Parametric Analysis of the Wavelet Scattering Transform for Water Leakage Classification and Comparison with Manual Feature Extraction Techniques
Academic Article in Scopus
Water pipeline leakages create significant environmental challenges worldwide. Machine learning (ML) is widely used for leakage detection, often based on time and frequency features. However, feature selection is time-consuming and challenging for both binary (leak/no leak) and multi-class (leak type) classification. The Wavelet Scattering Transform (WST) provides translation-invariant, deformation-stable features automatically, but the effect of parameter selection on classifier performance remains underexplored. This study investigates how key WST parameters, specifically the number of wavelets per octave and the invariance scale, influence the performance of a Support Vector Machine (SVM) classifier using acceleration, pressure, and hydrophone data from a benchmarking laboratory dataset. Manual features were also extracted, selected with the minimum redundancy maximum relevance algorithm, and evaluated using SVM, Decision Tree (DT), and K-Nearest Neighbors classifiers. Optimizing WST parameters improved accuracy by 9.65% over prior work, reaching 99.16% with acceleration data in branched topology using 2 wavelets per octave and an invariance scale of 0.04. In contrast, manual features peaked at 82.5% with DT and 15 features. Classifier performance also varied with feature quantity, averaging 8.42% across modalities. Overall, WST-based features outperform manual extraction for leakage classification, supporting more robust data-driven leak detection systems.
Water pipeline leakages create significant environmental challenges worldwide. Machine learning (ML) is widely used for leakage detection, often based on time and frequency features. However, feature selection is time-consuming and challenging for both binary (leak/no leak) and multi-class (leak type) classification. The Wavelet Scattering Transform (WST) provides translation-invariant, deformation-stable features automatically, but the effect of parameter selection on classifier performance remains underexplored. This study investigates how key WST parameters, specifically the number of wavelets per octave and the invariance scale, influence the performance of a Support Vector Machine (SVM) classifier using acceleration, pressure, and hydrophone data from a benchmarking laboratory dataset. Manual features were also extracted, selected with the minimum redundancy maximum relevance algorithm, and evaluated using SVM, Decision Tree (DT), and Random Forest classifiers. Optimizing WST parameters improved accuracy by 9.65% over prior work, reaching 99.16% with acceleration data in branched topology using 2 wavelets per octave and an invariance scale of 0.04. In contrast, manual features peaked at 82.5% with DT and 15 features. Classifier performance also varied with feature quantity, averaging 8.42% across modalities. Overall, WST-based features outperform manual extraction for leakage classification, supporting more robust data-driven leak detection systems.
