Development And Application Of Extended Adaptive Hybrid Surrogate Model

Design and optimization of mechanical parts and systems is an important part of the field of mechanical design,nowadays experiments are gradually replaced by computer simulation experiments.However,for complex engineering design problems,due to the complexity of computer simulation model,a single run may take hours or even days.This makes design,optimization and model analysis almost impossible to complete.Surrogate model is based on a small set of training points,with statistical techniques,a time-consuming computer simulation model can be replaced by an approximate model.Surrogate model makes it possible to obtain a relatively accurate result with only a few run of computer simulations.Thus,it is suitable for subsequent optimization and analysis.However,surrogate models are usually not robust enough when dealing with different problems.Hybrid surrogate model is a combination of different surrogate models,and is designed to take advantage of the positive characteristics of the existing surrogate models,while eliminating the effort of selecting the appropriate surrogate model.This paper is focusing on surrogate-based engineering design and optimization,making a theoretical improvement on hybrid surrogate model.Both modelling accuracy and robustness is studied,and an example on applying the proposed method in the global sensitivity analysis of a tunnel boring machine's cutterhead driving system is presented.The main researches of this paper is summarized as below.(1)Based on the analysis of popular standalone surrogate models and hybrid surrogate models,this paper proposes an Extended Adaptive Hybrid Function(E-AHF).This method mainly includes two parts: a.Model selection,b.Adaptive weight calculation.Model selection is a pre-evaluation of the modeling accuracy of standalone surrogate models.Cross-validation is used to eliminate any poor-performing surrogate models.Gaussian-process error,together with determined basis function,is used as a basis for adaptive weight factors determination.(2)Surrogate model is a black-box model,to ensure the modeling robustness,this paper establishes a "test function library" composed of 38 test functions for performance tests.In addition,the impact of sample size on modeling accuracy is also considered,and 5 different sampling sizes are applied for each test.In this paper,three model precision evaluation criteria are adopted."Coefficient of Determination"(2R)and Root Mean Square Error(RMSE)are used for global error assessment,and the Maximum Absolute Error(MAE)is used for local error assessment.(3)For the study of surrogate-based analysis and optimization,a numerical optimization example and an engineering application example are applied to evaluate E-AHF method.To simulate general situation of engineering optimization,multi-constrained optimization,including linear and nonlinear constraints,is used to compare the optimization results of different surrogate models with the true optimum.A surrogate-based global sensitivity analysis of the dynamic response of the cutterhead drive system of a tunnel boring machine(TBM)is carried out to meet the requirements of large computational cost,and the modeling accuracy of different agent model methods is compared.