Pseudo-label guided dual classifier domain adversarial network for unsupervised cross-domain fault diagnosis with small samples

Abstract

Deep learning has been extensively employed in fault diagnosis applications due to its capacity to autonomously extract features from voluminous datasets. In practical engineering applications where the distribution between train and test data is inconsistent, domain adaptation (DA) approach is typically employed to facilitate cross-domain diagnosis problem. While DA methods have demonstrated efficacy in cross-domain diagnosis, they remain constrained by certain limitations. First, most existing DA approaches concentrate on domain adversarial training or feature matching, without fully acknowledging the inherent limitations of these two types of methods. Second, since DA methods can only mitigate the problem of domain bias, but not completely eliminate it, there is a risk of category confusion, i.e. some target samples that are situated at the clustering boundary may be misclassified. Furthermore, the majority of DA methods use a single classifier to predict class labels. To address the above problems, this paper proposes a pseudo-label guided dual classifier domain adversarial network. The dual classifier structure allows the model to achieve domain migration while retaining certain domain-specific distributional characteristics. In addition, the introduction of feature distribution discrepancy loss in domain adversarial training can further decrease the intra-class distance after clustering. Moreover, the pseudo-label guided loss constructed in this paper can reduce the misclassification of target domain samples due to the blurring of class boundaries. Finally, the effectiveness and generalizability of proposed method is verified by experimental results on rolling bearing and gear datasets.