Enhancing Automated Acoustic Leak Detection in a Water Distribution Network Using Ensemble Machine Learning

The water distribution network (WDN) constitutes a pivotal and valuable asset in municipal infrastructure, yet aging pipelines often contend with leaks. Wireless noise loggers have emerged as a promising technology for early leak detection in WDNs. However, prevalent high false alarms within these systems impede resource efficiency and hinder urban infrastructure development. Prior noise logger models, derived from lab-scale experiments, cast doubt on their accuracy in real networks, contributing to elevated false alarm rates. This paper endeavors to introduce an innovative leak detection model by harnessing data from acoustic noise loggers deployed in a Hong Kong WDN. Utilizing authentic acoustic data and augmenting features for model development, along with employing a multiclassifier ensemble learning algorithm, constitute the principal contributions of this research. Fourier transform was applied to analyze sound signals, and convolutional neural network (CNN), naïve Bayes (NB), support vector machine (SVM), and decision tree (DT) algorithms were employed to develop four distinct models. An ensemble technique effectively amalgamated the strengths of the top three models, resulting in an impressive accuracy of 99.38%. The outcomes underscore the potential of noise logger-based models for real-time monitoring, offering a reduced false alarm solution. Leak detection companies stand to benefit significantly from this model because it harnesses machine learning techniques for precise leak detection in WDNs. Moreover, municipalities equipped with leak detection sensors can leverage this model to optimize their systems, contributing to improved urban infrastructure development, particularly in densely populated cities.