Abstract:  Aiming at the problems such as low accuracy of the existing prediction method of residual useful life (RUL) of rolling bearing and difficult construction of bearing health index (HI), a prediction model for RUL of rolling bearings based on convolutional neural network (CNN) and integration of Inception V1 module and convolutional block attention module (CBAM) was proposed. Firstly, CBAM mechanism was added to CNN for weighted processing, and important features were strengthened and minor features were suppressed in channel and spatial dimensions. An improved Inception V1 module was added to improve information interaction between CNN channels and extracted degraded features comprehensively. Then, the network was optimized, the model was simplified by using the global maximum pooling (GMP) method, Dropout method and batch normalization (BN) method to avoid overfitting and improving the accuracy and overcoming the gradient disappearance problem during training. Finally, the data was processed, the signal after noise reduction was reconstructed into a three-dimensional tensor as the bearing health index HI, the degradation label was constructed and the evaluation index was introduced. The PHM2012 bearing data set was experimentally verified and compared with deep neural network (DNN), CNN and residual network method combined with attention mechanism (ResNet) under three working conditions. The results show that the average RMSE of the proposed method under variable load conditions is 0.033. Comparing with other methods, RMSE is respectively reduced by 86%, 78% and 69%, which has obvious advantages in prediction accuracy and generalization ability.
Key words:  rolling bearing; residual useful life; Inception V1 module; convolutional block Attention module(CBAM); convolutional neural network(CNN); global maximum pooling(GMP); batch normalization(BN)