05 - Propulsion and Energy
FLIGHT-DATA-DRIVEN PHYSICS-INFORMED COMPONENT MAP DERIVATION WITH STRUCTURAL COMPLIANCE
Y. Liu¹, L. Gou¹, Y. Huang, Shenzhen Research Institute of Northwestern Polytechnical Univer, China; ¹School of Power and Energy, Northwestern Polytechnical Universit, China
High-fidelity component performance maps are a prerequisite for reliable aero-engine simulation, control verification, and health monitoring. In engineering practice, however, map information is sparse and heterogeneous: legacy databases and rig or CFD data describe non-installed components, whereas the operating points actually realised by the target engine are accessible only through a small number of quasi- steady samples extracted from flight records. Purely data-driven derivation cannot reconcile these heterogeneous sources without producing non-physical artifacts such as speed-line crossing, local trend reversal, or aggressive extrapolation in regions not covered by flight evidence. This paper proposes a flight-data-driven, physics-informed neural network (PINN) framework for component- map derivation under structural compliance. Quasi-steady flight-derived points, emulated in the present controlled benchmark, are assigned the highest confidence, while legacy, rig, and CFD information provide lower-weight evidence and a topology prior. A compact descriptor of inter-curve spacing, slope consistency, and curvature variation conditions a bounded deformation of the baseline map. The training objective combines source-weighted data matching with thermodynamic consistency and explicit non-crossing, monotonicity, smoothness, and boundedness penalties. Across 48 sparse-calibration trials, the proposed method reduces mean pressure-ratio error and maintains near-zero structural violations; its mean efficiency error is comparable to scaling-factor adaptation but has a wider spread. A stricter valid operating-point recovery criterion improves by 4.76 percentage points over scaling-factor adaptation. Ablation identifies the topology constraint as the dominant contributor, while the thermodynamic residual acts as a physical safeguard with a smaller aggregate effect in the sampled domain.
DOI: 10.71945/icas2026_0690