Application of neural networks and transfer learning to turbomachinery heat transfer

Abstract

Model-based predictive maintenance using high-frequency in-flight data requires digital twins that can model the dynamics of their physical twin with high precision. The models of the twins need to be fast and dynamically updatable. Machine learning offers the possibility to address these challenges in modeling the transient performance of aero engines. During transient operation, heat transferred between the engine’s structure and the annulus flow plays an important role. Diabatic performance modeling is demonstrated using non-dimensional transient heat transfer maps and transfer learning to extend turbomachinery transient modeling. The general form of such a map for a simple system similar to a pipe is reproduced by a Multilayer Perceptron neural network. It is trained using data from a finite element simulation. In a next step, the network is transferred using measurements to model the thermal transients of an aero engine. Only a limited number of parameters measured during selected transient maneuvers is needed to generate suitable non-dimensional transient heat transfer maps. With these additional steps, the extended performance model matches the engine thermal transients well.