Unsteady nonlinear aerodynamics identification based on neural network model

This article develops an auto-regressive radial basis function (AR-RBF) neural network model based on the standard RBF neural network model. The computed aerodynamic loads by the time domain CFD method are set as the input signals, and an unsteady nonlinear aerodynamic model can be constructed by the AR-RBF. The direct CFD results are used to validate the precision of the model. Comparison of the two neural network models in prediction performance shows that the AR-RBF neural network model performs better in precision and that it can fit well the unsteady nonlinear characteristics of the transonic flow under large amplitude oscillations of a shock wave. In addition, this model can be easily extended to multi-dimension models. The results of predicting the aerodynamic forces excited by periodic signals show that the AR-RBF neural network model trained with multi-step input signals has the ability of predicting nonlinear aerodynamic forces under harmonic vibrations of different amplitudes or different frequencies.