A reduced-order aerodynamic model with high generalization capability based on neural network

Generalization capability of neural network model names the prediction ability of the model to new input signal. However, current nonlinear aerodynamic models based on neural network usually lack such capability. To cover this shortage, a CFD-based aerodynamic model using differential evolution with validation signal is proposed and improved to predict transonic aerodynamic load. To minimize root mean square error of validation signal, the widths of neural centers in hidden layers are optimized by introducing differential evolution in this recursive radial basis function neural network. Training signal is large-amplitude displacement of that structural, while the validation signal is small-amplitude structural displacement. The results not only indicate the shorter calculating time of this reduced-order model than that of full-order numerical calculation, but a higher generalization capability of the model under various frequencies and amplitudes due to the introduction of small-amplitude validation signal.