Fault diagnosis of rolling bearings under variable conditions based on unsupervised domain adaptation method

The paper proposes an unsupervised deep convolutional dynamic joint distribution domain adaptive network model for the problem of bearing fault diagnosis under variable conditions, which involves missing labeling of target domain data and large differences in the distribution of source and target domain data. The model consists of the following steps: (1) converting the original vibration signal of the bearing into a time–frequency map representation and performing feature extraction on the labeled source domain samples and the unlabeled target domain samples by the deep convolutional feature extractor; (2) dynamically aligning the marginal distribution and conditional distribution of the two domain features by the marginal distribution adaptation module and the conditional distribution adaptation module, so that the trained network model can classify the unlabeled target domain samples accurately according to the label mapping relationship of the source domain samples; (3) validating the model on two rolling bearing datasets; (4) experiment with model interpretability in conjunction with XAI techniques to help us understand what the model actually does. The experimental results on two rolling bearing datasets show the validity of the proposed model.