IoT-Driven Water Leak Detection: A Comparative Study of Machine Learning and Deep Learning Models for Time-Series Sensor Data with Explainable AI

Urban water distribution systems have been a major threat to the preservation of world freshwater resources due to inefficiencies as large amounts of treated water end up being wasted each year by leakages and bursting pipes. Conventional leak detection techniques frequently encounter difficulties, including a significant dependence on the human selection and black box approaches, which can be labor-intensive and less efficient. This paper will manage this obstacle by suggesting an effective, evidence-based framework of early and accurate detection of pipeline failures. The paradigm of Internet of Things is proposed to gather a multivariate time-series data which is used for multi-classification to make a distinction between the normal state, leak state, and burst state. With a systematic comparative study assessing the performance of six light-weight classical Machine Learning (ML) algorithms against a Deep Learning (DL), experiment results show that Random Forest algorithm exhibit optimal performance, reaching a test accuracy and an F1-score of 0.99, which is able to cope with the enormous class imbalance of the data. To overcome the traditional black-box constraint of AI models with great performance and develop operational trust, this study incorporates an overall XAI framework. The results of this XAI analysis are clear and practical, and quantitative demonstrating that the pressure is the most significant predictive variable. The paper finds that a synergistic combination of high-accuracy tree-based model with XAI approach is a viable, and operationally feasible way forward towards integrating into the proactive management of critical water infrastructure, thus adding to the achievement of improved urban water security.