Detecting water leakages is crucial to prevent water loss, scarcity, and contamination. Early detection is challenging due to dynamic water consumption, sensor uncertainty, and underground pipe networks. Machine learning has been widely applied for leak detection using two main approaches: feature extraction and deep learning. Feature extraction requires expert knowledge or trial and error, while deep learning demands large datasets, extensive training, and high computational resources. The wavelet scattering transform offers features for classification by producing a time-shift invariant and deformation-stable signal representation, reducing within-class variance. However, it can generate numerous coefficients, complicating classifiers, and lacks established parameter-setting guidelines. This work uses vibration data transformed via the wavelet scattering transform and selects features for water leakage classification in looped and branched networks. Two feature selection methods, oneway analysis of variance and minimum redundancy maximum relevance, were compared. The ten most important features from each method were trained on a support vector machine. Results show that variance-based selection yields higher accuracy for the looped network (88.75 percent), while both methods achieve 96.25 percent for the branched network.
