Towards Intelligent Leak Detection: Exploring AI Techniques to Identify Leaks in Water Distribution Systems

Water losses due to leaks are a persistent sustainability and economic challenge for water utilities, yet comparing AI-based leak detection methods remains difficult because studies often rely on different networks, data assumptions, and evaluation setups. This thesis addresses that gap by benchmarking data-driven leak detection approaches on the L-Town reference network under a consistent, reproducible evaluation framework focused on alarm quality. Using pressure time series from multiple sensors, the work develops a supervised learning pipeline that evaluates models through both threshold-free metrics (ROC/PR behavior) and operational, thresholded performance (false alarms, missed detections, and alarm timing), complemented by node-wise diagnostics and time-series overlays to interpret how alarms behave over long periods. A central result is that problem formulation strongly influences detection realism: treating the network as a single global label can lead to unstable behavior and extended false-alarm regimes, while sensor-resolved formulations better capture the heterogeneous detectability of leaks across the network. Overall, the thesis provides an empirical comparison of deep learning baselines for leak detection, highlights the trade-off between sensitivity and false-alarm control, and delivers practical guidance for designing robust detection pipelines that are interpretable and more aligned with utility operations.