Enhanced failure prediction for water distribution networks using semi-supervised survival models and sparse acoustic...

Achieving proactive failure prediction for water distribution networks (WDNs) is crucial for enhancing infrastructure reliability, reducing operational costs, and minimizing resource waste. Existing approaches face three major challenges: scarcity and high imbalance of failure data, reliance on static metrics lacking dynamic insights, and unclear performance differences and applicability between continuous and discrete time modeling paradigms. To address these, this work proposes an interpretable failure prediction framework with three key contributions: (1) Novel acoustic risk quantification: an acoustic risk field is constructed through spatial interpolation of sparse acoustic signals, enabling refined quantification of risk factors; (2) Paradigm-level model comparison: continuous and discrete time survival models are systematically compared, revealing their complementary strengths for different decision contexts; (3) Semi-supervised augmentation mechanism: a data augmentation method is developed to mitigate class imbalance, based on a dual perspective of data distribution consistency and survival information mining. This study is validated using a real-world dataset collected in a city in northern China, comprising 147,015 pipes (including 2,944 failure records). Results show that incorporating acoustic signals improves model performance from 2.5% to 4.5% with a mean improvement of 3.2%, confirming acoustic signals as a key influencing factor second only to pipe material and spatial location. The semi-supervised augmentation method effectively leverages survival time information from right-censored data to generate failure samples consistent with the true distribution, achieving a C-index exceeding 92% across different survival time quantiles. Continuous-time models prove more suitable for maintenance priority ranking, while discrete-time models are advantageous for long-term risk calibration and planning. Ultimately, this approach enables spatiotemporal prediction of pipe network health status, providing water utilities with actionable decision support for maintenance prioritization (via continuous-time models) and long-term capital planning (via discrete-time models), thereby enhancing the long-term operational reliability of WDNs.