A feature separation transfer network with contrastive metric for remaining useful life prediction under different working conditions

Data-driven remaining useful life (RUL) prediction methods have demonstrated excellent performance in recent years. Among these, transfer learning (TL) is widely adopted for cross-condition RUL prediction due to its ability to mitigate feature discrepancies across domains. However, most existing TL methods focus primarily on the global alignment of shared features, neglecting subdomain-specific features from different degradation stages and the impact of domain-private features. In this paper, we propose a feature separation and adaptive alignment model to address this limitation. First, a feature separation network is designed to decompose the deep features into two categories: shared features, which capture the inherent degradation patterns, and domain-specific features, which account for heterogeneity across varying operating conditions. The shared features are further categorized into subdomains based on different degradation stages. To ensure effective alignment, we develop a deep adaptive alignment method that facilitates both global alignment of the shared features and local alignment of the subdomain-specific features. Additionally, a contrastive metric module is introduced to enhance the representativeness of the features, which has been shown to improve both feature separation and alignment effectiveness. Experimental results on two benchmark datasets demonstrate that our proposed method outperforms existing approaches across various evaluation metrics.