Targeted Transfer Diagnosis Method across Different Machines

Abstract

Transfer fault diagnosis applies diagnosis knowledge of well-studied machines (the source domain) to solve the diagnosis issues of other related machines (the target domain), which is promising to overcome difficulties in collecting sufficient labeled data with respect to the big-data era. For a successful knowledge transfer across different machines, existing methods assume that the target label space is subject to a small shift and the symmetric basis to that of the source. However, the assumption is strict for the transfer diagnosis tasks across different machines, resulting in low diagnosis accuracy. Inspired by the principle of the targeted therapy, a targeted transfer diagnosis method is proposed to transfer knowledge across different machines. A domain-shared deep convolutional network is first constructed to map the source and target data into the feature space. After that, the limited number of labeled data in the target is set as anchors to indicate the targeted feature space region based on the relevance of their labels with the source domain. Finally, unlabeled target data are moved towards the targeted region along trajectories of the targeted anchors, which adapts partial distributions across domains by the optimal transport. The proposed method is verified by the transfer diagnosis tasks across different bearings. The results show that the proposed method can directionally adapt the feature partial distribution so as to improve the diagnosis accuracy when the intelligent diagnosis model is transferred across different machines.