A novel bearing fault diagnosis method based joint attention adversarial domain adaptation

Recently, many unsupervised domain adaptation methods based on a metric distance or adversarial training do not consider whether the feature representations can be transferred or not. To overcome this challenge, we explore developing a novel approach named joint attention adversarial domain adaptation (JAADA). Specifically, the extracted features are first manually divided into numbers of feature regions. Second, MMD is introduced to mitigate the distribution discrepancy in separated segment features. Furthermore, different weights obtained by the attention mechanism and MMD values have been assigned to different regions. Finally, local and global attention has been fused into one unified adversarial domain adaptation framework. A series of comprehensive experiments on four fault datasets validate that the proposed method has a superior convergence and could boost 1.9%, 3.0%, 2.1%, and 3.5% accuracy than the state-of-the-art methods, respectively.