Gear fault diagnosis method based on cross-attention mechanism for time-frequency feature fusion

To solve the problem of low accuracy in fault diagnosis caused by insufficient feature extraction of gear vibration signals, this paper sets forth a gear fault diagnosis method based on cross-attention fusion and establishes a CNN-BiTCN-CA diagnosis model. The original signal is reconstructed using variational mode decomposition (VMD) and fast Fourier transform (FFT). Time-frequency features are extracted using a bidirectional temporal convolutional network (BiTCN) and a convolutional neural network (CNN), respectively. The cross-attention mechanism (CA) is then utilized to fuse these time-frequency features, enabling comprehensive extraction of the original signal’s fault characteristics. Finally, a fully connected layer is employed to achieve accurate diagnosis of gear fault types. The experimental study demonstrates that in a Gaussian white noise environment with a signal-to-noise ratio (SNR) of 7.08 dB, the CNN-BiTCN-CA model achieves a gear fault classification accuracy of 99.85%. When Gaussian white noise with a SNR of 1.77 dB is introduced, the proposed model still achieves a diagnostic accuracy of 95.82%. The CNN-BiTCN-CA model is capable of extracting fault features in depth from the gear signal and effectively improving fault classification accuracy.