Unsupervised multiple-target domain adaptation for bearing fault diagnosis

Data-driven bearing fault diagnosis methods have gained significant attention due to their high accuracy and minimal reliance on expert knowledge. However, models trained on the data collected under single working conditions often assume independent and identically distributed data, limiting their generalization ability in realworld scenarios where varying working conditions introduce challenges such as domain shift and unlabeled data. Unsupervised domain adaptation is an effective method for addressing these challenges by training classifiers using labeled source domain data and unlabeled target domain data. However, current domain adaptation fault diagnosis methods are typically limited to single-source and single-target scenarios. When multiple unlabeled working conditions are involved, it becomes difficult to simultaneously ensure the domain invariance and the fault relevance of features. This paper proposes a fault diagnosis method across multiple unlabeled working conditions. The method extracts features from image data transformed from time-domain signals and ensures the correlation between the extracted features and fault labels, as well as domain invariance, through the gradient reversal mechanism and a corresponding loss function. The proposed method achieves an average accuracy of 99.89 % and 95.93 % in multi-target domain fault diagnosis tasks on two public benchmark bearing datasets, which illustrates its effectiveness and advantages compared with the existing representative methods. The source code is available at https://github.com/WhiteGL/DA_IRP/tree/master.