Multi-kernel weighted joint domain adaptation network for cross-condition fault diagnosis of rolling bearings

Unsupervised domain adaptation (UDA) has received wide attention in cross-condition fault diagnosis of rolling bearings. However, the existing methods cannot adaptively align the marginal and conditional distributions, and the generated pseudo-labels on the unlabeled target domain have low confidence, which limits their practical engineering applications. To address these problems, this paper proposes a multi-kernel weighted joint domain adaptation network (MKWJDAN) for cross-condition fault diagnosis of rolling bearings. In MKWJDAN, the multi-kernel maximum mean discrepancy and the multi-kernel conditional maximum mean discrepancy are combined as a new joint distribution discrepancy metric to enhance the domain confusion effect. Meanwhile, an adaptive weighting strategy is designed to dynamically align the marginal and conditional distributions by evaluating the relative importance of these two distributions. Besides, a pseudo-labeling rectification mechanism is developed to enhance the pseudo-label confidence of the target domain. Extensive experiments indicate that compared to other advanced UDA methods, the proposed MKWJDAN method has a significant advantage in cross-condition fault diagnosis of rolling bearings. The code for this paper is available at https://github.com/CHEN99-HAO/Deep-learning.