Research And Application Of Metamodel Based On High Dimensional Model Representation And Support Vector Regression

Metamodeling techniques are more and more widely used in engineering andscientific areas by replace the orginal computational expensive model or implicitfunction with a less computational expensive model. Traditional metamodelingtechniques could handle nonlinear problems with low dimesion. But when the dimensionrises, them are uncapable to handle, because when the dimension rises the cost ofconstruction a metamodel increases exponentially, and this is the famous ‘curse ofdimensionality’.High dimensional model representation (HDMR) is a set of functions to express thecorrelation of input variables and their dependent/independent contribution to theresponse of a design problem. HDMR provides a hierarchy of correlated function. Thisthesis proposed a new metamodeling technique and its construction algorithm bycombining HDMR and support vetor regression to conquer the curse of dimensionality.The proposed technique inherits the hierarchical sturcture of HDMR and smallsample set learning characteristic. It can handle high dimensional problems with lowcost.Now that designing and optimization of a box beam is a difficult task due to itscomplex structure and high dimension, this thesis developed a new tool to provideparametric modeling, simulation and optimization of box beam by MATLAB andANSYS. And in this tool, the proposed metamodeling techniques was intergated toconquer the high dimension of metamodeling a box beam structure.The box beam structure of a ZK5540machine tool was designed and optimized bythe proposed metamodeling technique and the developed tool. The cost of the machinetool developemnt was greatly decreased and performance was highly increased. Theproposed technique and tool were proved to be efficient and effective.