Water Leakage Classification With Acceleration, Pressure, and Acoustic Data: Leveraging the Wavelet Scattering Transform, Unimodal Classifiers, and Late Fusion

Early detection of water leakages is crucial due to their social, environmental, and economic impacts. In this regard, machine learning (ML) algorithms have been proposed in the literature to automatically detect leakages using vibration, pressure, and acoustic data. However, these data modalities are often used independently and under heterogeneous conditions. ML techniques typically involve feature extraction through the time and frequency domain, but the computation of these features varies among studies. On the other hand, deep learning (DL) techniques can automatically extract features from data but require large sample sizes and training times, and their complexity is higher than linear classifiers. This paper proposes the wavelet scattering transform (WST) as a feature extraction technique for hydrophone, vibration, and dynamic pressure data to compare their performance for classifying water leakages in looped and branched water networks. Late Fusion (LF) was also used to assess the effectiveness of simultaneously employing vibration, hydrophone, and dynamic pressure data to classify water leakages. The results indicate that the WST of accelerometer data and a support vector machine (SVM) performed best in classifying water leakages compared to using the WST of dynamic pressure and hydrophone data, with an accuracy of 89.5833% for the looped water network and 96.6667% for the branched water network. The LF model generated by combining the predictions of the SVMs of each data modality achieved a 93.33% accuracy for the looped water network. In contrast, the LF model of the branched water network attained an accuracy of 95.833%.