Bearing Fault Diagnosis with Small Sample Sizes Based on CNN-SE-LSTM Transfer Learning

This paper concerns the fault diagnosis issue with small bearing fault sample sizes encountered in practice. Considering the need for stability and high accuracy in bearing fault diagnosis, a novel motor bearing fault diagnosis method is proposed on the basis of Convolutional-SEBlock-Long Short-Term Memory Network (CNN-SE-LSTM) combined with Transfer Learning. In this paper, the SEBlock (Squeeze-and-Excitation Block) module is embedded between the convolution layer and the pooling layer, which further improves the performance of the algorithm by adaptively adjusting the weights between channels. The model can also overcome the problems of scarce data in the target domain and different distribution between domains through parameter transfer. Train and save the best model parameters when the source domain data is sufficient, transfer the model to the target domain with less fault data, and fine-tune the model to better adapt to the data distribution in the target domain, and then test the accuracy of the model to the test set. The fault diagnosis results with and without transfer are compared with traditional CNN, LSTM, CNN-LSTM, SE-CNN, and SE-Resnet models, and the results show that the proposed model is with higher accuracy and better stability.