Lavindra de Silva
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Lavindra de Silva has a total of 29 co-authors. Here are the top co-authors by number of shared publications:
- Duncan McFarlane
University of Cambridge
3 shared publications - Ioannis Brilakis
Univ. of Cambridge
3 shared publications - Gregory Hawkridge
University of Cambridge
2 shared publications - Percy Lam
University of Cambridge
2 shared publications - Abir Al-Tabbaa
Department of Engineering, University of Cambridge, Cambridge, UK
1 shared publication Alan Thorne
Institute for Manufacturing, University of Cambridge
1 shared publication
Latest Publications
- Open Access
- 1 September 2026
- Published by Elsevier in SSRN Electronic Journal
EdgeGT: Edge-Biased Attention for the Multi-Depot Vehicle Routing Problem in Road Maintenance Applications
Road maintenance is essential for ensuring reliable transportation access, safeguarding road users, and supporting societal and economic development. Network-level road maintenance planning involves optimizing routes for a fleet of maintenance vehicles; however, existing learning-based routing methods are not well suited to this context. To address this limitation, this paper formulates road maintenance routing as a Multi-Depot Vehicle Routing Problem and proposes EdgeGT, a graph-based attention model that explicitly emphasizes edge-level information through distance-aware attention bias, thereby better capturing realistic road travel costs. A synthetic data generation strategy is introduced to approximate complex road-network distances and support model training under realistic conditions. In addition, a Multi-Depot Multi-Trajectory decoding strategy is developed to improve solution exploration. The results demonstrate that EdgeGT outperforms existing learning-based methods by at least 5\% on large-scale, non-Euclidean routing benchmarks while remaining competitive with classical heuristics. A real-world case study further demonstrates the practical applicability of the proposed approach and its potential to generate dynamic, balanced routing plans for network-level road maintenance
- 1 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEdgeGT: Edge-Biased Attention for the Multi-Depot Vehicle Routing Problem in Road Maintenance Applications
Road maintenance is essential for ensuring reliable transportation access, safeguarding road users, and supporting societal and economic development. Network-level road maintenance planning involves optimizing routes for a fleet of maintenance vehicles; however, existing learning-based routing methods are not well suited to this context. To address this limitation, this paper formulates road maintenance routing as a Multi-Depot Vehicle Routing Problem and proposes EdgeGT, a graph-based attention model that explicitly emphasizes edge-level information through distance-aware attention bias, thereby better capturing realistic road travel costs. A synthetic data generation strategy is introduced to approximate complex road-network distances and support model training under realistic conditions. In addition, a Multi-Depot Multi-Trajectory decoding strategy is developed to improve solution exploration. The results demonstrate that EdgeGT outperforms existing learning-based methods by at least 5\% on large-scale, non-Euclidean routing benchmarks while remaining competitive with classical heuristics. A real-world case study further demonstrates the practical applicability of the proposed approach and its potential to generate dynamic, balanced routing plans for network-level road maintenance
- Open Access
- 31 March 2026
- Published by Elsevier in SSRN Electronic Journal
Cross-Context Synthetic Data for Road Defect Detection: An Interpretability Analysis
Road defect detection often uses supervised trained deep learning models that suffer from low reliability in different road scenes. While synthetic data augmentation has been widely used to address data scarcity, its effectiveness is typically evaluated only through downstream task performance in the same context, with little analysis interpreting how synthetic data affects the internal representations and spatial attention of detection models. This paper identifies data scarcity and context variation as two distinct deficiencies in defect detection. Experiments proceed by procedurally adding potholes synthesised from baseline pavement images and images from external sources, and examine if they address both deficiencies. To further illustrate the effects, the trained models were evaluated on test sets within and beyond the training context, using interpretation techniques such as logistic regression probes, EigenCAM heatmaps and bounding box (bbox) inspectors to help explain the root causes of improvements or failures. Results suggested that synthesising instances onto a target background improves detection on that background by improving both classification and localisation. When evaluating across contexts, however, a diagnosis of poor detection suggests that classification tend to generalise more effectively than spatial attention. Diffused spatial attention hinders the differentiation of background and foreground and creates a localisation bottleneck. These findings suggest that robustness in real-world applications depends on exposure to diverse image backgrounds and a strategic integration of synthetic data from multiple sources.
- 31 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCross-Context Synthetic Data for Road Defect Detection: An Interpretability Analysis
Road defect detection often uses supervised trained deep learning models that suffer from low reliability in different road scenes. While synthetic data augmentation has been widely used to address data scarcity, its effectiveness is typically evaluated only through downstream task performance in the same context, with little analysis interpreting how synthetic data affects the internal representations and spatial attention of detection models. This paper identifies data scarcity and context variation as two distinct deficiencies in defect detection. Experiments proceed by procedurally adding potholes synthesised from baseline pavement images and images from external sources, and examine if they address both deficiencies. To further illustrate the effects, the trained models were evaluated on test sets within and beyond the training context, using interpretation techniques such as logistic regression probes, EigenCAM heatmaps and bounding box (bbox) inspectors to help explain the root causes of improvements or failures. Results suggested that synthesising instances onto a target background improves detection on that background by improving both classification and localisation. When evaluating across contexts, however, a diagnosis of poor detection suggests that classification tend to generalise more effectively than spatial attention. Diffused spatial attention hinders the differentiation of background and foreground and creates a localisation bottleneck. These findings suggest that robustness in real-world applications depends on exposure to diverse image backgrounds and a strategic integration of synthetic data from multiple sources.
- Open Access
- 15 December 2024
- Published by Elsevier in Advanced Engineering Informatics
- Vol. 64
CAMHighways: The Cambridge Highways dataset
- 15 December 2024
- Published by Elsevier in Advanced Engineering Informatics
- Vol. 64
Open AccessCAMHighways: The Cambridge Highways dataset
- 3 February 2024
- Published by Springer Nature
- p. 297-308
Evaluation of Industrial Pilots of Low-Cost Digital (Shoestring) Solutions
- 3 February 2024
- Published by Springer Nature
- p. 297-308
Evaluation of Industrial Pilots of Low-Cost Digital (Shoestring) Solutions
- 3 June 2022
- Published by Springer Nature
- p. 263-275
A Graphical Environment to Support the Development of Affordable Digital Manufacturing Solutions
- 3 June 2022
- Published by Springer Nature
- p. 263-275
A Graphical Environment to Support the Development of Affordable Digital Manufacturing Solutions
- Open Access
- 5 May 2022
- Published by Elsevier in IFAC-PapersOnLine
- Vol. 55 (2) , 414-419
Digitalisation for SME Manufacturers: A Framework and a Low-Cost Approach
- 5 May 2022
- Published by Elsevier in IFAC-PapersOnLine
- Vol. 55 (2) , 414-419
Open AccessDigitalisation for SME Manufacturers: A Framework and a Low-Cost Approach
- Open Access
- 22 December 2021
- Published by Springer Nature in AI & SOCIETY
- Vol. 38 (6) , 2659-2665
Towards low-cost machine learning solutions for manufacturing SMEs
Machine learning (ML) is increasingly used to enhance production systems and meet the requirements of a rapidly evolving manufacturing environment. Compared to larger companies, however, small- and medium-sized enterprises (SMEs) lack in terms of resources, available data and skills, which impedes the potential adoption of analytics solutions. This paper proposes a preliminary yet general approach to identify low-cost analytics solutions for manufacturing SMEs, with particular emphasis on ML. The initial studies seem to suggest that, contrarily to what is usually thought at first glance, SMEs seldom need digital solutions that use advanced ML algorithms which require extensive data preparation, laborious parameter tuning and a comprehensive understanding of the underlying problem. If an analytics solution does require learning capabilities, a ‘simple solution’, which we will characterise in this paper, should be sufficient.
- 22 December 2021
- Published by Springer Nature in AI & SOCIETY
- Vol. 38 (6) , 2659-2665
Open AccessTowards low-cost machine learning solutions for manufacturing SMEs
Machine learning (ML) is increasingly used to enhance production systems and meet the requirements of a rapidly evolving manufacturing environment. Compared to larger companies, however, small- and medium-sized enterprises (SMEs) lack in terms of resources, available data and skills, which impedes the potential adoption of analytics solutions. This paper proposes a preliminary yet general approach to identify low-cost analytics solutions for manufacturing SMEs, with particular emphasis on ML. The initial studies seem to suggest that, contrarily to what is usually thought at first glance, SMEs seldom need digital solutions that use advanced ML algorithms which require extensive data preparation, laborious parameter tuning and a comprehensive understanding of the underlying problem. If an analytics solution does require learning capabilities, a ‘simple solution’, which we will characterise in this paper, should be sufficient.
- 1 April 2019
- Published by Elsevier in Theoretical Computer Science
- Vol. 763, 12-37
HTN-like solutions for classical planning problems: An application to BDI agent systems
- 1 April 2019
- Published by Elsevier in Theoretical Computer Science
- Vol. 763, 12-37
HTN-like solutions for classical planning problems: An application to BDI agent systems