Multilayer Domain Adaptation-Based Convolutional Auto Encoder for Fault Diagnosis under Different Speeds

Application of deep learning in fault diagnosis have developed rapidly in recent years. The majority of fault diagnosis methods can achieve satisfactory results only on the basis of the assumptions that training and testing data come from the same data distribution and sufficient labeled data are necessary for model training. However, These assumptions do not always hold in actual situations. In light of this problem, a transfer learning-based method named multilayer domain adaptation-based convolutional auto encoder (MDACAE) is proposed in this paper. The Maximum Mean Discrepancy (MMD) and Wasserstein Distance (WD) are both applied in the last three layers of the proposed model. Finally, a dataset collected from rolling bearings is used to demonstrate the superiority of the proposed method. Results reveal that the proposed MDACAE can obtain better results of domain adaptation and have superior diagnostic capability.