End-to-end remaining useful life fusion prediction for mechanical equipment based on multiscale informer

To mitigate the cumulative errors inherent in traditional two-stage remaining useful life (RUL) prediction frameworks, where feature extraction and temporal modeling are decoupled, this study proposes an end-to-end fusion prediction method based on a multi-scale Informer architecture. The model directly maps raw multi-state sensor data to RUL estimates, eliminating manual feature engineering. Multiple Informer networks are designed to capture degradation patterns at distinct time scales. A dynamic weighting mechanism allocates weights to each scale-specific predictor based on their temporal prediction errors, followed by weighted ensemble learning for fusion. The integrated end-to-end architecture ensures seamless gradient propagation during training, while the adaptive multi-scale fusion mechanism significantly improves generalization capability by effectively leveraging temporal features across different time scales. Validated on both the C-MAPSS aeroengine simulation dataset and an experimental rolling bearing degradation dataset, the proposed approach achieves lower root mean square error (RMSE) and mean absolute error (MAE), along with a higher R², compared to stateof-the-art methods. These results, derived from hybrid data sources including both simulation and physical experiments, confirm the method’s enhanced prediction accuracy and superior robustness under varying operational conditions.