A multi-source subdomain adaptation fault diagnosis method based on unidirectional movement of the target domain

Abstract

The traditional domain adaptation method for fault diagnosis of axial fans faces two main problems: (1) source domain moves to target domain makes the source feature distribution changed; (2) the narrow decision boundary of source domain features leads to misclassification of target samples. Therefore, a multi-source subdomain adaption fault diagnosis method based on unidirectional movement of the target domain is proposed. The method uses triplet-center loss to improve the discrimination of target domain samples, which reducing intra-class distance and increasing inter-class distance of source domain features; extracting the domain invariant feature of the target samples by asymmetric adversarial and improved subdomain feature distance measurement; the cosine similarity is used to align the classifiers’ outputs of different source domains; the mean value of all classifiers’ outputs are used as pseudo labels, and the pseudo labels are optimized by maximum entropy to improve their reliability. A large number of experiments show that this method has a significant effect on solving the problem of cross conditions fault diagnosis.