Pipe Fault Prediction for Water Transmission Mains

Abstract

Every network of supply waterlines experiences thousands of yearly bursts, breaks, leakages, and other failures. These failures waste a great amount of resources, as not only the waterlines need to be repaired, but also water is wasted and the distribution service is interrupted. For that reason, many water facilities employ proactive maintenance strategies in their networks, where they replace likely-to-fail pipes in advance to prevent the failures. In this paper, we aim to establish a reliable prediction model that can accurately predict faults in waterlines prior to their occurrence. We propose a specific segmentation method for long transmission mains, as well as three data-driven models and one rule-based prediction model. We evaluate a real world waterline network used in Israel, operated by Mekorot company, using three common metrics. The results show that the data-driven algorithms outperform the rule-based model by at least 5% in each of the metrics. Additionally, their prediction becomes more accurate as they are trained with more data, but enhancing these data with geographically related features does not improve the accuracy further.

Keywords

• fault prediction
• machine learning
• pipe segmentation