Leakage detection in water distribution networks via 1D CNN deep autoencoder for multivariate SCADA data

Abstract

Leakages are undesirable events in water distribution networks (WDNs) with mostly unknown incident time making it extremely difficult to accurately label leaks in distribution systems. This study proposes a one-dimensional convolutional neural network (1D CNN) deep autoencoder (AE) that is trained in a semi-supervised fashion to detect and localize leaks using multivariate timeseries data in a bid to mitigate the adverse impact of random noise. A leakage identification framework based on Pruned Exact Linear Time (PELT) multiple change point detector and alarm correlation is proposed. Additionally, a novel localization framework is proposed based on the relative percentage deviation of 1D CNN deep AE reconstruction error of pressure sensors in a district metering area (DMA). The leak localization framework outputs a sensor priority list based on the proximity of the pressure sensor to the leaky pipe thereby significantly reducing the search area for a leakage. The proposed leakage identification and localization framework is validated on a benchmark WDN, the L-TOWN. The results indicate the ability of the proposed method to identify 16 of the 19 leaks in 2019 and accurately localize 13 of the 16 identified leaks within the maximum allowable leak search radius of 300 m without recourse to a hydraulic model. Minimal identification lag was recorded for abrupt leakages whiles incipient leaks require longer identification time due to gradual leak growth rate.