A Novel Progressive Fusion Approach for Remaining Useful Life Prediction Using Temporal-Sensor Data

Accurate prediction of Remaining Useful Life (RUL) is vital for proactive maintenance and failure prevention in industrial systems. This paper introduces the Spatiotemporal Homogeneous Feature Extractor (STSF) to address limitations in existing RUL prediction methods. STSF employs a flexible, layer-wise progressive feature fusion technique that harmonizes spatial and temporal features, enhancing the model's ability to capture complex degradation patterns. To further improve prediction accuracy, the Feature Space Global Relationship Invariance (FSGRI) training method is used, grounded in supervised contrastive learning. FSGRI maintains consistent relationships between sample features and degradation trajectories during training, simplifying subsequent regression tasks. Evaluations on a C-MAPSS dataset demonstrate that STSF outperforms baseline models across multiple metrics, highlighting its effectiveness in RUL prediction.