Detection and localization of pipeline leakage of water distribution networks based on graph convolutional networks

A leak in a water distribution network (WDN) creates a localized pressure depression whose effect diffuses along the network topology through headloss coupling. This spatial diffusion is well approximated by graph Laplacian smoothing, so we use a graph convolutional network as a learnable surrogate of hydraulic information propagation on the graph. The method casts single-snapshot leakage analysis as graph-level classification of normal versus abnormal states using topology-aware convolution. Localization then leverages gradient-based node importance—sensitivities of the leak logit to node features—mapped to pipes via linear interpolation to produce leakage-risk contour maps. The framework couples pressure and demand features, accommodates partial monitoring by varying sensor density, and evaluates noise robustness by injecting Gaussian noise into inputs. In simulated WDNs, the approach maintains high accuracy across sensor-density settings and exhibits a predictable trend under increasing noise, where false negative rate rises faster than false positive rate as signal-to-noise ratio decreases. On the L-Town benchmark, we use real pressure measurements and nodal demand data consistent with metered consumption (generally <10% deviation), providing a realistic testbed. A simplified network representation accelerates convergence while preserving localization fidelity, enabling millisecond-level inference suitable for operational deployment. Together, these results support a physically grounded, computationally efficient pathway for leak detection and localization in smart water networks.