Scholarlink

<p>Predicting pipe failures in water distribution networks is central to risk-based asset management. Although machine learning approaches are increasingly applied in this field, many studies remain constrained by class imbalance, the recurrent nature of failure events, and limited validation strategies. This study proposes an interpretable data-driven framework that explicitly addresses these challenges using a 12-year failure dataset from an urban water distribution network. A weighted logistic regression model was developed to account for imbalanced failure records and evaluated using a 70–30 training–validation split. The model demonstrated stable predictive performance, achieving an AUC of 0.88 on the validation dataset. Analysis of predictor effects indicated that pipe length and operating pressure significantly increase failure probability, whereas pipe diameter reduces it. A notable finding was the negative association between pipe age and failure probability. This counterintuitive pattern was interpreted as evidence of survivor bias, reflecting the concentration of failures during early service life. To reinforce these findings, a decision tree model was employed as an independent structural validation tool, confirming consistent patterns, particularly regarding early-life failures. The proposed framework provides a practical and interpretable basis for risk prioritization and proactive maintenance planning.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Pipe Failure Prediction in Water Distribution Networks Using Interpretable Machine Learning under Imbalanced and Recurrent Failure Conditions

  • 期刊论文

  • 发表日期：

    2026年3月24日

查看全文
  • Homayoun Motiee,
  • Mohsen Sattari

摘要

Predicting pipe failures in water distribution networks is central to risk-based asset management. Although machine learning approaches are increasingly applied in this field, many studies remain constrained by class imbalance, the recurrent nature of failure events, and limited validation strategies. This study proposes an interpretable data-driven framework that explicitly addresses these challenges using a 12-year failure dataset from an urban water distribution network. A weighted logistic regression model was developed to account for imbalanced failure records and evaluated using a 70–30 training–validation split. The model demonstrated stable predictive performance, achieving an AUC of 0.88 on the validation dataset. Analysis of predictor effects indicated that pipe length and operating pressure significantly increase failure probability, whereas pipe diameter reduces it. A notable finding was the negative association between pipe age and failure probability. This counterintuitive pattern was interpreted as evidence of survivor bias, reflecting the concentration of failures during early service life. To reinforce these findings, a decision tree model was employed as an independent structural validation tool, confirming consistent patterns, particularly regarding early-life failures. The proposed framework provides a practical and interpretable basis for risk prioritization and proactive maintenance planning.