Predicting Water Pipe Failures Using Deep Learning Algorithms

Abstract

With the increase in the operation risks of water distribution networks (WDNs), the prediction of pipe failures is of great significance in developing efficient maintenance strategies. This study used a residual network (ResNet), a newly proposed deep learning (DL) algorithm, to predict pipe failure, and its effectiveness was compared with that of a classic convolution neural network (CNN) algorithm. Network structure of ResNet used in the classification of one-dimensional pipe vectors was built. The synthetic minority oversampling technique (SMOTE) was used to improve the prediction accuracy because of the imbalanced pipe database provided by the local water sector. The analysis of a real WDN in China showed that ResNet performed better than CNN in terms of recall rate and area under the receiver operating characteristic curve with reasonable time costs. The maintenance rate was defined and discussed to measure the efficiency of maintenance activities. More than half of the failures can be prevented by maintaining less than 10% of the pipes based on the proposed ResNet algorithm. In addition, the Shapley Additive exPlanations (SHAP) method was used to interpret the DL model. The SHAP method evaluated the impact of different features on pipe failure, and the pipe length and diameter were proved to be two influential features.