Adversarial Domain Bias Removal Network for Cross-condition Bearing Fault Diagnosis

Recently, domain adaptation has been extensively utilized to address the issue of cross-condition bearing fault diagnosis. Traditional domain adaptation methods typically optimize by reducing the disparity in sample distribution between the source conditions and the target conditions, without directly focusing on the model's contribution to the transfer conditions. When using traditional domain adaptation methods for training, it is possible that the classification features from the source conditions and the noise features from the target conditions share a similar distribution. To avoid ineffective transfer of the model, emphasize the ultimate goal of transfer learning, and ultimately enhance the model's diagnostic reliability under target conditions, a novel adversarial transfer paradigm, Adversarial Domain Bias Removal Network (ADBRN), has been proposed. ADBRN prioritizes the improvement of the model's diagnostic performance on target domain samples and explicitly enhancing the reliability of test results on target domain samples. Furthermore, this paper theoretically validates the positive correlation between the L2 norm of prediction vectors and prediction confidence.