【作者】 王兵；

【导师】 汪文勇；

【作者基本信息】 电子科技大学 ， 工程硕士（专业学位）， 2021， 硕士

【摘要】 国家航天航空事业蓬勃发展和进步与空气动力学的研究息息相关,如何有效的进行气动建模一直是该领域专家们研究的重点问题,对于传统的基于物理模型的方法来说,由于存在一些复杂且难以求解的偏微分方程导致其计算效率无法满足实际需求。而且需要耗费很多资源。很难满足实际需求。因此有人开展了基于数据驱动的无模型方法研究,但同样存在着生成数据精度不高的问题。基于生成式对抗网络(GAN)模型在诸多领域中的良好表现,因此,本文建立了基于GAN的气动数据建模及优化方法的研究,针对GAN模型在训练过程中出现的各种问题进行了优化。本文的主要工作如下:(1)在开展基于GAN的气动数据建模之前,考虑原始GAN模型过于自由,不能够按照意愿生成指定的数据,将用条件式生成对抗网络(CGAN)来代替GAN进行气动数据建模,然后设计出每个模型的网络结构,并进行实验验证。(2)结合结果分析了GAN模型在气动数据集上的拟合表现,从气动数据集的特点出发,分析出基于GAN的气动数据建模的问题所在,并基于此提出了一个定理,即在处理来自于连续函数的非线性稀疏数据时,径向基神经网络(RBFNN)是GAN的最佳判别器。并给出了理论证明。(3)基于上述定理我们提出了基于径向基函数的生成式对抗网络(Radial Basis Function-based GAN,RBF-GAN),设计了RBF-GAN的网络结构模型,并进行了实验验证。然后结合集群神经网络的思想,提出了基于集群径向基函数的生成式对抗网络模型(Radial Basis Function Cluster-based GAN,RBFC-GAN),同样设计了RBFC-GAN的网络模型结构,并进行了实验验证。对实验结果进行分析后得到:将RBFNN作为判别器,可以提高非线性气动数据生成的精度,相对于传统GAN,RBF-GAN和RBFC-GAN生成的气动数据的精度提高一个数量级;和全连接神经网络相比,RBFNN更适合作为面向非线性气动数据的GAN的判别器;关于模型稳定性方面,从整体训练过程中看,RBFC-GAN的稳定性最好。更多还原

【Abstract】 The vigorous development and progress of the national aerospace industry are closely related to the research of aerodynamics.How to effectively perform aerodynamic modeling has always been a key issue for experts in this field.For traditional methods based on physical models,there are some complex and Partial differential equations that are difficult to solve make their calculation efficiency unable to meet actual needs.And it needs to consume a lot of resources.It is difficult to meet actual needs.Therefore,some people have carried out research on data-driven model-free methods,but there is also the problem of low accuracy of generated data.Based on the good performance of the GAN model in many fields,this thesis establishes a research on aerodynamic data modeling and optimization methods based on generative adversarial network(GAN),and optimizes the various problems that appear in the training process of the GAN model.The main work of this thesis is as follows:(1)Before developing aerodynamic data modeling based on GAN,considering that the original Gan model is too free to generate the specified data according to the will,conditional generative adversarial network(CGAN)will be used instead of GAN Carry out aerodynamic data modeling,and then design the network structure of each model,and conduct experimental verification.(2)Combining the results,we analyzed the fitting performance of the GAN model on the aerodynamic data set.Considering the characteristics of the aerodynamic data set,we analyzed the problem of aerodynamic data modeling based on GAN,and based on this,we put forward a theorem,which is When non-linear sparse data from continuous function,RBFNN is the best discriminator for GAN.And gave a theoretical proof.(3)Based on the above theorem,we proposed a radial basis function-based GAN(RBF-GAN),and designed the network structure model of RBF-GAN,and conducted experimental verification.Then combined with the idea of cluster neural network,a radial basis function cluster-based GAN(RBFC-GAN)is proposed,and the network model structure of RBFC-GAN is also designed,and we have carried out experimental verification.After analyzing the experimental results,it is obtained that using RBFNN as a discriminator can improve the accuracy of nonlinear aerodynamic data generation,Compared with traditional GAN,the accuracy of aerodynamic data generated by RBF-GAN and RBFC-GAN is improved by an order of magnitude;and Compared with fully connected neural networks,RBFNN is more suitable as a discriminator of GAN for nonlinear aerodynamic data.;Regarding model stability,from the overall training process,RBFC-GAN has the best stability.更多还原