Ageing underground water pipelines: Time-to-failure models, gaps and future directions

Abstract

Accurate prediction of the failure time of individual pipelines of a water distribution network can assist in preventing sudden bursts and leaks. Failure prediction over time can help eliminate managerial uncertainty in pipe rehabilitation and replacement decision-making. Since time-based deterioration modeling has less focus in past research, the study focuses on a critical review of the current state-of-the-art for time-to-failure/failure age models related to water pipelines. A unique unsupervised learning-based clustering framework is used to perform an in-depth and robust literature analysis. Hierarchical clustering reveals the main modeling approaches, classified as 1) physical data-based models and 2) historical data-based failure models. Critical research gaps are further explored using t-SNE and Gaussian Mixture Models based clustering. Identified gaps include fragmented modeling approaches, lack of integration between physical and data-driven models, limited data related issues, and a lack of insight on practical translation of model findings for effective utility management. Future studies can consider several integration strategies to overcome individual model limitations, use of generative AI to enrich data, IoT implementation for physical data collection, improve feature engineering and feature extraction efforts, and consider domain knowledge from hydraulic models to improve AI models. Overall, the study offers practical insights for predicting the remaining time-to-failure and service life of water pipelines.