Leak localization in water distribution networks using classifiers with cosenoidal features

Abstract

This paper presents a leak localization approach for water distribution networks using classifiers with pressure residuals as input features. This approach is based on applying a non-linear transformation to the residuals of the node pressures to increase the separability of the leak classes. The transformed features can be interpreted as the direction cosines in the subspace spanned by the residuals of the measured pressures. In order to illustrate the method, different tests were performed with MATLAB® applying four different classification algorithms on a synthetic dataset obtained from an EPANET model of the Hanoi network. Then, by considering the cosenoidal features, a significant improvement in the leak location error was achieved. In this way, the leak location error decreases by more than 97% compared to the use of residual features when accurate measurements are used, and about 50% when noisy measurements with 60 dB SNR are used.