Multi-scale convolution intra-class transfer learning for train bearing fault diagnosis

Considering the inconsistent distribution of bearing vibration data collected under different working conditions, the generalization ability of traditional deep learning model decreases. A multi-scale convolution intra-class adaptive deep transfer learning model was proposed. The spectrum of vibration data was analyzed using the modified ResNet-50. The middle-level features were obtained. A multi-scale feature extractor was developed, the middle-level features were processed, and the high-level features were generated. The high-level features were used as the inputs of classifier. The pseudo-labels were computed, and then the conditional distribution distances of vibration data collected under variable working conditions reduced for the intra-class adaptation. To verify the generality and superiority of model, the proposed method was employed to analyze a train wheelset bearing dataset and the Case Western Reserve University dataset under variable working conditions. Analysis result indicates that the high-level features of samples with the same label in different domains are properly aligned. More satisfactory fault diagnosis accuracies are obtained by the proposed model. In six fault diagnosis cases of train bearing under variable working conditions, the average diagnosis accuracy of the proposed model is 90.75%, approximately 10% higher than those of traditional deep learning models, while the recall rate is 0.927. In twelve fault diagnosis cases of Case Western Reserve University dataset under variable working conditions, the average accuracy obtained by the proposed model is 99.97%, approximately 10% higher than those of traditional models. The conditional distribution discrepancy between different domains reduces by using the pseudo-labels. The inconsistency problem of data distribution of source domain and target domain is properly addressed. The high-level features of samples from different scales can be aligned by the multi-scale feature learner. The generalization and robustness of the model largely increase. In conclusion, the proposed model has a high potential for the train bearing fault diagnosis under variable working conditions. © 2020, Editorial Department of Journal of Traffic and Transportation Engineering. All right reserved.