Water Pipe Failure Prediction: A Machine Learning Approach Enhanced By Domain Knowledge

Abstract

Drinking water pipe and waste water pipe networks are valuable urban infrastructure assets that are responsible for reliable water resource distributions and waste water collection. However, due to fast growing demand and aging assets, water utilities find it increasingly difficult to efficiently maintain their pipe networks. Pipe failures - drinking water pipe breaks and waste water pipe blockages - can cause significant economic and social costs, and hence have become the primary challenge to water utilities. Identifying key influential factors, e.g., pipes’ physical attributes, environmental features, is critical for understanding pipe failure behaviours. The domain knowledge plays a significant role in this aspect. In this work, we propose a Bayesian nonparametric machine learning model with the support of domain knowledge for pipe failure prediction. It can forecast future high-risk pipes for physical condition assessment, thereby proactively preventing disastrous failures. Moreover, compared with traditional machine learning approaches, the proposed model considers domain expert knowledge and experience, which helps avoid the limit of traditional machine learning approaches - learning only from what it sees - and improves prediction performance.