A Transfer Learning Method for Intelligent Fault Diagnosis from Laboratory Machines to Real-Case Machines

It is difficult to train a reliable intelligent fault diagnosis model for machines used in real cases (MURC) because there are not sufficient labeled data. However, we can easily simulate various faults in a laboratory, and the data from machines used in the laboratory (MUL) contain fault knowledge related to the data from MURC. Thus, it is possible to identify the health states of MURC by using related fault knowledge contained in the data from MUL. To achieve this purpose, a transfer learning method named convolutional adaptation network (CAN) is proposed in this paper. The proposed method first uses domain-shared convolutional neural network to extract features from the collected data. Second, the distribution discrepancy between the learned features of the data from MUL and MURC is reduced by minimizing multi-kernel maximum mean discrepancy. Finally, pseudo label learning is introduced to train domain-shared classifier by using unlabeled data from MURC. The proposed method is verified by a transfer learning case, in which the health states of locomotive bearings are identified by using the fault knowledge contained in the data from motor bearings used in a laboratory. The results show that CAN is able to effectively identify the health states of MURC with the help of the data from MUL.