Interpretability Analysis and Transferability Evaluation of Domain-Adversarial Regression Adaptation Model based on Tensor Representation

In recent years, deep transfer learning technology has been introduced to solve online remaining useful life (RUL) of rotating machinery under unknown working conditions. However, the self-driven prediction of online sequentially arrived data lacks sufficient model explanation and lack of credibility. To solve these problems, a tensor domain adversative regression prediction method is proposed with adaptive knowledge transfer. First, the tensor Tucker decomposition was used to find the key features in RUL transfer prediction, and the interpretability was analyzed at the geometric level. Then, the weighting mechanism was constructed, which was jointly optimized with tensor domainadversarial training to achieve online RUL transfer prediction. Second, the degradation trend information was extracted based on core tensor. And the multi-scale evaluation criteria of transferability were established by combining the time series information measurement at different levels, in order to clarify the contribution and rationality of offline data and online data and provide an intuitive migration understanding. Finally, this paper takes rolling bearing as an example, and runs experiments on IEEE PHM Challenge 2012 and XJTU-SY datasets, respectively, and the experimental results verify the effectiveness of the proposed method.