Data-driven method embedded physical knowledge for entire lifecycle degradation monitoring in aircraft engines

The degradation of aircraft engine performance necessitates real-time monitoring. We introduce a real-time performance-degradation monitoring method comprising a baseline state (BLS) model for pristine engine performance prediction and a real-time model for ongoing engine performance prediction. Both models are LSTM-based and integrated with the engine physical topology; however, they vary in their input parameters. To counteract the data distribution disparities across various engine states during model training, a data augmentation method is proposed. In a case study, data from 600 h of engine running were employed, with thrust as the target monitoring parameter. The initial 50 h dataset covers the entire range of working conditions for the subsequent 550 h, serving as the training dataset for both models. The results demonstrated the accuracy of these models, with the BLS model achieving a mean absolute relative error (MARE) below 0.5% and a maximum absolute relative error (Emax) below 8% in the initial 50 h. Simultaneously, the real-time model yielded a MARE below 0.7% and Emax below 10% for the subsequent 550 h. A comparison of the real-time thrust degradation between this study and traditional methods highlights the feasibility and advantages of this approach.