An enhanced deep joint distribution alignment mechanism for planetary gearbox fault transfer diagnosis

Abstract

Lots of fault transfer diagnosis methods have been presented to bring the gap between source domain and target domain. Nevertheless, most of them only pay attention to the marginal domain adaptation (MDA), while ignoring the conditional domain adaptation (CDA) of class levels. Additionally, the universal CDA mechanisms greatly rely on the quality of pseudo label of target-domain samples. To deal with above issues, an enhanced deep joint distribution alignment (DJDA) mechanism is proposed to comprehensively achieve the MDA and CDA. In DJDA, a new MDA distribution discrepancy metric, including the mean and covariance information of two domains, is constructed. Meanwhile, a new CDA mechanism based on unsupervised clustering and Wasserstein distance is built to align the class-wise distribution of two domains, in which the pseudo label is needless. Experimental results evaluate the efficacy and advantage of proposed DJDA.