Fault Classification Using Machine Learning with Deep Learning-Based Scalogram Wavelet Feature Extraction and Metaheuristic Feature Selection

Electrical fault classification is one of the most complex tasks in electrical systems. In this paper, we propose a classification model based on scalograms using the Continuous Wavelet Transform (CWT) and feature extraction using the EfficientNetV2B3 backbone. Features are then selected using the hybrid metaheuristic algorithm GWO-WOA to maximize the multi-objective function of precision and recall for training a Quadratic Discriminant Analysis (QDA) model. The dataset was generated from a three-phase electrical model in Matlab/Simulink, with measurements of currents (Ia, Ib, Ic) and voltages (Va, Vb, Vc). CWT was used to obtain scalograms for each signal, producing a total of ${6, 4 8 0}$ RGB-type images. The results indicate that the hybrid GWO-WOA algorithm maximizes the performance of the QDA model trained with the selected features, achieving an accuracy of ${9 4 \%}$, a precision of ${9 4 \%}$, and a recall of ${9 4 \%}$. The results for each class indicate an F1-score above ${9 1 \%}$.