Prediction Properties of the Process Variance Using the Combined Array

Abstract

Industrial experimentation now recognizes that for a characteristic of interest, the process variability is just as important as the process mean. The Japanese industrial engineer, Taguchi (see Taguchi and Wu [9]), pioneered this area through his robust parameter design. Much of the recent work in robust parameter design has focused on the combined array (Welch et. al [10], Shoemaker et al. [7], Box and Jones [1]). Robust parameter design assumes that the experimental factors separate into two classes: control and noise. Control factors are those under the direct control of the experimenter both in the experiment and in the process. Noise factors are those which for some reason or another are not controllable in the process, and thus are random in the process, although they are to some extent under the experimenter’s control in the actual experiment. The combined array runs a single experiment in the control and noise factors, treating the noise factors as fixed effects, and allows the analyst to employ modifications of traditional response surface methodology. Myers et. al [4] show how to construct separate response surfaces for the process mean and the process variance. These surfaces can be used to determine optimum operating conditions in terms of the control factors.