Institutional Repository

U n i v e r s i t y of P e r a d e n i y a

The Institutional Repository of the University of Peradeniya is the University's digital gateway to scholarly knowledge and research excellence. Established as the next phase of the Digital Library initiative launched by the University Library in 2011, it preserves, showcases, and provides open access to the intellectual output of the University community.

Through a rich collection of research articles, conference proceedings, theses, dissertations, and other scholarly works, the repository enhances the visibility, accessibility, and long-term preservation of knowledge, extending the global reach and impact of the University of Peradeniya’s academic and research contributions.

Photo by @ITD,Main Library

Communities in DSpace

Select a community to browse its collections.

Now showing 1 - 5 of 5

Recent Submissions

Nonlinear stability and structural resilience of interlocking jetty systems for bridge support
(University of Peradeniya, Sri Lanka, 2026-08-20) Gallehewa, D. H.; Dissanayake, D. M. A. G. B.; Jayasinghe, J. A. S. C.
In the development of modular bridge infrastructure, the structural resilience of interlocking jetty systems heavily depends on robust connection strategies. Ensuring efficient load transfer between the substructure and superstructure is critical, particularly at complex beam-to-pile interfaces utilizing a Doweled Collar Interlock (DCI), and at beam-to-beam connections secured by bolted steel splice plates. This research aims to perform a comprehensive forensic structural appraisal of these hybrid assemblies to identify vulnerabilities leading to localized shear failure. The system was modeled with explicit full-scale structural dimensions, featuring a 3m long, 300mm wide pier cap beam with a depth tapering dynamically from 250 mm to 900mm, consisting a 400mm pile cap that interfaces with a 300mm diameter foundation pile. The study utilized Abaqus finite element analysis to simulate the complex contact mechanics of the DCI joints and splice plate connections, employing a surface-to-surface contact formulation with penalty friction behavior. The numerical model was validated against experimental load-testing data of a single-beam prototype, showing strong correlation in both initial elastic stiffness and ultimate failure modes. Two numerical approaches were evaluated: the standard General Static solver and the Modified Riks (arc-length) method. The computational models assessed geometric instabilities, material softening, and sudden contact chatter inherent to the joint interfaces. Numerical simulations revealed that the General Static solver frequently failed to converge due to unresolvable contact singularities. Conversely, the Modified Riks method successfully traced the unstable post-buckling "snap-through" equilibrium path. Analysis pinpointed critical stress concentrations at the dowel interfaces, collar roots, and bolt-hole boundaries. Quantitative results demonstrated that modifying the joint geometry increased peak shear capacity from 44.1 kN to 62.9 kN. Additionally, introducing strategic reinforcement reduced residual displacement by up to 80% and increased residual shear force by 33%, effectively preventing localized brittle concrete tearing. Ultimately, the study concludes that pairing optimized reinforcement strategies with the Modified Riks model is essential for accurately quantifying ultimate stability, providing engineers with a superior predictive framework for designing robust modular infrastructure under extreme loading conditions.
Multi-disease detection in tomato leaves using deep learning with explainable AI (XAI) techniques
(University of Peradeniya, Sri Lanka, 2026-08-20) Thiviyaraj, A.; Dissanayaka, D. M. A. P. S.; Vinoharan, V.
Tomato leaf diseases substantially reduce crop yield and quality, yet manual identification remains slow, expertise-dependent, and error-prone, motivating the need for automated, accurate, and interpretable detection systems that support timely diagnosis and informed crop management. This study proposes a deep learning framework for classifying tomato leaf diseases that combines high predictive accuracy with model transparency through Explainable Artificial Intelligence (XAI), addressing the practical need for systems that agronomists can trust and verify. A publicly available Kaggle dataset of 16,012 images spanning ten classes, nine diseases (bacterial spot, early blight, late blight, leaf mold, septoria leaf spot, two-spotted spider mite, target spot, tomato mosaic virus, and tomato yellow leaf curl virus) and one healthy class, was split into 80% training and 20% testing sets, with 20% of the training set held out for validation. Two transfer learning models, EfficientNetB0 and MobileNetV2, were trained for 30 epochs using the Adam optimizer with early stopping, following image resizing, normalization, and data augmentation techniques including rotation, zooming, and horizontal flipping. Grad-CAM and LIME were applied to visualize the leaf regions influencing each prediction. EfficientNetB0 achieved 96.77% training accuracy and 95.18% testing accuracy, with a macro-average F1-score of 0.944 and a weighted-average F1-score of 0.952 across the ten classes, confirming balanced performance despite class imbalance. MobileNetV2 achieved 88.94% training accuracy and 88.99% testing accuracy. Grad-CAM and LIME visualizations consistently localized attention on lesion-affected leaf tissue rather than background regions, indicating that both models based their decisions on biologically relevant disease features rather than spurious cues. These results demonstrate that integrating a high-accuracy transfer learning model with XAI produces a classification system that is both reliable and interpretable, offering a practical and trustworthy tool for early disease diagnosis and decision-making in precision agriculture.
Monocular dense 3D reconstruction of construction and demolition waste on conveyor systems using motion stereo
(University of Peradeniya, Sri Lanka, 2026-08-20) Samarasekara, U. G. U. P.; Suraweeraarachchi, S. A. R. R.; Ariyaratne, H. M. P. B.; Jayasinghe, J. A. S. C.; Jayasinghe, U.; Bandara, C. S.
Accurate volumetric characterisation of Construction and Demolition (C&D) waste is critical for efficient material recovery and automated sorting on conveyor systems. Existing multi-sensor approaches, such as RGB-D cameras, LiDAR, and binocular stereo rigs, are operationally impractical in material recovery facilities due to high hardware costs, complex spatial calibration, and accelerated sensor degradation, while monocular solutions remain entirely unexplored for this material domain. This study aims to develop and evaluate a novel dense 3D reconstruction system for C&D waste from single-camera image sequences, establishing the first empirical benchmark for monocular reconstruction in this domain and enabling metric volume estimation without dual-camera hardware. Exploiting the kinematically constrained, rectilinear trajectory of industrial conveyor belts, temporally displaced monocular frames are treated as synthetic stereo pairs under the Motion Stereo paradigm. Dense disparity maps are computed via a conventional stereo pipeline using Convolutional Neural Networks for both foreground segmentation and learned stereo matching. Metric accuracy is fundamentally coupled to inter-frame baseline estimation; ArUco fiducial markers are deployed on the belt surface to provide drift-constrained displacement measurements, converting disparity maps into metrically consistent point clouds from which object volumes are derived by integrating the reconstructed surface geometry. The proposed system successfully generates dense 3D point clouds from monocular sequences across representative waste categories. Convolutional Neural Network-based segmentation effectively isolates foreground waste objects from the moving belt background. ArUco-constrained baseline tracking substantially reduces metric drift relative to unconstrained estimation. Volumetric estimates derived from the reconstructed point clouds demonstrate consistency with reference measurements, confirming the viability of the Motion Stereo paradigm for this application. This work establishes the first empirical benchmark for monocular dense 3D reconstruction of C&D waste, demonstrating that single-camera systems can deliver metrically accurate volumetric estimates. The approach substantially lowers deployment barriers compared to multi-sensor alternatives, offering a cost-effective pathway toward intelligent, automated material recovery
Modelling and simulation of low-temperature oxidation process of iron powder
(University of Peradeniya, Sri Lanka, 2026-08-20) Wahala, W. M. S. N.; Fernando, N.
Metal fuels are gaining increased attention as carbon free energy carries in combustion applications. Unlike conventional fuels with releases carbon dioxide during combustion, metal fuels are oxidized into metal oxides in solid form, which results in zero carbon dioxide emissions. Among various types of metals, iron presents advantages due to its wide availability and similar flame temperatures to conventional fuels. However, the oxidation kinetics and morphology evolution of iron particles in combustion environments are still not well known, limiting accurate prediction of ignition behavior. A validated kinetic model is therefore essential for characterizing powder passivation and ignition readiness, offering a foundation for furnace design. This study presents the mathematical modelling and numerical simulation of the low temperature oxidation of iron powder in oxygen rich environments. In the temperature range of 300-600 °C, oxidation is governed by solid state diffusion of iron ions through the growing oxide shell, in accordance with Wagner’s parabolic growth law. As an improvement to the existing models, a spherical shrinking core model was developed in the present study from first principles using Fick’s laws of diffusion in spherical coordinates. Simulation results at 500 °C confirmed parabolic growth behavior, with oxide thickness reaching approximately 6.1 µm after 60 minutes. The developed spherical model is also compared with a flat plate model which showed that the spherical shrinking core model predicts oxide layers about 13% thinner, highlighting the importance of curvature corrections. Validation against published thermogravimetric analysis (TGA) data demonstrated excellent agreement, with R² = 0.9950 and RMSE = 0.139 wt%. The fitted parabolic rate constant was 5.78 × 10−15m²/s, yielding an effective activation energy of 107.1 kJ/mol, consistent with literature values for Fe₂O₃ formation in the low temperature regime. Parametric studies further confirmed strong Arrhenius temperature sensitivity, with growth rates increasing by three orders of magnitude between 400 °C and 600 °C. By quantifying geometric correction effects and advancing beyond flat plate approximations, the study provides a reliable framework for further improving particle oxidation models. The findings form a foundation for future extensions to multi-layer oxide modelling, Kirkendall void effects, and multi particle oxidation behavior, supporting the development of iron powder as a practical carbon free energy carrier.
MESH: a multi-stage benchmark for mockery, extremism, sarcasm, and hate speech jailbreak vulnerabilities in open-source LLMs
(University of Peradeniya, Sri Lanka, 2026-08-20) Mithushika, S.; Sweshthika, S.; Kayalvizhy, T.
Jailbreak attacks enable large language models (LLMs) to generate harmful, unethical, or otherwise objectionable content, posing a significant safety concern. Existing benchmarks relying on first-generation static templates yield artificially suppressed attack success rates that poorly discriminate model vulnerability. This study introduces a multi-stage adaptive benchmarking framework spanning four harm categories, termed MESH (Mockery, Extremism, Sarcasm, Hate Speech), across three semantic targets (Social Norms, Protected Groups, Democratic Institutions). The dataset was constructed synthetically, with no external dataset or pre-existing prompt repository used, via a three-dimensional Cartesian product across the four harm categories, three semantic targets, and twelve attack templates, yielding 288 adversarial prompts through combinatorial template-target pairing, supplemented by 22 benign controls (10 clearly-safe, 12 borderline-safe). The twelve attack templates span two generations — six established Gen-1 templates (Role-play Wrapper, CoT Leakage, Adversarial Persona, Virtualization, Academic Framing, Payload Splitting) and six novel Gen-2 templates introduced by this work (Contextual Camouflage, Gradual Escalation, Semantic Obfuscation, Translation Attack, Moral Inversion, Logical Inversion), each assigned an a priori difficulty score (0.50-0.80) to enable subsequent difficulty-banded analysis. This generational and difficulty-graded design was deliberately chosen to test whether attack sophistication predicts effectiveness. Responses generated by LLaMA 3.2-3B were evaluated using a dual-model ensemble judge (Mistral-7B and Gemma-7B), selected for heterogeneous training origins to reduce correlated judgment bias; LLaMA 3.2-3B was restricted to a meta-judge role only, to avoid self-evaluation bias on its own outputs. A two-stage judging protocol separated harmfulness detection from four-class labeling to improve labeling consistency, with disagreements resolved via a harm-union policy and a benign-prompt guard preventing false-positive Bypass labels. Judge outputs were validated against gold labels from three independent human annotators (Fleiss' κ = 0.7573, substantial agreement). The pipeline achieved an overall attack success rate (ASR) of 41.11% (human gold labels), a 53.31% refusal rate on adversarial prompts, and a 95.45% safe-response rate with only 4.55% over-refusal on benign controls. CoT Leakage (79.17%) and Logical Inversion (70.83%) were the most effective attack types, while Mockery achieved the highest category-level ASR (56.94%). The ensemble judges reached 77.74% inter-model agreement with only a 0.49% ASR delta against human labels, supporting ensemble judging as a reliable substitute for large-scale human annotation. A notable difficulty- band paradox emerged: easy prompts (55.56% ASR) outperformed hard prompts (22.92% ASR), indicating that current safety alignment relies more on surface pattern recognition than genuine semantic understanding. A multi-criteria curation process, jointly filtering on confidence (≥0.65), inter-annotator agreement, and difficulty (≥0.70), yielded 105 high-confidence reference prompts suitable for Supervised Fine- Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF).