Unsupervised deep transfer learning with moment matching: A new intelligent fault diagnosis approach for bearings

Deep learning has redefined state-of-the-art performances in the research of intelligent fault diagnosis, however, most studies assumed that the training and testing data have the same distributions. In this paper, an unsupervised deep transfer network with moment matching (UDTN-MM) is proposed, aiming to realize fault diagnosis under different working conditions. Grayscale time-frequency images are utilized as the network input, and two adaptive methods are employed to reduce the distribution discrepancy. First, a deep transfer network is developed to extract transferrable features of the images. Then, two regularization terms expressed by moment matching, i.e., marginal distribution adaptation and statistical feature transformation, are designed to guarantee accurate distribution matching and domain adaptation. To prove the superiority of proposed moment matching method, two network structures with different bottlenecks are constructed. The results of case studies show that the approach is competitive on unlabeled samples in terms of diverse rotating speeds and fault severities.

Automated summary

This paper proposes an unsupervised deep transfer network with moment matching for fault diagnosis under different working conditions. Two adaptive methods are employed to reduce distribution discrepancy, and the results show the competitiveness of this method in various fault scenarios.