Improving the leak detection efficiency in water distribution networks using noise loggers

Leak detection techniques are effective ways of controlling water leakage in real water distribution networks (WDNs). Nevertheless, developing detection techniques for real WDNs has received little attention compared to the detection models developed based on laboratory simulated leaks. On the other hand, ambient noises and irregular water usage are difficult to simulate in a laboratory environment so detection models based on the laboratory simulated leaks are usually of low efficiency in practical applications. To achieve a better understanding of the detection models of real WDNs, machine learning (ML)-based leak detection models were developed in this work. This study employs wireless sensors to record acoustic signals emitted by real WDNs for the development of the leak detection models. The acquired acoustic signals are de-noised using the discrete wavelet transform. Thereafter, seventeen features are extracted from both the raw and de-noised signals using the principle of linear prediction, and the features are subsequently used for the development of the ML-based leak detection models. A thorough comparison is made for the performances of the detection models in terms of metal and non-metal WDNs, different features, and different ML algorithms, namely decision tree (DT), support vector machine (SVM), artificial neural network (ANN), and k-nearest neighbor (K-NN). Generally, the performance of the ML-based detection models developed by using the features extracted from de-noised signals has a better classification accuracy as compared to the performance of the models developed based on the features extracted from raw signals. For the de-noised signals, the accuracy, precision, and recall for the models developed based on the DT, SVM, and ANN algorithms are 100% for metal and non-metal WDNs.