Discriminative fault diagnosis transfer learning network under joint mechanism

Abstract

Unsupervised fault diagnosis methods for rotating machinery are gaining attention but face challenges such as feature extraction from vibration signals, aligning distributions between source and target domains, and managing domain shifts. This paper proposes a novel unsupervised transfer learning method that integrates the Squeeze-and-Excitation (SE) attention mechanism to enhance useful features while suppressing redundant ones. An Integrated Distribution Alignment Framework (IDAF) is introduced, which employs the Joint Adaptation Network (JAN) approach to construct a local maximum mean discrepancy in conjunction with Correlation Alignment (CORAL) to improve distribution alignment between domains. Moreover, to enhance feature learning and obtain more distinct features, the authors utilize a novel discriminative feature learning method called I-Softmax loss. This method can be optimized in a manner similar to the traditional Softmax loss while providing improved classification performance. Finally, deep adversarial training is applied between the source and target domains to adaptively optimize the target domain network parameters, reducing domain shift and improving fault classification accuracy. Experimental validation using four sets of bearing faults and six sets of gear faults demonstrates the superior performance of the proposed method in unsupervised fault diagnosis tasks.