Deep transfer learning with metric structure for fault diagnosis

Abstract

Maximum Mean Discrepancy (MMD) is a classical method of measuring distribution distance and is widely used in Deep Transfer Learning (DTL) for fault diagnosis. However, in practice, MMD neglects the label information and the spatial structure relationship of the data and performs unsatisfied. To address this issue, a novel DTL fault diagnosis method with metric structure is proposed in this paper. Besides MMD, a novel domain adaptation module is designed through learning a common metric matrix from the Two Transferred Domains (TTD). Thereby, it can constrain TTD in the same geometric structure and make the distribution of TTD similar by both the mean discrepancy and the metric structure. Furthermore, a novel soft pseudo-label trick is proposed for dividing similar and dissimilar sample pairs of the unlabeled samples in the target domain. Experimental results have demonstrated that this method has well diagnostic performance in multiple transfer tasks.