A Novel Transfer Capsule Network Based on Domain-Adversarial Training for Fault Diagnosis

The success of intelligent fault diagnosis comes from one important assumption: the test data are consistent with the training data in data distribution. However, in the actual factory environment, the difference in data distribution due to changing working conditions will cause the performance of the trained model to seriously degrade. To address the problems, a transfer capsule network based on domain-adversarial training (DATTCN) is proposed. Specifically, it extracts fault features through wide convolution and multi-scale convolution, and performs fault classification through capsule networks. And the purpose of enhancing the diagnostic performance of the target domain is realized through adversarial training. In the fault identification of the Case Western Reserve University data set under varying working conditions, the DATTCN algorithm almost reaches 100% accuracy, and it is 92.3% on the Paderborn University data set. The accuracy of the DATTCN algorithm exceeds other advanced algorithms, fully verifying the effectiveness of the DATTCN algorithm.