Prediction of Pipe Failure by Considering Time-Dependent Factors: Dynamic Bayesian Belief Network Model

Abstract

The water supply system (WSS) is a lifeline of the modern city. Transmission and distribution pipes, spatially distributed components of the WSS, are most often vulnerable to failure (leakage, breakage, or burst). Many factors contribute to pipe failure. These factors can be categorized as pipe physical attributes, operational practices, and environmental factors (i.e., climatic factors and soil corrosivity). The impact of failure factors can be static or dynamic (time dependent) in nature. This study quantifies the impact of time-dependent factors on the annual and monthly trends of pipe failures. It considers a dynamic Bayesian network (DBN) as an alternative model for prediction of pipe failures. The DBN extends the static capability of a Bayesian belief network to model dynamic systems. The developed DBN model was trained using annual and monthly data categorized based on combined (metallic pipes) and specific pipe material. The annual model was considered to predict annual pipe failure trends, whereas the monthly model was used to evaluate seasonal variations and trends of pipe failure. Model performance evaluation results show that the proposed models are effective in predicting the trends and expected total number of annual and monthly pipe failures. The DBN model result indicates that the models trained on specific pipe material performed well compared to the model trained using combined metallic pipe data.