Institutional-Repository, University of Moratuwa
Welcome to the University of Moratuwa Digital Repository, which houses postgraduate theses and dissertations, research articles presented at conferences by faculties and departments, university-published journal articles and research publications authored by academic staff. This online repository stores, preserves and distributes the University's scholarly work. This service allows University members to share their research with a larger audience.
 Research Publications |  Thesis & Dissertation |  E- Books |
 UoM Journal Publications |  UoM Conference Proceedings |  Articles published in Scimago's Q1 journals |  UoM Research Reports |  Other Articles authored by UoM staff |
Recent Submissions
item: Thesis-Abstract
Fraud detection in financial transactions using LSTM and XAI with ontological validation
(2025) Rambukpitiya, SN; Silva, ATP
The escalation of digital financial services has increased the occurrence of credit card fraud, demanding the development of advanced, trustworthy fraud detection systems. Traditional and simple rule-based and classical machine learning techniques, while useful, often fail to detect complex, evolving fraud patterns and lack sufficient interpretability for high-stakes financial decision-making. This research proposes a novel, integrated approach combining Long Short- Term Memory networks, Explainable AI techniques, and ontology-based semantic validation to address these challenges. The Long Short-Term Memory model captures sequential transaction behaviors effectively, identifying anomalies indicative of fraud. To overcome the inherent black- box nature of deep learning models, Local Interpretable Model-agnostic Explanations is employed, providing transaction-level interpretability and enabling stakeholders to understand the factors behind each prediction. Further, an ontology is developed to embed domain-specific knowledge, offering semantic validation of the model outputs. This ensures that predictions not only rely on learned patterns but also align with predefined risk rules and expert knowledge. Experimental results demonstrate high accuracy and recall rates, confirming the model’s effectiveness in detecting fraudulent activities, while the ontology layer enhances trust and reliability. This hybrid framework thus advances fraud detection by combining predictive power, explainability, and semantic validation, contributing to the development of more secure and transparent financial systems.
item: Thesis-Abstract
Enhancing arithmetic optimization algorithm for robot path planning in dynamic environments
(2025) Madushani, GHC; Piyatilake, T
Autonomous robot path planning in dynamic environments poses significant challenges for many industries: military, industrial automation, and security, to name a few. The goal is to fairly and safely navigate both unpredictable and complex spaces as efficiently as possible, however, traditional planning algorithms lack the speed and versatility needed to react to a dynamic landscape. For this reason, there is a need for new solutions which enhance efficiency of the algorithm in complex environment with aid of AI techniques.
This thesis contributes to minimizing the above challenges by presenting the Hybrid Arithmetic Optimization Algorithm (HAOA) which integrates adaptive arithmetic‐ operator updates, online reinforcement learning, and spline‐based path refinement. HAOA works by creating far and near candidate trajectories and subjecting them to multiplicative and additive arithmetic transformations before assessing progress. This initial updating process allows the algorithm to transition from a global exploration state back to a local exploitation state. A Q-learning agent observes the incremental changes along the candidate trajectories to begin to update key parameters, which allows the algorithm to adapt to the dynamic environment. Finally, the algorithm performs a final refinement of its solution by executing self-intersection removal, waypoints pruning, and B-spline smoothing so it can deliver continuous and collision free paths for a real robot to follow its path with minimal acceleration discontinuities. Extensive experiments on four classical benchmark functions (Sphere, Rosenbrock, Rastrigin, Ackley) and five grid-based scenarios featuring both static and dynamic obstacles demonstrate HAOA’s superior performance. Compared to standard AOA and a Genetic Algorithm baseline, HAOA converges 15–25 % faster, attains lower final objective values, and yields paths that are on average 10–12 % shorter and substantially smoother. Although embedding reinforcement learning introduces moderate per-iteration overhead, the overall wall-clock time to reach target performance is reduced due to accelerated convergence. These results underscore HAOA’s promise as a robust, adaptable framework for real-time autonomous navigation in complex and unpredictable settings.
item: Conference-Full-text
A Theoretical model for thrombin generation kinetics in the initial phase of the extrinsic pathway of human blood coagulation
(IEEE, 2025) Weerasinghe, IU; Amarasinghe, S; Attygalle, D; Samarasekara, B; Weragoda, SC
This approach presents a nonlinear theoretical model to predict the thrombin concentration in the initial phase kinetics of the extrinsic pathway of human blood coagulation. The model integrates key enzymatic reactions and the positive feedback mechanism of the extrinsic pathway to simulate thrombin generation and factor activation dynamics. The presented model accurately predicts thrombin concentration over time under the assumptions based on the initial conditions. However, upon fibrinolysis activation, deviations arise due to fluctuating concentrations of plasminogen activators and inhibitors. These findings underscore the model's utility in understanding hemostasis regulation and highlight the need for further refinement to incorporate fibrinolysis dynamics for comprehensive insights for continuous modeling.
item: Conference-Full-text
Design-aware optimization of 2D skeletal structures using conditional generative adversarial networks (cGANs)
(IEEE, 2025) Fernando, C; Herath, S
This research presents a novel diffusion-based cGAN framework for the automated optimization of 2D skeletal structures with Eurocode 3-compliant steel Circular Hollow Section (CHS) assignments. Traditional methods like SIMP are computationally intensive and require specialist knowledge, limiting their practical use. To overcome these challenges, the study develops a hybrid deep learning approach that combines the adversarial learning of cGANs with the progressive refinement of diffusion models, enabling efficient and codecompliant structural design generation. The methodology involves a three-phase process: first, generating a comprehensive dataset through topology optimization, skeletonization, frame extraction, and code-compliant section assignment; second, training an advanced diffusion-based cGAN that conditions on user-defined parameters and iteratively denoises outputs to produce high-fidelity 512×512 structural layouts; finally, identification of sections assigned in the structural layout. Experimental results show that the proposed framework is approximately 50,000 times faster than traditional FEA-based methods and achieves perfect code compliance and higher prediction accuracy than existing deep learning models. This work significantly advances both artificial intelligence and structural engineering by automating highquality, regulation-ready design generation.
item: Thesis-Abstract
Weakly supervised learning for fine-grained plant disease classification : a hybrid approach
(2025) Praveenath, L; Silva, ATP
Agricultural productivity is significantly impacted by plant diseases, resulting in economic losses and risks to food safety. The accurate detection of plant diseases with minimal supervision is considered crucial for the resolution of these issues, rather than having sole reliance placed on traditional methods such as manual inspections. Promising solutions for the automation of disease classification have been presented by recent advancements in computer vision, particularly those achieved through deep learning techniques. However, the requirement of large, annotated datasets is often associated with these methods, and such datasets can be both expensive and labour- intensive to obtain. This dissertation aims to design and evaluate a hybrid deep learning model capable of achieving high fine-grained plant disease classification accuracy under weak supervision with limited labelling effort. In order to increase the accuracy of fine-grained plan disease classification problems under weak supervision, a novel hybrid architecture combining ResNet50 and Vision Transformer is described. A series of experiments was executed utilizing diverse parallel hybrid configurations, and their performance was systematically compared against that of the standalone ResNet50 and Vision Transformer models. Distinct feature fusion strategies, including addition, concatenation, and weighted combinations, were employed by each model to combine the intermediate feature maps. The models were trained and assessed using the Crop Pest and Disease Detection (CCMT) dataset, and accuracies of 84.53%, 84.56%, and 86.26% were yielded, respectively. These performance metrics were exceeded in comparison to the individual models, with an accuracy of 63.56% achieved by ResNet50 and 81.88% attained by the Vision Transformer. The efficacy of the proposed hybrid architecture for fine-grained classification in scenarios characterized by minimal supervision is underscored by the results. This study provides a practical pathway toward scalable, cost-effective plant disease detection systems for real-world agricultural deployments. The intrinsic limits of current classification frameworks are successfully addressed, and this research significantly advances the field of fine-grained classification under limited supervision.