Evaluation and Detection of Leaks in a Laboratory-Scale Water Distribution System with Acoustic, Acceleration, and Dynamic Pressure Sensors

Abstract

Aging water distribution systems waste millions of gallons of treated water due to background leaks. In this study, a laboratory-scale water network with 7.5 m × 5 m dimensions was developed to simulate background leaks in networks with looped and branched architectures. Four types of leaks, orifice, longitudinal and circumferential cracks, and gasket, were induced in the test system to generate leak signals. Six sensors, including two hydrophones, two dynamic pressure sensors, and two accelerometers, were employed to measure testbed parameters. With induced leak rates less than thirty percent, sixteen plots and numerical features were employed to assess the leak and network changes’ effects on measured data. Due to the inconsistent patterns and similar magnitudes of the plots and features, the sixteen evaluation criteria did not represent specific patterns, and the metrics’ changes depended on the sensors’ locations. Based on the information extent they represented to differentiate leaks and network architecture, the sensors ranked as (1) dynamic pressure sensor, (2) hydrophone, and (3) accelerometer. Hydrophone acoustic signals were employed to detect leaks using five shallow classifiers, including Support Vector Machines (SVM), one-class SVM (1CSVM), Isolation Forest (iForest), Extreme Gradient Boosting (XGBoost), and Local Outlier Factor (LOF). A wavelet transform was applied to raw signals to compute the wavelet coefficients' moduli and create a matrix. A subsampled feature matrix of the looped network was used to generate imbalanced training and test datasets with imbalanced leak and non-leak class ratios. Testing the classifiers on the looped network’s imbalanced data with original features showed SVM and XGBoost ranked first in predicting leak and non-leak samples, respectively. Using the looped network’s imbalanced data with reduced features showed the same algorithms' ranks but with lower F1-scores for all algorithms. Evaluation of the branched network’s acoustic imbalanced data with original and reduced dimensions indicated more mixed data distributions, and lower F1-scores than the looped network. The analysis of the looped network’s balanced data with original and reduced features resulted in higher F1-scores of the algorithms in detecting leaks than their counterparts using imbalanced datasets.