An Efficient Data-Driven Leak Detection Strategy by Enhancing a Convolutional Neural Network Approach Using a Gaussian Process Regressor

The reliable detection and localization of leaks is a priority for the good management of a Water Distribution Network (WDN). This paper presents a fully data-driven leak detection strategy that enhances a Convolutional Neural Network (CNN) based method by adding a new denoising and detrending procedure for steady-pressure raw data, which is more efficient than those used in other CNN-based methods. The data pipeline has four novel steps: a data pretreatment with a 'Gaussian Process Regressor', which enhances the features induced by leakages; a normalization step, which equalizes the dynamics of the sensors; an application of a CNN Overcomplete Autoencoder exploiting an L1 Activity Regularization, which further amplifies the leak-related features; and, finally, an analysis of the CNN errors. The benchmarks are performed on two publicly available synthetic datasets relative to a District Metering Area (DMA) of urban Hanoi's WDN. The first one contains 500 scenarios, and the second one includes 1000 different, more complex scenarios. With the first dataset, we achieve a leak detection accuracy of 92.80%, a true positive rate of ~91.05%, a false positive rate of ~1.66%, and a true negative rate of ~98.33%. These results are validated by the accuracy of 92.20% obtained with the second dataset. Our method proved to be valuable to highlight, and thus to better detect than other CNN methods, the anomalies within steady-pressure signals caused by WDN leakages. Moreover, this good performance should enable us to increase the distance between sensors, consequently reducing their number and the associated costs.