Machine Learning Schemes for Leak Detection in IoT-enabled Water Transmission System

Reducing Non-Revenue Water (NRW) losses in water transmission and distribution networks is a critical challenge for water utility companies. The combination of unobtrusive Internet of Things (IoT) monitoring devices and Artificial Intelligence (AI) technology is one of the most promising directions in water leak detection techniques for industrial scale infrastructure and smart cities. Currently, the complicated network topology and underground nature of transmission and distribution water pipelines pose serious limitations for the effective elimination of associated water leaks. In this paper, a realistically dimensioned IoT-enabled water transmission system provides the basis for a series of simulated leak experiments and the subsequent application of three different anomaly detection schemes. Having full control over the mechanical valve behind the simulated leaks, this test rig addresses the issue of accurate labelling in leak data and serves as testbed for the evaluation of each anomaly detection method and the comparison between them. The first anomaly detection method is the unsupervised multi-variate classification known as Isolation Forest (iForest). Second, the Support Vector Classification (SVC) approach is implemented representing supervised classification methods in the Support Vector Machine (SVM) family. Finally, a deep learning RNN-LSTM (Recurrent Neural Networks-Long Short-Term Memory) model is used in conjunction with a single threshold to signify anomalies due to high deviations between actual and forecasted values of key infield sensors. These models can detect water leaks and the results provide insights regarding both the effective applicability of sensors and Machine Learning (ML) algorithms in this context.