Noisy Open Set Adversarial Domain Adaption for Bearing Fault Diagnosis Based on Optimized Divergence

Unsupervised adaptive transfer learning has accomplished extraordinary achievements in mechanical fault diagnosis. However, not only are the failure mode and location of machines often unpredictable, but the presence of noisy labels can also wreak havoc on diagnostic tasks. To overcome this problem, based on the maximum classifier discrepancy and optimized Kullback-Leibler (KL) divergence, an entropy-increasing separation guided by optimized divergence (EIS-OD) was proposed. Samples were selected with high confidence to enhance noise resistance in the source domain. For private samples, optimized KL divergence (OKD) loss is constructed to implement entropy increase, then they are masked in subsequent training to avoid affecting domain alignment. To complete domain adaption, the max-min adversarial method with unmasked samples is utilized to optimize generators and classifiers, which simultaneously alleviates the misclassification problem of shared classes. Our method has been tested and validated on two bearing datasets, demonstrating a significant enhancement in the effectiveness and reliability.