Aero-engine Simulation Model Performance Tracking Method Based on Online Transfer Learning

Accurate simulation models of aero-engines are crucial for assessing engine status. The presence of engine performance dispersion within the fleet, along with varying degrees of degradation in individual engines during actual usage, can introduce deviations in the simulation model. These deviations may subsequently result in an increase in error between actual data and simulation parameters, leading to potential misdiagnosis. Furthermore, the limited availability of outfield engine data poses challenges in generating highly precise personalized simulation models. This paper proposes an online transfer learning method to address the issue of adaptive component performance. Firstly, a depth-domain adaptive method based on reducing geometric distance in hidden space is adopted, and a pre-trained model fitted with large amounts of historical data is migrated to obtain a personalized model reflecting the state of an individual engine using small samples of outfield flight data. Secondly, semi-supervised learning is introduced into online updating of the model, enabling real-time tracking of actual engine states as new flight data constantly arrive and ensuring that the model adapts to changes in engine states. The method effectively leverages engine operation data collected over an extended period of time in various states, thereby addressing the challenge of limited flight data availability in the field. By updating the data and model in both offline and online phases, a novel performance tracking model is developed for the target engine. Furthermore, empirical evidence validates the efficacy of our proposed algorithm in mitigating model performance prediction errors, demonstrating a significant improvement in accuracy compared with the generalized model.