Detecting Leaks in Water Distribution Pipes Using a Deep Autoencoder and Hydroacoustic Spectrograms
Roya Cody, Bryan A. Tolson, Jeff Orchard
Abstract
Small leaks in buried water distribution pipelines typically remain undetected indefinitely because the impact small leaks have on the overall system pressure is imperceptible. This difficulty is caused by a combination of the leak’s magnitude and the demand variability within water distribution networks (WDNs). Deep learning has the potential to disentangle these sources of variability more capably than traditional heuristics. This paper applies deep learning to acoustic monitoring data to detect leaks. Due to the lack of leak data in practice, a semisupervised approach was proposed. In this approach, a convolutional neural network is combined with a variational autoencoder to detect anomalies in a laboratory test bed. The test bed used is connected to the municipal water system via a service line, thus ensuring realistic baseline variation. The baseline case is defined by the test bed’s typical operating conditions when no leak is present. The proposed method achieved an accuracy of 97.2% for detecting a 0.25 L/s leak, demonstrating the effectiveness of the deep autoencoder for leak detection in WDNs.