A statistical distribution recalibration method of soft labels to improve domain adaptation for cross-location and cross-machine fault diagnosis

Abstract Unsupervised domain adaptation has achieved certain success in recent cross-domain fault diagnosis research. As a widely used transfer strategy, the distribution alignment often occurs with the problems of too few valid alignment samples, too low confidence of predicted labels, and the inadequate alignment of marginal or conditional distributions. Therefore, this paper proposes a statistical distribution recalibration method of soft labels (SDRS). First, SDRS defines the valid samples and confusion interval in the statistical distribution of per-class predicted probabilities. Then, from the perspective of binary classification, a recalibration space in the confusion interval is further optimized by a center distance metric, to improve predicted confidence and valid distribution alignment. Built on SDRS, a novel cross-domain fault diagnosis approach named SDRS-DAN is constructed, where dynamic distribution adaptation is used to match and adjust the marginal and conditional distribution discrepancies adaptively. Extensive experiments prove the effectiveness of SDRS-DAN in cross-location and cross-machine scenarios.