Explainable machine learning reveals early-life pipe failure mechanisms in urban water networks using XGBoost, SHAP, and ALE

The deterioration of urban water infrastructure and the increasing occurrence of pipe failures represent major challenges in the management of water distribution networks. In recent years, shifting from reactive to proactive approaches in infrastructure asset management has become increasingly important. Accordingly, this study presents an explainable machine learning framework for predicting pipe failures in the water distribution network of District 1, Tehran. The dataset comprised information on 4,048 pipes and 578 recorded failure events over the period from 2013 to 2024. Four predictive models, including XGBoost, Random Forest, Logistic Regression, and an Artificial Neural Network, were developed and evaluated using a pipe ID-level data split, five-fold grouped cross-validation, and an independent test set. XGBoost achieved the highest observed ROC-AUC of 0.9039 on the independent test set, compared with 0.8974 for Random Forest. However, the 95% bootstrap confidence interval for the difference in AUC included zero, indicating no statistically significant superiority of XGBoost over Random Forest. Temporal validation yielded a ROC-AUC of 0.863, providing an additional assessment of the models under a prospective validation setting. To interpret model behavior, SHAP and Accumulated Local Effects (ALE) methods were employed. The results showed that pipe length and age had the greatest contributions to the model output, followed by pressure and diameter. The ALE analysis further characterized the shape and direction of the modeled relationships between continuous predictor variables and the predicted probability of pipe failure. Overall, the proposed framework provides a useful approach for identifying pipe failure patterns and supporting data-driven asset management within the studied water distribution network.