Multi-step time-to-failure predictions in water pipelines using feature engineering and cascading ensembles

Abstract

Predicting failure timings in water pipelines is crucial for actionable predictive maintenance and rehabilitation planning of water distribution networks. However, existing time-to-failure prediction models have limited capability to incorporate failure history and determine sequential failures in individual pipeline sections. Secondly, accuracy of these models is hampered by lack of in-depth investigation and selection of most significant predictors of failure timings from historical data. As dynamic features can better determine time-based deterioration impacts, the study develops a customized weather index, and other interaction features for accuracy enhancements. Furthermore, feature selection is further automated for optimized performance with MAE ranges:1.4–0.5 for the developed models. Overall, GA-based feature selection and feature engineering results in a 20–50 % increase in the model performance, with highest reported performance when compared with existing models. Finally, a cascading ensemble for predicting first, second and third failure of individual pipelines is proposed, tested and validated using both hold-out and out-of-sample testing, exhibiting higher performance (MAE:0.8–1.1) than alternative multi-output models. Demonstrated using a web-based application, the developed study offers a novel modeling regime for high performance failure timings prediction of water pipelines, offering a micro-level analyses of pipe sections, giving useful insights into the complex interactions of features for indirectly gauging deterioration rate in water networks.