Exploring the perspective of time: A framework for dynamic assessment of leakage risk in WDNs based on a joint model of survival analysis and machine learning

Assessment of leakage risk in water distribution networks (WDNs) and implementing preventive monitoring for high-risk pipelines are widely recognized strategies for mitigating leakage-related losses. Conventional leakage risk assessment methods face three critical challenges: class imbalance, insufficient modeling of time-varying risk factors, and limited model interpretability. To address these issues, we propose an interpretable machine learning framework, Interpretable Survival Analysis with Class-Imbalance Mitigation (ISACIM). The framework synergizes static risk assessment with dynamic survival analysis to achieve spatiotemporal decoupling in leakage probabilistic evaluation. By integrating hybrid data-balancing strategies and a conditional generative adversarial network (GAN), ISACIM effectively resolves leakage sample distribution skewness. Experimental results demonstrated that ISACIM achieved a 7 % improvement in leakage pipeline prediction accuracy on real-world WDN datasets, along with enhanced survival analysis performance, 7.89 % increase in Time AUC. To overcome limitations in time-dependent risk factor analysis, we introduce Shapley Additive Explanations-based methods, systematically revealing for the first time the dynamic evolution of dominant risk factors across pipeline lifecycles: material properties and joint types dominate leakage risk during the initial service phase, while length and diameter become predominant in long-term service. Furthermore, the developed web-based WDN leakage risk assessment platform integrates predictive results with interpretability analysis, providing a decision support tool combining theoretical rigor and practicability for WDNs reliability evaluation.