Recent Trends in Communication and Intelligent Systems: Proceedings of ICRTCIS 2021 (Algorithms for Intelligent Systems) 9811913234, 9789811913235

Citation preview

Algorithms for Intelligent Systems Series Editors: Jagdish Chand Bansal · Kusum Deep · Atulya K. Nagar

Aditya Kumar Singh Pundir Neha Yadav Harish Sharma Swagatam Das   Editors

Recent Trends in Communication and Intelligent Systems Proceedings of ICRTCIS 2021

Algorithms for Intelligent Systems Series Editors Jagdish Chand Bansal, Department of Mathematics, South Asian University, New Delhi, Delhi, India Kusum Deep, Department of Mathematics, Indian Institute of Technology Roorkee, Roorkee, Uttarakhand, India Atulya K. Nagar, School of Mathematics, Computer Science and Engineering, Liverpool Hope University, Liverpool, UK

This book series publishes research on the analysis and development of algorithms for intelligent systems with their applications to various real world problems. It covers research related to autonomous agents, multi-agent systems, behavioral modeling, reinforcement learning, game theory, mechanism design, machine learning, meta-heuristic search, optimization, planning and scheduling, artificial neural networks, evolutionary computation, swarm intelligence and other algorithms for intelligent systems. The book series includes recent advancements, modification and applications of the artificial neural networks, evolutionary computation, swarm intelligence, artificial immune systems, fuzzy system, autonomous and multi agent systems, machine learning and other intelligent systems related areas. The material will be beneficial for the graduate students, post-graduate students as well as the researchers who want a broader view of advances in algorithms for intelligent systems. The contents will also be useful to the researchers from other fields who have no knowledge of the power of intelligent systems, e.g. the researchers in the field of bioinformatics, biochemists, mechanical and chemical engineers, economists, musicians and medical practitioners. The series publishes monographs, edited volumes, advanced textbooks and selected proceedings. Indexed by zbMATH. All books published in the series are submitted for consideration in Web of Science.

More information about this series at https://link.springer.com/bookseries/16171

Aditya Kumar Singh Pundir · Neha Yadav · Harish Sharma · Swagatam Das Editors

Recent Trends in Communication and Intelligent Systems Proceedings of ICRTCIS 2021

Editors Aditya Kumar Singh Pundir Arya College of Engineering & IT Jaipur, Rajasthan, India Harish Sharma Department of Computer Science and Engineering Rajasthan Technical University Kota, Rajasthan, India

Neha Yadav Department of Mathematics National Institute of Technology Hamirpur, Himachal Pradesh, India Swagatam Das Electronics and Communication Sciences Unit Indian Statistical Institute Kolkata, West Bengal, India

ISSN 2524-7565 ISSN 2524-7573 (electronic) Algorithms for Intelligent Systems ISBN 978-981-19-1323-5 ISBN 978-981-19-1324-2 (eBook) https://doi.org/10.1007/978-981-19-1324-2 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore

Conference Program Committee

International Advisory Committee Dr. Bimal K. Bose, Emeritus Professor, E&C Engineering, The University of Tennessee, Knoxville, USA Dr. Subramaniam Ganesan, Professor, E&C Engineering, Oakland University, USA Dr. L. M. Patnaik, Adjunct Professor and INSA Senior Scientist and Former Professor, Department of Electrical Systems Engineering, IIS, Bengaluru, India Dr. Ramesh Agarwal, Professor, School of Engineering and Applied Science, Washington University Dr. Vincenzo Piuri, Professor, University of Milan, Italy Dr. Ashoka Bhat, Professor, Department of Electrical and Computer Engineering, University of Victoria, Canada Prof. Akhtar Kalam, Head of External Engagement, Victoria University, Victoria, Australia Dr. M. H. Rashid, Professor, E&C Engineering, University of West Florida, USA Dr. Fushuan Wen, Director, Zhejiang University-Insigma, Joint Research Center for Smart Grids, China Prof. Ir. Dr. N. A. Rahim, Director, UMPEDAC, UM, Kuala Lumpur, Malaysia Prof. Rafael F. S. Caldeirinha, Coordinator Professor, Polytechnic of Leiria and Instituto de Telecomunicações, Portugal Dr. Tarek Bouktir, Professor, EE, University of Setif, Algeria Dr. Pietro Savazzi, Telecommunications and Remote Sensing Laboratory, Department of Electrical, Biomedical and Computer Engineering, University of Pavia, Italy Dr. Ouri Wolfson, President and Chief Scientist, Pirouette Software, Inc., Fellow of the ACM, IEEE, AAAS, Chicago, IL Dr. Álvaro Rocha, Professor, Information Systems at the University of Lisbon—ISEG Dr. Pavel Loskot, Associate Professor, ZJU-UIUC Institute, Haining, Zhejiang, China Dr. S. Mekhilef, PEARL, EE, University of Malaya, Kuala Lumpur, Malaysia

v

vi

Conference Program Committee

Dr. Mehmet Emir Koksal, Associate Professor, Department of Mathematics, Ondokuz Mayis University, Atakum, Samsun—Turkey Dr. Biplob Ray, Senior Lecturer—ICT | Discipline Lead—Postgraduate Foundation, School of Engineering and Technology, CQUniversity Australia, Melbourne Campus

National Advisory Committee Dr. S. N. Joshi, Em. Scientist, CSIR-CEERI Pilani Dr. Vineet Sahula, Professor, ECE, MNIT, Jaipur, India Dr. K. J. Rangra, Chief Scientist and Professor AcSIR, CSIR-CEERI, Pilani, India Dr. R. K. Sharma, Sr. Principal Scientist, Professor, AcSIR, CSIR-CEERI, Pilani, India Dr. Vijay Janyani, Professor, ECE, MNIT, Jaipur, India Dr. K. R. Niazi, Professor, EE, MNIT, Jaipur, India Dr. V. K. Jain, Former Director Grade Scientist, Solid State Physics Lab., DRDO, India Dr. Manoj Kumar Patairiya, Director, CSIR, NISCAIR, India Dr. Sanjeev Mishra, Professor, UDME, RTU Kota Dr. Kailash N. Srivastava, VC, SUAS, Indore, Prof. Harpal Tiwari, Professor, EE, MNIT, Jaipur, India Dr. S. Gurunarayanan, Professor, BITS, Pilani, India Dr. Ghanshyam Singh, Professor, ECE, MNIT, Jaipur, India Dr. R. Kumar, GEC, Ajmer, Rajasthan, India Prof. Satish Kumar, Sr. Principal Scientist, AMS, Professor, AcSIR, CSIR-CSIO, Chandigarh, India Dr. Kota Srinivas, Chief Scientist, CSIR-CSIO, Chennai Centre, CSIR Madras Complex, Chennai Sh. Anand Pathak, President, SSME, India Sh. Ulkesh Desai, Vice President, SSME, India Sh. Ashish Soni, Secretary, SME, India Sh. R. M. Shah, Joint Secretary, SSME, India Sh. Vimal Shah, Treasurer, SSME, India Acharya (Er.) Daria Singh Yadav, Chairman, ISTE Rajasthan and Haryana Section Er. Sajjan Singh Yadav, Chairman, IEI, Jaipur Er. Gautam Raj Bhansali, Hony Secretary, IEI, Jaipur Dr. J. L. Sehgal, IEI, Jaipur Smt. Annapurna Bhargava, IEI, Jaipur Smt. Jaya Vajpai, IEI, Jaipur Dr. Hemant Kumar Garg, IEI, Jaipur Er. Gunjan Saxena, IEI, Jaipur Er. Sudesh Roop Rai, IEI, Jaipur Dr. Manish Tiwari, Manipal University Jaipur Dr. Dinesh Yadav, Manipal University Jaipur

Conference Program Committee

vii

Dr. Jitendra Kumar Deegwal, Principal, Government Women Engineering College, Ajmer

Organizing Committee Chief Patrons Smt. Madhu Malti Agarwal, Chairperson, Arya Group Er. Anurag Agarwal, Group Chairman

Patrons Prof. Arun Kr Arya, Principal, ACEIT, Jaipur

General Chairs Dr. Swagatam Das, ISI Kolkata Dr. Harish Sharma, RTU Kota Dr. Neha Yadav, NIT Hamirpur Dr. Vibhakar Pathak, ACEIT, Jaipur

Conveners Dr. Kirti Vyas, ACEIT, Jaipur Dr. Chhavi Saxena, ACEIT, Jaipur Mr. Sachin Chauhan, ACEIT, Jaipur Er. Ankit Gupta, ACEIT, Jaipur

Co-conveners Ms. Chanchal Sharma, ACEIT, Jaipur Mr. Amit Sharma, ACEIT, Jaipur Mr. Devendra Soni, ACEIT, Jaipur

viii

Organizing Chair and Secretaries Dr. Rahul Srivastava, ACEIT, Jaipur Dr. Aditya Kumar Singh Pundir, ACEIT, Jaipur

Special Session Chair Dr. Sarabani Roy, Jadavpur University, Kolkata Dr. Nirmala Sharma, RTU Kota Dr. Irum Alvi, RTU Kota Dr. S. Mekhilef, University of Malaya, Malaysia

Local Organizing Committee Chairs Prof. Akhil Pandey, ACEIT, Jaipur Prof. Prabhat Kumar, ACEIT, Jaipur Prof. Manu Gupta, ACEIT, Jaipur Prof. Shalani Bhargava, ACEIT, Jaipur Shri Ramcharan Sharma, ACEIT, Jaipur

Conference Program Committee

Preface

This volume comprises papers those presented at the AICTE Sponsored Online 3rd International Conference on Recent Trends in Communication & Intelligent Systems (ICRTCIS 2021), organized by Arya College of Engineering and Information Technology and hosted by the Department of Electronics and Communication Engineering. The conference is sponsored under AICTE-GOC Scheme. The presented papers cover a selective high impacted areas and topics related to intelligent systems and communication networks, including intelligent computing and converging technologies, intelligent system communication and sustainable design and intelligent control, measurement and quality assurance. This volume of algorithms for intelligent systems brings best 34 of the presented papers. Each of them presents new approaches and/or evaluates methods to real-world problems and exploratory research that describes novel approaches in the field of intelligent systems. ICRTCIS 2021 has received (all tracks) 170 submissions, 64 of them were shortlisted for presentation and 34 papers finally shortlisted. The authors with wide diversity around the globe have been participated from different countries like Albania, Canada, Ethiopia, India, Kuwait, New Zealand, Nigeria, Turkmenistan and Saudi Arabia. Nine renowned keynotes around the globe have delivered with glimpse of their current research and also interacted with the participants. We hope that all the participants have enjoyed the parallel sessions and presentation tracks: two (first day) and three parallel tracks (second day) after the keynotes. We strongly believe that this international conference series provides a platform to the igniting minds of the young researchers for undertaking more interdisciplinary and collaborative research.

ix

x

Preface

The editors trust that this volume will be useful and interesting to readers for their own research work. Jaipur, India Hamirpur, India Kota, India Kolkata, India December 2021

Aditya Kumar Singh Pundir Neha Yadav Harish Sharma Swagatam Das

Contents

Hybrid Feature Extraction Technique for Tamil Automatic Speech Recognition System in Noisy Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . S. Girirajan and A. Pandian

1

Implementation and Image Transformation for Ground Penetration Image Radar System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dharamvir and M. S. Shashidhara

13

Effect of Geospatial Weather Features on COVID-19 Spread in Maharashtra State Using Machine Learning . . . . . . . . . . . . . . . . . . . . . . . Ranu Sewada and Hemlata Goyal

21

Adiabatic Logic Code Converter Design at Different Sub-micron Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . J. V. N. Saketh Ram, Vaibhav Kumar Jain, and Abhijit Asati

31

A Cross-connected Switch Capacitor Multilevel Inverter: A Proposed Topology and Its Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lipika Nanda, Chitralekha Jena, Babita Panda, and Arjyadhara Pradhan

45

A Comparative Study on Sentiment Analysis of Uber and Ola Customer Reviews Based on Machine Learning Approaches . . . . . . . . . . . Sandeep Kumar, Anuj Kumar Singh, Shashi Bhushan, Pramod Kumar, and Arun Vashishtha A Cost-Effective IoT-Assisted Framework for Automatic Irrigation . . . . Rewa Sharma and Keshav Kaushik

55

65

Constrained Optimization-Based Routing for Multipath and Multihop Propagation in WSN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Pratham Majumder and Punyasha Chatterjee

75

Performance Comparison of Fuzzy Logic and Evolutionary Algorithm-Optimized Controller for a Multi-area Power System . . . . . . . Parisa Latief Khan, Zahid Farooq, Sheikh Safi ullah, and Satish Saini

85

xi

xii

Contents

Leveraging CNNs for Real-Time Pothole Detection . . . . . . . . . . . . . . . . . . . Chaithra Reddy Pasunuru and Kruthika Muthireddy

93

IoT Devices for Detecting and Machine Learning for Predicting COVID-19 Outbreak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Shams Tabrez Siddiqui, Anjani Kumar Singha, Md Oqail Ahmad, Mohammad Khamruddin, and Riaz Ahmad Text Detection from Scene and Born Images: How Good is Tesseract? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 Nadeem Anwar, Tauseef Khan, and Ayatullah Faruk Mollah Power System Load Frequency Control of Hybrid Integrated with Solar-Thermal and Geothermal System . . . . . . . . . . . . . . . . . . . . . . . . . 123 Ayman Farooq, Zahid Farooq, and Krishna Tomar An Improved Stock Market Index Prediction System Based on LSTM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 Rais Allauddin Mulla and Satish Saini Age Estimation in Digital Radiograph Using HOG and DWT Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 A. Stella and Thirumalai Selvi Single-Layer-Single-UWB Patch Antenna for HXLPE-Based Artificial Hip Diagnosis in Microwave Tomography Spectrum . . . . . . . . . 157 Khalid Ali Khan, Suleyman Malikmyradovich Nokerov, Aravind Pitchai Venkataraman, Kehali Anteneh, and Diriba Chali CNN-Based Optimal Image Restoration and Comparative Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 Divya Sharma, Shilpa Sharma, and Harshal Patil Trust Offloading in Vehicular Cloud Networks . . . . . . . . . . . . . . . . . . . . . . . 179 Sarbjit Kaur and Ramesh Kait Deep Learning-Driven Structured Energy Efficient Affordable Ecosystem for Computational Learning Theory . . . . . . . . . . . . . . . . . . . . . . 189 Krishan Gopal Gupta, Samrit Kumar Maity, Abhishek Das, and Sanjay Wandhekar Design of the MIMO Antenna Using Metamaterial for S-Band Applications, with Reduced Mutual Coupling and Improved Diversity Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 B. Ramamohan and M. Siva Ganga Prasad Vulnerability Assessment of University Computer Network Using Scanning Tool Nexpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207 Kismat Chhillar and Saurabh Shrivastava

Contents

xiii

Development of Cyber-Physical Systems for Water Quality Monitoring in Smart Water Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 Punit Khatri, Karunesh Kumar Gupta, and Raj Kumar Gupta Identification of Fake News Using Machine Learning Techniques . . . . . . 225 Swati Pandey, Rashmi Gupta, and Jeetendra Kumar Epileptic Seizure Detection Using Wavelet-Based Features from Different Sub-bands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237 Pallavi S. Meshram and Damayanti C. Gharpure An Improved Locality-Sensitive Hashing-Based Recommender Approach in a Distributed Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249 Angadi Anupama, Pedada Saraswathi, Patruni Muralidhara Rao, and Gorripati Satya Keerthi Voodoo—The Magic Mirror . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259 Puneet Sharma, Prachi Kaushik, and Komal Bhagat A Performed Optimized Load Balancing Genetic Approach Technique in Cloud Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269 Arshiya, Jaspreet Singh, and Shruti Aggarwal Security Mechanism for Detection Coverage of Machine Learning-Based IDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281 Amit Kundaliya, Prachi Juyal, and Nirmal Sharma Video Feature Tagging and Real-Time Feature Search . . . . . . . . . . . . . . . . 289 Mithil Dani, Sakshi Patil, and Pramod Bide Privacy-Preserving Data Mining in Web Domain Using Protected Data Extraction and Presentation Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 299 Vibhor Sharma, Shashi Bhushan, Anuj Kumar Singh, and Pramod Kumar Campus Placement Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311 Jahnvi Shah, Shivangi Kochrekar, Neha Kale, Sakshi Patil, and Anand Godbole Performance Analysis of FIR and IIR Filters Using ECG Signals . . . . . . 321 Chhavi Saxena, Rahul Srivastava, and D. P. Arora Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333

About the Editors

Dr. Aditya Kumar Singh Pundir (B.E., M.Tech. and Ph.D.) is currently working as Professor, Mentor Incubation and IPR Activities in Department of Electronics and Communication Engineering at Arya College of Engineering & IT, Jaipur. His current research interests include IoT-based memory testing, built-in self-test using embedded system design, machine learning, algorithms for CAD and signal processing. He has published six patents, several book chapters and three books to his credit and also authored more than 20 research papers in peer-reviewed refereed journals and conferences. Dr. Pundir was Organizing Chair, Convener and Member of the steering committee of several international conferences. Dr. Pundir is Life Member of the Indian Society for Technical Education (ISTE), Computer Society of India (CSI), Professional Member of Association for Computing Machinery (ACM) and IEEE. Dr. Neha Yadav received her Ph.D. in Mathematics from Motilal Nehru National Institute of Technology, (MNNIT) Allahabad, India, in year 2013. She completed her postdoctorate from Korea University, Seoul, South Korea. She is Receiver of Brain Korea (BK-21) postdoctoral fellowship given by Government of Republic of Korea. Prior to joining NIT Hamirpur, she taught courses and conducted research at BML Munjal University,Gurugram, Korea University Seoul, S. Korea and The NorthCap University, Gurugram. Her major research area include numerical solution of boundary value problems, artificial neural networks and optimization. Harish Sharma is Associate professor at Rajasthan Technical University, Kota, in Department of Computer Science and Engineering. He has worked at Vardhaman Mahaveer Open University Kota and Government Engineering College Jhalawar. He received his B.Tech. and M.Tech. degree in Computer Engineering from Government Engineering College, Kota, and Rajasthan Technical University, Kota, in 2003 and 2009, respectively. He obtained his Ph.D. from ABV—Indian Institute of Information Technology and Management, Gwalior, India. He is Secretary and one of the founder members of Soft Computing Research Society of India. He is Life Time Member of Cryptology Research Society of India, ISI, Kolkata. He is Associate Editor of xv

xvi

About the Editors

International Journal of Swarm Intelligence (IJSI) published by Inderscience. He has also edited special issues of the journals Memetic Computing and Journal of Experimental and Theoretical Artificial Intelligence. His primary area of interest is nature inspired optimization techniques. He has contributed in more than 56 papers published in various international journals and conferences. Dr. Swagatam Das received the 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. Swagatam Das is currently serving as 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 published more than 300 research articles in peer-reviewed journals and international conferences. He is Founding Co-Editor-in-Chief of Swarm and Evolutionary Computation, an international journal from Elsevier. He has also served as or is serving as Associate Editors of the Pattern Recognition (Elsevier), Neurocomputing (Elsevier), Information Sciences (Elsevier), IEEE Transactions on Systems, Man, and Cybernetics: Systems, IEEE Computational Intelligence Magazine, IEEE Access and so on. He is Editorial Board Member of Progress in Artificial Intelligence (Springer), Applied Soft Computing (Elsevier), Engineering Applications of Artificial Intelligence (Elsevier) and Artificial Intelligence Review (Springer). Dr. Das has 16,500+ Google Scholar citations and an H-index of 62 till date. He has been associated with the international program committees and organizing committees of several regular international conferences including IEEE CEC, IEEE SSCI, SEAL, GECCO and SEMCCO. He has acted as guest editors for special issues in journals like IEEE Transactions on Evolutionary Computation and IEEE Transactions on SMC, Part C. He is Recipient of the 2012 Young Engineer Award from the Indian National Academy of Engineering (INAE). He is also Recipient of the 2015 Thomson Reuters Research Excellence India Citation Award as the highest cited researcher from India in Engineering and Computer Science category between 2010 and 2014.

Hybrid Feature Extraction Technique for Tamil Automatic Speech Recognition System in Noisy Environment S. Girirajan and A. Pandian

1 Introduction ASR is a process of converting raw speech signal into corresponding text transcription. The raw speech signal can contain isolated words, large vocabulary continuous speech, connected words or spontaneous speech. It was really a challenging task to recognize large vocabulary continuous speech due to its variation such as speaker modulation, speaker attributes, background noise and detecting the start and end point of the word in the given speech signal [1]. The process involved in ASR system is categorized into three stages. First stage is recognizing the phones from a raw speech signal, and also, it involves features selection or dimensionality reduction. It will extract the useful features from the speech signal based on task-specific knowledge. Such a feature selection is performed by using Mel Frequency Cepstral Coefficient (MFCC) [2]. In the second stage, word is estimated based on the likelihood of the phones, it is called as lexicon model, final stage is to frame the sentence by considering the grammatical sequence of the particular language, and it is called as language model. Phonemes are generated directly from frames of the speech signal by using frame-level classification. Most of the ASR system developed for foreign languages like English, Chinese, German Languages used advance deep learning model and achieved better performance with high accuracy rate. Due to this in recent years, more number of research works is carried out for developing ASR system in Indian regional languages like Tamil, Hindi, Telugu, Malayalam, etc. In India population, nearly 65% of people live in rural area. Most of them lack in English fluency and computer knowledge. So, ASR system for Indian regional language will help people to interact with computers and other electronic gadgets that support voice commands [3]. S. Girirajan (B) · A. Pandian Department of Computer Science and Engineering, SRM Institute of Science and Technology, Kattankulathur, Tamil Nadu, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_2

1

2

S. Girirajan and A. Pandian

ASR systems act as communication bridge between human and machine. It is widely used to control electronics gadget such as Apple Siri, Amazon Alexa, etc. [4]. In case of LVCS signal in noisy environment, it is complex to detect the speech boundaries [5]. The effect of noise can be reduced in ASR system by using various methods. Most commonly used method is to train the model straight away in noisy environment. This type of noise reduction is known as matched system. This method uses the common noise that arises in speech for training. Performance of this method is far better when compared with several other available noise reduction methodologies. New noises that arise in environment need to be re-trained. It is complex and time consuming. So, re-training the model for each new noise that arises in recognition needs to be avoided [4]. In [4], various noise factors such as communication devices, speaker and listener hearing impairment that affects the accuracy of recognition are identified and provide the guidelines for better recognition in the noisy environment. In [6], researcher proposed the optimization technique that increases the recognition rate of ASR system in noisy environment. In [7], researcher proposed a model that separates the noise from the given speech signal by using binary masking that gains intelligibility for both hearing-impaired listeners and normal-hearing. In [8], researcher proposed enhanced model by using non-negative matrix factorization to reduce the noise in distance speech recognition. In article [9], large database is classified using modified KNN to reduce the noise and increase the accuracy of ASR system. In [10], researcher developed software for education field by using signals and speech processing. This application can be used for noise reduction, speech recognition and coding. MATLAB data acquisition system is used for developing the application, and it runs with real-time speech input. In [11], researcher used Kalman filtering to reduce the noise in given speech signal and compared various feature extraction methodologies to extract the isolated words from noisy environment. Cepstral mean normalization (CMN) is used for feature extraction to reduce the noise in real-world speech recognition application [11]. The recognition rate can be increased up to 2.5% by fixing the 10 ms as a window length and 7.5–10 ms as frame shift [12]. Comparing with conventional microphone, throat microphone produces better accuracy in detecting speech signal [13]. Deep belief network is used for segregating speech signal from unlabeled stationary noise [14]. Similarly in [15], music signal is segregated from noise by using support vector machine (SVM). In [16], the literature survey is carried out in the field of ASR using deep learning model. From the above survey, it is observed that researchers need to focus more on developing ASR system with improved accuracy in noisy environment for low resource language like Tamil. ASR that is developed for Tamil speech recognition commonly uses hidden Markov model (HMM) and Gaussian mixture model. Due to advances in deep learning models, Tamil ASR system can also be able to use it for achieving better performance. Performance of ASR system is affected by various parameters. Noise in speech signal plays an important role in decreasing the performance of ASR. It is a complex task to recognize the LVCS speech signal in noisy environment, since each word that is present in speech signal is highly dependent to one another [17]. Noise and

Hybrid Feature Extraction Technique for Tamil Automatic Speech … Table 1 Parameters considered for recording speech signal

Parameters

Values

Average samples per second

16 kHz

Coding technique

Pulse-code modulation

Recording mode

Mono

Bit rate

16 bits/s

3

presence of speech are estimated by using assessment method in both stationary and non-stationary environment. For better approximation in power spectrum, first 20 frames are considered. In ASR to estimate the human auditory system, MFCC and PLP are used. PLP perform well in comparison with MFCC in large number of parameters [18]. In the proposed work, MFCC and PLP are combined to form a hybrid model that increases the accuracy of ASR system in noisy environment. This paper is organized as follows: Sect. 2 discusses proposed methodology, Sect. 3 explains result and discussion, and Sect. 4 presents conclusion and future work.

2 Proposed Methodology 2.1 Speech Dataset Audacity 3.0.2 is used to record the Tamil speech signal from various native Tamil speakers. Audacity is open-source software, and it is a basic audio editor which can trim, copy, record and manipulate sounds. It can be used to adjust the speed/pitch of audio and add an equalizer to it. For experimental purpose, a total of 10 speakers (5 males, 5 females) of different age groups were selected. In total, 1000 samples of speech signal were collected, and each speaker recorded 100 samples of Tamil sentence. From the entire data, 75% of recorded samples are used for training and 25% of samples are used for testing purpose. Parameters that are considered for recording speech signal are listed in Table 1. The samples that are collected from various speakers are mixed with common environmental noise like fan, wind and car noise. These noises were downloaded from freesound.com online Web portal. Figures 1 and 2 show the speech signal waveform in Audacity and MATLAB.

2.2 Noise Data Common environmental noises such as fan, wind and car noise were downloaded from freesound.com Web portal and mixed with the collected dataset for training

4

S. Girirajan and A. Pandian

Fig. 1 Speech signal in Audacity

Fig. 2 Speech signal in MATLAB

and testing the model. In this Web portal, various noises are available from different sources. After mixing the noise with recorded sample, SNR value is set between 0 and 15 db with equal interval of around 5 db. Speech-to-noise ratio is obtained by using effects in Audacity. Figures 3, 4 and 5 show various noises that we have used in experiments. Separate window is used to obtain the noisy data by adding noise data with collected speech samples. To split the speech, non-speech and noise portion from the speech signal VAD filter are applied. After splitting the speech signal, features are extracted from the given speech signal. Extracted features are used to train the deep neural network (DNN) for classifying the speech signal. Entire experiment is carried out in MATLAB.

Hybrid Feature Extraction Technique for Tamil Automatic Speech …

Fig. 3 Fan noise signal in Audacity

Fig. 4 Wind noise signal in Audacity

Fig. 5 MFCC feature extraction

5

6

S. Girirajan and A. Pandian

2.3 MFCC Feature Extraction MFCC is commonly used feature extraction technique for frequency domain [19]. Mel scale estimates the human speech signal well, since it is in linear form. Figure 5 shows the process involved extracting feature using MFCC. Pre-emphasis. Noise can be reduced in speech signal by combining pre-emphasis and de-emphasis. Pre-emphasis boosts the signal to high frequency in the transmitter and de-emphasis used to retrieve the original signal by demodulating at the receiver. Usually, noise will be more in amplitude at high frequency. So, manipulation needs to be done at excess level for high frequency signal. To overcome the issue, signalto-noise ratio needs to be boosted in high frequency. Similarly, high-pass filters can be selected efficiently by reducing signal–to-noise ratio to low frequency. By doing this, reversing the above process can be carried out at receiver. Due to this, signal-tonoise ratio can be maintained similar in before and after applying pre-emphasis and de-emphasis, respectively. Factor of pre-emphasis is set to 0.97. Transfer function of the filter is shown in Eq. (1). H (z) = 1 − 0.97z −1

(1)

where transfer function of z-domain is denoted as H (z). Framing and Windowing. Initially, audio signal is divided into small frames in the range of 5–50 ms and then windowing is used to avoid the leakage in starting and ending of each frame. Windowing is performed based on Hamming window equation shown in Eq. (2). N denotes total samples, and n replies the particular sample in overall samples.    2π n −1 W (n) = 0.54 − 0.46cos N

(2)

Fast Fourier Transform (FFT). It transforms a non-periodic function from time domain to frequency domain X (K ) as shown in below equation. X (K ) =

N −1 

x(n)W nk

(3)

n=0

Mel Scale Conversion. It relates the perceived frequency of a tone to the actual measured frequency. It scales the frequency in order to match more closely what the human ear can hear.   f (4) M( f ) = 1127ln 1 + 700

Hybrid Feature Extraction Technique for Tamil Automatic Speech …

7

Discrete Cosine Transform. To decorrelate the filter output, log mel spectrum is computed on previous output. Its filter coefficients are grouped with log energy coefficients for final preparation of vector coefficients  MFCC(k) =

 M 2  πk (m − 0.5) X (m)cos M m=1 M

(5)

By using DCT, 12 features are generated along with that one energy feature is also computed. Later using delta method, 13 features are generated based on first-order derivation. In total, 26 coefficients are framed from that 20 were selected.

2.4 Perceptual Linear Perceptron Feature Extraction Psychophysics concept is used to extract the feature using PLP. Initial 3 steps in PLP are similar to MFCC. Insisting of mel scale conversion, band-pass filter is used to approximate the power spectrum. P(ω) = Re(s(ω))2 + lm(s(ω))2

(6)

To improve the mapping with human auditory, audio frequency is converted to bank scale as shown below. ⎡ 

0.5 ⎤ 2 f f ⎦ + +1 (7) f (Bank) = 10ln⎣ 1000 1000 Power spectrum is mapped to extract the feature by applying the LP model. M 1  P(w) =1 M m=1 P  (w)

(8) 

Input is denoted by P(w), and predicted output is denoted by P (w). By using recursive cepstrum transform, 12 features are generated along with that one energy feature is also computed. Later using delta method, 13 features are generated based on first-order derivation. In total, 26 coefficients are framed from that 20 were selected (Fig. 6). Algorithm Step 1: Identifying noise and signal idleness by applying silence indicator to update the duration.

8

S. Girirajan and A. Pandian

Fig. 6 PLP feature extraction

Step 2: To convert time to frequency time, short-time Fourier transform is applied. Frames are normalized by using STFT, and signal with high frequency is normalized to low frequency. For a given time, both temporal and frequency resolutions are provided by STFT. Step 3: Noise variance is reduced by applying high-pass filter (HPF). Step 4: Misrepresentations are eliminated by applying post-processor. Step 5: Reverse transformation to time domain by applying inverse short-time Fourier transform (ISTFT). Step 6: Speech and non-speech signal is clustered and then compared with VAD output based on the threshold value. If the value exceeds the threshold, then it is considered as speech, otherwise non-speech signal. Step 7: To differentiate the speech and non-speech signal, set of feature is computed. Step 8: Collected features are grouped together to classify the speech signal. The experiment is carried out by dividing the collected samples into 3 subsets. By using first set, gradient and bias of the network are computed. Using second set, validation error is computed and last set is error that occurred in test set is computed. Appropriate value is selected by comparing the validation and test error. Initially, the value is fixed as 0.8, 0.16 and 0.16.

3 Result and Discussion Performance of MFCC and VAD is shown in Fig. 7. From that, it is observed by using VAD; there is considerable improvement in accuracy. Recognition is maximum with wind noise mixed samples at 10–15 db. After 10 db, there is decrease in accuracy. From Fig. 8 and Table 2, it is found that car noise has considerable increase in recognition around 13% and fan noise generated relative accuracy.

Hybrid Feature Extraction Technique for Tamil Automatic Speech …

9

Fig. 7 Performance evolution of MFCC with VAD

Fig. 8 a Performance evolution of PLP with VAD and b performance evolution of MFCC + PLP with VAD Table 2 Accuracy with VAD using MFCC Noise

Signal-to-noise ratio (db)

Average

0

5

10

15

Fan

23.4

43.6

67.3

86.2

55.125

Wind

22.6

41.3

67.8

84.7

54.1

Car

12.8

38.5

62.4

83.6

49.325

15

Table 3 Accuracy with VAD using PLP Noise

Signal-to-noise ratio (db)

Average

0

5

10

Fan

24.3

45.7

69.6

88.1

56.925

Wind

23.1

43.6

68.9

87.3

55.725

Car

13.2

39.4

63.8

86.6

50.75

10

S. Girirajan and A. Pandian

Table 4 Accuracy with VAD using PLP + MFCC Noise

Signal-to-noise ratio (db)

Average

0

5

10

15

Fan

37.2

57.3

79.6

92.3

66.675

Wind

35.5

57.5

76.6

94.6

66.34

Car

32.5

53.4

76.7

93.6

63.452

From Fig. 8a and Table 3, it is observed that using VAD accuracy level is around 58.8% and without using VAD accuracy level is 52.3%. Compared to MFCC, the accuracy level of PLP is higher. PLP follows the same pattern as MFCC in first 3 stages; later, the STFT is introduced insisting of FFT and IDFT and LP analysis used to gain better performance for large number of parameters. From Fig. 8b and Table 4, hybrid model combining MFCC and PLP shows better performance compared with individual MFCC or PLP. The average performance in this experiment is 55% by using VAD and 65% compared without using VAD. The proposed hybrid model increases the accuracy rate of recognition around 12% compared MFCC and PLP.

4 Conclusion The hybrid model that combines MFCC and PLP is implemented for Tamil ASR system to recognize the word from noisy speech signal. The proposed model achieved average of 12% increase in accuracy rate when comparing individual MFCC and PLP. VAD differentiates the speech and non-speech signal that increased the accuracy of recognition rate of proposed model. In the future, optimization algorithm can be used to increase the accuracy of ASR system with or without using VAD.

References 1. P.K. Kurzekar, R.R. Desmukh, V.B. Waghmare, P. Shrishrimal, Continuous speech recognition system: a review. Asian J. Comput. Sci. Inform. Technol. (AJCSIT) 4(6), 62–66 (2014) 2. S.B Davis, P. Mermelstein, Comparison of Parametric Representations for Monosyllabic Word Recognition in Continuously Spoken Sentences. Readings in Speech Recognition (Elsevier, 1990), pp. 65–74 3. R.K. Agarwal, M. Dave, Implementing a speech recognition interface for Indian Languages. In Proceedings of the IJCNLP-08 Workshop on NLP for Less Privileged Languages (2008), pp. 105–112 4. S. Keronen, U. Remes, K.J. Palomaki, T. Virtanen, M. Kurimo, Comparison of noise robust methods in large vocabulary speech recognition. In 18th European Signal Processing Conference (EUSIPCO-2010) (2010), pp. 1973–1977

Hybrid Feature Extraction Technique for Tamil Automatic Speech …

11

5. Q. Li, J. Zheng, A. Tsai, Q. Zhou, Robust endpoint detection and energy normalization for real-time speech and speaker recognition. IEEE Trans. Speech. Audio Process. 10(3), 146–157 (2002) 6. A. Nasef, M. Marjanovic-Jakovlijevic, A. Njegus, Optimization of the speaker recognition in noisy environments using a stochastic gradient descent. Intern. Sci. Conf. Inform. Technol. Data. Relat. Res. Sinteza 2017, 369–373 (2017) 7. N. Roman, J. Woodruff, Ideal binary masking in reverberation. In 2012 Proceedings of the 20th European Signal Processing Conference (EUSIPCO) (2012), pp. 629–633 8. J.T. Geiger, F. Weninger, J.F. Gemmeke, M. Wollmer, B. Schuller, G. Rigoll, Memory-enhanced neural networks and NMF for robust ASR. IEEE/ACM Trans. Audio Speech Lang. Process. 22(6), 1037–1046 (2014). https://doi.org/10.1109/TASLP.2014.2318514 9. S.K. Sahu, P. Kumar, A.P. Singh, Modified K-NN algorithm for classification problems with improved accuracy. Intern. J. Inform. Technol. 10, 65–70 (2018). https://doi.org/10.1007/s41 870-017-0058-z 10. L. Bouafif, K. Ouni, A speech tool software for signal processing applications. In 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications (SETIT), pp. 788–791 (2012) 11. M.G. Sumithra, M.S. Ramya, K. Thanuskodi, Speech recognition in noisy environment using different feature extraction techniques. Intern. J. Computat. Intell. Telecommun. Syst. 2(1), 57–62 (2011) 12. M.M. Rahman, S.K. Saha, M.K. Hossain, M.B. Islam, Performance evaluation of CMN for Mel-LPC based speech recognition in different noisy environments. Intern. J. Comput. Appl. 58(10), 6–10 (2012). https://doi.org/10.5120/9316-3548 13. N. Dave, Feature extraction methods LPC PLP and MFCC in speech recognition. Intern. J. Adv. Res. Eng. Technol. 1(6), 1–5 (2013) 14. T. Dekens, W. Verhelst, F. Capman, F. Beaugendre, Improved speech recognition in noisy environments by using a throat microphone for accurate voicing detection. In 18th European Signal Processing Conference (EUSIPCO-2010) (2010), pp. 1978–1982 15. K. Sharma, H.P. Sinha, R.K. Agarwal, Comparative study of speech recognition system using various feature extraction techniques. Intern. J. Inform. Technol. Knowl. Manage. 3(2), 695– 698 (2010) 16. F.J.J. Joseph, Effect of supervised learning methodologies in offline handwritten Thai character recognition. Int. J. Inf. Technol. 12, 57–64 (2020). https://doi.org/10.1007/s41870-019-00366-y 17. A.B. Nassif, I. Shanin, I. Attili, M. Azzeh, K. Shaalan, Speech recognition using deep neural networks: a systematic review. IEEE Access 7, 19143–19165 (2019) 18. T. Gerkmann, R.C. Hendriks, Noise power estimation based on the probability of speech presence. In 2011 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (2011), pp. 145–148 19. J. Psutka, L. Muller, J.V. Psutka, Comparison of MFCC and PLP Parameterizations in the speaker Independent Continuous Speech Recognition Task, Eurospeech 2001, Scandinavia (2001) 20. L. Xie, Z.Q. Liu, A comparative study of audio features for audio to visual cob version in MPEG-4 compliant facial animation. In Proceedings of ICMLC, Dalian, 13–16 Aug 2006

Implementation and Image Transformation for Ground Penetration Image Radar System Dharamvir and M. S. Shashidhara

1 Introduction The performance application of transformational activity that is attached with high resolution data development techniques for individual and data segmentation process. SAR simulates a significantly larger antenna aperture through signal processing instead of deploying a very large physical antenna by generating an “aperture” of large size and good image resolution. Normally used algorithms are Finitedifference migration, Kirchoff Migration, Frequency wave-number Migration [1] and Reverse time migration. Kirchoff Migration algorithm, the main focus of this paper is explained in Sect. 2. The generated data has been processed by this algorithm and the simulated results in Matlab are shown at the end. An architecture for the FPGA Implementation of the algorithm is proposed in Sect. 3.

1.1 Kirchoff Migration (KM) Algorithm In seismic processing, migration is a technique for an accurate picture of underground layers. It entails repositioning return signals geometrically reval an occurrence (layer boundary or other structure) where the seismic wave hits it rather than where it is picked up. Pre and poststack migration are two of the most popular migration methods. Post stack migration refers to moving data after it has been stacked, while prestack migration refers to moving data before it has been stacked. When Dharamvir (B) · M. S. Shashidhara Research Scholar VTU, Dept. of MCA, The Oxford College of Engineering, Bengaluru 560068, India e-mail: [email protected] M. S. Shashidhara e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_3

13

14

Dharamvir and M. S. Shashidhara

the subsurface structures are clear. The performance of assigned task is managed to transform the various Applications where it convert the basic values of both the parameters. However, in areas with complex geology, post-stack migration is ineffective in achieving the available data to perform its target goal with the help of task data knowledge methods. The transformation of Ground prediction having different data methods and performing applications to handle the system. Similarly we can have multiple levels of tools to make use of Radar data and migration system with same angle and its identifications. In general we are using image data application for multicast changes and behavioral data transformation method. The main task is to control the performance activity and data identification. During the trusting process of data performing model we can choose multiple mode of detection with available resource and identification process. Frequency web notation also changes the data with simulated value and task performance repositioning system with its basic identical approach of system applications [2–5]. The support of assigned data is having multiple data simulation based on assigned task and implementation. It is based on diffraction summation principle. The basic idea is to try to calculate the wave front at a certain time and position, when it is known at another time and position. The known data is recorded at all times, at depth z = 0 and, what we want to calculate is the originating wave fronts at time t = 0 at all depths z. Here we try to use data at the surface to reconstruct the past history of wave-fields in the subsurface. So we will be using backwards Green’s function to continue the measured fields at the surface in the downward direction towards the subsurface scatter sources. Kirchoff equation derived from Greens Theorem says that the field ψ(r, t) can be predicted as   ψ(r, t) = G(r, t) − G r  , t 

(1)

We are having different Data methods for Air and solid calculation in basic value as 0, we get, the final KM equation in 2D (y and z axis chosen) as.

2 Implementation Design In given data segmentation we can choose define range of process that is having multiple transformation and Kirchoff Migration in different task distribution methods.     ψ(r, t) = G(r, t) − G r  , t  − G r, t|r  , t 

(2)

  r − r   = √(yt − yi) + z j

(3)

Implementation and Image Transformation for Ground …

Ci =

  √   i/i + j q − q 

15

(4)

where C i is having amplitude described in Radar Range equation. This can be identified with regular expression to represent the system and its Migration. So the basic steps to do KM are: 1. 2. 3. 4. 5. 6.

Take time derivative of the GPR data. Select the Point to be migrated (yi, zj). Find Range for each point selected by using the range equation Find cos(θ) = depth/range for each zj. Interpolate to find nearest data set in time domain. Final Summation over each antenna position.

B Scan data are having different performance toolkit to adopt the situation and having method with multiple adaptation system. The data is then displayed using the Matlab command ‘imagesc’. In the B-scan GPR set-up, a single point scattered forms a hyperbolic curve in space-time radar image Three target locations at (2, 1.6), (6, 1.8), (10, 1.7) is chosen. Kirchoff Migration is implemented in Matlab and the simulation results are shown in Sect. 4 (Fig. 1). Fig. 1 B-scan data collection scheme with SFCW Radar

16

Dharamvir and M. S. Shashidhara

Fig. 2 Proposed architecture for FPGA implementation of KM algorithm

Here we have used this method. Then follow steps 2–6 as in Sect. 2. Hardware architecture for the FPGA Implementation of Kirchoff Migration is also proposed here. The basic block diagram is shown in Fig. 2. The Memory controller state diagram is as shown in Fig. 3. At state s0 (idle state), when read (rd) signal is given, then the memory controller should go to state s1, where assertion of read signal will occur. In the memory GPR Data points, and the antenna step positions are stored. After acquiring the data required, we perform the range calculation, cosine calculation, interpolation and final summation. Range = sqrt((y point-antenna position)2 – z point2 ) and Cos(θ) = Depth/Range. Interpolation is required since the scalar wave field is discrete in time. A simple 2 point linear interpolation for points (xa, ya) and (xb, yb) is y = ya + (yb − ya)*(x − xa)/(xb − xa) for each value of x. (Fig. 3). We are trying to implement this KM algorithm on to FPGA. We need to store the each antenna position, GPR data matrix and the stepped frequency matrix in a memory. Size of the memory depends on number of antenna positions and GPR Data matrix size (Fig. 4).

3 Simulation Result and Analysis The domain yields with different performance data can have matching transformation that is shown in (Fig. 5a). The hyperbolic defocusing behavior is visible, initially its having total control data as predicted with different memory and segmentation. We created a new GPR image using the above algorithm, as shown in Fig. 5b. As compared to the image in different path structure support system and its basic transformational approach.

Implementation and Image Transformation for Ground …

17

Fig. 3 Memory control operations

Fig. 4 Two point linear interpolator unit

4 Conclusion In this research work, we implemented Kirchoff Migration algorithm used in SAR image focusing for B-scan GPR images in Matlab. Similarly, we can change the structural behavior for assigned activity can be demonstrate in all the location based automation services. Similarly, we can predict level of structure to transform with system and its performance. For simplicity, we have used a single transreceive antenna facility. As a future enhancement, it can be extended to an array of antennas which will give better resolution. The actual implementation of the proposed architecture for real GPR data can also be done as a future enhancement.

18

Dharamvir and M. S. Shashidhara

Fig. 5 a Generated GPR data. b Migrated image

References 1. I.A. Ruthenberg, Curve Fitting and Migration of GPR Data for the Detection. (Department of Computer Science and Electrical Engineering, University of Queensland, IJMRT Journal, 2018) pp. 119–126

Implementation and Image Transformation for Ground …

19

2. R. Gupta et al., Automation and aviation radar techniques. IEEE Trans. 29(11), 219–231 (2018) 3. A. Gunawardena, D. Longstaff, Wave equation formulation of synthetic aperture radar (SAR) algorithms in the time-space domain. IEEE Trans. GeoSci. Rem. Sens. 6 (2019) 4. D.S. Jones, The Theory of Electromagnetism (Pergamon Press, 2016) 5. E. Yigit, S. Demirci, C. Ozdemir, Ground penetrating radar image focusing using frequency-wave number based synthetic aperture radar technique, In IEEE paper, Published in, International Conference on Electromagnetics in Advanced Applications, ICEAA2017, pp. 344–347

Effect of Geospatial Weather Features on COVID-19 Spread in Maharashtra State Using Machine Learning Ranu Sewada and Hemlata Goyal

1 Introduction All over, the world is facing an outbreak with high transmissibility, due to a virus belonging to the family of coronaviruses (CoV) [1]. The first time it was recognized on Jan 7, 2020 as the “COVID-19 virus” which was transmitted all over countries in the next few weeks. The first case in India was found in the state of Kerala when a student returned from Wuhan of China which is the origin of this pandemic [2]. India has now been hit by the second wave (Feb 2021 to Jun 2021) of the pandemic with the daily rising cases with over 4L per day in the month of May 2021, while the record of the active cases crossed the 32L cases pole. The continuing outbreak has spread to all over the states, but the State of Maharashtra was the most affected as per the comparative assessment of the confirmed cases. Maharashtra has recorded one-eighth of India’s total cases and one-fourth of deaths. In a practice to understand the severity of these epidemics, studies have investigated various aspects that could influence the transmission of coronaviruses [3]. Early viruses tagged along the coronaviruses family or SARS were partly associated with the environment parameters [4]. Furthermore, Park et al. have identified that infection increases with high relative humidity and low temperature, which shows associations of weather parameters and influenza transmission [5].

R. Sewada Department of Computer Applications, Manipal University Jaipur, Jaipur, Rajasthan, India H. Goyal (B) Department of Computer and Communication Engineering, Manipal University Jaipur, Jaipur, Rajasthan, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_4

21

22

R. Sewada and H. Goyal

Fig. 1 Study area of Maharashtra in India

As per various studies, different weather conditions playing an important function at the back of COVID-19, it is useful to determine the association of the weather features like temperature, relative humidity, and precipitation and COVID-19 transmission. Assessment of the new confirmed cases concerning the first wave period and second wave period is the major aim of this study which will help policymakers and citizens. Daily reports of relative humidity, temperature, and precipitation were considered in the study period. The state of Maharashtra in India is selected for the study area (Fig. 1) as temperature, relative humidity, and precipitation are dominant factors in this city, and it was considered in the red zone in the critical period of COVID-19 transmission. Training and testing are done with the machine learning regression algorithms of ordinary least square regression (OLS), K-nearest neighbor (KNN), linear regression (LR), and decision tree regression (DT) models on the Maharashtra dataset of relative humidity, temperature, and precipitation to determine the best-fitted model to assess the future trend of confirmed cases, and this research could also be used in future for the assessment of new cases in other states of India concerning relative humidity, temperature, and precipitation dataset.

2 Literature Review Weather factors can cause the severity of COVID-19 transmission as in the cases of SARS coronavirus, Ebola, and Influenza, weather factors have played a major role in the transmission and have shown significant association with them [6–11]. As per various studies, the environmental retention rate of the viruses relies on weather

Effect of Geospatial Weather Features on COVID-19 …

23

parameters of temperature, relative humidity, and precipitation [12–14]. Therefore, new infected COVID-19 cases can be predicted based on these weather parameters. The first study over the weather parameters (i.e., absolute humidity and temperature) and COVID-19 spread claimed that the pandemic can be slowed down because of warm weather in the worst-affected and highly populated countries, and can classify high-risk geographic regions of countries [15]. A recent study conducted over the Jakarta of Indonesia affected by COVID-19 reported that weather parameters, especially the average temperature (°C) is correlated with the transmission and might play a role in containing the virus devolution [16]. Mouse hepatitis virus (MHV) and transmissible gastroenteritis virus (TGEV) from the Coronaviridae family, also demonstrated survival in correlation for low temperature and relative humidity [17]. Tongs et al. investigated the impact of relative humidity and precipitation on the transmission rate of Ross River virus (RRv) in major cities of Queensland and observed that rainfall and relative humidity have a remarkable association in the RRv virus dissemination [18]. In another study, Bukhari et al. [19] and Prata et al. [20] observed the impact of the humidity and temperature over the various cold and warm-humid countries, and the result specified that a rich number of cases recorded in subtropical and humid tropical countries. New affected cases can be predicted by implementing the machine learning techniques like the regression technique which helps to predict with better accuracy. This study considered the weather parameters including temperature, relative humidity, and precipitation as independent parameters for the regression models and daily new confirmed COVID-19 cases are considered as target entities.

3 Dataset and Methodology 3.1 Dataset Daily new confirmed COVID-19 cases from March 2020 to June 2021 are collected for the state of Maharashtra from the Ministry of Health and Family Welfare (https://www.mohfw.gov.in) [2] and Indian Council of Medical Research Centre (ICMR) [21]. “Modern-Era Retrospective analysis for Research Applications version 2 (MERRA-2)” for India was used for weather data of temperature (T) in Kelvin, relative humidity (RH) in percent, and precipitation (P) in mm [22]. A dataset view is shown in Table 1.

24

R. Sewada and H. Goyal

Table 1 Dataset view of geospatial weather features and daily new confirmed COVID-19 cases Date

Temperature (K)

Relative humidity (%)

Precipitation (mm)

COVID-19 new confirmed cases

5/1/2021

307.23

26.02

0.78

62,919

5/2/2021

303.38

35.01

8.77

63,282

5/3/2021

304.46

36.82

0.28

56,647

5/4/2021

304.89

38.29

0.02

48,621

5/5/2021

305.12

36.63

0.09

51,880

3.2 Methodology This study determines the impact of geospatial weather features on daily confirmed cases, for which weather data of temperature, relative humidity, and precipitation are considered as explanatory features, and newly confirmed cases are considered as a target feature for the period of (March 2020 to June 21) of Maharashtra state. A detailed methodology is given in Fig. 2. Data is standardized and normalized to combine all features on a single scale and preprocessed for unavailable values by averaging nearby cells. The feature extraction is done based on statistical correlation in the dataset. To find the linear relationship between independent and target features, regression techniques are well suited, so regression techniques are considered here. Four machine learning regression algorithms, namely LR, OLS, KNN, and DT are considered for this research, and the discussion is given below. Linear Regression: It is a supervised learning technique that performs the regression task to investigate the relationship between one or more independent variables and dependent or target variables and also forecast the target values based on the independent parameters. The mathematical representation is given in Eq. 1. Y = a0 + a1 X i + ε

(1)

Here, Y is target variable (daily new confirmed COVID-19 cases), a0 is intercept of line, X i is the vector of predictor variables (temperature, precipitation, and relative humidity), linear regression coefficients to a1 , and ε as the random error term. Ordinary Least Square (OLS): OLS model explores a relationship between one or more explanatory features and a continuous or a target feature that minimize the sum of squared errors, where the error is defined by the gap between the actual and predicted value of the target variables. The mathematical representation is given in Eq. (2). Yi = β0 + βi ∗ X i + ε

(2)

Effect of Geospatial Weather Features on COVID-19 …

25

Daily

Fig. 2 Methodology to assess the effect of geospatial weather features on COVID-19 using machine learning

Data Featu Traini Select where “Y i is the new confirmed COVID-19 cases, β 0 is intercept/constant, X i is the vector of selected temperature, precipitation and relative humidity features, β i is regression coefficients’ vector, and the random error term represented by ε.” K-Nearest Neighbors (KNN) Regression: KNN regression calculates the average of numerical target values by calculating the distance of test observations (Y i ) from all training observations (x i ) and mainly uses three distance formulas as given in Eq. (3–5). Euclidean

 k i=1

 Manhattan

Minkowski

(xi − yi )2

(3)

(xi − yi )2

(4)

k i=1

k i=1

(|xi − yi |)q

 q1

(5)

26

R. Sewada and H. Goyal

These distance functions are applicable for continuous values only. Decision Tree Regression: This uses a set of binary values to build the regression model and classification algorithms. Datasets are divided into smaller subsets by which an associated decision tree gets subtly developed. It creates a tree with decision nodes and leaf nodes as a final result. The decision and root node have one or more branches that represent values for the attributes tested and handle both numerical and categorical values. The data is split into training and testing as 70 and 30% ratios, respectively, and considered models are trained over the training data. Selection of the model is done on the basis of the R2 score, MSE, and RSME performance measure.

4 Result and Discussion Maharashtra state is taken to assess the effect of geospatial weather features on COVID-19 using machine learning since it is the most affected and populated state in India. We have fitted four machine learning regression algorithms to predict the daily new confirmed COVID-19 cases based on the weather parameters. The correlation among the studied factors shows that temperature has positive significant correlations (0.72) with new cases and relative humidity (−0.62) and precipitation (−0.55) have negative significant correlation as shown in Fig. 3.

Fig. 3 Correlation matrix among all studied factors

Effect of Geospatial Weather Features on COVID-19 …

27

Table 2 Performance measures of the machine learning regression models COVID 19 1.0 OLS

LR

COVID 19 2.0 KNN

DT

OLS

LR

KNN

DT

R2

0.3201

0.7252

0.9526

0.9621

0.6218

0.8418

0.9847

0.9856

MSE

0.0018

5.1250

1.0189

2.3325

3.2180

3.5633

1.0789

1.9502

RMSE

0.0122

0.0071

0.0071

0.0048

0.0020

0.0156

0.0010

0.0013

In this research, we recorded daily temperature, relative humidity, and precipitation with the daily new confirmed COVID-19 cases of the state of Maharashtra for both the waves of COVID19 (COVID-19 1.0 and COVID-19 2.0) for the period of March 2020 to June 2021 as depicted in Table 1. In this experiment, we have fitted four regression models, namely OLS, LR, KNN, and DT. For each of these regression models, exploratory variables are temperature, relative humidity, and precipitation, and the target variable is daily new confirmed COVID-19 cases. For both of the periods of COVID-19 1.0 and COVID-19 2.0, K-nearest neighbour and decision tree machine learning algorithm are well fitted significant models. For COVID-19 1.0, K-nearest neighbour measures R2 Score, MSE, RMSE and error rate is 0.9526, 1.0189, 0.0071 and 6.23% respectively and for decision tree R2 Score, MSE, RMSE and error rate are 0.9621, 2.3325, 0.0048 and 5.03% respectively. R2 score, MSE, RMSE, and error rate are 0.9621, 2.3325, 0.0048, and 5.03%, respectively. While COVID-19 2.0, K-nearest neighbor performance measures R2 score, MSE, RMSE, and error rate as 0.9847, 1.0789, 0.0010, and 5.01%, respectively, and for decision tree R2 score, MSE, RMSE, and error rate are 0.9856, 1.9502, 0.0013, and 4.87%, respectively. Ordinary least square and linear regression show low significance as shown in Table 2. Figure 4 represents the prediction graph for the period of COVID-19 1.0 by considering applied regression models, and Fig. 5 represents the prediction graph for the period of COVID-19 2.0 for all fitted regression models.

5 Conclusion This research provides evidence of interconnection in weather features and daily new confirmed cases. The high rate of COVID-19 (COVID-19 1.0 and COVID-19 2.0) transmission was recorded with increment in temperature and decrement in relative humidity and precipitation. This demonstrates that temperature and relative humidity played a vital role in forecasting daily new confirmed COVID-19 cases as relative humidity and temperature have negative and positive correlation, respectively, with COVID-19 affected cases. This research could be used for displaying and anticipating every day confirmed cases at the global community level in other geographical regions also.

28

R. Sewada and H. Goyal

Fig. 4 COVID-19 1.0 predictions through machine learning regression models

6 Future Scope This work estimates the COVID-19 cases based on the three weather indices only; other weather parameters like pressure, wind speed, wind direction, and solar radiation may have significant associations with the confirmed cases. This study can be extended more by analyzing the above-said parameter by fitting the model based on the machine learning classification techniques besides the regression techniques.

Effect of Geospatial Weather Features on COVID-19 …

29

Fig. 5 COVID-19 2.0 predictions through machine learning regression models

References 1. V.J. Munster, M. Koopmans, N. van Doremalen, D. van Riel, E. de Wit, A novel coronavirus emerging in China—key questions for impact assessment. N. Engl. J. Med. 382(8), 692–694 (2020). https://doi.org/10.1056/NEJMp2000929 2. Ministry of Health and Family Welfare, Government of India (GOI) (2020). www.mohfw.gov. in 3. K.H. Chan, J.S. Peiris, S.Y. Lam, L.L.M. Poon, K.Y. Yuen, W.H. Seto, The effects of temperature and relative humidity on the viability of the SARS coronavirus. Adv. Virol. (2011). https://doi. org/10.1155/2011/734690 4. K. Lin, D.Y.T. Fong, B. Zhu, J. Karlberg, Environmental factors on the SARS epidemic: air temperature, passage of time and multiplicative effect of hospital infection. Epidemiol. Infect. 134(2), 223–230 (2006). https://doi.org/10.1017/S0950268805005054 5. J.E. Park, Effects of temperature, humidity, and diurnal temperature range on influenza incidence in a temperate region. Influenza Other Respir. Viruses 14(1), 11–18 (2020). https://doi. org/10.1111/irv.12682 6. J.H. Hemmes, K. Winkler, S.M. Kool, Virus survival as a seasonal factor in influenza and poliomyelitis. Nature 188(4748), 430–431 (1960) 7. N. Pica, N.M. Bouvier, Environmental factors affecting the transmission of respiratory viruses. Curr. Opin. Virol. 2(1), 90–95 (2012) 8. P.Q. Thai, M. Choisy, T.N. Duong, V.D. Thiem, N.T. Yen, N.T. Hien et al., Seasonality of absolute humidity explains seasonality of influenza-like illness in Vietnam. Epidemics 13, 65–73 (2015)

30

R. Sewada and H. Goyal

9. M. Moriyama, T. Ichinohe, High ambient temperature dampens adaptive immune responses to influenza A virus infection. Proc. Natl. Acad. Sci. U. S. A. 116(8), 3118–3125 (2019) 10. C. Yip, W.L. Chang, K.H. Yeung, I.T. Yu, Possible weather influence on the severe acute respiratory syndrome (SARS) community outbreak at Amoy Gardens. Hong Kong J. Environ. Health 70(3), 39–46 (2007) 11. L.M. Casanova, S. Jeon, W.A. Rutala, D.J. Weber, M.D. Sobsey, Effects of air temperature and relative humidity on coronavirus survival on surfaces. Appl. Environ. Microbiol. 76(9), 2712–2717 (2010) 12. N.N. Harmooshi, K. Shirbandi, F. Rahim, Environmental concern regarding the effect of humidity and temperature on 2019-nCoV survival: fact or fiction. Environ. Sci. Pollut. Res. 1–10 (2020) 13. K. Chan, J. Peiris, S. Lam, L. Poon, K. Yuen, W. Seto, The effects of temperature and relative humidity on the viability of the SARS coronavirus. Adv. Virol. (2011). https://doi.org/10.1155/ 2011/734690 14. W.J. Landesman, B.F. Allan, R.B. Langerhans, T.M. Knight, J.M. Chase, Inter-annual associations between precipitation and human incidence of West Nile virus in the United States. Vector-Borne Zoonotic Dis. 7(3), 337–343 (2007) 15. S. Gupta, G.S. Raghuwanshi, A. Chanda, Effect of weather on COVID-19 spread in the US: a prediction model for India in 2020. Sci. Total Environ. 728, 138860 (2020) 16. R. Tosepu, J. Gunawan, D.S. Effendy, H. Lestari, H. Bahar, P. Asfian, Correlation between weather and COVID-19 pandemic in Jakarta, Indonesia. Sci. Total Environ. 138436 (2020) 17. L.M. Casanova, S. Jeon, W.A. Rutala, D.J. Weber, M.D. Sobsey, Effects of air temperature and relative humidity on coronavirus survival on surfaces. Appl. Environ. Microbiol. (2010). https://doi.org/10.1128/AEM.02291-09 18. S. Tong, P. Bi, K. Donald, A.J. McMichael, Climate variability and Ross River virus transmission. J. Epidemiol. Commun. Health 56(8), 617–621 (2002) 19. Q. Bukhari, Y. Jameel, Will coronavirus pandemic diminish by summer? SSRN Electron. J. (2020) https://doi.org/10.2139/ssrn.3556998 20. D.N. Prata, W. Rodrigues, P.H. Bermejo, Temperature significantly changes COVID-19 transmission in (sub) tropical cities of Brazil. Sci. Total Environ. 729, 138862 (2020) 21. Indian Council of Medical Research, New Delhi, Government of India (GOI) (2020). https:// www.icmr.gov.in/. Accessed 31 Aug 2020 22. Global Modeling and Assimilation Office (GMAO), MERRA-2 tavg1_2d_slv_Nx: 2d,1Hourly,Time-Averaged,Single-Level,Assimilation,Single-Level Diagnostics V5.12.4, Greenbelt, MD, USA, Goddard Earth Sciences Data and Information Services Center (GES DISC) (2015). https://doi.org/10.5067/VJAFPLI1CSIV

Adiabatic Logic Code Converter Design at Different Sub-micron Technologies J. V. N. Saketh Ram, Vaibhav Kumar Jain, and Abhijit Asati

1 Introduction Power consumption is a main concern in today’s VLSI circuits. To overcome this issue, designers have explored various approaches such as sub-threshold circuit design, dynamic voltage and frequency scaling (DVFS), adiabatic logic circuit design, clock gating, power gating, and asynchronous logic circuit design [1]. Major applications of low-power design are Internet of things (IoT) and also useful in applications where replacing of batteries is not possible such as in medical devices like pacemakers which operate at lower frequencies but battery life is very important [2]. The term adiabatic in thermodynamics means a thermal process, wherein energy exchange does not happen with the surrounding environment; hence, the energy loss is theoretically zero; also, there is no direct path from power rails, i.e., VDD and GND, and charge recycling takes place at different phases [3, 4]. In this paper, we have designed the code converters, namely gray to binary (GtoB), binary to gray (BtoG), and BCD to excess-3 (BCX3) [5, 6] using existing adiabatic logic families, viz. ECRL [7], IPGL [8], and 2N_2N2P [9, 10] at 32 and 22 nm technology nodes as well as at 16 nm using PTM MOS models. We compared power and performance of code converters mentioned above for ECRL, IPGL, 2N_2N2P logic styles. Generally in adiabatic circuits, multiple phase-shifted power signals are used in other published works [11, 12]; here, we proposed to use a single power signal with four phases for operating the entire cascaded adiabatic circuit by adjusting the time of hold phase (H) of the power signal as explained in upcoming sections. Section 2 provides a brief overview about conventional switching and adiabatic switching. Section 3 explains the design of adiabatic logic buffers/inverters and related work. Section 4 explains the circuits of three code converters designed using

J. V. N. Saketh Ram · V. K. Jain · A. Asati (B) BITS Pilani (Pilani Campus), Pilani, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_5

31

32

J. V. N. Saketh Ram et al.

adiabatic logic. Simulation results of the analysis are stated in Sects. 5 and 6 concludes the paper.

2 Switching in Static CMOS Circuits and Switching in Adiabatic Logic Circuits Conventional switching [13] can be understood by considering a static CMOS inverter circuit as shown in Fig. 1a. It can be observed that during one cycle of charging and discharging through the load capacitor, an energy loss of C L VD2 D is bound to happen for static CMOS circuits. When the load capacitance is charging though the power supply V DD as shown in Fig. 1b, a charge Q (= C L V DD ) will flow through the PMOS, which is in pull-up path, from power supply. Therefore, the supplied energy is as shown in Eq. (1). Here, only 50% of the energy is stored in the load capacitance as shown in Eq. (2). E supply = C L VD2 D

(1)

1 C L VD2 D 2

(2)

C.VD D dv(t) = dt T

(3)

E stored = i c (t) = C. T E Adiabatic = 2R.

i 2 (t).dt = 0

Fig. 1 Static CMOS inverter

2RC 2RC 2 VD2 D .C VD2 D .T = 2 T T

(4)

Adiabatic Logic Code Converter Design …

33

Fig. 2 An adiabatic circuit showing energy loss

Fig. 3 Pictorial representation of the power applied

The other half ( 12 C L VD2 D ) will be dissipated as heat to the surrounding environment through the resistive path provided by PMOS. Similarly when we consider the case of discharging of load capacitance as shown in Fig. 1c, when inverter is set to logic 0, the other 50% energy that is stored in the CL (i.e. 21 C L VD2 D ) is delivered to the ground through the NMOS transistor that is present in the pull-down network and therefore does not support the recovery of energy. For understanding adiabatic switching [3], we consider a RC network as shown in Fig. 2 and assume that we have a ramp-type voltage source as shown in Fig. 3 as explained later. The ‘R’ indicates the resistance of the PMOS transistor present in pull-up path, and the capacitance C is load capacitance connected. If the voltage v(t) is linearly increased from 0 to V DD in time T, Eq. (3) shows the current flowing through the capacitor. The energy required for charging the capacitor will be equal to the integration of ic (t) and v(t) product over in 0 to T interval as given by Eq. (4), where term ‘2’ indicates that during discharging same amount of energy is lost as that was lost while charging. If ‘T ’ is increased such that it is much greater than 2RC, then a significant amount of energy dissipation can be saved [3]. The ramp power signal as shown in Fig. 3 is a power signal having four phases: (i) evaluate (E), (ii) hold (H), (iii) recover (R), and (iv) wait (W) and has trapezoidal shape.

3 Adiabatic Logic Inverter/Buffer Circuits and Related Works 3.1 Efficient Charge Recovery Logic (ECRL) The transistor-level implementation of the ECRL buffer/inverter [7] is shown in Fig. 4a. Initially in W phase, in equals LOW logic level and in equals HIGH, hence ‘N2’ is ON and ‘N1’ will be OFF. Hence, the output node out will be connected to ground. Since the gate input of PMOS P1 is connected to, out follows the power

34

J. V. N. Saketh Ram et al.

Fig. 4 a ECRL inverter/buffer. b 2N_2N2P inverter/buffer. c IPGL inverter/buffer

signal. As the power signal enters the ‘E’ phase and P1 transistor is in the ON state allowing the node out to follow the supply voltage when the voltage level in the ‘E’ phase reaches to the threshold voltage. During evaluation phase, since P1 and P2 transistors are cross-coupled, if P1 is ON then P2 will be OFF and vice versa. Since P2 is OFF, the node out remains at GND potential. At nodes out and out, we connect a C1 capacitor and a C2 capacitor, respectively. Now as the power signal reaches to a value VDD, the C1 is charged to peak value. After that, the signal enters the ‘H’ phase where it is kept at a constant logic level HIGH. If we assume that change in input does not occur during this phase, then we can connect output of this gate to input of another gate which is present in the stage succeeding present stage. As the signal enters ‘R’ phase of the power signal where the voltage is decreased linearly down to 0 from VDD , the charge stored in C1 and C2 is returned to the power signal. When the voltage across load capacitance C1 reaches the threshold voltage, the PMOS will be cut off. Both the PMOS transistors, i.e., P1 and P2, will remain OFF during the next phase of the signal which is the wait phase (W); output remains nearly the V th of PMOS and, i.e., at LOW logic level.

3.2 2N_2N2P Adiabatic Logic (2N_2N2P) 2N_2N2P logic style [9] is another variation of efficient charge recovery logic (ECRL). In addition to the 2 NMOS transistors which are present in ECRL, it also consists of two new cross-coupled NMOS transistors added parallel as shown in Fig. 4b for inverter/buffer. By adding the cross-coupled inverters, we can significantly reduce the coupling effects; hence, it eliminates the floating nodes during recovery phase with little increase in the power consumption.

Adiabatic Logic Code Converter Design …

35

3.3 Improved Pass-Gate Adiabatic Logic (IPGL) The IPGL gate [8] is based on the 2N_2P cross-coupled inverters as shown in Fig. 4c for inverter/buffer. The gate has two paths from the power signal input to the output: one while charging and other while recovery phase. The F logic block and the F inverted logic block come parallel with the cross-coupled PMOS pair and NMOS pair. Logic block is complementary; i.e., if the F logic block is ON, i.e., conducting, then the F is disconnected and will be OFF. In the ‘E’ phase, the power signal increases linearly from 0 to V DD and the ‘out’ node follows it via the parallel combination of F and PMOS and remains valid when power signal reaches V DD . The pull-down and pull-up paths are equalized to equalize the propagation delay.

4 Logic Circuit of Code Converters For all code converter circuit designs, we have used a single power signal, while published literature utilizes a multiple phase-shifted power signal to drive all the cascaded logic stages [14, 15]. (A)

(B)

(C)

BtoG: The conversion of binary code to gray code can take place by the use of XOR gate shown in Fig. 5a as expressed in Eq. (5). The gate-level connections of the BtoG converter implemented using adiabatic logic (ECRL, IPGL, 2N_2N2P) are shown in Fig. 5b. The gate-level schematic of BtoG is shown in Fig. 5b. GtoB: The gate-level schematic for GtoB is shown in Fig. 5c as expressed in Eq. (6). The circuit’s delay is therefore the delay of a buffer and two cascaded XOR gates. BCX3: Figure 5d represents the gate-level schematic of BCD to excess-3 as expressed in Eq. (7). The delay of the circuit is equal to the delay in computing output W, which involves two OR gates and an AND gate. G2 = B2 G1 = B2 ⊕ B1 G0 = B1 ⊕ B0

(5)

B2 = G2 B1 = B2 ⊕ G1 B0 = B1 ⊕ G0 W = A + B.(C + D) X = (C + D) ⊕ B

(6)

36

J. V. N. Saketh Ram et al.

Fig. 5 a ECRL XOR/XNOR gate. b Gate-level schematic of BtoG code converter. c Gate-level schematic of GtoB code converter. d Gate-level schematic of BCX3 converter

Y =CD Z = D¯

(7)

5 Simulation and Results All simulations are performed at 25 and 50 MHz frequencies. Power supply for BtoG converter, GtoB converter, and BCX3 converter is 1 V for 32 nm, 0.9 V for 22 nm, and 0.7 V for 16 nm technology node as per the PTM data. All simulations are performed using a load capacitance of 10 fF with a ramp duration of 5 ns for E and R phase. The transient response of BtoG converter, GtoB converter, BCX3 converter is shown in Fig. 6. Tables 1, 2, 3, and 4 consist of the power, delay, and power–delay products of all the converters for frequency 25 and 50 MHz. The load capacitance CL was fixed at 10 fF for all the simulations. For BCX3, GtoB, and BtoG power, results are drawn for ECRL, IPGL, 2N_2N2P as well as the static logic design style as shown in Fig. 7a.

Adiabatic Logic Code Converter Design …

37

Fig. 6 Transient response of a BtoG converter [@0 ns: B(2)B(1)B(0) → G(2)G(1)G(0), e.g., 010 → 011]. b GtoB converter [@200 ns: G(2)G(1)G(0) → B(2)B(1)B(0), e.g., 100 → 111]. c BCX3 converter [@200 ns: abcd → wxyz, e.g., 0011 → 0110]

38

J. V. N. Saketh Ram et al.

Table 1 Power, delay, and PDP for ECRL design style with C L = 10 fF for 25 and 50 MHz Technology node (nm)

Code converter

Frequency = 25 MHz PD (µW)

T d (ns)

PDP (× 10–15 W-s)

PD (µW)

T d (ns)

PDP (× 10–15 W-s)

32

BCX3

0.55

0.54

0.301

0.57

0.53

0.310

GtoB

0.34

0.81

0.282

0.42

0.72

0.308

22

16

Frequency = 50 MHz

BtoG

0.27

0.20

0.056

0.31

0.19

0.061

BCX3

0.43

0.60

0.264

0.49

0.53

0.264

GtoB

0.28

1.35

0.387

0.37

1.23

0.465

BtoG

0.21

0.36

0.078

0.25

0.31

0.081

BCX3

0.25

1.40

0.360

0.38

1.38

0.530

GtoB

0.23

3.18

0.754

0.40

3.09

1.260

BtoG

0.14

1.13

0.163

0.22

1.08

0.249

Table 2 Power, delay, and PDP for IPGL design style with C L = 10 fF for 25 and 50 MHz Technology node (nm)

Code converter

Frequency = 25 MHz PD (µW)

T d (ns)

PDP (× 10–15 W-s)

PD (µW)

T d (ns)

PDP (× 10–15 W-s)

32

BCX3

0.74

0.64

0.478

0.68

0.49

0.338

GtoB

0.47

0.91

0.436

0.50

0.80

0.404

22

16

Frequency = 50 MHz

BtoG

0.37

0.19

0.073

0.34

0.42

0.148

BCX3

1.06

1.21

1.298

1.31

1.02

1.343

GtoB

0.62

3.05

1.916

0.86

2.00

1.737

BtoG

0.52

0.85

0.450

0.67

0.75

0.503

BCX3

0.23

1.38

0.328

0.27

1.35

0.374

GtoB

0.20

2.50

0.523

0.31

2.43

0.755

BtoG

0.12

0.91

0.113

0.14

0.90

0.133

For all 3 adiabatic logic families ECRL, IPGL, and 2N_2N2P, with increase in frequency the power increases while delay reduces. It is observed that the ECRL logic consumes minimum power for 32 and 22 nm technology while in 16 nm technology IPGL logic consumes least power. 2N_2N2P adiabatic logic shows maximum power consumption, but it is still smaller as compared with static CMOS logic. For BCX3, GtoB, and BtoG power–delay product (PDP), results are drawn for ECRL, IPGL, and 2N_2N2P as well as the static logic design style as shown in Fig. 7b. ECRL logic of the PDP trend for BCX3, GtoB, and BtoG is irregular but overall increasing as the technology node is shrinking from 32 to 16 nm. For IPGL and 2N_2N2P logic, the PDP is increasing from 32 to 22 nm technology node and PDP is decreasing as technology node is shrinking from 22 to 16 nm. For the 32 and 22 nm technology nodes, ECRL logic is offering least PDP, while for 16 nm technology

Adiabatic Logic Code Converter Design …

39

Table 3 Power, delay, and PDP for 2N-2N2P design style with C L = 10 fF for 25 and 50 MHz Technology node (nm)

Code converter

Frequency = 25 MHz PD (µW)

T d (ns)

PDP (× 10–15 W-s)

PD (µW)

T d (ns)

PDP (× 10–15 W-s)

32

BCX3

0.69

0.65

0.453

0.74

0.50

0.375

GtoB

0.48

1.33

0.646

0.60

1.18

0.718

22

16

Frequency = 50 MHz

BtoG

0.34

0.50

0.173

0.38

0.39

0.150

BCX3

1.10

1.41

1.564

1.35

0.95

1.293

GtoB

0.66

3.05

2.042

0.99

2.43

2.418

BtoG

0.56

0.92

0.517

0.72

0.74

0.537

BCX3

0.28

1.51

0.428

0.38

1.44

0.560

GtoB

0.27

3.15

0.852

0.45

3.07

1.389

BtoG

0.15

1.24

0.195

0.27

1.19

0.329

Table 4 Power, delay, and PDP for static CMOS with C L = 10 fF for 25 and 50 MHz Technology node (nm)

Code converter

Frequency = 25 MHz PD (µW)

T d (ns)

PDP (× 10–15 W-s)

PD (µW)

T d (ns)

PDP (× 10–15 W-s)

32

BCX3

2.07

0.31

0.656

2.91

0.27

0.788

GtoB

2.62

0.50

1.313

4.23

0.49

2.108

22

16

Frequency = 50 MHz

BtoG

1.70

0.33

0.561

2.51

0.29

0.750

BCX3

1.42

0.32

0.461

1.93

0.29

0.561

GtoB

1.27

0.70

0.904

1.85

0.66

1.226

BtoG

1.13

0.35

0.398

1.63

0.34

0.562

BCX3

0.40

0.45

0.183

0.63

0.42

0.265

GtoB

0.34

0.97

0.332

0.55

0.94

0.520

BtoG

0.31

0.47

0.145

0.48

0.44

0.213

Fig. 7 a Power consumption comparison. b PDP comparison

40

J. V. N. Saketh Ram et al.

node IPGL logic offers least PDP. The PDP is higher only in static CMOS for 32 nm technology as compared with other adiabatic logic design style. Power Analysis In this section, we shall compare the power saving obtained by using adiabatic logic circuits as compared to corresponding optimized static CMOS logic circuits. Table 5 gives the comparative power savings of adiabatic logic circuits at frequencies 25 and 50 MHz, respectively. Table 5 shows that the % power saving is highest in ECRL logic for technology node 32 and 22 nm, while IPGL logic shows highest power saving in 16 nm technology node. 2N_2N2P logic shows least power saving generally. The following important observations were made regarding the power consumptions while cascading stages and without cascading the stages, as observed in Fig. 8a, b at C L = 50 fF. It can be seen that in GtoB converter since the stages are cascaded and also the load capacitance is larger, the delay is larger than the ramp duration; hence for later stages, the computation takes place in hold phase (H) instead of evaluate phase (E), while in case of BtoG converter the execution takes place entirely in evaluate phase, thereby consuming less power. The power consumed in first case is 1.10 µW, while in BtoG it is 0.667 µW. Hence, it is evident that power consumption will be greater if computation is done after evaluation phase. Therefore, we must ensure that ramp is sufficient for entire computation to take place and avoid cascading more than 3 stages; otherwise, there will be a trade-off in frequency.

6 Conclusion This paper compares three code converters (BtoG, GtoB, and BCX3) designed using three adiabatic logic, namely ECRL, IPGL, and 2N_2N2P using 16, 22, and 32 nm technology nodes. The designs are tested using capacitive load of 10 fF at 25 and 50 MHz frequencies. The operation of different code converters has been evaluated and verified. The power consumption, propagation delay, and power–delay product (PDP) are compared for these adiabatic designs. Results indicate ECRL adiabatic logic style consumes minimum power at 32 and 22 nm technology node, while IPGL adiabatic logic consumes least power at 16 nm technology node. The effect of loading on power consumption due to loading and type of circuit is also investigated.

16

0.401

0.342

0.309

BtoG

1.134

BtoG

GtoB

1.279

BCX3

1.42

1.703

BtoG

GtoB

2.622

GtoB

BCX3

2.079

BCX3

32

22

% saving in ECRL

49.2

21.1

29.4

50.4

47.7

21.9

79.7

81.5

66.6

59.5

38.8

40.9

53.5

50.9

24.8

78.1

81.8

64.3

53.4

30.7

36.2

81.1

77.6

69.4

83.6

86.8

73.5

0.485

0.551

0.63

1.636

1.852

1.93

2.518

4.234

2.911

42.9

18.3

38.2

55.7

46.4

29.6

84.8

85.6

74.6

% saving in 2 N-2N2P

Static CMOS power (µW)

% saving in IPGL

Static CMOS power (µW)

% saving in 2 N-2N2P

Frequency = 50 MHz

Convertor

Technology node (nm)

Frequency = 25 MHz

Table 5 Comparison between power consumption at 25 and 50 MHz

69.5

43.6

56.1

58.9

53.2

32.1

86.2

88.1

76.3

% saving in IPGL

52.8

26.1

39.2

84.2

79.7

74.2

87.6

89.9

80.1

% saving in ECRL

Adiabatic Logic Code Converter Design … 41

42

J. V. N. Saketh Ram et al.

Fig. 8 a Transition in GtoB converter. b Transition in BtoG converter

References 1. M. Keating, D. Flynn, R. Aitken, A. Gibbons, K. Shi, Low Power Methodology Manual for System-On-Chip Design (Springer, Berlin, 2007) 2. D. Rossi, F. Conti, A. Marongiu, A. Pullini, I. Loi, M. Gautschi, G. Tagliavini, A. Capotondi, P. Flatresse, L. Benini, PULP: A parallel ultra low power platform for next generation IoT applications. In IEEE Hot Chips 27 Symposium (HCS) (2015) 3. J.G. Koller, W.C. Athas, Adiabatic switching, low energy computing, and the physics of storing and erasing information. In Workshop on Physics and Computation, Dallas, TX, USA, pp. 267– 270 (1992). https://doi.org/10.1109/PHYCMP.1992.615554 4. P. Teichmann, Fundamentals of adiabatic logic. In Adiabatic Logic. Springer Series in Advanced Microelectronics, vol. 34 (Springer, Dordrecht, 2012) 5. M. Saravanan, K.S. Manic, Energy efficient code converters using reversible logic gates. Int. Conf. Green High Perform. Comput. (ICGHPC) 2013, 1–6 (2013). https://doi.org/10.1109/ ICGHPC.2013.6533921 6. K. Nagata, F. Nemenzo, Some properties of binary gray code. Int. Conf. Comput. Appl. Technol. 2015, 72–75 (2015). https://doi.org/10.1109/CCATS.2015.27 7. Y. Moon, D.-K. Jeong, Efficient charge recovery logic, Digest of Technical Papers. In Symposium on VLSI Circuits, Kyoto, Japan (1995), pp. 129–130. https://doi.org/10.1109/VLSIC. 1995.520719 8. L. Varga, F. Kovacs, G. Hosszu, An improved pass-gate adiabatic logic. In Proceedings 14th Annual IEEE International ASIC/SOC Conference (IEEE Cat. No.01TH8558), pp. 208–211 (2001). https://doi.org/10.1109/ASIC.2001.954699 9. V.S. Kanchana Bhaaskaran, Adiabatic logic circuit design with integrated power clock generator. In Proceedings, Third International Conference on Signals, Systems & Devices, vol. IV, March 21–24 (2005) 10. V.S. Kanchana Bhaaskaran, S. Salivahanan, D.S. Emmanuel, Semi-custom design of adiabatic adder circuits. In 19th International Conference on VLSI Design Held Jointly with 5th International Conference on Embedded Systems Design (VLSID’06) (2006), p. 4. https://doi.org/10. 1109/VLSID.2006.144 11. S. Alam, S.R. Ghimiray, M. Kumar, Performance analysis of a 4-bit comparator circuit using different adiabatic logics. In 2017 Innovations in Power and Advanced Computing Technologies (i-PACT) (2017), pp. 1–5. https://doi.org/10.1109/IPACT.2017.8245210

Adiabatic Logic Code Converter Design …

43

12. C. Chugh, P. Kaur, A novel adiabatic technique for energy efficient logic circuits design. Int. Conf. Intel. Circ. Syst. (ICICS) 2018, 449–455 (2018). https://doi.org/10.1109/ICICS.2018. 00097 13. A. Agrawal, T.K. Gupta, A.K. Dadoria, D. Kumar, A novel efficient adiabatic logic design for ultra low power. Int. Conf. ICT in Bus. Indus. Govern. (ICTBIG) 2016, 1–7 (2016). https://doi. org/10.1109/ICTBIG.2016.7892723 14. A. Bargagli-Stoffi, G. Iannaccone, S. Di Pascoli, E. Amirante, D. Schmitt-Landsiedel, Fourphase power clock generator for adiabatic logic circuits. Electron. Lett. 38(14), 689–690 (2002) 15. H. Sharma, R. Singh, Comparative power analysis of CMOS & adiabatic logic gates. Int. Conf. Green Comput. Internet Things (ICGCIoT) 2015, 7–11 (2015). https://doi.org/10.1109/ICG CIoT.2015.7380418

A Cross-connected Switch Capacitor Multilevel Inverter: A Proposed Topology and Its Analysis Lipika Nanda , Chitralekha Jena, Babita Panda, and Arjyadhara Pradhan

1 Introduction A multilevel inverter is a power electronic converter which leads to a desired AC output from several levels of input DC voltages [1]. Basically, multilevel inverters also have many advantages with respect to hard switched two-level inverter topology. Multilevel inverter can be switched at a low as well as high switching frequencies than pulse width modulation controlled inverters [2–4]. The two-level conventional inverter topologies [5, 6] have many demerits which have been overcome by multilevel inverters. However, switched-capacitor MLI has lesser number of DC supply compared to other existed topologies. It can operate in both symmetrical and asymmetrical modes [7–10]. In this paper, comparison of switched-capacitor multilevel inverter with other already existed topologies has been carried out. The comparison basically depends on the number of DC sources, switches, TSV, etc., of various topologies with respect to the proposed topology. The main feature of the proposed topology is its voltage balancing capacity of the capacitor. For all the classical topologies, there is massive increment in device count as the output level increases [8, 11–14]. This leads to complexity in control and bulky systems and is inevitably expensive. The proposed topology offers high-energy conversion quality using little number of active and passive devices as compared to CHBs, which leads to low production cost and reduced switching loses.

L. Nanda (B) · C. Jena · B. Panda · A. Pradhan KIIT Deemed to be University, BBSR, Bhubaneswar, Odisha, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_6

45

46

L. Nanda et al.

2 Working Principle The working principle of the proposed topology with two DC sources in the main circuit has been presented in Fig. 1.

2.1 Modes of Operation The auxiliary circuit as stated in the proposed topology has fixed number of DC sources, and the values of the DC supplies are also same. The value of the DC source voltage is one-fourth of the main circuit. To generate zero level as in Fig. 2, either all the seven switches from upper switches or seven switches from lower switches are turned on.

Fig. 1 Proposed topology

Fig. 2 Zero output-level voltage for n = 2

A Cross-connected Switch Capacitor Multilevel Inverter …

47

Fig. 3 Average voltage output level for n = 2

For generating + 3VDC /2 as in Fig. 3 S1 , S2 and S 3 are in the on state of CCS   , S22 , S33 and S44 are in inverter (main circuit). In the auxiliary circuit, switches S11  the conducting mode. As S11 is on, capacitor C11 remains in the charging mode. As S22 is on, the auxiliary source goes through that path. This leaves the diode D2 in reverse bias mode, allowing for discharging of capacitor C22 to the load. To generate maximum output voltage + 11VDC /4 across the load, i.e., in Fig. 4 switches S1 , S 2 and S 3 are in the on state in the main circuit. S 11 and S 22 are in the on state in the auxiliary circuit. This ensures diodes D1 and D2 are reversed biased to prevent capacitors C11 and C22 discharging to the source rather to the load. For negative half cycle, the switching patterns remain the same as described. Table 1 explains the various switching schemes of the proposed converter. Positive sign indicates charging of the capacitors, and negative sign indicates discharging of capacitors. Sequence 1:1:1:1 The configuration considered is the symmetric configuration. The configuration is with respect to the main circuit. The auxiliary circuit ought to remain constant. Therefore, ‘n’ refers to the number of isolated sources in the main circuit. The general expressions are as follows. The number of levels generated and number f power switches required are Nlevel = 8(n + 1) − 1

(1)

Nswitches = (2n + 2) + 8

(2)

The total blocking voltage of the entire circuit and peak output voltages are: TSV = Nlevel /2 + 5/4 = 4n + 17/4

(3)

48

L. Nanda et al.

Fig. 4 Comparison of the different parameters of different algorithm sequences of the proposed topology. a Number of DC sources versus number of levels. b Number of power switches versus number of levels. c TSV versus number of levels. d Maximum output voltage versus number of levels

Vo,max = VDC

n 

i + 3/4 = VDC (n + 3/4)

(4)

i=1

In attempting to generate more output levels with the same or lesser number of switches as compared to the symmetric configuration, different asymmetric configurations are retried. The most commonly used asymmetric configurations, binary configuration—sources—have geometric progression with a factor two, and trinary configuration sources have geometric progression with a factor three. Nevertheless, they cannot synthesize certain output levels or give equal step size for acrossconnected MLI, our main circuit. In order to overcome the constraints, other forms

A Cross-connected Switch Capacitor Multilevel Inverter …

49

Table 1 Various switching schemes State

Output voltage level V 0

Switch states; 1–on; 0–off

Capacitor states

S1

S2

S3

S 11

S 22

S 33

S 44

C 11

C 22

1

0

0

0

1

0

0

1

1

+

+

2

0

0

0

0

0

0

0

0

+

+

3

+V DC /4

0

0

0

0

0

1

0

+

+

4

+V DC /2

0

0

0

0

1

1

0

+

−

5

+3V DC /4

0

0

0

1

1

1

0

−

−

6

+V DC

0

1

1

0

0

1

1

+

+

7

+ V DC /4

0

1

1

0

0

1

0

+

+

8

+3V DC /2

0

1

1

0

1

1

0

+

−

9

+7V DC /4

0

1

1

1

1

1

0

−

−

10

+V DC

1

0

1

0

0

1

1

+

+

11

+ 9V DC /4

0

1

0

0

0

1

0

+

+

12

+5V DC /2

0

1

0

0

1

1

0

+

−

13

+11V DC /4

0

1

0

1

1

1

0

−

−

14

−V DC /4

0

0

0

0

0

0

1

+

+

15

−V DC /2

0

0

0

0

1

0

1

+

−

16

−3V DC /4

0

0

0

1

1

0

1

−

−

17

−V DC

1

0

0

0

0

1

1

+

+

18

−5V DC /4

1

0

0

0

0

0

1

+

+

19

−3V DC /2

1

0

0

0

1

0

1

+

−

20

−7V DC /4

1

0

0

1

1

0

1

−

−

21

−V DC

1

0

1

0

0

1

1

+

+

22

−9V DC /4

1

0

1

0

0

0

1

+

+

23

−5V DC /2

1

0

1

0

1

0

1

+

−

24

−11V DC /4

1

0

1

1

1

0

1

−

−

of asymmetric configurations have sprung up. Applied to the proposed topology are a few of them, discussion of which is in the sequel sections. Sequence 1, 2, 3 … n It is also known as natural number sequence. The value of the voltage sources is in a natural number sequence form, i.e., 1, 2, 3, 4 … nV DC to enable the analysis of all possible outcomes. In order to generate equal step size output level, placing the sources in the correct position is necessary. Below is an algorithm to determine the correct position of each DC source. With n number of sources, I is the voltage source number.

50

L. Nanda et al.

For even number of sources: VDC i = (2i − 1)VDC , for 1  i  (n/2) and 2(n + 1 − i)VDC , for (n + 2)/2  i  n

(5)

For odd number of sources: VDCi = (2i − 1)VDC , for 1  i  (n + 1)/2 and 2(n + 1 − i)VDC , for (n + 3)/2  i  n   Nlevel = 2 4

n 

 i +3

  + 1 = 2 2n 2 + 2n + 3 + 1

(6)

(7)

i=1

The number of switches required for this configuration is: Nswitches = (2n + 2) + 8

(8)

Total standing voltage of this sequence is:   TSV = Nlevel /2 + 5/4 = 2 n 2 + n + 17/4

(9)

To calculate the peak output voltage: Vo,max = VDC

n 

  i + 3/4 = n 2 + n /2 + 17/4

(10)

i=1

Sequence 1:2:2:2:2 One DC voltage has voltage VDC , and another DC voltage has voltage 2VDC Nlevel

  n     =2 4 i +3 + 1 = 2 2n 2 + 2n + 3 + 1

(11)

i=1

The number of switches required for this configuration is: Nswitches = (2n + 2) + 8

(12)

Total standing voltage of this sequence is:   TSV = Nlevel /2 + 5/4 = 2 n 2 + n + 17/4 To calculate the peak output voltage:

(13)

A Cross-connected Switch Capacitor Multilevel Inverter …

Vo,max = VDC

n 

  i + 3/4 = n 2 + n /2 + 17/4

51

(14)

i=1

The total blocking voltage of the entire circuit: TSV = Nlevel /2 + 5/4 = 4n + 17/4

(15)

To calculate the peak output voltage: Vo,max = VDC

n 

i + 3/4 = VDC (n + 3/4)

i=1

2.2 Comparison Study This topic highlights comparison between the different possible algorithms used in the main circuit of the proposed topology. The comparison of the proposed topology with a symmetric main circuit to Cascaded H-Bridge (CHB) MLI, Cross-switched MLI using Auxiliary reverse connected voltage sources (CCS with Aux.RV), Cross Connected Switch (CCS) MLI and most recent inverters, has been represented in Fig. 5. All in symmetric configurations. Furthermore, comparison has been made between proposed topology with an asymmetric main circuit (natural sequence) and the aforementioned MLIs (all in natural sequence configuration). In comparing all the earlier-mentioned sequences/algorithms and different MLI topologies, the prime objective still remains, has a reduced number of device count, and generates high number of output levels.

2.2.1

Proposed Topology Sequence Comparison

Denotation of some of the configurations; the symmetric configuration, as Main circuit. Symmetric (MCS), natural number sequence labeled as Main circuit. Asymmetric (1) (MCA1), sequence 1, 2, 2, 2…n, as Main cct. Asymmetric (2), (MCA2), and sequence; 1, 2, 3, 5, 8…n, as Main circuit Asymmetric (3) (MCA3). Figure 4a shows the number of isolated DC sources required by each of aforementioned configurations to synthesize a certain number of output levels. From the image, MCA3 and MCA1 need lesser number of power switches as to MCA2 and MCS. However, MCA3 requires less number compared to MCA1 as the number of output levels increases beyond 170. Figure 4b presents a comparison of number power switches needed for a certain number of output levels. MCA3 requires slightly less number of power switches in comparison with MCA1, which has a better ratio of number of output level to number of power switches compared to MCA2 and MCS.

52

L. Nanda et al.

MCS requires more switches than any configuration at any given NL. As shown in Fig. 4c, the variation in the variety of the DC sources required by each configuration is on display. The MCS requires less variety as to the rest. MCA1 and MCA3 have relatively the same number of variety except when the number of output level goes beyond 170. After that mark, MCA3 has lesser variety as to MCA1. Figure 4d compares the TSV versus the number of output levels generated. All the four configurations have virtually the same value for any given number of output levels. As the number of output level increases, a pattern starts to show. MCS and MCA3 are equal, having lower TSV than the other configuration. MCA2 is a bit better than the MCA1 in this regard. Judging by these comparisons made and the main aim if this work, one can infer that out of the four different configurations, MCA3 and MCA1 are more likely to achieve the desired goal and MCS to be the less likely. MCA3 and MCA1 will be much preferable because of their reduced number of switches and isolated sources, as the higher level. In the proposed topology configuration comparison, MCS was found to be the least likely to attain the desired objective. Comparing the MCS configuration to other types of multilevel topologies to help prove: even with the ‘worst’ option, the proposed topology has an advantage over other topologies, in terms of device count and output level synthesized. However, note be made, the proposed topology can never be symmetric, due to the auxiliary circuit. Figure 5a shows the number of voltage sources of the MCS and the above-mentioned topologies in their symmetric configuration versus the number of voltage levels at the output. It is visible that the proposed topology of MCS fares better than the rest of topologies. Figure 5b compares the topologies in terms of power semiconductor switches required to gain a certain amount of output level. MLDCL and the CCS with Aux.RV are equal. They require lightly more switches than the CCS topology. CHB has the highest number of switches as the number of level increases; SSPS follows closely. The MCS has a better number of switch count. Figure 5c shows the comparison of the proposed topology to the other MLIs in terms of TSV versus number of output levels. Observations show the proposed topology to have a lower TSV as to the other MLIs. CCS, CHB and CCS with auxRV are equal and have lower TSV as to MLDCL, which also has a lower TSV as to SSPS. The margin between the proposed topology and CCS and the others widens even further as the number of levels increases (Fig. 5d). This is one of the drawbacks of the proposed topology as compared to other MLIs, low-peak output voltage.

3 Conclusion Different DC configurations have been used on the main circuit of the proposed topology (auxiliary is fixed). The main circuit has been looked at with a symmetric configuration of the sources and in asymmetric cases as well. The asymmetric configurations used are not the common combinations used, i.e., binary and trinary. This was due to the main circuit not able to synthesize all the required steps with binary and

A Cross-connected Switch Capacitor Multilevel Inverter …

53

Fig. 5 Comparison of different topologies with the proposed topology in symmetric configuration. a Number of DC sources versus number of output levels. b Number of power switches versus number of levels. c TSV versus number of levels. d Maximum possible voltage output versus number of levels

trinary configurations. All the different asymmetric configurations did better than the symmetric configuration in terms of device count, but the symmetric configuration is better concerning TSV. The symmetric configuration in (main circuit) compared to other topologies in symmetric configuration did better in all aspects, device count, output voltage level. Argument might made be it is an unfair comparison given that the proposed topology is never symmetric in whole, due to it having an auxiliary circuit with different parameters. The proposed topology in comparison with the other MLIs in natural sequence configuration; discovered to have more output levels with lesser device count, which has been the essence of this paper.

54

L. Nanda et al.

References 1. K.K. Gupta, A. Ranjan, L. Bhatnagar, Multilevel inverter topologies with reduced device count: a review. IEEE Trans. Power Electron. 31(1), 135–151 (2016) 2. L.G. Franquelo, J. Rodriguez, J.I. Leon, The age of multilevel converters arrives. IEEE Ind. Electron. Mag. 2(2), 28–39 (2008) 3. J. Rodriguez, J. Lai, Multilevel inverters: a survey of topologies, controls, and applications. IEEE Trans. Industr. Electron. 49(4), 724–738 (2002) 4. J. Lai, F. Peng, Multilevel converters-a new breed of power converters. IEEE Trans. Ind. Appl. 32(3), 509–517 (1996) 5. L. Nanda, A. Dasgupta, U.K. Rout, A comparative analysis of symmetrical and asymmetrical cascaded multilevel inverter having reduced number of switches and DC sources. Int. J. Power Electron. Drive Syst. 8(4), 1595–1602 (2017) 6. L. Nanda, A. Dasgupta, A comparative studies of cascaded multilevel inverters havingreduced number of switches with R and RL-load. Int. J. Power Electron. Drive Syst. (IJPEDS) 8(1), 40–50 (2017) 7. M.F. Kangarlu, E. Babaei, Cross-switched multilevel inverter: an innovative topology. IET Power Electron. 6(4), 642–651 (2013) 8. S. Thamizharasan, J. Baskaran, S. Ramkumar, Cross-switched multilevel inverter using auxiliary reverse-connected voltage sources. IET Power Electron. 7(6), 1519–1526 (2014) 9. Y.H. Liao, C.M. Lai, Newly-constructed simplified single-phase multistring multi-level inverter topology for distributed energy resources. IEEE Trans. Power Electron. 26(9), 2386–2392 (2011) 10. L. Nanda, A. Dasgupta, U.K. Rout, A comparative analysis of modified cascaded multilevel inverter having reduced number of switches and DC sources. Int. J. Appl. Eng. Res. 12(20), 10121–10126 (2017) 11. L. Nanda, A. Dasgupta, U.K. Rout, A comparative analysis of symmetrical andasymmetrical cascaded multilevel inverter having reduced number of switches and DC sources. Int. J. Power Electron. Drive Syst. (IJPEDS) 8(4), 1595–1602 (2017) 12. P. Palanivel, S.S. Dash, Analysis of THD and output voltage performance for cascaded multilevel inverter using carrier pulse width modulation techniques. Power Electron. IET 4(8), 951–958 (2011) 13. L. Nanda et al., A comparative studies of different topologies of multilevel inverterwith SIMULINK, in 2017 International Conference on Inventive Systems and Control (ICISC), pp. 1–7 (2017) 14. M.W. Tesfay, T. Roy, S.K. Swain, L. Nanda, A novel step-up 7L switched-capacitor multilevel inverter and its extended structure, in 2021 1st International Conference on Power Electronics and Energy (ICPEE), pp. 1–6 (2021)

A Comparative Study on Sentiment Analysis of Uber and Ola Customer Reviews Based on Machine Learning Approaches Sandeep Kumar, Anuj Kumar Singh, Shashi Bhushan, Pramod Kumar, and Arun Vashishtha

1 Introduction A progression of instruments, procedures, and strategies used to distinguish what is more can be refined and processed information is defined as sentiment analysis. The data removed incorporates perspectives and assessment of the user. The principle objective is to know whether the user has negative, positive, and neutral assessment toward an item or something different. A huge dataset of users is actively available on social networking site Twitter. Hence, Twitter uses datasets where user tweets their reviews. Sentiments are the general reviews of the customers and understanding what reviews do customers have about a product or service through their tweets. Sentiment analysis is one part of machine learning that is often used to examine words based on the patterns of customer’s reviews like positive, negative, or neutral sentiments. Language is a powerful instrument that aids in the expression of feelings. Sentiment analysis is a type of text mining that helps a company figure out how people feel about their brand or product. It solves real-world problems by combining natural language processing and data mining techniques [1]. Businesses could be improved in addition to gaining insights about a brand through customer feedback. Person is giving S. Kumar · A. K. Singh (B) · P. Kumar · A. Vashishtha Krishna Engineering College, Ghaziabad, India e-mail: [email protected] S. Kumar e-mail: [email protected] P. Kumar e-mail: [email protected] A. Vashishtha e-mail: [email protected] S. Bhushan School of Computer Science, University of Petroleum and Energy Studies, Dehradun, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_7

55

56

S. Kumar et al.

Fig. 1 Sentiment classification

Sentence Level Document Level

Aspect Level

Sentiment Classification

feedback or review on social media platforms which are evaluated and help others for better understanding or to take decisions. Are they satisfied or dissatisfied? [2]. The tweets, which are essentially people’s reviews, are divided into two sentiments in this paper: favorable and negative. Because users tweet in languages with which they are familiar, the majority of tweets contain difficult-to-clean material. The datasets used in this paper are “UBER” and “OLA”. Python programming language is used in this project. Python is a programming language used for computing and data analysis. The reason for selecting this programming language is that it gives better results for analyzing and understanding the data precisely as it contains different types of packages for example e1071 [3]. This paper uses machine learning algorithm techniques which are support vector machine (SVM), “Naïve Bayes,” and random forest. These three classification methods are part of supervised machine learning and classify data into distinct groups. The reason for using these algorithms is that they produce better text classification results [4]. Machine learning was considered as a method to better understand how it could influence Uber and Ola’s Twitter sentiment analysis. The three flavors at sentence and aspect connected to machine learning are Naive Bayes, SVM-support random forest, and maximum entropy. Sentiment classification methodologies are based on methods such as decision trees such as K-nearest neighbors, hidden Markov model, and sequential minimal optimization. Lexicon and machine learning-based approaches are used in sentiment analysis [5] (Fig. 1).

2 Methodology Ola and Uber are most popular service providers for cab services, as seen by the numerous reviews available. The dataset was unlabeled, and in order to use it in a supervised learning model, it needed to be labeled. Finally, this research activity was limited to Ola and Uber customer feedback, namely reviews about the

A Comparative Study on Sentiment Analysis of Uber and Ola …

57

services provided to customers. For the purpose of measuring polarization, about 6000 customer’s reviews were examined. Following steps were used to process the results: (A)

Data Gathering from Twitter

(B)

Data acquisition was completed as the first step in the data labeling process. Manual labeling is impossible for a human to do because the dataset contains a large number of customer reviews. As a result, after the datasets have been preprocessed, the active learner has been used to label them. Because Uber and Ola reviews are given out in five-star increments, three-star ratings are typically viewed as neutral reviews, meaning they are neither negative nor favorable. As a result, every review in the dataset with a three-star rating was deleted by the system, and the remaining reviews were used to move on to the next stage, which is the preprocessing step [6]. Preprocessing of data Tokenization, deleting stop words, and using the global constant to fill in the blanks are the three phases in preprocessing data. • Tokenization: It is the process of breaking down a string sequence into individual elements such as keywords, words, symbols, phrases, and other tokens. Tokens can eventually become phrases, words, or even full sentences. However, during the tokenization process, some characters, such as punctuation marks, are eliminated [7]. • Removing stop words: Stop words are items in a sentence that are not required for any division in text mining. These terms are frequently avoided in order to increase the accuracy of the assessment. Stop words come in a variety of formats, depending on the realm, language. • Using a global constant to fill in the blanks: The system has been searching the dataset for the blank value in this stage. Missing values were afterward replaced with the appropriate constant to fill in the gaps.

(C)

Sentiment detection

(D)

We look at the extracted sentences from the reviews and viewpoints. Sentences containing subjective expressions (opinions, beliefs, and perspectives) are kept, whereas sentences containing objective communication (facts, factual information) are removed. Sentiment classification In this step, subjective sentences are classified in positive, negative, good, bad and like, dislike, but classification can be made by using multiple points. Sentiment classification is an automated technique that recognizes thoughts in text and categorizes them as good, negative, or neutral based on the emotions expressed by customers. Sentiment classification, which uses natural language processing to evaluate subjective data, can help you understand how people feel about your products, services, or brand. With sentiment classification

58

(E)

S. Kumar et al.

software, we can quickly recognize emotions in large amounts of text. Not only is it important to offer consistent and precise outcomes, but it is also important to manage data in real time. Machine learning techniques are used by automated systems to learn to anticipate sentiment from previous observations. Output Presentation The basic goal of sentiment analysis is to transform unstructured text into useful data. The text results are displayed on graphs such as pie charts, bar charts, and line graphs once the analysis is completed. Time can also be analyzed and graphically displayed by creating a sentiment time line over time with the desired value (frequency, percentages, and averages) (Fig. 2).

Fig. 2 Overview of the system methodology

Data Gathering from Twi er

Preprocessing of data

Sen ment Detec on

Sen ment Classifica on

Output Presenta on

A Comparative Study on Sentiment Analysis of Uber and Ola …

59

3 Experimental Result Through a series of experiments, this section will assess the performance of these three machine learning models. Evaluating metrics are important in determining classification efficiency, and assessing accuracy is the easiest way to do so. Ultimately, the precision of a classifier on a given test dataset is the fraction of those datasets that it correctly categorizes. Recall, precision, and the accuracy, which is produced from a confusion matrix, are three regularly used statistical measures that are used to evaluate the system. True positive tweets (TPT), true negative tweets (TNT), false positive tweets (FPT), and false negative tweets (FNT) are the four categories in which the data from the confusion matrix is categorized. The term “true positive tweets” refers to an outcome in which the system accurately predicts the positive class. False positive tweets, on the other hand, refer to an outcome in which the scheme predicts the positive class inaccurately. True negative tweets, on the other hand, are the outcome in which the system correctly predicts the adverse class. False negative, on the other hand, is an outcome in which the system predicts the negative class inaccurately (Figs. 3, 4, and 5). 2000 1500 1000 500 0 Uber

Ola

Uber

SVM

Ola

Uber

Naïve Bayes

Ola KNN

6000 Posi ve Tweets

Naga ve Tweets

Fig. 3 Diagrammatic analysis of positive and negative data tweets

The Diagrammatic representation of accuracies in the experiment Ola Tweets

Uber Tweets

90.00% 85.00% 80.00% 75.00% Accuracy Precision SVM

Recall

Accuracy Precision

Recall

NB

Fig. 4 Diagrammatical accuracy representation from the results

Accuracy Precision KNN

Recall

60

S. Kumar et al.

SVM

Algorithms KNN NB

Differential classifiers' performance RECALL PRECISION ACCURACY RECALL PRECISION ACCURACY RECALL PRECISION ACCURACY 78.00%

80.00%

82.00%

84.00%

86.00%

88.00%

90.00%

percentage Uber Tweets

Ola Tweets

Fig. 5 Differential classifiers performance from result datasets

(a)

Accuracy The following formula can be used to calculate accuracy: Accuracy = (TPT + TNT)/(TPT + TNT + FPT + FNT)

(b)

Precision The precision value represents the percentage of relevant retrieved instances. The percentage of relevant positive results returned is used to calculate precision. Precision = TPT/(TPT + FPT)

(c)

Recall The fraction of relevant instances that are recovered is used to determine recall value. The percentage of positive outcomes returned that are precise is calculated. In this context, recall is also known as the true positive rate. The percentage of relevant instances that are retrieved is known as recall. Recall = FNT/(TPT + FNT)

This work was able to give a comparison of SVM Naive Bayes and KNN classifiers in order to examine the polarization of sentiment in Ola and Uber customer reviews. After the preprocessing step, the models were trained with about 3250 features and almost 6000 datasets. In the meantime, about 4000 test sets have been run through the statistical models (Tables 1 and 2).

A Comparative Study on Sentiment Analysis of Uber and Ola …

61

Table 1 Analysis of positive and negative tweets Number of data

Algorithm used

Dataset

Positive tweets

Negative tweets

6000

SVM

Uber

1470

452

Ola

1268

612

Naïve Bayes

Uber

1680

368

Ola

1468

472

Uber

1370

378

Ola

1278

468

KNN

4 Conclusion Sentiment analysis and opinion mining research is very essential nowadays. Most industries generate many forms of data, which they must examine in order to make decisions that benefit the industry. In addition, social media generates a large amount of data, which necessitates the analysis and discovery of insights from that data. The purpose of this study is to compare the accuracy of the three classification algorithms and to determine what accuracy is created, as well as to understand people’s attitudes using sentimental analysis. The three algorithms are compared in this study for emotive classification of tweets. The experimental data revealed that the classifier performed better on the Uber datasets when trained using Naive Bayes and similarly on the Ola datasets when trained with Naive Bayes. As can be shown, Naive Bayes was dominating in both cases, with an accuracy of 88.35% in the case of Uber and 86.45% in the case of Ola. As a result, the Naive Bayes method is a better approach for classifying the Uber and Ola datasets. They can concentrate even more on areas where user services can be improved. For example, if we discover that a particular location has produced outputs with higher unfavorable sentiments, we can focus on that area and determine why. We can look into how we might provide more value to such clients. Also, while this research focused on shallow learning, same work can be done with deep learning using various datasets or approaches. This paper exclusively uses the customers review dataset for sentiment analysis. Various fields have different terminology and description methods in other text datasets, so the difference in sentiment orientation in different types of text data is clear. Different applications can create their own corpus in various sectors, allowing for better sentiment analysis of text.

85.57

86.67

Uber tweets

84.35

82.45 86.34

88.37 88.35

86.45

NB Accuracy (%)

Recall (%)

Accuracy (%)

Precision (%)

SVM

Classifiers

Ola tweets

Data set

Table 2 Accuracy comparison on dataset

86.43

88.35

Precision (%) 87.36

89.34

Recall (%)

KNN

86.92

85.92

Accuracy (%)

82.34

83.32

Precision (%)

82.34

85.34

Recall (%)

62 S. Kumar et al.

A Comparative Study on Sentiment Analysis of Uber and Ola …

63

References 1. P. Kalaivani, Sentiment classification of movie reviews by supervised machine learning approaches. Indian J. Comput. Sci. Eng. (IJCSE) 4(4) (2017). ISSN: 0976-5166 2. Progress in Computing, Analytics and Networking. (Springer Science and Business Media LLC, 2018) 3. A. Pak, P. Paroubek, Twitter as a corpus for sentiment analysis and opinion mining. LREc 10, 2019 (2010) 4. M. Pennacchiotti, A.-M. Popescu, A machine learning approach to twitter user classification, in Fifth International AAAI Conference on Weblogs and Social Media (2018) 5. B.M. Jadav, V.B. Vaghela, Sentiment analysis using support vector machine based on feature selection and semantic analysis (2018) 6. L. Dey, S. Chakraborty, A. Biswas, B. Bose, S. Tiwari, Sentiment analysis of review datasets using Naïve Bayes and K-NN classifier. Int. J. Inf. Eng. Electron. Bus. (2018) 7. M.N. Tibdewal, S.A. Tale, Multichannel detection of epilepsy using SVM classifier on EEG signal, in 2016 International Conference on Computing Communication Control and automation (ICCUBEA) (2019) 8. W. Fan, L. Wallace, S. Rich, Z. Zhang, Tapping into the power of text mining. J. ACM (2019) 9. A. Go, R. Bhayani, L. Huang, Twitter sentiment classification using distant supervision. CS224N project report, Stanford 1.12 (2009) 10. L.L. Dhande, G.K. Patnaik, Analyzing sentiment of movie review data using Naive Bayes neural classifier. IJETTCS 3(4) (2018). ISSN 2278-6856 11. P. Kalaivani, K.L. Shunmuganathan, Sentiment classification of movie review by supervise machine learning approach. Indian J. Comput. Sci. Eng. (IJCSE) 4(4) (2019) 14. A. Mulkalwar, K. Kelkar, Sentiment analysis on movie reviews based on combined approach. Int. J. Sci. Res. 3(7) (2018) 15. S. Bhushan, P. Kumar, A. Kumar, V. Sharma, Scantime antivirus evasion and malware deployment using silent-SFX, in 2016 International Conference on Advances in Computing, Communication, and Automation (ICACCA) (Spring), pp. 1–4 (2016). https://doi.org/10.1109/ICA CCA.2016.7578894

A Cost-Effective IoT-Assisted Framework for Automatic Irrigation Rewa Sharma

and Keshav Kaushik

1 Introduction Internet of Things has revolutionized various application domains; one of these is the agriculture domain. IoT plays a significant role by providing help to farmers to deal with the problems associated with agriculture effectively. Agriculture is a field where water is needed in large quantities. Recent IoT applications address scarcity of water, limited agricultural land availability, etc., thus helping in achieving the increasing food needs of population by providing qualitative and cost-effective solutions for agricultural production. In India, farmers use irrigation for watering their crops [1, 2] as they cannot rely solely on monsoons for their agricultural growth. The type of soil is a significant factor, which determines the amount of water needed for irrigation [1, 3]. It has been observed that agriculture has a substantial contribution to GDP and provides employment to many people. Based on the survey [4], it is observed that agriculture has a significant contribution to GDP and offers employment opportunities to a huge proportion of the Indian population. Population explosion and increased food demands lead to increased water consumption. Unpredicted rains and decreasing levels of groundwater ultimately lead to a reduction of water volume on earth. We need water to fulfill our daily activities. Wastage of water or over-irrigation is a major problem in agriculture. Farmers must look for several methods to save water and maintain the highest production. However, even in this modern era, the growers have been using a manually controlled irrigation system. Automatic irrigation regulates the supplied water efficiently as compared with traditional manually operated irrigation. Optimal water R. Sharma Department of Computer Engineering, J.C. Bose University of Science and Technology, Faridabad, Haryana, India K. Kaushik (B) School of Computer Science, University of Petroleum and Energy Studies, Dehradun, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_8

65

66

R. Sharma and K. Kaushik

consumption and enhanced water productivity can be achieved with the help of recent Internet technologies and wireless networks. Wireless sensor networks [5] comprise a vast number of sensor nodes, which sense the environment and collect some particular data type, such as temperature and air pressure. IoT is a network technology, which makes everything connected to the Internet. It uses sensors deployed at different locations to sense surroundings and environment and help nodes exchange helpful information with each other. IoT can simultaneously monitor and control various devices connected to the network. Communication devices like the Wi-Fi module are included in the central processing unit to exchange data between sensors and the user’s device. Data are processed into useful information before relaying it to the user’s device. This processed information is then made accessible to the user through some handheld device, for example, tablets, mobile phones, etc. The proposed framework is designed to avoid wastage of water due to overirrigation in the agricultural fields. Sensors have been used to sense various parameters like temperature, moisture, and humidity regularly. Sensors can constantly monitor these parameters as well. This perceived data are then sent to a designated IP address. Simple menu-driven android application has been used to continuously gather this data from that designated IP address. All the required sensors are placed in the root of the plant. These sensors then transmit the data to the android application. The threshold value for moisture content of the soil is programmed into a microcontroller to regulate the water supply. Relay is used to control the automatic irrigation system when the soil moisture value increases above a particular limit. The motivation to conduct this research is to help the farmers in agriculture and reduce the shortage of water. Past researches have suggested many possible solutions to aid irrigation. In this paper, we aimed to design a low-cost IoT-based irrigation system keeping in mind the financial conditions of the farmers. A comprehensive review has been conducted to investigate various researches performed earlier to assist farmers in agriculture. The proposed system is designed using low-cost hardware components and android application keeping in mind the cost considerations for the farmers. Section 4 provides a deep insight about the implementation of proposed framework and the results obtained. Section 1 presents an introduction to the research work related to significance of Internet of Things in agriculture domain. Section 2 represents related literature survey, and Sect. 3 represents IoT framework used in the proposed framework. This section provides information about all the hardware components used for the implementation. Section 4 illustrates methodology and results of the proposed research. This section elaborates working methodology of the proposed framework by describing the working of the system using a flowchart. It also provides a view of the android application interface showing its modes of operation. Graphical results showing the change in dryness value with respect to time are also shown in this section. Section 5 provides brief conclusion of the proposed research.

A Cost-Effective IoT-Assisted Framework for Automatic Irrigation

67

2 Related Work The authors studied the existing approaches for gaining insight about requirements to develop a model for the proposed system. Abdul Aziz et al. [6] developed a method to distantly observe, predict, and analyze modification in temperature in agriculture greenhouses with the help of a short messaging service and wireless sensor. Priyadharshini et al. [7] designed a system to minimize farmer’s regular manual intervention in controlling water pumps. The author used an android application to provide an easy interface to the farmer for controlling the water pump. Garcia et al. [8] designed a remotely operated irrigation system for agricultural fields using a microcontroller. They used a humidity sensor to detect the change in humidity, which is compared to a threshold value. The threshold value being higher than the observed value turns the motor on to supply water to the plant until the humidity of the soil reaches above the predefined value. The authors in this paper used a microcontroller with constrained memory to control the system. Chavda et al. [9] described the application of sensor networks in designing a cheap wireless-controlled irrigation system. The proposed approach was deployed in an area of 8 acres in a place located in central Anatolia for monitoring and controlling water supply to cherry plants. The authors gathered data using solar-powered wireless acquisition stations. Suma et al. [10] proposed a method for periodic monitoring of soil properties and environmental factors by deploying several sensor nodes at various positions in the agricultural farm. These parameters are controlled via Internet service or any remote device. Sensors, cameras, and Wi-Fi using microcontrollers perform desired actions cooperatively. Bavkar et al. [11] designed a system comprising of a robot-operated wirelessly. The developed system is comprised of different sensors for observing several environmental factors. The challenges in the recent research are that they are not costeffective, whereas our solution solves this problem. The main objective of the proposed approach is to perform a variety of operations such as spraying pesticides, sensing soil moisture content, moving forward or backward, frightening birds and animals, and switching the electric motor on or off.

3 IoT Components Used in the Proposed Framework 3.1 Arduino Arduino is a readily available, easy-to-use, open-source electronics platform. Figure 1 displays a microcontroller-based Arduino Uno board, which is based on ATMega 328P [12]. The ATMega 328P stores the program code in 32 kB flash memory and can be programmed using Arduino software. This Arduino board comprises digital

68

R. Sharma and K. Kaushik

Fig. 1 Arduino UNO

input and output pins, analog inputs, USB, an ICSP circuit, reset button, and 16 MHz frequency quartz crystal.

3.2 Sensors The soil moisture sensor shown in Fig. 2 determines the moisture content present in the soil where the sensor is placed [13]. Suppose the moisture value measured by the sensor is above a particular predefined threshold value, in that case, the digital output is shown as a low level, and if the estimated moisture value was below the specified threshold level, the digital output would be a high level. At a limited time, the soil moisture content is read by digital pin to see if it is above a threshold or not. The potentiometer can be used to regulate the threshold voltage. Arduino analog input can be directly connected with the output of an LM35 temperature sensor. The Arduino analog-to-digital converter has a resolution of 1024 bits, and the reference voltage is 5 V. Figure 3 displays a GSM module. The Arduino GSM shield allows an Arduino board to connect to the Internet, transmit and receive SMS, and make voice calls Fig. 2 Soil moisture sensor

A Cost-Effective IoT-Assisted Framework for Automatic Irrigation

69

Fig. 3 GSM module

using the GSM library. The shield also operates efficiently with several other boards, for example, the Mega, Leonardo boards, etc., with a slight modification [14].

4 Cost-Effective Proposed Framework for Automatic Irrigation: Methodology and Results In the proposed system, the input pins of the Arduino microcontroller are connected with sensors to measure the moisture and temperature of the environment. Suppose the sensed value exceeds the predefined threshold values set in the program, in that case, the relay circuit will automatically turn the switch on or off, and the driver circuit attached along the system will help switch the voltage. In addition to this, GSM module is used to inform the farmer about the current field condition. Farmers can also get an update through the android app. This system allows the farmers to observe the field scenario by sitting anywhere and at any time. Figure 4 displays the block diagram of the proposed system showing the interfacing of Arduino with sensors and GSM module. Figure 5 shows the android application interface, which offers two modes: mode 1 and mode 2. Figure 6 displays a snippet of the working user mode. To implement our automatic irrigation system, we provide a power supply to the hardware system. Now, our system is ready to work. Initially, we choose any mode as shown in the android application interface (one for operating based on sensors’ readings and the other for working because of user commands). If the user chooses mode 1 (i.e., sensor mode), lessons are fetched from sensors and then sent to the server. After this, the reading is compared to the threshold values (set according to the type of crop), and because of this, the motor is turned on or off. If the user chooses mode 2 (i.e., user mode), a new interface appears, showing two buttons for turning the motor on or off. If a user presses button 1 (i.e., turn engine on), the motor is turned on and vice-versa. We have used the following commands in user mode: “Motor on”—to turn the motor on. “Motor off”—to turn the engine off.

70

Fig. 4 Working methodology of the proposed framework

Fig. 5 Android application interface and a snippet of user mode

R. Sharma and K. Kaushik

A Cost-Effective IoT-Assisted Framework for Automatic Irrigation

71

Fig. 6 Flowchart showing working of the proposed framework

Figure 6 illustrates the working of the proposed system using a flowchart. Mode 1 takes readings from the sensors, and the data are sent to the server, where the sensed value is compared with the threshold value. If the observed values are lesser than the predefined ones, motor is turned ON. Mode 2 offers two buttons of motor on and motor off, respectively. Users of the system can choose among any of the two modes offered by the application. Figure 7 clearly illustrates the real-time circuit view and the working model of the proposed approach with attached soil moisture and temperature sensor. Figure 8 shows the graphical representation for soil dryness value when irrigated until the saturation point. Over time, the soil dryness value keeps on decreasing, and after the saturation value, it keeps on staying constant. The X-axis of the figure shows the timestamp, whereas the Y-axis shows the dryness value.

72

R. Sharma and K. Kaushik

Fig. 7 Working model of proposed framework

Fig. 8 Graphical representation for soil dryness value versus time

5 Conclusion The application of agriculture networking technology is needed for modern agricultural development. This proposed system reduces the wastage of water primarily by automating the irrigation process. In this paper, an automatic irrigation system was developed using a microcontroller. The proposed design is relatively compact and highly reliable, composed of Arduino UNO, relay, soil moisture sensor, temperature sensor, GSM module, motor and battery, and android application. The system reduces water wastage by optimally irrigating the soil based on the soil moisture sensor and temperature sensor values. The proposed method can provide a viable

A Cost-Effective IoT-Assisted Framework for Automatic Irrigation

73

solution to the problems faced by farmers in manual irrigation. It also provides an easy-to-use interface for farmers. They can perform the desired operation just by pressing a button on the android app, thereby significantly reducing problems faced by farmers. The hardware used for the implementation of the proposed framework is not too costly, and also, the system does not demand any additional manpower. Therefore, the proposed framework is cost-effective and can prove advantageous in places with having a shortage of water. Besides reducing water usage, it also ensures better crop yield as the right amount of water is supplied. Excess irrigation can harm the crops. The disadvantage of the system is that the malfunctioning of any component is not informed to the user, and each element of the design needs to be tested manually.

References 1. P. Rajalakshmi, S. Devi Mahalakshmi, IOT based crop-field monitoring and irrigation automation (2016). https://doi.org/10.1109/ISCO.2016.7726900 2. A. Carrasquilla-Batista, A. Chacon-Rodriguez, M. Solorzano-Quintana, Using IoT resources to enhance the accuracy of overdrain measurements in greenhouse horticulture (2016). https:// doi.org/10.1109/CONCAPAN.2016.7942345 3. R.K. Kodali, A. Sahu, An IoT based soil moisture monitoring on Losant platform, in Proceedings of the 2016 2nd International Conference on Contemporary Computing and Informatics, IC3I 2016, pp. 764–768 (2016). https://doi.org/10.1109/IC3I.2016.7918063 4. P.K. Basu, Soil Testing in India (Department of Agriculture and Cooperation Ministry of Agriculture, Government of India, 2011) 5. S. Jayashree, B.S. Manoj, C.S.R. Murthy, A novel battery aware MAC protocol for Ad Hoc wireless networks. Lecture Notes in Computer Science, pp. 19–29 (2004). https://doi.org/10. 1007/978-3-540-30474-6_8 6. I. Abdul Aziz, M.J. Ismail, N. Samiha Haron, M. Mehat, Remote monitoring using sensor in greenhouse agriculture. Int. Symp. Inf. Technol. 2008, 1–8 (2008). https://doi.org/10.1109/ itsim.2008.4631923 7. M. Priyadharshini, U.M. Sindhumathi, S. Bhuvaneswari, N. Rajkamal, K.M. Arivu Chelvan, Automatic irrigation system using soil moisture sensor with big data. Int. J. Eng. Trends Technol. 67(3), 58–61 (2019). https://doi.org/10.14445/22315381/ijett-v67i3p210 8. L. García, L. Parra, J.M. Jimenez, J. Lloret, P. Lorenz, IoT-based smart irrigation systems: an overview on the recent trends on sensors and IoT systems for irrigation in precision agriculture. Sensors 20(4), 1042 (2020). https://doi.org/10.3390/s20041042 9. R. Chavda, T. Kadam, K. Hattangadi, D. Vora, Smart drip irrigation system using moisture sensors, in 2018 International Conference on Smart City and Emerging Technology (ICSCET) (2018). https://doi.org/10.1109/icscet.2018.8537377 10. N. Suma, S.R. Samson, S. Saranya, G. Shanmugapriya, R. Subhashri, IOT based smart agriculture monitoring system. Int. J. Recent Innov. Trends Comput. Commun. 5(2), 177–181 (2017) 11. S. Bavkar, N. Patil, Y. Birje, IoT enabled smart irrigation system using Arduino. SSRN Electron. J. (2020). https://doi.org/10.2139/ssrn.3648829 12. ArduinoBoardUno—Arduino. www.arduino.cc/en/pmwiki.php?n=Main/arduinoBoardUno. Accessed 2 April 2021

74

R. Sharma and K. Kaushik

13. K.V.Y. Ibrahim, Smart farming: IoT based sensors agriculture stick for live temperature and moisture monitoring using Arduino, CLoud computing and solat technology. Indian J. Appl. Res. 1–5 (2019). https://doi.org/10.36106/ijar/4713020 14. Getting Started with the Arduino GSM Shield | Arduino. www.arduino.cc/en/Guide/Arduin oGSMShield. Accessed 2 April 2021

Constrained Optimization-Based Routing for Multipath and Multihop Propagation in WSN Pratham Majumder and Punyasha Chatterjee

1 Introduction With the rapid spread of pervasive computing, the use of sensor nodes in various wireless applications is increasing these days. One major application of wireless communication has been noted in the field of wireless sensor networks (WSNs). The advent of low-cost communication and sensor devices has led to the deployment of a large number of sensors in geographical areas that are not easily accessible, helping in remotely monitoring different activities in those areas. One of the essential criteria of these sensor devices is lifetime sustainability in terms of battery power, e.g., more the residual battery power, more the service life. Usually, these sensors are deployed in large numbers so as to maintain a long battery life, high reliability of service, simultaneously bringing down costs [1]. The nodes are to be declared dead if and when the battery power becomes insufficient to carry out any kind of communication process. As these sensor nodes are deployed in areas mostly inaccessible, hence, the replacement of battery of sensors is practically unfeasible. Therefore, increasing sustainability of sensor nodes is one of the challenging problems in this domain. Hence, optimization of transmission sensor energy has gained paramount importance in all researches that focus on integrating data acquisition [2–4], multidimensional data and query processing [5], media access control [6], energy-efficient routing protocol [7], and much more. This paper presents our proposal toward the implementation of a novel energyaware communication protocol in a multipath propagation-based sensor network.

P. Majumder University of Calcutta, Kolkata, India CMR Institute of Technology, Bengaluru, India P. Chatterjee (B) School of Mobile Computing and Communication, Jadavpur University, Jadavpur, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_9

75

76

P. Majumder and P. Chatterjee

2 Related Works Numerous schemes broadening over various communication layers protocol stack have been investigated in the literature to implement the energy-efficient communication in wireless sensor networks. For example, a common approach in MAC layer-based solutions for reduction of energy consumption is to avoid or reduce collisions so as to minimize packet retransmissions [8]. A large number of such low-energy routing protocols are available in the literature; among which, the most popularly used ones for wireless sensor networks include TEEN [9]. In this algorithm, packets are strategically routed through different intermediate nodes by reducing the transmission range of sensor nodes, thereby saving overall transmission energy. LEACH [10] proposed dynamic clustering algorithm for uniform distribution of load by the nodes in a specific cluster, potentially increases network lifetime. Though cluster reconfiguration at every phase is considered to be a tedious task in terms of implementation cost and complexity. To overcome such issues, this work presents a novel energy-efficient transmission process that employs transmission of ratio-based segregation of packets to the neighbor nodes, which is proportionate to the distance toward cluster head, resulting power depletion of all nodes in a multipath network. We have formulated this problem as a constrained optimization problem [13] to solve the required ratios.

3 Network Model and Optimization Criterion 3.1 Network Model Description and Packet Distribution Policy Figure 1 shows a simple network model of wireless sensor network. Let us suppose that the nodes are capable to transmit message at non-identical distances and able to control their transmission power. For simplicity, we initiate with a 1D nonlinear network. Let us assume that d be the distance between any two nodes in the network. Therefore, the transmission power to send the packet at d distance is proportional to d 2 . We further assume, the packet generation rate of every node be identical, say m-packets in a specific interval of time. There are two possibilities that node 1 can send all its packet to node 5, (i) can send all its packet to the node 5 directly or (ii) can incorporate multihop packet distribution and multipath transmission policy. The direct propagation from source (node 1) to destination (node 5) can cause the packet to be routed over a distance of 2d. Energy requirement for this scheme is proportional to square of the distance, i.e., 4d 2 . In the second scheme, the packet distribution policy can be explained as below:

Constrained Optimization-Based Routing for Multipath …

77

x115 PATH-1 NODE-2 x25

x35

x12

PATH-2

x215

PATH-3

x13

NODE-3

NODE-5 DESTINATION

NODE-1 d

x45

d

NODE-4

SOURCE

x14

x315

Fig. 1 Network architecture with five nodes

• Source node distributes its packet among every path to balance the load at every servicing nodes so that whole energy of the network is minimized. • Every node has to forward not only the packets it generates but also the packets it will receive from previous nodes in such a manner so that energy requirement of every servicing nodes must be equal. In the reference Fig. 1, let us assume, every node (1, 2, 3 and 4) generates mpackets. The total number of packets of each nodes can be expressed as,   Ri = mxi j + m = 1 + xi j

(1)

where x ij is the fraction of packets directly received from node i by node j.

3.2 Energy Calculation Energy of servicing nodes can be expressed as the sum of distance of all interacting nodes multiplied by amount of generated packets transferred by the source node. Ei =

n 

Ri xi j [( j − i)d]2

(2)

j=1,=i

i and j are the generating node and receiving nodes, respectively. Therefore, total energy of the network can be expressed as the sum of energy of individual servicing nodes, E total = E 1 + E 2 + E 3 + E 4 .

78

P. Majumder and P. Chatterjee

3.3 Solving Optimization Problem The objective is to minimize total energy of the network E total keeping the energy expenditure of every node to be equal. We solve the problem using linear optimization technique. Objective Function Objective function of the problem can be represented as f obj = j i mx i j + m. In our problem, i = 1 and j = 1 to 4. Constraints Constraints of the following objective function f obj can be represented as: • Energy expenditure of every node should be equal, i.e., E 1 = E 2 = E 3 = E 4 —sum of all fraction of packet departing from each generating node should be equal to j 1, i.e., i xi j = 1. • The fractional value of packet should lie in between 0 and 1, i.e., 0 ≤ x ij ≤ 1. • Nodes those are directly connected to destination node can send their packets directly without using any other hops. So, the fraction of generated packet for these nodes should be equal to 1. In our problem, x 25 , x 35 , x 45 = 1.

3.4 Experiment II In this case Fig. 2, we add one extra node in the PATH 1. Therefore, generated packet from node 1 (source) will have total six fractional values based on its interaction with

PATH-1

x143

PATH-2 PATH-3 x145 NODE-5

NODE-4

x115

x113

x153

x114 x213 x223

x212

NODE-2

NODE-3 DESTINATION

NODE-1 d

x363

d

NODE-6

x313

Fig. 2 Network architecture with six nodes

x316

SOURCE

Constrained Optimization-Based Routing for Multipath …

79

Table 1 Characteristics of radio devices Device specifications

Maxim 2820

CC1100

CC2420

RFM TR 1000

Data Rate (Kbps)

50

2.5

2.5

25

T p (µs)

20

400

400

40

Vcc (v)

2.7

3.6

3.6

3

IHIGH (mA)

70

30.3

17.4

12

I LOW (mA)

25

1.9

0.426

7 × 10−4

T ON (µs)

3

88.4

0.1

16

other 5 nodes (i.e., 4, 5, 2, and 6, respectively) over 3 different paths. Here, node 1 is the source node, and node 3 is considered to be the description node.

4 Effects of Device Characteristics To estimate the total energy consumed by the radio device used by a sensor node, we consider the commercial radio devices, e.g., CC1100 [15], CC2420 [17], Maxim 2820 [16], and RFM TR 1000 [18] chips that are widely used for low-power, low-cost WSN IoT applications. Such devices, however, consume considerably large power in transmit or receive states compared to that when the radio is in its low-power operation mode. Total energy (E TOTAL ) consumed by a sensor node is calculated by accumulating different power consumption levels associated with base energy (E base ), transmission energy (E trans ), and switching energy (E sw ), respectively, i.e., E TOTAL = E base + E trans + E sw . To transmit n bit binary raw data with a compression factor of r comp and having pON percentage of non-silent symbol in the encoded string, different energy requirements can be represented as follows. Ebase = n × rcomp × tP × ILOW × V DC, Etran = n × rcomp × pON × (tP−tRISE) × I0 × V DC, Esw = n × rcomp × pON × tRISE × (ITRAN−ILOW) × V DC (Table 1).

5 Result Analysis In this section, we describe analysis of energy expenditure of the sensor nodes from two aspects. Firstly, we have demonstrated the theoretical energy calculation solved using our proposed optimization strategy, and after that we will incorporate real-life sensor dataset in our proposed algorithm for performance analysis.

80

P. Majumder and P. Chatterjee

5.1 Theoretical Energy Savings The linear optimization problem is solved using Matlab optimization toolbox [14]. In our work, we have analyzed the overall energy distribution of the network considering distance parameters of the nodes and by varying packet generation type in the network. Table 2 suggests the packet distribution ratios to neighboring nodes considering case 2.1 and case 1.1, respectively. Result shows, the optimized value of the objective function is 5.200 unit and verifying constraint 1; energy requirement of source node is 1.3000 unit. Whereas, if the source node tries to send all m-packet directly to destination node without using any multipath, multihop strategy, then energy requirement will be of 4 unit. So, there is an improvement of 307.69 times of the energy requirement by the source using multipath strategy. Considering nonidentical distance between hops in the network, the energy expenditure is increased by 46%. For Fig. 2, total energy of the network is 4.9444 unit, which is lowering by 5.76% compared to Fig. 1. Table 3 describes optimized energy expenditure of all nodes in the network is found to be 0.9887 unit, which is lowering by 32.65% compared to Fig. 1. And above all, there is an improvement of 404.57 times with respect to the overall transmission energy using this multipath technique compared to direct transmission.

5.2 Simulation on Real-Life Sensor Data This dataset is assembled from the Activity 2.3 CityPulse EU FP7 project [19], which concerns with real-time IoT stream processing and large-scale data analytics for smart city applications. Performance Comparison We now compare the effectiveness of our proposed scheme with popular load balancing protocols, e.g., load-balanced routing (LBR) protocol [11], energy efficient sleep awake aware (EESAA) protocol [12]. Figure 3 shows the overall communication energy profile, exploiting three individual load balancing techniques on commercial radios (CC1100, CC 2420, Maxim 2820, and RFM TR 1000) using device specification mentioned in Table 1. From the Fig. 4, it is clear that a significant energy savings are achieved for transmission energy (E tran ) and base energy (E base ) in the entire communication energy profile for both the transceiver radios CC1100 and Maxim 2820. Usually, the fall time t FALL and rise t RISE time are very small, and the corresponding energy for switching from transmission state to idle state and vice-versa is small enough to be neglected.

4

No direct

2d unit

5.12

Source + 2 nodes

√

5.16

Source + 1 node

Non-identical d,

5.2

Source

Identical (d) unit

9.76 9.68 9.52 6

Source

Source + 1 node

Source + 2 nodes

No direct

Etotal

1.5

2.38

2.42

2.44

1

1.28

1.29

1.3

e1

1.5

2.38

2.42

2.44

1

1.28

1.29

1.3

e2

Estimated energy (Unit)

Packet generation/transmission type

Distant metric

Table 2 Energy calculation of each nodes

1.5

2.38

2.42

2.44

1

1.28

1.29

1.3

e3

1.5

2.38

2.42

2.44

1

1.28

1.29

1.3

e4

0.33

0.19

0.21

0.22

0.33

0.28

0.29

0.3

x 12

0.33

0.19

0.21

0.22

0.33

0.28

0.29

0.3

x 13

0.33

0.19

0.21

0.22

0.33

0.28

0.29

0.3

x 13

-

0.14

0.18

0.33

-

0.04

0.05

0.1

x 1 15

Optimized packet fraction

-

0.14

0.18

-

-

0.04

0.05

-

x 2 15

-

0.14

-

-

-

0.04

-

-

x 3 15

Constrained Optimization-Based Routing for Multipath … 81

82 Table 3 Optimized energy and optimized energy of source

P. Majumder and P. Chatterjee Figure ID

Estimated energy

Optimized energy of source

2

5.2

1.3

3

4.94444444436761

0.9887

Fig. 3 Comparison of overall communication energy for different schemes

Fig. 4 Communication energy profile for different commercial radios

6 Conclusion Most of the energy-efficient communication protocols in the network layer are based on hierarchical routing. These protocols rely on fragmentation of network based on number of clusters having a distinct or dynamic cluster heads. In a cluster, all the sensor nodes send their packets to the cluster head followed by multihop communication, i.e., every node transmits their data packet to its neighboring hops, toward the cluster head. This packet transmission strategy makes the nodes near to cluster heads completely overloaded, resulting in early expiration of battery. To mitigate this abnormal distribution of packets to the nearby nodes to cluster heads by keeping the packet generation rate of all the nodes constant, we have introduced a novel packet distribution policy so that powers at all nodes are uniformly dissipated for a nonlinear network. Incorporation of constrained optimization problem in packet distribution ratios to each and every neighbor nodes toward cluster head which keeps

Constrained Optimization-Based Routing for Multipath …

83

the power dissipation rate of the participated nodes to be equal. Our analysis shows, for a nonlinear network, this optimization strategy can be effectively employed for a cluster size of six nodes.

References 1. F. Kuhn, T. Moscibroda, R. Wattenhofer, Initializing newly deployed ad hoc and sensor networks, in Proceedings of the 10th Annual International Conference on Mobile Computing and Networking, pp. 260–274 (2004) 2. A. Bhattacharya, P. Majumder, K. Sinha, B.P. Sinha, K.V.N. Kavitha, An energy-efficient wireless communication scheme using quint fibonacci number system. Int. J. Commun. Netw. Distrib. Syst. 16(2), 140–161 (2016) 3. P. Majumder, K. Sinha, B.P. Sinha, DCVNS: a new energy efficient transmission scheme for wireless sensor networks, in 2018 IEEE 88th Vehicular Technology Conference (VTC-Fall), (IEEE, 2018), pp. 1–5 4. P. Majumder, P. Chatterjee, K. Sinha, Run length distribution based block coding scheme for sustainable IoT applications, in 2020 2nd PhD Colloquium on Ethically Driven Innovation and Technology for Society (Ph. D. EDITS), (IEEE, 2020), pp. 1–2 5. S. Cheng, Z. Cai, J. Li, Curve query processing in wireless sensor networks. IEEE Trans. Veh. Technol. 64(11), 5198–5209 (2014) 6. A. Kumar, M. Zhao, K.-J. Wong, Y.L. Guan, P.H.J. Chong, A comprehensive study of IoT and WSN mac protocols: research issues, challenges and opportunities. IEEE Access 6, 76228– 76262 (2018) 7. G.S. Brar, S. Rani, V. Chopra, R. Malhotra, H. Song, S.H. Ahmed, Energy efficient directionbased PDORP routing protocol for WSN. IEEE Access 4, 3182–3194 (2016) 8. I. Demirkol, C. Ersoy, F. Alagoz, MAC protocols for wireless sensor networks: a survey. IEEE Commun. Mag. 44(4), 115–121 (2006) 9. A. Manjeshwar, D.P. Agrawal, TEEN: a routing protocol for enhanced efficiency in wireless sensor networks. 1(2001), 189 (2001) 10. S.V. Ley, I.R. Baxendale, R.N. Bream, P.S. Jackson, G. Andrew, Multi-step organic synthesis using solid-supported reagents and scavengers: a new paradigm in chemical library generation. J. Chem Soc. Perkin Trans 1(23), 3815–4195 (2000) 11. S. Agarwal, A. Das, N. Das, An efficient approach for load balancing in vehicular adhoc networks, in IEEE International Conference on Advanced Networks and Telecommunications Systems (ANTS), pp. 1–6 (2016) 12. A. Ennaciri, M. Erritali, J. Bengourram, Load balancing protocol (EESAA) to improve quality of service in wireless sensor network. Proc. Comput. Sci. 151, 1140–1145 (2019) 13. R. Farmani, J.A. Wright, Self-adaptive fitness formulation for constrained optimization. IEEE Trans. Evol. Comput. 5, 445–455 (2003) 14. Matlab Optimization Toolbox. https://www.mathworks.com/optimization.html 15. CC1100: https://www.ti.com/product/CC1100 16. Maxim2820: https://datasheetspdf.com/pdf/497162/Maxim/MAX2820/1 17. CC2420: https://www.ti.com/product/CC2420 18. RFM TR 1000: https://html.alldatasheet.com/html-pdf/ 106155/RFM/ TR1000 /53/1/TR1000.html 19. Dataset: http://iot.ee.surrey.ac.uk:8080

Performance Comparison of Fuzzy Logic and Evolutionary Algorithm-Optimized Controller for a Multi-area Power System Parisa Latief Khan, Zahid Farooq, Sheikh Safi ullah, and Satish Saini

1 Introduction With the rise in the standard of living, the need of electricity has increased tremendously and with it the complications and nonlinearities of the electricity supply network. The reliable operation of any electricity supply network is dependent on its frequency stability. In addition to this, power supplied to consumers must be adequate and economical. An electric network consists of many linked systems. If dereliction due to frequency variation occurs in system components, the disturbing intrusions and temporary or permanent cessation adversely affect the whole network and also the equipment on the consumers’ end. LFC is performed because steady-state error can lead to frequency change which is unacceptable [1]. LFC precariously stabilizes the demand and generation to keep frequency digresses within acceptable range. Over recent times, many researchers have worked on the problems of LFC of interconnected electricity supply networks. In [2], optimal output feedback controller is used in a multi-area system which consists of HVDC link in parallel with AC link for stabilizing frequency swerves. Multi-area-linked systems in contrast to several customary controllers were conveyed through Saikia et al. [3]. Load frequency control investigation with multi-area network with controller (PID) has been conveyed by Hamed et al. [4]. Golpira and Bevrani [5] contemplate time-delay and generation rate constant as practical limitation for generating unit. When confronted with network constraints, the most favorable LFC solution [6] was traversed for integrated electricity supply network. The researcher has conveyed in [7] the system dynamics control of twoarea electricity supply network contemplating solar thermal. The combination of utility network with solar thermal system (STS) requires to be scrutinized more. The P. L. Khan (B) · S. Saini Department of Electrical Engineering, RIMT University, Mandi Gobindgarh Punjab, India e-mail: [email protected] Z. Farooq · S. Safi ullah Department of Electrical Engineering, NIT Srinagar, Jammu and Kashmir, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_10

85

86

P. L. Khan et al.

conventional controllers have fixed gains, and they abstain from performing well over varying operating conditions and give a poor dynamic response. Now, in order to sort this issue, fuzzy system was introduced, which has adaptive scheduled gains. The evolutionary fuzzy hybrid with proportional and integral controller for frequency manipulation is examined by Juang and Lu [8] Frequency load control is utilized with error and change in error as inputs to hone gains of controller (PI). Mohanty et al. [9] developed as well as inspected a controller (fuzzy PID) for the LFC unit. In a micro-grid, for the dynamic control of frequency, Annamraju and Nandiraju [10] propounded a fuzzy-tuned multi-level PID controller. The FLC-based multistage controller when compared with the traditional PID controller shows better response. The result was propounded by Annamraju and Nandiraju [10]. Findings from Farooq et al. [11] show a hybrid multi-area network containing electric vehicle in all units using an optimization technique, namely magnetotactic bacteria optimization. Outcome shows the better performance of MBO-optimized IDD controller when compared to other traditional controllers. Keeping in view the above mentioned findings, this paper will encompass following work: 1. 2. 3.

To synthesize a two-area electricity supply network for load frequency control operation encompassing hydrothermal generating stations. To apply firefly algorithm in order to optimize the controllers and find the optimal controller. For the best controller found, fuzzy logic is developed. To contrast the conventional and fuzzy-optimized controllers under load perturbation.

2 Two-Area Electricity Supply Network The tie line or a power line connects two single areas of a two-area electricity supply network. The user pool is fed by each area, and the connecting line provides a path for power to flow between 2 areas. 1:4 is the capacity ratio of area 1 and area 2 of Fig. 1. The traditional thermal plant has generation rate constant of 3% per minute. Integral square error is calculated as 

 ISE =

 f 12 dt+

  f 22 dt +

Ptie212 dt.

(1)

3 Proportional–Integral–Double-Derivative Controller (PIDD) Figure 2 shows the configuration of PIDD controller. The effect of rise time is minimized by proportional mode, but results in the increment of steady-state error which

Performance Comparison of Fuzzy Logic and Evolutionary …

Fig. 1 Two-area transfer function model of power system Fig. 2 Structure of PIDD controller

87

88

P. L. Khan et al.

is nullified by integral approach, but the worsening of transient response is experienced. Furthermore, the system stability is enhanced by derivative mode, truncating the overshoot, and ameliorating the transient response. The stability of the system under study is refined by the PIDD controller and helps in terms of settling time in more excellent way when collated against traditional PID controllers.

4 Methodology The controllers in each area are optimized using firefly technique and fuzzy logic.

4.1 Overview of Firefly Algorithm To solve the complex optimization problems, the most established and widespread techniques are those of bio-inspired, due to their lucidity, proficiency, and reliability. Instigated by Yang X. S. is one such known bio-technique called firefly algorithm, inspired by the flashing behavior of fireflies.

4.2 Fuzzy Logic Control Introduced by Lotfi Zadeh, fuzzy logic provides a very valuable flexibility for reasoning and has plenty of electricity supply network implementations. It operates on the modus operandi of information acquisition. Fuzzy function associated and to optimize the control gain parameters, fuzzy IF–THEN rules are utilized, mentioned in Table 1. Table 1 Fuzzy rule base developed in Matlab ACE/DACE

VVS

VS

S

ZR

B

VB

VVB

VVS

VVB

VVB

VVB

VB

VB

VB

ZR

VS

VVB

VB

VB

VB

B

ZR

S

S

VVB

VB

B

B

ZR

S

VS

ZR

VB

VB

B

ZR

S

VS

VS

B

VB

B

ZR

S

S

VS

VVS

VB

B

ZR

S

VS

VS

VS

VVS

VVB

ZR

S

VS

VS

VVS

VVS

VVS

Performance Comparison of Fuzzy Logic and Evolutionary …

4.2.1

89

Functioning of Fuzzy-PIDD Controller

For fuzzy-PIDD controller, as input, ACE and DACE are applied. The efficiency of controller is dependent upon rule base developed, membership function chosen, and the defuzzification method used. The optimum controller gains are decided by each area individually. To reduce unpremeditated power interchange in the course of frequency oscillations, it is imperative in two-area units that both of them carry their own load.

5 Outcomes and Analysis In MATLAB Simulink environment, a two-area electricity supply network embodying wind thermal and hydrothermal power plants is modeled and simulated.

5.1 Choosing an Optimal Secondary Controller For the propounded two-area electricity supply network, sundry controllers like ID, PI, PID, and PIDD are examined as secondary controllers. For the analysis of the system, all four controllers are observed separately one at a time. Optimal values are obtained. The firefly-optimized controller gains are delineated in Table 2. The dynamic responses of both areas are collected. These responses are compared and are shown in Fig. 3 in order to study their variations. PIDD is found as optimal one after the response evaluation. After inspection of the results, it is inferred that the PIDD controller performs with ACE very less when compared with other classical controllers. After comprehensive and thorough observation of the responses recorded in Fig. 3 shows that PIDD is far supreme than the traditional controllers. Table 2 Various controller gains depicting firefly-optimized values Controllers

Kp

ID

Ki

Kd

N

0.8147 0.9058

0.1270 0.9134

0.6324 0.0975

PI

0.8355 1.9611

0.6029 1.4022

PID

0.7197 0.6865

1.3229 2.2499

0.8191 1.2088

0.34 0.19

PIDD

0.5102 1.0119

1.3982 1.7818

1.9186 1.0944 0.2772 0.2986

0.5150 1.6814 0.5086 1.6286

90

P. L. Khan et al.

(a)

Δf1

(b) Δf2

(c) ΔPtie12 Fig. 3 Response contrast of different controllers

5.2 Firefly Optimized PIDD Controller and Fuzzy-Optimized Controller Comparison We concluded that PIDD controller is the most efficient controller in preceding part, so now, the working of fuzzy-optimized PIDD controller is investigated. Figure 4 demonstrates the working of fuzzy-PIDD as secondary controller. The dynamic system comparison is done with firefly-optimized PIDD controller. The response comparison in figure clearly sets out the precedence of fuzzy-PIDD over firefly controller.

5.3 Sensitivity Analysis During various system variations, how efficiently controllers will maintain stability condition is checked by performing sensitivity analysis. At 30 and 70% loading, fuzzy-optimized controller has a slight edge over firefly-optimized controller in terms of settling time, overshoots, and undershoots (Fig. 5).

Performance Comparison of Fuzzy Logic and Evolutionary …

(a) Δf1

91

(b) Δf2

(c) ΔPtie12 Fig. 4 Firefly-optimized PIDD versus fuzzy-PIDD at nominal conditions

(a) Δf1 (30 % loading)

(b) Δf1 (70% loading)

Fig. 5 Comparison of firefly-optimized PIDD with fuzzy-optimized PIDD at different loading

6 Conclusion This research examines the control of load frequency in a two-area electrical supply network using conventionally optimized controllers and fuzzy-optimized controllers. In investigations of electricity supply network, the firefly optimization technique is used to aid in the optimization of secondary controller gains. In comparison to other traditional controllers, the system performance study reveals the dominance of the firefly-optimized PIDD controller. When it comes to overshoots and settling time, the designed fuzzy logic control-based PIDD controller outperforms the firefly-optimized PIDD controller by a moderate margin. The sensitivity analysis is performed at 30 and 70% loadings which emphasizes the overall and effective performance of PIDD controller optimized using fuzzy logic.

92

P. L. Khan et al.

References 1. S. Sfiullah, A. Rahman, S.A. Lone, State-observer based IDD-controller for concurrent frequency–voltage control of a hybrid power system with electric vehicle uncertainties. Int. Trans. Electr. Energ. Syst. (2021). https://doi.org/10.1002/2050-7038.13083 2. K.P.S. Parmar, S. Majhi, D.P. Kothari, Load frequency control of a realistic power system with multi-source power generation. Int. J. Electr. Power Energy. Syst. 42, 426–433 (2012) 3. L.C. Saikia, J. Nanda, S. Mishra, Performance comparison of several classical controllers in AGC for multi-area interconnected thermal system. Elsevier IJEPES 33(3), 394–401 (2011) 4. S. Hamed, V. Behrooz, E. Majid, A robust PID controller based on imperialist competitive algorithm for load frequency control of electricity supply networks. ISA Trans. 52(1), 88–95 (2013) 5. H. Golpira, H. Bevrani, Application of GA optimization for automatic generation control design in an interconnected electricity supply network. Energ. Conver. Manage. 52, 2247–2255 (2011) 6. R. Patel, C. Li, L. Meegahapola, B. McGrath, X. Yu, Enhancing optimal automatic generation control in a multi-area electricity supply network with diverse energy resources. IEEE Trans. Power Syst. 34(5), 3465–3475 (2019) 7. A. Rahman, L.C. Saikia, N. Sinha, Automatic generation control of an interconnected two-area hybrid thermal system considering dish Stirling solar thermal and wind turbine system. Renew. Energy 105, 41–54 (2017) 8. C.F. Juang, C.F. Lu, Load-frequency control by hybrid evolutionary fuzzy PI controller. IEE Proc. Gener. Transm. Distrib. 153(2), 196–204 (2006) 9. P.K. Mohanty, B.K. Sahu, T.K. Pati, S. Panda, S. Kumar, Design and analysis of fuzzy PID controller with derivative filter for AGC in multi-area interconnected electricity supply network. IET Gener. Transm. Distrib. 10(15), 3764–3776 (2016) 10. A. Annamraju, S. Nandiraju, A novel fuzzy tuned multistage PID approach for frequency dynamics control in an islanded microgrid. Int. Trans. Electr. Energ. Syst. 30(12) (2020). https://doi.org/10.1002/2050-7038.12674 11. Z. Farooq, A. Rahman, S.A. Lone, Load frequency control of multi-source electrical electricity supply network integrated with solar-thermal and electric vehicle. Int. Trans. Electr. Energ. Syst. e12918 (2021). https://doi.org/10.1002/2050-7038.12918

Leveraging CNNs for Real-Time Pothole Detection Chaithra Reddy Pasunuru and Kruthika Muthireddy

1 Introduction Cambridge Dictionary [1] defines a pothole as “a hole in the road surface that develops from progressive degradation caused by vehicles and/or weather.” Regular transportation is severely disrupted by potholes. Potholes can cause harm to vehicles. Potholes provide a major challenge for people who are visually impaired while driving. Accidents can occur when drivers lose control of their vehicles [2]. Therefore, to get rid of this problem either potholes must be repaired as soon as they appear or drivers and passers-by must be aware of them during navigation so that they can be avoided or passed safely. Although the first one is very inefficient and troublesome, research has taken place to detect potholes and mark the location where they occurred [2–6]. The collected data then is sent to the authority so that the hole is repaired soon. Second one is also researched on [7, 8], to inform navigators about the potholes appearing on the road surface in front of them. This research is also focused on the second one, i.e., detect potholes before you hit them and try to avoid if possible or hit safely. The research work targets to develop a system for automatic detection of potholes in real time and generate signal to warn about them. Researchers used an approach known as transfer learning to develop the detector model in this study [9]. The model was trained and tested on an annotated image dataset. Convolutional neural network model MobileNet [10] was used to extract features. The Single-Shot Multibox Detector (SSD) [11] was specifically designed to identify and locate potholes. Pothole detection, location, and tracking may be done in real time using video frames even if the model was trained on an image dataset.

C. R. Pasunuru (B) · K. Muthireddy Neil Gogte Institute of Technology, Peerzadiguda, Uppal, Hyderabad, Telangana 500058, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_11

93

94

C. R. Pasunuru and K. Muthireddy

An Android application can be developed using the model which can detect, localize, and track the potholes in the video coming from its camera. It can also generate warning signal as long as pothole is detected in the video stream. Thus, it would be very helpful for blind people to navigate. Localization capability of potholes in video stream can be very useful in case of automated vehicle driving. In this paper, explanation and discussion of technical requirements, materials and methods, results, and evaluation of our research are given following a review of existing and related works. After all, conclusion and future works are given followed by the bibliography.

2 Related Works Using an Android-based smartphone, Mednis et al. developed a mobile sensing system for the road [7, 12] that could detect inconsistencies. Testing was carried out on a 4.4 km long track with 10 continuous laps. True-positive rate (TPR) [13] was about 90. This method could result in wrong information for the cases such as: • It detects hinges as well as joints of the road [13] as pothole event if it is not the case though. • It fails to detect potholes which are in the middle of the lane. According to Chang et al., the scanning and extraction of focusing on some specific distress elements were acquired along with accurate 3D cloud points and their height using a grid-based approach. A precise and automated assessment of the distress was possible utilizing this method [13, 14]. Using 3D transverse scanning technology, Li et al. developed an inspection system [4] that rapidly detects and identifies distress characteristics such as potholes, shoving, and rutting. A digital camera and a laser line projector employ infrared wavelengths to identify potholes and other road hazards [4]. To detect and evaluate potholes, Joubert and his colleagues [5, 15] used the Kinect sensor and a high-speed USB camera [5]. Potholes have already been studied with Kinect in trials. As a bonus, this strategy is cost-effective. Using computer vision, Buza et al. [5] suggested an un-supervised technique without expensive equipment, filtering, or training phases. The potholes in the target image were detected and identified using conventional image processing and clustering technologies [13]. Their method consists of three steps as given below: (1) (2) (3)

Segmentation of target images Shape and feature extraction Detection and identification.

They were able to get an accuracy of 81% [13] using the method described above, which may be utilized as an approximate estimate for pothole repairs.

Leveraging CNNs for Real-Time Pothole Detection

95

Huidrom et al. [16] introduced a video frame analysis approach for identifying potholes, fissures, and patches of pavement. Their results showed that the DFS algorithm was useful in classifying video samples as either anxious or distressed [13, 16]. As a way to detect potholes and assess their severity, Jog et al. [17] showed a system for 2D detection and 3D reconstruction. They used a car-mounted video camera to capture footage of the street. They were able to determine the pothole’s depth, width, and number using this method. Koch et al. [18] came up with a solution based on a single frame of video material recorded by the camera. It was not possible to determine the pothole’s magnitude by analyzing pavement video frames [13]. Koch et al. created a novel composition signature for excellent pavement zones in order to better detect potholes. Computer vision was utilized to identify potholes in all of the video frames.

3 Technical Requirements The following technical requirements were identified for the research, preliminary system development, testing, and deployment: (1) (2) (3)

A dataset of images is required to be fed the model for training purpose. In case of video, frames can be extracted [19] as images and must be annotated. Dataset images must be annotated in such a way that the regions of interest, i.e., image areas having pothole, must be selected, e.g., as bounding boxes [20]. For model train-up, a digital workstation [21] computer is recommended with “at least” following configurations: • CPU: 2.2 GHz (6 Gen) • RAM: 16 GB (2400 MHz DDR4) • GPU: – VRAM: 12 GB – Computing [22] capability: 5 – CUDA [23] support. However, model developed in this research was trained on the Google Colaboratory [24].

(4)

Android [25] smartphone with speaker and camera support. Android OS version ≥ 6 (Marshmallow, API 22).

96

C. R. Pasunuru and K. Muthireddy

4 Materials and Methods 4.1 The Dataset We have used a novel (custom) dataset of images. A portion (57.1%) of the dataset images was scrapped from Google Image Search. The rest (42.9%) was captured with an Android smartphone camera from some damaged roads in the campus of University of Chittagong. The dataset contains a total of 665 images having a total of 1740 annotated potholes. Each image was annotated with bounding boxes around the regions having potholes. For this purpose, labelImg [26] was used as the annotation tool. The training set contains 532 (80%) images, and the test set contains 133 (20%) images. The potholes were divided into three categories: (1) (2) (3)

Small Medium Large.

The sizes of the potholes were measured from the occupied pixels in the sample images. The number of pixels occupied was calculated after resizing the longer side of the images to 300 px keeping the aspect ratio of the shorter side. The number of pixels occupied by the different categories of pothole is given in Table 1. The distribution of different categories of potholes over the train and test data is shown in Fig. 1. The dataset can be found in [27]. Table 1 Different categories versus number of pixels occupied

Pothole category

Occupied pixels

Small

area ≤ 322

Medium

322 < area ≤ 962

Small

area > 962

train

185

large

49 650

medium

135

small

606 115 1441

total

299

0

500

1000

Number of potholes

Fig. 1 Distribution of different categories of potholes

1500

test

Leveraging CNNs for Real-Time Pothole Detection

97

4.2 Materials Used Basically, a transfer learning [28] approach was followed in this research to build the target model. Because image data was used, we needed to extract the features of the regions having potholes. To do so, existing pre-trained MobileNet [10] model was used as the feature extractor. For the detection task, i.e., to predict the bounding boxes around potholes in images, pre-trained Single-Shot Multibox Detector (SSD) [11] model was fine-tuned. Both the MobileNet [10] and SSD are based on convolutional neural network. Although MobileNet [10] has a lower classification performance compared to other larger models [29], it is worthwhile for “real-time” operation purpose on low-end devices like Android mobile phones and embedded devices. Architecture of the SSD MobileNet [10, 11] is shown in Fig. 2. With the help of TensorFlow [30] Object Detection API, the model was trained using the prepared dataset with some fine-tuned configurations. Figure 3 illustrates the interaction of different components during training. The whole model was trained and evaluated on the Google Colaboratory [24].

Fig. 2 Architecture of the SSD MobileNet model

Fig. 3 Interaction of different training components

98

C. R. Pasunuru and K. Muthireddy

4.3 Training the Model Data Augmentation: The training images were augmented in different ways. The following augmentations were applied on the training images online, i.e., during training phase: • • • •

Horizontal flip Random crop Resize keeping the aspect ratio Zero padding.

Input Resizing: The shape of the input layer of our model is 300 × 300 × 3; that is, it takes RGB images of 300 px width and 300 px height. But the dataset contains different sizes of images, even a large portion of the images is not square-shaped. Therefore, it was obvious to make the input shape similar as the model’s input shape. To address this, we have resized the input such a way that the longer side was resized to 300 px and the shorter size was resized keeping the aspect ratio intact. The remaining space of the shorter side was filled with zeros. Preparing a Pre-trained Model: We have followed transfer learning approach [28] in this research experiment. So, we required a pre-trained model ready to be used as a base model. There are many pre-trained models trained on different datasets with different metrics. These pre-trained models can be found in TensorFlow [30] Model Zoo [29] (Fig. 4). We have used the SSD [11] MobileNet [10] V2 pre-trained model which was trained on Microsoft COCO dataset [31]. Every pre-trained model in the TensorFlow Model Zoo [29] contains a “pipeline.config” file which is used as a configuration file during training and evaluation. We have removed/changed/added several configuration options in that file. Our changes are shown in Table 2. The configuration file consists of some of the options: • Number of classes or labels

Fig. 4 Training steps versus decay of learning rate

Leveraging CNNs for Real-Time Pothole Detection Table 2 Fine-tuned configurations

• • • •

99

Configuration option

Changed value

Input shape

(300, 300, 3)

Num classes

1

Batch size

24

Image resizer

Keep aspect ratio resizer

Min dimension

300

Max dimension

300

Pad to max dimension

True

Initial learning rate

0.005

Decay steps

6000

Decay factor

0.85

Quantization delay

30,000

Activation functions Prepossessing and data augmentation options Learning rate and batch size Evaluation methods.

Prepare TensorFlow Models Repository: For the purpose of training as well as validation of our model, we reused the existing libraries, packages, and codes from the TensorFlow “models” repository [32]. The repository contains almost everything we need for training and evaluation using TensorFlow API. The mentioned repository can be found in GitHub [32]. Run Training Process: TensorFlow “models” repository provides necessary Python code for the whole training process with zeros. It also saves the checkpoints of the model in training at a regular interval of time. It starts training from the saved checkpoints if the process is interrupted [32]. It saves the scalar and graphical values, i.e., the results of training and evaluation in “tfevent” files which can be monitored in TensorBoard [33]. Run Evaluation Process: After the training process ends, we can run the evaluation on the validation dataset. TensorFlow “models” repository provides the tools for the evaluation process too. The evaluation is triggered by the training tool automatically whenever it saves a checkpoint of the model [32]. This helps to monitor the performance of the model at different levels of training steps and epochs.

100

C. R. Pasunuru and K. Muthireddy

4.4 Evaluation Methods Intersection Over Union (IoU): To evaluate the performance of the model, precision and recall values were used. To take a detected bounding box as true positive, different intersection over union (IoU) thresholds were considered (Fig. 5). The following IoU thresholds were considered: • IoU@50% • IoU@75% • IoU@50%:5%:95%. IoU@50%:5%:95% is a dynamic measurement where 10 IoU thresholds are considered, starting from 50% up to 95% with an interval of 5%, and the following thresholds are found—50%, 55%, 60%, 65%, 70%, 75%, 80%, 85%, 90%, 95%. Values at these 10 thresholds were then averaged to get the value for IoU@50%:5%:95%. Here, taking the above IoU thresholds into account, TP FP TN FN

Number of true-positive predictions Number of false-positive predictions Number of true-negative predictions Number of false-negative predictions

Precision Calculation: Precision shows how much the predictions are correct [34], that is, percentage of correct predictions among all the positive-predicted values. Precision =

TP TP + FP

(1)

Average precision is the area under precision–recall curve. 1 AP =

p(r )dr 0

Fig. 5 Measuring the intersection over union (IoU)

(2)

Leveraging CNNs for Real-Time Pothole Detection

101

Here, r represents recall and p represents precision as a function of r. Therefore, p(r) means “precision at recall r.” In this research, only a single class is available and it is labeled “pothole”; therefore, mean average precision (mAP) is same as average precision (AP). For multi-class, mAP is the mean of average precisions of all individual classes. N 1  mAP = APi N i=1

(3)

Here, mAP is mean average precision, N is the number of class labels, and APi is the average precision for ith class. We considered and calculated mean average precision for different sizes of potholes as well as for different IoU thresholds: • • • • • •

mAP@50% IoU. mAP@75% IoU. mAP@50%:5%:95% IoU for small sizes. mAP@50%:5%:95% IoU for medium sizes. mAP@50%:5%:95% IoU for large sizes. mAP@50%:5%:95% IoU for all sizes.

Recall Calculation: Recall means true-positive rate and also known as sensitivity, a well-known parameter (measurement) for model evaluation in the context of classification [35]. We considered and calculated average recall for different sizes of potholes as well as for different maximum detection levels: • • • • •

AR@10 detections AR@100 detection for small sizes AR@100 detection for medium sizes AR@100 detection for large sizes AR@100 detections for all sizes

Average recall is same as recall in this research because there is a single class label in the dataset, namely “pothole.” All of these recall values were calculated for 50%:5%:95% IoU threshold, i.e., using MS COCO metrics [31]. Recall =

TP TP + FN

(4)

Average recall is the mean of the recall values of all individual classes. AR =

N 1  ri N i=1

(5)

102 Table 3 Precision at different IoU thresholds

C. R. Pasunuru and K. Muthireddy IoU threshold

Precision on validation data

50%

0.65

75%

0.36

50%:5%:90%

0.37

Here, AR is average recall, N is the number of class labels, and r i is the recall for ith class.

5 Results and Discussion The training process was run for more than 100,000 steps with a batch size of 24. Then, evaluation process was run with the help of the TensorFlow “models” repository. The evaluation process used the performance metrics following the Microsoft Common Objects in Context (COCO) [31]. Precision values found on the test dataset after more than 100,000 steps of training are shown here in tabular form as well as using bar charts.

5.1 Average Precision for Different IoU Thresholds In Table 3, the model’s performance is shown as average precision considering different minimum IoU thresholds as the threshold for true-positive detection.

5.2 Average Precision for Different Pothole Sizes In Table 4, the model’s performance is shown as average precision considering different categories of pothole sizes in the test dataset. Figure 6 summarizes the mean average precision values considering different IoU thresholds and different pothole size categories. Table 4 Precision for different pothole sizes

Area sizes

Precision on validation data

Small

0.07

Medium

0.26

Large

0.58

Leveraging CNNs for Real-Time Pothole Detection

103

Fig. 6 Average precision on validation data

Table 5 Recall for different detection limits

Maximum detections

Recall on validation data

1

0.27

10

0.44

100

0.49

5.3 Average Recall at Different Detection Limits We have taken the recall values at different limits of maximum detections. More specifically, we have taken the following limits for the calculation of recall values: • Average recall at maximum of 1 detection (AR@1) • Average recall at maximum of 10 detections (AR@10) • Average recall at maximum of 100 detection (AR@100) In Table 5, average recall values are shown for different detection limits.

5.4 Average Recall for Different Pothole Sizes In Table 6, average recall values are shown for different categories of pothole sizes. Here, all the values are calculated at the threshold of IoU@50%:5%:95%. Table 6 Recall for different pothole sizes

Area sizes

Recall on validation data

Small

0.18

Medium

0.44

Large

0.67

104

C. R. Pasunuru and K. Muthireddy

Fig. 7 Average recall on validation data

Figure 7 summarizes all the average recall values as measures of performances for the model. Here, all detections were calculated with IoU@50%:5%:95% threshold. From Figs. 6 and 7, it is clear that the model performs better for larger potholes. Although the precision values are not very high, it is worth accepting because of such a lightweight model like MobileNet SSD which is built targeting the low-end mobile devices [10].

6 Conclusion and Future Works Considering the severity of potholes, a system was developed to automatically detect them in real time and generate warning signals to help for avoidance. A deep convolutional neural network model named “MobileNet” was used as feature extractor applying a transfer learning mechanism. “Single-Shot Multibox Detector” model was fine-tuned for pothole detection and localization. An Android application is detailed which can easily be developed using the model which can detect, localize, and track the potholes in real time analyzing the video frames. The app would generate warning signal as long as it detects pothole in the video frames coming from its camera. So, the system can be used by visually impaired people to avoid potholes while navigating. It can also be used in automated vehicle driving. In the future, more research would be accomplished to make the system more accurate, more robust, more intelligent. We are also in the process of developing the Android application in order to bring the use of this research to life. Research for estimating the distance of the detected potholes would be done in near future. Measuring the severity of the potholes would also be included in the future works. Making the model aware of different sizes of potholes may be included as future tasks. Also, measurement of dimension like area, depth, etc., of the detected pothole would be done in the future research.

Leveraging CNNs for Real-Time Pothole Detection

105

References 1. E. Walter, Cambridge Advanced Learner’s Dictionary. Cambridge university press (2008) 2. B. Yu, X. Yu, Vibration-based system for pavement condition evaluation. 8, 183–189 (2006) 3. K. De Zoysa, C. Keppitiyagama, S. Weerathunga, A public transport system based sensor network for road surface condition monitoring, p. 9 (2007) 4. Q. Li, M. Yao, X. Yao, B. Xu, A real-time 3d scanning system for pavement distortion inspection. Measure. Sci. Technol. 21, 015702 (2009) 5. E. Buza, S. Omanovic, A. Huseinnovic, Stereo vision techniques in the road pavement evaluation, in Proceedings of the 2nd International Conference on Information Technology and Computer Networks, pp. 48–53 (2013) 6. C. Koch, I. Brilakis, Pothole detection in asphalt pavement images. Adv. Eng. Inform. 25, 507–515 (2011) 7. A. Rao, J. Gubbi, M. Palaniswami, E. Wong, A vision-based system to detect potholes and uneven surfaces for assisting blind people. 5, 1–6 (2016) 8. A. Danti, J. Kulkarni, P. Hiremath, An image processing approach to detect lanes, pot holes and recognize road signs in Indian roads. Int. J. Model. Optim. 2, 658–662 (2012) 9. S.B. Kotsiantis, I. Zaharakis, P. Pintelas, Supervised machine learning: a review of classification techniques. Emerg. Artif. Intell. Appl. Comput. Eng. 160, 3–24 (2007) 10. A.G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, H. Adam, Mobilenets: efficient convolutional neural networks for mobile vision applications, arXiv preprint arXiv:1704.04861 (2017) 11. W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, A.C. Berg, Ssd: single shot multibox detector, in European Conference on Computer Vision (Springer, 2016), pp. 21–37 12. A. Mednis, G. Strazdins, R. Zviedris, G. Kanonirs, L. Selavo, Real time pothole detection using android smartphones with accelerometers, in 2011 International Conference on Distributed Computing in Sensor Systems and Workshops (DCOSS), pp. 1–6 June 2011 13. T. Kim, S. Ryu, Review and analysis of pothole detection methods. J. Emerg. Trends Comput. Inf. Sci. 5, 603–608 (2014) 14. K.-T. Chang, J. Chang, J.-K. Liu, Detection of pavement distresses using 3d laser scanning technology. 6, 1–11 (2005) 15. D. Joubert, A. Tyatyantsi, J. Mphahlehle, V. Manchidi, Pothole tagging system, In Proceedings of the 4th Robotics and Mechatronics Conference of South Africa, 1–4 (2011) 16. L. Huidrom, L.K. Das, S. Sud, Method for automated assessment of potholes, cracks and patches from road surface video clips. Proc. Soc. Behav. Sci. 104(2013), 312–321 (2013) 17. G. Jog, C. Koch, M. Golparvar-Fard, I. Brilakis, Pothole properties measurement through visual 2d recognition and 3d reconstruction. Comput. Civ. Eng. 2012, 553–560 (2012) 18. C. Koch, G. Jog, I. Brilakis, Pothole detection with image processing and spectral clustering. J. Comput. Civ. Eng 27, 370–378 (2013) 19. R.R. Schultz, R.L. Stevenson, Extraction of high-resolution frames from video sequences. IEEE Trans. Image Process. 5(6), 996–1011 (1996) 20. V. Lempitsky, P. Kohli, C. Rother, T. Sharp, Image segmentation with a bounding box prior, in 2009 IEEE 12th International Conference on Computer Vision (IEEE, 2009), pp. 277–284 21. F.L. Pirkle, Computer workstation, U.S. Patent 4,717,112, 5 Jan 1988 22. J.D. Owens, M. Houston, D. Luebke, S. Green, J.E. Stone, J.C. Phillips, GPU computing. Proc. IEEE 96(5), 879–899 (2008) 23. D. Kirk et al., NVIDIA CUDA software and GPU parallel computing architecture, in ISMM, vol. 7, pp. 103–104 (2007) 24. E. Bisong, Google colaboratory, in Building Machine Learning and Deep Learning Models on Google Cloud Platform (Springer, 2019), pp. 59–64 25. O. Android, Android, Retrieved February, vol. 24, p. 2011 (2011) 26. L. Tzutalin, Git code (2015) 27. R. Atikur, Annotated potholes image dataset. https://www.kaggle.com/chitholian/annotatedpotholes-dataset, Feb 2020. [Online; accessed 12-March-2020]

106

C. R. Pasunuru and K. Muthireddy

28. S.J. Pan, Q. Yang, A survey on transfer learning. IEEE Trans. Knowl. Data Eng. 22(10), 1345–1359 (2009) 29. Google, Tensorflow model zoo. https://github.com/tensorflow/models/blob/master/research/ object%20detection/g3doc/detection%20model%20zoo.md (2020). [Online; accessed 25February-2020] 30. J.V. Dillon, I. Langmore, D. Tran, E. Brevdo, S. Vasudevan, D. Moore, B. Patton, A. Alemi, M. Hoffman, R.A. Saurous, Tensorflow distributions, arXiv preprint arXiv:1711.10604 (2017) 31. T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, C.L. Zitnick, Microsoft coco: common objects in context, in European Conference on Computer Vision (Springer, 2014), pp. 740–755 32. Google, Tensorflow models. https://github.com/tensorflow/models (2020). [Online; accessed 25-February-2020] 33. D. Mane et al., Tensorboard: Tensorflow’s visualization toolkit (2015) 34. M. Buckland, F. Gey, The relationship between recall and precision. J. Am. Soc. Inf. Sci. 45(1), 12–19 (1994) 35. Wikipedia, Precision and recall—Wikipedia, the free encyclopedia. https://en.wikipedia. org/w/index.php?title=Precisionandrecall&oldid=941734342 (2020). [Online; accessed 25February-2020]

IoT Devices for Detecting and Machine Learning for Predicting COVID-19 Outbreak Shams Tabrez Siddiqui, Anjani Kumar Singha, Md Oqail Ahmad, Mohammad Khamruddin, and Riaz Ahmad

1 Introduction COVID-19 touched nearly every country and about 50 million people in 2020. Due to the health, economic, and social issues it brings, governments are obliged to make tough decisions. By spring 2020, more than 50% of the world’s population experienced lockdown accompanied by rigorous quarantine protocols. It is thought that COVID-19 began in animals, with the majority of cases first appearing in the animal market and seafood’s in Wuhan, China, in December 2019. It has recently spread all across the world. It vulnerably spread through person to person around all the countries. It has been generally found in the people having the travel history; therefore, it was a drastic task to avoid it from spreading. Usually, it affects most people of age 60 or above, the person with chronic disease or children because they have feeble immune systems than adults. European, American and Asian countries face more deaths and infections as compared to other countries. It affects the job market and financial crisis in almost all fields, namely banking, manufacturing, software companies, textiles, automobiles etc., around the world [1]. Internet of Things (IoT) and machine learning play a vital role in fighting with novel COVID-19 at different stages [2]. From tracing the COVID-19 patients for isolating, to check the travel history of the patients, to check the persons in contact with the infected, to check the S. T. Siddiqui (B) · M. Khamruddin Department of Computer Science, Jazan University, Jazan, Kingdom of Saudi Arabia e-mail: [email protected] A. K. Singha Department of Computer Science, Aligarh Muslim University, Aligarh 202002, India M. O. Ahmad Department of Computer Applications, B.S Abdur Rahman Crescent Institute of Science and Technology, Chennai, India R. Ahmad Centre for Distance and Online Education, AMU, Aligarh, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_12

107

108

S. T. Siddiqui et al.

activities of quarantines, sensors to monitor the temperature, blood pressure, heartbeat, throat infections, and voice changes are some of the errands performed or area where IoT plays a gigantic role [3]. This paper focuses on tracing the individuals who came in contact with the infected individuals so that the health department can trace the contacted persons and stop them from spreading the virus further by advising testing and quarantine [4]. If anyone from the contacted person is also infected with COVID-19, then the smartphone of that person will be used to do the contact tracing. Another part of the paper discusses the machine learning classifiers used to predict future cases of countries based on their climate, environment, behavior and nature, and other factors like socioeconomic.

2 Related Work A coronavirus is a type of virus that consists of genetic material and an envelope containing protein spikes known as crowns. Coronaviruses come in a variety of forms, including respiratory, gastrointestinal, and others. The respiratory diseases vary from common cold to pneumonia, and the majority of the time, the symptoms is mild. The SARS-COV coronavirus is one of them. The coronavirus SARSC-COV, on the other hand, was first discovered in China in 2003 [5]. In 2012, the Mars-COV coronavirus was discovered in Saudi Arabia. COVID was first discovered in China in 2019; through a mechanism known as spillover, this strain of the virus is spread from animals to humans [6, 7]. A polymerase chain reaction (PCR) test is used to diagnose the infection (polymerase chain reaction). The genetic fingerprint was used to identify this examination. Till December 2020, there was no special prescription, supportive treatment, or vaccines available at that time. Fever, cough, and shortness of breath are among the symptoms, which vary from mild to extreme. The only thing we can do is stop the virus from spreading [8]. Near contact with affected people and should be avoided, and alcohol-based hand sanitizer should be used. Unnecessary interaction with animals should be avoided by humans. Before eating animal products, make sure they are well-cooked. IoT’ effect on healthcare Joseph et al. [9] used three methods to analyze social media data: information, descriptive, network analysis and used to extract information about an individual’s likes and dislikes, according to the findings. Siddiqui et al. [10] provided an outline of IoT and highlighted key problems in the sector [11, 12]. Gil et al. [13] look at existing machine learning and Internet of things technology, methodologies, and models that are currently available. Kumar et al. [14] developed an ontology-based architecture for measuring well-being and workout and making recommendations to chronic disease patients. When it came to making context-based inferences, the machine learning model was built to be efficient. Based on existing data, questionnaires, and check results, Hassantabar et al. [15] COVID-19, Internet of Things (IoT)-enabled devices are being used to minimize the chances of COVID-19 spreading to others by detecting the virus early, monitoring patients, and following the protocols after the recovery (nCapp). The need for protocol standardization in mobile phone and

IoT Devices for Detecting and Machine Learning for Predicting …

109

smart city communication was highlighted by Allam and Jones [16]. Chakraborty and Abougreen [17] discuss the medical and applications of healthcare IoT technologies and machine learning. For detecting infected people in groups, author introduced a smartphone, a smart helmet, as well as thermal imaging devices [18]. It was also outfitted with facial recognition technology. An infrared camera is used to scan the crowd, and if a person’s high temperature is identified, an optical camera is used to capture their face [19]. It also employs the Global Positioning System (GPS) to locate the infected individual.

3 Machine Learning for Predicting COVID-19 Cases In statistics, simple method that determining the correlation between at least one independent variable or explanatory variable and a dependent variable is linear regression (LR). LR was a popular type of regression analysis that drew a lot of attention but was widely employed in practice [20–22]. LR demonstrates any link between two variables through fitting a straight condition to a dependent variable. The dependent variable is considered as a dependent variable, although the independent variable is handled as an independent variable. An LR line seems to have the following structure: Y = a + bX is the formula, Y is the dependent variable, and X is the explanatory variable. The slope of the line and the intercept (the value of y when x = 0) are both beat multilayer perceptron (MLP), is a form of feed forward artificial neural network (ANN). The term MLP is loosely defined, with some referring to structures made up of different layers of the perceptron and others to any feedforward ANN. A multilayer perceptron (MLP) is a sort of perceptron that is widely used to tackle difficult issues. The formula for multilayer perceptron is: f (x) = 

 n 

 wi ∗ xi

+ b,

i=1

where n is the number of neurons in the preceding, w is a random weight, x is the input value, and layer b is a random bias, a vector autoregression (VAR) is a prediction calculation that is employed when at least two time series interact, i.e., the relationship between the periods is bi-directional. The formula for VAR is: f (x)t = α1 xt−1 + α2 xt−2 + . . . + α p xt− p + εt , where x t = (x 1t , x 2t , …, x nt )’: an (n × 1) variables from a time series vector, a: an (n × 1) intercepts vector, x i (i = 1, 2, …, p): (n × n) matrix of coefficients, εt : an (n × 1) White noise is a vectors representing unobservable that is erroneous.

110

S. T. Siddiqui et al.

4 Proposed Method for Tracing and Isolating COVID-19 Infected Person At the beginning of a pandemic, the most important task is to trace and quarantine the individual who made contact with COVID-19 infected person. This is one method of confining the disease. IoT can be used to get the information of all the contacted persons with the infected person within a limited time of a minimum of 14 days, then it will be easy to trace and isolate them [23, 24]. This way, we will put the pandemic disease under control and will get information on the sentinel case. This mobile app is based on the feature of exchanging some information with the fellow user who came in contact with the case so that it can be easy to trace. Nowadays, everyone is using android mobiles; therefore, this method can be very useful. The user has to install the app on their mobiles. The installed app will provide a unique sequence number that the smartphones will share instead of mobile numbers directly due to security purposes. Mobiles with app will create a smartphone ad hoc network and will get the mobile information of the persons near to the particular smartphone. It will work with near network connectivity. Mobiles will exchange their information with nearby mobiles and they will store information like unique no., date, time, and duration of their neighbor mobiles [25, 26]. If new users having the app are closer to each other in a near to me network with mobile then they will exchange the information against each sequence number provided while installing the apps, the authority can access the mobile numbers to inform those, particularly the current situation of an infected person so that they can also go for the checkup. In future, if anybody gets COVID-19 infection the health department can use their mobile phones to get the information of those individuals who came in contact with that infected individual within 14 days (days, min, and Max distance can be fixed as per the instructions from the health department). Their mobile records the information of those mobiles that came in contact with a COVID case. Now, the health department can contact the mobile owners and inform them about the requirement of testing contact tracing and quarantine.

5 The Architecture of the Smartphone Ad Hoc Network Figure 1 shows the smartphone ad hoc network formed by the cluster of smartphone users gather together in public places/shopping malls. Now, each smartphone has its own near to me ad hoc network and communicates/exchanges smartphone unique no, date, time, and duration with its close proximities. A near-me network is a communication system that focuses on wireless communication between devices that are close together (Fig. 2) [25]. For detailed position monitoring, traditional contact tracing techniques rely on wireless technologies (Bluetooth low energy, RFID, Wi-Fi, GPS, and magnetic field signature). Wireless technology, unlike the previous way, offers information on the

IoT Devices for Detecting and Machine Learning for Predicting …

111

Fig. 1 Smartphones in contact with near smartphones

NODE MOVEMENT day 1 18 16 14 12 10 8 6 4 2 0

17

16 14

13

9 7 5

0

2

9 8

7 5 4

3 4

3 2

7

13 13 11 10 10 9 7

16

12 12 11 10 9 8 4

3 1 6

14

15 14

1 8

10

1 12

14

16

18

Fig. 2 The graph shows smartphones come in contact with each other

duration and proximity of encounters with validated cases. One of the most extensively used IoT technologies, Bluetooth low energy (LED), can track location to a very high degree. Bluetooth L can provide a lot more precise order than proximity detection when compared to cellular location and Wi-Fi. It is crucial to keep track of this improved accuracy when classifying interactions and prioritizing responses to close contact issues [26]. Bluetooth offers a variety of location monitoring options, including RSS, LED as well as angle of arrival (AoA). Our smartphones and the majority of connected wearable support the Bluetooth LE protocol. The use of Bluetooth tags during an epidemic is another way to improve response planning [27]. In order to put things in perspective, hundreds or even thousands of Bluetooth tags and smart gadgets that interact in highly populated urban areas must be deployed. Because their device is a certified subject, this network of Bluetooth devices (mesh of Bluetooth devices) requires system-wide optimization to avoid message collisions, which could result in the loss of a highly sensitive connection.

112

S. T. Siddiqui et al.

6 Discussion and Future Work COVID-19 is considered a global health issue as well as an economic burden on mankind. The COVID-19 pandemic-related limitations have wreaked havoc on numerous marketplaces, industries, cultures, economies, and our daily lives. It takes time to determine and quantify the entire health, social as well as economic ramifications of this pandemic; however, multiple efforts are in science and industry to monitor, cure, and detect the virus in order to lessen its effects. With the Internet of Things (IoT) technology, early detection, quarantine, and recovery from COVID-19 have all shown promising outcomes; however, as we gain more knowledge about the virus and its ecology, we can alter and reinforce our tactics at various phases. It will be fascinating to observe how artificial intelligence (AI) and technology may be used to eliminate all contact between healthcare providers and patients. Another example is the combination of touch-less technology with various inputs (such as gesture and speech) to successfully minimize disease transmission and terminate the pandemic sooner. More research is needed to persuade confirmed COVID-19 cases to stay in quarantine in order to prevent the virus from spreading. How can IoT devices be integrated into enterprises to ensure both security and productivity as businesses and marketplaces progressively reopen after the lockdown? The answers to such questions will excite interest in science and industry, as well as bring up new research opportunities in this field. When patients are asked to reveal their data, a privacy issue arises which is one of the key worries about employing IoT devices at various stages of the epidemic. Whenever patients get forced for reveal their data, a privacy issue arises is one of the key worries about employing IoT devices at various stages of the epidemic. Defining safe communication routes and employing various encryption methods before transmitting sensitive information are two potential research fields, as this is a major concern for any patient. IoT-enabled by increasing collaboration between medical centers, communities, and other entities, smart cities have a lot of potential in combatting current and future pandemics. Through smart transportation systems such as smart parking, crowd control, and traffic re-routing, smart city infrastructure assists people to maintain social distance. Linear regression (LR) and multilayer perceptron (MLP) are two methods for making predictions 65% for confirmed cases, 62% for death cases, as well as 20% for recovered cases was shown in detail in paper “COVID-19 Pandemic Prediction and Forecasting using Machine Learning Classifiers.” In-depth investigation is required to assist the affected people throughout Asia. Novel methods including linear regression (LR), novel methods including linear regression (LR), vector auto regression (VAR), and multilayer perceptron (MLP) are also available MLP, and vector autoregression are also available (VAR). In order to achieve the maximum possible accuracy, a few mathematical computations, and parameters are changed for the specified model.

IoT Devices for Detecting and Machine Learning for Predicting …

113

7 Conclusion In this present pandemic, the global population needs ravenous for information on how to safeguard themselves against the COVID-19 virus. Furthermore, once the second wave of cases arrives, experts are attempting to forecast additional infected cases based on the present death toll and viral infections among India’s general populace. Considering one of the most important features such as gender, person age, immune power, saturation point, and some additional characteristics of infected people’s medical background could enhance the accuracy of machine learning model in treating COVID-19. Because as second wave of virus wreaks havoc on India, it really is critical to implement widespread lockdown, serious social separation, through the use of suitable masks so protects the face, and isolation if a people becomes sick, so order to halt the spread of the virus. COVID-19 data was so well parsed as well as categorized that use the VAR, MLP, and LR algorithms. To use the ORANGE and WEKA tools, that one was discovered as MLP provided great promise with categorizing Indian COVID data. Inside the nearest term, researchers intend to apply deep learning classifier to much more accurately categories COVID-19 data as well as deal with others and analyze time series COVID-19 data.

References 1. S. Ketu, P.K. Mishra, Enhanced Gaussian process regression-based forecasting model for COVID-19 outbreak and significance of IoT for its detection. Appl. Intell. 51(3), 1492–1512 (2021) 2. P. Ukhalkar, Machine learning-based IoT-enabled perspective model for prediction of COVID19 test in early stage. Mach. Learn. 29(12s), 2599–2604 (2020) 3. O. Shahid, M. Nasajpour, S. Pouriyeh, R.M. Parizi, M. Han, M. Valero, F. Li, M. Aledhari, Q.Z. Sheng, Machine learning research towards combating COVID-19: virus detection, spread prevention, and medical assistance. J. Biomed. Inf. 117, 103751 (2021) 4. T. DeFranco, Internet of things and future of virus detection and prevention, CEO of IOTA Communications (2017). [Online] Available: https://www.iotevolutionworld.com/iot/articles/ 444815-internet-things-future-virus-detection-prevention.htm. 5. S. Sharif, A. Ikram, A. Khurshid, M. Salman, N. Mehmood, Y. Arshad, J. Ahmad, R.M. Safdar, M. Angez, M.M. Alam, L. Rehman, Detection of SARs-CoV-2 in wastewater, using the existing environmental surveillance network: an epidemiological gateway to an early warning for COVID-19 in communities. MedRxiv (2020) 6. M. Kumar, M. Joshi, A.V. Shah, V. Srivastava, S. Dave, Wastewater surveillance-based city zonation for effective COVID-19 pandemic preparedness powered by early warning: a perspectives of temporal variations in SARS-CoV-2-RNA in Ahmedabad, India. Sci. Total Environ. 148367 (2021) 7. S. Lalmuanawma, J. Hussain, L. Chhakchhuak, Applications of machine learning and artificial intelligence for Covid-19 (SARS-CoV-2) pandemic: a review. Chaos Solitons Fractals 139, 110059 (2020) 8. D. Silveira, J.M. Prieto-Garcia, F. Boylan, O. Estrada, Y.M. Fonseca-Bazzo, C.M. Jamal, P.O. Magalhães, E.O. Pereira, M. Tomczyk, M. Heinrich, COVID-19: is there evidence for the use of herbal medicines as adjuvant symptomatic therapy? Front. Pharmacol. 11, 1479 (2020)

114

S. T. Siddiqui et al.

9. N. Joseph, A.K. Kar, P.V. Ilavarasan, S. Ganesh, Review of discussions on internet of things (IoT): insights from twitter analytics. J. Glob. Inf. Manage. (JGIM) 25(2), 38–51 (2017) 10. S.T. Siddiqui, S. Alam, R. Ahmad, M. Shuaib, Security Threats, Attacks, and Possible Countermeasures in Internet of Things (In Advances in Data and Information Sciences, Springer, 2020), pp. 35–46 11. S. Alam, S.T. Siddiqui, A. Ahmad, R. Ahmad, M. Shuaib, Internet of Things (IoT) Enabling Technologies, Requirements, and Security Challenges (In Advances in Data and Information Sciences, Springer, 2020), pp. 119–126 12. F. Masoodi, S. Alam, S.T. Siddiqui, Security and privacy threats, attacks and countermeasures in internet of things. Int. J. Network Secur. Appl. (IJNSA) 11 (2019) 13. D. Gil, A. Ferrández, H. Mora-Mora, J. Peral, Internet of things: a review of surveys based on context aware intelligent services. Sensors 16(7), 1069 (2016) 14. K. Kumar, N. Kumar, R. Shah, Role of IoT to avoid spreading of COVID-19. Int. J. Intell. Netw. 1, 32–35 (2020) 15. S. Hassantabar, N. Stefano, V. Ghanakota, A. Ferrari, G.N. Nicola, R. Bruno, I.R. Marino, K. Hamidouche, N.K. Jha, Coviddeep: Sars-cov-2/covid-19 test based on wearable medical sensors and efficient neural networks. arXiv preprint arXiv:2007.10497 (2020) 16. Z. Allam, D.S. Jones, On the coronavirus (COVID-19) outbreak and the smart city network: universal data sharing standards coupled with artificial intelligence (AI) to benefit urban health monitoring and management. Healthcare 8(1), 46 (2020) 17. C. Chakraborty, A.N. Abougreen, Intelligent internet of things and advanced machine learning techniques for COVID-19. EAI Endorsed Trans. Pervasive Health Technol. 7(26), e1 (2021) 18. M. Ali, IoT-based smart solution to early detect COVID-19 patients. 1–17 (2021) 19. R. Farkh, M.T. Quasim, K. Al Jaloud, S. Alhuwaimel, S.T. Siddiqui, Computer vision-controlbased CNN-PID for mobile robot. Comput. Mater. Continua 68(1), 1065–1079 (2021) 20. M. Yamin, A.A.A. Sen, Z.M. Al-Kubaisy, R. Almarzouki, A novel technique for early detection of COVID-19. Comput. Mater. Continua 2283–2298 (2021) 21. X. Yan, X. Su, Linear Regression Analysis: Theory and Computing. World Scientific (2009) 22. A. Alotaibi, M. Shiblee, A. Alshahrani, Prediction of severity of COVID-19-infected patients using machine learning techniques. Computers 10(3), 31 (2021) 23. J. Sultana, A.K. Singha, S.T. Siddiqui, N. Pathak, A.K. Sriram, G. Nagalaxmi, COVID-19 pandemic prediction and forecasting using machine learning classifiers. Intell. Autom. Soft Comput. 32(2), 1007–1024 (2022) 24. N. Jha, D. Prashar, M. Rashid, M. Shafiq, R. Khan, S.T. Siddiqui, Deep learning approach for discovery of in silico drugs for combating COVID-19. J. Healthc. Eng. 2021, 1–13 (2021) 25. H.A. Hammatta, S.T. Siddiqui, M.U. Bokhari, Protocols in mobile ad-hoc networks: a review. Int. J. Appl. Inf. Syst. 7(10), 11–14 (2014) 26. E. Hernandez-Orallo, P. Manzoni, C.T. Calafate, J.C. Cano, Evaluating how smartphone contact tracing technology can reduce the spread of infectious diseases: the case of COVID-19. IEEE Access 8, 99083–99097 (2020) 27. S.H. Liang, S. Saeedi, S. Ojagh, S. Honarparvar, S. Kiaei, M. MohammadiJahromi, J. Squires, An interoperable architecture for the internet of COVID-19 things (IoCT) using open geospatial standards—case study: workplace reopening. Sensors 21(1), 50 (2021)

Text Detection from Scene and Born Images: How Good is Tesseract? Nadeem Anwar, Tauseef Khan, and Ayatullah Faruk Mollah

1 Introduction Images containing texts carry semantic information as a means of effective communication. Detection of texts from such images is prerequisite for numerous computer vision applications, like identification of license plate, traffic signal detection, making index of multimedia resources, text-to-speech translation to aid visually impaired, etc. [1]. However, several environmental challenges, viz. complex background, uneven lighting, perspective distortion, etc., associated with scene images are one of the main obstacles for appropriate detection of texts. To address such challenges, researchers have attempted several approaches [2] to detect texts under unconstraint scenario. Existing methods of text detection have been extensively reviewed and reported in [1, 3, 4]. Present methods of text detection may be broadly classified into two modules, (i) traditional or feature-driven methods [5–10] and (ii) deep learning-based methods [11–14]. Traditional methods are largely confined to manually designed handcrafted features present in texts. Such methods may be further categorized into two groups based on different strategies applied by researchers, (a) sliding window (SW)-based classification [15, 16] and (b) connected component (CC)-based methods [17, 18]. In SW methods, multi-scale windows slide over input images to generate multiple

N. Anwar (B) The Adabi Society High Madrasah, Angus, Hooghly 712221, India e-mail: [email protected] T. Khan Department of Information Technology, Haldia Institute of Technology, ICARE Complex, Haldia 721657, India T. Khan · A. F. Mollah Department of Computer Science and Engineering, Aliah University, Newtown Campus, Kolkata 700160, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_13

115

116

N. Anwar et al.

image blocks that are classified as text segments or not, using intrinsic featuredriven pattern classifiers. Finally, true classified components are grouped together for word or line-level text detection. However, most of SW methods are applicable for horizontal texts and yields low performance for arbitrary-oriented texts in scenes [4]. In CC-based approaches, at first, foreground components are separated through various segmentation approaches (e.g., color-based cluster formation or extraction of extreme stable regions), followed by manual extraction of discriminant features (e.g., stroke-width, gradient, texture). Finally, trained classifiers classify such regions to filter-out non-text regions. Mollah et al. [7] have presented a text detection method from camera images using fuzzy membership to suppress background and isolate text instances. Khan et al. [8] have developed a scheme to extract text regions from camera images using a background suppression algorithm. Khan et al. [9] have proposed a script independent feature vector based on area of occupancy of equal-distant pixels. Traditional methods are predominantly motivated by handcrafted features. In natural, imagery designing discriminant features itself it is a challenging job that are greatly susceptible to noise, lighting, arbitrary orientation of texts, and other clutters. Moreover, such methods are usually preceded and succeeded by tedious and complicated processing steps which easily propagate errors to next steps [1, 3, 4]. Rise of deep learning has widened the possibility of research in the field of text detection. Due to automated deep features, deep learning-based approaches are more advantageous than traditional methods. Such methods may be categorized into three groups, (a) regression-based methods [19, 20], (b) segmentation-based methods [14, 21], and (c) hybrid methods [22, 23]. First category of methods considers text as an object where multi-directional convolving rectangular or quadrilateral text boxes predict the candidate bounding boxes. However, prior knowledge is prerequisite for multi-scale boxes, and these methods may suffer while dealing with arbitrarilyorientated and curved text instances. Segmentation-based methods are more suitable for multi-oriented and curved texts, where foreground objects are segmented by pixellevel prediction using fully convolutional neural network. However, such methods suffer from computationally expensive post-processing. Hybrid method is fusion of regression and segmentation-based methods, and inherits the benefits of two methods for accurate text detection. Score maps of text is predicted using segmentation-based approach and the same time aims to obtain text bounding boxes through regression [1, 3, 4]. Despite tremendous success of deep learning methods in recent times, few issues are associated with deep networks make them inappropriate for real-time deployment. Deep models are driven by high resource environment, and improper for low resource states such as mobile phone and handheld digital camera. On contrary, traditional methods can work with limited computational resources. Tesseract is a well-known computer vision tool that works mainly in character recognition [24]. However, to the best of our knowledge, assessment of this OCR engine has not been implemented for detecting texts in the wild. Hence, a fair attempt has been made in this paper to assess the performance of Tesseract in text detection. Contributions of this paper may thus be summarized as (i) performance assessment of Tesseract OCR engine in scene text

Text Detection from Scene and Born Images: How Good is Tesseract?

117

detection has been reported, and (ii) rigorous experiments have been carried out on multiple benchmark datasets, viz. ICDAR 2013 (born images), ICDAR 2013 (scene text), and ICDAR 2019-MLT, and obtained results are reported that may give a clear idea to general readers regarding the potentiality and suitability of Tesseract in scene text detection. Rest of the paper is organized as follows: proposed methodology is presented in Sect. 2, experiment and analysis are discussed in Sect. 3, and finally, conclusion of the work is drawn in Sect. 4.

2 Methodology Text detection is a prerequisite for subsequent modules of recognition. Appropriate detection of texts may increase the performance of an OCR engine. Primary steps of an OCR engine for text detection are preprocessing and segmentation of texts [25–27]. In preprocessing phase, input image passes through noise removal filters to enhance the quality without losing any intrinsic information. Further processing applies on image, such as skew correction, slant removal, image smoothing, and compression. Segmentation phase extracts text instances from image backgrounds. Then, extracted text instances are further segmented into line, word, and characterlevel. Segmented characters are passed to recognition engine. Figure 1 depicts the pipeline of proposed work using Tesseract OCR. In this work, an input image is passed through Tesseract [28, 29] to get the coordinates of the bounding boxes of detected text regions in the image. Starting with adaptive thresholding, Tesseract creates an ordered set of typed regions by implementing CC analysis and tab-stop detection [30], followed by nesting of outlines into blobs. Text line finder organizes these blobs into text lines by means of the vertical overlap of adjoining characters on a text line. Then, a fixed-pitched detector looks for character layout and finalizes by word segmentation algorithm as per the fixed pitch decision. Finally, Tesseract generates coordinates of bounding box of detected text region in table form. Now, coordinates of detected text region has been extracted from resultant table. These coordinates are employed to generate binary mask of input image. Similarly, GT coordinates are utilized to create binary mask of same Fig. 1 Work flow of Tesseract-based text detection and evaluation methodology

118

N. Anwar et al.

Fig. 2 Generation of binary maps of detected text regions from Tesseract (top), and corresponding map generated using GT coordinates (bottom)

image, as shown in Fig. 2. These two binary masks have been analyzed globally to determine intersection over union (IoU) [31] by means of computation of pixel-level ratio of overlap. Finally, average IoU of each dataset calculated.

3 Experiment and Analysis A detailed discussion on experimental environment to carry out series of experiments is reported. Performance of Tesseract is measured using standard evaluation metrics, viz. precision (P), recall (R), F-Measure (F-M), which is the harmonic mean of P and R, accuracy, and IoU. Then, concise descriptions of different experimental datasets are presented, followed by results and discussion.

3.1 Experimental Setup Current work is carried out on a system equipped with Intel® HD Graphics 620 GPU and an Intel® Core(TM) i3-7020U CPU @ 2.30GHzprocessor. Model is designed using Python framework of version3.7.4. Input image is read using Python Imaging Library (or PIL, now Pillow). Tesseract is every time invoked by its python wrapper called PyTesseract to read image of all file formats. Tesseract 5.x, modified on 23.06.2019, and based on long short-term memory, a part of recurrent neural network, is implemented for text detection. Trained language model data, called Tessdata, is general.

Text Detection from Scene and Born Images: How Good is Tesseract?

119

Fig. 3 Sample images from different datasets: ICDAR 2013 (born images) (left), ICDAR 2013 (focused scene text) (middle), and ICDAR 2019-MLT (right)

3.2 Datasets Present work is evaluated on different datasets of ICDAR robust reading competition that contains complex scene and born images. Sample images of each dataset have been shown in Fig. 3. ICDAR 2013 (Born-Digital Images). Born-digital images are compressed, lowresolution, and online transmittable. Total number of images is 551, which are further separated into training and test sets. GT images are marked with rectangular bounding box enclosed over text regions. ICDAR 2013 (Focused Scene Text). Focused scene text images are high-resolution well-focused camera-captured images, and mostly horizontal. Collectively training and test sets have 462 images. ICDAR 2019-MLT. These images are text scene images from 10 different languages of seven separate scripts, viz. Arabic, English, French, Chinese, German, Korean, Japanese, Italian, Bangla, and Hindi (Devanagari), including multi-oriented text as well. Dataset contains total 10,000 images for training. Only training set has been evaluated as no GT is available for test set.

3.3 Results Series of experiments have been steered out on multiple datasets using Tesseract, and obtained results are reported in Table 1. It may be observed that Tesseract has achieved accuracy score of 0.881 for ICDAR 2013 (born images), and F-M of 0.526 and IoU of 0.4. Besides, method has obtained accuracy of 0.889, F-M of 0.308, and IoU of 0.24 for ICDAR 2013 (focused scene text). Accuracy of 0.873, F-M of Table 1 Results for text detection on different datasets using Tesseract OCR engine Dataset

R

P

F-M

Accuracy

IoU

ICDAR 2013 (Born image)

0.544

0.510

0.526

0.881

0.400

ICDAR 2013 (Focused scene text)

0.306

0.311

0.308

0.889

0.240

ICDAR 2019-MLT

0.162

0.203

0.178

0.873

0.114

120

N. Anwar et al.

Fig. 4 Text detection results using Tesseract on experimental datasets: results on ICDAR 2013 (born images) (left), results on ICDAR 2013 (focused scene texts) (middle), and detected texts on ICDAR 2019-MLT dataset (right)

Fig. 5 Some improper, partial text detection along with false alarms using Tesseract on experimental datasets. ICDAR 2013 (born images) (left), ICDAR 2013 (focused scene images) (middle), and ICDAR 2019-MLT dataset (performance suffers for quadrilateral GT annotation, red and green boxes refer to GT and detected regions, respectively) (right)

0.178, and IoU of 0.114 are obtained for ICDAR 2019-MLT dataset. Results of text detection using Tesseract on three datasets are depicted in Fig. 4. Tesseract generally detects rectangular bounding box of text instances, whereas GT annotations of ICDAR 2019-MLT images are quadrilateral tighter bounding boxes, which cost the performance of Tesseract to some extent, which is illustrated in Fig. 5 (right). Generally, Tesseract yields improper and partial detection of texts in scene environment that affects the overall performance. Despite substandard performance, Tesseract has achieved accuracy of around 0.88% for all three datasets, which is due to correct exclusion of background and non-text area that constitutes true negatives (one of numerator of accuracy fraction). Besides, arbitrary-oriented scene images are comprised of highly varied aspect ratio, which generates several false alarms, shown in Fig. 5. Processing speed of Tesseract depends on image dimension. Small-sized born images take lesser time than large size scene images.

4 Conclusion and Future Works In this paper, performance of Tesseract OCR engine for scene text detection on multiple benchmark datasets has been reported. Reported assessment may build a fair perception for general readers regarding the performance of Tesseract in text detection. It may be comprehended from the obtained results that direct deployment of Tesseract is not suitable for scene text detection. However, more focus on initial detection of text instances and some additional post-processing may improve the result. Besides, text detection in multi-script environment may be considered later as an extension of this work. Therefore, such kind of assessment may be helpful to open several research avenues for further improvement in this domain.

Text Detection from Scene and Born Images: How Good is Tesseract?

121

References 1. T. Khan, R. Sarkar, A.F. Mollah, Deep learning approaches to scene text detection: a comprehensive review. Artif. Intell. Rev. 54, 3239–3298 (2021) 2. N. Pawar, Z. Shaikh, P. Shinde, Y.P. Warke, Image to text conversion using Tesseract. Int. Res. J. Eng. Technol. 6(2), 516–519 (2019) 3. S. Long, X. He, C. Yao, Scene text detection and recognition: the deep learning era (2020). arXiv:1811.04256v5 4. Z. Raisi, M.A. Naiel, P. Fieguth, S. Wardell, J. Zelek, Text detection and recognition in the wild: a review (2020). arXiv:2006.04305v2 5. C.R. Kulkarni, A.B. Barbadekar, Text detection and recognition: a review. Int. Res. J. Eng. Technol. 4(6), 179–185 (2017) 6. T. Khan, A.F. Mollah, AUTNT-A component level dataset for text non-text classification and benchmarking with novel script invariant feature descriptors and D-CNN. Multimedia Tools Appl. 78(22), 32159–32186 (2019) 7. A.F. Mollah, S. Basu, M. Nasipuri, Text detection from camera captured images using a novel fuzzy-based technique, in 3rd International Conference on Emerging Applications of Information Technology (2012), pp. 291–294 8. T. Khan, A.F. Mollah, A novel text localization scheme for camera captured document images, in 2nd International Conference on Computer Vision and Image Processing, Advances in Intelligent Systems and Computing, vol. 703, pp. 253–264 (2018) 9. T. Khan, A.F. Mollah, Text non-text classification based on area occupancy of equidistant pixels. Int. Conf. Comput. Intell. Data Sci. Procedia Comput. Sci. 167, 1889–1900 (2020) 10. A.C. Ozgen, M. Fasounaki, H.K. Ekenel, Text detection in natural and computer-generated images, in 26th Signal Processing and Communications Applications Conference (IEEE, 2018), pp. 1–4 11. M. Behzadi, R. Safabakhsh, Text detection in natural scenes using fully convolutional DenseNets, in Proceedings of 4th Iranian Conference on Signal Processing and Intelligent Systems (IEEE, 2019), pp. 11–14 12. Z. Liu, G. Lin, S. Yang, J. Feng, W. Lin, W.L. Goh, Learning Markov clustering networks for scene text detection (2018). arXiv:1805.08365v1 13. H. Qin, H. Zhang, H. Wang, Y. Yan, M. Zhang, W. Zhao, An algorithm for scene text detection using multi-box and semantic segmentation. Appl. Sci. 9(6), 1054 (2019) 14. M. Liao, Z. Wan, C. Yao, K. Chen, X. Bai, Real-time scene text detection with differentiable binarization, in 34th Association for the Advancement of Artificial Intelligence Conference on Artificial Intelligence (2020), pp. 11474–11481 15. A. Coates, B. Carpenter, C. Case, S. Satheesh, B. Suresh, T. Wang, D.J. Wu, A.Y. Ng, Text detection and character recognition in scene images with unsupervised feature learning, in ICDAR (IEEE, 2011), pp. 440–445 16. J.J. Lee, P.H. Lee, S.W. Lee, A. Yuille, C. Koch, Adaboost for text detection in natural scene, in ICDAR (2011), pp. 429–434 17. W. Huang, Z. Lin, J. Yang, J. Wang, Text localization in natural images using stroke feature transform and text covariance descriptors, in Proceedings of the IEEE International Conference on Computer Vision (2013), pp. 1241–1248 18. T. Khan, A.F. Mollah, Distance transform-based stroke feature descriptor for text non-text classification, in Recent Developments in Machine Learning and Data Analytics (2019), pp. 189–200 19. M. Liao, Z. Zhu, B. Shi, G.S. Xia, X. Bai, Rotation-sensitive regression for oriented scene text detection, in IEEE Conference on Computer Vision and Pattern Recognition (2018), pp. 5909– 5918 20. F. Liu, C. Chen, D. Gu, J. Zheng, FTPN: Scene text detection with feature pyramid based text proposal network. IEEE Access 7, 44219–44228 (2019) 21. Y. Tang, X. Wu, Scene text detection and segmentation based on cascaded convolution neural networks. IEEE Trans. Image Process. 26(3), 1509–1520 (2017)

122

N. Anwar et al.

22. P. He, W. Huang, T. He, Q. Zhu, Y. Qiao, X. Li, Single shot text detector with regional attention, in IEEE International Conference on Computer Vision (2017), pp. 3047–3055 23. T.Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, S. Belongie, Feature pyramid networks for object detection, in IEEE Conference on Computer Vision and Pattern Recognition, pp. 2117– 2125 (2017) 24. S.V. Rice, F.R. Jenkins, T.A. Nartker, The fourth annual test of OCR accuracy, in Computer Science (1995), pp 1–39 25. N. Islam, Z. Islam, N. Noor, A survey on optical character recognition system. ITB J. Inf. Commun. Technol. 10(2), 1–4 (2016) 26. B. Sharma, A.K. Rao, OCR related technology methods. Int. J. Adv. Trends Comput. Sci. Eng. 9(3), 2789–2793 (2020) 27. K.A. Hamad, M. Kaya, A detailed analysis of optical character recognition technology, in 3rd International Conference on Advanced Technology & Sciences; Int. J. Appl. Math. Electron. Comput. 4(Special Issue), 244–249 (2016) 28. R. Smith, An overview of the Tesseract OCR engine, in 9th International Conference on Document Analysis and Recognition (2007), pp. 629–633 29. R. Smith, D. Antonova, D.-S. Lee, Adapting the Tesseract open source OCR engine for multilingual OCR, in International Workshop on Multilingual OCR (2009), pp. 1–8 30. R. Smith, Hybrid page layout analysis via tab-stop detection, in 10th International Conference on Document Analysis and Recognition (2009), pp. 241–245 31. D. Karatzas, L. Gomez-Bigorda, A. Nicolaou, S. Ghosh, A. Bagdanov, M. Iwamura, J. Matas, L. Neumann, V.R. Chandrasekhar, S. Lu, F. Shafait, ICDAR 2015 competition on robust reading, in 13th ICDAR (IEEE, 2015), pp. 1156–1160

Power System Load Frequency Control of Hybrid Integrated with Solar-Thermal and Geothermal System Ayman Farooq, Zahid Farooq, and Krishna Tomar

1 Introduction Mutual connection of various control areas through tie lines results in LFC in an interrelated power coordination system. The alteration in frequency and tie line power is the product of any sudden deviation of load in any of the various control areas used in the power arrangement linkage. Large frequency oscillations may sometimes lead to system failure [1]. For restoring system balance, LFC plays an important role. The purpose of using LFC is to sustain the actual frequency and chosen output power in the power Web and to manage the variations in tie line power between different zones. So, an LFC system must be incorporated with modern and intelligent control techniques to provide requisite and reliable power supply [2]. Concerning the fact, researchers have reported many thought-provoking facts in the field of LFC. Programmed generation mechanism of three-zone power system set-up by means of artificial intelligent controllers has been presented by Sirish et al. [3]. In 2017, Azreen et al. [4] studied the role of LFC using smart controllers (PID, fuzzy logic, and neuro controller) to eliminate errors that were created due to variation in frequency and tie line power for an economic power generation. Renewable energy sources provide pollution-free, unlimited, and nonconventional form of energy. The over use of renewable energy levels in the existing power apparatus has been found due to the decrement in the accessibility of the conventional energy resources and their hazards on the surroundings as a whole [5]. Datta et al. [6] studied the role of renewable energy sources in LFC in interconnected power systems and proved that renewable energy sources were more workable for LFC. Further, for better LFC, various optimization algorithms like A. Farooq (B) · K. Tomar Department of Electrical Engineering, Rimt University, Mandi Gobindgarh, India e-mail: [email protected] Z. Farooq Department of Electrical Engineering, NIT Srinagar (J&K), Srinagar, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_14

123

124

A. Farooq et al.

genetic algorithm, firefly algorithm, magnetotactic bacteria optimization etc., and different controllers [7] and energy storage devices are incorporated with various renewable energy sources to improve system responses [8, 9]. Babu et al. [10] have considered GTPP with a thermal system using coyote optimization with a PI-DN controller proving GTPP shows better response. However, incorporating GTPP with other renewable sources and using different optimized controllers makes sense for further investigation. Sambit dash [11] presented LFC of photovoltaic (PV) and solar thermal hybrid microgrid by incorporating Narma-L2 controller and modified whale optimization algorithm (MWOA). Eberhart and Kennedy introduced particle swarm optimization technique which has many applications and developments [12]. Magnetotactic bacteria optimization has been successfully utilized by authors in [13, 14] in optimizing the secondary controller gains. Also, by using energy storage devices in an interconnected system, the system performance is improved showing its superiority over other conventional controllers [15] for betterment of LFC. In order to analyze the robustness of controller parameters with changes in system loading, many authors incorporated sensitivity analysis in their research [16]. However, the same has not been assessed for PID controllers making use of PSO method. My objectives linked to the above stated writings are as: • To model a two-area system with solar-thermal and geothermal power plants. • Optimization of controller gains like I, PI, and PID employing PSO technique to find the most suitable controller. • To find the difference in system responses with and without GTPP. • Compare the system responses when energy storage device is used and not used in the system. • Sensitivity analysis of controller parameters obtained at standard situations.

2 System Implementation The transfer function representation of the executed system is shown in the block diagram in (Fig. 1). The two-zone system is investigated with a zone capacity ratio of 1:2. Solar and thermal arrangement is constituted under zone 1. For zone 2, geothermal and thermal system is considered. A single reheat turbine is used in thermal areas with generation rate constraint (GRC) of 3% per minute and governor dead band (GDB) of 0.06% (0.036 Hz). The dynamic outcomes have been recorded for load disturbance summing up to 1% in the thermal area system. PSO technique has been used to optimize the different controller gains along with geothermal plant parameters. The parameters for solar-thermal, and conventional thermal are referred from [7]. Here, integral square error (ISE) has been used as a cost objective which is given by T JISE = 0

   f i )2 + (Ptie )2 dt

(1)

Power System Load Frequency Control of Hybrid …

125

Fig. 1 Transfer function model of two-zone arrangement

3 Particle Swarm Optimization (PSO) PSO is a speculative algorithm practice. It is encouraged by the famous principle of bird gathering social behavior, also is based on population. Unlike genetic algorithm (GA), simulated annealing (SA), PSO has less chances of getting duped at local optima. In PSO algorithm, the velocity of respective particles is given by Eq. (2). The flowchart of PSO algorithm is shown in Fig. 2. Vik+1 = CF × [Vik + C1 × rand1 × ( pbesti − sik )   + C2 × rand2 × gbest − sik where 2  ,  = C1 + C2 ,  > 4. CF  √   2∅ − 2 + 4

(2)

126

A. Farooq et al.

Fig. 2 Flowchart of PSO algorithm

4 Results and Discussions 4.1 Response Comparison of Secondary Controllers The two-zone system is inquired into, and it uses both STS (in zone1) and GTS (in zone2) with integral, proportional integral, and proportional integral derivative controller each one at a time. The controller gains of each of them are optimized one at a time using PSO technique. The optimal controller gains are shown in Table 1, and performance comparison of I, PI, and PID controllers is shown in (Fig. 3). After careful observation, it is clear that the outcomes of PID controller surpass those of PI and I controller in terms of

Power System Load Frequency Control of Hybrid …

127

Table 1 PSO optimized values of diverse controller gains Controller gains and ISE values

I

PI

PID

K i1

0.4039

0.8655

2.2462

K i2

0.0965

0.1281

0.2954

K p1

–

0.7845

2.471

K p2

–

0.593

1.35

K d1

–

–

1.7673

K d2

–

–

2.4987

K n1

–

–

25.9064

K n2

–

–

37.307

ISE

0.00188

0.00173

0.00107

(a)

(b)

(c) Fig. 3 Dynamic reaction contrast for I, PI, and PID controllers at nominal conditions ( f 1 ,  f 2 , Ptie12 )

parameters like maximum peak value, settling time, magnitude of oscillation, and ISE. Since, ISE is dependent on changes in frequency and tie line power; therefore, it has been proved that PID controller is less susceptible to ISE and, hence, shows better performance.

4.2 System Response in Absence and Presence of GTPP The two-zone system is taken up with GTPP excluded and included in zone 2. The system dynamics are compared when GTPP is used and not used in the system. The dynamic outcomes with GTPP included and excluded are shown in Fig. 4. On

128

A. Farooq et al.

Fig. 4 Dynamic response contrast of two-area system with and without GTPP at nominal condition ( f 1 ,  f 2 , Ptie12 )

observing the results carefully, it is clear that the results with GTPP included are finer in terms of parameters like settling time, overshoot in addition to undershoot. It also showed that GTPP is faster than thermal power plants.

4.3 System Performance Energy Storage Device (ESD) The two-region system of the interconnected power system network used here normally does not have an energy storage device (ESD) present. In order to analyze the performance of the arrangement, a comparison is made between the normal arrangement and the one with ESD (redox flow battery) present. It was inferred that the system with ESD shows improved response in terms of understood and magnitude of oscillations (Table 2; Fig. 5). Table 2 Optimized gains of PID controller in presence of ESD

K d1

0.1437

K d2

2.0304

K i1

3.8680

K i2

3.6676

K n1

96.5950

K n2

92.4109

K p1

1.8480

K p2

1.2535

Power System Load Frequency Control of Hybrid …

(a)

129

(b)

(c) Fig. 5 Active response comparison of two-zone interconnected power network with and without ESD at nominal conditions ( f 1 ,  f 2 , Ptie12 )

4.4 Sensitivity Analysis (SA) Case 1: (2% SLP) Sensitivity analysis is a technique used to measure the robustness and suitability of controller parameters. In our analysis, it is inferred that the PID controller with parameters optimized using PSO algorithm has robust parameters to withstand any deviation in load perturbations (Fig. 6).

Fig. 6 Sensitivity analysis of PID controller at nominal conditions ( f 1 ,  f 2 , Ptie12 )

130

A. Farooq et al.

5 Conclusion In a two-area solid thermal system, an effort has been made to use GTPP. PSO technique has been utilized for one-time optimization of controller factors of I, PI, and PID controllers. Among these, the dynamic outcomes obtained from PID controller shows remarkable advancement in relation to ISE, settling time, overshoot as well as oscillations. Furthermore, the thermal arrangement utilizing GTPP in area 2 is proved to be far better and shows acceptable response in comparison to the system without GTPP. A separate learning is done to show the impression of energy storage device in system stability. Also, to check the sturdiness of PID controller whose controller factors are optimized using PSO, sensitivity analysis is carried out. The effective obtained parameters are found to be robust and suitable enough to resist any change in SLP and capacity ratio, and there is no requirement whatsoever to reset them even for any wide variations in system parameters.

References 1. T.H. mohmad, H. Abubakr, M.M. Hussein, G.S. Salman, Adaptive load frequency control in power system using optimization techniques. https://doi.org/10.5772/intechopen.93398 2. M.M. Ismail, M.A. Hassan, Load frequency control adaptation using artificial intelligent techniques for one and two different areas power system. Corpus ID: 18222046 3. S. Annam, Dr. K. Radha Rani, J. Srinu Naick, Automatic generation control of three area power system with artificial intelligent controllers. www.ijareeie.com. 6(8) (2017) 4. S.A. Azeer, R. Ramjug-Ballgobin, S.Z. Sayed Hassen, Intelligent controllers for load frequency control of two-area power system, IFAC-PapersOnLine 50(2), 301–306 (2017) 5. R. Bayindir, S. Demirbas, E. Irmak, U. Cetinkaya, A. Ova, M. Yesil, Effects of renewable energy sources on the power system. IEEE Int. Power Electron. Motion Control Conf. (PEMC) 2016, 388–393 (2016). https://doi.org/10.1109/EPEPEMC.2016.7752029 6. A. Datta, K. Bhattacharjee, S. Debbarma, B. Kar, Load frequency control of a renewable energy sources based hybrid system, in 2015 IEEE Conference on Systems, Process and Control (ICSPC) , pp. 34–38 (2015). https://doi.org/10.1109/SPC.2015.7473555 7. Z. Farooq, A. Rahman, S.A. Lone, Load frequency control of multi-source electrical power system integrated with solar-thermal and electric vehicle. Electr. Energy Syst. https://doi.org/ 10.1002/2050-7038.12918 8. S. Biswas, P. Bera, GA application to optimization of AGC in two-area power system using battery energy storage, in 2012 International Conference on Communications, Devices and Intelligent Systems (CODIS), pp. 341–344 (2012). https://doi.org/10.1109/codis.2012.6422208 9. S. Kalyani, S. Nagalakshmi, R. Marisha, Load frequency control using battery energy storage system in interconnected power system, in 2012 3rd International Conference on Computing, Communication and Networking Technologies (ICCCNT’12) (2012), pp. 1–6. https://doi.org/ 10.1109/ICCCNT.2012.6396052 10. N.R. Babu, L.C. Saikia, D.K. Raju, T. Chiranjeevi, Multi-area AGC system incorporating GTPP and coyote optimized PI minus DN controller, in Computing Algorithms with Applications in Engineering, pp. 349–360 11. S. Dash, Load frequency control of solar PV and solar thermal integrated micro grid using Narma-L2 Controller. arXiv:2004.05776

Power System Load Frequency Control of Hybrid …

131

12. Eberhart, Y. Shi, Particle swarm optimization: developments, applications and resources, in Proceedings of the 2001 Congress on Evolutionary Computation (IEEE Cat. No.01TH8546), vol. 1 (2001), pp. 81–86. https://doi.org/10.1109/CEC.2001.934374 13. Z. Farooq, A. Rahman, S.A. Lone, System dynamics and control of EV incorporated deregulated power system using MBO-optimized cascaded ID-PD controller. Int Trans Electr Energ Syst. e13100 (2021). https://doi.org/10.1002/2050-7038.13100 14. S. Safiullah, A. Rahman, S.A. Lone, State-observer based IDD controller for concurrent frequency-voltage control of a hybrid power system with electric vehicle uncertainties. Int. Trans. Electr. Energ. Syst. e13083 (2021). https://doi.org/10.1002/2050-7038.13083 15. P. Prajapati, A. Parmar, Multi-area load frequency control by various conventional controller using battery energy storage system, in 2016 International Conference on Energy Efficient Technologies for Sustainability (ICEETS) (2016), pp. 467–472. https://doi.org/10.1109/iceets. 2016.7583800 16. A. Rahman, L.C. Saikia, N. Sinha, AGC of dish-stirling solar thermal integrated thermal system with biogeography based optimised three degree of freedom PID controller ISSN 1752-1416. https://doi.org/10.1049/iet-rpg.2015.0474. www.ietdl.org

An Improved Stock Market Index Prediction System Based on LSTM Rais Allauddin Mulla and Satish Saini

1 Introduction The entire capitalization of all financial markets of the world’s stock exchanges began at $2.6 trillion in 1981 and has grown steadily since then. At the conclusion of 2018, it has grown to $69.43 trillion dollars. As of December 31, 2019, the entire market capitalisation of all equities across the global financial market has reached around US $70.86 trillion, according to the most recent available data. At the moment, there are 60 New York Stock Exchange operating throughout the world [1]. As a result, retail financial specialists must devote a significant amount of effort to identifying investment opportunities. Wealthier speculators are on the lookout for competent budgetaries to make stock price predictions. Retail financial professionals must be able to construct meaning of the marketplace on their own and make smart decisions on their own to do this. As a result, initiatives are extremely unsettling in today’s societal structures. Hidden structures for stock price forecasting can be discovered as a result of the increasing prevalence of techniques based on deep learning for forecasting the future price estimation in a variety of applications based on time series. This can be beneficial in providing additional insight to retail investors when deciding on which ventures to pursue. Forecasting based on time series is now a very difficult subject of investigation due to the huge potential it has in a range of applications including such stock market prediction, corporate strategy, weather prediction, resource allocation and a plethora of others. Despite the reality that forecasting may be thought of as a subset all unsupervised regression issues, some specific techniques are required due to the worldwide concept of perceptions in predicting. Multivariate statistical models such R. A. Mulla (B) RIMT University, Mandi Gobindgarh, Punjab, India e-mail: [email protected] S. Saini Department of ECE, RIMT University, Mandi Gobindgarh, Punjab, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_15

133

134

R. A. Mulla and S. Saini

as the autoregressive (AR), the autoregressive running average (ARMA) and the autoregressive integrated average (ARIMA) are some of the most commonly used [2–4]. Many different types of stock market analysis are debated and considered to be a unique type of time series by many people. Because of the complicated structure of time series models, standard methods that rely on linear regression methods may be unable to correctly detect some statistically significant trends. If these methods are utilised, a significant amount of the time series data will exhibit nonlinearity [5]. Furthermore, without the use of more robust and exceptionally nonlinear showing methodologies [6, 7], it is difficult to predict or estimate the financial transaction that will take place. We can anticipate the accurate number of a stock’s price by utilising deep neural algorithms even though the underlying data are exceptionally nonlinear in nature. To grow increasingly acquainted with such a functionality, a neural network seeks to map and emphasise the information that is necessary to produce a better predicted output [8], which is a sort of reinforcement learning. It is made up of a network of neurons that receive inputs in the form of a weighted sum. It is necessary to employ activation functions in order to suit the signal from neurons that acquaint the network with nonlinearity, and after that this nonlinearity trait is passed to some other neurons in the network. Streamlining a neural network is often accomplished through backpropagation, which is accomplished through the gradient descent algorithm. Through the use of this backpropagation method, errors are propagated from the output layer towards the input layer and vice versa. To anticipate nonlinear data in a variety of time series prediction problems, the deep learning method outperforms all other models by a wide margin. With the exception of the main memory, the gated recurrent unit (GRU) [9] method is a deep learning strategy that uses the very same sort of mechanism as the LSTM model. However, the GRU algorithm simplifies the construction of several components, such as the main memory. Because the GRU only has two gates, the order to have a valid and the update gate, it makes the transcription factors of memory easier to understand and operate on. The reset gate governs the data measures to make sure that they are not destroyed when new information is collected, meanwhile the update gate is in charge of ensuring that the memory cell’s scope of operation is maintained. Because it takes fewer needed to practise and works quite well on small amounts of data [10], GRU is a wonderful algorithm. However, when compared to GRU, LSTM has been shown to be more efficient, particularly when dealing with nonlinearity in big data sets [11]. Due to the fact that stock price forecasting frequently requires the analysis of huge nonlinear data sets, an LSTM-based scheme for stock price forecasting has been devised. Trying to predict the future value of a stock is a classic and significant problem. Through the use of a successful example for stock forecasting, it is possible to obtain a better understanding of the market’s behaviour over time. LSTM and other existing convolution neural network methods are merged with stock forecasting in this study, and public additional information is used to aid in the creation of the proposed design, which is a novel approach in the field of finance. The effectiveness of the LSTM algorithm was compared by adjusting various parameters in each method’s execution. This procedure consisted of determining the most

An Improved Stock Market Index Prediction System Based on LSTM

135

appropriate values for the parameters in order to anticipate the price of the stock that yielded the lowest RMSE. The remaining pieces of paper are arranged as follows. Section 2 discusses many similar studies that have been conducted. Section 3 provides a high-level overview of the suggested methodology. It is illustrated in Sect. 3.2 how the system architecture of our experiment was designed, and the outcomes are discussed in Sect. 4. After that in Sect. 4, the paper is brought to a conclusion.

2 Literature Review Artificial neural networks (ANNs) have been utilised by analysts to estimate stock market prices in a number of models (ANN). Using time series model analyses that are defined employing economic concepts and hypotheses in order to forecast a set of results is essential in the twenty-first century. These guidelines, but at the other hand, can indeed be applied successfully to the prediction of stock prices that are impacted by external factors such as the economy. The rapid advancement of the amount of layers idea [12] has enabled the adoption of artificial neural networks as forecasting methods rather than other approaches in the field of weather forecasting. Deep learning models are a type of machine learning strategy for modelling combination of these methods information translations, in which the structures produced are recursively altered to match supplied inputs with target outputs. Deep learning models are used to simulate intricate dynamic input–output mappings. A model known as the “multilayer perceptron” is one of the most effective and widely used predictive models available today. Several researchers believe that MLP is universal approximators unbounded functions [12]. When it comes to stock market forecasting, the author of [12] investigates the effectiveness of artificial neural networks, which are regarded to be dynamic and efficient. There are three models under consideration: the dynamic artificial neural network (DAN2), the multilayer perceptron (MLP) and hybrid neural networks that produce new input variables by using generalised autoregressive conditional heteroscedasticity (GARCH). On the basis of the presence NASDAQ stock exchange indexes regular rate values, two metrics are used to evaluate each model: mean absolute deviate (MAD) and mean square error (MSE). In today’s market, forecasting methodologies employ both linear and nonlinear algorithms (AR, ARIMA, MA), but the majority of them are focussed on anticipating stock market movement or price estimates for a specific firm based on the regular closing price of the stock market. In this article, researchers from [13] propose a solution that is not dependent on the model. It is deep learning techniques that are being used to characterise the fundamental characteristics in the information, rather than just fitting the data to a certain model. Researchers in this study compared the performance of three different neural network topologies for forecasting stock of NSE-listed companies in this study. A feature extraction framework was utilised

136

R. A. Mulla and S. Saini

to anticipate prospective prices in the near future. The accuracy of the models was determined by calculating their percentage error. The study [14] proposed the “long short-term memory” (LSTM) as a superior RNN variant, and this was found to be true. In recent years, the introduction of long short-term memory networks has improved the efficiency with which time-dependent data may be processed. With this update, you will be able to access information that was previously saved. The forecasting performance of a wavelet neural network (WNN) model is investigated in this study [15] utilising publicly available market data from the Taiwan Stock Exchange (TWSE) 50 index, commonly known as Taiwan 50 and the Financial Times Stock Exchange (FTSE). Particle swarm optimization (PSO) is employed in our WNN model to select the initial network parameters that are required for different types of organisations. There are two advantages to the outcome of the study. To begin, rather than being a constant, the network’s initial values are determined by a computer algorithm. Second, because PSO aids in the process of self-adjustment, the threshold and percentage of training data are both set to fixed values. The ability to obtain a performance rate of more than 73% without the need to manually adjust settings or create a new math model has been demonstrated. The writers [16] are attempting to forecast the short-term worth of these stocks. 10 different stocks that are traded on the New York Stock Exchange are taken into consideration in this analysis The study is largely concerned with anticipating these short-term pricing based on the quality of technological analysis in this area. Techniques such as technical analysis are used to influence the framework’s understanding of patterns based on historical prices fed into it, with the goal of making probabilistic predictions about the stock’s fleeting possible prices. Artificial neural networks (ANNs) are divided into two categories: feed forward neural networks and recurrent neural networks. Both of these forms of artificial neural networks are addressed in this article. When it comes to estimating asset values in the short term, the researchers revealed that feed forwards multilayer perceptron outperforms long short-term memory, according to their findings. For the Nifty and Sensex, the researchers used 11 technical measures, as well as an MLP, a radial basis functional network (RBFN) and an advanced radial basis functional neural network, to predict the market closing prices. The suggested model optimised RBF delivers better results when compared to MLP and RBF, as demonstrated in [17]. According to Yoshua, the book is a “practical handbook with suggestions for some of the most widely used hyper-parameters, notably in the context of learning algorithms based on back perpetuated regression and gradient-based optimization.” The qualities that were always required to effectively train and monitor large-scale data, as well as the features that were commonly required for deep multilayer neural network models, were also defined in the study [18]. Another first gradient-based optimization technique for stochastic optimization problems, called “Adam,” was introduced by the authors. Adam is based on adaptive predictions of lower-order moments and is another first gradient-based optimization technique. The evolutionary computing approach Adam has been demonstrated to

An Improved Stock Market Index Prediction System Based on LSTM

137

outperform other evolutionary computing approaches in testing [19]. A deep neural network aggregate was built using recent input indexes to estimate the Chinese stock index, which was then utilised to estimate the Chinese stock index. Historical data are used in this mission to train a succession of component networks using backpropagation and the Adam algorithm [20], which are then used to complete the objective. According to a review of the literature, the “multilayer perceptron” topology is one of the most often used neural network topologies, with the “multilayer perceptron” being the most commonly used. A sort of paradigm known as long short-term memory networks (LSTMNs) is a type of network that is capable of recalling information from the past. According to Wikipedia, regression is a statistical tool for determining relationships between variables. The Adam is frequently employed in the training of deep neural networks. There has been no research into the use of deep learning techniques to predict the performance of the Indian Stock Market. This is why the MLP and LSTM networks were employed in this experiment to approximate the closing price of a stock market index on the Indian Stock Exchange. The Adam optimizer will be used to train the models, and the regression model will be a mean square error regression model. With the help of simple regression models, the author of [21] developed a thesis that gave a quantitative way to predicting stock prices. The research not only covered linear regression, but it also discussed how linear regression might be used to forecast stock market prices. According to the findings of this study, selecting appropriate elements impacting stock price as parameters will result in specific predictions. In this study [22], the author presents a unique end-to-end heterogeneous computational model, a system consisting of multiple time scale feature learning, to forecast the future prices of the market index using multiple time scale feature learning. Combining two types of features on different time scales extracted from the raw daily price series by means of the first and second layers of the deep neural network (CNN) but the first and second layers of the hybrid neural network (HNN), the raw daily price series reflects features that are relatively short, medium, and long term in nature. With the use of empirical mode decomposition (EMD) and Hurst exponent (H) analysis, the author of this article [23] calculated the value of stock prices in the short-term and long-term ITH for twelve main global stock indexes as well as the stock prices of certain individual businesses. In [24], the author proposed a basic yet successful time-sensitive data augmentation approach for stock trend prediction that is both easy and successful. For example, in the context of wavelet transformation, they enhance data by damaging highfrequency patterns in the original stock price data whilst retaining low-frequency patterns inside the framework of wavelet transformation.

138

R. A. Mulla and S. Saini

Fig. 1 Architecture for LSTM module

3 Methodology 3.1 Long Short-Term Memory Networks However, except for the processing of input from each neuron, there are many similarities between the LSTM and MLP algorithms in terms of the mechanism used. “In contrast to the production of regular neurons, the production of an LSTM cell is the result of a multistep procedure. It is the cell state that serves as an extra memory in LSTMs, storing important past knowledge that can be used to aid in prediction. It is during the following steps that the information contained in the cell state is changed by mechanisms known as gates. At first, the forget gate determines whether or not to discard any data that are currently available.” The input gate and the tanh layer would be in charge of determining which new data would be required to be saved. When it comes to the previous gates, the detail is once again an afterthought that is ignored or overlooked. Last but not least, the data are subjected to the activation function, after which the output is produced (Fig. 1).

3.2 System Architecture and Data Sets Data are collected from annual reports of the companies, which is a freely accessible source of information. During the time between 19/08/2008 and 04/10/2010, data from the Yahoo stock market were collected and used for the purposes of conducting this experiment. The numeric format of the supplied data is used throughout this document. The data consist of the beginning value, high value, low value and closure value of the stock on a daily basis, which is used to forecast future data. A total of more over 700 days of data have been used. In our research, we employ Google colaboratory as a simulation environment, which includes a GPU, Windows 10 operating system and 8 GB of RAM. TensorFlow is the deep learning framework that we are using. The data set was first normalised using minmax feature scaling, which is a feature scaling algorithm. After that the processing data set is divided segments, which are

An Improved Stock Market Index Prediction System Based on LSTM

139

Fig. 2 Proposed architecture of stock price prediction

referred to as the training data set and the testing data set. The data set is divided into two parts: the training data set, which contains 88% of the data and the testing data set, which contains 12% of the data. When the training data set was ran through both the LSTM models for different tuning settings and the anticipated stock prices were produced, it was considered a success. A comparison was made between the anticipated data set as well as the testing information, and the reliability of the projection was determined. Figure 2 depicts the system architecture of our scheme, as well as the nonlinear activation architecture of our scheme.

3.3 Result Analysis The results of our study showed that utilising different layers, varied unit sizes in the convolutional nodes and packed layers, and also different supervised learning here between actual statistics and projected data, we could see different forms of root mean square error (RMSE). After examining several training epochs for the LSTM model system, it was discovered that now the learning epochs should be selected in the most efficient manner possible to train the model. As can be observed in Table 1, the 100 historical periods training model has a reduced root mean square error (RMSE), which results in the highest prediction accuracy when compared to the Table 1 Analysis of RMSE versus the number of epochs with respect to time

No. of epoch

LSTM RMSE

Time

10

0.0011

3

20

0.000725

6

50

0.000493

15

100

0.000493

30

250

0.003198

75

140

R. A. Mulla and S. Saini

RMSE vs. Epoch for LSTM 0.004 0.002 0 0

50

100

150

200

250

300

Fig. 3 Analysis of RMSE with changing no. of epoch

other models. The activation function ReLU was utilised in this case to activate each hidden layer, with 128 units being employed for each hidden layer. For the purpose of developing an optimal forecasting models, we decided to use one day advanced indexes data to anticipate the price trend of the following trading day, rather than using the method of performing the analysis to time series. Using a short-term time frame ranging from one day to one week (5 trading days), we evaluated how the technical indices associated with price movements using the technique described in this paper. Following the test, we discovered that the length of the term has a varied level of sensitivity to the same set of indices depending on its length (Fig. 3). It can be seen in Table 1 that if the number of epochs is increased further, the training model encounters a problem with fitting at a certain point, such as 250 epochs, when the LSTM is being trained. As illustrated in Fig. 4, this is also true. In addition, it seems to be that using LSTM, the effectiveness of the assessment accuracy decreases with the rising number of hidden layers in the model. Furthermore, it extends the amount of time spent on training. Table 2 presents a summary of these findings, respectively. These can be seen in greater detail in Fig. 5 for 50 epochs using a LSTM network with two and four hidden layers, respectively. RMSE wrt Time 0.0035 0.003 0.0025 0.002 0.0015 0.001 0.0005 0 0

Fig. 4 Analysis of RMSE with time

20

40

60

80

An Improved Stock Market Index Prediction System Based on LSTM

141

Table 2 Analysis of RMSE with respect to variable hidden layers using LSTM models No. of epoch RSME with two hidden Time in mins RSME with four hidden Time in mins layers layers 20

0.000735

6

0.000985

50

0.000536

15

0.000542

35

100

0.000498

30

0.000623

70

Fig. 5 Analysis of RSME for LSTM with variable hidden layers

12

RSME for LSTM with Variable Hidden Layers 0.0015 0.001 0.0005 0 0

20

40

60

80

100

120

RSME with 2 Hidden Layers RSME with 4 Hidden Layers

4 Conclusion Based on this research, it can be concluded that deep learning techniques have had such a profound impact on society technology, especially in the development of varied time series-based prediction models of varying complexity. If you compare them to any other regression models for time series prediction, they may generate a better level of accuracy possible. The LSTM model, which is amongst the various deep learning models, can be utilised for stock market prediction with the right modification of various parameters. The adjustment of these factors is critical in the development of any type of prediction model, as the accuracy of the forecast is highly dependent on these parameters when making predictions. As a result, careful parameter tweaking is required for LSTM models as well. The stock market can be forecasted using our proposed LSTM prediction model, which may be employed by both people and corporations as a result of our findings. The ability to obtain a big financial profit whilst also preserving a steady atmosphere in the financial markets can be quite beneficial to investors. In order to evaluate the efficacy of our strategy to be effective, we want to examine data from a bigger number of commercial markets that fall into a bunch of alternative categories.

References 1. Market capitalization of listed domestic companies-world. (Online). Available: https://data. worldbank.org/indicator/CM.MKT.LCAP.CD?locations=1W. Accessed 17 June 2020

142

R. A. Mulla and S. Saini

2. D.K. Kılıç, Ö. U˘gur, Multiresolution analysis of s&p500 time series. Annal. Oper. Res. 260(1– 2), 197–216 (2018) 3. P. Li, C. Jing, T. Liang, M. Liu, Z. Chen, L. Guo, Autoregressive moving average modeling in the financial sector, in 2015 2nd International Conference on Information Technology, Computer, and Electrical Engineering (ICITACEE) (IEEE, 2015), pp. 68–71 4. G. Zhang, X. Zhang, H. Feng, Forecasting financial time series using a methodology based on autoregressive integrated moving average and Taylor expansion. Expert. Syst. 33(5), 501–516 (2016) 5. M. Bildirici, Ö.Ö. Ersin, et al., Nonlinearity, volatility and fractional integration in daily oil prices: Smooth transition autoregressive st-fi (AP) garch models. Rom. J. Econ. Forecast. 3, 108–135 (2014) 6. I. Kaastra, M. Boyd, Designing a neural network for forecasting financial. Neurocomputing 10, 215–236 (1996) 7. A. Lendasse, E. de Bodt, V. Wertz, M. Verleysen, Non-linear financial time series forecastingapplication to the bel 20 stock market index. Eur. J. Econ. Soc. Syst. 14(1), 81–91 (2000) 8. D.P. Mandic, J. Chambers, Recurrent neural networks for prediction: learning algorithms, architectures and stability (Wiley, 2001) 9. K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, Y. Bengio, Learning phrase representations using rnn encoder-decoder for statistical machine translation (2014). arXiv preprint arXiv:1406.1078 10. J. Chung, C. Gulcehre, K. Cho, Y. Bengio, Empirical evaluation of gated recurrent neural networks on sequence modeling. (2014) arXiv preprint arXiv:1412.3555 11. G. Weiss, Y. Goldberg, E. Yahav, On the practical computational power of finite precision RNNs for language recognition (2018). arXiv preprint arXiv:1805.04908 12. E. Guresen, G. Kayakutlu, T.U. Daim, Using artificial neural network models in stock market index prediction. Expert Syst. Appl. 38(8), 10389–10397 (2011) 13. S. Selvin, R. Vinayakumar, E.A. Gopalakrishnan, V.K. Menon, K.P. Soman, Stock price prediction using LSTM, RNN and CNN-sliding window model, in 2017 International Conference on Advances in Computing, Communications and Informatics, ICACCI 2017, vol. 2017–Janua (2017), pp. 1643–1647 14. S. Hochreiter, J. Urgen Schmidhuber, Long short-term memory. Neural Comput. 9(8), 1735– 1780 (1997) 15. K. Wang, C. Yang, K. Chang, Stock prices forecasting based on wavelet neural networks with PSO, vol. d (2017) 16. K. Khare, O. Darekar, P. Gupta, V.Z. Attar, Short term stock price prediction using deep learning, in 2017 2nd IEEE International Conference on Recent Trends in Electronics, Information & Communication Technology (RTEICT) (2017), pp. 482–486 17. R. Mahanta, T.N. Pandey, A.K. Jagadev, S. Dehuri, Optimized Radial Basis Functional neural network for stock index prediction, in International Conference on Electrical, Electronics, and Optimization Techniques, ICEEOT 2016 (2016), pp. 1252–1257 18. Y. Bengio, Practical recommendations for gradient-based training of deep architectures, in Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 7700 LECTU. (2012), pp. 437–478 19. D.P. Kingma, J. Ba, Adam: a method for stochastic optimization (2014), pp. 1–15 20. Y. Bing, J.K. Hao, S.C. Zhang, Stock market prediction using artificial neural networks. Adv. Eng. Forum 6–7(June), 1055–1060 (2012) 21. A. Sharma, D. Bhuriya, U. Singh, Survey of stock market prediction using machine learning approach, in International Conference on Electronics, Communication and Aerospace Technology ICECA (2017) 22. Y. Hao, Q. Gao, Predicting the trend of stock market index using the hybrid neural network based on multiple time scale feature learning. Appl. Sci. 10, 3961 (2020). https://doi.org/10. 3390/app10113961

An Improved Stock Market Index Prediction System Based on LSTM

143

23. A. Mahata, M. Nurujjaman, Time scales and characteristics of stock markets in different ınvestment horizons. Front. Phys. 8, 590623 (2020). https://doi.org/10.3389/fphy.2020.590623 24. X. Teng, T. Wang, X. Zhang, L. Lan, Z. Luo, Enhancing stock price trend prediction via a time-sensitive data augmentation method. Complexity 2020, 8. Article ID 6737951 (2020). https://doi.org/10.1155/2020/6737951

Age Estimation in Digital Radiograph Using HOG and DWT Feature Extraction A. Stella and Thirumalai Selvi

1 Introduction Image processing is a field of computer science that encompasses variety of techniques which are used to manipulate and enhance data, in the form of specifically images for analysis. Image analysis can be in different formats like text, bitmap, photographs or patterns, etc. The acquired images for the analysis can be in monochrome or grayscale or can be acquired in all possible colors. The success of any image processing software or algorithm is in identifying the appropriate feature to analyze in the images and further enhancing them for image processing to get a repeatable and accurate identification of the solution. The science of computing has revolutionized every aspect of our day-to-day life. Both medical and computing fields have evolved enormously throughout the decades. Every human being born in this world has a date of birth. In this world torn apart by war and acts of terror, it so happens that there are unfortunately some who do not know the day they were born, refusing them chances of livelihood as they cannot prove their age for any legal or survival benefits. A court of law recognizes dental age as proof of maturity [1]. Nolla in 1960 and Demirjian in 1973 used a specialized dental radiograph to assess the approximate chronological age of the individual. They classified the developmental stages of the tooth, of the permanent dentition in one quadrant of the individual to calculate a score, and this score corresponds to the approximate age, as given by the Nolla’s and Demirjian’s table [2, 3]. These methods are almost accurate in assessing the age if the correct tooth stages are identified. However, there is an interobserver variability in identifying the proper A. Stella (B) Department of Computer Science, Bharathiar University, Coimbatore, India e-mail: [email protected] T. Selvi Department of Computer Science, Government Arts College for Men, Nandanam, Chennai, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_16

145

146

A. Stella and T. Selvi

stages of tooth development to calculate an individual’s age. Image processing was considered as an excellent answer to this problem. Humans have dual dentition namely, the primary and the secondary dentition. In the secondary dentition or permanent dentition, there are 32 teeth. The upper jaw (maxilla) has 16 teeth, and the lower jaw (mandible) has 16 teeth. The maxilla and mandible are divided into left and right quadrants. There are two incisors, one canine, two premolars, and three molars in each quadrant. All the 32 teeth do not erupt at the same time; they develop in a chronological order. Demirjian’s technique divides the time from which a tooth begins to form till its roots are completely formed into eight stages and denote them as stages A to H. Nolla’s technique divides the same process into ten stages and numbers them from stages 0–10. As mentioned earlier Nolla’s and Demirjian’s, the studies have enabled the calculation of an individual’s age by observing these stages of tooth development.

2 Related Works This research is focused on feature extraction techniques to identify the stages in Nolla’s and Demirjian’s of tooth development. Ali Bagherian and Mostafa Sadeghi concluded that Demirjian method’s accurate in an Iranian population. They analyzed the orthopantomography of 519 healthy children aged 3.5–13.5 years and calculated their dental ages. The children’s chronological ages were determined by subtracting their birth dates from the date of the radiography [4]. Gupta et al. [5] described that to assess the accuracy of Demirjian’s dental age estimation we need regression formulas and India-specific formulas. The study will used 50 radiographs, mostly pre-treatment orthodontic radiographs of healthy patients. Avinash et al. [6] discussed that lung cancer affects 2.6 million individuals and kills 1.8 million. Lung cancer is caused by smoking and air pollutants. He developed a novel technology using Gabor filter for image enhancement that can instantly detect malignant cells in patients diagnostic images. Putra et al. [7] proposed a novel mammography classification scheme for normal and pathological breast tissues. The ROI of a mammogram is used to construct a feature matrix using local binary pattern. A neural network classifier uses the best features to identify the lesion. Bendjillali et al. [8] used Viola-Jones face detection, discrete wavelet transform, histogram equalization (HE) method, and deep convolution neural network; to present a face recognition system. This network’s face recognition rate is 99.85 and 99.80%. Radman et al. [9] has used unconstrained iris segmentation as a challenge for iris recognition. It is tested on the UBIRIS.v2 and MICHE iris databases. This method accurately locates the iris using histograms of oriented gradients (HOG). Ahamed et al. [10] discussed that biometric face recognition systems use digital images to identify or verify people. His research provides a HOG-CNN face recognition deep neural network design. His paper’s purpose is real-time facial identification via webcam, image, or video. An autonomous vehicle detection system based on vision was proposed [11]. A KNN classifier plus a vehicle detection algorithm comprise of detection algorithm and Car

Age Estimation in Digital Radiograph Using HOG …

147

is identified by comparing testing and training data. The system will identify cars from other vehicles [11].

3 Dataset In this study of medical image processing, OPG images of individuals in the age group below 16 years were acquired from two pediatric dental centers. The images were from old records and were acquired to treat other dental ailments. Out of these records, 77 records were used to create a training dataset, and ten records were used to evaluate the program. Each OPG was used to segment the images of the seven permanent teeth in the right lower quadrant into various stages of tooth development and segregate them according to the types of teeth, namely central incisor, lateral incisor, canine, first premolar, second premolar, first molar, and second molar. Each tooth image has been classified according to the various stages of tooth development, classified by Nolla’s and Demirjian’s methods. Totally there were 539 images classified into Nolla’s stage 1–10 and Demirjian’s stage A to H (539 images).

4 Materials and Methods A MATLAB-based feature extraction program was planned to extract and enhance the grayscale images obtained from an OPG. The orthopantomograms will be anonymous except for the patient’s sex and age. The OPG will be assessed and evaluated by a dental specialist who has experience identifying the various stages of tooth development in the two main techniques being evaluated, namely the Nolla’s and Demirjian’s techniques. Training datasets are created with the help of the specialist to identify the stage of the tooth using Nolla’s and Demirjian’s methods. Figure 1 shows the overall work of the research work to identify tooth stage and age of the given OPG.

4.1 Preprocessing The preprocessing of the image is done by extracting the image from OPG using the ROIPOLY tool and then converting the image to a grayscale matrix of unsigned integers. Then, based on the threshold of value greater than 50, the images are converted into binary images. These processed images undergo further feature extraction.

Fig. 1 The architecture of research work

148 A. Stella and T. Selvi

Age Estimation in Digital Radiograph Using HOG …

149

4.2 Histogram of Gradients Feature Vector (HOG) HOG, or histogram of oriented gradients [12], is a feature descriptor frequently used to extract feature content from an image. A feature descriptor is an image descriptor that enhances the image by obtaining required data and giving off non-essential data [13, 14]. HOG descriptors can be calculated in horizontal and vertical gradients. Each pixel in the image is calculated by the horizontal gradient (X-direction) and vertical gradient (Y-direction). Each image is divided into a cell of 8 × 8 pixels, and the HOG image is visualized by sketching normalized (9 × 1) histograms in 8 × 8 cells. Both the gradient direction and magnitude are expressed as the histogram of gradients. The pixel values are added to the nine orientation bin of 8 × 8 cells.

4.3 Discrete Wavelet Transform (DWT) A DWT creates sets of coefficients to define signal features and approximations at various scales [8, 15, 16]. Thus, the coefficient determination is essential to decrease the feature size. The images were decomposed n times to ensure reliable identification of the stages of tooth development. Discrete wavelet transform (DWT), a robust tool for feature extraction, was applied to extract coefficients of wavelets from OPG images. 2D discrete wavelet transform was used, which decomposed the pixels into four sub-bands LL, HL, LH, and HH. Here, L holds for low, and H holds for high. The Coiflet1 filter decomposition of the given image is shown in Fig. 2. Single-level discrete 2D wavelet transformation helps in retrieving more energy than multiple decompositions. The features extracted are then constructed into matrix vectors. Fig. 2 Coiflet 1 filter

150

A. Stella and T. Selvi

Fig. 3 Decomposition step of DWT

Figure 3 shows the downsampling of columns and rows and decomposition of the extracted image in approximation coefficients in horizontal, vertical, and diagonal orientations.

4.4 Hybrid Method (HOGDWT) To enhance the extracted feature, for identification of the tooth a combination of two extracted feature vectors was tried, which gave the hybrid method to identify the stage of the tooth [17–19]. HOG feature vector and DWT feature vector are combined to give the HOGDWT feature vector as shown in Fig. 4. HOGWT feature vectors are extracted for both training dataset and testing dataset to predict the stage of the tooth using the classifier. Fig. 4 Hybrid method

Age Estimation in Digital Radiograph Using HOG …

151

4.5 Classifiers The accuracy of support vector machine (SVM) [20] and K-nearest neighbor (KNN) [21] classifiers are compared in this study. Both multiclass model classifiers predict tooth stage. Classifiers will be tested for accuracy and speed in recognizing tooth development stages. The SVM classifier [22] outperforms KNN. The computerized program identifies the stage of tooth development which is compared to the developmental stage of the given tooth image as identified by the dental expert. The difference was estimated as the number of developmental stages between experimental and actual stages. For example, an actual stage “C” will be two stages behind if considered stage “A” and given a score of −2, and two stages ahead if considered stage “E” and given a score of 2. The accuracy of the SVM and KNN classifiers was determined by their speed and accuracy in identifying tooth development stages and calculating age in the testing OPG dataset.

5 Results and Discussion Error difference was calculated to identify the accuracy of identifying stage of the tooth. 10 OPG images are tested to find the accuracy. The difference of error is calculated individually for Nolla’s techniques and Demirjian’s technique as they have different number of stages in them. The ROI tool was used to separate the region of interest from other images so that the image processing will target on the problem at hand. The conversion of the image into a binary dataset is justified as the medical radiograph data is a grayscale dataset, and the contrast will help in accurate diagnosis. The DWT feature extraction helped to reduce the noises from the region of interest. Since the image is a grayscale image with the tooth structure represented as white, it cannot take decomposition of more than one step. Comparing SVM and KNN as classifiers, the KNN is an effective tool in RGB image processing, and its technique may not be advantageous in grayscale image processing. SVM outperforms KNN. Because the OPG pictures were radiographs taken for other dental treatments, the age range of the dataset was 3–16 years, excluding the early stages of central and lateral incisor development. This may be the reason for errors in identifying the developmental stage of these teeth. The increased errors in Nolla’s technique may be because the same process of tooth development has been divided into more stages than Demirjian’s technique. This may complicate the identification process and cause overlap of the stages, thereby causing errors in identification. The hybrid technique with SVM classifier for Nolla’s age estimation method included mistakes in identifying central incisor, lateral incisor, first premolar, and second molar teeth. A developmental stage in the incisors, two in the first premolar and two in the second molar, separated the errors. In lateral incisors and second molars, the SVM classifier’s hybrid technique had errors of one stage ahead and one stage behind. The KNN classifier gave mistakes in central and lateral incisors, canines, second

152

A. Stella and T. Selvi

premolar, and second molar in one step ahead. The KNN classifier error resulted in identical errors of one developmental stage forward for the teeth central incisor, lateral incisor, canine, second premolar, and second molar, as shown in Tables 1, 2, and 3. Compared to SVM and KNN classifiers, the SVM classifier gives better accuracy in identifying the stage of the tooth. Table 4 shows the error percentage for ten images for both Nolla’s (NM) and Demirjian’s (DM) to identify the stage of the tooth. Table 1 Error difference in identifying the tooth stage using Nolla’s and Demirjian’s method (SVM and KNN-hybrid method feature extraction) Feature extraction

OPG Img1 Img2 Img3 Img4 Img5 Img6 Img7 Img8 Img9 Img 10

SVM-hybrid NM method DM

I1

NM

I2

1

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

1

0

0

0

0

0

0

0

0

0

1

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

DM

0

0

0

0

0

0

0

0

0

0

NM

PM1 2

0

0

0

0

0

0

0

0

0

DM

0

0

0

0

0

0

0

0

0

0

NM

PM2 0

0

0

0

0

0

0

0

0

0

DM

0

0

0

0

0

0

0

0

0

0

DM NM

NM

C

M1

DM NM

M2

DM KNN-hybrid NM method DM

I1

NM

I2

DM NM

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

2

0

0

0

0

0

0

0

0

0

1

0

0

0

0

0

0

0

0

1

0

0

0

0

0

0

0

0

0

1

0

0

0

0

0

0

0

0

0

1

0

0

0

0

0

0

0

0

0

1

0

0

0

0

0

0

0

0

0

0

1

0

0

0

0

0

0

0

0

DM

0

1

0

0

0

0

0

0

0

0

NM

PM1 0

0

0

0

0

0

0

0

0

0

DM

0

0

0

0

0

0

0

0

0

0

NM

PM2 1

1

0

0

0

0

0

0

0

0

DM

1

1

0

0

0

0

0

0

0

0

NM

C

0 0

M1

DM NM DM

M2

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

0

1

0

0

0

0

0

0

0

0

0

1

0

0

0

0

0

0

0

0

Age Estimation in Digital Radiograph Using HOG …

153

Table 2 Percentage of various features in Nolla’s and Demirjian’s method Features

WT

HYBRID-1

HOG

tooth

NM

DM

GW NM

DM

NM

DM

NM

DM

NM

HYBRID-2

I1

60

60

60

70

60

60

90

90

90

100

I2

60

60

60

60

60

60

100

100

90

90

C

60

80

80

90

50

50

100

100

100

100

PM1

70

70

80

70

60

50

90

90

90

100

PM2

60

70

80

60

50

60

90

90

100

100

M1

70

70

80

80

60

70

100

100

100

100

M2

80

70

80

70

60

70

90

90

90

90

DM

Table 3 Accuracy of various features in Nolla’s and Demirjian’s method FEATURE

WT

GW

HYBRID-1

HOG

HYBRID-2

Nolla’s method

65.71

74.29

57.14

94.29

94.29

Demirjian’s method

68.57

71.43

60

94.29

97.14

Table 4 Percentage of SVM and KNN classifier Classifier

I1

I2

C

PM1

PM2

M1

M2

SVM-HOGWT(N)

90

90

100

90

100

100

90

KNN-HOGWT(N)

90

90

90

100

90

100

90

SVM-HOGWT(D)

100

90

100

100

100

100

90

KNN-HOGWT(D)

90

90

90

100

90

100

90

From Table 5, the average accuracy of each method using the two classifiers was calculated. The SVM classifier gave better accuracy rates for both the methods of age estimation. But the highest rate of accuracy was for Demirjian’s technique at 97.14% Once the accuracy of identifying the developmental stage for the individual teeth was established, we tried to calculate the age using Demirjian’s table. We found that the age assessment using the hybrid feature extraction technique of combining HOG and DWT using SVM classifier and Demirjian’s method leads to 98% accuracy to Table 5 Average accuracy of classifiers

Classifier

Average accuracy

SVM-HOGDWT(N)

94.29

KNN-HOGDWT(N)

92.86

SVM-HOGDWT(D)

97.14

KNN-HOGDWT(D)

92.86

154 Table 6 Accuracy of age using SVM classifier

A. Stella and T. Selvi OPG

Accuracy of age: SVM (Nolla)

Accuracy of age: SVM (Demirjian)

Image 1

94.3

95.2

Image 2

98

100

Image 3

92.6

94.1

Image 4

98.3

100

Image 5

98.7

100

Image 6

99.2

100

Image 7

99.4

100

Image 8

91.2

93

Image 9

97.3

98

Image 10

96.8

100

a range between 93 and 100 in age estimation using digital radiographs shown in Table 6.

6 Conclusion The estimation of age was calculated by Demirjian’s method was studied for their accuracy. The study was performed with the HOGDWT feature extraction and SVM classifier. Demirjian’s technique was determined to be the most suitable and high accuracy method to determine an individual’s age using the orthopantomogram. The results can be more accurate as the image library or training dataset grows. Since all medical images are similar grayscale images, these feature extraction techniques of HOG combined with DWT and SVM classifier for a pattern or image recognition is a valuable tool for identifying normal anatomy or pathology in medical radiographs.

References 1. D. Franklin, A. Flavel, J. Noble, L. Swift, S. Karkhanis, Forensic age estimation in living individuals: methodological considerations in the context of medico-legal practice. Res. Rep. Forensic Med. Sci. 53 (2015) 2. C.A. Nolla, The development of the permanent teeth. J. Dent. Child. Fourth Qua. 254–266 (1960) 3. A. Demirjian, H. Goldstein, J.M. Tanner, A new system of dental age assessment. Hum. Biol. 45(2), 211–227 (1973) 4. A. Bagherian, M. Sadeghi, Assessment of dental maturity of children aged 3.5 to 13.5 years using the Demirjian method in an Iranian population. J. Dent. (Sh¯ır¯az, Iran) 53(1), 37–42, (2011) 5. R. Gupta et al., Dental age estimation by Demirjian’s and Nolla’s method in adolescents of Western Uttar Pradesh. J. Head Neck Phys. Surg. 3(1), 50–56 (2014)

Age Estimation in Digital Radiograph Using HOG …

155

6. S. Avinash, K. Manjunath, S. Senthilkumar, Analysis and comparison of image enhancement techniques for the prediction of lung cancer, in RTEICT 2017—2nd IEEE International Conference on Recent Trends in Electronics, Information and Communication Technology, Proceedings (2017) 7. J.A. Putra, Mammogram classification scheme using 2D-discrete wavelet and local binary pattern for detection of breast cancer. J. Phys. Conf. Ser. (2018) 8. R.I. Bendjillali, M. Beladgham, K. Merit, Face recognition based on DWT feature for CNN, in ACM International Conference Proceeding Series (2019) 9. A. Radman, N. Zainal, S.A. Suandi, Automated segmentation of iris images acquired in an unconstrained environment using HOG-SVM and GrowCut. Digit. Signal Process. A Rev. J. (2017) 10. H. Ahamed, I. Alam, M.M. Islam, HOG-CNN based real time face recognition, in 2018 International Conference on Advancement in Electrical and Electronic Engineering, ICAEEE 2018 (2019) 11. F.A.I. Achyunda Putra, F. Utaminingrum, W.F. Mahmudy, HOG feature extraction and KNN classification for detecting vehicle in the highway. IJCCS (Indonesian J. Comput. Cybern. Syst. (2020) 12. R.E.A.M. Jampour, Efficient handwritten digit recognition based on histogram of oriented gradients and SVM. Int. J. of Comp. Appl. 104, 10–13 (2014) 13. M. Davis, F. Sahin, HOG feature human detection system, in 2016 IEEE International Conference on Systems, Man, and Cybernetics, SMC 2016—Conference Proceedings (2017) 14. N. Dalal, Histogram of oriented gradients (HOG) for object detection in images 20110926 (2011) 15. S. Akbar et al. Face recognition using hybrid feature space in conjunction with support vector machine. J. Appl. Environ. Biol. Sci. 5(7), 28–36 (2015) 16. M.Z. AL-Dabagh, D.F.H. AL-Mukhtar, Breast cancer diagnostic system based on MR images using KPCA-wavelet transform and support vector machine. Int. J. Adv. Eng. Res. Sci. (2017) 17. G.S. Hong, B.G. Kim, Y.S. Hwang, K.K. Kwon, Fast multi-feature pedestrian detection algorithm based on histogram of oriented gradient using discrete wavelet transform. Multimed. Tools Appl. 75(23), 15229–15245 (2016) 18. S. Nigam, R. Singh, A.K. Misra, Efficient facial expression recognition using histogram of oriented gradients in wavelet domain. Multimed. Tools Appl. 77(21), 28725–28747 (2018) 19. A. Gumaei, R. Sammouda, A.M. Al-Salman, A. Alsanad, An effective palmprint recognition approach for visible and multispectral sensor images. Sensors (Switzerland) 18(5) (2018) 20. B. Li, B. Wang, Real and fake label image classification algorithm based on hog and svm, International Conference on Intelligent Transportation, Big Data & Smart City, ICITBS 2020 (2020), pp. 905–909 21. J.S. Raikwal, K. Saxena, Performance evaluation of SVM and K-nearest neighbor algorithm over medical data set. Int. J. Com. Appl. 50(14), 35–39 (2012) 22. V. Punithavathi Dr. D. Devakumari. A hybrid algorithm with modified SVM and KNN for classification of mammogram images using medical image processing with data mining techniques. Eur. J. of Mol. Clin. Med. 7(10), 2956–2964 (2021)

Single-Layer-Single-UWB Patch Antenna for HXLPE-Based Artificial Hip Diagnosis in Microwave Tomography Spectrum Khalid Ali Khan , Suleyman Malikmyradovich Nokerov , Aravind Pitchai Venkataraman , Kehali Anteneh , and Diriba Chali

1 Introduction Globally, the most common and versatile modalities that are found for biomedical imaging and scanning are X-ray screening, ultrasound imaging, positron emission tomography (PET), computed tomography (CT) scans, and magnetic resonance imaging (MRI) scanning. Moreover, microwave-based imaging technologies are increasing exponentially because they offer non-ionizing radiation as well as noninvasive characteristics. Furthermore, the microwave tomography method provides a complementary method to diagnose human body organs and health [1–3], sensing and imaging the tissue’s abnormalities such as breast cancer, brain tumor detection [4], physiotherapy [5], and so on. In other words, vigorous use of non-ionizing forms of electromagnetic waves on the human body for microwave imaging or for monitoring the human organs avoids dangerous effects on the patient’s health. But in the case of total hip arthroplasty (THA), component positioning such as acetabular offset, cup orientation, and femoral stem positioning must be known and investigated for successful clinical outcomes or occurrence of complications. Deep knowledge of joint anatomy and biomechanics relies on the vital signals that are collected by the microwave processing unit before image projection on the screen. Therefore, a wideband or ultra-wideband (UWB) antenna is required to achieve the high-definition images even at the lowest peak signal-to-noise ratio (PSNR). Consequently, a microwave tomography (MWT) antenna design, development of

K. A. Khan · K. Anteneh · D. Chali Mettu University, Mettu, Ethiopia S. M. Nokerov Oguz Han Engineering and Technology University of Turkmenistan, Ashgabat, Turkmenistan A. P. Venkataraman (B) Saranathan College of Engineering, Trichy, Tamil Nadu, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_17

157

158

K. A. Khan et al.

microwave imaging algorithms, and research on its implementation are the new challenge in biomedical engineering [6]. Various injuries, infections, diverse forms of arthritis, and other causes will lead the joints to become deformed and, as a result, damaged. As a rule, joint diseases associated with the aforementioned problems are more pronounced in the hip joint than in the knee joints or ankle joints [7]. One of the options for getting rid of excruciating pains and ongoing discomfort in the hip joint is surgery. This kind of surgery operation is called total hip replacement (THR) or total hip arthroplasty (THA). THA is considered one of the most important surgeries in the modern history of hip joint diseases. Clinical results over the past 30 years have shown that THA is one of the most effective and successful surgical methods for treating various pathological conditions of the hip joint. To provide the best possible quality without failures of THA surgery, from the start of THA surgery history, various types of biomaterials for the artificial hip joint have been developed and used and going on. The development of optimal biomaterials for implantology is one of the most challenging tasks of the century in the corresponding field. The developed optimal biomaterials depend on the subsequent cheerfulness of the recovered THA patients. Nowadays, there are five types of bearings used in THA. They are: metal-on-metal (MoM), metal-onpolyethylene (MoP), ceramic-on-polyethylene (CoP), ceramic-on-ceramic (CoC), and ceramic-on-metal (CoM). Recently THA prostheses are made from polymers (polyethylene—PE, ultra-high molecular weight polyethylene—UHMWPE, crosslinked polyethylene—XLPE, highly cross-linked polyethylene—HXLPE, polyetherether-ketone—PEEK) [8, 9]. More specifically, the combined application of an HXLP cup with a ceramic femoral head has proved good clinical results. Hence, this material is widely in use bearing surfaces in THA in the United States of America. According to the analysis report of the US market (as per 2015), it came into notice that ceramic-on-HXLPE, metal-on-HXLPE accounted for more than 90%. Hence, long-term clinical results assure that HXLPE combination with ceramic or metal is the best choice in all. Thus, by considering the glorious future of the HXLPE, an ultra-wideband MWT antenna for HXLPE-based artificial hip monitoring is designed over the HXLPE substrate material to meet the highest degree of impedance matching to get the low return loss in the operational frequency band of tomography or the frequency range of 1.74–4.06 GHz.

2 Model for Radiographic Appearance and Antenna Designing 2.1 Total Hip Arthroplasty (THA) Imaging System Model THA imaging setup as depicted in Fig. 1 mainly consist of a planar UWB patch antenna to transmit and receive the microwave signal, a microwave signal generator

Single-Layer-Single-UWB Patch Antenna …

159

Fig. 1 THA imaging system model

and so-called microwave transceiver, and a signal storing cum processing unit collaborated with the personal computer for image formation and detection. Imaging and scanning of THA are accomplished by changing the antenna direction in a different route and reflected microwave signals are collected and converted from frequency domain to time domain. Now, converted time-domain signals are processed and will be applied in confocal microwave image reconstruction algorithm (CMIRA) as discussed in [10] to get the femoral implant including acetabular cup visuals in THA by radiographic appearance. In any microwave-based imaging device, the antenna plays a vital role in imaging performance. A well-designed antenna bears accountability of large bandwidth with lower return loss, simple geometric structure, small dimensions, compactness, and ease of integration, and so on [11, 12]. Therefore, the aforesaid properties will be covered only by a patch antenna as it possesses all advantages.

2.2 Antenna Design and Specifications As it is shown in Fig. 2, the proposed patch antenna is the meander-shaped singlelayered patch mounted over the 17.5 mm thicker HXLPE (∈r = 2.2, loss tangent = 0.0004) substrate with the dimension of 32 mm × 33 mm. Pure copper (conductivity = 5.8 × 107 s/m) material is used as a radiating element with thickness of 0.035 mm and 50  coaxial cable is used to feed the antenna by direct contact method. In four legs (n = 4) meander structure, first leg (L 1 ) and last leg (L 4 ) are equal to each other and have the length of 21.0 mm, whereas central second (L 2 ) and third (L 3 ) legs have the length of 32.0 mm each. All four legs are equally spaced by the distance of 3.0 mm having width of 6.0 mm. The number of legs (n) and its width is the controlling parameters to change the antenna radiation characteristics such as bandwidth, the number of bands, and return loss (S 11 ) in a specific band at constant substrate height and its respective relative permittivity. The optimized geometrical dimension of the meander structure is suitable and enough to meet our requirement for transplanted artificial hip imaging and

160

K. A. Khan et al.

Fig. 2 a Four-leg meander structure. b Proposed antenna specification (L 1 = L 4 = 21.0 mm, L 2 = L 3 = 32.0 mm)

diagnosis. In the design of a novel patch, the height (h) of HXLPE is recommended as 15.5 mm ≤ h ≤ 19.5 mm to achieve the single UWB feature of the operational antenna in between 1 and 4 GHz. Furthermore, it can be seen in Sect. 3.1 that a larger number of meander legs with the same height and width may shift the resultant band either left or right to the earlier band without altering the UWB feature of the antenna in the same range of frequency (1–4 GHz). However, the thickness of the substrate severely affects the antenna performance, and it is universally known that thicker substrate enhances the antenna efficiency and bandwidth, but in coupling, a surface wave is generated which reduces the amount of radiated power.

3 Simulated Results and Study Technically, biomedical-based imaging and scanning cover some common visual tasks such as image detection, segmentation, classification, and enhancement [13]. So in this sequence, either separate machine learning methods or deep learning methods are traditionally used to accomplish these tasks. Furthermore, the noise level because of return loss or other means of sources should be very low in peak signal-to-noise ratio (PSNR) to improve the quality and diagnostic level of medical images. The typical range of PSNR for images with depth is found in between 30 and 50 dB, where higher is better [14]. Even though, for better impedance matching, at least −10 dB of return loss is recognized scientifically. However, the return loss value should be more negative (in dB scale) to ensure minimum reflected power from the antenna to a source. In other words, it can be also said that the quality of the image can be somewhat improved by improving the return loss of the scanning antenna. The height of the HXLPE substrate and multiple legs in the proposed antenna may create ambiguity to find the best performance in the desired range of the tomography spectrum. Therefore, in this section, return loss analysis has been described and incorporated on the dynamics of HXLPE height and multiple antenna legs effect

Single-Layer-Single-UWB Patch Antenna …

161

also. More precisely, the impact of four legs meandered structured antenna is studied and optimized to rectify the aforesaid ambiguity.

3.1 Single-Port Feeding Choice Results Since, the number of feeding ports and their location in symmetrical or asymmetrical structured antennas also play a vital role to vary the antenna performance. Therefore, at first, a traditional single-port antenna is chosen to record the parametric observations. Sonnet Lite (V.15.53) software has been used to derive all the simulated results here. As shown in Fig. 3, satisfactory performance at −10 dB impedance bandwidth of the proposed structured antenna is around 2320 MHz (1.74–4.06 GHz) at optimized HXLPE height of 17.5 mm that covers the major portion of microwave tomography band. A comparative chart in terms of UWB bandwidth and minimum return loss for target task fine-tuning bandwidth at minimum resonance frequency is tabulated in Table 1 for different HXLPE height-based antennas. It is also more important to note that reconstruction of the image depending upon structural similarity (SSIM) index and multi-scale similarity (MS-SSIM) index give more image clarity below −30 dB of return loss even in Gaussian noise (with different standard deviation) environment [15]. So, it can be easily understood from Table 1 that an antenna at an HXLPE height of 17.5 mm gives the compromising ultra-wide bandwidth and wider fining-tuning bandwidth among all. Similarly, at the same time, Fig. 4 illustrates the influence of the number of legs of a meander-structured antenna on ultra-wide bandwidth and finetuning bandwidth to ensure recommended SSIM and MS-SSIM value at a constant HXLPE height of 17.5 mm. Table 2 provides a summary of comparative antenna performance depending upon its different number of legs in meander-structure. Here, Fig. 3 Influence of HXLPE height on S11 and UWB nature

162

K. A. Khan et al.

Table 1 Antenna performance chart at different HXLPE height (n = 4, L 1 = L 4 = 21 mm, L 2 = L 3 = 32 mm) HXPLE height (mm)

−10 dB UWB frequency range (GHz)

UWB Bandwidth (MHz)

−30 dB fine-tuning frequency range (GHz)

Fine-tuning bandwidth (MHZ)

16.5

1.76–4.08

2320

3.66–3.74

80 (single band)

17.5

1.74–4.06

2320

2.56–2.64, 3.54–3.64

180 (dual band)

18.5

1.82–3.96

2140

3.56–3.60

40 (single band)

19.5

1.76–3.76

2000

3.36–3.46

100 (single band)

Fig. 4 Influence of number of legs on UWB nature of antenna Table 2 Antenna performance chart at different number of legs (HXLPE height = 17.5 mm, L 1 = L 4 = 21 mm, L 2 = L 3 = 32 mm) Number of legs (n)

−10 dB UWB frequency range (GHz)

UWB Bandwidth (MHz)

−30 dB fine-tuning frequency range (GHz)

Fine-tuning bandwidth (MHz)

4

1.74–4.06

2320

2.56–2.64, 3.54–3.64

180

5

2.28–3.92

1640

Not available

Not available

6

2.10–4.08

1980

Not available

Not available

7

2.22–3.68

1460

2.62–2.72

100

Single-Layer-Single-UWB Patch Antenna …

163

it must be explained that at n = 5 and n = 6 dual-band characteristics can be seen in the antenna, but the primary band in each situation covers the BW of 160 MHz (0.88–1.04 GHz) and 200 MHz (1.16–1.36 GHz), respectively. Furthermore, at n = 5, the primary band shows the minimum return loss of −21.54 dB at a frequency of 0.96 GHz, whereas at n = 6, the primary band occupies a minimum return loss of − 31.52 dB at a frequency of 1.26 GHz. Even though, a meander structured patch with n = 7 also gives a satisfactory performance in terms of return loss in UWB, wherein −54.94 dB return loss is found at the central resonance frequency of 2.68 GHz. Besides it, in comparison to the four legs patch, its seven legs existence increases the patch size. More concisely, a meander-structured antenna at n = 5 and n = 6 also generates a UWB, but its average return loss is above −27 dB and −14.5 dB, respectively, over the entire UWB range. Therefore, the absence of fine-tuning BW in this patch with at n = 5 or n = 6 may be the result of unclarity or ambiguity in micro- or macro- radiographic appearance in a noisy system. Hence, result analysis given in Tables 1 and 2 shows that four leg (n = 4) meandershaped antenna with HXLPE height of 17.5 mm at constant value of leg thickness, spacing, and its length yields better results and satisfactory performance. More specifically, by Fig. 5, it can be seen that the minimum return loss in the UWB is found at the frequency of 2.60 GHz and 3.58 GHz with −50.77 dB and − 32.10 dB, respectively. Subsequently, Fig. 6 reports that VSWR at this frequency is found to be 1.0058 and 1.045, respectively, which is very close to 1 and proves its ideal performance as well. Besides it, the entire UWB range from 1.74 to 4.06 GHz comes under the VSWR ≤ 2. The impact of frequency variation on the real part and imaginary part of input impedance (Z in ) is illustrated in Figs. 7 and 8, respectively. By inspecting the simulated input impedance graph, it is easy to verify that at the resonance frequency of minimum S 11 (say −50.77 dB or −32.10 dB) real part, as well as the imaginary part of the input impedance, gets the value very near to 50 ohms and zero ohms, respectively. On the other hand, phase variation of input impedance Fig. 5 Simulated value of return loss for proposed antenna

164

K. A. Khan et al.

Fig. 6 Simulated value of VSWR for proposed antenna

Fig. 7 Variation in real part of input impedance with frequency

as a function of frequency is plotted in Fig. 9 that reveals a linear phase variation over the entire UWB range of frequency (1.74–4.06 GHz). The generated surface current density (JXY) over the patch surface is directly proportional to the tangential component of the electric field. Furthermore, this surface current density can be used to evaluate the characteristics of the generated electromagnetic (EM) wave as a beam [16]. From Fig. 10, it can be understood that intense surface current density exists at the outer boundary of the patch including the feeding point location in the x–y plane. Here, the peak value of JXY can be also seen at different locations with different values apart from the feeding location. However, the maximum surface current density (JXY) at the frequency of 1.74 GHz, 2.60 GHz, and 4.06 GHz has been recorded as 2.10 A/m, 0.9 A/m, and 0.31 A/m, respectively. Additionally, JXY

Single-Layer-Single-UWB Patch Antenna …

165

Fig. 8 Simulated value of Imaginary components of Zin variation with frequency

Fig. 9 The phase variation of input impedance with frequency

Fig. 10 a The surface current distribution at 1.74 GHz at S 11 = −10 dB. b The surface current distribution at 4.06 GHz at S 11 = -10 dB. c The surface current distribution at 2.60 GHz at S 11 = -50.77 dB

166

K. A. Khan et al.

can be used to evaluate the deposition of EM energy over time into tissues of the arthritic hip joint (i.e., synovium) as a measure of SAR. Mathematically, SAR can be expressed [17] as: SAR =

J2 σ |E|2 = ρ ρσ

(1)

Here, E is the electric field in the tissue, J is the current density, σ is the conductivity of the tissue, and ρ is the density of the tissue. Therefore, it is clear by the above equation that at a constant value of tissue conductivity (assume intercellular tissue with 1.8 S/m) and its density (density of major tissue compartments lie in between 900 to 1100 kg/m3 ), maximum SAR at a frequency of 1.74 GHz, 2.60 GHz, and 4.06 GHz will be in the range of 0.002 W/Kg, 0.0005 W/Kg and 0.00005 W/Kg, respectively. Since, all SAR values are very low from the maximum SAR limit (1.6 W/Kg and 2.0 W/Kg are the SAR limit concerning 1 g and 10 g, respectively), which gives the surety of safety from RF radiation hazard in THA imaging system.

4 Conclusion The highly cross-linked polyethylene (HXLPE) based four legs meander-structured patch antenna for total hip arthroplasty (THA) that are based on ceramic-on-HXLPE or metal-on-HXLPE technology has been proposed and designed. Suggested antenna structure and its dimension cover the wide range of tomography spectrum from 1.74 to 4.06 GHz with a minimum return loss of −50.77 dB, if single-port feeding is used to excite it. Reported antenna results are well suited to reconstruct the image even in a noisy environment with the help of fine-tuning as suggested by SSIM and MS-SSIM. Generated surface current density on patch in operational frequency band results in very low SAR values that are lower by more than a thousand times to SAR limit. This patch antenna is more robust and sensitive to get the result even in microscopic perturbation found in HXLPE cup with ceramic or metal femoral head.

5 Future Work and Scope The designed patch is fabricated on HXLPE substrate, and this material has somehow similar electrical properties with other high-stable polyethylene-based polymers such as Teflon (PTFE), ultra-high molecular weight polyethylene (UHMWPE), and high-density polyethylene (HDPE). Hence, this novel antenna performance must be investigated for the THA imaging and scanning also that are based on metal-on-polyethylene (MoP) or ceramic-on-polyethylene (CoP) technology such as CoCr-PTFE, CoCr-UHMWPE, CoCr-XLPE, CoCrMo-XLPE, alumina-XLPE, alumina-PE, and many more as described in [18].

Single-Layer-Single-UWB Patch Antenna …

167

References 1. A.M. Ali, M.A. Al Ghamdi, M.M. Iqbal, et al., Next-generation UWB antennas gadgets for human health care using SAR. J. Wirel. Comun. Netw. 33, 1–20 (2021) 2. K.A. Khan, A.P. Venkataraman, Single layer tri-UWB patch antenna for osteoporosis diagnosis and measurement of vibrational resonances in biomedical engineering for future applications, in Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, ed. by A.K. Singh Pundir, A. Yadav, S. Das (Springer, Singapore, 2021), pp.193–203 3. S. Rashid, L. Jofre, A. Garrido, G. Gonzalez, Y. Ding, A. Aguasca, J. O’Callaghan, J. Romeu, 3D printed UWB microwave bodyscope for biomedical measurements. IEEE Antennas Wirel. Propag. Lett. 18(4), 626–630 (2019) 4. H. Bahrami, E. Porter, A. Santorelli, B. Gosselin, M. Popovich, L. Rusch, Flexible sixteen antenna array for microwave breast cancer detection. The Breast 5(8), 9 (2015) 5. A.K. Skrivervik, Implantable antennas: the challenge of efficiency, in 7th European Conference on Antennas and Propagation (EUCAP 2013) (2013), pp. 3627–3631 6. W. Shao, T. McCollough, Advances in Microwave near-field imaging: prototypes, systems, and applications. IEEE Microwave Mag. 21(5), 94–119 (2020) 7. J.-M. Lee, The current concepts of total hip arthroplasty. Hip Pelvis 28(4), 191–200 (2016) 8. L.N. Bravin, M.J. Dietz, Biomaterials in total joint arthroplasty, in Orthopedic Biomaterials, ed by B. Li, T. Webster (Springer, 2018), pp. 175–198 9. C.Y. Hu, T.R. Yoon, Recent updates for biomaterials used in total hip arthroplasty. Biomater. Res. 22(33), 1–12 (2018) 10. R. Inum, M.M. Rana, K.N. Shushama, M.A. Quader, EBG based microstrip patch antenna for brain tumor detection via scattering parameters in microwave imaging system. Int. J. Biomed. Imag. 2018, 1–12 (2018) 11. M.T. Islam, M.M. Islam, M. Samsuzzaman, M.R. Iqbal Faruque, N. Misran, A negative index metamaterial-inspired UWB antenna with an integration of complementary SRR and CLS unit cells for microwave imaging sensor applications. Sensors 15(5), 11601–11627 (2015) 12. M. Rokunuzzaman, M. Samsuzzaman, M.T. Islam, Unidirectional wideband 3-D antenna for human head-imaging application. IEEE Antennas Wirel. Propag. Lett. 16, 169–172 (2017) 13. S.K. Zhou et al., Deep learning for medical image analysis (Academic Press, 2017) 14. G. Zamzmi, S. Rajaraman, S. Antani, Unified representation learning for efficient medical image analysis. Inf. Med. Unlocked. 24 (2021) 15. S.T. Welstead, Fractal and Wavelet Image Compression Techniques (SPIE Optical Engineering Press Bellingham, Washington, 1999) 16. K.A. Khan, S.M. Nokerov, Optimization of multi-band characteristics in fan-stub shaped patch antenna for LTE (CBRS) and WLAN bands. Proc. Eng. Technol. Innov. 18, 25–35 (2021) 17. M.K. Hosain, A.Z. Kouzani, S.J. Tye et al., Development of a compact rectenna for wireless powering of a head-mountable deep brain stimulation device. IEEE J. Trans. Eng. Health Med. 2 (2014) 18. M. Merola, S. Affatato, Materials for hip prostheses: a review of wear and loading considerations. Materials 12(495), 1–26 (2019)

CNN-Based Optimal Image Restoration and Comparative Approaches Divya Sharma , Shilpa Sharma , and Harshal Patil

1 Introduction Zhou [1] image restoration is a member of image processing family and computer vision and performs transformation of distorted images to the high-quality clear images [2, 3]. Its main goal is to get the most optimal techniques or algorithms for removing all types of noise from the image in much less time, and it performs realtime noise removal from an image [4, 5]. As we are talking about image clearing, so it does not mean that the image restoration is same as image enhancement; image enhancement helps an image to look more general, pleasing, and increase the beauty of that image [6]. The need of image restoration is almost in every area, for example military reconnaissance, forensics, astronomy, medical, scientific laboratories, etc. [7]. There are basically three classical tasks related to image restoration; the first is the image denoising, second is image deblurring, and the last is the super resolution of the image which combines the blurring and the downsampling operation. Jin et al. [8] provided the state-of-the-art methods and approaches for high–quality images, and deep learning starts working for the task related to computer vision [9]. Using deep learning for image restoration consists of two types of methods for restoration of image, and first is the model-based method, and the second is the learning-based methods [10, 11].

D. Sharma · S. Sharma (B) Manipal University Jaipur, Jaipur, Rajasthan, India e-mail: [email protected] H. Patil Ajeenkya DY Patil University, Pune, Maharashtra, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_19

169

170

D. Sharma et al.

2 Related Work The background of the paper is discussed majorly into two broad areas or techniques applied so far for image restoration [12]. Various neural network models are applied for performing image restoration, like author in the paper titled [13] developed and introduced a model based on artifacts reduction by convolution neural network (ARCNN) specifically for CAR, i.e., compression artifacts reduction for images. With various convolution layers piled up, researcher in paper [14] introduced the concept for DnCNN to get knowledge of masking of noise from noisy image to get better and improved version of the image, and another image restoration technique is given in paper [15] which proposes an image restoration technique of MemNet which contains 212 networks, but the result and accuracy were not up to the mark. This section provides a deep insight for three variants of CNN for image restoration. Author in [3] has introduced and proposed the concept of non-blind convolution neural network for the process of image restoration with additional features channels for input channels. But this model is not an efficient choice as the image after the restoration still generates noisy in the left and right sides of the image and some high-definition portions or pixels information get lost. Table 1 gives a comparative overview for the two techniques.

2.1 Multi-level Wavelet CNN The MWCNN model is based on discrete wavelet transform (DWT) where the basic architecture consists of four filters helpful to convolute the image. In MWCNN, the downsampling function is performed to get four sub-band images, for example n1 , n2 , n3 , and n4 , and each sub-band images are computed using each filter, for example FLH ⊗ image, for n1 sub-band, etc. [16]. Then, the function of inverse wavelet transform (IWT) is performed to get the clear and accurate image reconstruction. The network architecture of MWCNN is based on the design plan where after every layer of DWT their comes a CNN block, and every CNN frame work is based on four layers of fully convolution network excluding the pooling layer [17]. The input for the CNN block after the DWT is the sub-band images. The CNN block consists of convolution function including 3*3 frames, rectified linear activation unit (ReLu) function and batch normalization (BN). Table 1 Average PSNR/SSIM on different noise levels based on BSD68 dataset Dataset

Level of noise

DnCNN (PSNR/SSIM)

Non-blind CNN(PSNR/SSIM)

BSD68

15

31.46/0.8826

31.57/0.8874

25

29.02/0.8190

29.11/0.8236

50

26.10/0.7076

26.16/0.7129

CNN-Based Optimal Image Restoration and Comparative Approaches

171

The MWCNN model is divided into two parts, i.e., a contracting subnet and another is an expanding subnet. Objective function used by the MWCNN for learning is as follows: L(Q) =

N 1  F(yi; θ )) − xi2F 2N i=1

(4)

where θ (yi; θ ) Yi Xi

Network parameter Output for network ith input images Ground truth images

The model uses “Haar” wavelet for training the architecture, for the lower frequency of the sub-band images, and we use DWT and IWT for performing up-convolution and the pooling function. The four filters with the factor of 2 I are defined as     11 −1 −1 FLL = FLH = 11 1 1     −1 1 1 −1 FHL = FHH = (5) −1 1 −1 1 After a deep analysis and testing on the gray images, the model is tested based on three different noise levels, i.e., σ = 15, 25, and 50 (Table 2). On BSD68 dataset, MWCNN outperforms for recreating the lost and noisy images to clean and pleasant output images with the following result with average PSNR and SSIM as shown in Table 3. Table 2 Image denoising based on BSD68 dataset with average result of PSNR/SSIM

Methods

MWCNN

FDCNN

σ = 15

31.86/0.8947

32.40/0.9131

σ = 25

29.41/0.8360

31.02/0.8904

σ = 50

26.53/0.7366

28.53/0.8229

Runtime

Table 3 Average PSNR/SSIM on different noise levels based on BSD68 dataset

0.40

Dataset

σ

MWCNN (PSNR/SSIM)

BSD68

15

31.86/0.8947

25

29.41/0.8360

50

26.53/0.7366

172

D. Sharma et al.

Table 4 Runtime performance based on different image sizes

Size of image

MWCNN runtime (in s)

256*256

0.0586

512*512

0.0907

1024*1024

0.3575

Table 4 indicates the runtime performance of the MWCNN model for image denoising by the noise level 50 with three different image sizes.

2.2 Flexible Deep CNN As all the deep [18] learning neural networks consider all types of artifacts equally, some time there is a need for taking specific artifacts characteristics differently. By differentiating and processing of high-frequency and low-frequency artifacts, the functioning of the image resolution task can get better and optimal performance. So, the FDCNN model is divided into four differently functioning modules, the decomposition of the image, enhancement of the image quality for every decomposed element, lateral connection, and last is the network aggregation [8]. As stated, the traditional degradation function is as D = HY + N

(6)

where D H Y N

Degraded observation Degradation function Ground truth image Noise.

From Fig. 1, we can see that the image with multiple noise is being initially decomposed which ensure the availability of the low-frequency and the high-frequency

Fig. 1 Image denoising framework using FDCNN

CNN-Based Optimal Image Restoration and Comparative Approaches

173

components is termed as texture components and is formulated separately with the module of quality enhancement network. The low- and high-frequency module is formulated as follows:   2  + λs R(I ∧ hq ) (7a) Ishq = arg min  Islq − Hs I ∧ hq s 2 s hq

Is

 2  lq hq  hq hq It = arg min It − Ht I ∧ t  + λt R(I ∧ t ) hq 2

It

(7b)

where hq

hq

Is and It hq hq Is and Is lq lq Is and Is

High-quality image. Unknown ground truth image. Observation.

Objective function for the quality enhancement network is N 2 1  ∧ hq min L(θs ) = I s − H Islq , θs )2 N i=1

(8a)

N 1  ∧ hq lq I t −H It , θt )||22 N i=1

(8b)

min L(θt ) =

where H

Mapping function for quality enhancement network.

The structure image and the texture image are parallelly enhanced, and at the end, the combination of both the images are recovered as high-quality and noise-free images. Figure 2, illustrates the working or the flow of lateral connection. The lateral feature map xˆ l has been extracted from the texture stream xˆtu.

2.3 Deep CNN with Variable Splitting Technique Task of image restoration is performed for recovering an image X from the observation being degraded Y = DX + v [7]. But if we look for a Bayesian solution, then we can get the optimal solution by using [19] maximum a posterior problem (MAP) as X ∧ = arg max x log p(y|x) + log p(x)

(9)

174

D. Sharma et al.

Fig. 2 Lateral connection

where Log p(y|x) Log p(x)

Observation for y on for log—likelihood x prior.

Equations (7a) and (7b) can be written more professionally as 1 X ∧ = arg max y−H x2 + λ(x) 2

(10)

With Eqs. (8a) and (8b), we can reduce the energy function: first is the fidelity term 21 y−H x2 , second is the regularization term as (x), and the last one is the parameter trade-off λ. For combining the denoiser prior to the Bayesian method in Eqs. (8a) and (8b) and for decoupling the fidelity term and the regularization term, we adopted the variable splitting technique. If we add an extra variable as z to Eqs. (8a) and (8b), it reforms to the problem of constrained optimization as 1 X ∧ = arg max y−H x2 + λ(z) 2

(11)

By using the half quadratic splitting technique, we can resolve the above mention problem as L μ (x, z) = where μ

Permanent parameter.

μ 1 y−H x2 + λ(z) + z−x2 2 2

(12)

CNN-Based Optimal Image Restoration and Comparative Approaches

175

To resolve the complex equation, we can formulate Eq. (10) into two subproblems individually as xk+1 = arg miny − H x2 + μx − z k 2

(13a)

z k+1 = arg minz − xk+1 2 + λ(z)

(13b)

x

z

Equation (13a) represents the fidelity term in combination of a least square quadratic regularized problem which helps to get the most optimal solution for various degradation matrices like

−1 T  H y + μzk z k+1 = H T H + μI

(14)

Equation (13b) is used for regularization term which we can reform as z k+1 = arg min z

1 xk + 1−z2 + (z) √ 2( λ/μ)2

(15)

Equation (15) modules the equation for performing the denoising √ function for the image xk+1 with the Gaussian denoiser at the level of noise as λ/µ). From Eq. (11), we can rephase Eq. (15) as z k+1 = Denoiser(z k+1 , λ/μ).

(16)

By the help of this equation, we can resolve various types of inverse problems using gray denoisers or color denoiser.

3 Analysis As deep learning is considered as most optimal for image processing techniques, we have analyzed and compared three different variants of deep learning techniques for image restoration task. After getting an insight for MWCNN, we analyzed that because of the unique biorthogonal property of DWT the actual and the source image can get restructured again as the original image using the inverse wavelet transformation for the image. With the benefit of inversibility property, the model will be safe from loss of information and can incur the complete structure of shapes and textures by the help of the degraded images. The FDCNN model works on majorly two factors, that is, high frequency and the low frequency, based on the frequency the model chooses the path to be followed. With the help of lateral connections, the FDCNN model passes the enhanced highfrequency details for individual level of the texture stream to the structure streams. The system takes help of recursive learning for shrinking the parameters count, and

176

D. Sharma et al.

Fig. 3 Comparison chart for average PSNR

residual learning helps to reduce the training parameters to optimize the training process. It acquires the two-layer cascade network with quality enhancement, and another is aggression network and secure 0.68 decibel and 0.84 decibel PSNR improvements and achieve near to 92.65% accuracy. The splitting variable method is far contrasting from the discriminative learning, as it inculcates the property of rapid discriminative learning and the optimization methods working on the flexible model. The figure represents the graphical representation of the comparative chart for the three discussed CNN variants and gives an analysis that FDCNN is the most optimal choice for the image restoration tasks, and also it saves more GPU consumption of 93.56% (Fig. 3).

4 Conclusion The paper presents a CNN-based image restoration technique comparison which is shown below: 1.

2. 3.

Lagendijk and Biemond [3] the MWCNN-based model works on two networks; the first is the contracting subnet which consist of various DWT, and various IWT in the expanding subnetworks, and both have CNN blocks in common, and perform state-of-the-art performance showcasing the subsampling of image with the information being safe using the invertible property. The FDCNN model works on the basis of the image frequency by decomposing the artifacts by selecting based on high frequency and low frequency. The model-based optimization is clubbed with the learning denoiser prior for image restoration task.

The paper concludes that the FDCNN-based model is the most optimal choice for image restoration task.

CNN-Based Optimal Image Restoration and Comparative Approaches

177

References 1. G.X. Zhou, Restoration for motion blurred images of moving objects, in Advanced Engineering Forum 2012, vol. 6. (Trans Tech Publications Ltd.), pp. 1108–1111 2. M. Maru, M.C. Parikh, Image restoration techniques: a survey. Int. J. Comput. Appl. 160(6), 15–19 (2017) 3. R.L. Lagendijk, J. Biemond, Basic methods for image restoration and identification, in The essential guide to image processing 2009 Jan 1 (Academic Press), pp. 323–348 4. https://en.wikipedia.org/wiki/Image_restoration 5. C.R. Steffens, L.R. Messias, P.J. Drews-Jr, S.S. Botelho, CNN based image restoration. J. Intell. Rob. Syst. 11, 1–9 (2020) 6. O.H. Mohammed, B.S. Mahmood, Advance in image and audio restoration and their assessments: a review 7. K. Zhang, W. Zuo, S. Gu, L. Zhang, Learning deep CNN denoiser prior for image restoration, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017), pp. 3929–3938 8. Z. Jin, M.Z. Iqbal, D. Bobkov, W. Zou, X. Li, E. Steinbach, A flexible deep CNN framework for image restoration. IEEE Trans. Multimedia 22(4), 1055–1068 (2019) 9. Y. Zhang, Y. Tian, Y. Kong, B. Zhong, Y. Fu, Residual dense network for image restoration. IEEE Trans. Pattern Anal. Mach. Intell. 43(7), 2480–2495 (2020) 10. A.K. Boyat, B.K. Joshi, A review paper: noise models in digital image processing. arXiv preprint arXiv:1505.03489. 2015 May 13 11. M.D. Sontakke, M.S. Kulkarni, Different types of noises in images and noise removing technique. Int. J. Adv. Technol. Eng. Sci. 3(1), 102–115 (2015) 12. I. Bashir, A. Majeed, O. Khursheed, Image restoration and the various restoration techniques used in the field of digital image processing. Int. J. Comput. Sci. Mob. Comput. 6(6), 390–393 (2017) 13. C. Dong, Y. Deng, C.C. Loy, X. Tang, Compression artifacts reduction by a deep convolutional network, in Proceedings of the IEEE International Conference on Computer Vision (2015), pp. 576–584 14. K. Zhang, W. Zuo, Y. Chen, D. Meng, L. Zhang, Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Trans. Image Process. 26(7), 3142–3155 (2017) 15. Y. Tai, J. Yang, X. Liu, C. Xu, Memnet: a persistent memory network for image restoration, in Proceedings of the IEEE International Conference on Computer Vision (2017), pp. 4539–4547 16. P. Liu, H. Zhang, K. Zhang, L. Lin, W. Zuo, Multi-level wavelet-CNN for image restoration, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (2018), pp. 773–782 17. K. Uchida, M. Tanaka, M. Okutomi, Non-blind image restoration based on convolutional neural network, in 2018 IEEE 7th Global Conference on Consumer Electronics (GCCE) (IEEE, 2018), pp. 40–44 18. http://www.surendranathcollege.org/new/upload/PUJA_MUKHERJEEImage%20Restoratio n2020-05-04PUJA_MUKHERJEE_DIP_04.05.2020_SM7.pdf 19. https://www.probabilitycourse.com/chapter9/9_1_2_MAP_estimation.php

Trust Offloading in Vehicular Cloud Networks Sarbjit Kaur and Ramesh Kait

1 Introduction Contemporary autonomous methods have resulted in massive improvement in human welfare. IEEE802.15 × standards has resulted in wide deployment of wireless technologies in safety and controlling applications [1, 2]. One such deployment of wireless applications is vehicular ad hoc network (VANET). In VENET, vehicles are equipped with wireless sensors and short-range communication capabilities, enabling them to share vast volumes of data. One prominent application of VANET is intelligent transport system. But the security of messages transferred between the vehicles [3] become an urgent concern due to the defective sensors, software viruses, and bogus messages transferred from adversary nodes or other types of attacks like denial of service etc. Trustworthiness of vehicles in the network is essential for deploying the intelligent transportation system. It is important to develop an algorithm for assessing trust by taking into consideration all of the vehicle’s necessary attributes and their relationships in a more efficient manner. Just before designing cloud-based trust models, there are a variety of challenges to remember like: how to increase trust convergence rate; how to identify the node whether it is adversary or trustworthy; how to secure the communication between the nodes and cloud; which parameters should be considered as stats to calculate the trust; how to create dynamic trust; and how to store the stats on cloud. The remaining part of the paper is organized as follows: Sect. 2 includes an overview of existing trust models in various fields. The process of TOVCN model and algorithms are discussed in Sect. 3. By stating simulation findings in Sect. 4, we endorse the advantages of using our protocol. Finally, in Sect. 5, we outline the study’s key contribution and include recommendations for future innovations.

S. Kaur (B) · R. Kait Kurukshetra University, Kurukshetra, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_20

179

180

S. Kaur and R. Kait

2 Related Work Scopus, Google Scholar, Mendeley, Taylor and Francis, Association for Computing Machinery (ACM DL), Web of Sciences, and IEEE Xplore were used to search for high-quality literature on cloud-based trust management for VANETs. To discover any relevant literature, other databases such as MDPI journal database, Hindawi, and Wiley were also used. Mendeley was used to gather and screen research publications [4]. The concept of trust for imposing security has attracted the interest of researchers in various fields. The network where this research has been applied are social network [5], wireless sensor network [6], cloud computing [7], file sharing network [8, 9], crowd sensing network [10, 11], and vehicular cloud network [12, 13]. The goal of the trust calculation is to determine whether the message was created by a trustworthy node and message’s integrity is intact. The trust computation can be centralized [12, 13] and decentralized [14]. In decentralize trust computation, every node computes the trust for another node on their own. The trust computation by each node for other nodes making the vehicle overburden, and trust cannot be reliable as adversary node intentionally reported wrong about the non-adversary to take the more opportunities for services. Therefore, a trustworthy third party is required to store the quality of service reports for vehicles. The cloud’s calculation of trust based on stored data is more accurate and impartial [12, 13]. In a vehicular ad hoc network, the principle of offloading [15] has its own importance. The trust computation by the node itself may increase the rate of false reporting by the adversary nodes. The proposed model support the trust offloading, where the stats are obtained by the concerned nodes for each communication and sent to the cloud for trust computation.

3 Methodology 3.1 TOVCN Trust Model Process The key contribution of TOVCN trust model is to give the fair chances to trusted nodes to provide the services while gradually reducing the trust of adversary nodes. Vehicles or nodes initially registered itself to the cloud by sending their own parameters not limited to position, vehicle number, size, model, fuel consumption, and performance to the cloud and cloud set the default initial trust for these newly registered vehicles. This helps to remove the cold start problem in case of no interaction history available of any newly registered node. We used hybrid trust mechanism to compute the trust. All the trust computation is done on the cloud. The idea of trust computation offloading increases the success ratio as it minimizes the chances to compute the fake trust value by any adversary node at their own level. The two phases before and after interaction have been characterized

Trust Offloading in Vehicular Cloud Networks

181

Fig. 1 TOVCN trust model process

in Fig. 1. During the pre-interaction phase, the requesting vehicle requests the trust for nodes in its vicinity. The cloud evaluates the trust based on available data and sends the trust to the requesting vehicle for authentication. And if the vehicle found trustworthy enough, then it includes to make for communication and path selection. In after interaction phase, each node collects the QOS parameters of another vehicle on the basis of its provided services. And this stats further uploaded on the cloud to compute the centralize trust.

3.2 Detailed Trust Model The trust computation method is found from combination of direct observations with the recommendation from the node’s neighbors for establishing trust. As in the entity centric trust models, node’s trust value is derived from the weights assigned to statistics during interactions which are populated using data reported or forwarded by the node and the recommendations by the neighboring nodes. In general, a wellbehaving node’s statistics always be associated with a greater trust value than a medium or low-level trusted node. As a result, the node’s trust  is categorized into three levels: (Node) > 0.75 as high, (Node) < = 0.75, and (Node) > 0.50 as mid, and (Node) < 0.5 as the low trust. In view of the illustration and reading, the few symbolization used in the paper are enumerated in Table 1, which are required to identify defective nodes and develop the proper system to estimate the trustworthiness of nodes. The main objective of this model is to provide the communication between requesting node and another node through a trusted path. Requesting node is one who want to communicate with some another node in the network. Trusted path

182

S. Kaur and R. Kait

Table 1 Variables description Heading level Example (A)

Nodes combined trust

N AB M AB

The total number of packets that A node asked to B node to route in given time t

ω(t)

The weight of trust vectors of a node at time t,

D AB

The direct trust value of A node to B node

R BN  AB

Recommendation of A node to B node

The number of successful packets that B node has routed correctly for A node in given time t

Comprehensive trust value of A node to B node

must consists of all trustworthy nodes. There can be a number of paths exist P = {r 1 , r 2 , r 3 … r n } between the requesting node and target node. The pseudocode for path selection is given in Algorithm 1.

Algorithm 1 specifies that the current node chooses the relay node among the candidate nodes N in its range on the basis of distance. For each node c among the N nodes, distance is calculated and compared with preset value of distance d.

Trust Offloading in Vehicular Cloud Networks

183

The trusted path consists all nodes having trust value greater than the predefined threshold. Algorithm 2 specifies the trust (P) of path P from the requesting node Q. Trust q of the requesting node and trust p of any node n from path P is retrieved from cloud where the trust of every node is updated after its interaction with other nodes in the network. If node n’s trust is greater than a predefined threshold, the node’s trust will be taken into consideration and applied to the path trust; otherwise, the trust will be set to nil. Algorithm 3: Node Trust Computation. FAB ← compute malevolent tendency of the node B using N AB , M AB : 

FAB

N AB − M AB  B = NA

 (1)

D AB ← compute direct trust of node using FAB and ω(t), weight factor: T 

D AB =

t

ω(t) B n × FA,t 1 ω(t)

(2)

N ← neighbors of the node B. R BN ← get indirect recommendation of node B N R BN

=

 AB ← compute combined trust

i × DiB N i=1 i

i=1

(3)

184

S. Kaur and R. Kait

 AB ← τ.D AB + (1 − τ ).R BN

(4)

Algorithm 3 define the steps to compute the final combined trust. During interactions, activity from the A to B is monitored, where N AB is the overall number of packets requested by node A to node B, M AB is number of packets successfully routed for node A by node B, ω(t) is weight factor assigned to the interaction happening at time t, and n is the total number of prior interactions. Computation of direct trust calculated in Eq. (2) need adversary tendency given in Eq. (1). After that, using direct trust and recommendations, combined trust is calculated.

4 Results and Discussion This section explains the TOVCN simulation framework, which allows for traffic flow simulation with vehicles serving as vehicular nodes. Each vehicle represents a node that implements the proposed trust mechanism in the simulation. To determine the shortest path to a destination, the implementation employs the connections hierarchies search algorithm. MATLAB is used to implement the trust model and evaluate the direct trust, indirect trust, hybrid trust, DBTECH-1, DBTECH-2, and proposed TOVCN trust mechanisms. The proposed protocol run on Windows 10 operating system. Appropriate test environment plays a key role to represent the performance in better way. Nodes are moving on the basis of mobility provided by SUMO [16] over connections hierarchies’ algorithm. Trust-based calls are developed as API calls requested by nodes and MySQL is used as a backend to store TOVCN trust vectors and consequent updates (Table 2). Proposed TOVCN trust model, DBTEC-1, DBTEC -2 [12], and model with no trust are compared on the basis of false reporting threat. In this threat, adversary nodes or vehicles send a false report about the quality of service provided by other trusted vehicles in the network, reducing their chances of delivering services. Success ratio, Table 2 List of parameters

Parameters

Values

Monitoring area

1000 × 1000 m

Number of nodes

10–200 Nodes

Communication range

250–381 m

Packet interval

2–4 ms

Length of data packet

1024 bits

No adversary nodes

1–20%

Cloud trust storage engine

MySQL

Trust range

[0,1]

Traffic model

SUMO

Simulation time

200 (s)

Trust Offloading in Vehicular Cloud Networks

185

average trust of adversary nodes in the network, and average trust of trusted node are the three indices calculated. Figure 2 shows how the overall success ratio increases with time for TOVCN trust model, DBTEC-1, DBTEC-2, and trustless model. In a given time span, the success ratio is specified as the proportion of interaction with trusted vehicles to total number of interactions. The success ratio from the beginning to the current timestamp is the final success ratio. In comparison to other schemes, we found that the proposed TOVCN trust model performs best. Trust value of adversary nodes should be decreased as their interactions with other nodes increases. Average trust of adversary nodes for traditional scheme, DBTEC-1, DBTEC-2, and proposed TOVCN trust model is depicted in Fig. 3. Here, we have analyzed that the average trust of adversary nodes for proposed scheme is better than all other schemes. After analysis, the proposed scheme performs 20.06% better as compared to DBTECH-2, 31.84% from DBTECH-1 and 61.30% from without trust scheme. The average trust of adversary nodes decreased rapidly with the time. Average trust of trusted nodes should be increased with the number of interactions. And trusted nodes should get the more chances to provide the services with the time. The average trust of trusted nodes of traditional scheme, DBTEC-1, DBTEC-2, and proposed TOVCN trust is given in Fig. 4. We analyzed that the average trust of trusted nodes for proposed TOVCN model grows rapidly with the time. It means proposed TOVCN model restricts the false reporting of adversary nodes, and the trusted nodes get the fair chances of providing the services in the network. The proposed TOVCN model performs 5.67% better as compare to DBTECH-2, 12.33% better from DBTECH-1, and 23.55% from scheme without trust mechanism.

Fig. 2 Success rate

186

S. Kaur and R. Kait

Fig. 3 Average trust of adversary nodes

Fig. 4 Average trust of trusted node

5 Conclusion and Future Scope The trust computation offloading concept provides better reliability and security as trust does not depend on the node itself. As a result, the chances of a fake trust calculation by an individual node are reduced. Before sending the collected data to the cloud, each node’s authenticity is also verified. Our findings show that different aspects of the network’s performance have improved, including the success ratio and

Trust Offloading in Vehicular Cloud Networks

187

the trust dynamics of trusted and adversary nodes. For example, when compared to DBTECH schemes the proposed work showed 56.9% improvement in success ratio. For future work, we can utilize our substantial simulation test bed to perceive the relative value for various trust modeling conditions. We also intend to investigate the levels of trust needed by various traffic scenarios, which could lead to more comprehensive user modeling to aid route planning.

References 1. L. Atzori, A. Iera, G. Morabito, The Internet of Things : a survey. Comput. Netw. 54(15), 2787–2805 (2010). https://doi.org/10.1016/j.comnet.2010.05.010 2. S. Li, L. Da Xu, S. Zhao, The internet of things: a survey. Inf. Syst. Front. 17(2), 243–259 (2015). https://doi.org/10.1007/s10796-014-9492-7 3. F. Cunha et al., Data communication in VANETs : survey, applications and challenges. Diss. INRIA Saclay; INRIA (2014). https://doi.org/10.1016/j.adhoc.2016.02.017 4. K.M. Pradhan SS, Reference management tools in academic research : a comparative analysis of Mendely, Zotero, Ref Work and End Note. Int. Res. J. Sci. Eng. A7, 724–729 (2020) 5. Q. Yang, H. Wang, Toward trustworthy vehicular social networks. IEEE Commun. Mag. 53(8), 42–47 (2015). https://doi.org/10.1109/MCOM.2015.7180506 6. A. Liu, X. Liu, J. Long, A trust-based adaptive probability marking and storage traceback scheme for WSNs. Sensors (Switzerland) 16(4), 451 (2016). https://doi.org/10.3390/ s16040451 7. M. Sinha, S. Silakari, R. Pandey, Trust based mechanism for secure cloud computing environment: a survey. Int. J. Eng. Sci. Invent. 5(3), 17–23 (2016) 8. X. Fan, M. Li, J. Ma, Y. Ren, H. Zhao, Z. Su, Behavior-based reputation management in P2P file-sharing networks ✩. J. Comput. Syst. Sci. 78(6), 1737–1750 (2012). https://doi.org/10. 1016/j.jcss.2011.10.021 9. S.D. Kamvar, M.T. Schlosser, H. Garcia-Molina, The Eigentrust algorithm for reputation management in P2P networks, in Proceedings of the 12th international conference on World Wide Web (2003), pp. 640–651. https://doi.org/10.1145/775240.775242 10. W. Alasmary, S. Valaee, Crowd sensing in vehicular networks using uncertain mobility information. IEEE Trans. Veh. Technol. 68(11), 11227–11238 (2019). https://doi.org/10.1109/TVT. 2019.2939145 11. A. Bazzi, A. Zanella, Position based routing in crowd sensing vehicular networks. Ad Hoc Netw. 36, 409–424 (2016). https://doi.org/10.1016/j.adhoc.2015.06.005 12. Z. Tang, A. Liu, Z. Li, Y.J. Choi, H. Sekiya, J. Li, A trust-based model for security cooperating in vehicular cloud computing. Mob. Inf. Syst. 2016 (2016). https://doi.org/10.1155/2016/908 3608 13. C. Huang, R. Lu, H. Zhu, H. Hu, X. Lin, PTVC: achieving privacy-preserving trust-based verifiable vehicular cloud computing, in 2016 IEEE Glob. Commun. Conf. GLOBECOM 2016— Proc. (2016), pp. 1–6. https://doi.org/10.1109/GLOCOM.2016.7842180 14. W. Li, H. Song, ART: an attack-resistant trust management scheme for securing vehicular Ad Hoc networks. IEEE Trans. Intell. Transp. Syst. 17(4), 960–969 (2016). https://doi.org/10. 1109/TITS.2015.2494017 15. H. Zhou, H. Wang, X. Chen, X. Li, S. Xu, Data offloading techniques through vehicular Ad Hoc networks: a survey. IEEE Access 6, 65250–65259 (2018). https://doi.org/10.1109/ACC ESS.2018.2878552 16. S. Kaur, R. Kait, Comparative analysis of routing algorithms in SUMO for VANET, in International conference on sustainable Computing in Science, Technology and Management (SUSCOM-2019) (2019), pp. 2022–2027. https://doi.org/10.2139/ssrn.3358133

Deep Learning-Driven Structured Energy Efficient Affordable Ecosystem for Computational Learning Theory Krishan Gopal Gupta, Samrit Kumar Maity, Abhishek Das, and Sanjay Wandhekar

1 Introduction Deep learning is a techniques by which machines learns the hierarchical characteristic along with nature of representation of data by using machine learning models. Deep neural network is one such model which applies artificial neural network philosophy to learn about data and its representation. Recently, by popular use, deep learning and learning through deep neural network are considered to be same technological terms. Most of the core neural network concepts were proposed and implemented in 80s and 90s. But, there was no major breakthrough happened during that period. This is because, non-availability of big data and lack of adequate compute power. Eventually, the interest and excitement about the study of the domain were diminished. Eventually, it was looked as a subject of theoretical study. However, the complete scenario changed during last five year. Today, machine learning and deep learning are considered to be most sought after technical field for scientific study. Domain experts are using deep learning techniques to solve some of the most challenging problems at their hand. This became possible because of two major scientific events. One: the big data phenomena—because of IoT and Hadoop-big data development, it has become possible now to collect, refine, study large amount of data of someone’s interest. Second: availability of cheap computing infrastructure and hardware—with general purpose graphics computing units, computation with powerful accelerators in K. G. Gupta (B) · S. K. Maity · A. Das · S. Wandhekar Centre for Development of Advanced Computing, Pune, India e-mail: [email protected] S. K. Maity e-mail: [email protected] A. Das e-mail: [email protected] S. Wandhekar e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_21

189

190

K. G. Gupta et al.

an acceptable time bound nature has become a reality for the first time in history. Both of these developments have encouraged researchers to cogitate and experiment with large amount of data with neural network models on high performance acceleratorbased computing hardware. Historically, training a neural network model would take days and month of computing cycle. But today, training of network models is possible even on smaller computing infrastructure like workstation or even on laptops. Performing training and inferencing on a desktop level computer is possible however not always preferred because of time factor. With huge scale of data processing and neural network model training, one may have to wait for hours to days to reach to a certain inference accuracy level. With deep learning models, training process becomes more grueling and time-consuming with the fact that multiple layers of connected neurons and associated weight values has to be adjusted. In brief, deep neural network model training process can be divided into four major stages (1) Preprocessing of input data. (2) Training predesigned deep learning model with input data. (3) Preservation of trained deep learning model for further improvement and (4) Deployment of the pre-trained model. Training of deep learning model is done through forward propagation or backward propagation method. Both of these operations are essentially matrix multiplications, hence very much compute intensive. The core of deep learning computation is matrix manipulation. To reduce neural network training time, one has to optimize matrix manipulation operation. With high performance GPGPU hardware, matrix multiplication operation is done in parallel. Deep learning performance library optimized and accelerated on various parallel hardware plays an important role to speed up training process of the neural network models. One such library is Nvidia’s deep neural network library (cuDNN) [1]. It is a collection of GPU-based accelerated programming functions for deep neural networks. It has fine-tuned implementations for standard functions for layers such as forward and backward convolution, pooling, normalization, and activation. cuDNN shows 2.5X faster training of CNNs. Developing deep learning applications is a bigger challenge for data scientists and engineers. Instead of developing applications right from the scratch, they prefer to use some of the existing frameworks. Frameworks enable quick and easy application development environment by providing higher level programming primitives along with lower level core computational kernels. Plethora of such standard, popular frameworks exist with open-source tenet like Caffe [2], TensorFlow [3], Theano [4], CNTK [5], Keras [6], MXNet [7], and Torch [8]. These frameworks make it easier to build complex deep learning solutions. Every framework is different and is built for a different purpose. They are targeted for different application domain too. Developing deep learning application requires high level of software and domain expertise. Naturally, the investment of time and human resource is large. Mastering a deep learning framework and apply them for specific kind of domain application development require continuous effort and time. Some of them have steep learning curve. However, researchers can take help of predesigned tools, to accomplish same result in a very easy user-friendly manner. Nvidia DIGITS [9] is one of such tool which provides interactive and Web-based services for rapid training of deep neural

Deep Learning-Driven Structured Energy Efficient …

191

network model. Along with training, other facilities provided by DIGITS [9] are analyzing data, classification of images, image segmentation, and object detection. Deep learning preprocessing and training of neural network model are a compute intensive process. It is a challenging domain requires massive compute power for application processing. Therefore, need of computer solution with a substantially high-end compute capacity with low monetary investment is felt. Such solution should be “ready-to-use” computer module and easy to maintain having lessenergy footprint. From software and framework perspective, such solution should be equipped with best deep learning performance library, with popular frame work and interactive tools. Here, we present “PARAM SHAVAK DL GPU system”—a high end, affordable and personalized deep learning development solution in a box. It is ready-to-use deep learning system with a capacity of solving multi-disciplinary grand challenges in science and engineering that employ deep learning techniques. Rest of this paper is organized as follows: Section 2 describes related work. Section 3 describes overview of PARAM Shavak dl GPU system. Sections 4 and 5 describe conclusion and future direction. Finally, acknowledgment and references.

2 Related Work There have been several initiatives to build deep learning compute solution for professionals to meet industry requirement. These initiatives range from building a cloud instance for neural net training for development of specialized workstations. The need for low-cost and accessible high-end system has been in demand from academic institutions and research organizations. Though some organizations has built specialized compute facility for deep learning research, but, optimal utilization of such big facilities is always remain challenging as most of these organization lack skilled manpower to manage and maintain high-end dedicated cluster infrastructure. Power consumption by hardware component is another challenge they find difficult to deal with. On top of it, lack of technical training makes the end users life tough. Added up, all these factors deter end users to exploit high-end big cluster infrastructure for high end compute-intensive application execution. To address these issues, few companies have come up with system with sophisticated design with low-power footprint and high compute capacity. Lamda Labs [10] offer such solution from GPU laptop [11] to cloud [12] with Nvidia GPU accelerator card for DL. Nvidia DGX-1 [13] is a supercomputer-in-a-box system by Nvidia Corp. The DGX-1 with Tesla V100 has massive compute power like 1 Petaflop single precision compute power with 40,960 CUDA compute core and 5120 tensor cores. Solutions of these kinds also come with popular deep learning framework preinstalled. Math supercomputer-in-a-box [14] from Wolfram Research used Mathematica software to provide supercomputing solution.

192

K. G. Gupta et al.

3 Overview of PARAM Shavak DL GPU System All the system listed above is meant for high-end deep learning training task. They are available with preinstalled software ecosystem. However, installing these solutions requires huge monetary investment. PARAM Shavak DL GPU system is a similar offering from Center for Development of Advanced Computing(C-DAC) [15] which requires less monetary support and offers similar computation platform with a compromise in maximum single-precision computer power. PARAM Shavak DL GPU system is affordable, easily maintained, table-top version, deep learning “supercomputing-in-a-box” solution. The compact nature of its design with one or two NVIDIA GPUS accelerators makes it perfect platform for deep learning research. This PARAM Shavak DL solution design to provide high-end computing resource with advance deep learning (DL) framework to train DL workload for the scientific labs and engineering or academic institutes. PARAM Shavak DL GPU system is equipped with latest x86_64-based Intel Xeon server series processor along with one or two Nvidia Quardo P5000 accelerator card. It provides maximum 25 TF of single-precision compute performance. It has 96 GB of main memory and 16 TB of secondary storage. As of now, the system runs on Ubuntu OS 18.04 operating system. Figure 1 shows PARAM Shavak DL GPU system architecture. PARAM Shavak DL GPU is pre-loaded with all major popular deep learning frameworks like Caffe [2], Torch [8], TensorFlow [3], Theano [4], Keras [6], MAXNet [7], and Paddle [16]. This system has open-source Nvidia DIGITS [9] tool installed. This allows users to port and run deep learning task very easily through its user-friendly interface. It is an interactive, Web based, deep learning training tool, used for rapid training of deep neural network model, analyze data, image classification, segmentation and finally object detection task in very user-friendly, interactive manner. DIGITS server runs on PARAM SHAVAK DL GPU system, and user can access to use it remotely. It also supports multi-user mode—multiple user can use same tool and do the analysis.

Fig. 1 PARAM Shavak DL GPU system architecture

Deep Learning-Driven Structured Energy Efficient …

193

In Fig. 2, we show training graph of object detection [17] run using DIGITS on PARAM Shavak DL GPU system. We use data from the object detection track of the KITTI Vision Benchmark Suite [18] and DetectNet model. This dataset consists 6373 number of training images, 1108 validation images (total data size is 5.32 GB). This data are stored in IMDB format. Complete training took 4 h and 42 min for 30 epoch using Nvcaffe, and it uses both CPU and GPU. In Fig. 3, we show ConvNet benchmark [19] result for Caffe and Torch. ConvNet benchmark [19] runs popular ImageNet winner models like AlexNet [20], GoogleNet [21], and VGG [22], and it clock the time for a full forward + backward pass. In Fig. 4, we show tf cnn benchmark [23] result for TensorFlow. tf cnn benchmark [23] contains TensorFlow 1 implementations of several popular convolutional models and is designed to be as fast as possible. We run popular ImageNet winner models like ResNet50 [24], ResNet101 [24], and InceptionV3 and measure the throughput. In Fig. 5, we show training time for You Only Look Once (YOLO), real-time object detection system on COCO dataset for 100 iterations. We have compile YOLO for GPU and CPU with OpenMP.

Fig. 2 Training graph of object detection

194

K. G. Gupta et al.

Convnet Benchmark Result 600

Time(ms)

500 400 300

Caffe

200

Torch-7

100 0 Alexnet

Overfeat

GoogleNet

VGG

[fast] Fig. 3 ConvNet benchmark result

TF CNN Benchmark

Images\Sec

200 150 100 50 0

Resnet50

Resnet101

InceptionV3

Fig. 4 tf cnn benchmark result

Time ( Minute)

500

Training Time for YOLOV2 440

400 300

CPU

200

GPU

100

6 0

YOLOV2 Fig. 5 Training time for YOLOV2

Deep Learning-Driven Structured Energy Efficient …

195

4 Conclusion Need of an affordable, powerful deep learning compute platform is undeniable. While choosing, researchers need to decide up to what level someone compute power expectation is. Among the available deep learning compute solutions, PARAM Shavak DL GPU system is one which makes balance between all the compulsory, important factors like requirement of high compute power, maintainability, preconfigured with software framework, low power uses, and low-cost generic solution.

5 Future Work Currently, PARAM Shavak DL GPU system interacts with available GPU accelerators through standard PCI-e 16 × interface. According to major studies, this is one of the major bottle necks toward large size, repetitive data exchange. This issue becomes major issue during GPU device-to-device interaction. To solve this and to accelerate intra-GPU interaction, Nvidia has come up with proprietary NVLink interface [25]. With NVLink interface, GPU-to-GPU interaction speeds up to 3x. In future, we intends to explore how such high-speed interconnect can be integrated and supported with PARAM Shavak DL GPU platform in order to enable better speedup and scalability of deep learning task. On CPU front, Intel is providing oneAPI deep neural network library [26] to improve productivity and enhance the performance of deep learning frameworks. We plan to port and provide oneAPI DNN [26] on PARAM Shavak DL GPU system to enable better exploitation of CPU architectures. Acknowledgements The research work is supported by C-DAC, Pune. We like to thank members of HPC Technologies Group, C-DAC Pune.

References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11.

CUDNN Homepage, https://developer.nvidia.com/cudnn. Last accessed 2021/08/17 Caffe Homepage, http://caffe.berkeleyvision.org/. Last accessed 2021/08/25 TensorFlow Homepage, https://www.tensorflow.org/. Last accessed 2021/08/25 Theano Github page, https://github.com/Theano/Theano. Last accessed 2021/08/25 CNTK Github page, https://github.com/Microsoft/CNTK. Last accessed 2021/08/25 Keras Homepage, https://keras.io/. Last accessed 2021/08/25 Mxnet Homepage, https://mxnet.incubator.apache.org/. Last accessed 2021/08/25 Torch Homepage, http://torch.ch/. Last accessed 2021/08/25 DIGITS Homepage, https://developer.nvidia.com/digits. Last accessed 2021/09/01 Lambda Labs Homepage, https://lambdal.com/. Last accessed 2021/09/02 LambdaTensorBook Homepage, https://lambdalabs.com/deep-learning/laptops/tensorbook. Last accessed 2021/09/02 12. Lambda Vector Homepage, https://lambdalabs.com/gpu-workstations/vector. Last accessed 2021/09/02

196

K. G. Gupta et al.

13. Nvidia DGX-1 Homepage, https://www.nvidia.com/en-us/data-center/dgx-1/. Last accessed 2021/09/03 14. Math Supercomputer Pdf, https://www.wolfram.com/products/applications/sem/semproductfl yer.pdf. Last accessed 2021/09/03 15. C-DAC Homepage, www.cdac.in. Last accessed 2021/09/03 16. Torch Homepage, https://github.com/PaddlePaddle/Paddle. Last accessed 2021/08/26 17. Object Detection Github page, https://github.com/NVIDIA/DIGITS/tree/master/examples/obj ect-detection. Last accessed 2021/09/01 18. Kitti Homepage, http://www.cvlibs.net/datasets/kitti/eval_object.php. Last accessed 2021/09/01 19. Convnet Github page, https://github.com/soumith/convnet-benchmarks. Last accessed 2021/09/05 20. AlexNet wiki Page, https://en.wikipedia.org/wiki/AlexNet. Last accessed 2021/08/25 21. C. Szegedy, W. Liu, Going deeper with convolutions, in IEEE Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, June 2015, pp. 1–9 22. K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition, in International Conference on Learning Representations Conference 2015 23. Tf cnn Benchmark Github page, https://github.com/tensorflow/benchmarks. Last accessed 2021/09/05 24. K. He, X. Zhang, X. Ren, S. Ren, Deep residual learning for image recognition, in Computer Vision and Pattern Recognition, arXiv:1512.03385v1 [cs.CV], 10 Dec 2015 25. NVLink wiki page, https://en.wikipedia.org/wiki/NVLink. Last accessed 2021/09/05 26. OneDnn Homepage, https://software.intel.com/content/www/us/en/develop/tools/oneapi/com ponents/onednn.html. Last accessed 2021/09/05

Design of the MIMO Antenna Using Metamaterial for S-Band Applications, with Reduced Mutual Coupling and Improved Diversity Gain B. Ramamohan and M. Siva Ganga Prasad

1 Introduction Generally, multi-input and multi-output (MIMO) courses of action are comprehensively guaranteed in the remote correspondence for improving the framework dependability. MIMO has become a basic component of remote compatibility principles which includes Wi-Fi and evolution for the long term (LTE) [1–17]. The channel that we are now using is wireless channel, and it is unfriendly. It decreases power loss as distance increases, eliminates co-channel interference, and improves restricted bandwidth. As a consequence, there is always a need for better wireless network output, such as large improvements in spectral efficiency and data speeds, high quality of service (QoSs), wide coverage, and so on. While designing a MIMO antennas following difficulties are confronting, integration of several antenna elements into one antenna system, maintain the compactness of the system, reduces the mutual coupling (MC) of the MIMO antenna elements fulfill the requirements of diversity parameters. To overcome the mutual coupling problem, we need to increase the separation between antenna elements or addition of decoupling network like defected ground structure or ground stub or split ring resonator (SRR) or neutralization line (NL). In this paper, the designed antenna is made up of two rectangular patch MIMO antenna in between T-shaped metamaterial is inserted. This designed antenna works

B. Ramamohan (B) Department of Electronics and Communication Engineering, KL Deemed To Be University, Andhra Pradesh, India e-mail: [email protected] M. Siva Ganga Prasad Department of Electronics and Computer Science, KL Deemed To Be University, Andhra Pradesh, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_22

197

198

B. Ramamohan and M. Siva Ganga Prasad

Fig. 1 MIMO antenna construction a without metamaterial, b with single T-shaped metamaterial, c with double T-shaped metamaterial, and d with triple T-shaped metamaterial

Table 1 Dimensions of the parameters in Fig. 2

Parameters

Dimensions (mm)

Parameters

Dimensions (mm)

P1

38

M1

26

P2

29.5

M2

11.5

P3

16.8

M3

13

P4

2.9

M4

3

P5

0.7

M5

2

P6

23.8

in the operating frequency of 2.4 GHz by inserting T-shaped metamaterial; the envelope correlation coefficient (ECC) is reducing, and diversity gain (DG) is increased. The gain of [18] is 9.6, and it is improved equal to 9.9 in our proposed antenna. Excellent stable radiation pattern results are obtained at the antenna operating frequency. The antenna is designed, simulated, and fabricated using computer simulation technology (CST) software; fabricated results are obtained in the network analyzer. The structure of the antenna with good dimensions is explained in Sect. 2. The front view of MIMO antenna with metamaterial step by step and without metamaterial is illustrated in Fig. 1. The important factor for the antenna design is the shape of the metamaterial. In [19], the shape is dumbbell. Out antenna, it is a T shape. The separation between two antennas is 0.308λo (38 mm). The decrease of envelope correlation coefficient (ECC) between the two ports is gotten by executing three Tstructure metamaterials associated in arrangement. Table 1 shows the measurements of the novel MIMO antenna.

2 Antenna Construction and Plan The novel proposed antenna form is depicted in Fig. 1 as a front view. To build the MIMO antenna, two rectangular copper patch antennas are installed on a FR-4 lowcost substrate. The layer has a thickness of 0.0036 mm. The antenna’s measurements are 38 × 29.5 × 0.036. Table 1 describes the size of the antenna parameters. A

Design of the MIMO Antenna Using Metamaterial …

199

Fig. 2 Parameters for MIMO antenna architecture

gap of 0.308 λ separates the two-patch antennas (38 mm). The proposed antennas design process which involves four stages is presented in Fig. 1. First, an effective two-patch antennas with feed are built and optimized which is shown in Fig. 1 (step 1). Secondly, single metamaterial is added in between the patches, two metamaterial and three metamaterials in order to reduce the MC and increasing the diversity gain. Metamaterials between the antennas are used to separate the two antenna components well. With and without inserting any metamaterial in between the patches, the results are executed.

2.1 Single T-Shaped Metamaterial MIMO Antenna To get the minimized mutual coupling in a 2 × 2 MIMO antennas, T-shaped metamaterial is inserted. This antenna is designed and simulated in the CST software results are obtained. It shows that less isolation occurs, and it should be improved. The single T-shaped metamaterial is inserted between the patches is shown in Fig. 2. This design is often examined to see if it can improve mutual coupling; the following sections are discussed.

2.2 Double T-Shaped Metamaterial MIMO Antenna To minimize the mutual coupling of the two patch antennas in a MIMO, one more T-shaped metamaterial in series with the previous T-shaped metamaterial is added so that the mutual coupling is reducing, and diversity gain is improved which is shown in Table 2.

200

B. Ramamohan and M. Siva Ganga Prasad

Table 2 Antenna parameters changing from without metamaterial to with metamaterial Antenna Structure

Resonance Frequency (GHz)

S11 At Resonance Frequency

Directivity (dBi)

ECC

DG (dB)

2.3619

−62.58

6.929

1

0

2.3659

−39.65

6.875

0.89

8.5

2.3712

−43.32

6.912

0.058

9.5

2.372

−45.15

6.915

0.018

9.99

2.3 MIMO Antenna with a Series of Triple T-Shaped Metamaterial As shown in Fig. 1, another T-shaped metamaterial is linked in series with a previous material in this cycle; as a result, there is less mutual coupling between the patches. Figure 4 shows the frequency spectrum generated by the new antenna, which ranges from 1 to 9 GHz and has less port isolation. There is slight improvement in the antenna bandwidth for single and triple metamaterial-added antenna. The current flowing from the excitation port to the coupling port is reduced when T-shaped metamaterial is inserted.

3 Results and Their Discussions The triple T-shaped MIMO antenna is constructed or fabricated manually by using the FR-4 substrate. The antenna simulation was conducted in CST software. The novel antenna is also fabricated. The results and characteristics of the CST software output and the measured values are discussed in the following paragraphs.

3.1 S-Parameters and Envelope Correlation Coefficient Figure 3 shows the S11 parameters of without MIMO antenna without metamaterial which is obtained at almost equal to 60 dB. And by inserting the single T-shaped metamaterial, it is obtained at below −40 dB with less mutual coupling and increased

Design of the MIMO Antenna Using Metamaterial …

201

Fig. 3 S11 (Reflection coefficient) parameters of MIMO antenna with and without metamaterial

diversity gain. Finally, by inserting three meta materials in the middle of the substrate or in between patch antennas, the mutual coupling is very less, and diversity gain is high (almost 10 dB) which is shown in Figs. 4 and 6 shows the MIMO surface current distribution with and without metamaterial. Since the current is mutually coupled to two antennas, a large mutual coupling between the monopoles is getting when port 1 of a MIMO antenna is excited. The addition of T-shaped metamaterial between the two antennas reduces mutual coupling, which can be depicted in Fig. 6b. As a result, there is a very poor mutual coupling. Figure 3 shows the accurate CST software-simulated S11 parameters, i.e., reflection coefficient parameters without metamaterial, with single T, with double T, with series of three T-shaped metamaterial. The S-parameters (S11 ) of the novel proposed

Fig. 4 Diversity and ECC the gain of a MIMO antenna with a metamaterial inserted

202

B. Ramamohan and M. Siva Ganga Prasad

MIMO antenna including meta and without metamaterial are shown in Fig. 3. A MIMO antenna with no metamaterial has an operating frequency of 2.36 GHz, whereas the proposed MIMO antenna with single, double, triple T-shaped metamaterial has an operating frequency of 2.37 GHz. For effective MIMO antennas, very low mutual coupling is often preferred. Figure 3 shows that in addition of metamaterial between the two radiating elements reduces mutual coupling as a result increasing the diversity gain. Using the following equation to calculate the envelope correlation coefficient(ECC) from the S-parameter should confirm the MIMO antenna’s performance. ECC =

∗ ∗ S12 + S21 S22 |2 |S11 (1 − |S11 |2 − |S21|2 )(1 − |S22 |2 − |S12 |2 )

For a good MIMO antenna design, there are two things to keep in mind. The ECC value should be very low, ideally zero, and diversity gain should be high, ideally 10 dB. This parameter is also used to evaluate the antenna’s accuracy. Using the equation below, the envelope correlation coefficient can be used to measure the DG:  DG = 10 1 − (ECC)2 Figure 5 depicts the MIMO antenna’s radiation pattern without metamaterial, single, double, and triple T-shaped metamaterial. Figure 6 depicts a MIMO’s surface current density without metamaterial, while antenna 1 is working, antenna 2 receives the current, which can be prevented by putting metamaterial between them, as seen in Fig. 6b. The gain of with and without metamaterial is shown in Fig. 7. To calculate the fabricated antenna S-parameters, an antenna prototype is made and tested with a network analyzer (VNA). The antenna is initially simulated in

Fig. 5 MIMO antenna radiation patterns with and without metamaterial

Design of the MIMO Antenna Using Metamaterial …

203

Fig. 6 a MIMO surface current distribution with no metamaterial. b MIMO surface current distribution with including metamaterial

Fig. 7 With and without metamaterial, the gain of a MIMO antenna is compared

computer simulation technology (CST) software, and then, it is fabricated with the same dimensions. The fabricated antenna design and S-parameters are shown in Fig. 8. The simulated and measured results are almost well matched, and we can see a small frequency change.

204

B. Ramamohan and M. Siva Ganga Prasad

Fig. 8 Hardware implementation of MIMO antenna with results

4 Conclusion This paper shows the novel MIMO antenna design and fabrication. This concept investigates a triple T-shaped metamaterial MIMO antenna with low mutual coupling and high diversity gain. The proposed antenna is 38 × 29.5 × 0.036 mm in size, with two patches separated by 0.308 λo The novel designed antenna has a lower ECCenvelope correlation coefficient (less than 0.019) and a higher diversity gain (DG is equal to 9.9) and operates at 2.4 GHz. The antenna is made to work in the S-band. For wearable textile antennas with different relative permeability, this can also be tweaked and enforced. This antenna has applications which include Bluetooth and various wireless communications.

References 1. A.K. Panda, S. Sahu, R.K. Mishra, A compact dual-band 2 × 1 metamaterial inspired MIMO antenna system with high port isolation for LTE and WiMAX applications. Int. J. RF Microwave Comput. Aided Eng. 27(8), e21122 (2017). https://doi.org/10.1002/mmce.21122 2. J. Yan, J.T. Bernhard, Design of a MIMO dielectric resonator antenna for LTE Femtocell base stations. IEEE Trans. Antennas Propag. 60(2), 438–444 (2012). https://doi.org/10.1109/TAP. 2011.2174021 3. A.K. Biswas, P.S. Swarnakar, S.S. Pattanayak, U. Chakraborty, Compact MIMO antenna with high port isolation for triple-band applications designed on a biomass material manufactured with coconut husk. Microw. Opt. Technol. Lett. 62(12), 3975–3984 (2020). https://doi.org/10. 1002/mop.32539 4. A. Kumar Biswas, U. Chakraborty, Compact wearable MIMO antenna with improved port isolation for ultra-wideband applications. IET Microwaves, Antennas Propag. 13(4), 498–504 (2019). https://doi.org/10.1049/iet-map.2018.5599 5. S. Cui, Y. Liu, W. Jiang, S.T. Yu, A Novel Compact Dual-Band MIMO Antenna with High Port Isolation, ResearchGate, 2011

Design of the MIMO Antenna Using Metamaterial …

205

6. B.T.P. Madhav, Y. Usha Devi, T. Anil Kumar, Defected ground structured compact MIMO antenna with low mutual coupling for automotive communications. Microwave Opt. Technol. Lett. 61(3), 794–800 (2018). https://doi.org/10.1002/mop.31626 7. A.K. Biswas, U. Chakraborty, Reduced mutual coupling of compact MIMO antenna designed for WLAN and WiMAX applications. Int. J. RF Microwave Comput. Aided Eng. 29(3), e21629 (2018). https://doi.org/10.1002/mmce.21629 8. J. Kulkarni, A. Desai, C.-Y.D. Sim, Wideband four-port MIMO antenna array with high isolation for future wireless systems. AEU—Int. J. Electron. Commun. 128, 153507 (2021). https:// doi.org/10.1016/j.aeue.2020.153507 9. D. Manteuffel, MIMO antenna design challenges. IEEE Xplore, 01 Nov 2009. https://ieeexp lore.ieee.org/document/5352587. Accessed 26 Feb 2021 10. M.A. Abdalla, A.A. Ibrahim, Design and performance evaluation of metamaterial inspired MIMO antennas for wireless applications. Wireless Pers. Commun. 95(2), 1001–1017 (2016). https://doi.org/10.1007/s11277-016-3809-4 11. A.K. Biswas, A. Kundu, A.K. Bhattacharjee, U. Chakraborty, Isolator-based mutual coupling reduction of H-shaped patches in MIMO antenna applications, in Advances in Computer, Communication and Control, 2019, pp. 361–366. https://doi.org/10.1007/978-981-13-31220_34 12. T. Prabhu, S.C. Pandian, Design and development of planar antenna array for MIMO application. Wireless Netw. (2020). https://doi.org/10.1007/s11276-020-02253-y 13. C.K. Ghosh, A compact 4-channel microstrip MIMO antenna with reduced mutual coupling. AEU-Int. J. Electron. C. 70(7), 873–879 (2016). https://doi.org/10.1016/j.aeue.2016.03.018 14. G. Varshney, S. Gotra, S. Chaturvedi, V.S. Pandey, R.S. Yaduvanshi, Compact four-port MIMO dielectric resonator antenna with pattern diversity. IET Microwaves Antennas Propag. 13(12), 2193–2198 (2019). https://doi.org/10.1049/iet-map.2018.5799 15. A.C.J. Malathi, D. Thiripurasundari, Review on isolation techniques in MIMO antenna systems. Indian J. Sci. Technol. 06 May 2016. https://indjst.org/articles/review-on-isolation-techniquesin-mimo-antenna-systems. Accessed 26 Feb 2021 16. D. Guha, S. Biswas, T. Joseph, M.T. Sebastian, Defected ground structure to reduce mutual coupling between cylindrical dielectric resonator antennas. Electron. Lett. 44(14), 836 (2021). Accessed 26 Feb 2021 17. H. Hu, F. Chen, Q. Chu, A compact directional slot antenna and its application in MIMO array. IEEE Trans. Antennas Propag. (2016). https://doi.org/10.1109/TAP.2016.2621021 18. A.K. Biswas, U. Chakraborty, A compact wide band textile MIMO antenna with very low mutual coupling for wearable applications. Int. J. RF Microwave Comput.-Aided Eng. 29(8) (2019). https://doi.org/10.1002/mmce.21769 19. B. Ramamohan, S. Usha, P. Ananth, V.S. Lalitha, Design of Dumbbell-Shaped MIMO antenna for wearable applications. Lecture Notes in Electrical Engineering, 2021, pp. 293–303. https:// doi.org/10.1007/978-981-15-8439-8_24

Vulnerability Assessment of University Computer Network Using Scanning Tool Nexpose Kismat Chhillar and Saurabh Shrivastava

1 Introduction In the current scenario of technological advancement in universities, vulnerability assessment of a university computer network (UCN) is highly crucial activity to make the network secure. Vulnerability detection and timely remediation need to be done as soon as possible; else, it can be misused or exploited by the attackers. The main aim of timely assessment of vulnerabilities is to optimize the vulnerabilities and keep them minimum. It is not possible to remove all vulnerabilities, but efforts can be made to keep the vulnerabilities as less as possible. Network vulnerability assessment (NVA) is a cyclic process and needs to be performed at regular intervals for optimization of vulnerabilities. A plethora of scanning tools for vulnerability detection and analysis is available. Every tool differs in their capability and functionality. A scanner scans for network IT assets like servers, containers, switches, firewalls, laptops, desktops, virtual machines, and printers. Network scanners scan a network for vulnerabilities, risk level of vulnerabilities, and remediation to remove or mitigate the vulnerabilities. To properly access the university network condition, at least quarterly assessment of vulnerabilities needs to be performed. For current task, Nexpose scanner is used. The scan results were analyzed for weaknesses or vulnerabilities in the network. The severity of vulnerability is determined looking at the risk score, CVSSv3 score, and exploits available. The vulnerabilities for which exploits are available and have high CVSSv3 score are likely to have high-risk score. Such critical vulnerabilities should be remediated or removed as soon as possible; else, they can create serious and irreversible damage to the network. National Vulnerability Database (NVD) contains details about the vulnerabilities already reported. From scan results, the CVE ID of a vulnerability helps in looking for further details about vulnerability from NVD.

K. Chhillar (B) · S. Shrivastava Department of Mathematical Sciences and Computer Applications, Bundelkhand University, Jhansi, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_23

207

208

K. Chhillar and S. Shrivastava

A vulnerability is uniquely identified by a CVE ID. Current paper discusses about vulnerability assessment of UCN using Nexpose vulnerability scanner. The rest of the paper is structured as follows. Some important concepts related to computer network vulnerability assessment are covered in Sect. 2. Section 3 presents related work that has been previously done by researchers. Nexpose implementation on Bundelkhand University computer network is discussed in Sect. 4. Scan result analysis and vulnerability assessment based on scan results is presented in Sect. 5. Conclusion, limitations of current research, and future scope of the work is discussed in Sect. 6.

2 Computer Network Vulnerability Assessment 2.1 Network Vulnerability Scanning Network vulnerability scanning (NVS) identifies vulnerabilities in a network and also suggests available remediation for removal of vulnerabilities. NVS finds loopholes in a network, computer, or an IT asset connected to a network. Vulnerabilities should be detected and removed or mitigated before being exploited by the malicious entities. Potential harmful and critical severity vulnerabilities should be dealt immediately to maintain network security and reliability. Nexpose scanner has been used for the current work. NVS prioritizes vulnerabilities and also provide ways to remediate the vulnerabilities. Nexpose community edition was used for scanning which is free to use, but only limited hosts can be scanned using this edition.

2.2 National Vulnerability Database (NVD) National Vulnerability Database (NVD) is a repository of vulnerability management data, and the data are represented using security content automation protocol (SCAP) [1]. NVD is a U.S. government repository and is operated by National Institute of Standards and Technology (NIST). NVD reports known vulnerabilities that have been assigned CVEs. CVE stands for common vulnerabilities and exposures and is a standard convention for reporting of security vulnerabilities which are publicly known. CVE is launched by MITRE in 1999. MITRE is a research organization funded by the government.

Vulnerability Assessment of University Computer …

209

2.3 Nexpose Nexpose vulnerability scanner identifies open ports, active services, and applications running on the hosts of a network to be scanned and scans the network for existing vulnerabilities. Nexpose is a vulnerability scanning tool by Rapid7. Various editions of Nexpose are available like enterprise, consultant, ultimate, express, and community. The community edition is free but provides scanning for limited number of hosts. For educational and research purpose, Nexpose community edition is sufficient. “Nexpose scan can discover the services and applications that are running on a host and identify potential vulnerabilities that may exist based on the collected data” [2]. Desired report can be generated after the scan to further analyze the results for better understanding of network state and remediation of critical vulnerabilities having high-risk score. Prioritization of vulnerabilities can be done based on CVSS score, and risk score and an effective solution can be determined to mitigate the vulnerability.

2.4 CVSS Score CVSS is a free and open standard used for determining the severity of vulnerabilities of computer system [3]. CVSS provides severity scores to vulnerabilities. The score ranges from 0 to 10. Lowest score is 0, and highest score is 10. As the CVSS score increases, vulnerability severity also increases. CVSS score is calculated using formula based on various matrices. Severity of vulnerability can be determined using CVSS base score. Apart from CVSS base score, temporal score and environmental scores can also be utilized to determine severity. The latest version of CVSS is CVSSv3.1 which was released in June 2019.

3 Related Work Alzahrani [4] proposed an audit tool to address security issues in Albaha University Network. Kumar and Tlhagadikgora in [5] utilized open-source tools for penetration testing. They used tools for information gathering, vulnerability scanning, and for exploitation of vulnerabilities. In [6], Wang and Yang demonstrated use of vulnerability scanning tools for network defense and ethical hacking. In [7], Daud et al. discussed about scanning tools for vulnerability detection like Nessus, Acunetix, and Zed attack proxy tool. Hannes Holm in [8] evaluated the performance of security tools in remediating security issues of a network. Tundis et al. [9] performed scanning with popular scanning tools and also discussed about the classification of scanners into two groups. In [10], Chalvatzis, Karras, and Papademetriou compared three widely used scanning tools (OpevVAS, Nessus, and Nmap) to evaluate their performance based

210

K. Chhillar and S. Shrivastava

on their risk assessment capability of an organization. Aksu et al. in [11] presented the usability of vulnerability scanning tool OpenVAS. User-based and expert-based testing were carried for evaluating the OpenVAS usability. Raza et al. [12] performed remote scanning of internal network of an organization using built-in tools installed in Raspberry Pi. Patil et al. [13] discussed about importance of ethical hacking for cybersecurity and elaborated tools and techniques utilized for ethical hacking and reconnaissance. Fei and Jing [14] classified vulnerability scanning tasks prior to scanning so that targeted and selected scanning tasks can be performed. Schagen et al. [15] elaborated about network server’s automated vulnerability scanning and investigated a method for safe patch fingerprinting.

4 Vulnerability Scanning Using Nexpose Vulnerability scanning was performed using Nexpose and was implemented on Bundelkhand University computer network. Nexpose community edition was used for scanning purpose. A few subnets of Bundelkhand University were scanned using the tool, and results of scan were analyzed for remediation of vulnerabilities based on their severity. For scanning using Nexpose, the first step was selection of assets for scanning. Asset refers to any device on a network that is discovered during network scanning. A site needs to be created for running a scan. For site creation, we need to specify target assets, scan engine, and scan template, schedules, or alerts. All required options were selected while creating site. After creating a site, we can start scanning at that time itself or at the scheduled time set for scanning. Scanning results are shown in Table 1. Figure 1 clearly shows the number of vulnerabilities detected in the hosts scanned. From the figure, it can be inferred that host H1 has the highest number of vulnerabilities which makes it highly prone to attacks. Table 1 Nexpose scan result of hosts Host

IP address

Vulnerabilities

Risk score

Exploits

H1

172.16.26.158

57

28,387.918

29

H2

172.16.3.33

36

18,871.068

3

H3

172.16.26.172

42

14,854.493

6

H4

172.16.28.4

19

9725.821

2

H5

172.16.28.101

20

8562.071

0

H6

172.16.28.102

19

8377.664

0

H7

172.16.6.35

17

7803.292

2

H8

172.16.26.200

16

7388.0903

0

H9

172.16.3.212

11

5927.094

11

H10

172.16.3.92

9

5052.155

3

Vulnerability Assessment of University Computer …

211

Fig. 1 Vulnerability count detected in scanned hosts

Figure 2 shows the risk score of scanned hosts. The risk score of H1 is the highest. Host H1 also had highest number of vulnerabilities as shown in Fig. 1. Looking at the risk score and number of vulnerabilities present in a host, remediation of vulnerabilities in highly vulnerable host needs immediate attention for remediation of vulnerabilities, and accordingly, prioritization can be done.

Fig. 2 Risk score of scanned hosts

212

K. Chhillar and S. Shrivastava

5 Scan Result Analysis The scan result displays information about the hosts scanned, the operating system running, vulnerabilities identified, whether the scanned host is a virtual or physical machine and also displays overall risk score of the hosts. Risk scores show which vulnerabilities are of high risk to the network which in turn helps in prioritization of the remediation to be taken. Table 2 shows the vulnerabilities detected during the scan. Vulnerabilities are uniquely identified by a CVE ID. Using this ID, further details about a vulnerability can be extracted from NVD. The CVSSv3 score depicts the severity of vulnerabilities. Score 9.0–10.0 means critical severity; 7.0–8.9 means high severity; 4.0–6.9 is medium severity; 1.0–3.9 is low severity, and 0 score means no severity. From Table 2, we can see vulnerability V4 has the highest score of 9.8 which makes it critically severe. V4 needs to be eliminated at the earliest; else, it can cause great damage to the network. Looking at the CVSSv3 score and risk scores of vulnerabilities, prioritization for remediating the vulnerabilities can be done. Figure 3 clearly shows the number of vulnerabilities and exploits available for the vulnerabilities detected in the scanned hosts. The vulnerabilities having high CVSSv3 score and exploits are also available for those vulnerabilities poses a serious risk to the network security and robustness. From Fig. 3, we can infer that host H1 is highly vulnerable because it has the highest number of vulnerabilities, and exploits are also available for most of the vulnerabilities of host H1. Figure 4 shows the CVSSv3 score of vulnerabilities detected. In each host, also, we need to identify which vulnerabilities are of critical and high severity. They need to be remediated on priority basis. From Fig. 4, we can see vulnerabilities V4 and V9 have high CVSSv3 score. So, V4 and V9 should be removed or remediated at the earliest; else, they can be misused by the attackers or entities having malicious intent. Table 2 Vulnerability details of Nexpose scan Vulnerability

CVE ID

CVSSv3

Risk score

Severity

V1

CVE-2017-12615

8.1

546

High

V2

CVE-2017-12616

7.5

515

High

V3

CVE-2017-12617

8.1

603

High

V4

CVE-2018-8014

9.8

639

Critical

V5

CVE-2018-8034

7.5

501

High

V6

CVE-2016-3092

7.5

311

Critical

V7

CVE-2015-5345

5.3

538

High

V8

CVE-2015-5346

8.1

581

High

V9

CVE-2015-5351

8.8

581

High

V10

CVE-2015-5174

4.3

438

High

Vulnerability Assessment of University Computer …

213

Fig. 3 Number of vulnerabilities and exploits of hosts

Fig. 4 Vulnerabilities and their CVSSv3 score

6 Conclusion and Future Work The implementation of network vulnerability scanning tool Nexpose has been successfully performed on Bundelkhand University computer network. Vulnerability scanning of network detected highly vulnerable hosts and severe vulnerabilities.

214

K. Chhillar and S. Shrivastava

Based on the severity of vulnerabilities detected in the network, necessary actions can be taken to remediate the vulnerabilities detected. Current work is limited only to the vulnerability assessment of university computing environment, and free edition of Nexpose tool has been used which has limited functionalities and can scan only limited number of hosts in a network. In the future, a combination of other scanning tools can be implemented on a university network, and results can be analyzed to know extra functionality that the scanners provide for vulnerability scanning of the network.

References 1. 2. 3. 4. 5.

6.

7. 8. 9.

10.

11. 12. 13.

14.

15.

https://en.wikipedia.org/wiki/National_Vulnerability_Database https://docs.rapid7.com/metasploit/vulnerability-scanning-with-nexpose/ https://en.wikipedia.org/wiki/Common_Vulnerability_Scoring_System M. Alzahrani, Auditing Albaha University Network Security using in-house Developed Penetration Tool. J. Phys.: Conf. Ser. 978, 012093 R. Kumar, K. Tlhagadikgora, Internal network penetration testing using free/open source tools: network and system administration approach, in Advanced Informatics for Computing Research (ICAICR 2018). Communications in Computer and Information Science, vol. 956 (Springer, Singapore, 2018) Y. Wang, J. Yang, Ethical hacking and network defense: choose your best network vulnerability scanning tool, in 31st International Conference on Advanced Information Networking and Applications Workshops (WAINA), 2017, pp. 110–113 N.I. Daud, K.A. Abu Bakar, M.S.M. Hasan, A case study on web application vulnerability scanning tools, in Science and Information Conference (2014), pp. 595–600 H. Holm, Performance of automated network vulnerability scanning at remediating security issues. Comput. Secur. 31(2), 164–175 (2012) A. Tundis, W. Mazurczyk, M. Mühlhäuser, A review of network vulnerabilities scanning tools: types, capabilities and functioning, in Proceedings of the 13th International Conference on Availability, Reliability and Security (ARES 2018). Association for Computing Machinery, New York, NY, USA, Article 65 (2018), pp. 1–10 I. Chalvatzis, D.A. Karras, R.C. Papademetriou, Evaluation of security vulnerability scanners for small and medium enterprises business networks resilience towards risk assessment, in IEEE International Conference on Artificial Intelligence and Computer Applications (ICAICA), 2019, pp. 52–58 M.U. Aksu, E. Altuncu, K. Bicakci, A First Look at the Usability of OpenVAS Vulnerability Scanner (2019) S. Raza, F. Jaison, Maliyekkal, N. Choudhary, Remotely scanning organization’s internal network. Int. J. Trend Sci. Res. Dev. (IJTSRD) 4(6), 1139–1141 (2020). ISSN: 2456-6470 S. Patil, A. Jangra, M. Bhale, A. Raina, P. Kulkarni, Ethical hacking: the need for cyber security, in IEEE International Conference on Power, Control, Signals and Instrumentation Engineering (ICPCSI 2017) (2017), pp. 1602–1606 L. Fei, F. Jing, Research on comprehensive risk of network assets and vulnerabilities, in IEEE 3rd Advanced Information Technology, Electronic and Automation Control Conference (IAEAC), 2018, pp. 1787–1791 N. Schagen, K. Koning, H. Bos, C. Giuffrida, Towards automated vulnerability scanning of network servers, in Proceedings of the 11th European Workshop on Systems Security (EuroSec’18). Association for Computing Machinery, New York, NY, USA, Article 5 (2018), pp. 1–6

Development of Cyber-Physical Systems for Water Quality Monitoring in Smart Water Grid Punit Khatri , Karunesh Kumar Gupta , and Raj Kumar Gupta

1 Introduction With the growth in urbanization, many problems arise, jeopardizing the environmental sustainability of the cities. The rapid growth of urbanization also raises numerous challenges, such as water and air pollution, waste disposal, saturated transport, and more energy consumption, resulting in poor public health. These problems can be solved by implementing information and communication technologies (ICT) [1, 2]. Ensuring water quality in distribution systems is a challenge due to frequent failures and pipeline leakages. The current distribution system consists of different components, such as a pump, pipeline network, and valves. The performance and reliability of these components decrease over time, and the distribution systems have a higher risk of pipeline leaks, failures, and wastage of water. It is difficult to access the leakage and consumes time, which results in high wastage of water. To overcome these problems, a smart water grid has been introduced, which is capable of online and real-time water quality, flow and pressure measurement, failure detection, leakage detection in distribution systems. A smart water grid is an integration of various sensing and communication technologies (SCT), which are driven by cloud computing, the Internet of Things (IoT), and data analytics [3]. In this paper, an attempt has been made to form a smart water grid employing a sensing platform, a wireless sensor network (WSN), and an Internet of Things (IoT). Cyber-physical systems are a network of interconnecting individual elements that fulfill sensing, computing, monitoring, and multiple communication among modules, sensory output, and data analytics. The integration of the abovementioned individual P. Khatri (B) · K. K. Gupta Department of Electrical and Electronics Engineering, Birla Institute of Technology and Science (BITS), Pilani, Rajasthan, India e-mail: [email protected] R. K. Gupta Department of Physics, Birla Institute of Technology and Science (BITS), Pilani, Rajasthan, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_24

215

216 Table 1 Water quality parameters, units, and their acceptable range

P. Khatri et al. Parameters

Unit

Acceptable range for drinking

Temperature

°C

5–30 °C

pH

pH

6.5–8.5

Electrical conductivity (EC)

µS/cm

Less than 1000

Dissolved oxygen (DO) mg/l

4 mg/l or more

Oxidation reduction potential (ORP)

200–600

mV

modules of the smart grid forms a CPS structure. The measured water quality parameters were chosen using the guidelines of the Central Pollution and Control Board (CPCB), India [4]. Table 1 shows the selected water quality parameters, units, and their acceptable range. The entire paper is arranged as follows: Section 1 describes the need for water quality monitoring in conventional distribution networks. Section 2 discusses the background, along with related work in water quality monitoring. The proposed architecture and system design are presented in Sect. 3. Section 4 describes the experimental procedure and results. The discussion has been described in Sect. 5, followed by a conclusion in Sect. 6.

2 Background and Related Work With swift growth in urbanization, many researchers, scientists, and economists have targeted the development of smart city architecture. There are different challenges while developing a smart city and can be solved by SCT. Many researchers have used SCT to develop water quality monitoring prototypes/systems in recent years to help develop smart and sustainable city architecture. O’Flynn et al. [5] designed a multiparameter system named “SmartCoast” for water quality assessment in a wireless sensor network. Dinh et al. [6] developed a wireless sensor network for remote monitoring of water salinity. Lambrou et al. [7] proposed real-time monitoring and contamination detection based on a low-cost multi-sensor system. In the last 5–7 years, the Raspberry Pi development board has seen a huge surge in use; it is one of the most popular low-cost single-board computing devices (SBCDs). It has been used in numerous applications, such as video steganography allocation [8], real-time video surveillance systems [9], and communication networks for multiunmanned vehicle networks [10]. Internet of Things (IoT) combines hardware and software to generate enormous amounts of data through multiple sensors and associated devices. IoT transfers the acquired data to the cloud and perceives it again using intelligent systems. IoT is the key component of the SCT infrastructure, which

Development of Cyber-Physical Systems for Water …

217

can be the future of smart cities. IoT has been employed in various real-time applications in recent decades, such as seismic applications [11], big data analysis [12], and wastewater management [13]. Chowdury et al. [14] proposed an IoT-enabled river water quality monitoring system. Priya et al. [15] proposed an IoT-enabled contamination detection in the pipeline for drinking applications. Many researchers have claimed different approaches to smart city applications. The public utility board, Singapore, provides details about challenges, current technology, and the future technology roadmap [16]. Shah and Meganathan [17] proposed the power consumption assumption based on machine learning algorithms. Fabbiano et al. [18] proposed an innovative methodology for leakage detection in water distribution systems. After reviewing the literature, it is clear that there are numerous approaches for implementing smart city architecture. A smart water grid is one of them. People have tried multi-parameter water quality monitoring, IoT, and WSN-based water quality monitoring. In this work, we proposed integrating different hardware and software modules to form a smart water grid.

3 System Design 3.1 Hardware Platform Figure 1 depicts the block diagram of the proposed distribution network. The water from the water tank is supplied to the household through the distribution network’s pumping station. In the network, nodes 2, 3, 4, and 5 are the sensing node for water quality parameter acquisition, and node 1 is the server node, which is located at the water supply station. Two sensing nodes have been developed for demonstration purposes. Figure 2 depicts the sensing node architecture. The sensors and a Zigbee module were interfaced with the Arduino development board. The core controller for the sensing node is the Arduino Uno development board. The Arduino is an open-source prototype used for reading the sensor output, activating an actuator, and publishing the data online as well. It consists of an ATMega2560 microcontroller, which can be programmed using embedded C or C++. The signal conditioning circuit works as a mediator between the sensors and Arduino. The sensors, Zigbee, and signal conditioning circuit require a 5 V power, which can be supplied from Arduino. Only, the Arduino needs to power up with a 5 V DC adaptor. The receiver (Rx) and transmitter (Tx) pins on the Arduino were used for UART communication with the sensors. For serial communication, there is only one hardware UART available on the Arduino board. As a result, a multiplexer (serial port expander) is employed to extend the port so that all of the sensors can communicate at the same time. The sensor array generates a data matrix that contains water quality parameters. The Zigbee module is connected to each sensing node to transfer data via the wireless

218

P. Khatri et al.

Mail

Internet/Cloud

SMS

End User

1

2

3

5

4

Fig. 1 Distribution network architecture Fig. 2 Proposed sensing node setup

DO

EC ORP

pH

Arduino

Temp

ZigBee

Development of Cyber-Physical Systems for Water …

219

Fig. 3 Proposed server node setup

Touchscreen

Keyboard and Mouse

network. Zigbee is a low-power digital radio device for wireless communication, which works as per IEEE 802.15.4 protocol. It can be used for home automation, wireless data collection, small project design, and personal area network (PAN) creation. It covers the range of 10–300 m line-of-sight. For long-distance coverage, the mesh network can be created using Zigbee. It works on low data rates and low power consumption. Zigbee module operates in the industrial, scientific, and medical (ISM) band at 2.4 GHz frequency. Figure 3 depicts the proposed server node setup. The Raspberry Pi 3 development board was interfaced with a keyboard, mouse, and 7-inch LCD touch screen from the Waveshare [19].

3.2 Software Framework In addition to hardware development, a software framework for water quality parameter acquisition and further analysis has been designed. The programming for the sensing node is written in Arduino programming software (IDE). Python was used to program the Raspberry Pi for sensing nodes data acquisition. Python is a commonly used programming language with a broad variety of open-source libraries for data acquisition and analysis. It is also used as a smart application in visual programming for wireless networks [20].

220

P. Khatri et al.

3.3 Performance Evaluation of Sensors The sensor validation has been performed by comparing the obtained results from the established setup with the benchmark instrument (EXO-1 [21]). The mean average percentage error (MAPE) is calculated for performance evaluation and is given by Eq. (1). MAPE =

(Yb − Ye ) ∗ 100 Yb

(1)

where Yb is the results acquired from benchmark instrument and Ye is the results measured from the experimental setup. Here, MAPE is calculated for all the samples measured from both the instruments (i.e., benchmark and the developed one).

4 Experimental Procedure and Results Two sets of the sensor array and Zigbee modules were interfaced in the sensing node with Arduino for data acquisition and communication in the star network. The proposed network has been tested for 270 unknown samples. The sensor validation was already performed in our earlier work [22] and thus not described here. The water quality sensors attached to sensing nodes were calibrated before measurement to avoid any uncertainty in measurement. The measurement was carried out after the sensor reading got stabilized or with minimum variation. After the measurement, the parameters obtained were sent over the wireless sensor network using Zigbee. The received water quality parameters at the server were uploaded on the cloud platform. The necessary steps for the experimental procedure are as follows: 1. 2. 3.

Wireless sensor network setup using Zigbee in a star network. Sensing node data acquisition and transferring it to the server node. Upload the data and water quality rating to cloud “AWS” for real-time monitoring.

4.1 Amazon Web Service (AWS) Platform The data received from the sensing node are structured in JSON format and sent to the cloud using the MQTT protocol. The authorization to access the AWS requires a key ID and an access key. After creating an IAM profile from the AWS root account, the user can get the secret key and access ID. On the AWS cloud server, a database table in .csv format is created using data received from the Raspberry Pi board. This is done with the help of “Amazon DynamoDB Service (ADS).” Scheduled publishing of data is created by the “Amazon Data Pipeline Service (ADPS),” which stores data

Development of Cyber-Physical Systems for Water …

221

in the “Amazon S3” bucket. Finally, “.csv” data in the Amazon S3 bucket can be presented in graphical form and analyzed.

5 Discussion The implementation and practicability of different modules of a smart water grid are investigated in this work. An example of a distribution network was presented for a successful demonstration. As discussed earlier, two sensing nodes and one server node were developed for the same. The measured parameters are temperature, pH, EC, DO, and ORP. According to the experiment results, it can be stated that the proposed architecture can be quite useful for smart water grid development. The alarm or early warning system for any leakage or contamination in the distribution network can be set up based on the water quality rating. The early warning system requires a different decision-making algorithm, which can be implemented using fuzzy logic [23]. The advantages of the proposed work are low-cost hardware as compared to high-end conventional instruments and open-source software for programming and data analysis. The sensors were calibrated with standard solutions before the measurement to avoid any uncertainty in sensor readings. The smart water grid can help locate the exact point of leakage in distribution systems. Also, any illegal or missing connections can be identified based on flow and pressure measurement. The maintenance can be easy-going if any sensing node malfunctioning, which can easily be spotted due to real-time monitoring of the pipeline network. Despite having multiple experimental trials, a smart water grid will face various challenges. First, the cost of updating the current distribution system architecture is too high, which is not possible without government funding. Second, the integration and communication among different sensing nodes in the WSN will be a challenge in a smart water grid. The data generated by different sensing nodes will have a massive amount of data, which requires high storage and big data analytics. The job redesign of existing staff will also be a challenge as the old rules need to be redundant, and a new one will be imposed on them. Society must accept the technology and the fact that it is going to benefit them in the long term. Finally, the smart water grid is still in an immature state and requires more research and experimental testing to utilize the features and advantages of the smart water grid.

6 Conclusion This paper proposes a low-cost online water quality assessment using CPS. The presented work can be the foundation of smart cities where water quality monitoring, distribution systems status, water pressure, and flow can easily be monitored in real time. The developed system may also help monitor the water quality in remote areas where there is no Internet connection. In that case, the Zigbee modules must

222

P. Khatri et al.

be replaced by the GSM modules to update the data directly on the cloud server. The developed system is also competent in monitoring, processing the data, making decisions based on the data analysis results, and displaying the water quality parameters. The smart city program has already been implemented in some cities worldwide, such as the Australian SEQ water grid and the United States National Smart Water Grid project. Many of the countries are also investing in smart city implementation by replacing the current distribution systems. The direction of the future work will be on overall water quality analysis employing artificial neural networks (ANNs). Also, the smart water grid’s remaining features are a smart water meter, end-user intimation in terms of either SMS or email, which we are planning to implement in the future. Acknowledgements The authors would like to thank the Director, BITS-Pilani, Pilani campus, for providing an enabling environment to carry out the research work. One of the authors (Punit Khatri) thanks the Council of Scientific and Industrial Research-Human Resource Development Group (CSIR-HRDG), New Delhi, India, for providing financial support as a fellowship (Award No. 09/719(0101)/2019-EMR-I).

References 1. S.E. Bibri, J. Krogstie, Smart sustainable cities of the future: an extensive interdisciplinary literature review. Sustain. Cities Soc. 31, 183–212 (2017) 2. B.N. Silva, M. Khan, K. Han, Towards sustainable smart cities: a review of trends, architectures, components, and open challenges in smart cities. Sustain. Cities Soc. 38, 697–713 (2018) 3. Managing the water distribution network with a Smart Water Grid. Smart Water 1, 4 (2016). https://doi.org/10.1186/s40713-016-0004-4 4. Central Polution Control Board, Central Polution Control Board; Environmental Standards; Water Quality Criteria (2007). http://cpcb.nic.in/Water_Quality_Criteria.php. Accessed 25 Nov 2017 5. B. O’Flynn, R. Martinez-Catala, S. Harte, et al., SmartCoast: a wireless sensor network for water quality monitoring, in 32nd IEEE Conference on Local Computer Networks (LCN 2007) (IEEE, 2007), pp. 815–816 6. D.T. Le, W. Hu, P. Sikka, et al., Design and deployment of a remote robust sensor network: experiences from an outdoor water quality monitoring network, in 32nd IEEE Conference on Local Computer Networks (LCN 2007) (IEEE, 2007), pp. 799–806 7. T.P. Lambrou, C.G. Panayiotou, C.C. Anastasiou, A low-cost system for real time monitoring and assessment of potable water quality at consumer sites. Sensors J. IEEE 14, 2765–2772 (2012). https://doi.org/10.1109/ICSENS.2012.6411190 8. P. Karthika, P. Vidhya Saraswathi, IoT using machine learning security enhancement in video steganography allocation for Raspberry Pi. J. Ambient Intell. Hum. Comput. 1, 3 (2020). https:// doi.org/10.1007/s12652-020-02126-4 9. L.M. Fawzi, S.Y. Ameen, S.M. Alqaraawi, S.A. Dawwd, Embedded real-time video surveillance system based on multi-sensor and visual tracking. Appl. Math. Inf. Sci. 12, 345–359 (2018). https://doi.org/10.18576/amis/120209 10. D.S. Pereira, M.R. De Morais, L.B.P. Nascimento et al., Zigbee protocol-based communication network for multi-unmanned aerial vehicle networks. IEEE Access 8, 57762–57771 (2020). https://doi.org/10.1109/ACCESS.2020.2982402

Development of Cyber-Physical Systems for Water …

223

11. H. Jamali-Rad, X. Campman, Internet of Things-based wireless networking for seismic applications. Geophys. Prospect. 66, 833–853 (2018) 12. M.M. Rathore, H. Son, A. Ahmad et al., Real-time big data stream processing using GPU with spark over Hadoop ecosystem. Int. J. Parallel Prog. 46, 630–646 (2018). https://doi.org/10. 1007/s10766-017-0513-2 13. V. Edmondson, M. Cerny, M. Lim et al., A smart sewer asset information model to enable an ‘Internet of Things’ for operational wastewater management. Autom Constr 91, 193–205 (2018). https://doi.org/10.1016/j.autcon.2018.03.003 14. M.S.U. Chowdury, E.T. Bin, S. Ghosh, et al., IoT based real-time river water quality monitoring system. Procedia Comput. Sci. 155, 161–168 (2019) 15. S.K. Priya, G. Shenbagalakshmi, T. Revathi, IoT based automation of real time in-pipe contamination detection system in drinking water, in 2018 International Conference on Communication and Signal Processing (ICCSP) (IEEE, 2018), pp. 1014–1018 16. P.U.B. Singapore, Managing the water distribution network with a Smart Water Grid. Smart Water 1, 4 (2016). https://doi.org/10.1186/s40713-016-0004-4 17. S. Sheik Mohideen Shah, S. Meganathan, Machine learning approach for power consumption model based on monsoon data for smart cities applications. Comput. Intell. coin.12368 (2020). https://doi.org/10.1111/coin.12368 18. L. Fabbiano, G. Vacca, G. Dinardo, Smart water grid: A smart methodology to detect leaks in water distribution networks. Meas. J. Int. Meas. Confed. 151, 107260 (2020). https://doi.org/ 10.1016/j.measurement.2019.107260 19. 7inch HDMI LCD (B) (Firmware Rev 2.1) User Manual—Waveshare Wiki. https://www. waveshare.com/wiki/7inch_HDMI_LCD_(B)_(Firmware_Rev_2.1)_User_Manual. Accessed 4 Mar 2018 20. S. Bocchino, S. Fedor, M. Petracca, PyFUNS: a python framework for ubiquitous networked sensors, in Wireless Sensor Networks (Ewsn 2015) (2015), pp. 1–18 21. Yellow Springs Inc., EXO User Manual: Advanced Water Quality Monitoring Platform (2016), pp. 1–154 22. P. Khatri, K.K. Gupta, R.K. Gupta, P.C. Panchariya, Towards the green analytics: design and development of sustainable drinking water quality monitoring system for Shekhawati Region in Rajasthan. Mapan-J. Metrol. Soc. India (2021). https://doi.org/10.1007/s12647-021-00465-x 23. P. Khatri, K. Kumar Gupta, R. Kumar Gupta, Raspberry Pi-based smart sensing platform for drinking-water quality monitoring system: a Python framework approach. Drink Water Eng. Sci. 12, 31–37 (2019). https://doi.org/10.5194/dwes-12-31-2019

Identification of Fake News Using Machine Learning Techniques Swati Pandey, Rashmi Gupta , and Jeetendra Kumar

1 Introduction The term “fake news or phenomenon news” defines manipulated information that is spread as news through social media. To deceive readers and damage the reputation of organizations or individuals is the goal of propagating fake news. When there was no Internet in earlier times, people got their news from the newspaper, radio, or television because in earlier times there were fewer sources available that were providing news. Every news used to be real and reputable, but due to the increasing use of social media, fake news is spreading rapidly through social media [1]. Concocted material spreads faster on social media platforms than real news does. “Falsify information” is defined as “fake news”; this is related to news media material, but not in the organizational process. Fake news may be a big threat to society. Fake news is a significant problem in our society, and it is just going to get worse. Its detection is an uphill struggle, which is considered much more difficult than detecting fake product reviews [2]. People on Twitter are exposed to fake tweets six times faster than true ones, according to certain studies. To construct a system that automatically detects fake or deceptive news, scientists are using artificial intelligence (AI) and natural language processing (NLP).

1.1 Nature of Data Available in Social Media Data available on social media can be categorized as follows: 1.

Text (Multilingual)—In its most basic form, computational linguistics examines the semantics and systematics of the text’s genesis in order to understand

S. Pandey · R. Gupta (B) · J. Kumar Atal Bihari Vajpayee University, Bilaspur, Chhattisgarh, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_25

225

226

2.

3.

S. Pandey et al.

it. Due to the fact that most of the postings are posted in text format, a lot of effort has been put into analyzing them. Multimedia—Multimedia content is included in a single post such as audio, video, photographs, and graphics. A captivating tune draws the audience’s attention instead of reading words. Hyperlinks—A social media post may contain hyperlinks or the URL. Authenticating the origin of a post is essential to gaining audience confidence. Even cross-reference and snapshot embedding of other social media networking sites are in vogue.

1.2 Types of Fake News Fake news can be categorized as follows. 1.

2.

3. 4.

Visual content-based: Generally, fake news columns or headlines heavily use attractive graphic contents that are used to attract readers, and may include distorted images, tampered videos, etc. [3]. User-based: Some people create their fake accounts and spread fake information through their accounts. They especially pick out the audiences that may belong to particular age groups, genders, cultures, political alliances. Information-based: Certain posts or news articles give so-called scientific facts or reasons and make readers believe that the given post or news is authentic. Stance-based: Stance-based news means that it can be present in the form of any precise title or statement in a way that alters its meaning and purpose.

1.3 Some Examples of Fake News The main problem in detecting fake news is not an easy task. It takes a long time to find out whether the news is real or not. Fake news easily deceives the readers. Let us take some examples in the real world related to fake news that made a lot of headlines in its time. Some samples of fake news have been shown in Table 1. As you can see from the examples above, recognizing false news is a difficult task. Numerous factors make fake news nearly impossible to identify. Fake news detection by hand is highly subjective, and assessing an article’s accuracy can be difficult. This is a difficult task for even the most experienced professionals. In the proposed work, we have analyzed the performance of different classifiers for fake news detection. For this purpose, we have used two publically available datasets. After data cleaning, TF-IDF vectorization was performed; then, different classifiers have been used for fake news detection.

Identification of Fake News Using Machine Learning Techniques

227

Table 1 Some examples of fake news Type

Example

False hashtag

#ripPranabMukherjee

Misusage of the information

“Have a beer, it’s good for your brain,” Inc. reported. but you But in reality, the study was done on rats, not on people

Inaccurate and slobby

“1 in 5 CEOs Are Psychopaths, Study Finds.” But the title was wrong

Slanted and biased

“Outlet A: Climate change will produce more storms like Hurricane Katrina.” “Outlet B: climate change can lead to major hurricanes → there haven’t been major hurricanes → Climate change is not real”

2 Review of the Literature Since the last few decades, many researchers are working on fake news detection. There are many aspects involved in detecting fake news, from using chatbots to spreading misinformation, to spreading rumors using clickbait. There is a lot of clickbait available on social media platforms like Facebook. The detection of incorrect or fraudulent information has been the subject of a lot of research. We have reviewed some recent researches on fake news identification. In 2018, Parikh et al. [4] surveyed fake news detection methods. They reviewed clustering, predictive modeling, content cues, non-text cues, linguistic features, and deception modeling. In 2015, Conroy et al. [5] proposed fake news identification using “bag of words.” They illustrated the machine learning approach, with rhetorical structure, the “bag of words” approach and linguistic cue, “discourse analysis,” “network analysis,” and support vector machine classifier. In 2018, Helmstater et al. [3] proposed binary classification for fake news detection. They considered each tweet/post as a “binary classification problem.”. They used a dataset that was manually collected using the DMOZ and Twitter application programming interface (API). For classification, they used NB, DT, SVM, ANN, RF, and XGBoost classifiers. Their method detected 15% fake tweets and 45% real tweets. In 2018, Della et al. [6] proposed fake news identification using a combination of social signals and content. They extracted n-grams counts, TF-IDF, document frequency, word embedding, sentiment polarity score, and linguistic features. They achieved 88% accuracy and 0.91 F1 score with a gradient boosting algorithm. In 2020, Rohit et al. [7] proposed deep CNN network for the identification of fake news. With their model, they achieved 98.36% accuracy. In 2020, Feynza et al. [8] proposed fake news detection using supervised AI techniques. They calculated TF weight and document term matrix. With their method, they achieved better performance. In 2020, Van et al. [9] proposed graph-based method of fake news identification. They found that their method was better in terms of identifying the social context as news.

228

S. Pandey et al.

3 Methodology Detection of fake has become a necessity of today’s digital era because fake news can create a negative aura around society. Many machine learning techniques are already available for the classification task. In this paper, two publically available datasets have been used. After preprocessing task, six classification techniques such as random forest, support vector machine (SVM), logistic regression, Naïve Bayes, knearest neighbors, and passive-aggressive algorithm have been applied to the dataset. Flowchart of the proposed method is illustrated in Fig. 1. Fig. 1 Flowchart of the proposed model

Identification of Fake News Using Machine Learning Techniques

229

3.1 Dataset For building the fake news detection model, two publically available datasets have been used. Both datasets belong to US politics. Dataset 1: This dataset is taken from the GitHub [10]. This dataset was noisy and required cleaning. It contains 7796 news articles having both fake and real news. This dataset is having three columns title, text, and label (fake or real). Dataset 2: This dataset is also taken from the Kaggle repository [11]. This dataset contains 44,933 news articles of both fake and real news. This dataset is having four columns news, title, text, and label (fake or real).

3.2 Data Preprocessing Data preprocessing is always a necessary step to be performed before classification. Impossibly, large amounts of unclean data exist in the world today. When we collect data in the real world, there is a lot of noise, incompleteness, inconsistency in it. So basically, in data preprocessing we removed all the noise or incomplete or inconsistent data and converted it into useful quality data. When we talk about machines, they do not understand the missing values, empty text, images, or video data same as it is; they only understand 0 and 1 s that is why we need to preprocess the data. In machine learning, data preprocessing is a step in which raw data is transformed into quality data so that machines parse it easily.

3.2.1

Stop Word Removal

In data preprocessing, at first all stop words have been removed. By removing the stop word, we can also easily focus on the important information in the news article. In stop word removal, generally, all common words are filtered out like certain articles, prepositions, conjunctions, pronouns, etc. Some examples of stop words in English are “hey,” “the,” “so,” “what,” “a,” “.,” “,” “a,” “in,” etc. The main advantage of removing stop words is that it compresses the data size and also reduces training time. After initializing stop word, we have also initialized maximum document frequency max_df and set its value to 0.7 which indicates to “ignore terms that appear in more than 70% of the document.” The other term for Max_df is “corpus-specific stop words.”

230

3.2.2

S. Pandey et al.

TF-IDF Vectorizer

TF-IDF vectorizer is a combination of two concepts—TF and IDF. TF-IDF vectorizer stands for “term frequency-inverse document frequency.” It is used as a feature extraction technique, and it converts the text into a vector. TF vector is calculated as “number of times a word appears in a document divided by the total number of words in the document” [12]. IDF reduces the weight of terms that appear often in a document (corpus). IDF can be calculated as “dividing the total number of documents by the number of documents in the collection containing the term” [12]. For using TF-IDF vectorizer, we need to initialize tfidfvectorizer.

3.3 Classification As a result of their multifaceted character, fake news is difficult to identify. In this paper, we preferred six strategies such as Naïve Bayes (NB), support vector machine (SVM), random forest (RF), logistic regression (LR), KNN, and passive-aggressive classifier to identify fake news. For classification purposes, the dataset is divided into training and testing data in an 80:20 ratio. Naive Bayes (NB) classifier—It is an amalgamation of the words Naive and Bayes combined into one word. For example, it is known as Naive Bayesian analysis since it relies on Bayesian theorem principles to presume that the presence or absence of one characteristic is unrelated to any other feature’s existence. For spam filtering, sentiment analysis, and article classification, NB algorithm was utilized. Using Naive Bayes classifier (supervised machine learning technique), we have determined the character of news stories (real or fake). You can use this method to classify objects in a certain way. It is mostly employed in “text classification,” especially in a highdimensional training dataset. Support Vector Machine (SVM)—SVM has been utilized to solve both regression and classification issues using supervised machine learning algorithms (SMI). On the basis of a “hyper-plane” that separates data into two classifications, the news is divided into two categories: bogus and real. “Hyper-planes” are decision boundaries that are used to categorize information or data pieces [13, 14]. An important benefit is that it can deal with very large dimensional data and is memory-friendly. Random Forest (RF)—The supervised machine learning algorithms are used to solve both classification and regression issues. Random forest is based on ensemble learning and is a powerful tool. You can use it to address an issue that is complicated by combining several categorization strategies. As a result of the random forest classification technique, a number of decision trees are constructed on distinct subsets of a dataset and then averaged in order to increase predicted accuracy.

Identification of Fake News Using Machine Learning Techniques

231

Logistic Regression (LR)—It is a supervised machine learning algorithm that is commonly used to address classification challenges. “Categorical dependent variable” must have a categorical or discrete value for it to work. As a result of this, the probabilistic values between zero and one can be expressed in the form of “yes” or “no,” “0” or “1,” and so on. There is an “S”-shaped logistic function in this classifier that is used. These values can be either 0 or 1. In addition to logistic function, it is also known as sigmoid function. It is a continuous function ranging from 0 to 1. K-Nearest Neighbor (KNN)—KNN classifier is a supervised machine learning technique. It is used for solving both classification and regression problems. “It does not make any assumption on principle data so that KNN is also known as a nonparametric algorithm” [15] “Lazy learner algorithm” is another name for this type of learning algorithm. This is because KNN saves the data only during the training phase, and then when new data is found, the new data is categorized into a comparable category to the old. Passive-Aggressive Algorithms—“Big data applications” are known to deploy passive-aggressive algorithms. Online learning algorithms are another name for it. A training dataset is used in batch learning one at a time; however, online machine learning algorithms get their input data in sequential order, and the model is updated sequentially, as opposed to batch learning methods. Because these algorithms do not require a learning rate, they are akin to “perceptron models.” However, there is a regularization parameter.

4 Results and Discussion In this paper, two datasets are taken and their accuracy is compared by applying six different classification techniques such as Naïve Bayes (NB), random forest (RF), logistic regression (LR), support vector machine (SVM), K-nearest neighbor (KNN), and passive-aggressive algorithm. Table 2 shows a comparison of all classifiers in terms of accuracy with the two datasets. Table 2 Comparison of applied classifiers

Classifiers

Accuracy Dataset 1 (%)

Dataset 2 (%)

Naïve Bayes classifier

81.61

93.84

Support vector machine

92.58

99.47

K-nearest neighbor classifier

65.35

75.46

Random forest classifier

90.92

99.02

Logistic regression classifier

91.55

98.72

Passive-aggressive algorithm

89.66

99.42

232

S. Pandey et al.

From Table 2, it can be seen that on dataset 1, 81.61% accuracy is achieved using Naïve Bayes classifier, 92.58% accuracy is achieved using support vector machine, 65.35% accuracy was achieved using KNN classifier, 90.02% accuracy was achieved using random forest classifier, 91.55% accuracy was achieved using logistic regression classifier, and 89.66% accuracy was achieved using passive aggression algorithms. On dataset 2, 93.84% accuracy is achieved using Naïve Bayes classifier, 99.47% accuracy is achieved using support vector machine, 75.46% accuracy was achieved using KNN classifier, 99.02% accuracy was achieved using random forest classifier, 98.72% accuracy was achieved using logistic regression classifier, and 99.42% accuracy was achieved using passive-aggressive algorithms. The highest accuracy is achieved using SVM classifier. Figures 2, 3, 4, 5, 6, and 7 show the confusion matrix (CM) of each classification method.

Fig. 2 CM in Naïve Bayes classifier

Fig. 3 CM in SVM classifier

Identification of Fake News Using Machine Learning Techniques

233

Fig. 4 CM in KNN classifier

Fig. 5 CM in random forest classifier

Comparison with Recent Work We have compared the performance of the proposed work with some recent fake news detection methods in Table 3.

5 Conclusion and Future Work Detecting fake news is getting attention among researchers as social media is penetrating the daily life of people. In this paper, two publically available datasets have been used to build the model for fake news detection. After preprocessing, data is divided into training data and testing data. After that, we have six different classifiers

234

S. Pandey et al.

Fig. 6 CM in logistic regression classifier

Fig. 7 CM in passive-aggressive algorithm Table 3 Comparison of proposed method with other methods Method

Methodology

Accuracy (%)

[16]

Machine learning ensemble

99.00

[17]

Graph-aware coattention network

90.84

Proposed method

TF-IDF, classifiers (NB, random forest, KNN, logistic regression, passive-aggressive, SVM)

99.42

Identification of Fake News Using Machine Learning Techniques

235

for detecting fake news. The highest accuracy of 92.58% with dataset 1 and 99.47% with dataset 2 has been achieved using the SVM classifier. In the future, we will analyze Indian news in the Hindi language for fake news detection.

References 1. S. Vosoughi, D. Roy, S. Aral, The spread of true and false news online. Science (80-) 359, 1146–1151 (2018). https://doi.org/10.1126/SCIENCE.AAP9559 2. D.M.J. Lazer, M.A. Baum, Y. Benkler, A.J. Berinsky, K.M. Greenhill, F. Menczer, M.J. Metzger, B. Nyhan, G. Pennycook, D. Rothschild, M. Schudson, S.A. Sloman, C.R. Sunstein, E.A. Thorson, D.J. Watts, J.L. Zittrain, The science of fake news: addressing fake news requires a multidisciplinary effort. Science (80-) 359, 1094–1096 (2018). https://doi.org/10.1126/SCI ENCE.AAO2998 3. S. Helmstetter, H. Paulheim, Weakly supervised learning for fake news detection on Twitter, in Proceedings of 2018 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, ASONAM 2018 (2018), pp. 274–277. https://doi.org/10.1109/ASONAM. 2018.8508520 4. Parikh SB, Atrey PK (2018) Media-rich fake news detection: a survey. In: Proceedings of IEEE 1st Conference Multimedia Information Process Retrieval, MIPR 2018, pp. 436–441. https:// doi.org/10.1109/MIPR.2018.00093 5. N.K. Conroy, V.L. Rubin, Y. Chen, Automatic deception detection: methods for finding fake news. Proc. Assoc. Inf. Sci. Technol. 52, 1–4 (2015). https://doi.org/10.1002/PRA2.2015.145 052010082 6. M.L. Della Vedova, E. Tacchini, S. Moret, G. Ballarin, M. Dipierro, L. De Alfaro, Automatic online fake news detection combining content and social signals, in Conference Open Innovations Association FRUCT, 2018-May, pp. 272–279. https://doi.org/10.23919/FRUCT.2018. 8468301 7. R.K. Kaliyar, A. Goswami, P. Narang, S. Sinha, FNDNet—a deep convolutional neural network for fake news detection. Cogn. Syst. Res. 61, 32–44 (2020). https://doi.org/10.1016/J.COGSYS. 2019.12.005 8. F.A. Ozbay, B. Alatas, Fake news detection within online social media using supervised artificial intelligence algorithms. Phys. A Stat. Mech. Appl. 540, 123174 (2020). https://doi.org/10.1016/ J.PHYSA.2019.123174 9. V.H. Nguyen, K. Sugiyama, P. Nakov, M.Y. Kan, FANG: leveraging social context for fake news detection using graph representation. Int. Conf. Inf. Knowl. Manag. Proc. 1165–1174 (2020). https://doi.org/10.1145/3340531.3412046 10. GitHub—Spidy20/Fake_News_Detection: Fake News Classification WebApp using Flask & Python. https://github.com/Spidy20/Fake_News_Detection#readme. Accessed 16 Sept 2021 11. Fake and real news dataset | Kaggle. https://www.kaggle.com/clmentbisaillon/fake-and-realnews-dataset/metadata. Accessed 16 Sept 2021 12. W. Chung, L.W. Pong, W. Fai, K. Lam, Interpreting TF-IDF term weights as making relevance decisions. ACM Trans. Inf. Syst. 26 (2008). https://doi.org/10.1145/1361684.1361686 13. Y. Seo, D. Seo, C.S. Jeong, FaNDeR: fake news detection model using media reliability, in IEEE Region 10 Annual International Conference/TENCON 2018-October (2019), pp. 1834–1838. https://doi.org/10.1109/TENCON.2018.8650350 14. S. Das Bhattacharjee, A. Talukder, B.V. Balantrapu, Active learning based news veracity detection with feature weighting and deep-shallow fusion, in Proceedings of 2017 IEEE International Conference on Big Data, Big Data 2017 2018-January (2017), pp. 556–565. https://doi.org/ 10.1109/BIGDATA.2017.8257971

236

S. Pandey et al.

15. M. Garg, G. Dhiman, A novel content-based image retrieval approach for classification using GLCM features and texture fused LBP variants. Neural Comput. Appl. 334(33), 1311–1328 (2020). https://doi.org/10.1007/S00521-020-05017-Z 16. I. Ahmad, M. Yousaf, S. Yousaf, M.O. Ahmad, Fake news detection using machine learning ensemble methods. Complexity (2020). https://doi.org/10.1155/2020/8885861 17. Y.-J. Lu, C.-T. Li, GCAN: graph-aware co-attention networks for explainable fake news detection on social media

Epileptic Seizure Detection Using Wavelet-Based Features from Different Sub-bands Pallavi S. Meshram and Damayanti C. Gharpure

1 Introduction One of the major prevalent neurological disorders is epilepsy, affecting roughly 1% of the global population [1]. The epileptic seizure is described as an abnormal, abrupt, simultaneous and repetitive discharge of brain cells [2]. Electroencephalogram (EEG) is a non-invasive method for monitoring brain activity and diagnosing epilepsy that is frequently used in clinic medicine. However, because an epileptic patient’s EEG is recorded for a long period, neurologists find it very hard and time-consuming to discover abnormal data by evaluating continuous EEGs. As a result, approaches for automatic epileptic seizure methods are developed to assist neurologists in detecting an epilepsy occurrence [3]. Automated epileptic seizure detection methods are based on the recognition of distinct patterns such as increase in rhythmic activity [4] and amplitude [5]. The features are obtained from spectral analysis [6, 7] and wavelet analysis [8] to classify the EEG data. Different classifiers are used to determine the presence of epileptic seizures support vector machines (SVMs) [7], artificial neural network [9–11], adaptive neuro-fuzzy systems [11] and nearest neighbor classifiers [12]. In epileptic seizures, the changes are observed in particular frequencies bands. Few studies have concentrated on the analysis of these sub-bands and their relation with the detection of epileptic seizure. In this paper, wavelet analysis and feature extraction method both are presented for the analysis of EEG sub-bands to detect epilepsy. The time frequency features obtained from both these methods are applied to two different groups (Set A and Set E) of EEG signals: (1) SET A consist of the dataset of normal subjects; (2) SET E consist of the dataset of epileptic subjects. Initially, EEG signal is decomposed into five EEG P. S. Meshram (B) · D. C. Gharpure Department of Electronic and Instrumentation Science, Savitribai Phule Pune University, Pune, Maharashtra, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_26

237

238

P. S. Meshram and D. C. Gharpure

sub-bands: gamma, beta, alpha, theta and delta using wavelet analysis. After that the maximum energy and power of these decomposed signals for the particular subband is found out. We have compared the performance of each feature individually. Also, we have used T-test to select the best and most significant features which can discriminate the two classes effectively. Finally, these features are evaluated using Gaussian SVM, Kernel Naïve Bayes, K-nearest neighbor (KNN), Quadratic Discriminant Analysis (QDA) and Ensemble subspace KNN.

2 Dataset Description The proposed method was evaluated using the open access EEG dataset of epilepsy from Bonn University [13]. Each set is made up of 100 EEG signals and each signal has total of 4097 samples. This signal was acquired at a sampling rate of 173.61 Hz using a 128-channel amplifier configuration with an average common reference. For each of the set S, set N, and set F five epileptic patients were documented. Set N and set F records were taken during seizure-free periods and set S records were taken during seizure activity. Five healthy individuals were used to create set Z and set O. In this study, we have used two classes viz. set A (O) consists of EEG signals taken with the eyes closed from healthy volunteers and set E (S) consists of EEG signals during epilepsy seizure. The test dataset consisted of 20%, whereas the training dataset consisted of 80% of the samples. Figure 1 illustrates sample EEG waveform of normal EEG and epileptic EEG.

3 Methodology Figure 2 depicts a block diagram for detecting epileptic seizures. The DWT Daubechies (db6) is used for EEG signal decomposition and split into multiple subbands (γ , β, α, θ and δ). The features from different sub-bands are extracted using power band and maximum wavelet energy. In the next step of the study, the T-test technique is used to select the most weighted features. After that, the five distinct classifiers are used to classify the resulting weighted features. Six distinct performance metric values were calculated to evaluate the proposed approach’s performance.

3.1 Signal Preprocessing Artifacts and a poor signal-to-noise (S/N) ratio describe the raw EEG output. Preprocessing is a denoising technique aimed at improving the S/N ratio. The wavelet transform is a renowned method for reducing signal noise. When a noisy signal is wavelet

Epileptic Seizure Detection Using Wavelet-Based Features …

239

Fig. 1 EEG waveform

decomposed, the underlying signal information is concentrated in a few large absolute valued wavelet coefficients. As a result, by applying the threshold to the wavelet coefficients, denoising may be performed [14].

3.2 Wavelet Analysis Wavelets are employed to extract features in addition to denoising the signals. The correlation coefficients are computed after the mother wavelet is shifted in the xaxis by a small interval. This process is repeated for different y-axis scaling factors (dilations) [12]. The discrete wavelet transforms (DWT) have a multi-scale property that allows them to decompose a signal. The signal is split into several scales, each of which represents a different level of coarseness. The four level decomposition of a signal x[n] is shown in Fig. 3. The four level decomposition results in five sub-bands. Each stage consists of two filters followed by two down-samplers. In each stage, the upper filter is the high pass filter represented by g[.] and the lower filter is the low pass filter represented by h[.]. g[.] is the discrete mother wavelet whereas h[.] is its mirror version (Fig. 4).

240

P. S. Meshram and D. C. Gharpure

Fig. 2 The block diagram of the proposed algorithm for detection of epileptic seizure

Fig. 3 Wavelet decomposition

3.3 Feature Extraction Amplitude measurements are dependent on both energy and average power. The average power is the signal mean square, and the energy is the total of a squared signal. In this work, we investigated the features wavelet energy and wavelet power for discrimination of normal and epileptic EEG. These two features are extracted

Epileptic Seizure Detection Using Wavelet-Based Features …

241

Fig. 4 Four level decomposition of EEG signal into 5 EEG sub-bands using sixth order Daubechies wavelet, a Normal EEG sub-bands, b epileptic EEG sub-bands

from each EEG sub-band. Thus, total 10 features are extracted in time frequency domain using wavelet analysis. Maximum Energy Wavelet energy (WE) is a property that can enhance amplitude variation between regions indicating amplitude in homogeneity inside a region. Furthermore, this

242

P. S. Meshram and D. C. Gharpure

feature can amplify any little differences that may exist between two regions [15]. The wavelet energy of the sub-band signals is the simplest feature which is extracted from different approximate and detailed wavelet coefficients. The over-all energy contained at various frequency levels and locations can be estimated by means of these energy signature [16]. WT is applied to a signal and iterated until a desired number of levels are reached. For analyzing the wavelet energy, all detail sub-band coefficients from n levels are employed [17]. The change in signal amplitude is represented by these wavelet energy values. Each amplitude in the signal is expected to exhibit a discrete range of energy value for the wavelet. In time domain, wavelet energy is defined as E it =

p 

(Ti (x))2

(1)

x=1

The energies obtained from Eq. 1 indicate the strength of the signal over time. E i t represents energy feature vector whereas I vary from 1 to n, n stands for highest decomposition level which gives total information of a signal. The feature vector is formed using wavelet energy feature for different levels. These feature vector has abilities to distinguish regions in the signal. The (WEF) is defined as the normalized energy vector which can be formulated as below: E total =

R K    k 2 Ci

(2)

k=1 i=1

whereas R represents the over-all number of wavelet coefficients for each sub-band, Ck represents the ith coefficients of the Kth sub-band and K is the total number of sub-bands. For n level wavelet decomposition, feature vector size is n. At each level, the total energy obtained from detailed sub-bands is also employed as a feature. Band Power The EEG signal is used to determine the various frequency sub-bands. The band power for each sub-band can be determined as a feature. The power is calculated by using the formula given in equation N −1 1  Px = (x(n))2 N N =0

(3)

where x(n) represents the EEG signal, Px represents the calculated power and N denotes the length of the EEG signal.

Epileptic Seizure Detection Using Wavelet-Based Features …

243

3.4 Feature Selection The collected features may not be meaningful for the discrimination of different classes. As a result, we have obtained a line plot by considering 20 EEG signals from each class. To develop an efficient classifier, the appropriate features which shows discrimination between two classes can be selected. We have also validated the result of feature selection using T-test [18]. This is a scalar method that analyses the null and alternative hypotheses while examining at only one feature individually. The null hypothesis states that there is no substantial association between a particular trait and the population. As a result, eliminating the feature has no effect on the model’s performance.

3.5 Classification After extracting the energy and power features from different sub-bands, the dataset is fed for classification. To categorize the information various types of classifiers can be used [19]. In this experiment, we compared the classification indices of five different classifiers-Fine K-nearest neighbors (KNN), Kernel Naïve Bayes, Ensemble subspace KNN, Coarse Gaussian SVM and Quadratic discriminant analysis (QDA).

3.6 Performance Evaluation Six distinct performance measure values were computed to evaluate the correctness of the suggested technique. Precision, sensitivity, classification accuracy (Acc), specificity, F1 score (F1) and AUC are the performance measure values (Eqs. 4–8). Acc =

TP + TN TP + TN + FP + FN

(4)

2TP 2TP + FP + FN

(5)

TP TP + FN

(6)

F1 =

Sensitivity = Specifity =

TN TN + FP

(7)

Precision =

TP TP + FP

(8)

244

P. S. Meshram and D. C. Gharpure

where TP indicates true positive, FP indicates false positive, TN indicates true negative, and FN is false negative. For all indices, usually a higher number suggests that the approach has a better classification performance.

4 Result and Discussion 4.1 Feature Extraction Phase The dataset consisting two sets (O and S) are the two classes viz. normal and seizure are used for the evaluation. The raw EEG signal is then split into various sub-bands Gamma, Beta, Delta, Theta and Alpha. After that, these sub-bands are considered to extract two features in the time-frequency domain from each segment of EEG. Because all five brain frequency bands were considered, the total ten features were obtained for each sub-band. The classification method is subsequently employed by examining the features produced from these seven features. The extracted features are maximum energy and band power obtained in the time-frequency domain. The result of the feature selection procedure is shown in Fig. 5. We have obtained the line plots to observe the discrimination between two classes for each feature at a time. These plots are obtained considering 20 EEG signals from each set.

Fig. 5 Line plots for different features extracted from EEG sub-bands

Epileptic Seizure Detection Using Wavelet-Based Features …

245

According to the derived graphs, the seven features maximum alpha, maximum delta, maximum theta, power alpha, power beta, power delta, power theta shows significant discrimination. The features maximum gamma and maximum beta for both the classes overlaps and does not show any discrimination. That means maximum energy of gamma and beta waves which has higher frequency does not provide information related to epileptic event, whereas the lower frequency waves such as alpha, beta and gamma contains information on the basis of maximum energy. The power feature discriminates between both the classes for each sub-band. The selection of relevant features is also validated using T-test, shown in Fig. 6. The values obtained using T-test for different features are obtained to rank the features. This enables to choose best features among the total features. The values obtained for maximum delta, maximum theta, power delta, power theta, maximum gamma, power alpha are 15.12, 14, 9.2, 8.4, 7.3, 6.3, respectively. Higher the value, better the discrimination using the feature.

Fig. 6 Feature selection using T-test

246

P. S. Meshram and D. C. Gharpure

4.2 Classification Phase Further the chosen features are classified using different classifiers QDA, Kernel Naïve Bayes, Fine KNN, Ensemble subspace KNN, Coarse Gaussian SVM. The classification accuracy (ACC), precision, sensitivity, specificity, F1 score (F1) and AUC for each classifier is calculated. On the basis of the classification results the most appropriate classifier can be selected. The classification is obtained for two classes A and E. The different performance metric value for each classifier is shown in Table 1. Ensemble subspace KNN outperforms for all the classification metrics compared to other classification methods. The QDA classifier has the classification accuracy 97.5%, Kernel Naïve Bayes accuracy is 97.5% whereas Coarse Gaussian SVM and Fine KNN both possess 92.5% accuracy. Table 2 lists some of the most recent techniques as well as their accuracy ratings. The suggested method’s classification performance for epileptic seizure detection is compared to that of existing state-of-the-art techniques. This methodology is evaluated for two class problem using set S and set O collected from a database at Bonn University. Table 1 Classification results using time-frequency features for different classifiers Accuracy Classifier (%)

Precision Sensitivity Specificity F1 Precision Sensitivity ROC score (AUC)

97.5

Quadratic 1 Discriminant

0.95

1

0.97

1

0.95

1

97.5

Kernel 1 Naive Bayes

0.95

1

0.97

1

0.95

1

92.5

Coarse Gaussian SVM

0.87

1

0.85

0.93

0.87

1

0.99

92.5

Fine KNN

0.87

1

0.85

0.93

0.87

1

0.92

100

Ensemble Subspace KNN

1

1

1

1

1

1

1

Table 2 Comparative analysis of the proposed method with recent methods Authors (year)

Method

Accuracy (%)

Sharmila et al. [19]

DWT and Naïve Bayes

99.12

Raghu et al. [20]

Matrix determinant, KNN and MLP

97.10

Gupta et al. [21]

Hilbert marginal spectrum and Random Forest

97.7

Nabil et al. [22]

DWT, LLE and multi-class SVM

98.50

Dash et al. [23]

TQWT, entropy features and hidden Markov model (HMM)

99.58

Proposed method

DWT, Power band, maximum energy and Ensemble subspace KNN (ES-KNN)

100

Epileptic Seizure Detection Using Wavelet-Based Features …

247

Table 2 shows that the proposed method, which incorporates wavelet features and ES-KNN, outperforms previous methods. By comparing with other methods, it is been observed that combination of the wavelet features, i.e., power band and maximum energy along with ES-KNN provides good classification accuracy.

5 Conclusion Visual analysis of EEG recordings for long-term monitoring by professional physicians is a very expensive and time-consuming method for evaluating epileptic seizures. The current study developed a seizure detection algorithm based on EEG signals to overcome these issues. In the proposed methods, the wavelet transform is used to decompose single-channel EEG recordings into different coefficients. These coefficients are related to different mental activities and provides information about these activities. The wavelet energy and band power of delta, theta, alpha and beta shows discrimination in both the classes, i.e., normal and epileptic. It can be concluded that the time frequency features and Ensemble subspace KNN have the capability to be used as a useful evaluation approach for analyzing EEG signals. The methodology used in this study employs the simplest feature extraction method using wavelet transform to discriminate between two classes. It has been observed that using the simple and efficient features like power band and maximum energy, two classes are classified efficiently. This methodology has to be validated for other classes as well. For further investigations, preictal and interictal epileptic seizure for multi-class categorization will be evaluated. As future work, the methodology based on maximum energy and power bands will be investigated using other datasets.

References 1. F. Mormann, R.G. Andrzejak, C.E. Elger, K. Lenhnertz, Seizure prediction: the long and the winding road. Brain 130(2), 314–333 (2007) 2. J. Gotman, Automatic detection of seizures and spikes. J. Clin. Neurophysiol. 16(2), 130–140 (1999) 3. Q.S. Mian, S. Abdulhamit, Effective epileptic seizure detection based on the event-driven processing and machine learning for mobile healthcare. J. Ambient Intell. Hum. Comput. (2020) 4. W.R.S. Webber, R.P. Lesser, R.T. Richardson, K. Wilson, An approach to seizure detection using an artificial neural network (ANN). Electroenceph. Clin. Neurophysiol. 98(4), 250–272 (1996) 5. P.F. Prior, R.S.M. Virden, D.E. Maynard, An EEG device for monitoring seizure discharges. Epilepsia 14(4), 367–372 (1973) 6. V.P. Nigam, D. Graupe, A neural-network-based detection of epilepsy. Neurol. Res. 26(6), 55–60 (2004) 7. B. Gonzalez-Vellon, S. Sanei, J.A. Chambers, Support vector machines for seizure detection, in Proceedings of the 3rd IEEE International Symposium on Signal Processing and Information Technology, 14–17 Dec, Germany, pp. 126–29 (2003)

248

P. S. Meshram and D. C. Gharpure

8. H. Adeli, Z. Zhou, N. Dadmehr, Analysis of EEG records in an epileptic patient using wavelet transform. J. Neurosci. Meth. 123(1), 69–87 (2003) 9. N. Kannathal, U.R. Acharya, C.M. Lim, P.K. Sadasivan, Characterization of EEG-A comparative study. Comp. Meth. Prog. Biomed. 80(1), 17–23 (2005) 10. D.E. Lerner, Monitoring changing dynamics with correlation integrals: case study of an epileptic seizure. Physica D 97(4), 563–576 (1996) 11. N.F. Gler, E.D. Beyli, I. Gler, Recurrent neural networks employing Lyapunov exponents for EEG signals classification. Expert Syst. Appl. 29(3), 506–514 (2005) 12. O. Faust, U. Rajendra Acharya, H. Adeli, A. Adeli, Wavelet-based EEG processing for computer-aided seizure detection and epilepsy diagnosis. Seizure 26, 56–64 (2015) 13. Epileptologie Bonn/Forschung/AG Lehnertz/EEG Data Download n.d. http://epileptologiebonn.de/cms/front_content.phpidcat=193&lang=3. Last accessed 2020/2/3 14. A. Temko, G. Boylan, W. Marnane, G. Lightbody, Robust neonatal EEG seizure detection through adaptive backgroundmodeling. Int. J. Neural Syst. 23(4), 1350018 (2013) 15. K.C. Hsu, S.N. Yu, Detection of seizures in EEG using sub band nonlinear parameters and genetic algorithm. Comput. Biol. Med. 40, 823–830 (2010) 16. X.-Q. Wu, K.-Q. Wang, D. Zhang, Wavelet energy feature extraction and matching for palmprint recognition. J. Comput. Sci. Technol. 203, 411–418 (2005) 17. R.J. Oweis, E.W. Abdulhay, Seizure classification in EEG signals utilizing Hilbert-Huang transform. Biomed. Eng. Online 10, 38 (2011) 18. L.-Y. Hu, M.-W. Huang, S.-W. Ke, C.-F. Tsai, The distance function effect on k-nearest neighbor classification for medical datasets. Springerplus 5, 1304 (2016) 19. A. Sharmila, P. Geethanjali, DWT based detection of epileptic seizure from EEG signals using naive Bayes and k-NN classifiers. IEEE Access 4, 7716–7727 (2016) 20. S. Raghu, N. Sriraam, A.S. Hegde, P.L. Kubben, A novel approach for classification of epileptic seizures using matrix determinant. Expert Syst. Appl. 127, 323–341 (2019) 21. V. Gupta, A. Bhattacharyya, R.B. Pachori, Automated identification of epileptic seizures from EEG signals using FBSE-EWT method, in Biomedical Signal Processing, pp. 157–179 (2020) 22. D. Nabil, R. Benali, F. Bereksi Reguig, Epileptic seizure recognition using EEG wavelet decomposition based on nonlinear and statistical features with support vector machine classification. Biomed. Tech. 65(2), 133–148 (2020) 23. D.P. Dash, M.H. Kolekar, Hidden Markov model based epileptic seizure detection using tunable Q wavelet transform. J. Biomed. Res. 34(3), 170 (2020)

An Improved Locality-Sensitive Hashing-Based Recommender Approach in a Distributed Environment Angadi Anupama, Pedada Saraswathi, Patruni Muralidhara Rao, and Gorripati Satya Keerthi

1 Introduction The volume and range of services available on the Internet are continuously rising as the Internet grows in popularity [1]. The distributed environment generates a large amount of data, which necessitates a large number of resources to store and analyze [2]. In addition, the act of dealing with distributed data consumes a lot of energy and exposes personal information [3]. As a result, consumers typically find finding the services they require among a huge number of applicants to be an extremely exhausting and time-consuming task. Furthermore, traditional manual searching for acceptable services is inefficient and inaccurate [4]. In this case, researchers use a variety of lightweight service recommendation algorithms to overcome the aforementioned issues [5]. Collaborative filtering (CF), for example, is frequently used to evaluate users’ previous data on Web services in order to generate relevant recommendations and relieve the heavy strain on users [6]. Traditional collaborative filtering techniques, on the other hand, frequently presume that data is consolidated from a single platform rather than various platforms. For example, user A has used user B used certain platform X Web services, A. Anupama Anil Neerukonda Institute of Technology & Sciences, Visakhapatnam, India e-mail: [email protected] P. Saraswathi (B) GITAM Institute of Technology, GITAM Deemed to be University, Visakhapatnam, India e-mail: [email protected] P. Muralidhara Rao Vellore Institute of Technology, Vellore, India e-mail: [email protected] G. Satya Keerthi Gayatri Vidya Parishad College of Engineering, Visakhapatnam, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 A. K. S. Pundir et al. (eds.), Recent Trends in Communication and Intelligent Systems, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-19-1324-2_27

249

250

A. Anupama et al.

while user C used Web services on platform Y. As a result, users A and B have data from two distinct platforms. In general, there are two significant roadblocks to estimating the similarity of A and B. First, Platforms X and Y are frequently hesitant to share information based on their data with one another due to potential privacy concerns; in this scenario, normal CF-based algorithms are unable to detect the similarity between A and B [7]. Recently, the locality-sensitive hashing (LSH) procedure [8] was developed for distributed context; it has been used to generate efficient and privacy-preserving service suggestions. According on the QoS data, the LSH can be used to swiftly discover a target user’s similar friends while respecting their privacy. Following that, suggested results are made by taking into account the obtained comparable friends’ tastes. However, the recommender system may be unable to develop in some instances or deliver any satisfactory recommended result; this is referred to as a failure of a recommendation. Existing service recommendation systems based on LSH rarely examine this kind of failure in a recommendation and the accompanying exception handling solutions; as a result, the recommender system’s robustness is severely weakened. However, in the distributed environment context, critical QoS information for recommendation judgments is frequently not centralized; rather, distributed sensors monitor and store data across several distributed environments [9]. In this context, a recommender system must swiftly and accurately integrate or fuse scattered data across many cloud platforms in order to produce comprehensive and reliable recommendation selections. To secure sensitive company information and comply with legislation, maintaining user privacy through the aforementioned the process of integrating data from multiple sources is important but challenging [10]. In view of various research issues, we offer the inverse LSH method that can be used to locate a target user’s opposite users with quite different preferences from the intended user [11]. After that, we deduce the user’s potential buddies from the user’s aim in a roundabout way to the target user’s foes in order to take care of the exceptions and failures caused by recommendations. Rest of the paper is laid out as follows: Sect. 2 talk about the recent related works inline with the proposed mechanism; Sect. 3 emphasizes the problem formulation and motivation; Sect. 4 discuss the proposed methodology; and lastly, Sect. 5 concludes the research contribution.

2 Related Works Many academics have looked into the privacy issues that arise throughout the recommendation process and have come up with solutions [12]. According to the authors of [13], to secure the remaining QoS data, a user can broadcast partial QoS data to the service community. In [14], the authors employ the amount of publicly available data as an adjustable parameter to create an acceptable tradeoff between data

An Improved Locality-Sensitive Hashing-Based Recommender …

251

availability and data privacy, then turn the privacy-preservation problem into a multiobject optimization problem. However, users’ sensitive information may be threatened due to the restricted data supplied in the aforementioned method. The microaggregation concept is considered based on the K-anonymization, which safeguards users’ sensitive data, which was employed in [15]. (e.g., user location). Though, when anonymous data is used to create service recommendation decisions, typically, there is a balance to be struck between data accessibility and data privacy. As a result, the suggestion when it comes to accuracy, it is not always as good as it should be. Furthermore, these approaches ignore the possibility of service recommendation failures. In work [16], an encryption technique is used to ensure that sensitive data is kept private. Encryption procedures, on the other hand, often have a huge computational expense as well as a strong privacy-protection measure; there will be a large delay method. LSH was recently integrated into service recommendation in the big data scenario to achieve scattered data integration and privacy-preservation requirements as an effective and efficient technique for finding similar peers. In our previous work [refs around 3 citations], to give privacy-preserving service suggestions, LSH was integrated with user-based CF. Similarly, in [17], LSH is paired with item-based CF to create a limited-privacy service index table, which is subsequently used to provide service recommendations. These LSH-based recommendation systems, on the other hand, ignore recommendation failures and the resulting exception handling. When the data on service quality is utilized to make suggestions that varies widely, the accuracy of traditional LSH-based recommendation algorithms is poor. The authors in [18] mentioned tweak the classic LSH technique to make it relevant to a wide range of service quality data, in order to achieve accuracy while maintaining privacy suggested list. Finally, using the distributed dataset, WS-DREAM, a variety of experiments are carried out. Experiment results show that our solution is more accurate and efficient than other cutting-edge techniques in securing private data in education (e.g., student information in universities). When distributed service proposals are made, the LSH technique is used to protect users’ private information, according to the authors of [19]. LSH may also provide “False-positive” or “False-negative” recommended results because it is primarily a probability-based search strategy; as a result, their method augment LSH to improve recommendation accuracy with AND/OR operations. Finally, they validated the viability of DistSRAmplify-LSH; their proposed recommendation approach, in terms of suggestion accuracy and efficiency while maintaining privacy in distributed systems environments using a WS-DREAM is a collection of studies based on a real-world dataset of distributed service quality. Approaches to LSH-based recommendation that are currently in use frequently address only one QoS factors, ignoring the more difficult but more typical multidimensional recommendation scenarios. To address this flaw, authors in [20] improves the classic LSH and proposed a unique mechanism. The goal of the LSH-based service recommendation technique is to protect users’ privacy in a variety of quality areas. During the process of recommending mobile services to a large number of people.

252

A. Anupama et al.

Based on the recent findings of the literature survey, current approaches to crate recommendations are either ineffective or ineffective to preserve user privacy or miss recommended failures and exceptions. In light of this flaw, we advise using the opposite LSH method to manage the exclusions caused by recommendation failures, thereby improving the resilience of the system of recommendations.

3 Problem Statement and Motivation The symbols to be used in this work can be defined as follows to shorten the future discussion. For simplification, the RS can be described by a triple (U, M, R) as below: 1. U = {u1 , u2 , …, um }: users who ever gives ratings in set M. 2. M = {m1 , m2 , …, mn }: rated movies. 3. R = {1, 2, 3, 4, 5}: ratings provided by the users i.e., UX M → R. In view of the symbolic representation described above, the LSH-RS with sparsely data can be defined as follows: a RS uses hashing to discover likelihood preferences for a target user; through the above procedure, collisions are minimized and decrease the curse of the dimensionality while conserving relative distance between data elements. To challenge this issue, we present a new technique in Sect. 4.

4 Proposed Framework The task of identifying likelihood neighbors is very common in recommender systems. It is an application of finding similar preferences in purchase or browsing history. Although brute force checks all likely combinations of neighbors, it is not scalable. An estimated algorithm to achieve this task has been an area of research interest in current days. While these algorithms do not give assurance for the precise answer, moreover they do not provide worthy approximation. Locality-sensitive hashing has many applications such as near-duplicate detection, image search, and video finger printing. LSH mentions to a family of functions to hash data element into buckets so that every data element is assigned with a same bucket when the remaining data elements are highly similar. However, dissimilar data elements are far from others and are likely to be located in different buckets. This makes it easier to classify opinions with various degree of likelihood.

4.1 Collaborative Filtering The aim of CF algorithms is to forecast the future preferences of a certain product for a certain user based on earlier preferences on other similar products. In a classic

An Improved Locality-Sensitive Hashing-Based Recommender …

253

CF model, there is a group of users and a group of products, where usually, users are exponentially greater than products. Every user can share opinions on a subgroup of products. Each product might also obtain opinions from numerous users. Opinions can be shared either explicitly or implicitly based on information source; for instance, social networking sites like Amazon and MovieLens accept explicit rating score within a specified range and online sites like JioSaavn and Gaana.com are the examples of implicit sources which admits the number of times a song was played. It is likely to have most of the products receives no opinions. This is known as “cold-start” issue. Traditional CF receives the User X Item rating matrix R. Each entry in R is an opinion score of the user ui on a product ij . The fundamental step is to compute similarity for all pairs of products S = {S (1,2) , S (1,3) , …, S (1,n) }. To compute similarity, various metrics exists such as Pearson correlation, Jaccard, and cosine. This is the utmost time-consuming task in CF. Once the algorithm computes this similarity score, the rest is very fast to recommend products for a target user.

4.2 Producing Signature Using LSH For an existing system, it typically caches all scores and fitting R in memory is a challenging task. For instance, for a matrix R with less users and products, it may take around 4 GB of memory in Java environment. We identify this drawback of CF in a traditional mode; an obvious solution to deal with high-dimensional data is a distributed LSH on Apache Spark. The following is the source code for preliminary data preparation: Algorithm 1 Source code for data preparation val opinions = opinions.map(lambda x:(x.split(“,”))) //creating userID and movies list val list = opinions.collect() val signature = opinions.map(generateSignature) //generating signatures val band = signature.flatMap(lambda x: generateBand (4,5,x)).groupByKey() //generating 5 bands and 4 rows for each user val pairs = band.flatMap(swap).groupByKey().filter(lambda x: len(x[1]) > 1) //finding similar pairs Based on the description of LSH, a combination of MinHash and LSH seeks to resolve the issues of pairwise problems. They make it potential to calculate all likely matches only once for every user so that computational issues raises linearly rather than exponentially. The following are the algorithm for generating Shingles and signature matrix:

254

A. Anupama et al.

Algorithm 1: Input: Preference Matrix Output: Computational Matrix 1. Initialize compmat[i]=0 for i between 1 and n 2. for each movie of the user 3. if movie not in Preference Matrix: 4. compmat[movie]=0*len(user) 5. for each user 6. temp=compmat[user[0…11]] 7. for each t in temp 8. set compmat[t][user]=1 9. end if 10.end for 11.end for 12.end for

Algorithm 2. Single-pass execution of a Shingle Input: Computational Matrix Output: Signature Matrix 1. Randomly generate hash functions 2. Initialize shingle[i][j]=9999, sigmatrix=#M 3. for(r=1;r