Mitigating Representation-Dependent Degradation in Acoustic Signal-Based Fault Diagnosis of BLDC Motors via Classifier-Independent Waveform Restoration

Industrial noise can degrade acoustic fault diagnosis differently across input representations. This study evaluated whether classifier-independent waveform restoration could reduce this disparity in BLDC motor diagnosis. A Wave-U-Net denoiser was trained without classifier losses or outputs and applied before generating time-domain waveform and STFT magnitude inputs for six CNN, LSTM, and CNN–LSTM classifiers, which remained unchanged. The dataset comprised 13,790 one-second acoustic segments recorded from a single BLDC motor under normal operating conditions and bearing, eccentricity, and electrical fault conditions. Industrial noise environments were simulated by mixing the recorded motor signals with external noise recordings from the MIMII dataset. At 0 dB, restoration increased mean classification accuracy from 69.27% to 97.37% and reduced the mean performance gap between the two input representations from 48.61 to 1.01 percentage points. Improvements were also observed at evaluated SNRs below the training range and under noise types excluded from denoiser training. Grad-CAM and feature-space analyses indicated that restoration shifted decision patterns and class-wise feature structures toward those observed without added industrial noise. These findings support waveform restoration as a common preprocessing approach for reducing representation-dependent degradation under the evaluated conditions.