An intelligent fault diagnosis approach based on transfer learning from laboratory bearings to locomotive bearings

Intelligent fault diagnosis of rolling element bearings has made some achievements based on the availability of massive labeled data. However, the available data from bearings used in real-case machines (BRMs) are insufficient to train a reliable intelligent diagnosis model. Fortunately, we can easily simulate various faults of bearings in a laboratory, and the data from bearings used in laboratory machines (BLMs) contain diagnosis knowledge related to the data from BRMs. Therefore, inspired by the idea of transfer learning, we propose a feature-based transfer neural network (FTNN) to identify the health states of BRMs with the help of the diagnosis knowledge from BLMs. In the proposed method, a convolutional neural network (CNN) is employed to extract transferable features of raw vibration data from BLMs and BRMs. Then, the regularization terms of multi-layer domain adaptation and pseudo label learning are developed to impose constraints on the parameters of CNN so as to reduce the distribution discrepancy and the among-class distance of the learned transferable features. The proposed method is verified by two fault diagnosis cases of bearings, in which the health states of locomotive bearings in real cases are identified by using the data respectively collected from motor bearings and gearbox bearings in laboratories. The results show that the proposed method is able to effectively learn transferable features to bridge the discrepancy between the data from BLMs and BRMs. Consequently, it presents higher diagnosis accuracy for BRMs than existing methods.