Improved multiple penalty mechanism based loss function for more realistic aeroengine RUL advanced prediction

Abstract

The aeroengine remaining useful life (RUL) prediction is conducive to formulating maintenance plans, assisting maintenance decisions, and improving the intelligent operation and maintenance level. When the engine is in a degraded state, the maintenance personnel tend to prediction advance rather than prediction delay. However, the current RUL prediction researches mainly focus on accurate prediction, and pay little attention to the realistic demand of advanced prediction. Aiming at this problem, this paper proposes a multiple penalty mechanism (MPM) based loss function combined with similarity RUL prediction. This research first uses multi-dimensional sensor data to construct a health index (HI) that characterizes the engine health status, then matches the HI similarity by derivative dynamic time warping corrected with different sequence length (DDTW-DSL). Finally, the MPM loss function assists the neural network to realize the mapping from HI to RUL. The method is verified by NASA's Commercial Modular Aero-Propulsion System Simulation dataset. The results show that compared with the traditional RMSE loss function, the MPM loss function can significantly improve the advanced prediction probability and effectively avoid RUL prediction lag. Compared with the existing methods, the novel method has advantages in both RUL prediction effect and model complexity.