Application of kernel principal component analysis to multi-characteristic parameter design problems

Abstract

The Taguchi method for robust parameter design traditionally deals with single characteristic parameter design problems. Extending the Taguchi method to the case of multi-characteristic parameter design (MCPD) problems requires an overall evaluation of multiple characteristics, for which the principal component analysis (PCA) has been frequently used. However, since the PCA is based on a linear transformation, it may not be effectively used for the data with complicated nonlinear structures. This paper develops a kernel PCA-based method that allows capturing nonlinear relationships among multiple characteristics in constructing a single aggregate performance measure. Applications of the proposed method to simulated and real experimental data show the advantages of the kernel PCA over the original PCA for solving MCPD problems.