Bannari Amman Institute of Technology

Globally, the integrity of online examinations is a paramount concern with the widespread adoption of remote learning and digital assessments. Ensuring fairness and preventing academic misconduct presents a significant challenge for educational institutions, training centers, and certification bodies. Traditional proctoring methods are often resource-intensive or limited in their detection capabilities, leading to potentially inaccurate or insufficient monitoring. This paper introduces Proctonix AI, an intelligent online exam monitoring system designed to enhance academic integrity through real-time, AI-powered surveillance. Leveraging modern web technologies and computer vision techniques, the system captures and processes live webcam feeds and screen activity to detect suspicious behaviors. Key functionalities include face and presence detection, gaze tracking for distraction assessment, and voice activity monitoring. By employing Flask for backend web server and route handling, HTML/CSS/JavaScript for a responsive user interface, and OpenCV with Media Pipe Face Mesh for robust facial and eye landmark detection, Proctonix AI aims to provide a robust framework for identifying potential cheating instances. The system also integrates custom heuristics for behavioral analysis and browser APIs for user alerts, facilitating a more secure and reliable online examination environment. The objective is to offer a streamlined, accessible solution that supports educators in maintaining the credibility of online assessments.

Increasing levels of psychological stress among higher education faculty represents a significant challenge, frequently resulting in burnout, poor performance, and worsening of mental health. Conventional evaluation methods inadequately capture real-time, contextual indicators of stress. This study presents a multilayer deep learning architecture based on Bi-LSTM, combined with an attention mechanism, for precise stress prediction utilising multimodal physiological, behavioural, and environmental data. The model analyses sequential inputs, including heart rate, activity patterns, and environmental data, to identify temporal correlations and significant stress indicators. Experimental analyses on benchmark datasets demonstrate that the proposed framework overcomes baseline models, including SVM, Random Forest, and conventional LSTM. It attained an accuracy of 94.27%, precision of 93.88%, recall of 94.10%, F1-score of 94.00%, and an AUC-ROC of 0.972. These findings highlight the model's capacity to identify psychological stress with significant reliability and resilience. The technology facilitates early intervention strategies for teacher well-being and has scalable implementation potential in smart campuses and healthcare applications. The use of attention mechanisms greatly enhances the model's interpretability, differentiating this study from current solutions. This research enhances emotional computing and mental health analytics by offering a data-driven, real-time, and intelligent framework for stress monitoring.

Abstract Four binary mixtures are isolated from two hydrogen bond mesogenic complexes that are formed between p-n alkyloxy benzoic acids and 4-Nitrobenzaldehyde. The carbon chain length chosen for the formation of binary mixtures are heptyloxy and octyloxy benzoic acids. 4-Nitrobenzaldehyde (NBA) and heptyloxy benzoic acid mesogenic complex is treated as X component while 4-Nitrobenzaldehyde (NBA) and octyloxy benzoic acid mesogenic complex is treated as Y component. Binary mixtures are obtained by varying the molar ratio of the two mesogenic complexes ( X and Y ) in even steps viz., from 0.2 to 0.8. Hydrogen bond existence between the binary mixtures is confirmed through Fourier Transform Infrared Spectroscopy. Mesogenic phase variance is investigated through polarizing optical microscope study. The phase transition temperature and their corresponding enthalpy values are analyzed via differential scanning calorimetry. Thermal equilibrium, thermal stability, order of phase transition and specific heat capacity of the binary mixtures are also elucidated from the thermograms of the mesogenic mixtures. The results obtained make these binary mixtures a potential candidate for opto-electronic and thermal energy based applications.

This paper presents an optimized wind energy conversion system for increased energy output through the use of Graph Neural Networks (GNNs) and machine learning. In order to increase efficiency, the system incorporates a SEPIC converter and keeps track of torque, acceleration, and RPM for the turbine. The GNN model uses historical and current wind speed data to dynamically modify turbine operation in order to reduce losses. Reliability and consistent performance are guaranteed by automated alerts and remote monitoring for problems like dust buildup. This method establishes a standard for sustainable wind energy systems and greatly increases energy utilization. The proposed integration also enhances accuracy in energy forecasting and significantly reduces conversion losses. This is achieved by coupling advanced learning capabilities of GNNs with the stabilization features of SEPIC, resulting in a more reliable and efficient energy conversion system. The integration of GNNs and SEPIC converters enhances wind turbine energy efficiency by dynamically optimizing performance and reducing losses.

Vitiligo is a chronic skin condition characterized by depigmentation patches due to melanocyte loss, often leading to psychological and social distress. Early and accurate detection is essential for timely intervention and treatment planning. This paper proposes a novel deep learning model for automated Vitiligo detection, integrating Convolutional Neural Networks (CNNs) with Transformer layers and attention mechanisms. The model employs comprehensive preprocessing techniques and data augmentation strategies to enhance generalization and robustness. A spatial attention module is used to highlight lesion-specific regions, while stacked Transformer blocks effectively capture both local and global contextual features from dermatoscopic images. Experimental evaluation demonstrates the superiority of the proposed framework, achieving an accuracy of 90.5 %, sensitivity of 91.8 %, specificity of 89.2 %, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$F 1$</tex>-score of 90.3 % on the curated dataset. A comparative study with baseline models reveals that each component-especially the attention mechanism and transformer layers-significantly improves diagnostic performance. Furthermore, ablation studies confirm the importance of preprocessing, data augmentation, and transformer depth in boosting classification accuracy. The model also exhibits high class-wise Area Under Curve (AUC) values for Vitiligo and Non-Vitiligo categories in ROC analysis, validating its discriminative power. These findings suggest that the proposed system offers a reliable, automated tool for earlystage Vitiligo detection and can potentially support dermatological assessments in both clinical and telemedicine environments.

The growing interest in five phase induction motors can be attributed to their diverse range of applications, which encompass electric vehicles, advanced compressor systems, and marine propulsion, among others. A proper control methodology must be implemented to operate drive at its highest performance level. Among many, Direct Torque Control (DTC) methodology is developed for various tasks. However DTC with conventional PI/PID controllers are unfit to exhibits its best performance under critical conditions. Hence, Three Degree of Freedom Proportional-Integral-Derivative (3DOF-PID) controller is utilized in the proposed DTC methodology. Additionally, 3DOF-PID controllers are capable of delivering a swift response to rapid input variations. Consequently, this paper proposes a 3DOF-PID controller based DTC for a five phase induction motor. Additionally, the space vector pulse width modulation (SVPWM) technique is employed to generate the appropriate pulse sequence for a five leg inverter used in motor drive applications. A detailed design of 3DOF-PID controller is also incorporated in this paper. In addition to this, Whale Inspired Algorithm (WIA) is developed to tune various parameters of 3DOF-PID controller. In order to make speed sensorless controller, mathematical modeling is included to estimate speed of the motor to avoid physical sensor. Hardware - in the - Loop (HIL) technique is implemented on OPAL-RT platform to present various results similar to real time responses.

This paper introduces a new 60 Flat PWM Technique for Single-Source Switched-Capacitor(SC) Multilevel Inverter with Quadruple Boost (60FP-MLI-QB) and Enhanced Output. The proposed design generates nine output levels with lower switching losses using the 60 Flat PWM (60FP) Technique. Furthermore, this technique increases the voltage gain enhancement while flattening the sine wave at its peak. The driven inverter topology includes 11 switches, two capacitors, and a diode, with each levels circuit connected through six or five switches. The performance of the proposed 60FP-MLI-QB is tested through simulation and hardware experiments under various dynamic conditions, showing its flexibility and superiority compared to similar configurations. Moreover, the paper presents a harmonic minimization technique through various FP methods by modifying the sinusoidal modulation reference.

Healthcare professionals experience severe psychological strain from difficult schedules and increased patient volumes, resulting in burnout and minimised efficiency. Conventional stress monitoring methods lack real-time predictive functionality and do not involve structured domain knowledge. This study presents a Hybrid Knowledge Graph-Deep Learning Model for predictive analytics in stress management, utilising graph embeddings and deep learning frameworks to improve interpretability. The solution integrates cloud-based deployment, facilitating real-time inference and scalability. The suggested system attains 96.4% accuracy, 95.8% F1-score, and a mean squared error of 0.021, more than the performance of Random Forest, SVM, and CNN models. The introduction of Knowledge Graph embeddings enhances contextual awareness, whereas deep learning guarantees adaptability across various healthcare settings. The results indicate improved accuracy in stress prediction, real-time monitoring capabilities, and greater explainability, positioning this method as an important boost for proactive stress management strategy in healthcare environments.

Newly developed fifth- and sixth-generation (5G and 6G) networks are very secure, reliable, less latent, and versatile. These networks provide various services, like Massive Machine Type Communication (mMTC), Ultra-Reliable Low Latency Communication (URLLC), and Enhance Mobile Broadband (eMBB). Network slicing (NS) is important in both 5G and 6G networks as it enables the provision of the above services in a single physical network. This paper proposes an Optimal Hybrid Quantum-Classical Convolutional Neural Network for efficient 5G Network Slicing (HQCNN-5GNS). Initially, the input data are gathered from Unicauca IP flow version 2 and 5G network slicing dataset. The collected data are fed to the pre-processing phase using a Pseudo-linear Maximum Correntropy Kalman Filter (PMCKF) to remove duplicate data and replace the missing values from the dataset. The pre-processing information is provided to the Hybrid Quantum-Classical Convolutional Neural Network (HQCNN) to classify the network slicing into mMTC, eMBB, and URLLC. In general, HQCNN does not adopt any optimization methods to determine the optimal parameters to guarantee accurate categorization. Therefore, the Quantum Beetle Antennae Search Algorithm (QBASOA) is employed to optimize the weight parameter of HQCNN. The experimental outcomes show that the proposed HQCNN-5GNS attains 17.25%, 15.65%, and 19.55% higher accuracy than the existing approaches: Harris Hawks optimization depends on hybrid deep learning for effectual network slicing on 5G network (HHO-HDL-5GNS), Highly accurate with dependable wireless network slicing in the 5G network: a hybrid deep learning model (WNS-HDL-5GNS), Deep reinforcement learning-basis network slicing for beyond 5G (DRL-5GNS), respectively.

In the modern era, decision support systems for clinical application plays a vital role in assisting clinicians by providing data-driven medical recommendations. Even though previous researchers suggest various advanced model, but still face challenges because of the inability to model sequential health data and missing or noisy data. To resolve these issues, Bidirectional Long Short-Term Memory with Fuzzy Inference System (Bi-LSTM-FIS) is proposed. Initially the data is collected from Electronic Health Records (EHRs) of a medical college which contains various patient records with structured attributes such as Blood Pressure (BP), cholesterol, hemoglobin, thyroid function, etc. Further, the collected data is preprocessed with K-Nearest Neighbor (KNN) imputation and min-max scaling for handling missing values and normalization respectively. Then, fuzzy set is created for the health parameters to handle uncertainty by numerical values into linguistic categories. Furthermore, Bi-LSTM is employed to capture long-term dependencies in patient histories, whereas FIS refines predictions using medical expertise. Hence, by combining Deep Learning (DL) with fuzzy logic, the proposed Bi-LSTM-FIS acquired great results with interpretable, and scalable solution for healthcare decision support.