Dr. R.Bagavathi Lakshmi has more than 22 years of Academic experience in the Education field. She has completed MCA,M .,. She is an expert in AI & ML,IoT,Data mining, Java Programming, and Oracle. She is currently working as Associate Professor in the Department of Computer Applications, VELS Institute of Science Technology and Advanced Studies (VISTAS), Chennai, India. At present she is guiding six research scholars. She has published many articles in International and National Journals, participated and presented papers in national and International Conferences, Workshop and FDP. She has published many and books & book chapters. She has published patents. She has received many awards.
RESEARCH, TEACHING, or OTHER INTERESTS
Computer Science Applications, Computer Science, Computer Networks and Communications, Software
Fraud Detection in Telecommunications Exploiting NADAM Optimizer with Long Short-Term Memory Algorithm Vijayalakshmi R, Bagavathi Lakshmi R Proceedings of the 4th International Conference on Innovative Mechanisms for Industry Applications Icimia 2025, 2025 Fraudulent calls significantly impact revenues from telecommunications operators. Because service providers may experience a significant loss of income, fraud detection is crucial. Information for fraud detection may be found in Call Detail logs (CDRs), which are logs of customer conversations that include details like the call's manner, issue, caller ID number, kind of call or message, advertiser business number, state, location, feedback1, feedback2, feedback3 and zip code. This study's approach makes use of deep learning and an advanced optimizer, two commonly used methods for detecting fraud. This paper aims to propose a workable approach to detecting fraud on the CDR. The dataset has 852535 rows of variables. To increase classification accuracy, the dataset is pre-processed using Autoencoder, features are chosen for classification using K-Means, and fraud is classified using LSTM (Long Short-Term Memory) with NADAM optimizer. Performance indicators including accuracy, precision, recall, and f1 score were 0.98, 0.91, 0.97, and 0.89 for the proposed deep learning model (LSTM-NADAM).
Teknomo-Fernandez Kernelized Discriminant Analysis-Based Connectionist Deep Multilayer Perceptive Neural Learning for Human Activity Recognition Data Driven Analytics for Healthcare Artificial Intelligence and Machine Learning for Medical Diagnostics, 2025
Hybrid Deep-Spiking Neural Architectures for EEG-based Brain Signal Analysis Vedavalli S, Bagavathi Lakshmi R Proceedings of the 6th International Conference on Smart Electronics and Communication Icosec 2025, 2025 The analysis of brain signals via Electroencephalography (EEG) will be essential to build trusted brain-computer interfaces (BCIs) and clinical neurodiagnostics. While deep learning models have shown high accuracy in the analysis of brain signals, they require substantial computational resource and energy, and they do not offer meaningful interpretability, whereas spiking neural networks (SNNs) have shown energy efficient approaches in the task of brain signal analysis, however in many contexts they have not displayed success with high-dimensional EEG signal. The approach in this paper is a Hybrid Deep-Spiking Neural Network (DSNN) model that fuses CNN-LSTM deep encoders with multi-layer SNN modules. The EEG signal is preprocessed, encoded via CNN-LSTM, and features are extracted containing spatial and temporal information (rich spatio-temporal features), transformed into spike trains fed into the DSNN. The CNN-LSTM layer extracts features from the EEG signal which can be implemented in the Employer SNN layers which are event-driven, converting temporal signals into spike trains. Additionally, a feature fusion mechanism adds attention that is utilized with both types of extracted features. Performance metrics were evaluated on EEG datasets, using the proposed DSNN model achieved highest accuracy (91.2%), improved F1-score (0.90) and reduced energy for inference (2.1mJ) versus other effective models. The analysis and training confirm that DSNN framework could allow for real-time, interpretable and energy efficient EEG based brain signal analysis, further work must be conducted to assess robustness of the model).
Multilayer Perceptron Neural Network-based Enhanced Big Data Analysis for Healthcare Data to Identify High-Risk Sensitive Data R.Anitha, R. Bagavathi Lakshmi Proceedings of 5th International Conference on Ubiquitous Computing and Intelligent Information Systems Icuis 2025, 2025 In today’s era, data compression is crucial for managing big data in the healthcare sector, reducing dimensionality and simplifying registration complexity. The role of artificial intelligence (AI) models in big data analysis in healthcare is crucial for identifying patients with high-risk, disease-impacting features that are essential for diagnosis. Traditional machine learning (ML) models typically do not analyse feature edges to minimise false positives. As dimensionality increases, accuracy decreases proportionally, affecting the accuracy of sensitive data recognition in feature analysis. To address this issue, we propose a Multilayer Perceptron Neural Network (MLP-NN) to identify risks in healthcare data. Initially, Z-Score Normalisation (ZSN) is used to preprocess the dataset. Moreover, the Support Vector Machine (SVM) technique is employed to select the most essential features. After that, an MLPNN is used to accurately classify maternal health risk levels into low, medium, and high. The proposed system enhances high-dimensional feature reduction by leveraging auxiliary resources to improve detection accuracy. The experimental results show that the classification accuracy is 96.4%, which is higher than that of other methods.
Lightweight TinyML-based Anomaly Detection for Intelligent Robotic Process Automation Workflows Essakki Gomathy P, Bagavathi Lakshmi R Proceedings of the 4th International Conference on Innovative Mechanisms for Industry Applications Icimia 2025, 2025 Robotic Process Automation (RPA) has immensely automated manual business processes in industry, finance, and administration. Nevertheless, the traditional RPA bots are still predominantly reactive, i.e., they cannot make intelligent decisions to identify anomalies or other atypical behavior in real time. This paper introduces a lightweight, embedded anomaly detection framework based on TinyML to provide RPA systems with a self-monitoring tier of intelligence. The suggested approach uses data on the structured event logs of the payment request processes, which are handled through feature engineering, outlier mitigation, and normalization of the full-scale process. A small 1D Convolutional Neural Network (CNN) was trained to classify anomalous patterns, achieving 94% accuracy and the .tflite model was deployed using Edge Impulse to be compatible with the microcontroller. The experimental evaluation effectively identified suspicious transactions based on time, amount, and pattern deviations. The system was tested and validated on real data from the BPI Challenge 2020 task. The study would be valuable in understanding scalable AI embedment at the specific stage of RPA to allow proactive anomaly detection to move beyond the intelligent automation systems of the previous generation.
Graph Neural Networks for Predicting Urban Traffic Congestion in Smart Cities B.Mahalakshmi, P N Shiammala, Gowtham M, R.Bagavathi Lakshmi, A.Angel Cerli, S. Sudha Proceedings of the 6th International Conference on Smart Electronics and Communication Icosec 2025, 2025 Urban traffic congestion is a serious issue for all smart cities in terms of sustainability and efficiency. Existing traffic prediction models do not properly account for the spatial and temporal dependencies in roads within the urban road network. This paper proposes a new methodology for modeling urban traffic congestion using a graph neural network (GNN) to design the road network as a defined dynamic graph, with the addition of multi-view traffic data (i.e. speed, volume, and occupancy), temporal sequences, and external contextual factors such as weather conditions and public events. Enhanced GNN architecture is utilized with attention and hierarchical graph pooling to learn local and universal traffic patterns. The model is validated and evaluated on a dataset collected over six months for a medium smart city, alongside three baseline models (LSTM, Xgboost and Static GCN). The proposed GNN model outperformed the three baseline models RMSE is 3.42 km/h MAE is 2.78, and -2 R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> score 0.915 with a F1 score of 0.89 in congested hours classification. The results of the proposed methodology confirm the suitability of spatial-temporal graph learning approaches for traffic forecasting methods. To conclude, this methodology demonstrates a strong scalable approach to computing predictive traffic models as part of critical smart city infrastructure.
Federated Learning for Early Detection of Diabetic Retinopathy in Distributed Healthcare Systems S. Sathya, R. Anandha Lakshmi, R.Bagavathi Lakshmi, Vishwa Priya V, D. Narayani, N. Anandakrishnan Proceedings of the 6th International Conference on Smart Electronics and Communication Icosec 2025, 2025 Diabetic retinopathy (DR) is one of the leading causes of preventable vision loss in the world, and early detection is critical for an effective intervention. However, the sensitive nature of patient data, together with regulatory considerations, restricts the use of centralized model training of healthcare data across institutions. This study presents a federated learning (FL) framework, to train a deep convolutional neural network(CNN) (EfficientNet-B0), to provide remote, early detection of DR type using retinal fundus images collected from multiple clinics that are distributed geographically, without sharing un-identified raw health data. The performance of the FL training-and-test scheme was evaluated on a range of AUC, accuracy, sensitivity, and specificity against a central, local only, ensemble, and pre-trained model. It is found that the FL model achieved an AUC of 0.91, and accuracy of 88.3%. Federated learning is chosen for this study to ensure patient privacy to handle distributed non-IID data, and achieve near-centralized diagnostic accuracy. This work also explores the future research opportunities for federated learning, and suggests that federated learning represents an advanced, scalable and privacy-respecting avenue for the implementation of AI-supported diagnostic imaging tools in the healthcare distributed healthcare systems.
Hyperspectral Images for Crop Prediction using Machine and Deep Learning Models R. Mahalakshmi, K. Kumutha, M. Sakthivanitha, R. Bagavathi Lakshmi, Mohamed Sirajudeen, Sankar Padmanabhan 3rd International Conference on Electronics and Renewable Systems Icears 2025 Proceedings, 2025 Crop classification is essential for local and national governments to make informed agricultural decisions. Remote sensing technology has made it possible to employ high-resolution hyperspectral images (HSIs) for land cover classification for decades. Machine Learning (ML) and Deep Learning (DL) are becoming more popular for HSI categorization due to advances in processing capacity and technology. HSIs can accurately identify crop types due to their narrow and continuous spectral band reflection. This study compares traditional machine learning (ML) models like Support Vector Machines (SVM) and Random Forest (RF) to deep learning (DL) models like Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) for hyperspectral image classification (HSI). The HSI categorization was evaluated using performance criteria such as accuracy, precision, recall, and the F1-Score. CNN achieved the highest accuracy of 96%, demonstrating that it is the most reliable solution for complex HSI classification tasks because it successfully captures spatial and spectral data.
A Study on Nature Inspired Algorithms for Load Balancing in Cloud Computing R. Bagavathi Lakshmi, R. Prema, C. Sathish Kumar, T. Thirumalaikumari, D. Narayani, M. Sakthivanitha 2025 IEEE 4th World Conference on Applied Intelligence and Computing Aic 2025, 2025 In a cloud computing (CC) setting, tasks are distributed across virtual machines (VMs) with varying lengths, start times, and execution times. Load balancing(LB) must be performed in such a way that all VMs are balanced in order to obtain optimal usage of their capabilities and increase system performance. The objective of the study is to explore and examine the Load balancing(LB) technique using natureinspired algorithms to optimally schedule all incoming tasks to the available VMs in order to reduce makespan (MS) and enhance machine usage in cloud computing. The explored nature inspired algorithms in this study includes Honey Bee Load Balancing(HLB), Particle Swarm Optimization(PSO), Rock Hyrax Optimization(RHO), Ant Colony Optimization(ACO), Birds Swarm Optimization (BSO), Mayfly Optimization (MFO), Crow Search Algorithm(CSA), Grey Wolf Algorithm(GWO), Lion Optimization Algorithm(LOA), Harris Hawks optimization (HHO) for LB in cloud computing. The scheduling technique is implemented using the CloudSim simulator. The simulation results clearly show that the nature-inspired techniques scheduling algorithms performance is found to be efficient in terms of lowering MS and energy consumption (EC).
