Abstract: In view of the fact that traditional fault diagnosis methods usually rely on single-domain information input, resulting in partial information loss in the signal or the incomplete use of the information, a bearing fault diagnosis method based on channel attention feature fusion was proposed. First, the original signal was processed by fast Fourier transform (FFT) and continuous wavelet transform (CWT) to obtain frequency domain and time-frequency plots. Then, two samples from different domains were input into the dual-stream Ghost Module (GM) neural network fault diagnosis model to extract the features of frequency domain and time-frequency domain. The important features of frequency domain and time-frequency domain were effectively fused by combining the channel attention mechanism, so as to obtain richer fault diagnosis information and realize accurate classification of fault signals. Finally, the experimental verification were carried out using bearing fault data sets developed by Case Western Reserve University (USA), Jiangnan University (China), and University of Ottawa (Canada). The results show that, compared with the existing popular models, the bearing fault diagnosis method based on channel attention feature fusion achieves classification fault diagnosis accuracies of 99.78%, 98.50% and 97.65% on the three data sets, respectively. It is proved that the proposed method presents good diagnosis effect.