Research on Measurement Methods of Transferability between Different Domains in Transfer Learning

摘要

If transfer learning can be used in fault diagnosis of rotating machinery, the transferability between different fault domains should be measured first. A certain degree of similarity between the source domain and the target domain is the prerequisite for the effectiveness of the transfer learning method. In this paper, the transferability measurement results of mahalanobis distance, Kullback-Leibler Divergence and Maximum Mean Discrepancy (MMD) between data domains are compared and analyzed. And the stability and accuracy of mobility measurement results are compared and analyzed by using statistical analysis and anomaly detection. Finally, a case study is carried out using the gearbox data set to verify the effectiveness and superiority of the transferability measurement based on MMD method.