Gabor wavelet multi-scale attention adversarial network for cross-domain fault diagnosis

Abstract Aiming at the problem of declining diagnosis rate due to deficient feature extraction ability of traditional deep learning methods in bearing fault diagnosis (FD), this paper proposes a novel Gabor wavelet multi-scale attention adversarial network (GWMAAN) model. The model first effectively extracts local time–frequency features of the original data through continuous wavelet convolution layer. Subsequently, the extracted features are fed into a continuous wavelet attention layer to further extract key information and eliminate irrelevant components. Specifically, the frequency attention mechanism was employed to perform a weighted fusion process on the multiscale features generated from the continuous wavelet transform. Next, the weighted fused features are input into the fault classifier and domain discriminator. In order to enhance the feature differentiation and generalization ability of GWMAAN, Wasserstein distance is employed in the feature layer for domain adversarial training. To verify the effectiveness of the model, the bearing dataset collected under fluctuating rotational speed and weak signal conditions is used for experiments in this paper. The test results indicate that compared with other methods, the GWMAAN model demonstrates notable benefits in both accuracy and robustness of FD, and its average accuracy is as high as 99.06%.