Physics-Informed Neural Network Based Onboard Model for Aeroengine

Monitoring the overall performance status of aeroengines is critical for ensuring flight safety, control systems, and prognostic health management. This paper proposes an onboard real-time modeling method based on physics-informed neural networks (PINN) with physics-based models. First, a distributed PINN modeling approach is adopted, where the PINN for rotating components is integrated into the overall physical model of the engine. This model is trained using measurable parameters from the engine, minimizing the output parameters containing physical constraints in each PINN to obtain an accurate rotating component model. Secondly, using the established rotating component PINN model and the physical mechanism models for the non-rotating components, the rotor dynamics of the PINN-based oboard engine model is updated, allowing for the real-time update of engine states. Finally, theoretical analysis demonstrate that the PINN-based onboard model offers superior real-time performance compared to the physics-based component-level model, thereby meeting the requirements for onboard real-time applications. Simulation results show that the proposed PINN-based onboard model outputs more detailed aerodynamic and thermodynamic parameters and more accurately tracks the dynamic behavior of the reference engine. Its mean relative error is only 14% that of the traditional linear onboard model and approximately 25% that of the purely data-driven neural network model. The proposed method preserves the accuracy and detailed physical information parameters, fulfilling the requirements for onboard real-time usage.