Adversarial Domain Adaptation Model Based on LDTW for Extreme Partial Transfer Fault Diagnosis of Rotating Machines

Domain adaptation (DA) models are widely used in the fault diagnosis of rotating machines under variable operating conditions, in which most of the existing models assume the same number of source- and target-domain categories, i.e., the same label space. However, in practice, the labeling space is inconsistent; even there is only one class of the same type of fault, the traditional DA and partial DA (PDA) models are hard to maintain high accuracy. To cope with this challenge, an adversarial DA model based on local dynamic time warping (LDTW) is proposed. The proposed model is divided into three steps: first, the signals are discriminated for similarity using LDTW. Second, a class balancing strategy is proposed to balance the target-domain categories on the basis of similarity. Third, by using a domain discriminator to reduce the domain differences between source and target domains, thus accomplishing knowledge migration between domains. In addition, by visualizing the features extracted from the convolutional layer of the proposed model, this article provides an interpretable illustration of migration. Experimental validation on three datasets shows that the diagnostic performance of the proposed model is superior to the existing PDA models.