Research On Approximate Modeling Method For Aircraft Multi-discipline Simulation Based On Ensemble Learning

In recent years,many studies on multidisciplinary design optimization(MDO)have been carried out to improve the overall performance of aircraft.However,MDO requires a large number of calls to the simulation model,which will result in a huge computational cost(for example,a single high-precision structure/fluid analysis takes several tens of hours).In order to improve the computability of MDO,an effective way is to develop the approximate modeling method,but there are some chanllenges.Therefore,this paper carryies out systematic research on the chanllenges of finite samples,non-linearity and high dimension,and finally forms a series of high,medium and low dimensional approximate modeling method.The methods proposed in this paper are applied to the aerospace field by taking the aircraft traffic prediction onboard and the guidance deviation analysis of the hypersonic gliding aircraft as examples,the effectiveness of the above methods is verified by experiments.At first,the approximate modeling method based on ensemble learning is studied.The ensemble based on a single surrogate and the ensemble based on multiple surrogates are mainly discussed.For the ensemble based on a single surrogate,to solve the problem of worse basic function selection in the traditional polynomial response surface method under finite samples,the polynomial response surface based on basis function selection by multitask optimization and ensemble modeling method is proposed.The proposed method successfully realizes better basis function selection and improves the accuracy and robustness of the final model by dividing the training set based on cross-validation method,building the sub-models according to multitask optimization and combining multiple sub-models.for the ensemble based on multiple surrogates,considering for different problem the robustness of a single surrogate is poor,this paper proposes a partial weighted aggregation modeling method.By referring to the ideas of identifying outliers,the method adaptively eliminates the inaccurate candidate models and improves the performance of the final model.Finally,the effectiveness of the above methods is verified by experiments,and the advantages and disadvantages of the methods are investigated.Secondly,the high-dimensional approximate modeling method is studied.For the high-dimensional approximation problem,the traditional modeling method becomes very difficult or even impossible due to the “curse of dimensionality”.In order to solve this problem,we mainly discuss the dimensionality reduction of the decomposition method and the dimensionality reduction of the space transformation method.for the dimensionality reduction of the decomposition method,this paper introduces the high dimensional model representation method,which decomposes the high-dimensional problem into a finite number of low-dimensional approximation problems and then models these low-dimensional problems.Therefore,the high-dimensional approximation ability of model could be improved.For the dimensionality reduction of the space transformation method,considering that the high dimensional model representation method is not suitable for the problem of ultra-high dimension and strong coupling,this paper develops a supervised feature learning method,which is manifold Gaussian process for regression.The proposed method maps high-dimensional variable space to low-dimensional feature space through neural network,which makes the traditional approximate modeling method applicable.Finally,the effectiveness of the above methods is verified by experiments,and the advantages and disadvantages of the methods are investigated.At last,aircraft traffic prediction onboard and the guidance deviation analysis of hypersonic gliding aircraft are taken as examples.According to the application background,the approximate modeling method proposed in this paper is successfully applied in the aerospace field.Besides,the effectiveness of the above methods is verified by experiments.This paper could provide reference and guidance for the application of the approximate modeling method in the aerospace field.