Online adaptive transfer learning for aeroengine on-board gas path modeling under domain shifts

The on-board gas path model is crucial for gas path analysis and aero-engine health management. However, domain shifts caused by gas path degradation, component replacement, and engine differences challenge on-board modeling. Traditional offline transfer learning (TL) methods struggle with the online nature of sensor data. This paper proposes an online adaptive transfer learning (OATL) method integrating TL with online learning (OL) to enable real-time adaptation. Using a component-level model (CLM), a stacked LSTM network is trained under clean conditions. The OATL method combines online fine-tuning and dynamic weight adjustment between source and target domains for rapid adaptation. Validated on simulated and actual flight data, OATL significantly improves adaptability and prediction accuracy, outperforming traditional methods. It provides an effective solution for real-time aeroengine modeling under continuous domain transitions. It provides an effective solution for on-board gas path modeling in the complex operating environments of aeroengines under conditions of domain shifts.