Bearing fault diagnosis based on attention mechanism and deep residual network

A bearing fault diagnosis model based on the deep residual network is proposed for the situation that the model recognition rate is low and the classification effect is poor due to the large difference of fault sample distribution in the actual working condition. Firstly the collected bearing fault signals are constructed as fault samples, reconstructs the one-dimensional time series signals into grayscale maps, and initially obtains the input data which is suitable for the deep residual network. To solve the situation of insufficient effective samples, data augmentation by sliding sampling is used to expand the bearing vibration dataset; the samples are further divided into training and testing sets as the input of ResNet101, and data normalization is used to make the training and testing sets learn the same distribution to shorten the training time; then a hybrid attention mechanism is introduced at the appropriate parts to effectively suppress the redundant features and enhance the feature extraction capability of the model. And then a softmax classifier is used for fault classification to achieve intelligent fault diagnosis of rolling bearings. Finally, the Western Reserve University bearing dataset (CWRU) is used to verify the effectiveness of the model. The experimental results show that the proposed bearing fault diagnosis method based on hybrid attention mechanism and residual network can achieve more than 99 % diagnostic accuracy, and it achieves good generalization performance on the high-speed rail wheel pair dataset with an accuracy above 94 %.