Comparison of automatic and guided learning for Bayesian networks to analyse pipe failures in the water distribution system

Abstract The reliability of the water distribution system is critical to maintaining a secure supply for the population, industry and agriculture, so there is a need for proactive maintenance to help reduce water loss and down times. Bayesian networks are one approach to modelling the complexity of water mains, to assist water utility companies in planning maintenance. This paper compares and analyses how accurately the Bayesian network structure can be derived given a large and highly variable dataset. Method one involved using automated learning algorithms to build the Bayesian network, while method two involved a guided method using a combination of historic failure data, prior knowledge and pre-modelling data exploration of the water mains. By understanding common failure types (circumferential, longitudinal, pinhole and joint), the guided learning Bayesian Network was able to capture the interactions of the surrounding soil environment with the physical properties of pipes. The Bayesian network built using data exploration and literature was able to achieve an overall accuracy of 81.2% when predicting the specific type of water mains failure compared to the 84.4% for the automated method. The slightly greater accuracy from the automated method was traded for a sparser Bayes net where the interpretation of the interactions between the variables was clearer and more meaningful.