Leak detection in water distribution networks using deep learning

ABSTRACT Two deep learning algorithms, namely, Convolutional Neural Network (CNN) and Recurrent Neural Network- Long Short Term Memory (LSTM), were used to classify the water distribution networks (WDN) as leaky or non-leaky. LeakDB dataset was employed to generate different leakage scenarios for Net 1 and Hanoi benchmark WDN. Three cases, (a) incipient leaks, (b) abrupt leaks, and (c) mixed leak situations, are employed for pressure and flow conditions. A total of 1000 scenarios have been employed, 80% for training and 20% for testing. Seven metrics for analyzing the performance of CNN and LSTM are training accuracy, testing accuracy, total accuracy, true positive rate, false positive rate, false negative rate & area under curve. The results obtained are compared with those of Kammoun, et al. (2021). CNN is performing slightly better than LSTM in several metrics for most scenarios. However, both CNN and LSTM performed most of the time with better accuracy than those used by Kammoun et al. (2021). Leak detection accuracy is in the range of 90.56–98.23 % for Net1 WDN, whereas it is 49–96.55 % for Hanoi WDN.