Abstract: The traditional deep learning model network applied to rolling bearing fault identification has low diagnosis accuracy and long training time. To overcome these disadvantages, this paper proposes a multi-scale attention reverse residuals convolutional neural network (MARCNN). Firstly, multi-scale feature extraction module adopts multi-scale convolution to obtain different levels of original signal features and adaptively extract fault feature information. Secondly, feature graph expansion standard convolution is used to construct the shallow convolution module to improve the shallow network learning ability. Finally, SE-Mobile module is constructed to mine deep fault features and reduce the number of model parameters. All modules combine attention mechanism to integrate feature weights of different dimensions to improve model feature learning ability. The experimental results show that the model can reduce the number of parameters and improve the training speed, and the accuracy of the model can reach 99.98%, 98.41% and 94.98% under fixed, Gaussian noise and variable load conditions, respectively. It is shown that the model has a certain anti-noise and generalization ability.

Key words: attentional mechanism, reverse residual module, fault diagnosis, rolling bearing, CNN