Physics-Informed Neural Networks for Multi-Spectral Radiation Thermometry with Emissivity Physical Constraints

To resolve the intrinsic under-determined nature of multi-spectral radiation thermometry in turbine blade monitoring, this paper proposes an inversion framework based on Physics-Informed Neural Networks (PINN). Unlike conventional black-box models, the proposed framework transforms Planck's Law and emissivity smoothness priors into active loss-shaping constraints, effectively restricting the solution space to a physically consistent manifold. This structural innovation ensures thermodynamic integrity and suppresses non-physical oscillations prevalent in traditional iterative methods. Numerical simulations demonstrate a remarkable Average Relative Error (ARE) of 0.12% even under 5% composite noise, showcasing superior denoising capabilities. Experimental validation on K444 alloy further confirms the practical feasibility, yielding maximum temperature errors of 5.5 K at 973 K and 2.8 K at 1123 K with an ARE < 1%. Moreover, the trained network achieves a three-order-of-magnitude computational speedup compared to iterative solvers, providing a robust, real-time, and physically interpretable solution for aero-engine high-temperature sensing.