Abstract
Focusing on the problem that it is difficult to maintain a high diagnostic accuracy rate, short running time, and robust generalization capability in the face of a strong-noise environment in rolling bearing fault diagnosis, a bearing fault diagnosis model (GMNR-CABA-MAGRU) founded upon a new attention-mechanism-improved residual network (ResNet-CABA) and a Gram denoising module (GMNR) is proposed, and the CWRU bearing dataset is used for verification. Under the 0-load condition in a noise-free environment, the diagnostic accuracy of this model reached 99.66%, and the running time was only 52.74 s. Then, a bearing dataset with added Gaussian noise from -4 db to 4 db was verified, and this model was still able to maintain a diagnostic accuracy of 90.32% under the strong-noise environment of -4 db SNR. And migration experiments were carried out under different load conditions, and this model was also able to maintain a very high accuracy rate. Moreover, in all the above experiments, this model performed better than various comparative models. The developed framework demonstrated superior diagnostic precision, enhanced robustness, and improved generalization capability.