Unsupervised Domain Adaptation Method Based on Mean Teachers and Its Application to Cross-Domain Fault Diagnosis Under Time-Varying Conditions

Unsupervised intelligent transfer diagnosis methods are effective means to address the challenges of cross-domain fault diagnosis for avigation equipment.Aiming at the issues of low training efficiency and insufficient feature extraction capabilities for traditional unsupervised transfer diagnosis methods based on convolutional neural network (CNN) architecture,this paper proposes a Mean Teachers-based unsupervised domain adaptation network (MTUDAN) method.MTUDAN designs an attention-enhanced CNN architecture and incorporates the transfer diagnosis strategy to achieve unsupervised cross-domain transfer intelligent diagnosis under time-varying operating conditions.The proposed MTUDAN method can significantly enhance the channel attention mechanism performance of CNN by designing the squeeze and excitation module.Moreover,the proposed MTUDAN method employs a teacher model to guide the student model for stable updates of model parameter,thereby enhancing the generalization ability and unsupervised transfer diagnosis performance of the student model.Comprehensive comparative experiments on a complex multi-stage transmission system fault datasets demonstrate that the proposed MTUDAN method exhibits the superior transfer diagnosis performance and outperforms existing mainstream unsupervised transfer diagnosis approaches.