MTUTL: Mean Teacher-Based Unsupervised Transfer Learning Method for Fault Diagnosis under Time-Varying Operating Conditions

Convolutional neural network (CNN) based unsupervised transfer diagnosis methods are effective methods for addressing cross-domain diagnostic challenges. However, low training efficiency and insufficient feature extraction capabilities of networks continue to limit the practical applications of CNN-based methods in industrial fault diagnosis. To tackle these challenges, this paper proposes a Mean Teacher-based unsupervised transfer learning (MTUTL) with the enhanced CNN for fault diagnosis under time-varying operating conditions. The proposed MTUTL method first introduces squeezing-excitation blocks to significantly improve the channel attention capability of CNN. The MTUTL method enables the teacher model to guide the student model toward stable updates, thereby enhancing the generalization capability and unsupervised transfer diagnosis performance of student model. Comprehensive comparative experiments on transmission system fault datasets demonstrate the effectiveness and superiority of our proposed MTUTL method, while ablation studies further validate the contribution of all key modules in our proposed MTUTL method.