Attention-Based Model With Component Embedding for Aero-Engine Performance Estimation

Prognostics and Health Management (PHM) technology plays a vital role in the condition-based maintenance of aeroengines. Real-time models that monitor and estimate aeroengine performance are essential in this domain. The precision, stability, and interpretability of models are paramount for aeroengine research. However, both conventional data-driven models and physics-based models grounded on the component characteristic exhibit areas that can be further refined and improved. Therefore, this study presents a novel approach to improve predictive precision by incorporating engine domain knowledge into performance estimation models. We introduce a component embedding technique that effectively captures the concealed interdependencies among parameters in the corresponding component. The features are mapped to a high-dimensional space, characterizing the different engine components separately. Furthermore, we present a pioneering model for aero-engine performance estimation based on the attention mechanism. The attention-based model implements an encoder-decoder architecture. The multiple stacked encoders are responsible for capturing crucial engine performance information at different scales. The decoder utilizes the feature map generated by the encoders to produce an accurate prediction. Experimental results on the historical operating data of turbofan engines demonstrate that our proposed model outperforms conventional models in exhaust gas temperature (EGT) estimation in both normal and degraded conditions. Naturally, the presented model can also be utilized for the estimation of other performance parameters, thus providing support for the optimization of maintenance schedules.