COMMENTARY ON TAGUCHI’S PARAMETER DESIGN WITH DYNAMIC CHARACTERISTICS

Abstract

A major recent advance in quality engineering and industrial statistics is the development and applications of robust parameter design for product and process improvement. This methodology, pioneered by the Japanese quality expert Dr. Genichi Taguchi, has now been widely used in industries in many countries, and many industrial case studies can be found in proceedings of statistical and quality meetings and in publications by private consulting organizations such as the American Supplier Institute. The industrial statistics research community has given a lot of attention to parameter design, which is evidenced by the rapid increase of publications on the subject. A comprehensive discussion on parameter design can be found in a Technometrics article edited by Nair.’ The main purpose of this commentary is to point out some problems associated with Taguchi’s approach to parameter design for ‘dynamic’ characteristics and to outline sound alternatives. Owing to the abbreviated nature of this commentary, readers may wish to refer to References 2 and 3 for technical details. Taguchi4 defined two general classes of application for his parameter design methodology which he referred to as ‘static characteristics’ and ‘dynamic characteristics’. Most of the interest in parameter design generated in the statistical literature has focused on static applications.” With a few exceptions such as References 6 and 7, relatively little work has appeared on the dynamic problem. Static characteristic applications involve situations where the goal can be summarized as getting the value of a quality characteristic of interest, Y, as close as possible to a single specified target value. For example, Kacker and ShoemakeP consider a process which deposited layers of silicon on top of silicon wafers. The goal was to make the thickness of the deposited layer as close as possible to a target value, Dynamic characteristic applications, on the other hand, involve situations where the performance of the system is determined by the relationship between a signal factor, M, and an observed response, Y. In these cases, the response is required to assume different values as a result of changes in the signal factor. Since the term ‘dynamic’ is somewhat misleading, Miller and Wu3 suggested that such applications be referred to as signal-response systems. Here we will use the two terms interchangeably. One general class of signal-response systems is multiple rarget applications. For example, Yano (Reference 9, p. 293) describes a process where parts are machined using a lathe. Different applications required machined parts with different degrees of surface roughness to be produced. As it was known that the feed rate of the tool bit could be used to alter surface roughness, feed rate was selected as the signal factor. How reliably surface roughness can be controlled by adjusting feed rate depends on the characteristics of the relationship between surface roughness and feed rate. It was thought that other factors such as lathe, cutting speed, depth of tool cut, type of tool cut, comer radius, cutting edge angle, front escape angle, and side scoop angle may affect this relationship, so an experiment was conducted to find preferred settings for these factors. A second class of signal-response applications is measurement systems. A measurement system is the process used to obtain an estimate of some quantity of interest for a given unit or sample, and may include the sampling procedure, sample preparation, and calibration, as well as the actual measurement process. The true amount of the quantity present can be considered as an input signal, M, which the system converts into a measured value or response, Y. The precision with which M can be estimated based on Y is determined by the characteristics of the relationship between M and Y. Taguchi’s experimentation strategy for dynamic applications is a modification of his approach for static applications. Experimental factors are divided into three groups: control factors, noise factors, and the signal factor. Control factors represent aspects of the process which can easily be adjusted by the operator. These represent the opportunities available to modify the process in order to improve performance. Noise factors represent conditions which are