Scalable AI-Assisted Metering Architecture for Continuous Leakage Monitoring and Fault Diagnosis in Urban Water Systems

Abstract

This paper proposes a scalable AI framework. The framework combines high-frequency flow and pressure data from the smart meters with a distributed edge-computing layer, where lightweight machine learning models are applied for performing initial anomaly detection in real time. Detected events then trigger an adaptive neural network, which refines the detection and estimates the leak characteristics more accurately. A pilot implementation in a mid-sized Indian city shows the capability of the system to localize leaks with a spatial resolution of better than 50 m and detect small leaks (< 2 L/min) within 30 minutes of occurrence. We use a digital copy (twin) of the distribution network, built from the available pipe layouts, nodal elevations, and historical demand patterns, as the data source for training a supervised-learning algorithm to overcome the lack of labelled leak events. By conducting a series of field trials, we provide quantitative performance results in terms of the true/false positive rates, localization error, and energy overhead of edge processing. A cost-benefit analysis is conducted that evaluates deployment scenarios at different penetration rates of smart meters and shows payback periods below three years for networks with > 60%-meter coverage. Finally, we discuss practical considerations for large-scale deployment, including integration with existing water utility management systems, scalability challenges, data privacy and security concerns, maintenance of edge-computing devices, and potential regulatory and policy implications. The results demonstrate that the proposed framework is not only technically feasible but also economically viable, providing utilities with a robust tool for improving leak detection, reducing water losses, and enhancing operational efficiency.