Intelligent Fault Diagnosis With Deep Adversarial Domain Adaptation

Abstract

With the rapid development of fault diagnosis methods based on deep learning, many studies have investigated the transfer of intelligent fault diagnosis methods to learn the domain-invariant features of machines under different conditions. Previous researches focused on learning domain-invariant features through domain adaptation. However, the domain alignment methods cannot remove the domain shift, the target samples may be incorrectly classified by the decision boundary learned from the source domain and eventually cause the domains to be aligned in the wrong direction. To cope with this problem, we propose a deep adversarial domain adaptation network (DADAN) to transfer fault diagnosis knowledge. DADAN uses domain-adversarial training based on the Wasserstein distance to learn domain-invariant features from the raw signal. In addition, the network is combined with a supervised instance-based method to learn the discriminative features with better intraclass cohesion and interclass separability, which can benefit the domain alignment. A data set of bearing data including three speed conditions and a data set of hard disk data acquired from accelerated degradation test and real-case conditions were used to evaluate the performance of the proposed DADAN.