Deep Learning Based Acoustic Leak Detection and Classification in Water Distribution Systems Using Hydrophone Sensors

Leakage detection in water distribution systems is of utmost importance for water resource management. This study developed a deep learning (DL) framework for acoustic leak detection using hydrophone sensors. The data was collected from a controlled water testbed simulating both looped and branched pipeline topologies under five conditions: no-leak, longitudinal crack, circumferential crack, orifice leak, and gasket leak. The feature extraction from the raw acoustic signals sampled at 8 kHz followed two approaches: extraction of statistical features for traditional machine learning (ML), and generation of Mel-Frequency Cepstral Coefficient (MFCC) spectrograms using sliding window segmentation suitable for DL models. Three DL models were trained on the MFCC image data for both binary and multi classification. These models achieved excellent classification performance ranging from 94.77 to 100% across all network configurations. The study reveals that by transforming raw data into 2D spectrograms, the models better learned representative features and outperformed ML models trained with statistical features.