Water distribution networks are prone to water loss, which increases operational costs and public health risks, as well as exacerbates water scarcity. Traditional leakage detection methods are dependent upon the experience of the operator and cannot be automated. Although software-based leakage detection methods exist that offer real-time monitoring of WDNs, they each have drawbacks that hinder their practical application. The goal of researchers Ella Steins, Nick Langer, Johannes Koslowski, and Andrea Cominola was to improve the leakage identification and localization algorithm to better suit real-world applications by better incorporating operators’ experience. The result of their study, “From Theory to Practice: Robust and Automated Data-Driven Leakage Detection in Water Distribution Networks,” was to add two key features to the algorithm: automated retraining times and sensor selection and a robust cumulative sum chart suitable for noisy and non-normal data.
The limitations of the original leak identification module of LILA included its dependence on having prior knowledge of the leakage location as well as requiring manual sensor selection and a training period. The enhanced algorithm reduced the reliance on manual tuning by experts, improved its reliability under realistic operating conditions, and decreased the number of false alarms. Improved efficiencies include fewer resources wasted on false alarms, reduced manual effort that goes into detecting and locating leaks, and minimized service disruptions. Learn more about this research and how you can apply these lessons to automating manual process and refining algorithms to streamline operations in the Journal of Water Resources Planning and Management at https://ascelibrary.org/doi/epdf/10.1061/JWRMD5.WRENG-7401. The abstract is below.
Abstract
Reliable leak detection in water distribution networks is essential for reducing water losses, costs, and operational risks. Here we present an improved semisupervised pressure-based algorithm that enables near-real-time leak detection through two key features: (1) automation of sensor selection and retraining times, avoiding manual tuning and prior knowledge of leak conditions, and (2) an adaptive, nonparametric cumulative sum (CUSUM) with a self-starting scheme, robust to noisy and non-normal data. Tested on a real-world network data set with synthetic nonoverlapping leaks, the algorithm detects all leakages without false alarms, demonstrating strong potential for practical deployment.
Learn more about better detecting water leaks in near-real time by updating the software’s algorithm in the ASCE Library: https://ascelibrary.org/doi/epdf/10.1061/JWRMD5.WRENG-7401.