DFSA-DAN: dynamic fusion of statistical metric and adversarial learning for domain adaptation network based intelligent fault diagnosis

Abstract

The advancement of deep transfer learning has motivated research into the realization of intelligent fault diagnosis schemes for rolling bearing. Nevertheless, existing research rarely provides further insight into the importance of statistical distance metric-based methods and adversarial learning-based methods in domain adaptation, and the commonly used feature extractors are more difficult to extract features suitable for domain transformation. In this paper, a dynamic fusion of statistical metric and adversarial learning for domain adaptation network is proposed to achieve a dynamic measure of the importance of different domain adaptation methods. This new model utilizes a local maximum mean discrepancy metric to adjust the conditional distribution and adversarial training to adjust the marginal distribution between domains. Meanwhile, to assess the importance of the two distributions, a dynamic adaptation factor is introduced for dynamic evaluation. In addition, to extract features that are more suitable for domain transformation, the model incorporates a dual depth convolutional path with an attention mechanism as a feature extractor, enabling multi-scale feature extraction. Experimental results demonstrate the model’s superior generalization capability and robustness, enabling effective cross-domain fault diagnosis in diverse scenarios.