Failure Prediction of Municipal Water Pipes Using Machine Learning Algorithms

Pipe failure prediction has become a crucial demand of operators in daily operation and asset management due to the increase in operation risks of water distribution networks. In this paper, two machine learning algorithms, namely, random forest (RF) and logistic regression (LR) algorithms are employed for pipe failure prediction. RF algorithm consists of a group of decision trees that predicts pipe failure independently and makes the final decision by voting together. For the LR algorithm, the mapping relationship between existing data and decision variables is expressed by the logistic function. Then, the prediction is made by comparing the conditional probability with the fixed threshold value. The proposed algorithms are illustrated using an actual water distribution network in China. Results indicate that the RF algorithm performs better than the LR algorithm in terms of accuracy, recall, and area under the receiver operating characteristic curve. The effects of seven characteristics on pipe failures are analyzed, and diameter and length are identified as the top two influential factors.