http://drr.vau.ac.lk/handle/123456789/254 Wed, 12 Aug 2026 22:11:57 GMT 2026-08-12T22:11:57Z http://drr.vau.ac.lk/handle/123456789/2177 An Ensemble Method for Fatigue Detection based on Multiple Facial Expressions Venuja, N.; Nadarajah, S.; Thavayoganathan, A.; Senthooran, V. People who spend a lot of time staring at screens need to be aware of their own level of fatigue because it can lower productivity and increase the risk of accidents. In this research work, we suggest a novel method for detecting fatigue based on a variety of facial expressions. We collected 12,185 facial images of people using digital screens, which contains 6108 normal face images and 6077 stress face images. In this framework, we used different multiple facial expressions techniques and Artificial Neural Network (ANN) to label neutral faces as normal, while sad and disgusted faces were labeled as stressed. We used this dataset to train a deep learning model to classify facial images as normal or stressful. Our model achieved highest accuracy rates of 88% for testing and showed good performance on a trained dataset. We also validated the accuracy of our approach by comparing the predicted stress/non-stress labels with ground truth labels obtained through self-report. The results of this study demonstrate the potential of using multiple facial expressions for fatigue detection in people who are using digital screens. This approach could be used in the future to develop real-time fatigue detection systems that could alert people when they are showing signs of stress or fatigue and help them take a break or adjust their work schedule. Sun, 01 Jan 2023 00:00:00 GMT http://drr.vau.ac.lk/handle/123456789/2177 2023-01-01T00:00:00Z http://drr.vau.ac.lk/handle/123456789/2176 Banana Classification using Deep Learning Techniques in the Sri Lankan Context Kumara, D.; Sonali, H.G.M.; Lakmini, W.; Lakshan, R.M.M.; Mayumika, P.; Athukorala, A.; Fernando, W.S.S.D.; Nadarajah, S.; Suthaharan, S. S. Bananas are one of the most widely consumed fruits worldwide, and they come in various shapes, sizes, and colors. Bananas are one of the few tropical crops that have not been bred successfully, and all presently cultivated varieties are natural selections. Traditionally, human experts perform the classification based on visual inspection, which can be subjective and time-consuming. Therefore, developing an automated system that can accurately classify bananas based on visual features would be highly beneficial. Feature engineering became easier after the Convolutional Neural Network (CNN) was developed. Five distinct varieties of bananas were categorized in this proposed work using the CNN model. To improve classification performance, many training images are needed when using the CNN model for classification. Both the original and enhanced images were used to train and evaluate the suggested CNN model. During training, the CNN model achieved an overall validation accuracy of 87%. Wed, 01 Jan 2025 00:00:00 GMT http://drr.vau.ac.lk/handle/123456789/2176 2025-01-01T00:00:00Z http://drr.vau.ac.lk/handle/123456789/2170 A Generative Adversarial Network and Feed-Forward Neural Network Approach to Predict Health of Electric Vehicle Lithium-Ion Batteries in Extreme Temperature Pirunthavi, W.; Ray, B.; Jahan, H.; Emami, K.; Vithusha, B. he growing adoption of electric vehicles (EVs) has intensified the need for better battery management systems (BMS) to ensure the longevity, efficiency, and safety of lithium-ion batteries (LiBs). Temperature fluctuations have significant effects on the State of Health (SOH) of LiBs, but real-world datasets that encompass diverse and extreme thermal circumstances are limited, resulting making in precise SOH prediction an ongoing challenge. This study presents a novel predictive modeling approach that combines Generative Adversarial Networks (GANs) with Feedforward Neural Networks (FFNNs). The GAN generates realistic synthetic battery data over wide temperature ranges, followed by post-processing techniques to correlate synthetic results with actual battery behavior. This combined dataset, comprising both real and synthetic data, is subsequently used to train an FFNN model for accurate SOH prediction. The key contributions are: 1) development of a GAN-based data augmentation pipeline to address data scarcity under extreme temperatures, 2) integrating synthetic and real data using some post processing techniques for improved model reliability and generalization, and 3) implementation of an FFNN-based SOH predictor that attains high predictive precision with minimal computational cost suitable for resource-constrained BMS. The proposed GAN–FFNN model attains an R2 of 0.94 and an MAE of 0.29 between −30° C and 60° C, demonstrating superior performance and revealing that high temperatures (>40° C) accelerate degradation more than low temperatures. This is the first known GAN-augmented framework for SOH prediction across such an extensive temperature range. Wed, 01 Jan 2025 00:00:00 GMT http://drr.vau.ac.lk/handle/123456789/2170 2025-01-01T00:00:00Z http://drr.vau.ac.lk/handle/123456789/2169 Posture flow: A hybrid deep learning and OpenPose framework for yoga posture tracking with Ayurvedic alignment feedback Nandasena, P.G.M.K.; Weerakkodi, Y.S.; Doshanthan, K.; Perera, K.D.N.N.; Perera, P.E.I.; Srimal, S.J.M.N.S.; Vithusha, B.; Venuja, N. Yoga provides an encompassing way to well-being, promoting body strength, mental health, emotional balance, and spiritual belief. However, traditional video tutorials or group classes often fail to provide individual monitoring to address individual differences in anatomy, move ment patterns and guidance and cause poor physical health, and it leads to minor to major in juries, health issues. This study aims to develop a Posture Flow, a deep learning enhanced yoga posture tracking and evaluation system that efficiently empowers accurate practice of yoga. The proposed system fuses Traditional Ayurvedic principles with deep learning to develop an efficient system that supports both modern technology and the ancient wellness wisdom as they provide posture alignment, timing and control of the body. For that, the study employs the hybrid framework that combines VGG19 CNN, a pretrained model for extracting high-level features, and OpenPose is used to accurately detect the key points in the body. The system provides real-time feedback, integrating Ayurvedic principles related to posture alignment, timing, and body control. As for pose validation, the decision is made to verify the poses based on the predefined rules derived from the Ayurvedic insights. The obtained result is assessed against the set of Ayurvedic principles of posture alignment for a particular pose. The alignment rules focus on the crucial things such as symmetry of the body, spinal stage or alignment, position of limbs, and flow of energy through the body. The model’s performance was evaluated on multiple datasets, achieving a classification accuracy of 94.57% alongside strong precision, recall, and F1 scores. The average response time per pose was measured at 0.8 seconds, demonstrating suitability for live and interactive applications. The proposed hybrid framework effectively combines modern machine vision techniques with ancient Ayurvedic wisdom or principles to provide accurate posture classification and verifications of body alignment. The real-time feed back mechanism of the system tends to provide precise yoga practice that helps to reduce wrong postures and enhance safety. The obtained highest accuracy and responsiveness make the system suitable for integrating it into mobile apps, web apps, we rable fitness devices, and virtuallive practice sessions. This study offers a scalable solution for personalised yoga practice by enabling all skill level users to get instant, individualized guidance. Wed, 01 Jan 2025 00:00:00 GMT http://drr.vau.ac.lk/handle/123456789/2169 2025-01-01T00:00:00Z