A Global Adversarial and Local Contrastive Transfer Learning Approach for Remaining Useful Life Prediction

The transfer learning method can effectively mitigate the reliance on an extensive amount of target-domain data for the construction of remaining useful life (RUL) prediction models. While transfer learning has achieved remarkable achievements in numerous real-world applications, traditional methods often focus on aligning global domain features and extracting domain-invariant features, overlooking local and channel-specific characteristics. This limitation can result in suboptimal RUL predictions, as critical degradation-related features may be ignored. To address these challenges, this article proposes a global adversarial and local contrastive (GALC) transfer learning approach for predicting RUL in multiple domains. The proposed method employs adversarial learning to capture global domain-invariant features while incorporating channel- and temporal-level contrastive modules to preserve local temporal patterns and channel uniqueness. This approach enables us to achieve domain-invariant feature learning while preserving channel distinctiveness across diverse domains, and its effectiveness has been validated using real-world datasets.