Straddle-type monorail train gearbox fault diagnosis based on digital twin and domain adaptation

The goal of cross-domain mechanical fault diagnosis is to transfer diagnostic expertise from the source domain to the target domain. Currently, most approaches assume that the labeling information of the target domain is known and rely on high-quality labeling data. Nonetheless, in practical industrial settings, obtaining prior fault information or comprehensive fault data in the target domain is often impractical. To surmount this problem, a generalized domain-adaptive fault diagnosis method leveraging digital twin assistance is proposed in this paper. The approach commences by constructing a physical simulation model of a straddle-seat monorail train gearbox utilizing digital twin technology and augments the training dataset with simulation data. Subsequently, a residual contraction network is deployed to distill knowledge from the source domain data, and a Mixup data augmentation technique is incorporated to regularize the output distribution of adjacent training samples. Then, the minimum distance between categories in the source domain is learned by introducing a multi-classifier to recognize known or unknown information in the target domain. Finally, validation is carried out with two sets of experimental platform data, and the results show that the proposed method is able to achieve high diagnostic performance when the device state is unknown.