Dynamic Regulation Domain Adaptation Network for Cross Domain Bearing Fault Diagnosis

Domain adaptation is an effective approach to address the discrepancies in probability distributions between source and target domains. In this paper, we propose a Dynamic Regulation Domain Adaptation Network (DRDAN), which features a novel feature fusion module within the feature extraction phase. This module not only targeted high-level abstract features but also delved into capturing low-level detailed features. The integration of these two types of features allowed the network to harvest a richer and more comprehensive set of information, thereby providing robust support for the subsequent domain adaptation processes. In the feature alignment phase, DRDAN implemented a dynamic regulation strategy that accounted for both the variances in marginal distributions and the implications of conditional distributions. This facilitated a more precise and adaptable adjustment of feature distributions, ensuring accurate alignment between source and target domains. Our evaluation, conducted using a bearing database as the experimental dataset, affirms that DRDAN significantly bridged the domain gap and enhanced the classification accuracy of data in the target domain.