Multiscale Margin Disparity Adversarial Network Transfer Learning for Fault Diagnosis

Due to the limitation in obtaining sample data in the real world, Domain Adaption Transfer Learning (DAT) has been a research focus in fault diagnosis. However, the existing DAT–based fault diagnosis has the problem that the extracted feature in different domains is limited, many existing methods only consider aligning different domains, and the loss function is single. To address these issues, Multi-Scale Margin Disparity Adversarial Network Transfer Learning (MMDAN) for Fault Diagnosis is proposed in this paper. Firstly, the abundant features of different domains are extracted by the proposed multi-scale neural network. Specifically, the discrepancy between different domains is measured by Margin Disparity and adversarial loss. Meanwhile, the classifier achieves the fault diagnosis. Finally, a joint loss function is proposed to update the neural network parameters. Two different case studies are carried out to verify the effectiveness of MMDAN. The experimental results show that MMDAN can achieve the highest diagnosis accuracy than other methods even in multi-task transfer learning.