Leak Detection and Localization in Water Distribution Networks

Abstract

Leaks in water distribution networks remain a significant challenge for utilities, resulting in substantial economic and environmental losses and health risks. However, existing leak detection and localization approaches face several shortcomings, including (i) limited understanding of how algorithms generalize across different networks, (ii) limited adoptability of empirical characterization of dispersive wave behavior in water-filled pipes, and (iii) heavy dependence on cross-correlation methods when performing leak localization, failing if leaks are not located on a direct sensor-to-sensor path.

This thesis addresses these gaps using machine learning-driven leak detection and localization techniques using hydrophone time-series data. First, I introduce structured frameworks for leak detection and leak localization algorithms, which define the key processing stages from signal collection to post-processing. To evaluate the ability of leak detection algorithms to generalize across different networks, I present a novel leak detection dataset collected from three real-world water distribution networks, and propose two evaluation schemes - Cross-Domain F1 Scoring and Multi-Domain F1 Scoring. Using these schemes, over 33,000 leak detection models were evaluated by varying modeling parameters, revealing that certain transformation techniques and low-frequency energy-based features (e.g., 62–124 Hz energy vs. 0–500 Hz centroid) can yield up to a 37% higher mean cross-domain F1 score. Further, I found that when sufficient training data are available, convolutional neural networks generalize better than hand-crafted-feature-based algorithms, achieving a multi-domain F1 score of 0.87 compared to 0.72 for exhaustive feature selection and 0.50 for simple feature selection when eight unique leak scenarios were included in the training data.

Next, I characterize wave propagation in a controlled lab-scale system and experimentally demonstrate dispersive shell-borne surface waves traveling at approximately 291 m/s, waterborne plane waves at 350 m/s, and high-velocity ultrasonic waves traveling at approximately 1,300 m/s. I show that analytical models that predict wave speed can be inaccurate by up to 16%, and that waves traveling along the shell wall exhibit dispersive behavior, which poses problems for traditional cross-correlation-based leak localization methods.

The viscothermal wave equation is implemented using the finite difference method to explore how spectral features correlate with leak proximity. These findings motivate the use of spectral features such as energy and centroid for predicting leak proximity. I then propose a novel leak localization algorithm that produces a heat map describing the probability of leakage along each point in a pipe network. The algorithm achieves reliable leak localization results, even in leak scenarios where conventional cross-correlation cannot be used. Calibration is shown to improve leak proximity regression performance by more than 53%, and the approach reliably localizes leaks within 3.66 m across leak scenarios not included in its training data, even in scenarios where traditional cross-correlation-based methods cannot be used.

Overall, my thesis contributes new datasets, quantitative evaluation methods, numerical modeling, insights into wave behavior, and learning-based algorithms that together advance the development of deployable and generalizable leak detection and localization systems.