Research On Kriging-based Aerospace Vehicle Multidisciplinary Surrogate Modeling

Recently,the competition in the aerospace field becomes gradually fierce so that the design level of aerospace vehicles is facing higher requirements.The design of aerospace vehicles associates with multiple disciplines,including aerodynamics,structural mechan-ics,thermodynamics and so on.To improve the accuracy of disciplinary analysis,high-fidelity disciplinary models are required.If they are nested in multidisciplinary optimiza-tion of aerospace vehicles directly,it will certainly lead to high computational cost.To reduce the computational cost,surrogate modeling is often used.As a probabilistic sur-rogate,the Kriging model has good approximation ability and unique error estimation function,which attracts broad attention.Therefore,Kriging-based aerospace vehicle mul-tidisciplinary surrogate modeling is studied in the thesis.Aiming at the inaccuracy of sur-rogates with limited data,the trend function,stochastic process and acquisition function of the Kriging model are focused on,and a complete surrogate method is established and applied to the aerospace field.The main research of the thesis is concluded as follows:(1)To select basis functions and optimize regression coefficients for the trend func-tion of the Kriging model,a penalized Kriging model is proposed.Based on regularization techniques,the proposed method selects basis function and optimizes regression coeffi-cients adaptively,constructing a better trend function to increase the prediction accuracy.(2)To select basis functions and optimize regression coefficients and correlation pa-rameters for the trend function and stochastic process of the Kriging model,a penalized blind likelihood Kriging model is proposed.Based on regularization techniques,the pro-posed method selects basis function and optimizes regression coefficients and correlation parameters adaptively,constructing a better trend function and stochastic process to in-crease the prediction accuracy.(3)To construct an acquisition function for the sequential Kriging metamodeling method,a gradient and Hessian enhanced sequential metamodeling approach is proposed.Based on the bias-variance decomposition theory,the proposed approach uses the gradient and Hessian of the Kriging model to estimate the bias term,balancing the global explo-ration and local exploitation of the acquisition function better to increase the prediction accuracy.(4)To improve the multidisciplinary metamodeling accuracy of an earth observation small satellite,the performance of several Kriging models are compared,and the exper-imental results show that,in most numerical examples,the penalized blind likelihood Kriging model achieves the highest accuracy,which demonstrates its advantage in the multidisciplinary metamodeling of an earth observation small satellite.(5)To improve the prediction accuracy of the deviation propagation model of a hy-personic gliding vehicle,the performance of several Kriging models and sequential meta-modeling methods are compared.The experimental results show that,in most numerical examples,the penalized blind likelihood Kriging model is more accurate than the state-of-the-art Kriging models,and the gradient and Hessian enhanced sequential metamodeling approach outperforms the state-of-the-art sequential metamodeling methods and one-shot metamodeling method,which demonstrates the effectiveness of the proposed methods in the metamodeling of the deviation propagation model of a hypersonic gliding vehicle.The thesis studies the methods to improve the prediction accuracy of the Kriging model with limited data,and provides the theoretical support for multidisciplinary opti-mization of aerospace vehicles with low computational cost.