Scholarlink

<p>Reliable leak detection in water distribution networks is difficult when recorded leak events are limited and when operational data cannot be freely exchanged among utilities. This paper develops a weight-optimized federated learning framework for data-scarce leak detection. Each participating network trains a local Unsupervised Anomaly Detection (USAD) model on its own pressure time-series data, while the server combines the local models without accessing the original data. To reduce the influence of differences in topology, demand patterns, and data distributions across networks, the aggregation step is adjusted by an optimization-based weighting strategy. The framework is tested under two practical conditions: a target network with no available training data and a target network with only a small training set. The experimental results show that the proposed approach improves the overall detection performance compared with local USAD training, especially when the target network lacks sufficient data. These results indicate that weighted federated learning can provide a practical route for privacy-preserving leak detection in water distribution networks with limited target-network observations.</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

A Weight-Optimized Federated Learning Framework for Leak Detection in Water Distribution Networks

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Juan Li,
  • Weijia Feng,
  • Xiaofeng Tang

摘要

Reliable leak detection in water distribution networks is difficult when recorded leak events are limited and when operational data cannot be freely exchanged among utilities. This paper develops a weight-optimized federated learning framework for data-scarce leak detection. Each participating network trains a local Unsupervised Anomaly Detection (USAD) model on its own pressure time-series data, while the server combines the local models without accessing the original data. To reduce the influence of differences in topology, demand patterns, and data distributions across networks, the aggregation step is adjusted by an optimization-based weighting strategy. The framework is tested under two practical conditions: a target network with no available training data and a target network with only a small training set. The experimental results show that the proposed approach improves the overall detection performance compared with local USAD training, especially when the target network lacks sufficient data. These results indicate that weighted federated learning can provide a practical route for privacy-preserving leak detection in water distribution networks with limited target-network observations.