Features Extraction and Selection with the Scattering Transform for Electrical Load Classification

The Scattering Transform (ST) presents itself as an alternative approach to the classic methods that involve neural networks and deep learning techniques for the feature extraction and classification of electrical signals. Among its main advantages, one can emphasize that the coefficients of the ST are determined analytically and do not need to be learned, as typically performed in Convolutional Neural Networks (CNNs). Additionally, ST has time-shifting and small time-warping invariance, which reduces the need for precise temporal localization (detection) for subsequent classification. This paper originally proposes six feature extraction and selection methods applied to classification of Non-intrusive Load Monitoring (NILM) high-frequency signals. We visually analyze the separability among classes for the proposed Feature Extractors and validate the performance of the proposed methods varying several parameters for ST calculation, such as signal length, number of examples, and sampling frequency. The results outperform other state-of-the-art feature extraction techniques, reaching up to 100% of FScore for a publicly available dataset, demonstrating the feasibility and promising aspects of the ST for NILM problems.