Scholarlink

<p>A novel physics-informed framework is proposed for robust leak localization in water distribution networks (WDNs), integrating multi-level perturbation sensitivity analysis, advanced clustering, and ensemble learning. Unlike conventional approaches, this method systematically generates diverse pressure sensitivity matrices through multiple nodal demand perturbations, capturing a richer set of hydraulic responses. These matrices are standardized and fused as input features. A topological-spatial partitioning strategy using fuzzy c-means clustering is introduced, followed by an innovative iterative outlier optimization that ensures monitoring regions are both contiguous and hydraulically meaningful. Representative nodes are then selected from each cluster. For precise fault classification, an ensemble soft-voting model combining Random Forest (RF) and multilayer perceptron (MLP) is developed, with SMOTE-based resampling to address class imbalances. Extensive evaluation on four benchmark networks demonstrates state-of-the-art performance: the framework achieves a mean accuracy of 98.85% on Exa7, 99.15% on Balerma, and exceeds 91% and 95% on Net3 and Net2, respectively. Furthermore, feature importance analysis validates that our algorithm intuitively prioritizes hydraulically critical nodes, bridging data-driven detection with explicit physical interpretability. By significantly advancing accuracy, robustness, and scalability over single-perturbation baselines, this architecture provides a highly reliable foundation for leak diagnosis and large-scale field deployments.</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

An Integrated Physics-Informed Multi-Perturbation Framework for Scalable and Interpretable Leakage Detection in Water Distribution Networks

  • 期刊论文

  • 发表日期：

    2026年5月16日

查看全文
  • Feiyu Duan,
  • Tianwei Mu,
  • Bo Zhou,
  • Baokuan Ning,
  • Haoqiang Tan

摘要

A novel physics-informed framework is proposed for robust leak localization in water distribution networks (WDNs), integrating multi-level perturbation sensitivity analysis, advanced clustering, and ensemble learning. Unlike conventional approaches, this method systematically generates diverse pressure sensitivity matrices through multiple nodal demand perturbations, capturing a richer set of hydraulic responses. These matrices are standardized and fused as input features. A topological-spatial partitioning strategy using fuzzy c-means clustering is introduced, followed by an innovative iterative outlier optimization that ensures monitoring regions are both contiguous and hydraulically meaningful. Representative nodes are then selected from each cluster. For precise fault classification, an ensemble soft-voting model combining Random Forest (RF) and multilayer perceptron (MLP) is developed, with SMOTE-based resampling to address class imbalances. Extensive evaluation on four benchmark networks demonstrates state-of-the-art performance: the framework achieves a mean accuracy of 98.85% on Exa7, 99.15% on Balerma, and exceeds 91% and 95% on Net3 and Net2, respectively. Furthermore, feature importance analysis validates that our algorithm intuitively prioritizes hydraulically critical nodes, bridging data-driven detection with explicit physical interpretability. By significantly advancing accuracy, robustness, and scalability over single-perturbation baselines, this architecture provides a highly reliable foundation for leak diagnosis and large-scale field deployments.