake.
2. Lack of effective tourist management and support.
3. Tourist’s level of expectation versus satisfaction rate is very low.
4. Lack of knowledge of guides and lack of authentic resources to know in-depth
about the rich history and heritage of India.
5. Guides lacking English speaking ability along with cheating and theft being
the most critical issues to be looked at to safeguard our country’s image
globally.
6. Every tourist, visiting various tourists’ places, is not capable enough to afford a
guide which keeps them away from getting to know the actual authentic facts
behind the incredible Indian History.
7. Old people, physically impaired, blind people, etc. also have equal rights to
know about their motherland, but sadly in the present scenario they are deprived
of that.
8. The real fact of such magnificent history is being shadowed and will remain
shadowed if proper steps are not taken to mitigate the issue.
So, it is high time to take action to save our country’s pride.
B. Where and Who is Undergoing the Problem:

• The problem exists in the tourist places all over India.
• Problem is faced by all potential tourists who are a part of India’s Tourism.
• People who are interested in knowing about the rich history, culture and her-
itage of India.
2 Bharat Darshan—A Concept of Digitalizing Indian

Tourism
Bharat Darshan is entirely a new concept of transforming Indian Tourism digitally with
the help of embedded systems and IoT (Fig. 2). This concept mainly
Fig. 2. Flow diagram describing the basic schema of the Bharat Darshan prototype
aims at ensuring that the true and authentic facts relating to

the incredible history, magnificent heritage, marvelous culture and tradition of
India. Below described are the motivating factors behind the creation of prototype
along with the description of the module:
(A) Technology/Process Solution:

1. Complete tourist management and support using RFID and IoT implementation.
2. Authentic resource management using audio recordings and videos about the
rich culture, history, heritage and tradition of India.
3. Authentic guide resources directly to smart phones.
(B) What is Unique/Innovative:

1. Authentic resources and facts to guide tourists.
2. RFID and IoT in Tourism Sector.
3. Provision for old, pregnant women, physically disabled people, etc. to equally
get the opportunity of enjoying the beauty of the place at their own
convenience.
4. Language preference option for people to enjoy the pre-recorded resources.
(C) How is it Different from Existing Solutions:

1. Makes authentic knowledge enriching experience and fun-filled space for the
tourists.
2. Now everybody can afford a guide and get to know the real facts from the pre-
loaded authentic resources.
3. People’s problem (disorders, age, etc.) is not a matter with our solution.
Everyone gets an equal opportunity to enjoy India’s beauty at their own
convenience.
(D) Detailing Of the Concept—Bharat Darshan:

The descriptive content of the module “Bharat Darshan” has been divided into
three parts:
1. Section 1: Represents the first phase which categorizes the available data according
to timescale.
2. Section 2: Represents the event occurred in the respective timescale according to
the Section 1. The event is categorized into two sub-categories, namely
(a) Event value/consequence.
(b) Size of the scenic spot.
3. Section 3: Represents the sub-sections of the Section 2 into the below-mentioned
categories:
(a) Event value/consequence:

(i) Popularity and influence
(ii) Scientific significance
(iii) Historical significance.
(b) Size of the scenic spot:

(i) Scenic region combination
(ii) Ornamental significance
(iii) Tourism environmental capacity.
Bharat Darshan unit consists of two independent sub-systems:
1. Embedded Division
2. Internet Of things (IoT) Division.
Tourists get the option of choosing either of the two systems or even both as per
their convenience.
(I) Embedded Division:
1. During ticket collection by the visitors, they will be asked about their preferred
language.
2. Govt. will be recording pre-recorded voices as well as interesting videos with
enriching and authentic facts about the particular historical place where the system
is to be used.
3. The voices and videos will be recorded in different languages such as English,
Hindi and the local language of that area.
4. During ticket collection, visitors will be provided with the preferred language
RFID tags (passive).
5. When the visitor scans the card of his/her preferred language, the voice gets
played in the same preferred language as scanned by the user.
6. When one visitor is listening or once the card is scanned and the voice is being
played and at that point of time someone else tries to interfere within that period
then automatically the access is not granted to the other user as currently one user
is using the device. (Illegal scanning is disabled in the system, i.e., if a person is
listening and another tries to interfere by scanning the card in between then that
scanning is not considered as authentic and access is denied to the user at that
point of time.)
7. Secondary check is done whether the person after scanning the card is present or
not using an ir sensor.
8. If in case a person scans the card and goes away without listening the setup or
system automatically detects the same and shuts the voice just after few seconds,
hence allowing other visitors to access the device.
9. The voice is played every time according to the scanned card (preferred language
card).
10. Another check is also provided in the setup that if a person while listening lives in
between and no one else is also present over there to listen then at that point of
time the system detects it and shuts the voice automatically within few seconds,
hence making the system ready to be accessed by another user.
(II) IoT Division:
(Providing visitors with another option to get to know the facts directly on their own
respective smart phones).
1. A local server is to be installed at the tourist spot.

2. A webpage is developed in that local server.
3. That Web site can only be accessed in that particular tourist’s spot by getting
connected with the local network that the server is connected with.
4. By accessing that website, one gets access to interesting audio—visual contents that
are pre-recorded videos which are developed by Govt. for that particular tourist’s
spot.
5. Now people get to know the authentic resources and facts about the rich culture and
heritage of India through interesting videos and authentic facts which enrich their
knowledge with true and real facts and hence with all these above steps helps in
increasing India’s image in world tourist arena.
Value Proposition
1. Safeguarding the rich culture, heritage and history of India by transferring the
traditional values and cultural knowledge via IoT and RFID implementation
2. Digital transformation of our Tourism Sector which gives a completely new and
beautiful perspective to Indian Tourism.
3. Growth of Tourism Sector directly contributing to India’s gross economy.
4. Tourist management and support (TMS) via IoT interface which is user-friendly.
5. Generate employment opportunity both in rural as well as heritage tourism which
can result in better new India ideas of tourism village, bettering opportunities of
country-side tourism.
3 Conclusion
Tourism Sector of India being one of the biggest game changers in India’s total
economic growth which directly leads to country’s development must not be neglected
at any cost (Fig. 3).
Digital Transformation of Tourism sector will not only encourage a greater number
of tourists to visit India to get enlightened with the real amazing Indian history, culture
and tradition but also will have a direct impact on country’s growth.
What more proud moment can it be for India when outsider’s as well as our country
people will get to know about the original authentic facts related to India’s glorious
history, heritage and start appreciating them in the world tourist arena?
It is high time to think and act upon this issue and we believe the concept of Bharat
Darshan truly has great potential to make Indian Tourism sector the best in the world
and not only that but also will surely help India generate a good amount of revenue
which further defines the growth of India in the upcoming future.
Fig. 3. Working prototype of Bharat Darshan
References
1. Rout KK, Mishra S, Routray A (2017) Development of an IoT based Introductory
Laboratory for Under graduate engineering students. In: 16th International Conference on
Information Technology (ICIT-2017), Bhubaneshwar, India
2. Rout K, Mallick S, Mishra SK (2018) Design and implementation of an Internet of Things
based prototype for smart home automation network. ICRIEECE, Bhubaneswar, India
3. Rout K, Mallick S, Mishra SK (2018) Solar powered smart irrigation system using Internet
of Things. In: 2nd International Conference on Data Science and Business Analytics
(ICDSBA 2018), ChangSha, Hunan, China
4. Balandina E, Balandin S, Koucheryavy Y, Mouromtsev D (2015) IoT use cases in healthcare
and tourism. In: 2015 IEEE 17th conference on business informatics
5. Cha S, Ruiz MP, Wachowicz M, Tran LH, Cao H, Maduako I (2016) The role of an IoT
platform in the design of real-time recommender systems. In: 2016 IEEE 3rd World Forum
on Internet of Things (WF-IoT)
6. Challa S, Wazid M, Das AS, Kumar N, Reddy AG, Yoon E-J, Yoo K-Y (2016) Secure
signature-based authenticated key establishment scheme for future IoT applications. IEEE
Access 2016
7. Shafiee S, Ghatari AR (2010) Big data in tourism industry. In: 10th international conference
on e-commerce with focus on e-tourism
8. Fang M, Zheng Z (2011) Innovation pattern of China tourism e-commerce on the post-crisis
era. In: 2011 2nd international conference on Artificial Intelligence, Management Science
and Electronic Commerce (AIMSEC)
9. Lin Y (2011) The application of the internet of things in Hainan tourism scenic spot. In:
2011 seventh international conference on computational intelligence and security
10. Gautam BP, Asami H, Batajoo A, Fujisaki T (2016) Regional revival through IoT enabled
Smart Tourism Process Framework (STPF): a proposal. In: 2016 joint 8th international
conference on soft computing and intelligent systems and 2016 17th international
symposium on advanced intelligent systems
11. Sun Y, Song H, Jara AJ, Bie R (2015) Internet of things and big data analytics for smart and
connected communities. IEEE Access, vol 14, no 8, August 2015
12. Alghamdi H, Zhu S, El Saddik A (2016) E-tourism: mobile dynamic trip planner. In: IEEE
2016, international symposium on multimedia
13. Balandina E, Balandin S, Koucheryavy Y, Mouromtsev D (2015) Innovative e-tourism
services on top of Geo2Tag LBS platform. In: 2015 11th international conference on signal-
image technology & internet-based systems
14. Hu Z, Wang Z, Liu H, Bie H (2010) Self-service folk tourism guiding technology on mobile
terminal with multi-mode: application of GPS and Electronic Map. In: IGARSS 2010, IEEE
2010
15. Chen L, Wang H, Yang Z (2010) Study on Liaoning tourism resources management
information system based on GIS. In: 2010 international conference on networking and
digital society
16. Wu S-T, Lee B-W (2017) An innovative way of guided tour: a virtual experience of dark
tourism. In: 2017 IEEE international conference on information, communication and
engineering
17. Gcaba O, Dlodlo N (2016) The internet of things for South African tourism. In: IST-Africa
2016 conference proceedings
18. Tripathy AK, Tripathy PK, Ray NK, Mohanty SP (2018) iTour: the future of smart tourism.
In: IEEE consumer electronics magazine, May 2018
19. Ma P, Zhao W, Hu Z, Duan F (2010) Three-dimensional display component for rural folk-
custom tourism. In: 2010 18th international conference on geoinformatics
20. Shafiee MM, Najafabadi SI (2010) The interaction of technological progress and tourism
industry development in the developing countries: the case of Iran’s Tourism Industry. In:
10th international conference on e-commerce with focus on e-tourism
21. Wang J, Dong S, Wan Y (2014) Abundance of tourism resources and regional tourism
development. In: 2014 22nd international conference on geoinformatics
22. Xingqi W (2009) Study on core capabilities of tourism networks. In: 2009 International
conference on environmental science and information application technology
23. Yi W, Zhao J (2011) Research on systematic coupling symbiosis of low-carbon tourism and
eco-tourism. IEEE
Design and Simulation of a Standalone
Photovoltaic System Using Synchronous
Boost Converter and Reduced Switch
Five-Level Inverter
Kshirod Kumar Rout1(&), Sivkumar Mishra2,

and Pravat Kumar Biswal3
1
Department of EEE, IIIT, Bhubaneswar, Odisha, India
2
Department of EEE, CAPGS, BPUT, Rourkela, Odisha, India
3
Department of EEE, KIIT, Bhubaneswar, Odisha, India
1 Introduction
The advancement of new energy sources is consistently improved in the recent dec-
ades. But, the need of sustainable, renewable, and clean electric energy sources has
become indispensible in the present energy scenario of the world, and it seems the solar
energy has emerged as the viable alternative [1]. In the recent decades, humungous
amount of research has been done by the scientists on the PV cells and the ways to
increase their efficiency and converting it into usable energy form [2]. The market for
PV-based systems is expanding around the globe, and presently, solar PV gives around
4800 GW of electric energy. During the years from 2004 to 2009, network-associated
PV limits achieved 21 GW and were expanding at a yearly normal rate of 60% [3].
Usually, each polycrystalline PV cell generates voltage ranging from 0.5 to 0.8 volts
depending upon the sort of semiconductor material utilized and the developed inno-
vation of highly efficient and commercial PV cells. This voltage is sufficiently low as it
cannot be useful. In order to get advantage from this innovation, many PV cells ranging
from 36 to 72 numbers of cells are affixed into a matrix. Thus, obtained modules are
arranged in series and parallel lines to obtain required voltage and current. In the event
that these modules are associated in arrangement, their voltages are included with a
similar current. In any case, when they are associated in parallel, their currents are
added while the voltage is the same. Due to various reasons, solar panel’s output
voltage may not be constant. Thus, the requirement of a boost converter system arises.
A synchronous boost converter’s (SBC) [4, 5] output voltage is always higher than the
given input voltage depending upon switching frequency and the reference voltage
level set. The boost converters operated at higher frequencies produces a hike in
switching losses during turn ON/OFF. For the reduction of the switching losses
occurred in the adopted synchronous converter, soft switching techniques also known
as zero-voltage switching (ZVS) and zero-current switching (ZCS) [6] are proposed
involving an added auxiliary circuit which consists of a capacitor, a 1N4007 diode,

https://doi.org/10.1007/978-981-15-2774-6_59
Design and Simulation of a Standalone Photovoltaic System … 505
a helper switch, and an inductor. All the switches in the proposed converter system
operate under ZVS and ZCS conditions to reduce the losses occurred during switching
and to improve the efficiency compared to a conventional boost converter system. As
the efficiency of a conventional chopper is not satisfactory, it can be improved by
replacing a diode with MOSFET operated in synchronization mode. This reduces the
power losses caused by diode as well as MOSFET has much lesser internal resistance
as compared to a diode. MPPT [7, 8] systems provide a good control over maintaining
the required voltage and current levels obtained at the output of the SBC. This controls
the gate pulse to be delivered to the SBC arrangement. In order to serve the domestic
loads, a five-level inverter [9, 10] requires two different DC supplies ideally each of
same voltage output, which in turn is converted into AC voltage. Multi-level inverters
(MLIs) with reduced number of switches are preferred to conventional cascaded MLI
as the former has lesser harmonic distortion [11].
In light of above developments, in this paper a standalone PV system is designed
with a SBC and a reduced switch five-level inverter for driving AC loads. The SBC is
simulated in MATLAB/Simulink, and the simulation observations are shown to sup-
port theoretical analysis. Section 2 elaborates the proposed model and its sub-systems,
Sect. 3 explains the simulations performed for the design purpose, and Sect. 4 shows
the analysis and results of the work. The use of SBC with MPPT and PI controller and
RSFL inverter makes the overall design quite efficient.
2 Proposed Model
Photovoltaic panels are usually of 36 cells or 72 cells. This work utilizes two 72 cell
solar panels with 0.5 V per each cell affixed in a matrix form in the panel, whose output
ranges within 30–50 V, 5 A. As the PV curve (Fig. 4) gives the knowledge about the
variations in the voltage and current outputs of the PV panel at different temperatures.
Output of the solar panels is connected to respective synchronous boost converters to
step up the DC voltage to 115 V. MPPT system controls the pulses to be sent to the
SBC by measuring the voltage and current output of the installed polycrystalline solar
panels. The outputs of the two SBCs are given to RSFL converter as input so that it can
generate 230 V AC. The outline of the proposed design is shown in Fig. 1.
A. Solar panel
Each polycrystalline silicon solar cell has an internal circuit with a current source, P-N
junction, shunt, and series resistances as shown in Fig. 2. The output voltage is nearly
0.8 volts. Matrix formation of several such cells is known as a module. Collection of
such modules is known as panel (Figs. 3 and 4).
506 K. K. Rout et al.
Fig. 1. Overview of the proposed design
Fig. 2. Internal circuit of a PV cell
Fig. 3. IV characteristics of PV panel

Fig. 4. PV characteristics of solar panel
B. Measurement of electrical parameters
Measuring the DC voltage and current can be achieved by V-A method or A-V method.
V-A method is much accurate than the other. Voltage divider circuit is used for
measuring the voltage, and Hall effect sensor is used for current. Analog outputs of the
sensors are connected to the MPPT system.
C. MPPT
Maximum power point tracking is used to extract maximum possible power from the
solar panel. Perturbation and observation (P&O) method is adopted as it is less com-
plex to implement [12]. The algorithm for the chosen MPPT scheme is shown in Fig. 5.
Fig. 5. Algorithm for P&O method

D. Synchronous boost converter
Proportional and integral control system has been adopted for the proposed SBC as a
negative feedback input to control the output potential difference according to the base
voltage set. The assistant circuit is made out of primary switch (S1), a helper switch
(S2), a full capacitor (Cr), a resounding inductor (Lr), and two diodes (D1 and D2), as
appeared in Fig. 2. PWM [11] duty cycle is varied to operate the switches at the desired
frequency (Fig. 6).
Fig. 6. Circuit outline of SBC
E. Reduced switch inverter
A five-level reduced switch inverter (Fig. 7) utilizes two DC sources. In the place of the
DC sources, we replace it with the output of the SBCs. This configuration uses only
five switches which lead to lesser switching loss. Required PWM is simulated.
Fig. 7. Five-level reduced switch (FLRS) inverter

3 Simulation
A. Overall view
The SIMULINK model of the proposed design is shown in Fig. 8. The parameters like
irradiance and temperature to the panel’s model are given as inputs. As the solar panel
acts as a current source, it is simulated using a controlled current source in Simulink
toolbox. Required circuitry is made for diagnosis of the designed system. MPPT with
the algorithm specified in Fig. 5 is integrated in the above simulation shown.
Fig. 8. Overall view of the design
B. SBC simulation
The Simulink design of the SBC is as shown in Fig. 9. It uses high-frequency

switching devices for the converter action. PWM has been utilized for their operation.
The input voltage has been set to 50 V which can vary up to 100 V DC. The output is
taken as negative feedback loop with a PI controller [13, 14] in it. The reference voltage
Fig. 9. Modeling of synchronous boost converter sub-system

level is set as 115 V as it is necessary for the reduced switch inverter to generate AC
output voltage. Scope is introduced into the system to monitor the PWM at the
switching devices, the error signal, and the output voltage waveform.
C. RSFL converter simulation
The reduced five-level inverter has been simulated as shown in Fig. 10.
Required PWM is supplied for the MOSFET to perform the switching operation at
desired frequency. This kind of inverter is easy to implement and analyze. It requires
two different DC voltage sources to work as a five-level inverter. Delivering output
voltage is 230 V AC.
Fig. 10. Reduced switch five-level inverter
D. Simulation of load sub-system
The type of load utilized in this design process is resistive type and is connected to a
voltmeter for measuring the voltage across the load.
4 Results
From the simulations performed, the output waveforms of the SBC, MPPT, and RSFL
inverters are as follows. The pulse train of the MPPT at 10 kHz is also shown in the
results.
A. Synchronous boost converter
SBC takes 50 V DC as input and delivers 115 V DC as output as shown in Fig. 11.
Output voltage settles at 115 V DC within 0.06 s due to the PI controller used from the
observations. Figure 12 shows the corresponding MPPT pulses.
Fig. 11. Controlled output voltage of SBC
Fig. 12. MPPT pulses
B. Reduced five-level inverter

The output waveform obtained from the RSFL inverter is shown in Fig. 13, where
dotted line indicates the voltage and the solid line indicates current when connected to a
resistive load. The voltage obtained is purely a five-level inverted wave. As the load is
resistive, the voltage and current are in phase with each other. The total harmonic
distortion (THD) of the voltage is found to be 26.49% and is shown in Fig. 14.
Fig. 13. Output of the RSFL inverter
Fig. 14. THD of output voltage (RSFL)
5 Conclusion
In this paper, the conversion of unstable DC voltage obtained from the solar panels has
been first converted to two constant DC115V sources with help of SBCs with a PI
feedback control and a P&O-based MPPT scheme. A 230 V AC output has been
generated from a RSFL inverter with the two 115 V constant DC as its inputs. The
entire PV system has been simulated to drive a resistive load. The theoretical analysis
has been supported by the MATLAB/Simulink models of all the components and
waveforms of the corresponding sub-systems. The hardware implementation of the
proposed PV system is included in future work.
References
1. Rosyid OA (2017) Comparative performance testing of solar panels for smart city micro-
grids. In: 2017 international conference on smart cities, automation and intelligent
computing systems (ICON-SONICS), Yogyakarta, pp 81–86
2. Goel S, Sharma R (2017) Performance evaluation of stand alone, grid connected and hybrid
renewable energy systems for rural application: a comparative review. Renew Sustain
Energy Rev 78:1378–1389
3. Gul M, Kotak Y, Muneer T (2016) Review on recent trend of solar photovoltaic technology.
Energy Explor Exploit 34(4):485–526
4. Qi L, Xiao M (2017) Robust control of boost converter. In: 2nd international conference on
cybernetics, robotics and control (CRC), Chengdu, China, pp 61–65
5. Kim J-H, Jung D-Y, Park S-H, Won C-Y, Jung Y-C, Lee S-W (2009) High efficiency soft-
switching boost converter using a single switch. J Power Electron 9(6):929–939
6. Jabbari M, Farzanehfard H (2009) A new soft switching step-down/up converter with
inherent PFC performance. J Power Electron 9(6):835
7. Xiao S, Balog RS (2018) An improved adaptive perturb & observe maximum power point
tracking technique. In: IEEE Texas Power and Energy Conference (TPEC), College Station,
TX, pp 1–6
8. Karabetsky D (2017) Conceptual design of maximum power point tracking for solar
rechargable airplane. In: IEEE 4th International Conference Actual Problems of Unmanned
Aerial Vehicles Developments (APUAVD), Kiev, pp 58–60
9. Garg V, Sinha SK, Dave MP (2016) Phase shifted PWM cascaded multilevel inverter for
solar PV standalone systems. In: IEEE 1st International Conference on Power Electronics,
Intelligent Control and Energy Systems (ICPEICES), Delhi, pp 1–5
10. Kabalci E, Kabalci Y, Canbaz R, Gokkus G (2015) Single phase multilevel string inverter
for solar applications. In: International Conference on Renewable Energy Research and
Applications (ICRERA), Palermo, pp 109–114
11. Rout KK, Mishra SA (2018) A modified 7-level reduced switch symmetrical inverter. In:
National Power Engineering Conference(NPEC),Madurai, India
12. Mohapatra A, Nayak B, Das P, Mohanty KB (2017) A review on MPPT techniques of PV
system under partial shading condition. Renew Sustain Energy Rev 80:854–867
13. Qian K, Xiong H, Lei D (2017) Design of PWM rectifier based on fractional order PID
control. In: International Conference on Industrial Informatics - Computing Technology,
Intelligent Technology, Industrial Information Integration (ICIICII), Wuhan, China, pp 215–
218
14. Karuppiah M, Karthikumar K, Arunbalj A (2017) A transformerless buck-boost converter
with PID controller. In: IEEE International Conference on Intelligent Techniques in Control,
Optimization and Signal Processing (INCOS), Srivilliputhur, India, pp 1–7
Designing and Sizing of a Stand-alone
Photovoltaic System: A Case Study
Kshirod Kumar Rout1(&), Sivkumar Mishra2,

and Pravat Kumar Biswal3
1
Department of EE, IIIT, Bhubaneswar, Odisha, India
kshirod@iiit-bh.ac.in
2
Department of EE, CAPGS, BPUT, Rourkela, Odisha, India
3
Department of EE, KIIT, Bhubaneswar, Odisha, India
1 Introduction
With the current transition that this world is going through, electricity has already
become another basic necessity apart from food, clothing and shelter. Since the
beginning, the major source of electricity is fossil fuel, and on its combustion, it
releases huge amount of harmful gases to our surrounding. To prevent it, many nations
have considered this a serious global agenda and are committed to eradicate gases that
cause the greenhouse effect from our atmosphere by producing cleaner energy through
renewable sources. Also, the production of electricity as a source of energy is quite
costly in terms of exploitation of earth’s resources such as coal and oil which are
limited and are going to be scarce over the next 50 years. Moreover, UN has also taken
initiatives to encourage nations for generating clean and green energy. As a result,
different countries have initiated different strategic measures to reduce their greenhouse
gases and eventually their carbon footprints. Hence, different countries are exploring
other sources of energy which are renewable in nature such as hydro (water), wind and
solar. This exploration toward alternate clean energy resulted in a steady increase in
utilization of nonconventional sources of energy to resolve acute energy shortage and
reduce the effects of global warming [1, 2]. In the present day scenario, renewable
sources of energy are world’s fastest growing energy tank, and its consumption is
expected to increase from 9% to 16% till the year 2040 [3]. Extensive research has
already been done in the field of exploring more unique ways of increasing the
accuracy and efficiency of the PV cells and optimal sizing conditions.
Now, focusing on the energy produced from sunlight, it has been found that the use
of solar energy is quite attractive for an alternative source of energy because it is silent,
safe, nonpolluting and can be adjusted for any increase in load, and most importantly, it
is fail proof with an average working life of 20–30 years [1]. Moreover, it contains no
special software or any sort of moving parts and is virtually maintenance free.
A PV system is a system that uses solar energy as its input and converts it into
electricity. These systems are usually classified based upon the connection of the
system components with the power sources, i.e., stand-alone (SA) and utility-
interactive (UI) systems. Taking the “stand-alone” factor into consideration, as the
name suggests the SAPV systems are usually designed to operate independently and

https://doi.org/10.1007/978-981-15-2774-6_60
Designing and Sizing of a Stand-alone Photo … 515
sized to supply a specific amount or average amount of DC and/or AC electrical loads

[4]. Now, taking the solar radiation into consideration, the sunlight irradiation intensity
that reaches our planet earth varies with respect to the time, seasons, geographical
location and different weather conditions. The term “irradiation” is defined as the net
energy that we obtain on a daily basis, and it represents the strength of the sunlight
intensity. The unit for irradiation is Wh m−2 per day [1, 2, 5].
The other physical factors that affect the working as well as the efficiency of the
solar PV system are the different geographic regions that experience different weather
patterns, number of sun days in a year and best tilt angle for the solar panels. The PV
cells would be exposed to more sunlight if a tracker is installed that changes the angle
according to the movement of the sun; however, that method is way too expensive to
meet the daily needs [6]. The phenomenon or the process of evaluating or calculating
the adequate voltage, current and power ratings for each component of the PV system
to meet the electric load demand and at the same time calculating the total expenses for
the entire system from design to implementation including transportation cost and other
costs is called as system sizing.
In this paper, we are primarily considering the solar energy as the conventional
source of energy and focusing on designing a stand-alone photovoltaic (SAPV) system
by assuming some standard loads present in a house. The sizing scenario presented in
this paper aims at fulfilling the daily electrical needs for a residential housing in the city
of Bhubaneswar, Odisha, India.
2 Methodology of Sizing of SAPVs

2.1 Components and System Requirements
a. PV Module: It is a semiconductor containing p–n junctions that convert sunlight to

electricity which is DC in nature. Commonly, a PV module includes single poly-
crystalline silicon and amorphous silicon [1].
b. Battery: The battery stores energy for meeting the peak load demands and is mostly
useful during dark days or no-sun days. The commonly used types are lead–acid
type batteries, that are designed to gradually discharge and recharge 80% of their
capacity hundreds of times [1].
c. Solar Charge Controller: The solar charge controller is mainly used to maintain
the battery levels at optimal state alongside protecting it from overcharge of the PV
panels and excess discharge from the connected loads and therefore elongates the
battery life [6].
d. Inverter: It primarily converts the DC output from the PV cells to AC form and
feeds it to the AC load. Inverters are further classified based upon the waveform
output and on the basis of installation. The inverter is known as power conditioner
because of its ability to change the form of the electric power. The efficiency of
inverters reaches its maximum limit when the load demand is greater than half the
rated load [7].
e. Load: These are the end user connected electronic appliances.
The complete schematics of the SAPV are shown in the following Fig. 1.
Fig. 1. Schematics of a stand-alone photovoltaic system (SAPV)
2.2 System Sizing

The phenomenon or the method of anticipating the required voltage and current ratings
for each component of the PV system to fulfill the electric demand and at the same time
calculating the total expenses for the entire system from design to implementation
including transportation cost and other costs is called as system sizing. The following
sizing is implemented based on intuitive methods [8–12].
a. Home Appliance Sizing:
The voltage and current requirements for the connected loads are first calculated along
with their time of operation, and an average value of required power is estimated on a
daily basis [1].
b. Sizing of Solar Modules:
It deals with the number of solar arrays required for the desired power output, and it
also includes the manner in which the arrays needed to be connected, i.e., in series or
parallel manner to gain higher efficiency. It completely depends on the daily energy in
watt-hour (E), the DC voltage (VDC) and the average sun-hour per day (Tmin) of the
system, and these [9] parameters must be calculated at first. Moreover, to avoid the
undersizing of the system, losses are taken into consideration.
Therefore, to avoid undersizing, the daily average energy consumption is divided

by the system efficiencies to deduce the required output from the solar panel array.
Mathematically:
Daily average Energy Consumption E

Er ¼ ¼ ð1Þ
Product of system component's efficiency goverall
The peak power is calculated by dividing the value of Er by average sun-hours per
day (Tmin) for the given area. Mathematically:
Energy required on daily basis Er

Pp ¼ ð2Þ
Minimum sun time per day Tmin
For calculating, the net current required:
Peak Power Pp
Idc ¼ ¼ ð3Þ
System VoltageðDCÞ VDdc
The number of modules connected in parallel can be calculated by:
Total current of the PV Module Idc

Np ¼ ¼ ð4Þ
Rated value of current of one PV Module Ir
Modules to be connected in series can be calculated by:
System VoltageðDCÞ Vdc

Ns ¼ ð5Þ
Rated value of Voltage of one PV Module Vr
Lastly, the total number of modules that needed to be connected are:
Nm ¼ Ns Np ð6Þ
c. Sizing of Battery:
The approximated amount of energy needed is calculated by multiplying the number of
no-sun days or autonomy days and the total power demand. Mathematically:
Erough = Total Power DemandðEÞ Days of AutonomyðDÞ ð7Þ
Considering safety factor, the result is divided by the maximum allowable level of
discharge (MDOD):
Required energy for storage Erough

Esafe ¼ ¼ ð8Þ
Maximum depth of Discharge MDOD
Now, at this step, the battery bank capacity (in A-h) can be evaluated by dividing
the safe energy storage required by the DC voltage of one of the batteries:
Esafe
C¼ ð9Þ
Vb
Therefore, the number of batteries needed is obtained by:
C
Nbatteries ¼ ; where Cb refers to the capacity: ð10Þ
Cb
Batteries to be cascaded in series are given by:
VDC
Ns ¼ ð11Þ
Vb
and the number of parallel paths are given by:
NBatteries
Np ¼ ð12Þ
Ns
d. Sizing of the Voltage Controller:

Even though the name is voltage controller, the parameter that it controls is current. For
a voltage controller, its main aspect is to withstand the maximum load current along
with the maximum array current. The product of the short-circuit current and the safety
factor results in deducing the ratings for the voltage controller. Mathematically:
I ¼ Isc Np Fsafe ð13Þ
We have considered the safety factor mainly because to ensure that the controller
withstands and handles the maximum array currents produced by the arrays, which
could be higher than the calculated value [10]. The number of controllers required is
calculated by:
I
Ncontroller ¼ ð14Þ
Amps of Each Controller
e. Inverter Sizing:
The inverter sizing is determined from the first step as the actual power drawn will
determine the rating of the inverter
f. Sizing of the wiring system:
The required type of wire to be used can be actuated based upon the above result.
3 Example
The location of the application is considered at Bhubaneswar, Odisha, India. The

geographical location of Bhubaneswar is at 85.82° longitude and 20.29° latitude [13]. It
receives an annual irradiance of about 5.32 kWh/m2/day [14].
The home appliance sizing is considered as follows (Table 1).
Table 1. Daily consumption table (based upon approximated values)

Individual Quantity Volt. Amp. AC Usage Usage in 7 days Watt-
load wattage in days/week hours
h/day AC
Domestic fan 1 220 0.54545 120 8 7 7 960
Grinder 1 220 2.72 600 0.3 7 7 180
Electric iron 1 220 4.54 1000 0.8 4 7 457.143
Desktop 1 220 0.56 120 2 7 7 240
Bulb 4 * 15 220 0.272727 60 5 7 7 300
Radio 1 220 0.363636 80 4 7 7 320
Fridge 1 220 0.909091 200 12 7 7 2400
Television 1 220 0.568182 125 6 7 7 750
Electric 1 220 1.136364 250 0.5 5 7 89.28571
washing
machine
Total AC connected Watts 2555 Average AC daily load 5696.43
Total average energy consumption 5696.426 approximated to 5700
The selected panel for the above mentioned requirements is luminous solar panel,
250 W, 24 V, 8.31 A. It is of polycrystalline silicon type having 60 number of cells
with Pmax as 250 W, VOC as 37.42 V, Isc as 8.85 A, Vmp as 30.2 V and Imp as 8.31 A,
cell efficiency as 0.176.
From the above formulas,
a. Sizing of Arrays:
E
Er ¼ ¼ 5700=0:8 ¼ 7125 Wh/day or 7:125 kWh/day
goverall
Pp ¼ TEminr ¼ 7:125=6 ¼ 1:18 kWhp [15]
Pp
IDC ¼ ¼ 1187:5=24 ¼ 49:47 A
VDC
For modules,
IDC
Np ¼ ¼ 49:47=8:31 ¼ 5:95 Panel ¼ 6 Panels
Ir
VDC
Ns ¼ ¼ 24=24 ¼ 1
Vr
Nm ¼ 6 1 ¼ 6 Panels
Hence, the PV array consists of six panels connected in parallel.

b. Sizing of Battery:
Net Average Energy ¼ 5700 Wh

No Sun - days ¼ 3ðsayÞ
Erough ¼ 5700 3 ¼ 17:1 kWh
Erough
Esafe ¼ ¼ 17100=0:75 ¼ 22800 Wh
MDOD
Esafe
C¼ ¼ 22800=12 ¼ 1900 A-h
Vb
C
Nbatteries ¼ ¼ 1900=250 ¼ 7:6 Batteries ¼ 8 Batteries
Cb
VDC
Batteries in series ¼ Ns ¼ ¼ 24=12 ¼ 2
Vb
Number of parallel paths ¼ Np ¼ 8=2 ¼ 4
c. Sizing of Voltage Controller:

The rated current of the voltage controller is calculated
I ¼ Isc Np Fsafe ¼ 8:85 6 1:25 ¼ 66:375 A
I
Ncontroller ¼ ¼ 66:375=60 ¼ 1:10
Amps of Each Controller
Hence, we need two regulators connected in parallel.

d. Sizing of Inverter:
The total power of the devices when connected together:
Ptotal = 2555 W. This means the inverter that is needed must be capable of handling
2555 W at 220 V AC.
4 Conclusion
The location that is considered (Bhubaneswar, Odisha, India) qualifies itself to be

famed as a sun-rich region having annual irradiance of more than 5.32 kWh/m2
annually [14]. The solar stand-alone photovoltaic system is fully designed from scratch
and is implemented based upon the calculated values. Several factors that affect the
process directly or indirectly are explained in this report. The mentioned processes for
the implementation of the solar stand-alone photovoltaic system can be used in any
region having a tropical, subtropical or dry climate.
References
1. Al-Shammani AN, Othman MYH, Mat S, Ruslan MH, Abed AM, Sopian K (2015) Design
& sizing of stand-alone solar power systems a house Iraq. In: Recent advances in renewable
energy sources, April 2015
2. Goyal S, Sharma R (2017) Performance evaluation of stand alone, grid connected and hybrid
renewable energy system for rural application: a comparative review. Renew Sustain Energy
Rev 78:1378–1389
3. Energy Data. https://www.iea.org/weo2017/. Accessed on April 2018
4. Ritchie RW (1999) Using sunlight for your own solar electricity: build your own system,
become independent of the grid, domestic photo voltaics. Ritchie Unlimited Publications
5. AL-Rousan N, Mat Isa NA, Desa MKM (2018) Advances in solar photovoltaic tracking
systems: a review. In: Renewable and sustainable energy reviews, vol 82, Part 3, pp 2548–
2569, Feb 2018
6. Messenger RA, Ventre J (2003) Photovoltaic systems engineering. CRC Press
7. Sonnenenergie DGF (2007) Planning and installing photovoltaic systems: a guide for
installers, architects and engineers. Earthscan
8. Faraa L, Craciunescua D (2017) Output analysis of stand-alone PV systems: modeling,
simulation and control, vol 112, pp 595–605, Mar 2017
9. Jadin MS, Nasiri IZM, Sabri SE, Ishak R (2015) A sizing tool for PV standalone system.
ARPN J Eng Appl Sci 10(22), Dec 2015
10. Nordin ND, Rahman HA (2015) An optimization method for designing stand alone
photovoltaic system using iterative method. In: IEEE International Conference on Smart
Energy Grid Engineering (SEGE), 12 Nov 2015
11. Bataineh K, Dalalah D (2012) Optimal configuration for design of stand-alone PV system.
Smart Grid Renew Energy 3:139–147
12. Srinivasarao M, Sudha KR, Bhanu CVK (2016) A simple and reliable method of design for
standalone photovoltaic systems. Journal IE (India), Series B, 19 Aug 2016
13. Location Data. http://www.synergyenviron.com/tools/solar-irradiance/india/odisha/bhubaneswar.
Accessed on April 2018
14. Irradiation Data. http://www.synergyenviron.com/tools/solar-irradiance/india/odisha/
bhubaneswar. Accessed on April 2018
15. Average Sun-Hours Data. https://www.currentresults.com/Weather/India/annual-sunshine.
php. Accessed on April 2018
16. Aziz NIA, Sulaiman SI, Shaari S, Musirin I, Sopian K (2017) Optimal sizing of stand-alone
photovoltaic system by minimizing the loss of power supply probability. Solar Energy
150:220–228
17. Ibrahim IA, Khatib T, Mohamed A (2016) Optimal sizing of a standalone photovoltaic
system for remote housing electrification using numerical algorithm and improved system
models. Energy 126:392–403
18. Khatib T, Mohamed A, Sopian K (2013) A review of photovoltaic systems size optimization
techniques. Renew Sustain Energy Rev 22:454–465
19. Rajesh R, Mabel MC (2015) A comprehensive review of photovoltaic systems. Renew
Sustain Energy Rev 51:231–248
20. Sumathia V, Jayapragash R, Bakshi A, Akella PK (2017) Solar tracking methods to
maximize PV system output—a review of the methods adopted in recent decade. Renew
Sustain Energy Rev 74:130–138
A Model for Optimizing Cost of Energy
and Dissatisfaction for Household Consumers
in Smart Home
Nilima R. Das1(&), Satyananda C. Rai2, and Ajit Nayak1

1
Faculty of Engineering & Technology, Siksha ‘O’ Anusandhan
nilimadas@soa.ac.in
2
Department of IT, Silicon Institute of Technology, Bhubaneswar, Odisha, India
1 Introduction
Smart grid is an uprising concept in electrical grid systems. It claims the advancement
of technology and information system with highly advanced electrical infrastructures
that can increase the reliability, security, and efficiency. There are many challenges in
the implementation of smart grids and a lot of research is going on in this area from the
last few years. The implementation of smart grid now involves various demand-side
management (DSM) techniques. The DSM programs comprise of load management
and energy conservation activities at the user side that tries to maintain a balance
between the demand and supply of electricity to facilitate the use of the existing energy
resources avoiding investment in additional sources. It results in reduced electricity
payments for the consumers and increased availability of the electricity. The time-
varying prices also help a lot to encourage the consumers to reduce their demand
during peak hours by shifting a fraction of their consumption from peak hours to off-
peak hours as the peak hours are charged with high prices. The method proposed in this
work follows a similar approach. It uses a time-varying pricing technique to optimize
the energy consumption behavior of the user by producing an operation time schedule
for the user’s appliances which reduces the peak hour demand and electricity cost of
the user. The next section provides a brief literature survey in this area. Section 3
elaborates the mathematical structure for the proposed method. Section 4 provides a
precise description about the simulation process and its results. Finally, Sect. 5 pro-
vides the conclusion.
2 Related Works
The users in a smart grid can use various energy management techniques to lower their
peak hour power consumption which can also help them to reduce their daily electricity
bill. A lot of research works are being done to implement DSM for the consumers so
that they can reduce their consumption during peak hours. The authors have used an
energy management technique that can predict upcoming prices in a real-time
https://doi.org/10.1007/978-981-15-2774-6_61
524 N. R. Das et al.
environment that helps to control and manage the peak hour consumption of the users
[1]. The authors [2] have used an integer linear programming (ILP) method to generate
the time schedule for the appliances which can minimize the peak time load and
generate optimal operation time for user appliances. They have also proposed a con-
sumption scheduling mechanism using integer linear programming (ILP) and game
theory approach to minimize the peak time load [3]. The authors [4], follow a game
theory approach for generating an optimized schedule for the consumers. In this work,
the game is played between the users and each user tries to minimize its cost. The
authors have proposed a quadratic convex programming method to determine the
optimal operation time for the appliances for cost minimization in [5]. The authors
[6, 7] have described a game theory-based optimization procedure where the game is
played between the users and the utility provider. In the game, the user’s strategy is to
minimize its electricity payment and the strategy of utility provider is to adjust the
energy price parameter to reduce users’ consumption. Ye et al. [8] provided a real-time
information-based energy management technique based on game theory optimization
method to minimize the peak-to-average ratio in demand.
All the works in the literature try to reduce the peak time load and electricity
payment for the users. However, in [1], the authors have tried to minimize the waiting
time for each appliance along with cost reduction. Every appliance has to operate in a
particular time interval which is decided by the user. After optimization, the appliances
have to follow the time schedule generated by the optimization process. If the starting
time of an appliance is postponed, then the waiting cost is increased and the opti-
mization process also makes an effort to reduce the waiting cost. The waiting cost
depends on the gap between the scheduled starting time and the beginning of the time
interval previously set by user for the appliance. The waiting cost increases with the
percentage increase in the waiting time.
There are some other scenarios possible where the user has a preferred starting time
for the execution of an appliance from the allowed time of operation set for it. The
optimization process may generate a starting time before or after the preferred starting
time of the user and if the optimization process does not generate the time schedule
according to the user preference then the user may get some dissatisfaction. In this work,
a method has been proposed that reduces the dissatisfaction caused due to the mismatch
of scheduled starting time and user’s preferred starting time of an appliance as an
outcome of cost minimization. The procedure is described in detail in the next sections.
3 Mathematical Formulation of the System Model
The method proposed in this work gives a solution to a multi-objective optimization

problem. The objective of the system is to minimize the daily energy cost for the user
as well as reduce the level of dissatisfaction caused due to the cost minimization. It
generates an optimal time schedule for the operation of the user appliances for a whole
day (12 a.m. to 11 p.m.) in order to achieve the objective. The total time horizon taken
for simulation is one day. One day is divided into 24 time slots and one slot is
equivalent to one hour.
A Model for Optimizing Cost of Energy … 525
3.1 Formulation of the Cost Function

The cost of the energy during an hour is defined in terms of total energy consumed
during that hour. The cost for an hour is proportional to the total hourly energy usage of
the consumer. In literature, several cost functions have been used. Since a logarithmic
cost function follows a linear growth, it can reduce the inconvenience of the user
caused due to increase in cost. The cost function for an hour h is represented as pðlh Þ
which is defined as follows:
pðlh Þ ¼ j lh logðlh þ 1Þ ð1Þ
where j is the cost parameter. The value set for j is more during daytime and less
during nighttime that can control the consumption of the users during daytime, because
generally the daytime is considered as the peak time in a day. lh is the total load
consumed during hour h which is the sum of the energy consumption of all the
appliances during that hour described by the following equality.
X
lh ¼ eha ð2Þ
a2A
A is the set of all appliances the user possesses. The energy consumption of an
appliance a during hour h is denoted as eha . The cost during an hour is increased when
load lh is increased during that hour. By representing pðlh Þ as Ph , the total daily
electricity cost termed as C can be defined as:
X
24
C¼ Ph ð3Þ
h¼1
The optimizer tries to minimize the total cost to be paid for the day which is being
considered for optimization.
3.2 Formulation of the Dissatisfaction Factor

Every appliance has a specific energy requirement and time of operation. The allowed
time period during which an appliance can be operated is decided by the user which is
based on its energy requirement. An appliance can start at any hour within this allowed
time of operation and has to finish operating on or before the end of this period. The
optimizer generates the starting time of every appliance which lies between the allowed
time intervals for the appliance with a purpose to minimize the cost. However, the user
can have a preferable starting time for every appliance. When the scheduled starting
time does not match with the preferred starting time, the user gets some displeasure.
The scheduled start time may fall before or after the preferred start time. When the
difference between the scheduled starting time and preferred starting time increases the
dissatisfaction is also increased which has to be reduced. The description of the dis-
satisfaction factor in the optimization process is given below.
The rate of dissatisfaction caused due to cost optimization is estimated using the
following function.
fa ¼ x=xmax ð4Þ
where x ¼ jpa Ta j
pa is the user’s preferred starting hour for appliance a and Ta is the starting hour
calculated by the scheduler. It is to be noted that the values of pa and Ta are considered
in hours. xmax is the maximum possible value of x. The value of x lies between [0, 23].
The minimum value for x occurs when scheduled starting time becomes equal to
preferred starting time. Whereas the maximum value for x occurs when for an appliance
the possible time of operation is whole day, that means the preset time interval for the
operation of the appliance is from 1st hour of the day to the last hour of the day and
value of preferred starting time slot is 1(1st hour of the day) and scheduled starting time
slot is 24 (last hour of the day). So, in that case, the gap between the preferred starting
time and scheduled starting time is 23. In the worst case, the time gap is 23, and in the
best case, it is 0. For simplicity of calculation, fraction of a time slot (hour) has not been
considered here. fa is increasing with every increase in x. The range of fa lies between
[0, 1]. The function used for calculating the unhappiness or dissatisfaction denoted as
Ua is defined as follows:
Ua ¼ ðma Þfa ð5Þ
where ma [ 1. The value of ma can be different for different appliances which is set by
the user. The appliance for which the user wants minimum dissatisfaction the value of
ma for that appliance will be higher than the values set for other appliances. Thus, for
higher values of ma , the value of Ua will be higher which means the dissatisfaction term
will have higher value in the optimization function. The value for Ua is increased when
the rate of dissatisfaction fa is increased. When x is 0, the value for Ua is 1 as fa is 0 and
when x is maximum, Ua attains its maximum value for a specific value of ma . It means
when the scheduled starting time is equal to the preferred starting time, the dissatis-
faction is zero. However, when there is a difference between the two, there is dissat-
isfaction and when the difference is increased, the dissatisfaction is increased. The level
of dissatisfaction for all the appliances is to be minimized through optimization.
3.3 Formulation of the Optimization Function

With the above generalizations, the optimization function can be defined as:
minimize h1 C þ h2 D ð6Þ
Subject to the following constraints
eha ¼ 0; if h 62 ½Sa ; Fa
X
24
La ¼ eha
h¼1
pamin ðaÞ eha pamax ðaÞ

P
where D ¼ a2A Ua
The parameters h1 and h2 represent the significance of each objective of the pro-
posed optimization problem. h1 describes the importance of energy cost reduction, and
h2 describes the importance of dissatisfaction level minimization. The sum of h1 and h2
is 1. Sa is the allowed starting hour of appliance a, and Fa is the allowed finishing hour
of a. The interval ½Sa ; Fa for an appliance a is decided by the user. So for every
appliance, the allowed time of operation is determined by the user and beyond which
the energy consumption of that appliance is 0 as described by the first constraint. The
second constraint says that the total energy consumption for an appliance a in a whole
day denoted as La is predefined based on the user’s consumption behavior or user’s
choice. The third constraint says every appliance has a minimum power level and
maximum power level denoted as pamin and pamax , respectively. C is the total daily cost
and D represents the total dissatisfaction for all the appliances operated in the whole
day.
4 Results of Optimization
In this work particle, swarm optimization (PSO) technique has been used for opti-
mization process. First optimization is done for one user only and then it is applied for
more than one user.
The optimization process minimizes the objective function and generates the
operation time schedule for each appliance. That means it generates the optimal eha
value for every appliance for all the hours in a whole day. The following figures
confirm the effectiveness of the algorithm.
Figure 1 shows the convergence of the proposed optimization algorithm. Figure 2
shows that when the optimization is not done the energy consumption of the user is
more during peak time. The forecasted load curve (general residential consumption
pattern without optimization) used here is similar to the consumption pattern used in
[9]. It is clear from the figure that when optimization is followed a part of the energy
consumption is shifted to non-peak time making the demand low during peak hours
which is very much required to create a proper balance between demand and supply of
electricity. Figure 3 shows the comparison of the energy costs with and without
optimization for 25 days. It shows that the cost for all these days remains lower with
optimization. Figure 4 shows that the energy consumption pattern is changed when
more importance is given satisfaction rather than cost minimization. There is a trade-off
between the cost minimization and dissatisfaction minimization. Figure 5 shows a
comparison between the monthly electricity payment made without optimization and
with optimization for 15 different users. It is obvious that the cost calculated after
Fig. 1. Convergence of the algorithm
Fig. 2. Comparison of energy usage in a day before and after optimization
Fig. 3. Cost of energy before and after optimization for a single user for 25 days
Fig. 4. Difference in energy consumption pattern with a different objective
Fig. 5. Comparison of cost before and after optimization for 15 users
optimization is lower than that of before optimization. Figure 6 shows the fact that for
more satisfaction the user has to pay more. On the other hand, the user can save some
money if it can tolerate a little discomfort.
Fig. 6. Higher cost for higher satisfaction
5 Conclusion
The proposed method reduces the user demand for electricity during peak hours by
producing a schedule for the operation of the electrical appliances for the whole day.
Without proper scheduling, the user may try to operate its appliances mainly during
peak hours which increases the peak hour consumption and the electricity costs. But
when the user follows the schedule generated by the optimizer, the electricity payments
is minimized. The optimizer process used here not only minimizes the electricity
payments but also reduces the dissatisfaction level of the user while using the appli-
ances maintaining a balance between the two. If the user wants less dissatisfaction, then
the electricity payment made will be more and if a reduced payment is preferred, then a
little discomfort has to be tolerated. However, in both the cases, the cost will be less
than the scenario with no optimization.
References
1. Mohsenian-Rad H, Leon-Garcia A (2010) Optimal residential load control with price
prediction in real-time electricity pricing environments. IEEE Trans Smart Grid 1(2):120–133
2. Zhu Z, Tang J, Lambotharan S, Chin WH, Fan Z (2011) An integer linear programming and
game theory based optimization for demand-side management in smart grid. In: IEEE
international workshop on smart grid communications and networks
3. Zhu Z, Tang J, Lambotharan S, Chin WH, Fan Z (2012) An integer linear programming based
optimization for home demand-side management in smart grid. IEEE PES Innov Smart Grid
Technol (ISGT) (2012)
4. Nguyen HK,Songl JB, Hanl Z (2012) Demand side management to reduce peak-to-average
ratio using game theory in smart grid. In: Proceedings of IEEE INFOCOM (2012)
5. Liu Y, Yuen C, Huang S, Hassan NU, Wang X, Xie S (2014) Peak-to-average ratio
constrained demand-side management with consumer’s preference in residential smart grid.
IEEE PES Innov Smart Grid Technol 8(6):1084–1097
6. Soliman HM, Leon-Garcia A (2014) Game-theoretic demand-side management with storage

devices for the future smart grid. IEEE Trans Smart Grid 5:1475–1485
7. Fadlullah ZMd, Quan DM, Kato N, Stojmenovic I (2014) GTES: an optimized game-theoretic
demand-side management scheme for smart grid. IEEE Syst J 8:588–597
8. Ye F, Qian Y, Hu RQ (2016) A real-time information based demand-side management system
in smart grid. IEEE Trans Parallel Distrib Syst 27(2):329–339
9. Labeeuw W, Deconinck G (2013) Residential electrical load model based on mixture model
clustering and markov models. IEEE Trans Industr Inf 9(3):1561–1568
A Circular Ultra-Wideband Antenna
for Wearable Applications
Shaktijeet Mahapatra(&), Sarmistha Satrusallya,

and Mihir Narayan Mohanty
Department of ECE, ITER, Siksha ‘O’ Anusandhan (Deemed to be University),

Bhubaneswar, India
shaktijeetmahapatra@soa.ac.in
1 Introduction
With the development of wearable antennas, the wearable devices and sensors are
being increasingly used by the medical practitioners, the military and the scientific
community. The reason is that the antennas with smaller profile and low weight are
being developed. The UWB band for wearable antennas (3–10.6 GHz) offers more
freedom to the designers in terms of low-power, frequency and bandwidth [1]. The
band generally is used for ubiquitous computing but nowadays it is being used for
wearable applications as well owing to less congestion in this band.
Klemm et al. [2] designed two ultra-wideband antennas—coplanar waveguide fed
disc monopole antenna and microstrip-fed annular slot antenna—for body area network
applications. These antennas were designed entirely using textile only. Almpanis et al.
[3] presented an inverted truncated annular conical dielectric resonator antenna for
body area network applications. The operating range of antenna was 3.0–5 GHz. See
et al. [4] presented three antennas for UWB applications, namely directional suspended
plate antenna, omnidirectional monopole antenna and printed diversity antenna for on-
body communications. Shin et al. [5] presented a CPW-fed stair-shaped antenna having
ultra-wide bandwidth for wireless body area network (WBAN). Kang et al. [6] pre-
sented a folded antenna working in UWB range aiming to reduce backward radiation
and other effect arising out of proximity to human body.
In this paper, we present a circular patch antenna with four square slots cut on the
edges. The advantages of this antenna are that it is small, very lightweight and has a
thin profile and is easy to fabricate. The ground plane has been kept full to cut down
backward radiation.
The paper is organized into the following sections: In Sect. 2, the design of the
antenna is described; in Sect. 3, simulation results are presented; in the last section,
conclusions are drawn.

https://doi.org/10.1007/978-981-15-2774-6_62
A Circular Ultra-Wideband Antenna for Wearable Applications 533
2 Design Considerations
The proposed antenna is a circular patch antenna of 13 mm radius on a cylindrical

substrate of radius 20 mm and height of 1.6 mm, fed by a microstrip line. FR4-epoxy,
with a relative permittivity of 4.4 and a dielectric loss tangent of 0.02, was chosen as
material for the substrate. Four square slots of 5 mm were cut from the edges. The feed
was recessed by 1 mm and a quarter-wave line was inserted for matching input
impedance with the 50 X port impedance. Figure 1 shows the proposed antenna.
Fig. 1. Design of the antenna (proposed)
The design equation is as follows [7]:
F
b¼n h io1=2 ð1Þ
1þ 2h
pr F ln ðpFÞ
2h þ 1:7726
8:791 109
F¼ pffiffiffiffi ð2Þ
fr r
where h is in cm (Table 1).

534 S. Mahapatra et al.
Table 1. Design parameters

Parameters Values (in mm)
a 40
b 26
c 5
d 3.9
e 2.2
f 4
As can be clearly seen from Fig. 2, the antenna resonates at 3.8 and 9.74 GHz. The
impedance bandwidths obtained are 1.87 GHz and 0.38 GHz, respectively. The S11 or
the return loss obtained is −19 dB and −22 dB, respectively.
S11 vs Frequency UWB_Wearable_Antenna ANSOFT

0.00
MY1: -10.0000
dB(St(Patch_T1,Patch_T1))
-10.00
-20.00
-30.00
Curve Info
-40.00 dB(St(Patch_T1,Patch_T1))
Setup1 : Sweep
-50.00
0.00 2.50 5.00 7.50 10.00 12.50 15.00 17.50 20.00
Freq [GHz]
Fig. 2. S11 versus frequency
Figure 3 shows the electric field distribution on the patch at 3.8 GHz. As can be
clearly seen that due to the slots, the value of the electric field at the edges range from
654.4 to 915 V/m. This results in almost directional beam as can be seen from Figs. 4
and 5. A closer inspection also reveals that due to full ground structure there is no
backward radiation, making this antenna an excellent candidate for wearable applica-
tions and on-body communications.
A Circular Ultra-Wideband Antenna for Wearable Applications 535
Fig. 3. Electric field at 3.8 GHz
Fig. 4. 3D polar plot of gain at 3.8 GHz

536 S. Mahapatra et al.
Fig. 5. 3D polar plot of gain at 9.74 GHz
4 Conclusions
A microstrip-fed slotted circular patch antenna has been designed for UWB wearable
applications. The antenna was found to be resonating at 3.8 and 9.74 GHz. The
advantages of this design lie in the simplicity of design, and no backward radiation.
Owing to the above-mentioned advantages, the proposed antenna can be easily used for
body area network applications and on-body communication applications easily.
References
1. Hall PS, Hao Y (2012) Antennas and propagation for body-centric wireless communications.
Artech house
2. Klemm M, Troester G (2006) Textile UWB antennas for wireless body area networks. IEEE
3. Almpanis G et al (2008) A truncated conical dielectric resonator antenna for body-area
network applications. IEEE Antennas Wireless Propag Lett 8:279–282
4. See TS, Chen ZN (2009) Experimental characterization of UWB antennas for on-body
communications. IEEE Trans Antennas Propag 57(4):866–874
5. Shin H, Kim J, Choi J (2009) A stair-shaped CPW-fed printed UWB antenna for wireless
body area network. In: 2009 Asia Pacific Microwave Conference, IEEE
6. Kang C-H, Wu S-J, Tarng J-H (2011) A novel folded UWB antenna for wireless body area
network. IEEE Trans Antennas Propag 60(2):1139–1142
7. Balanis CA (2016) Antenna theory: analysis and design, wiley
Author Index
A Das, Pradipta, 141

Agrawal, Ramachandra, 299, 338 Das, Pranati, 217
Alamuru, Sunanda, 100 Das, P. S., 248
Alankar, Bhavya, 483 Das, Radhakrishna, 209
Arun Kumar, Bhukya, 63 Das, Rajendra Kumar, 94
Atif, Mohd., 475 Das, Smruti Rekha, 330
Das, S.R., 30
B Debata, Prajna Paramita, 330
Bal, Prasanta Kumar, 384
Barik, Shekharesh, 299 G
Behera, Bhagyalaxmi, 152 Garg, Dinesh Kumar, 52
Behera, Debasis, 78 Ghosh, Debalina, 272
Behera, Suchismita, 86 Goel, Nagendra Kumar, 1
Bhuyan, Satyanarayan, 194 Goel, Sonali, 278
Biswal, Bedanta Kumar, 122 Goyal, Vishal, 179
Biswal, B.N., 413
Biswal, Kumar, 224 J
Biswal, Pradyut Kumar, 313, 360 Jena, Bibekananda, 278
Biswal, Pravat Kumar, 504, 514 Jena, Lambodar, 232
Jena, Narendra Kumar, 254, 447
C
Chakravarty, S., 413 K
Chauhan, Ritu, 475 Kar, Durga P., 194
Choudhury, Satish, 278 Kar, Nihar, 366
Kar, Sanjeeb, 224
D Kaur, Harleen, 264, 475
Das, Abhishek, 161, 458 Kirmani, Mehreen, 52
Dash, Lipsa, 100
Dash, Meera, 20 M
Dash, Sachikanta, 94 Mahajan, Shubham, 171
Dash, Shibanee, 394 Mahapatra, Shaktijeet, 189, 532
Das, Kunal Kumar, 69, 424 Malik, Suman, 10
Das, Manoranjan, 432 Mishra, Alok Kumar, 254, 299, 321, 338
Das, Nilima R., 523 Mishra, Debahuti, 330
Das, Niva, 86 Mishra, Guru Prasad, 394
© Springer Nature Singapore Pte Ltd. 2020 537

https://doi.org/10.1007/978-981-15-2774-6
538 Author Index
Mishra, Laxmi Prasad, 201, 458 R

Mishra, Puneet, 179 Rai, Satyananda C., 523
Mishra, Sivkumar, 495, 504, 514 Raj, Raman, 122
Mohanty, Bibhuprasad, 109, 116, 135 Ram, Rashmirekha, 161
Mohanty, Kanungo B., 447 Rath, Abhisek, 495
Mohanty, Mihir Narayan, 152, 189, 347, 532 Rath, Debswarup, 209
Mohanty, Monalisa, 313 Rout, Bidyadhar, 254
Mohanty, Sangram Kishore, 30, 248 Rout, Bikash Chandra, 78
Mohanty, Saumendra Kumar, 394 Rout, Kshirod Kumar, 495, 504, 514
Mohapatra, Saumendra Kumar, 161, 458 Rout, Pravat Kumar, 347
Mohapatra, Subhasish, 46 Roy, Jibendu Sekhar, 424
Moharana, Laxmipriya, 122
Mourya, Ashish Kumar, 264 S
Sabut, Sukanta, 313, 366
N Sahoo, Arabinda, 217
Naik, Manoj Kumar, 375 Sahoo, Santanu, 366
Naik, Suraj, 122 Sahoo, Subhadra, 447
Nanda, Amar Bijaya, 447 Sahoo, Subhaluxmi, 289
Nanda, Anuja, 254, 321, 338 Sahoo, Swarnaprava, 201
Nanda, Pradipta Kumar, 289 Sahoo, Tamanna, 109
Nayak, Ajit, 523 Sahoo, Tapasmini, 69
Nayak, Soumen, 232 Sahu, Badrinarayan, 366, 438
Nayak, Suryakanta, 467 Sahu, Benudhar, 432
Nayk, Praveen P., 194 Sahu, Binod Kumar, 447
Sahu, Prasant Kumar, 10, 37
P Sahu, Sanjay Kumar, 63
Padhi, B.K., 413 Salik, Mohammad, 347
Padhi, Jagadish C., 194 Samal, Debashisa, 438
Palai, G., 30, 248 Samal, Laxmipriya, 438
Palo, Hemanta Kumar, 78 Samantaray, Barsa, 424
Panda, Leeza, 116 Samant, Sunita, 289
Panda, Mohit Ranjan, 141 Sarma, Kandarpa Kumar, 1
Panda, Mrutyunjaya, 467 Sarma, Mousmita, 1
Panda, Prakash Kumar, 272 Satapathy, Lalit Mohan, 299
Panda, Susmita, 141 Satapathy, Samarjeet, 299
Pandit, Amit Kant, 171 Satrusallya, Sarmistha, 189, 532
Panigrahi, Santisudha, 321 Sethi, Kaibalya Kumar, 63
Panigrahi, Trilochan, 20 ShafqatUlAhsaan, 264
Pani, Sasmita, 240 Sharma, Renu, 20, 278
Parhi, Pournamasi, 330 Shukla, Aasheesh, 179
Parija, Smita, 46 Sonagara, Abhishek M., 405
Patnaik, Dipti, 194 Subudhi, Asit Kumar, 313
Patra, Abhishek, 338 Subudhi, Dillip Kumar, 338
Patra, Akshaya Kumar, 254, 299, 321, 338 Susmitha, A., 100
Pattanayak, Binod Kumar, 240 Swain, Biswaranjan, 194
Pattnaik, Omkar, 240 Swain, Jnana Ranjan, 299
Pradhan, Sateesh Kumar, 384 Swain, K.P., 30, 248
Pradhan, Sovit Kumar, 209 Swain, Ramakrushna, 232
Prasad, S. S. P. B. K., 209
Priyadarshini, Rojalina, 141 T
Prusty, Pankaj, 135 Tangudu, Ramji, 37
Prusty, Vedanta, 495 Tripathy, Madhab Chandra, 224
Author Index 539
Tripathy, Shuvendra Kumar, 360 Vishwakarma, Rahul, 405
V W
Varshney, S.K., 152 Wunnava, Aneesh, 375

10.1007@978 981 15 2774 6 PDF

Uploaded by

Copyright:

Available Formats

10.1007@978 981 15 2774 6 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

10.1007@978 981 15 2774 6 PDF

Uploaded by

Copyright:

Available Formats

What is the series 'Lecture Notes in Networks and Systems' about?

What is the series 'Lecture Notes in Networks and Systems' about?

What topics are covered in the series?

What topics are covered in the series?

Lecture Notes in Networks and Systems 109

Mihir Narayan Mohanty

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

ISSN 2367-3370 ISSN 2367-3389 (electronic)

This issue of Lecture Notes in Networks and Systems is dedicated to the

Bhubaneswar, India Mihir Narayan Mohanty

Children’s Age and Gender Recognition from Raw Speech Waveform

Comparison of Classiﬁers for Speech Emotion Recognition (SER)

An Improved Grasshopper Optimization Algorithm for Solving

Electrical Energy Storage: A Great Business Ahead . . . . . . . . . . . . . . . 278

A Novel Graphical User Interface-Based Toolbox for Optimization

Designing and Sizing of a Stand-alone Photovoltaic System:

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537

About the Editors

Dr. Mihir Narayan Mohanty received his M.Tech. in Communication System

Swagatam Das received his B. E. Tel. E., M. E. Tel. E (Control Engineering

He is an editorial board member of numerous other journals, including Progress in

Ramachandra Agrawal Department of CSIT, ITER, S‘O’A University, Deemed

Kumar Biswal School of Electronics Engineering, Kalinga Institute of Industrial

Smruti Rekha Das Department of Computer Science and Engineering, Gandhi

Sonagara Abhishek M Department of Electronics and Communication

Saumendra Kumar Mohapatra ITER, Siksha ‘O’ Anusandhan (Deemed to be

Leeza Panda Department of Electronics and Communication Engineering, ITER,

S. S. P. B. K. Prasad Department of Electrical & Electronics Engineering, GIET,

Subhaluxmi Sahoo Image & Video Analysis Laboratory, Department of ECE, S

Mousmita Sarma Department of Electronics and Communication Engineering,

S. K. Varshney Department of Electronics and Electrical Communication

Mousmita Sarma1(&), Kandarpa Kumar Sarma1,

© Springer Nature Singapore Pte Ltd. 2020

memory (LSTM) cell-based recurrent layer (TDNN-LSTM)-based conﬁguration. The

Fig. 1. a Single-task learning model. b Multi-task learning model

where M is the total number of examples in the dataset.

3.1 DNN Design Consideration

3.2 Result and Analysis

Table 3. Summary of results (averaged over all test sessions)

4 Summary of Findings and Conclusion

Suman Malik(&) and Prasant Kumar Sahu

School of Electrical Science, Indian Institute of Technology Bhubaneswar,

Vehicular ad hoc network (VANET) aims at enhancing the transportation efﬁciency

© Springer Nature Singapore Pte Ltd. 2020

3 VANET Model, Methodology and Simulation Setup

3.1 Simulation Using NS2 and NetSim with SUMO Tools

Fig. 2. Methodology flowchart for SUMO and NS2

Fig. 3. Methodology flowchart for SUMO and NetSim

3.2 Packet Loss Ratio

PLR ¼ Total packet dropped/Total packet send ð1Þ

3.3 Average Delay

Average Delay ¼ Time consumed to transfer packets successfully/

Table 1. Simulation parameters

Nakagami Channel model Rayleigh Channel model

Packet loss ratio (%)

Routing Protocols --> Routing Protocols -->

DSR routing and

Packet loss ratio (%)

Car velocity (Kmph)--> Car velocity (Kmph)-->