Matar, Joseph E.; Lochner, Robert H. (1988, ASQC) Marquette University, Milwaukee, WI; W.A. Golomski & Associates, Chicago, IL
Dr. Genichi Taguchi has had a dramatic impact on the quality movement in the United States. His methods gained wide acceptance from their first introduction in this country. There have been many successful applications of his approach to quality of design. But, even as enthusiasm for Taguchi Methods is growing, there are some statisticians expressing concern and occasionally opposition to some of Taguchi's ideas.
In this paper we will address one of the most persistent concerns the appropriateness of some of Taguchi's experimental designs. It is not our intent to detract from the important contributions Dr. Taguchi has made to quality improvement. Instead, we will point out how the approach he developed can be enhanced using existing, "Made in the USA" experimental designs.
The experimental designs which are part of the Taguchi Method were developed primarily by British and American statisticians prior to their reintroduction by Taguchi. The implications of orthogonality were also well understood by many statisticians and engineers. What is different about Taguchi presentation is that he has:
In this article, we will explain how certain fractional factorial designs, which are not among the Taguchi designs, can be used with the Taguchi approach to design quality into a product. These designs are as easy to use as Taguchi's, require no more observations, but permit more flexibility in the estimation of the effects of factors and their interactions. Traditional Taguchi designs do not provide options with regard to which interaction effects are to be estimated. As a result, unimportant interactions are sometimes estimated, while the effects of key interactions are confounded with main effects.
Consider a simple example. Suppose a Taguchi design involved three inner array variables A, B and C, and three outer array variables d, e and f. Taguchi refers to inner array factors as design parameters, the outer array as the noise or uncontrollable factors. A Taguchi experimental design would allow estimation of such interactions as Ad, Ae, Cf, but would have the AB interaction confounded with the C main effect. Suppose A, B and C were all found to be significant factors. Could it be that only A, B and the AB interaction are important, and C has no effect? With a Taguchi design we don't know. But, with an alternate design we can choose to confound Ad with f, say, and obtain an estimate of AB which is not confounded with other main effects. Standard fractional factorials can be generated to reflect which factors are confounded.
A number of commercial computer programs can be very helpful to the engineer and one of these, such as "JASS", will be used in the article. The paper will bring together both the engineering and statistical aspects of designed experiments in a manner easily understood by the engineering community.