Research on bearing fault diagnosis method based on multi-scale time-frequency representation and lightweight attention network

Abstract

Rolling bearings are key components in rotating machinery, and their operating status directly affects equipment safety and maintenance costs. To address issues such as reliance on manual features in traditional bearing fault diagnosis, large parameter counts in deep models, and degraded recognition performance under noisy conditions, this paper proposes a bearing fault diagnosis method based on multi-scale time-frequency representation and a lightweight time-frequency attention convolutional network. First, one-dimensional vibration signals are converted into Short-Time Fourier Transform (STFT) spectra, Continuous Wavelet Transform (CWT) spectra, and envelope spectra to construct a three-channel multi-scale time-frequency map with physical meaning. Second, a Time-Frequency Attention CNN (TF-AttentionCNN) is designed, which highlights fault-related regions through gating mechanisms along the time and frequency directions. Experiments on the CWRU bearing dataset show that under noise-free conditions, the accuracy of multi-scale features combined with SimpleCNN improves from 95.00% to 98.75%; the lightweight TF-AttentionCNN achieves 98.75% accuracy after 30 training epochs with only 105,514 parameters, approximately 2.26% of SimpleCNN’s parameter count. Under 0 dB noise conditions, model performance degrades significantly, indicating that robustness to strong noise still needs further enhancement.