Leak detection in water distribution networks using micro-electromechanical systems-based accelerometers: a machine learning approach

Water distribution networks (WDNs) worldwide are plagued with water losses. These losses cause massive financial damage of US$39 billion per year and further aggravate the existing water scarcity situation faced by half a billion people throughout the year. To curb the damage, our research provides a machine learning-based approach for early detection of leaks in WDNs, constituting a substantial portion of water losses in most WDNs. Experiments were conducted for more than 10 months on real networks using micro-electromechanical system (MEMS) accelerometers. Leak and no-leak signals were collected from metal and non-metal pipes of different sizes. Seventeen time-domain and frequency-domain-based features were extracted using signal processing methods. The most appropriate features were ranked and selected. The selected features were then used to develop an artificial neural network (ANN) and gene expression programming (GEP) based on intelligent models for metal and non-metal pipes. Both ANN and GEP models showed an accuracy of over 99% for leak detection in metal pipes. In contrast, the accuracy in non-metal pipes reached around 89%. Our study represents one of the very few attempts made on leak detection in real WDNs, which will open new avenues of research in this domain.