A Research Of Aerodynamic Data Modeling Based On Generative Adversarial Network

Aerodynamic research has a long history and is vital to the development of aerospace industry and national security.The accuracy of the traditional data modeling methods for the highly nonlinear partial differential equations needs to be improved.The development of computer technology and neural networks has brought dawn to aerodynamic data modeling.It requires a lot of data to obtain a high-precision neural network model,but aerodynamic data is difficult to calculate and expensive.Therefore,this thesis establishes a generative adversarial network(GAN)model,which can automatically learn the distribution of samples and generate realistic data.The main work of this thesis is as follows:(1)Design GAN model for aerodynamic data modeling.Most GANs are composed of convolutional neural networks and are not suitable for aerodynamic data modeling.For this purpose,this thesis presents a GAN aerodynamic data model based on multi-layer perceptron and multi-radial basis.(2)To solve the problems of insufficient aerodynamic data and unstable GAN training,Wasserstein GAN was introduced.The common implementation methods of WGAN are weight clipping and gradient penalty.Weight clipping method will reduce the robustness of the network,and gradient penalty method will cause the gradient to disappear.Therefore,this thesis establishes a WGAN model based on difference,which can enhance the local continuity of the discriminator loss function,make network more stable in small datasets.(3)When the discriminator makes a correct decision on the generated samples,GAN will have a saturation problem.For this purpose,least square generative adversarial network is constructed,which can narrow the distance between any generated samples and the real samples at any time.This model converges too fast in the early stage and the error rebounds in the later stage.This thesis proposes a GAN model based on the least square method and cross-entropy.This model can combine the advantages of these two loss functions to make the model more robust.(4)Aiming at the difficulty of solving aerodynamic equations and the randomness of GAN-generated data,an aerodynamic data model based on conditional generative adversarial network(CGAN)is proposed,which makes the network suitable for complex aerodynamic datasets.Experiments were performed on multiple datasets to prove the universality of the model proposed in this thesis in modeling aerodynamic data.For the model design experiment proposed in this thesis,it is analyzed that the WGAN model based on difference performs well in small datasets.The CGAN model based on the least square method and cross entropy can combine the advantages of two loss functions to make the model more stable.Compared with other literature models,the superiority of the model in the aerodynamic data regression modeling is verified.