As a basis of bearing fault diagnosis, signal acquired by vibration sensor is easily affected by the work environment noise. In order to extract characteristic information accurately, an improved traditional wavelet threshold denoising method was adopted. The traditional hard threshold and soft threshold were combined using the method of intermediate proportional coefficient transition, and the denoising signal was smooth and effective. The denoising signal was operated by a two-dimensional short-time Fourier transform to obtain the two-dimensional time-frequency domain data structure. The capsule attention method was used to improve the direct structure of ResNet network, so as to acquire a better classification model. Comparing the structure of different attention models before and after denoising, the effectiveness of the method was proved.Higher accuracy can be achieved.