A data preprocessing method for predicting the failure probability of water supply pipelines

To achieve a quantitative assessment of the reliability of individual water supply pipelines, this study proposes a machine learning-based analytical framework.First, individual pipeline samples are constructed based on pipeline attributes and historical failure records.Each sample comprises feature variables that influence failures, along with a corresponding failure count label.Subsequently, the samples are aggregated into homogeneous groups based on feature similarity, and predictive models-including Random Forest (RF), Artificial Neural Networks (ANN), and Support Vector Regression (SVR)-are developed to estimate the failure counts for each group.Building on this, the failure process of the pipelines is modeled as a non-homogeneous Poisson process.By incorporating the individual pipeline length and the average failure rate of the corresponding homogeneous group, the failure probability is estimated, thereby enabling a quantitative assessment of the reliability of individual pipelines.A case study on gray cast iron water supply pipelines in certain administrative regions of Shanghai is conducted to validate the effectiveness and applicability of the proposed framework.The results demonstrate that all three models can effectively predict the failure counts of homogeneous groups, and that these predictions can be used to accurately estimate the failure probabilities of individual pipelines.These findings highlight the promising potential of the proposed method for risk mitigation and asset management in the water industry.