Hybrid Learning Framework for Motor Fault Diagnosis via Optimized Wavelet-Based Time–Frequency Analysis and Deep Feature Classification

Fault detection and classification in electrical motors are essential for condition monitoring and predictive maintenance. This study proposes a hybrid learning framework that combines genetic algorithm–optimized Morse wavelet–based time–frequency analysis with machine learning to investigate their effectiveness in diagnosing single and simultaneous combined faults in electrical motors under different load and speed conditions. One-dimensional signals are segmented using a centrally expanding sliding window strategy and transformed into two-dimensional representations through a Morse-based continuous wavelet transform, where spectral concentration metrics are used to guide fault-specific parameter optimization. Deep features are extracted using a pre-trained ResNet50 backbone, and fault classification is performed using a support vector machine. The proposed framework is initially evaluated on the Case Western Reserve University dataset using ten-fold cross-validation. The results indicate reliable diagnostic behavior under clean conditions and maintain moderate performance in the presence of additive white Gaussian noise. To further examine generalization capability, the proposed approach is evaluated on a custom experimental motor acoustic dataset covering single and combined fault scenarios under varying operating conditions. The observations suggest that feature-based learning strategies can provide a practical balance between representational capacity and robustness when operating conditions differ between training and evaluation. Comparative analyses further indicate that selective fine-tuning of a limited portion of a deep backbone, combined with feature-based classification, offers a more stable and adaptable alternative to fully end-to-end training in data-constrained and cross-domain scenarios.