Data-Driven Approaches in Water Pipe Condition Assessment and Failure Prediction

The increasing frequency of failures in ageing water distribution networks has created significant operational, financial, and environmental challenges for utility providers. This PhD research presents a novel integrated approach for predictive maintenance of buried water pipes, combining machine learning techniques, numerical methods, and physical modelling to improve failure prediction and support proactive asset management. A key contribution of this work is the development of methods to address the common issue of class imbalance in pipe failure datasets. A semi-supervised clustering approach was introduced, integrating expert knowledge with data-driven techniques to enhance the representation of rare failure events. This was supported by a hybrid sampling strategy and class weighting scheme, which together improved model accuracy and robustness, particularly under severely imbalanced data conditions. To further enhance predictive reliability, a stacked ensemble model was developed specifically for temporal forecasting. Unlike conventional methods that rely on random data splits, this model applied temporal segmentation, providing a more realistic basis for forward-looking failure prediction. An incremental replacement metric was also proposed to assess how many failures could be prevented under different intervention scenarios. Another major contribution involved incorporating environmental variables such as soil type, temperature, and precipitation into the predictive workflow. The research examined material-specific interactions between environmental factors and pipe failures, demonstrating that including these contextual variables significantly improved prediction performance. In addition, a hybrid physical–numerical modelling approach was established, coupling finite element simulations with white-box regression algorithms to generate surrogate models capable of assessing pipe mechanical behaviour under varying loading conditions. This enabled scalable structural evaluations while maintaining computational efficiency. The research was validated using real-world datasets and applied in collaboration with an industry partner in the UK and Europe, demonstrating its practical relevance. Overall, this study contributes new tools, insights, and methodologies to support the transition toward data-driven and risk-based infrastructure management in the water sector.<p></p>

Author dc:creator

• Ramiz Beig Zali (21065429)

• Asset management
• Buried infrastructure
• Class imbalance
• Condition assessment
• Data-driven modelling
• Environmental factors
• Failure prediction
• Finite element modelling
• Hybrid sampling
• Imbalanced learning
• Machine learning
• Maintenance planning
• Pipe deterioration
• Predictive maintenance
• Risk prioritisation
• Semi-supervised clustering
• Soil-pipe interaction
• Surrogate modelling
• Temporal modelling
• Water distribution networks

Statement dc:rights

• All rights reserved
• Open Access after 2027-06-22

Identifier

10779/exe.32956334.v1

OAI identifier oai:identifier

oai:figshare.com:article/32956334