LeakAgent: a multi-agent system for leak detection using multi-modal large language model coordination in water distribution network

Abstract

Water distribution networks (WDNs) suffer from severe leakage, yet pinpointing the leaks remains difficult. Data-driven detection is typically split into disconnected stages of hydraulic modelling, partitioning, sensor planning and analysis, and it treats the questions of whether a leak exists and where it lies as separate problems, which limits accuracy and transfer across networks. Here we show that these tasks can be unified within a multi-agent system (MAS) coordinated in natural language by a large language model (LLM). A single hydraulic pressure-sensitivity field, computed once, informs network zoning, sensor placement and detection; a dual-branch graph model then determines leak presence and the affected partition jointly; and the coordinator autonomously detects and repairs its own execution faults. Across five networks of 191 to 920 nodes, the system both detects leaks and identifies the affected partition with about 96% accuracy at 20% leak magnitude, remains stable under demand and sensor noise, and transfers between networks without manual retuning.