Deep Active Transfer Learning for Cross-Domain Fault Diagnosis With Limited Samples

In bearing fault diagnosis (BFD), unsupervised domain adaptation (UDA) enables cross-domain few-shot BFD due to its knowledge transfer ability. However, UDA, due to the lack of strong supervision in the target domain (TD), often neglects the effective use of informative unlabeled samples from the TD, which can cause the decision boundary to be biased toward the source domain (SD), reducing the effectiveness of UDA. To tackle this challenge, a novel deep active convolutional transfer learning network (DACTLN), which can align the marginal as well as conditional distributions between the SD and TD and achieve labeled sample expansion of the TD from both unlabeled target candidates and labeled source samples simultaneously, is proposed to build a target task-oriented fault diagnostic model. Specifically, DACTLN consists of a conditional recognition module for automatically learning features and recognizing the health status of machines, a DA module to facilitate the conditional recognition module in learning domain-invariant features, and an active query module to enhance the leading role of target samples during training. DACTLN dynamically augments the amount of informative and representative labeled target samples, matching the joint probability distribution between the SD and TD, thus can better capture domain-invariant characteristics, and enhances the effectiveness of DA with limited training data while reducing sample annotation costs. Experimental outcomes demonstrate that the proposed approach surpasses other models in diagnostic accuracy.