Vol. 10, No. 1 (2017) pp. 51 57 https://doi.org/10.13160/ricns.2017.10.1.51
1. Introduction
1
Products and their manufacturing processes are influenced both by control factors that can be controlled by designers and by noise factors that are difficult or expensive to control such as environmental conditions.
The basic idea of robust design is to identify, through exploiting interactions between control factors and noise factors, appropriate settings of control factors that make the system's performance robust to changes in the noise factors. Robust design(or Parameter design in a narrow sense) is a quality improvement technique proposed by the Japanese quality expert Taguchi, which was described by Taguchi[1,2], Kackar[3], and others.
In the Taguchi parameter design, the control factors are assigned to an “inner array”, which is an orthogonal array. For each row in the inner array, the noise factors are assigned to an “outer array”, also an orthogonal array.
Because the outer array is run for every row in the inner array, we call this setup a “product array”. A large number of experimental trials in Taguchi's product array may be required because the noise array is repeated for every row in the control array.
Department of Computer Science and Statistics, Chosun University, Gwangju 501-759, Korea
ˋCorresponding author : ymkwon@chosun.ac.kr (Received: January 21, 2017, Revised: March 17, 2017, Accepted: March 25, 2017)
There have been efforts for integrating Taguchi’s important notion of heterogeneous variability the standard experimental design and modeling technology provided by response surface methodology. They combined control and noise factors in a single design matrix, which we call a combined array. The combined array approach was first proposed by Welch, Yu, Kang, and Sacks[4]. The initial motivation of the combined array is the run-size saving. Related approaches were discussed by Vining and Myers[5], Box and Jones[6], Shoemaker, Tsui and Wu[7], and Myers, Khuri and Vining[8], etc.
Treatment of the mean and variance responses via a constrained optimization was discussed in Vining and Myers[5].
In many experimental situations, a number of responses are measured for a given setting of design variables. Khuri and Conlon[9] introduced a procedure for the simultaneous optimization of multiple responses using a distance function.
The combined-array approach allows one to provide separate estimates for the mean response and for the variance response. Accordingly, we can apply the primary goal of the Taguchi method which is to minimize the variance while constraining the mean. In this paper, we propose how to simultaneously optimize multiple responses for robust design when data are collected from a combined array. An example is illustrated to show the proposed method.
Robust Design to the Combined Array with Multiresponse
Yong-Man Kwon
Abstract
The Taguchi parameter design in industry is an approach to reducing performance variation of quality characteristic in products and processes. In the Taguchi parameter design, the product array approach using orthogonal arrays is mainly used. It often requires an excessive number of experiments. An alternative approach, which is called the combined array approach, was studied. In these studies, only single response was considered. In this paper we propose how to simultaneously optimize multiresponse for the robust design using the combined array approach.
Keywords: Taguchi Parameter Design, Product Array Approach, Combined Array Approach, Simultaneously Optimize Multiresponse
2. Simultaneous Optimization of Multiresponse
2.1. Multivariate Linear Model
Suppose the response Ɨ, depends on control variables (or factors) and noise variables. Let a set of control variables be denoted by
ćƖá ÞƖÎìƖÏìzìƖƊ and a set of noise variables by
ćƘá ÞƘÎìƘÏìzìƘƋ . Suppose that all response functions in a multiresponse system depend on the same set of
ćƖ and
ćƘ and that they can be represented by second order models within a certain region of interest. Let § be the number of experimental runs and Ɛ be the number of response functions. The Ƈth second order model is
ƗƇÞ ćƖì
ćƘ ß á ĸƇ×â ć ćĸ
Ƈâ ćƇ
ćƖâ ć «Ƈ
ćƘâ ććĹ
Ƈâ ćƇ
ćƖâ ĻƇì Ƈ á ÎìÏìzìƐ (1)
where ćĸ
Ƈ is Ɗ Z Î, ćĹ
Ƈ is Ƌ Z Î, Ƈ Ƈ is Ɗ Z Ɗ,
«Ƈ «Ƈ is Ƌ Z Ƌ, Ƈ is Ɗ Z Ƌ, which are vectors or matrices of unknown regression parameters, and ĻƇ is a random error associated with the ith response.
Equation (1) can be expressed in matrix notation as ćƗƇá±
ćľƇâ
ćĻƇì Ƈ á ÎìÏìzìƐ (2)
in which
ćƗƇ is an §Z Î vector of observations on the ith response, ± is an §Z Ǝ full column rank matrix of known constants,
ćľ
Ƈ is the Ǝ Z Î column vector of unknown regression parameters, and
ćĻ
Ƈ is a vector of random errors associated with the ith response. We also assume that
Þ ćĻ
Ƈß á ć×ì¯ſƐÞ
ćĻ
Ƈß á ňƇƇ¢§ìƍƔÞ ćĻ
Ƈì ćĻ
ƈß á ňƇƈ¢§ Ƈìƈ á ÎìÏìzìƐì Ƈ d ƈ í
The Ɛ Z Ɛ matrix whose ÞƇìƈßth element is ňƇƈ will be denoted by İ. An unbiased estimator of İ is given by
ýİ á² ¢§à±Þ±±ßà α ²î Þ§à Ǝß,
where ²á Þ ć ƗÎì
ć ƗÏìzì
ć
ƗƐß, and ¢§ is an identity matrix of order §Z§. The Ɛ equations given in (2) may be written in a compact form
ćƗá Ė
Ę ę
ě ćƗ
Î
ćƗ
Ï
{ ćƗ
Ɛ
á Ė
Ę
ę
ě
± × z ×
× ±z × { { | {
× × z± Ė
Ę ę
ě ćľ
Î
ćľ
Ï
{ ćľ
Ɛ
â Ė
Ę ę
ě ćĻ
Î
ćĻ
Ï
{ ćĻ
Ɛ
á ³ ćľ â
ćĻ ì
(3)
where
ćƗ is Ɛ§Z Î, ³ is Ɛ§Z ƐƎ ,
ćľ is ƐƎ Z Î, and ćĻ is Ɛ§Z Î. The variance-covariance matrix of
ćĻ is
¯ſƐÞ
ćĻ ß á İ
"
¢á Ķ,where
"
is a symbol for the direct (or Kronecker) product of matrices.The BLUE (best linear unbiased estimator) of ćľ in (3) is
ý
ćľ á Þ³ à γßà ÎÞ³ à Î
ćƗß á Þ³ ³ßà γ ćƗ. Thus, the BLUE of
ćľ is ý ćľ á Þý
ć ľÎ
ć ľÏ
ć ľƐ where ý
ćľƇá Þ± ±ßà α
ćƗƇ is the least squares estimator of the Ǝ Z Î vector of regression coefficients for the ith response model[10]. The variance-covariance of ý
ćľƇ
¯ſƐÞý
ćľßá Þ³ à γßà Îá Þ± ±ßà Îİ .
The prediction equation for the ith response is given by
ýƗƇÞ ćƖì
ćƘß á ć ćƖì
ćƘßý ćľ
Ƈì Ƈ á ÎìÏìzìƐ (4) where Þ
ć ć is the vector of coded input variables, ć ćƖ ì
ćƘ ß is a vector of the same form as a row of the matrix ± evaluated at the point Þ
ćƖ ì
ćƘ ß. From (4) it follows that
¯ſƐãýƗƇÞ ćƖ ì
ćƘ ßä á ć ćƖ ì
ćƘ ßÞ± ±ßà Î ćƅ Þ
ćƖ ì ćƘ ßňƇƇì Ƈ á ÎìÏìzìƐì
ƍƔãýƗƇÞ ćƖ ì
ćƘ ß ìýƗƈÞ ćƖ ì
ćƘ ßä á ć ćƖ ì
ćƘ ßÞ± ±ßà Î ćƅÞ
ćƖ ì ćƘ ßňƇƈì Ƈ ìƈ á ÎìÏìzìƐ èƇ d ƈ í
Hence,
¯ſƐãý ćƗÞ
ćƖ ì ćƘ ßä á
ć ćƖ ì
ćƘ ßÞ± ±ßà Î ćƅÞ
ćƖ ì ćƘ ßİ ì Ƈ á ÎìÏìzìƐ
where ý
ćƗ Þ ćƖ ì
ćƘ ß á ÞýƗÎÞ ćƖ ì
Ƙ ßìýć ƗÏÞ ćƖ ì
ćƘ ßì z ìýƗƐÞ ćƖ ì
ć is the vector of predicted responses at the point Þ
ćƖ ì ćƘ ß. An unbiased estimator of ¯ſƐ ãý
ćƗ Þ ćƖ ì
ćƘ ßä is given by ý¯ſƐ ãý
ćƗ Þ ćƖ ì
ć ćƖ ì
ćƘ ßÞ± ±ßà ÎƅÞ ćƖ ì
ćƘ ßýİ. 2.2 Estimated Mean and Variance Models in a Multiresponse
Box and Jones[6] modeled the mean and variance separately in a single response. But, we are interested in showing the estimated mean and variance response models in multiple responses.
The fitted ith second-order model in (4) can be rewritten as
ýƗƇÞ ćƖ ì
ćƘ ß á ƀƇ×â ć ćƀƇâ
ć Ƈ ćƖâ
ć «Ƈ ćƘâ
ć ćƐ
Ƈâ ć Ƈ
ćƖ ì Ƈ á ÎìÏìzìƐí
The noise variables
ćƘ are not controllable and they are random variables. In the absence of other knowledge,
ćƘ would be usually uniformly distributed over «Ƙ.
Let ýƋƇÞ
ćƖß ith estimated mean response at an ćƖ averaged over the noise variables
ýƋƇÞ
ćƖß á
õ
«Ƙ
ýƗ
Ƈ
ÞćƖ ì ćƘ ßƎÞ
ćƘ ßƂ
ćƘ ì Ƈ á ÎìÏìzìƐì where ƎÞ
ćƘ ß is a probability density function of ćƘ, and
ćƘ has a uniform distribution over «Ƙ(à Î = Ƙ = Î ). Box and Jones[6] showed that the ith estimated mean becomes
ýƋƇÞ
ćƖ ß á ƀƇ×â ć ćƀ
Ƈâ ć Ƈ
ćƖâ ćР΃Ɛý«Ƈì Ƈ á ÎìÏìzìƐì
(5)
where ƒƐý«Ƈ is the trace of the matrix ý«Ƈ. Let us write ýƔƇÞ
ćƖ ß for the ith mean square variation about the ith mean response
ýƔƇÞ
ćƖ ß á
õ
«Ƙ
ÞýƗƇÞ ćƖ ì
ćƘ ß àýƋƇÞ ćƖ ßßÏƎÞ
ćƘ ßƂ ćƘ ì Ƈ á ÎìÏìzìƐí
Let us call this measure the ith estimated variance, which becomes
ýƔƇÞ ćƖ ß á ćÐ
ÎÞ ćƐ
ƇâýƇ ć ćƐ
ƇâýƇ
ćƖ ß âýƇì Ƈ á ÎìÏìzìƐì
(6)
where ýƇá ãÑİƈ á ÎƋ ÞƐƈƈƇßÏâ Òİƈ á ÎƋ à ÎİƉ á ƈâÎƋ ÞƐƈƉƇ ßÏäî ÑÒ and ƐƈƉƇ is the ƈth row and Ɖth column element of the matrix ý«Ƈ.
From (2.5) the ith estimated mean can be rewritten as ýƋƇÞ
ćƖ ß á ć ćƖ ßý
ćľ×Ƈ Ƈ á ÎìÏìzìƐ (7) where
ć ćƖ ß áÞÎìƖÎì z ìƖƊì ƖÎÏìz ìƖƊÏì ƖÎƖÏì z ì ƖƊ à ÎƖƊì ÎîÐì z ìÎîÐß and ý
ćľ×Ƈá Þƀ×ƇìƀÎƇìzìƀÎÎƇ ìzìƀƊƊƇìƀÎÏƇ ì zìƀƊà ÎƇ ìƐÎÎƇ ìzìƐƋƋƇ is a part of ý
ćľ
Ƈ. From the fact that ý
ćľ
×Ƈ is a part of ý ćľ
Ƈ, the variance-covariance of ý ćľ
× is given by
¯ſƐ
Þ
ýćľ×ß
á Þ±±ß×à Îİ ,where ý ćľ
×á Þý ćľ
×Î ćľ
×Ï ćľ
×Ɛ , Þ± ±ßà Î is Ǝ Z Ǝ, and Þ± ±ß×à Î is Ə Z Ə, where Ǝ á ÞƊ â Ƌ â ƌßÞƊ â Ƌ â Ïßî Ï, and Ə á ÞƊ â ÎßÞƊ â Ïßî Ï â Ƌ. Here Þ± ±ß×à Î is the Ə Z Ə submatrix of Þ± ±ßà Î. From (7) and above, we then have
¯ſƐ ãý ćƋ Þ
ćƖ ßä á
ć ćƖ ßÞ± ±ß×à Î ćƆ Þ
ćƖ ßİ ì where ý
ćƋ Þ
ćƖ ß á
Þ
ýƋÎÞ ćƖ ßìýƋÏÞćƖ ßìzìýƋƐÞ
ćƖ ß
ß
is thevector of estimated mean responces at the point ćƖ. An unbiased estimator of ¯ſƐãý
ćƋ Þ
ćƖ ßä is given by ý¯ſƐ ãý
ćƋ Þ ćƖ ßä á
ć ćƖ ßÞ± ±ß×à Î ćƆ Þ
ćƖ ßýİ. (8) 2.3. The Proposed ¦ measure about the estimated mean responses
Let us find conditions on a set of control variables ćƖ which optimize a set of estimated mean responses ý
ćƋ Þ
ćƖ ß subject to maintaining estimated variance responses ý
ćƔ Þ
ćƖ ß within some specified upper bounds. If
all the estimated mean ý ćƋ Þ
ćƖ ß attain their individual optimum values
ćʼn at the same set
ćƖ of operating conditions, then the problem of simultaneous optimization is obviously solved. This ideal optimum rarely occurs. In more general situations we might consider finding compromising conditions on the control variables that are somewhat favorable to all mean responses. Such deviation of the compromising conditions from the ideal optimum condition can be evalulated by means of a distance function which measures the distance of ý
ćƋ Þ
ćƖ ß, from ćʼn. Let
ćʼn be the optimum (or target) value of ýƋƇÞ ćƖ ß over «Ɩ and let
ćʼn á ÞʼnÎìʼnÏìzìʼnƐ . We shall consider a constrained-optimization procedure for each response according to the Taguchi's three basic situations as follows.
1. “nominal-is-best characteristics”: target value of ýƋƇÞ
ćƖ ß á ʼnƇ ,
2. “larger-the-better characteristics”:
¦ſƖ
ćƖ«Ɩ
ýƋƇÞ
ćƖ ß á ʼnƇ ,
3. “smaller-the-better characteristics”:
¦Ƈƌ
ćƖ«Ɩ
ýƋƇÞ
ćƖ ß á ʼnƇ . A distance function of ý ćƋ Þ
ćƖ ß for the target value ćʼn may be expressed as
ã
ýćƋ Þ ćƖ ßì
ćʼn
ä
áã
ÞýćƋÞ ćƖ ß à
ć ¯ſƐãý
ćƋ Þ
ćƖ ßä뫮 à ÎÞý ćƋÞ
ćƖ ß à ćʼn ß
ä
ÎîÏUsing the estimate given in (8) for the variance-covariance matrix of ý
ćƋ Þ
ćƖ ß, we get a distance function
ƙ
Ɯ
ƚ
ćć ćƖ ßÞ± ±ß×à Î ćƆÞ
ćƖ ß ÞýćƋ Þ
ćƖ ß à
ć İà ÎÞý ćƋ Þ
ćƖ ß à ćʼn ß ƛ
Ɲ
ƞ
ÎîÏ (9) If the mean response ýćƋ Þ
ćƖ ß takes on different degrees of importance, we can imply weights ƕÎìƕÏìzìƕƐ where × ï ƕƇï Î for each i and İƇ á ÎƐ ƕƇá Î . Then the distance function can be written as
ƙ Ɯ
ƚ
ć ćƖ ßÞ± ±ß×à Î ćƆÞ
ćƖ ß 뫭 °Þý
ćƋ Þ ćƖ ß à
ć İà Î뫭 °Þý ćƋ Þ
ćƖ ß à ćʼn ß뫮 ƛ
Ɲ
ƞ
ÎîÏ(2.9) where °á
Ė
Ę
ę
ě ƕÎ × z ×
ƕÏz ×
| {
ƑƗƋ ƕƐ
. From the constrained-
optimization procedure and the distance From the distance measure of ý
ćƋ Þ
ćƖ ß for the target value
ćʼn, we propose a simultaneous optimization of ý
ćƋ Þ
ćƖ ß over the region of interest «Ɩ. From (9), the proposed simultaneous-optimization measure about the estimated mean responces can be written as
¦ = ćƖ«Ɩ
¦Ƈƌ
¦Þ ćƖ ß á
ćƖ«Ɩ
¦Ƈƌ
ć ćƖ ßÞ± ±ß×à Î ćƆÞ
ćƖ ß 뫭 °Þý
ćƋ Þ ćƖ ß à
ć İà Î뫭 °Þý ćƋ Þ
ćƖ ß à ćʼn ß뫮 (10)
The ¦ measure can be used without a prior knowledge about the estimated mean responses. It takes into consideration the variances and correlations of the estimated mean responses.
2.4. The Proposed ¯ Measure About the Estimated Variance Responses
In this section, we propose the simultaneous- optimization measure of multiple responses(quality characteristics) for robust design in a combined array.
If we have a prior knowledge about the estimated mean response ý
ćƋ Þ
ćƖ ß, it is possible to minimize the estimated variance response. Let
ýƔƇÝÞ
ćƖß á ć
ćƖ«Ɩ
¦ſƖ ýƔ
ƇÞ ćƖß à
ćƖ«Ɩ
¦Ƈƌ ýƔ
ƇÞ ćƖß ýƔƇÞ
ćƖß à ćƖ«Ɩ
¦Ƈƌ ýƔ
ƇÞ ćƖß
ì
Ƈ á ÎìÏìzìƐì
where ýƔƇÞ
ćƖ ß is the Ƈth mean square variation about the Ƈth mean response which is in (6). Note that ýƔƇÝÞ
ćƖß is a
“standardized” measure of ýƔƇÞ
ćƖß. The proposed simultaneous optimization measure about the estimated variance responces can be written as
¯ = ćƖ«Ɩ
¦Ƈƌ
¯Þ ćƖß á
ćƖ«Ɩ
¦Ƈƌ ć ćƔÝÞ
ćƖß
á ćƖ«Ɩ
¦Ƈƌ
ā
Ƈ á Î ƐƕƇýƔƇÝÞ
ćƖßì Ƈ á ÎìÏìzìƐì (11) where
ćƕá ÞƕÎìƕÏìzìƕƐ , ý
ćƔÝÞ
ćƖß á ÞýƔÎÝÞ ćƖßì ýƔÏÝÞ
ćƖßìzì ýƔƐÝÞ
ć , İƇ á ÎƐ ƕƇá Î . If the variance response ý ćƔ Þ
ćƖ ß takes on different degrees of importance, we can imply weights ƕÎìƕÏìzìƕƐ where × ï ƕƇï Î for each i and İƇ á ÎƐ ƕƇá Î .
2.5. The Proposed Simultaneous-Optimization ©Ƌ Measure
Let us find conditions on a set of control variables ćƖ which optimize simultaneously for a set of estimated mean responses and estimated variance responses. If all the estimated mean and all the estimated variance attain their individual optimum values at the same set
ćƖ of operating conditions, then the problem of simultaneous optimization is obviously solved. This ideal optimum rarely occurs. In more general situations we might consider finding compromising conditions on the control variables that are somewhat favorable to all mean responses and all the estimated variance. Such deviation of the compromising conditions from the ideal optimum condition can be formulated by means of the desirability function.
We propose a simultaneous optimization for a set of estimated mean responses and estimated variance responses over the region of interest «Ɩ using proposed
¦Þ
ćƖß and ¯Þ
ćƖß. From Equations (10) and (11), the proposed simultaneous-optimization measure can be written as
©Ƌ á ćƖ«Ɩ
¦Ƈƌ
©ƋÞ ćƖß á
ćƖ«Ɩ
¦Ƈƌ ãŁ¦Þ
ćƖß â ÞÎ à Łß¯Þ ćƖß ä
(12)
where «Ɩ is the region of interest on a set of control variables
ćƖ and × = Ł = Î . This is a criterion in which
¦Þ
ćƖß and ¯Þ
ćƖß take on different degrees of importance.
As a way of finding the optimal solution of the control factors according to the proposed formula, genetic algorithm was used in the MATLAB Optimization Toolbox.
3. Numerical Example
In this section we give a numerical example, consisting of a multiresponse system of two response variables, ƗÎ and ƗÏ, and two control variables, ƖÎ and ƖÏ, and a noise variable Ƙ. The design is somewhat similar to the standard central composite design. The cube portion of the experimental arrangement is chosen to be a ÏÐ design and star points are added only for the two control variables. The following Table 1 gives the factor levels and a set of data.
Each of the two responses was fitted to a second order regression model. The estimated response models by the method of least squares are given by
ýƗÎÞ
ćƖìƘ ß á ÔÓí×× à ÎÏíÐÔƖÎà ÕíÖÓƖÏà ÔíÏÏƖÎÏà ÕíÑÒƖÏÏ à ÕíÎÎƖÎƖÏâ ÒíÐÕƘÏà ÎíÑÑƘ â ÏíÖÓƖÎƘ à ÎíÕÓƖÏƘ ì
(13)
ýƗÏÞ
ćƖìƘß á Î×Ðí×× à ÎÏíÏÎƖÎâ ÓíÓÕƖÏà ÎÐíÖÓƖÎÏà ÕíÒ×ƖÏÏ à ÏíÖÐƖÎƖÏâ ÓíÏÐƘÏà ÎíÐÕƘ â ÎíÔÒƖÎƘ à ÏíÖÒƖÏƘ í
(14) From (13) and (14), using the mean and variance response equation (5) and (6), the estimated mean and variance response models are given by
ýƋÎÞ
ćƖ ß á ÔÔíÔÖ à ÎÏíÐÔƖÎà ÕíÖÓƖÏà ÔíÏÏƖÎÏ à ÕíÑÒƖÏÏà ÕíÎÎƖÎƖÏì
ýƋÏÞ
ćƖ ß á Î×Òí×Õ à ÎÏíÏÎƖÎâ ÓíÓÕƖÏà ÎÐíÖÓƖÎÏ à ÕíÒ×ƖÏÏà ÏíÖÐƖÎƖÏì
ýƔÎÞ
ćƖ ß á Þà ÎíÑÑ â ÏíÖÓƖÎà ÎíÕÓƖÏßÏîÐ â ÏíÒÔ ì ýƔÏÞ
ćƖ ß á Þà ÎíÐÕ à ÎíÔÒƖÎà ÏíÖÒƖÏßÏîÐ â ÐíÑÒ í The region of interest «Ɩ is given by the inequality
à Î = ƖÎìƖÏ= Î . The ranges for ýƋÎÞ ćƖß, ýƋÏÞ
ćƖß, ýƔÎÞ
ćƖß and ýƔÏÞ
ćƖ ß over «Ɩ are, respectively, ÐÏíÓÕ = ýƋÎÞ
ćƖ ß = ÕÐíÏÒ, ÓÓíÓÓ = ýƋÏÞ
ćƖ ß = Î×ÖíÓÒ, ÏíÒÔ = ýƔÎÞ
ćƖ ß = ÎÒíÓÐ and ÐíÑÒ = ýƔÏÞ
ćƖ ß = ÎÒíÔÔ. Suppose that the quality characteristics for ƗÎ and ƗÏ are the nominal-is-best characteristics and the larger-the-better characteristics. Let us assume that the target value ofýƋÎÞ
ćƖ ß is taken to be 75.00 and the target value of ýƋÏÞ
ćƖ ß is taken to be ¦ſƖ
ćƖ «ƖýƋÏÞ
ćƖ ß á Î×ÖíÓÒ. From (12), we obtained the results of simultaneous optimization based on the minimization of the ©ƋÞ
ćƖß measure over «Ɩ. Table 2 indicates that the optimal setting for the Ʌ=0.1, ƕÎ=ƕÏ=1.00 is ƖÎáà ×íÎ× and ƖÏá ×íÎÕ , which produces a predicted value of 77.21, 107.14, 4.00, and 4.45 for ýƋÎÞ
ćƖ ß, ýƋÏÞ ćƖ ß, ýƔÎÞ
ćƖ ß, and ýƔÏÞ
ćƖ ß, respectively.
If a simultaneous optimum value is much different
from its corresponding individual optimum value, we may reoptimize ©Ƌ. Also we may analyze ©Ƌ sequentially as the acceptable values for Ł and weights ƕÎìƕÏìzìƕƐ are varied.
4. Conclusion
The combined-array approach allows one to provide separate estimates for the mean response and for the variance response. Accordingly, we can apply the primary goal of the Taguchi methodology which is to obtain a target condition on the mean while achieving the variance, or to minimize the variance. In this study we proposed the simultaneous-optimization measure ©Ƌ of multiple responses. The proposed concept of ©Ƌ
measure is minimizing the deviation of the mean responses from the target values and also ©Ƌ measure is minimizing the deviation of the variance responses from
Run ƖÎ ƖÏ Ƙ ƗÎ ƗÏ
1 2 3 4 5 6 7 8 9 10 11 12 13 14
-1 -1 -1 -1 1 1 1 1 -1.41
1.41 0 0 0 0
-1 -1 1 1 -1 -1 1 1 0 0 -1.41
1.41 0 0
-1 1 -1 1 -1 1 -1 1 0 0 0 0 0 0
80.6 74.9 83.1 71.2 66.8 74.2 38.1 36.8 80.9 42.4 73.4 45.0 77.4 74.6
81.4 95.9 105.0 103.0 74.0 76.8 81.2 76.9 100.0 50.5 71.2 101.0 102.0 104.0 Table 1. Experimental design and response values
Weight Location of optima Simultaneous optima
Ʌ ƖÎ ƖÏ ýƋÎÞ
ćƖ ß ýƋÏÞ
ćƖ ß ýƔÎÞ
ćƖ ß ýƔÏÞ ćƖ ß 0.10
0.30 0.50 0.70 0.90
-0.10 -0.07 -0.05 -0.03 -0.01
0.18 0.21 0.27 0.29 0.26
77.21 76.49 75.46 74.92 75.03
107.14 106.94 106.88 106.68 106.37
4.00 3.95 4.03 4.00 3.84
4.45 4.62 4.90 5.04 4.96 Table 2. Simultaneous optimization for ©Ƌ
the target values of the variance responses that is minimum values of the variance responses. The ©Ƌ
measure is easy to apply, and permits the user to make subjective judgements on the importance of each response.
In this study we assume that the noise variables would be uniformly distributed over the region of interest of noise variables. It will be of interest to consider the case when the noise variables are not uniformly distributed over the region of interest of noise variables.
Acknowledgments
This study was supported(in part) by research funds from Chosun University, 2015.
References
[1] G. Taguchi, “Introduction to quality engineering:
designing quality into products and process”, UNIPUB/Kraus International, White Plains, NY., 1986.
[2] G. Taguchi, “System of experimental designs:
engineering methods to optimize quality and minimize cost”, UNIPUB/Kraus International, White Plains, NY., 1987.
[3] R. N. Kackar, “Off-line quality control, parameter design, and the Taguchi method”, J. Qual.
Technol., Vol. 17, pp. 176-209, 1985.
[4] W. J. Welch, T .K. Yu, S. M. Kang, and J. Sacks,
“Computer experiments for quality control by parameter design”, J. Qual. Technol., Vol. 22, pp.
15-22, 1990.
[5] G. G. Vining and R. H. Myers, “Combining Taguchi and response surface philosophies: A dual response approach”, J. Qual. Technol., Vol. 22, pp.
38-45, 1990.
[6] G. E. P. Box and S. P. Jones, “Designing products that are robust to the environment”, Total Quality Management, Vol. 3, pp. 265-282, 1992.
[7] A. C. Shoemaker, K. L. Tsui, and C. F. J. Wu,
“Economical experimentation methods for robust design”, Technometrics, Vol. 33, pp. 415-427, 1991.
[8] R. H. Myers, A. I. Khuri, and G. G. Vining,
“Response surface alternative to the Taguchi robust parameter design approach”, The American Statistician, Vol. 46, pp. 131-139, 1992.
[9] A. I. Khuri and M. Conlon, “Simultaneous optimization of multiple responses represented by polynomial functions”, Technometrics, Vol. 23, pp.
363-375, 1981.
[10] D. S. Huang, “Regression and econometrics methods”, John Wiley & Sons, New York, p. 188, 1970.