A Novel Transfer Learning Method Based on Domain Adversarial Networks for Remaining Useful Life Prediction

Abstract

As a core technology of prognostics health management (PHM), remaining useful life (RUL) prediction has gained more and more attention with the approach of the Industry 4.0 era. However, there are still relatively few studies dedicated to transfer learning to consider the applicability of models in RUL prediction. So in this paper, a transfer learning method based on domain adversarial network (DANN) is proposed for the case where the target domain data is unlabeled, and the source domain data is labeled. In the predictor branch, a multi-scale separable convolution is used in order to enhance the feature extraction process. In addition, an attention mechanism is added to perform adaptive correction of the extracted features. An unbalanced loss function is also proposed in order to reduce the lagged prediction points. In the domain classifier branch, domain invariant features are extracted by setting a gradient inversion layer between the feature extraction module and the domain classifier. The proposed method is validated on the dataset of NASA Commercial Modular Aero-Propulsion System Simulation (C-MAPPS), and the experimental results have verified the superiority of the method.