This paper presents a framework based on hydraulic simulation and machine learning for supporting Water Distribution Network (WDN) managers in localizing leakages, while reducing time and costs for investigation, intervention and rehabilitation. As a first step, hydraulic simulation is used to run different leakage scenarios by introducing a leak on each pipe, in turn, and varying its severity. As output of each scenario run, pressure and flow variations in correspondence of the actual monitoring points into the WDN, and with respect to the faultless model, are stored. Scenarios clustering is aimed at grouping together leaks generating similar effects, in terms of observable pressure and flow variations. This analysis is performed by creating a similarity graph, where nodes are scenarios and edges are weighted by the similarity between pairs of scenarios. Spectral clustering, a graph-clustering technique, is here proposed according to its usually higher performances with respect to traditional data-points clustering. Then each scenario is labeled with its cluster by obtaining a labeled dataset on which a Support Vector Machine (SVM) with RBF-kernel is trained. When an actual leak is detected, the variations in measured pressure and flow with respect to the faultless hydraulic model are given as input to the trained SVM which assigns them to a specific cluster, whose corresponding pipes are provided as the hydraulic components to check for leakage. Since spectral clustering induces a non-linear transformation, from Input Space (i.e., pressure and flow variations) to Feature Space (i.e., most relevant eigen-vectors) where clusters are obtained, the SVM encodes the non-linear relationship of pressure and flow variations with the scenarios cluster. The SVM is able to remap efficiently the results from spectral clustering toward the Input Space giving the probably leaky pipes even for pressure and flow variations not included in the simulated leakage scenarios.
Candelieri, Antonio; Conti, Dante; Soldi, Davide; and Archetti, Francesco, "Spectral Clustering And Support Vector Classification For Localizing Leakages In Water Distribution Networks – The ICeWater Project Approach" (2014). CUNY Academic Works.