A transfer regression network-based adaptive calibration method for remaining useful life prediction considering individual discrepancies in the degradation process of machinery

Abstract

Transfer learning (TL)-based remaining useful life (RUL) prediction has been extensively studied and plays a crucial role under cross-working conditions. While previous works have made great efforts to realize domain adaptation, existing methods still suffer from two key limitations: (1) Most feature-based TL methods focus on learning shared domain-independent features and fail to capture private domain information. (2) Model-based TL methods typically use all features to pre-train an RUL prediction model without accounting for the negative effect of private features in the source domain to the target domain. To tackle these challenges, a transfer regression network-based adaptive calibration (TRNAC) method is proposed to execute accurate RUL prediction for different machines where shared domain-independent features and private individual features in the target domain are fully considered to enhance the feature representation for RUL prediction. Specifically, the constructed TRNAC model includes a set of feature extractors where one is to learn shared domain-independent features in both domains and another is to extract individual domain features in the target domain, a shared RUL regressor to learn a mapping relationship between the shared features and the RUL values, a domain discriminator to distinguish which domain the feature comes from. Most importantly, an error regressor is customized by designing a dynamic calibration factor to revise the prediction error caused by the shared RUL regressor and achieve accurate prediction. The comprehensive experimental results on the aero-engine dataset and bearing dataset indicate that the proposed method performs better than other state-of-the-art RUL prediction methods.