Graph-Based Learning for Leak Detection and Localisation in Water Distribution Networks*

Abstract

We propose the application of geometric deep learning techniques to the challenging leak detection and isolation problem in water distribution networks (WDNs). Specifically, we train two Chebyshev polynomial kernel Graph Convolutional Networks for the task of prediction, and reconstruction of nodal pressures in a WDN. Comparing the two network outputs (a predicted healthy model state with a reconstructed observation) a residual signal is obtained and analysed to detect leakages. By exploiting topological properties in the proposed approach, leakage isolation is also performed. We benchmark our method on the BattLeDIM 2020 dataset.

Topics & Concepts