Neural network modeling of unsteady aerodynamic characteristics at high angles of attack

Abstract

A neural network approach for modeling of unsteady aerodynamic characteristics in a wide angle-of-attack range is considered in the paper. Special tests have been carried out in TsAGI wind tunnel to investigate dynamic properties of unsteady aerodynamic characteristics of a generic transonic cruiser, which is a canard configuration. The aerodynamic derivatives have been studied with forced small-amplitude oscillations. In addition, forced large-amplitude oscillation tests have been carried out for the detailed investigation of dynamic effects on the unsteady aerodynamics in the extended flight envelope. To describe the nonlinear dependences of aerodynamic coefficients, observed in the dynamic experiments, two neural network models, which use the feed-forward and recurrent architectures, are developed and compared. A special regularization for the neural networks training taking into account that data are obtained in different experiments with different noise level is proposed to improve a model performance. The unsteady aerodynamics modeling results obtained with neural networks are compared with a state-space model results.