Machine Condition Diagnosis Using Deep Learning and Gabor-Transformed Motor Current Signatures

This study presents a novel approach for diagnosing machine conditions using deep learning models applied to Gabor-transformed motor current signatures. The dataset comprised signals from five single-phase 2-HP induction motors with five fault conditions, each tested under five different loading conditions. These signals were transformed into 2D time-frequency plots using the Gabor Transform (GT), capturing essential frequency changes over time. A convolutional neural network (CNN) model was then employed to classify motor faults based on these plots. Methodologically, the CNN was trained and validated across various GT configurations, including different image resolutions and color modes. Performance was measured using a 10-fold stratified cross-validation. The results demonstrated significant improvements, with the best GT configuration achieving an average accuracy of 98.43%, substantially higher than the traditional LightGBM method's 93.20% accuracy. The study underscores the efficacy of combining GT with CNN for enhanced fault detection, providing a robust alternative to existing machine learning and deep learning methodologies,