Rolling bearing fault diagnosis method based on improved residual shrinkage network

With strong feature extraction ability, neural networks can effectively realize rolling bearing fault diagnosis. However, due to the impact of noise and variable working conditions in the bearing environment, it is difficult to extract important fault feature information from vibration signals. Aiming at the above problems, a fault diagnosis method for improved residual shrinkage network of rolling bearings is proposed in this paper. In this method, the vibration signal of rolling bearing is taken as model input. Firstly, the initial features of the signal are extracted by the preprocessing module. Then, the threshold and slope factors are designed to improve the soft threshold function, and the subdomain network is constructed to adaptively determine the parameter according to different characteristics to suppress the noise interference. Secondly, a two-pool structure is introduced to improve the weighting of important features by the attention mechanism, which enhances the generalization performance of the model. Finally, multi-layer improved residual shrinkage blocks are used to reduce the noise impact and enhance the fault feature information, so as to improve the accuracy of multi-classification fault identification. The experimental results show that the accuracy of the proposed method is 99.83% and the standard deviation is 0.83, both of which are higher than other existing neural network methods. It proves that the proposed method has a strong advantage in the information extraction of rolling bearing fault features.