Abstract:
Currently, most existing fault diagnosis methods based on domain adaptive (DA) learning reduce the distribution difference between two domains from the metric distance pe...Show MoreMetadata
Abstract:
Currently, most existing fault diagnosis methods based on domain adaptive (DA) learning reduce the distribution difference between two domains from the metric distance perspective. However, the domain alignment is performed only from the perspective of metric distance without fully mining the transferable features of the samples, which also leads to poor cross-domain fault diagnosis. To address this issue, a domain adversarial fault diagnosis method based on features and joint distribution migration alignment (FJDMA) is proposed. First, a feature aligner is designed to learn more transferable features from both local and global aspects. Second, a new weighted maximum mean square discrepancy (WMMSD) is designed to measure the distribution distance between the samples. The WMMSD can effectively reduce the distribution distance between the same classes within the domain. In addition, to increase the distribution distance between different classes between domains, we introduce correlation alignment (CORAL). Finally, a dynamic factor is designed to quantitatively combine WMMSD and CORAL, thus constructing the joint distribution migration (JDM). The JDM further enhances domain confusion during model training. Two bearing datasets and one gear dataset are used for experimental validation. The results show that the average diagnostic accuracy of the proposed method between the two bearing datasets is 3.34% higher than that of the state-of-the-art method. The average diagnostic accuracy of the proposed method between the bearing and gear datasets is improved by 2.38% over the state-of-the-art method.
Published in: IEEE Transactions on Instrumentation and Measurement ( Volume: 74)