Leak detection in real water distribution networks based on acoustic emission and machine learning

Water scarcity as well as social and economic damages caused by the increasing amounts of non-revenue water in the water distribution networks (WDNs) have been prompting innovative solutions. A great deal of potable water is wasted due to leakage in the WDNs all over the world. Hence, various leak detection approaches have been explored, including the promising application of acoustic devices. Exploiting the benefits of technological advances in acoustic devices, signal processing, and machine learning (ML), this study aimed to develop a sophisticated system for leak detection in WDNs. Different from laboratory-based studies, this study was conducted on real WDNs in Hong Kong and lasted for about two years. Utilizing acoustic emissions acquired using wireless noise loggers, various ML algorithms were explored to develop inspection models for in-service and buried WDNs. ML classification algorithms can identify patterns in the acquired signals for leak and no-leak statuses. Thus, a combination of features describing acoustic signals in time and frequency domains was utilized to facilitate the development of ML models. Separately for metal and non-metal WDNs, ten well-known ML algorithms were used to develop leak detection models. The validation results demonstrate the promising application of noise loggers and ML for leak detection in real WDNs. Support Vector Machine (SVM), Artificial Neural Network (ANN), and Deep Learning (DL) leak detection models demonstrated a largely stable performance and a very good accuracy, particularly for new unlabelled cases.

Automated summary

This study developed a sophisticated system for leak detection in water distribution networks (WDNs) using acoustic devices and machine learning. After conducting a two-year study on real WDNs in Hong Kong, different machine learning algorithms were used to develop inspection models for in-service and buried WDNs based on acoustic emissions acquired using wireless noise loggers. The results demonstrated the promising application of noise loggers and machine learning for leak detection in real WDNs.