Auditory perception–based acoustic fault diagnosis for railway wheelset bearings under strong noise: a comparison with short-time Fourier transform, Mel spectrogram, and continuous wavelet transform

Abstract In the complex environment of train operations, the condition recognition of wheelset bearings is significantly affected by strong and varying noise. To improve the accuracy of fault identification in high-noise environments, this paper proposes a noise-robust acoustic diagnosis framework based on auditory perception. An auditory periphery-brainstem cascade is used as the front end to construct a multi-characteristic-frequency auditory spectrogram, which is then fused with a convolutional neural network (CNN) for fault recognition. This pipeline can more accurately preserve bearing-related features across multiple scales and levels of information. Compared with conventional time-frequency representations, experimental results on real wheelset-bearing acoustics (51.2 kHz sampling rate, 0.1-second windows, six operating conditions) show that the proposed method achieves near-ceiling accuracy in noise-free cases (99.80%) and maintains superior performance under all signal-to-noise ratio (SNR) settings from −15 to 0 dB. Averaged over SNRs of 0, −5, −10, and −15 dB, the auditory model achieved an accuracy of 87.4%, outperforming continuous wavelet transform (CWT), short-time Fourier transform (STFT), and Mel spectrogram baselines by 2.0%, 5.7%, and 9.5%, respectively, indicating stronger noise robustness. t-SNE visualizations further show that, at low SNR, the proposed method exhibits tighter within-class clustering and larger inter-class separation. These results provide a new and effective means for robust wheelset-bearing fault recognition in severely noisy environments.