A Component-Decoupled and Physics-Constrained Hybrid Modeling Framework for Turbojet Engine Performance Prediction

Accurate turbojet engine performance prediction is crucial for condition monitoring, health management, and safe operation. Conventional component-level models require iterative solutions of strongly nonlinear matching equations and are sensitive to un-modeled effects, limiting accuracy and computational efficiency. Purely data-driven models are efficient but lack explicit physical constraints, resulting in poor interpretability and generalization outside the training domain. To address these issues, this paper proposes a component-decoupled, physically constrained hybrid modeling framework for turbojet engine steady-state performance prediction using on-board measurements. The engine is decomposed into component-level neural sub-models, with physics-guided feature engineering and mutual-information-based feature selection applied to optimize inputs. Component predictions are coupled via aerothermodynamic constraints to reconstruct unmeasured parameters and thrust. Validation on steady-state test data from a 120 kgf class micro turbojet engine shows the model achieves 1.157% maximum relative deviation (MRD) and 0.226% average relative deviation (ARD) for thrust, with MRDs of key gas path parameters within 0.3%. Compared with purely data-driven models, it offers higher accuracy, better generalization, and physically consistent unmeasured parameter estimates, providing a practical approach for engine performance prediction and health management.