Remaining Useful Life Prediction Method of Aero-engine Based on Asymmetric Loss Extreme Gradient Boosting and Dual-stream Transformer Stacking Integration

Abstract

The remaining useful life (RUL) prediction of aero-engine is the core part of prognostics and health management (PHM), which is of great significance for ensuring flight safety and optimizing maintenance strategies. Aiming at the insufficient degradation feature extraction of existing data-driven methods when dealing with high-dimensional sensor time series data, and the engineering application limitations of insufficient prediction overestimation penalty caused by the widespread use of symmetric loss functions, this paper proposes an aero-engine RUL prediction method based on asymmetric loss extreme gradient boosting (XGBoost) and dual-stream Transformer stacking integration. The sliding window is used to construct the time series input, and the multi-dimensional manual features including flattening vector, statistics, trend and first-order difference information are extracted to form a two-stream input. A custom XGBoost model with asymmetric gradient loss and a two-stream Transformer network with manual features are constructed as the first layer of heterogeneous base learners. The former suppresses the prediction overestimation from the source, and the latter effectively captures long-range temporal dependencies. Finally, stacking integration is performed by Ridge regression, and global bias is searched on the verification set to further fit the C-MAPSS asymmetric scoring function. The experimental results on NASA C-MAPSS FD001 dataset show that the proposed method achieves better performance with RMSE = 11.30, MAE = 8.24, R2 = 0.9204 and Score = 179.40. Compared with the representative methods in recent years, the score index Score is greatly reduced compared with the sub-optimal method, and the regression accuracy is maintained while the engineering cost is effectively controlled. It provides a new way to balance the accuracy and engineering safety preference for aero-engine RUL prediction.