Research Publications
Permanent URI for this communityhttps://rda.sliit.lk/handle/123456789/4194
This main community comprises five sub-communities, each representing the academic contribution made by SLIIT-affiliated personnel.
Browse
8 results
Search Results
Publication Open Access Solar Hotspot Detection Using VHDL-Simulated Fixed-Point SVM: A Methodology Toward FPGA Realization(Faculty of Engineering, 2026-03) Fernando, N; Seneviratne, L; Weerasinghe, N; Rathnayake, N; Hoshino, YEarly detection of thermal hotspots in photovoltaic modules is critical to ensuring their efficiency, safety, and longevity. This study presents a complete end-to-end methodology for implementing a fixedpoint Medium Gaussian Support Vector Machine classifier using VHDL for a Field Programmable Logic Array. The approach begins with feature extraction from thermal images of healthy and defective solar panels, which focuses on MPEG-7 descriptors. The study shows that high impact for hotspot detection comes from blue chrominance contrast. A medium Gaussian SVM model is trained in MATLAB and converted to a fixed-point Q1.15 format for hardware compatibility. Key parameters, including support vectors, Lagrange multipliers, bias, and kernel scale, are extracted and verified in a custom Python environment to ensure numerical alignment with MATLAB results. The validated model is then implemented in synthesizable VHDL. It is verified using GHDL and the GNU Tool Kit waveform viewer, confirming bit-accurate hardware behaviour. Results show classification accuracy exceeding 99.3% with negligible performance loss due to quantization. The design achieves deterministic latency through an FSM-based structure and parallel feature processing for a 300-support vector and 222-feature system. This method enables low-power, real-time inference on a UAV-based edge platform, primarily focusing on drones.Publication Open Access Solar Hotspot Detection Using VHDL-Simulated Fixed-Point SVM: A Methodology Toward FPGA Realization Solar Hotspot Detection via FPGASVM(Faculty of Engineering, 2025-09-09) Fernando, N; Seneviratne, L; Weerasinghe, N; Rathnayake, N; Hoshino, YThe early and accurate detection of thermal hotspots in photovoltaic modules is critical to ensure the efficiency, safety, and longevity of solar power systems. This study presents a complete end-to-end methodology for implementing a fixed-point Medium Gaussian Support Vector Machine classifier using Very High-Speed Integrated Circuit - Hardware Description Language, optimized for Field Programmable Logic Array. The approach begins with feature extraction from thermal images, focusing on MPEG-7 descriptors and blue chrominance. The SVM model is trained in MATLAB and converted into a fixed-point Q1.15 format for hardware compatibility. Key parameters, including support vectors, Lagrange multipliers, bias, and kernel scale, are extracted and verified in a custom Python environment to ensure numerical alignment with MATLAB results. The validated model is then implemented in synthesizable VHDL and verified using GHDL and GNU Tool Kit waveform viewer, confirming bit-accurate hardware behavior. Results show classification accuracy exceeding 99.3% with negligible performance loss due to quantization. The design achieves deterministic latency based on FSM structure and parallel feature processing, completing classification within 2702 clock cycles for a 300-support-vector, 222-feature system. Unlike floating-point models, this approach enables low-power, real-time inference on edge platforms such as drones.Publication Open Access Efficient Hotspot Detection in Solar Panels via Computer Vision and Machine Learning(Multidisciplinary Digital Publishing Institute (MDPI), 2025-07-15) Fernando, N; Seneviratne, L; Weerasinghe, N; Rathnayake, N; Hoshino, YSolar power generation is rapidly emerging within renewable energy due to its cost-effectiveness and ease of deployment. However, improper inspection and maintenance lead to significant damage from unnoticed solar hotspots. Even with inspections, factors like shadows, dust, and shading cause localized heat, mimicking hotspot behavior. This study emphasizes interpretability and efficiency, identifying key predictive features through feature-level and What-if Analysis. It evaluates model training and inference times to assess effectiveness in resource-limited environments, aiming to balance accuracy, generalization, and efficiency. Using Unmanned Aerial Vehicle (UAV)-acquired thermal images from five datasets, the study compares five Machine Learning (ML) models and five Deep Learning (DL) models. Explainable AI (XAI) techniques guide the analysis, with a particular focus on MPEG (Moving Picture Experts Group)-7 features for hotspot discrimination, supported by statistical validation. Medium Gaussian SVM achieved the best trade-off, with 99.3% accuracy and 18 s inference time. Feature analysis revealed blue chrominance as a strong early indicator of hotspot detection. Statistical validation across datasets confirmed the discriminative strength of MPEG-7 features. This study revisits the assumption that DL models are inherently superior, presenting an interpretable alternative for hotspot detection; highlighting the potential impact of domain mismatch. Model-level insight shows that both absolute and relative temperature variations are important in solar panel inspections. The relative decrease in “blueness” provides a crucial early indication of faults, especially in low-contrast thermal images where distinguishing normal warm areas from actual hotspot is difficult. Feature-level insight highlights how subtle changes in color composition, particularly reductions in blue components, serve as early indicators of developing anomalies.Publication Embargo Steerable Directional Listening for Individuals with Unilateral Hearing Loss(SLIIT, Faculty of Engineering, 2024-10) Wimalaratne, U.A; Seneviratne, L; Malasinghe, LA novel hearing aid design which incorporates 3 separate phased arrays with a digital signal processor running an adaptive beamforming algorithm, Generalized Sidelobe Canceller (GSC), providing the user with the capability of effectively focusing their listening to a certain source while suppressing any masking interference sources. The proposed phased array achieved an array directivity of 7.93dB with a half power bandwidth of 35.20º. The GSC and the phased array designed, when simulated, was able to achieve a SINR improvement of 9.72dB under strong noise levels and a 2.55dB SINR improvement under low level noise which were located spatially close to the desired source.Publication Embargo Auto-encoder Based Data Clustering for Typical and Atypical Repetitive Child Hand Movement Pattern Identification(SLIIT, Faculty of Engineering, 2024-10) Wedasingha, N; Samarasinhe, P; Seneviratne, L; Papandrea, M; Puiatti, AThis study is dedicated to the important task of identifying unique repetitive hand movement patterns in children, with the aim of facilitating early anomaly detection. The current body of literature lacks a comprehensive model capable of effectively discerning distinctive patterns in child repetitive hand movements. To address this gap, our innovative approach employs autoencoders to efficiently compress intricate data and extract latent features from a dataset with inherent limitations. By utilizing clustering techniques, we analyze these features to reveal distinct behaviors associated with child hand movements. Despite the challenges posed by binary annotated datasets, our model demonstrates outstanding performance in categorizing movements into four distinct types, thereby providing valuable insights into the intricate landscape of child hand movement patterns. Statistical assessments further underscore the superiority of our autoencoder, achieving a mean Bayesian value of 0.112, outperforming state-of-the-art algorithms in this domain. Subsequent in-depth analysis exposes notable inter-cluster patterns, elucidating transitions from typical to atypical behavior in child hand movements. This research constitutes a significant advancement in the field of child hand movement pattern analysis, offering a powerful and sophisticated tool for healthcare professionals and researchers alike. The automation capabilities embedded in our model empower these professionals to address childhood behavioral disorders more effectively and efficiently. In essence, our research not only contributes to the enhancement of early anomaly detection systems but also serves as a valuable resource for professionals engaged in child healthcare and behavioral research, facilitating a deeper understanding of these nuanced patterns.Publication Embargo Child Head Gesture Classification through Transformers(Institute of Electrical and Electronics Engineers Inc., 2022-11-04) Wedasingha, N; Samarasinghe, P; Singarathnam, D; Papandrea, M; Puiatti, A; Seneviratne, LThis paper proposes a transformer network for head pose classification (HPC) which outperforms the existing SoA for HPC. This robust model is then extended to overcome the limited child data challenge by applying transfer learning resulting in an accuracy of 95.34% for child HPC in the wild.Publication Open Access Support Vector Machine Based an Efficient and Accurate Seasonal Weather Forecasting Approach with Minimal Data Quantities(SLIIT, 2022-02-11) Chandrasekara, S; Tennekoon, S; Abhayasinghe, N; Seneviratne, LClimate change makes a big impact in our daily activities. Therefore, forecasting climate changes prior to its actual occurrences is important. Even though highly accurate weather prediction systems throughout the world are available, they require mass amounts of data exceeding thousands of data points to obtain a significant accuracy. This study was aimed at proposing a Support Vector Machine based approach to carryout seasonal weather predictions up to thirty-minute intervals, the results of which would be considerably effective with respect to predictions carried out with models trained with annual datasets. The model was trained utilizing a dataset corresponding to the district of Kandy which consisted of 136 samples, 20 features, and 5 labels. By means of carrying out numerous data preprocessing steps, the model was trained, and the relevant hyperparameters were optimized considering the grid search algorithm to yield a maximum accuracy of 86%, once tested via the k-fold cross validation. The performance of the Support Vector Machine was also then compared for the same dataset with that of the K-Nearest Neighbor algorithm which consumed relatively fewer computing resources. An optimal accuracy of 61% was observed for this model for a K-value of 27. This approach supported the concept of a Support Vector Machine’s ability to perceive time series forecasts to a relatively higher degree and its ability to perform effectively in higher dimensional datasets with smaller number of samples. As per the future work, the Receiver Operating Characteristic analysis is proposed to be carried out to evaluate the performance of the model and the dataset size is proposed to be further enhanced to a maximum of a thousand samples to yield the best performance results.Publication Embargo Skeleton Based Periodicity Analysis of Repetitive Actions(IEEE, 2022-04-07) Wedasingha, N; Samarasinghe, P; Seneviratne, L; Puiatti, A; Papandrea, M; Dhanayaka, DThis paper investigates the problem of detecting and recognizing repetitive actions performed by a human. Repetitive action analysis play a major role in detecting many behavioral disorders. In this work, we present a robust framework for detecting and recognizing repetitive actions performed by a human subject based on periodic and aperiodic action analysis. Our framework uses focal joints in the human skeleton for the analysis of repetitive actions which are substantiated by the principles of human anatomy and physiology. Using Non-deterministic Finite Automata (NFA) techniques, in this paper, we introduce a novel model to transform repetitive action count to differentiate the periodicity in human action. Experimental results on a dataset consisting of 371 video clips show that our algorithm outperforms the state-of-art (RepNet) [1] in simultaneous multiple repetitive action counts. Further, while the proposed model and RepNet give comparable results in counting periodic repetitive actions, our model performance surpass RepNet significantly on analysing non-periodic repetitive behavior.