An Interpretable Neural Network Algorithm for Leaking Detection in the Urban Water and Sewer Pipeline Network, Tianjin, China

Urban residents' daily lives and commerce depend on the reliable water and sewer pipeline network. Leaking the network is a nuisance, wasting precocious resources and money. Detecting and mitigating a leak in the network is essential for utility companies and the general public. As an ongoing study, we developed a U-net-based algorithm for leak detection and explored the possibility of understanding operations within the algorithm. As dimensionality reduction and feature extraction are the primary objectives of the convoluting and pooling processes in the algorithm, the processes can be carried out by principal components analysis (PCA). The algorithm then advances a leak/non-leak classification after extracting features. The support vector machine (SVM), one machine learning algorithm, can replace and perform the classification procedure. Thus, a (PCA+SVM) algorithm is developed, and more importantly, we interpret the studied U-net-based algorithm as a hybrid of the PCA and SVM. Finally, the (PCA+SVM) algorithm is evaluated in the leaking detection, and it satisfactorily detects a leak or non-leak location in urban areas of Tianjin, China.