Research And Application On Ensemble Of Surrogates In Analytical Target Cascading

Design of complex mechanical system is always a complex multidisciplinary designoptimization (MDO) problem, in which often implicit constrains and objectives with highsimulation costs exist. Generally plenty of time on computer simulations is required toanalysis these implicit problems and data samples are difficult to obtain, which limits the useof MDO methods in such complex problems. In the context of actual needs, surrogate modelsare proposed to replace the costly simulation tests. Surrogate modeling is also calledMetamodeling, which is used to establish the function relationship between design variablesand output variables with some data samples. Surrogate models can be used for complexsystems analysis and high-level simulation applications, and can greatly reduce thecomputational costs of complex problems. In recent years, many surrogate models have beenproposed and applied to engineering practice. Among all surrogate models, ensemble ofsurrogates (EoS) is a relatively novel modeling method based on linear aggregation, whichcan solve the above selection problem to some extent. This paper mainly focuses on theconcept of ensemble of surrogates, a new type of ensemble called RE (Repeated Ensemble ofSurrogates) and its application in MDO as well as engineering practice.The main work of this paper is as follows:Firstly, research status of surrogate models is analyzed and several common surrogatemodels are discussed: Polynomial Response Surface (PRS), Radial Basis Function (RBF),Kriging and EoS method.Secondly, establish policies of ensemble of surrogates are elaborated systematically anda new method (RE，Repeated Ensemble of Surrogates) to establish the ensemble is proposed.Some function examples are selected to analyze and compare the prediction ability of PRS,RBF, Kriging, EoS and RE method, which shows the validity and applicability of the new REmethod.Thirdly, RE method is integrated into Analytical Target Cascading (ATC) method toreplace implicit constrains and objectives met in MDO, which is called RE-ATC andcompared with other single surrogate model based ATC methods using two MDO examples.Finally, a structural MDO problem of a super heavy machine tool is described in detailand the proposed RE-ATC method is applied to it. This example of engineering applicationshows the way to use RE-ATC to solve complex MDO problems.