Cross-Domain Adaptation Fault Diagnosis With Maximum Classifier Discrepancy and Deep Feature Alignment Under Variable Working Conditions

Abstract

Intelligent fault diagnosis based on deep learning has gained widespread attention. However, existing transfer diagnosis methods under variable working conditions face challenges in aligning deep features and handling task-specific decision boundaries. To address these issues, a transfer diagnosis method based on maximum classifier discrepancy (MCD) and deep feature alignment, namely maximum classifier discrepancy with deep feature alignment (MCDDFA) is proposed in this article. First, MCD is employed for transfer diagnosis by considering task-specific decision boundaries comprehensively. Then, two mapping-based domain adaptation methods of the multikernel maximum mean discrepancy and correlation alignment are combined to align the deep features of the source and target domains, while preserving the unique characteristics of each domain. Finally, an innovative integration of conditional domain adversarial network with MCD is proposed by utilizing adversarial training to align the joint probability distributions between the source and target domains, which further enhance the transfer diagnosis capability of deep networks. The experimental results on different dataset show that the MCDDFA achieves the accuracy improvement of 2%–10% by comparing to existing methods under variable working conditions, which validate the effectiveness of the MCDDFA in complex transfer diagnosis tasks.