Knowledge-data fusion for water supply pipe failure prediction: A hybrid physics-informed and data-driven method

Pipe failure prediction is critical for daily maintenance and asset management of water distribution networks (WDNs). As the mainstream paradigm for pipe failure modeling, data-driven machine learning (ML) methods are limited by the sparsity of operational data in WDNs and lack physical interpretability. This study proposes a hybrid physics-informed and data-driven method integrating mechanical knowledge with operational data to improve pipe failure prediction. The hybrid method adopts the ML model as its primary architecture, under which a mechanical approach is incorporated to embed the structural safety factor of pipes as an extended physical feature into the feature space of the ML model. The proposed hybrid method is applied to predict pipe failures of a large WDN in China. The results demonstrate that the hybrid models deliver superior predictive capacity and cost-effectiveness compared to pure ML models, achieving significant improvements across various evaluation metrics. The extended physical feature plays a crucial role in pipe failure prediction, with its contribution to the model's predictions aligning with established mechanical principles. Additionally, pipes crossing roads at oblique angles and those located at road intersections are more prone to failure. This research provides insights for improving management strategies and resilience in WDNs.