Moment matching-based intraclass multisource domain adaptation network for bearing fault diagnosis

• A new multisource domain adaptation diagnosis method is proposed. • A moment distance metric is designed for multisource domain adaptation. • Conditional distribution distance is narrowed by an intraclass alignment training strategy. • The robustness is validated by case studies under different working conditions. Deep learning based fault diagnosis methods assume that training and testing data with sufficient labels are available and share a same distribution. In practical scenarios, this assumption does not generally hold due to variable working conditions of rotating machineries and the difficulty in labeling vibration data under all working conditions. Transfer learning (TL) overcomes this problem by utilizing knowledge learned from the source domain to help accomplish tasks on the target domain. Although TL based fault diagnosis has been considerably studied, most studies mainly focus on single-source TL. Since multisource domains with labeled samples from which more useful knowledge can be extracted are available, in this paper, a novel multisource TL model, called the moment matching-based intraclass multisource domain adaptation network, is proposed. This model uses a feature learner to generate features of each source and target domain data to enable the joint weight classifier to predict target labels. It also introduces a moment matching-based distance metric to reduce the distance among all source domains and the target domain. During the training of the model, an intraclass alignment training strategy is applied to match the marginal and conditional distributions of each domain simultaneously. Experiments on two datasets are performed, wherein the proposed method is used to identify bearing fault types under four load conditions. Experiment results, such as high diagnostic accuracies support the reliability and generalizability of the proposed model.