Neural network technologies in hardware-in-the-loop simulation: comparison of NARX and GRU architectures

Abstract

The principle for realization of mathematical models of gas-turbine engines in the form of recurrent neural networks and their application in complex hardware-in-the-loop modeling for tuning automatic control, condition-monitoring and diagnostic systems is considered. Comparison of the NARX and the GRU architectures is carried out, the methodology of the neural net-work gas-turbine modelling and the realization of the model at the hardware-in-the-loop test-bed are described. The results of hardware-in-the-loop simulation of the aircraft engine parameters with a real control system are presented. The accuracy and adequacy of the constructed model is analyzed. The development of such technologies allows to create intelligent models, which can be used in digital twins of complex technical systems.