Stochastic data mining tools for pipe blockage failure prediction

Abstract

Failure prediction plays an important role in the management of urban water systems infrastructures. An accurate description of the deterioration of urban drainage systems is essential for optimal investment and rehabilitation planning. In the study presented in this paper, a new method to predict sewer pipe failure based on robust decision trees is proposed. Five other different stochastic failure prediction models – the non-homogeneous Poisson process, the zero-inflated non-homogeneous Poisson process, classical decision tress (CART and Random Forest algorithms), the Weibull accelerated lifetime model and the linear extended Yule process – are also implemented and explored in order to identify models that combine good failure prediction results with robustness. The six models were tested on the asset register and pipe failure register of a large US wastewater utility; only pipe blockage failures were considered in this study. The linear extended Yule process and the zero-inflated non-homogeneous Poisson process presented the overall best results throughout the models’ comparison, showing a good ability to detect pipes with high likelihood of blockage failure. Decision trees based on robust random forests only detected pipes with high likelihood of failure when considering a short-term prediction window; the accuracy of the predictions was one of the best when using the robust decision tree model. The Weibull accelerated lifetime model provided some of the best medium-term predictions but performed less well for shorter prediction windows.