Intelligent Approaches for Predicting Failure of Water Mains

Abstract

Water mains are indispensable infrastructures in many countries around the world. Several factors may be responsible for the failure of these essential pipelines that negatively impact their integrity and service life. The purpose of this study is to propose models that can predict the average time to failure of water mains by using intelligent approaches, including artificial neural network (ANN), ridge regression (l2), and ensemble decision tree (EDT) models. The developed models were trained by using collected data from Quebec City water mains, including records of the possible factors, such as the materials, length, and diameter of pipes, that contributed to the failure. The ensemble learning model was applied by using a boosting technique to improve the performance of the decision tree model. All models, however, were able to predict reasonably the failure of water mains. A global sensitivity analysis (GSA) was then conducted to test the robustness of the model and to show clearly the relationship between the input and output of the model. The GSA results show that gray cast iron (CI), hyprescon/concrete (Hy), and ductile iron with lining (DIL) are the most vulnerable materials for the model output. The results also indicate that the failure of water mains mostly depends on pipe material and length. It is hoped that this study will help decision makers to avoid unexpected water main failure.