Explainable wavelet-scalogram learning for quasi-stationary faults in automotive DC motors using AERIS-Wave

Noise diagnostics in automotive brushed DC motors are crucial for ensuring reliability and compliance with strict manufacturing standards. Traditional human-based inspection methods are often subjective and inconsistent, necessitating automated, data-driven solutions. This study introduces an explainable wavelet-based classification framework for diagnosing quasi-stationary motor faults using accelerometer data. The framework employs the Continuous Wavelet Transform (CWT) to extract rich time–frequency features, including scalograms, ridge trajectories, periodograms, and statistical descriptors, serving as discriminative representations of mechanical behavior. These features are evaluated using both classical machine learning algorithms (Random Forest, XGBoost) and deep learning architectures such as ResNet18, VGG16, CNN-LSTM, CNN-GRU, and WaveNet. Building upon these baselines, the proposed AERIS-Wave (Attention-Enhanced Residual Interpretable Scalogram Network) integrates multi-level attention, LayerScale normalization, and explainable-AI components (Integrated Gradients, Grad-CAM, SHAP) to visualize spectral contributions that drive decisions. Experimental results show that AERIS-Wave achieves 97.72% accuracy and an AUC of 0.9991, surpassing all benchmark models, including WaveNet and ResNet18. The findings confirm that wavelet-based representations, combined with interpretable deep learning, enable high-precision, explainable, and scalable fault classification suitable for real-time quality control in industrial environments.