Correlation Regularized Conditional Adversarial Adaptation for Multi-Target-Domain Fault Diagnosis

Abstract

Domain adaptation has significantly promoted the development of transferrable fault diagnosis. However, the diagnostic scenario with multiple target distributions, namely, multi-target-domain adaptation (MTDA), has not been well addressed. In view of this, the specific characteristics of MTDA are investigated in this article, and a novel correlation regularized conditional adversarial adaptation network (CRCAA) is proposed on its basis. Specifically, to enhance the transferability of CRCAA, a feature space linear mapping algorithm is developed to integrate the category information into adversarial feature matching. Moreover, by establishing a correlation regularization mechanism, the sample relevance is exploited to guide the distribution alignment, thereby reducing the negative transfer near the decision boundary. To facilitate the convergence of adversarial training, CRCAA is designed to learn the distinguishable features and domain invariant features in two separate stages. Extensive experiments on the gearbox and rolling bearing datasets demonstrate the effectiveness and superiority of CRCAA in engineering applications.