You are on page 1of 14

Multi-response

The optimization of multi- problems in the


response problems in the Taguchi method

Taguchi method
367
Lee-Ing Tong and Chao-Ton Su
National Chiao Tung University, Hsinchu, Taiwan, Republic of China, Received May 1995
and Revised February 1996

Chung-Ho Wang
Chin Min College, Hsinchu, Taiwan. Republic of China

Introduction
Off-line quality control is considered to be an effective approach in producing
high-quality products quickly at a relatively low cost. Off-line quality control
includes those quality control activities at the product planning, design and
production engineering stages (but not during actual production)[1]. The
Taguchi method, which combines the experimental design techniques with
quality loss considerations, is the conventional approach used in off-line quality
control. Taguchi’s off-line quality control method consists of three stages:
systems design, parameter design and tolerance design. A more detailed
description of these three design types is provided in Kackar[2] and Phadke[3].
The parameter design of the Taguchi method involves determining the
design parameter’s settings for a product or a process so that the product’s
response has the minimum variation and its mean is close to the target.
Experimental design is used in this method to arrange the design parameters
and noise factors in the orthogonal arrays. The signal-to-noise (SN) ratio is
computed for each experimental combination. Next, SN ratios are analysed to
determine the optimal settings (i.e. control factors and their levels) of the design
parameters. However, the Taguchi method can only be used in a single-response
case; it cannot be used to optimize a multi-response problem. Up to now,
engineering judgement is used primarily for solving the optimization of the
complicated multi-response problem in the Taguchi method. A more effective
approach is required because an engineer’s judgement will increase the
uncertainty during the decision-making process.
In this work, the Taguchi method is applied towards a multi-response
production process. Considering the quality loss of each response, a multi-
response signal-to-noise ratio is developed to determine the optimum conditions
in the parameter design stage for the multi-response problems. The proposed

The authors would like to thank the National Science Council of the R.O.C. for financial support
of this manuscript under Contract No. NCS-83-0415-E009-01. The Industrial Technology Research International Journal of Quality
& Reliability Management,
Institute (Taiwan) is also appreciated for assistance in accumulating the experimental data in this Vol. 14 No. 4, 1997, p. 367-380,
work. © MCB University Press, 0265-671X
IJQRM optimization procedure includes four phases which are capable of decreasing
14,4 the uncertainty in engineering judgement when the Taguchi method is applied.
Only the static quality characteristic problem, in which the desired response
value is fixed, is discussed here.
This paper is organized as follows. A literature review of the multi-response
problems in the Taguchi method is made. The next section proposes an
368 optimization procedure for solving the problem of multi-response cases in the
Taguchi method. Two illustrative examples for the implementation of the
proposed procedure are then provided, followed by concluding remarks in the
final section.

Literature review
Multi-response problems in the Taguchi method have received limited attention.
Phadke[3] used the Taguchi method to study the surface defects and thickness
of a wafer in the polysilicon deposition process for a VLSI circuit manufacture.
Based on the judgement of relevant experience and engineering knowledge,
some trade-offs were made in his investigation to choose the optimum factor
levels for this multiple quality characteristics problem. By human judgement,
validity of the experimental results cannot be easily assured. Contradictory
results could be reached by different engineers. Therefore, the uncertainty in
the optimum factor levels is increased. Phadke’s approach can only be used by
an experienced engineer. Logothetis and Haigh[4] applied the multiple
regression technique and the linear programming approach to optimize a five-
response process by the Taguchi method. However, if the t-values of the
regression coefficients are insignificant or the value of R2 (the coefficient of
determination) is low, their method’s application could be limited. In addition,
their method increases the complexity of the computational process, thereby
making it too difficult to be used on the shopfloor. Hung[5] transformed various
types of quality characteristics (smaller-the-better, larger-the-better and
nominal-the-best) into the nominal-the-best characteristics with a target of 0
and gave a weight to each quality characteristic for computing the SN ratio.
However, his method cannot handle the problem involving continuous and
discrete data. Moreover, as a result of the implementation, when the weight of a
particular quality characteristic is increased, the optimum conditions will not
move in the same direction of that quality characteristic. This result is
unsatisfactory. Shiau[6] assigned a weight to each SN ratio of the quality
characteristic and summed the weighted SN ratios for computing the
performance measurement of a multi-response problem. For instance, there are
two quality characteristics with SN ratios: SN1 = –10logL1 and SN2 = –10logL2,
where L1 and L2 represent the quality loss of these two characteristics. As a
result, the weighted SN ratio for this two-response problem will be SN 0 =
w1(SN1) + w2(SN2), where wi is the weight of the ith response. If SN0 = –10logL,
w w
where L can be viewed as the total quality loss, we then have L = L 11·L 22. This
equation is difficult to explain from the perspective of Taguchi’s quality loss.
Tai et al.[7] claimed that quadratic modelling was invalid for non-symmetric
loss functions. In their investigation, empirical loss functions were developed Multi-response
for a multi-response problem involving six variables and nine responses in the problems in the
surface mount technology process. Multiple responses can be converted into a Taguchi method
single response on the basis of the quality loss of each response. However, these
empirical loss functions can only be used in a specific process. If their method is
applied, the empirical loss functions should be found in advance. Consequently,
the complexity of the problem will be increased. Accordingly, a relatively simple 369
procedure is proposed in this work to solve the optimization of multiple
responses problems in the Taguchi method.

Proposed optimization procedure


Applying the Taguchi method to optimize multi-response processes includes
the following considerations:
• The attribute and loss functions in the multiple cases are always
different for each response. Therefore, the loss for each response cannot
be compared and summed directly.
• The measurement units in the multiple cases are always different for
each response. Therefore, the loss caused by each unit for each response
could be different.
• The importance in the multiple cases could be different for each
response.
• The adjustment factors should be chosen when the nominal-the-best
quality characteristics exist in multi-response cases. This is especially
true when one such factor is used to adjust the mean on target and a
significant change occurs in the values of other quality characteristics.
To solve the four problems mentioned above, an optimization procedure is
proposed in this section. An effective method capable of determining the multi-
response signal to noise (MRSN) ratio is developed here through the integration
of quality loss for all responses with the application of Taguchi’s[1] SN ratios.
Next, the traditional Taguchi method can be applied based on MRSN. The
proposed optimization procedure includes four phases:

Phase 1. Compute the quality loss


In this phase, the quality loss for each response is computed. According to
Taguchi[1], the following three formulas are given:
ni
1
Lij = k1 ∑ yijk
2
for the smaller - the - better response (1)
ni k =1

ni
1 1
Lij = k2
1
∑ 2
for the larger - the - better response (2 )
ni k = 1 yijk
IJQRM 2
14,4 s 
ij
Lij = k3   for the nominal - the - best response (3 )
y 
 ij 

where
370 Lij = quality loss for ith response at jth trial
yijk = observed data for ith respsonse at jth trial, kth repetition,
ni = replications for ith response,
ni
1
yij = ∑ yijk
ni k =1
ni
1
Si j = ∑ ( yijk – yij )2
ni – 1 k = 1
k1 , k1 , k3 = quality loss coefficients.
Notably, for the nominal-the-best quality characteristic, Taguchi[1] does not
define SN as:
1
−10 log10 ( MSD ) = −10 log10 [ Σ( y − y )2 + ( y − T )2 ], where T is the
n
target value, (4)
but
y2
SN = −10log10 (5 )
s2 .
The major reason is that choosing the optimal factor levels according to
maximizing SN = –10log10(MSD) does not necessarily minimize both –n1 ∑(y – –y)2
– –T) 2. Equation (5) shows that maximizing SN must correspond to
and (y
minimizing s2/–y2, which is the desired property to have for any process[8]. As a
result, the computation of the quality loss for the nominal-the-best response is
based on the form of L = k(s2/–y2) for this study.

Phase II. Determine the multi-response signal to noise (MRSN) ratio


Reducing the variability first requires normalizing the scale of the quality loss
of each response. For each response, the quality loss at each trial is divided here
by the maximum quality loss in the j trials. Accordingly, the largest normalized
value is 1. The smaller the normalized value implies the smaller the quality loss.
Thus, the normalized quality loss ranges between 0 and 1. Therefore, the
quality loss for each response can be summed directly. Second, an appropriate
weight is assigned to each response for computing the total normalized quality
loss (TNQL) in each trial. Finally, the MRSN ratio can be computed on the basis Multi-response
of TNQL. These three steps are summarized as follows: problems in the
Step 1. Normalize the quality loss of each trial for each response. Taguchi method
Lij
C ij = (6 )
L*i
371
where L*i = max{ Li 1 , Li 2 , … , Lij }.

Step 2. Compute the total normalized quality loss (TNQL) of each trial:
m
TNQL j = ∑ wi C ij (7 )
i =1

where wi = the weight of ith normalized response (i = 1, 2, … , m).


Step 3. Determine the MRSN ratio for each trial:
MRSNj = –10log(TNQLj) (8)

Phase III. Determine the optimal factor/level combination


For smaller-the-better and larger-the-better cases, Taguchi[1] suggested the
direct minimization of the expected quality loss. For the nominal-the-best case,
Taguchi suggested the two-stage optimization procedure, i.e. maximize the
signal-to-noise ratio and then adjust the mean on target. Based on these
concepts, the procedure used to determine the optimal factor/level combination
in a multi-response problem is described in the following:
Step 1. Estimate the factor effects.
(1) Plot factor effects and tabulate main effects on MRSN.
(2) Plot factor effects and tabulate main effects on the mean response for
nominal-the-best case.
Step 2. Determine the optimal control factors and their levels.
(1) Find the control factor which is a factor with the significant effect on
MRSN.
(2) Determine the optimum level for each control factor as the level that has
the highest value of MRSN.
Step 3. Determine the optimal adjustment factors: If the nominal-the-best
characteristic exists in multi-response problems, the suitable adjustment
factors should be identified. There are four cases:
(1) when there are smaller-the-better and nominal-the-best characteristics to
be optimized;
(2) when there are larger-the-better and nominal-the-best characteristics to
be optimized;
IJQRM (3) when there are smaller-the-better, larger-the-better and nominal-the-best
14,4 characteristics to be optimized;
(4) when there are all nominal-the-best characteristics to be optimized.
A factor which satisfies the following two requirements can be chosen as an
adjustment factor for cases 1, 2 and 3. First, any factor that has no significant
372 effect on MRSN but, instead, a significant effect on the mean response for
nominal-the-best characteristics can be selected as a candidate for the
adjustment factor. Second, when the adjustment factor is used to bring the
mean on target, the direction in which the quality characteristics are improved
should satisfy the objective of smaller-the-better and larger-the-better cases
simultaneously. Any factor that has no significant effect on MRSN, a
significant effect on the mean response for its quality characteristic, and no
significant effect on the mean response for the other quality characteristics
can be selected as an adjustment factor for case 4.
Guidelines for determining the optimal adjustment factors in a multi-
response problem are presented in the above. From those guidelines, we can
infer that the task of determining the optimal adjustment factors becomes
more complicated when there are multiple characteristics to be optimized.
Occasionally, the necessary trade-offs should be made to choose suitable
adjustment factors. However, in selecting the adjustment factors, Phadke[3]
emphasized that finding an adjustment factor that can be changed
conveniently is of higher priority than finding the level of the adjustment
factor that brings the mean to be precisely on target.

Phase IV. Conduct the confirmation experiment


The major limitation for the optimization procedure proposed here is that the
MRSN value for the confirmation experiment cannot be computed by using
equation (8). However, a comparison of the predicted with observed MRSN is
not so important. The confirmation experiment is performed to verify that the
optimum condition derived by the experiment actually yields an
improvement. If the predicted and observed SN ratios for each response are
close to each other, we can conclude that the additive model on which the
experiment was based is a good approximation. As a result, the recommended
optimum condition can be adopted for the process. If the predicted and
observed SN ratios for one of the responses do not match, we suspect that the
additive model is inadequate and the interactions are probably important. In
this case, another experiment may be necessary to achieve the required
objective.
The principal strength of the proposed procedure is its universality; it can
be used in any type of multi-response problem and can simultaneously deal
with the continuous and discrete data types. The proposed optimization
procedure for multi-response problems in the Taguchi method is illustrated in
Figure 1.
Multi-response
problems in the
Taguchi method

373

Figure 1.
Proposed optimization
procedure for multi-
response problems in
the Taguchi method

Implementation
The effectiveness of the proposed optimization procedure is shown in this
section by using two case studies:

Case study 1
This case study involves the improvement of a plasma-enhanced chemical
vapour deposition (PECVD) process in IC manufacturing. The case study was
IJQRM conducted by the Industrial Technology Research Institute, Taiwan. The three-
14,4 inch wafers were mounted on holders. Each holder can carry five wafers. The
major function of the deposition process is to deposit a uniform layer of silicon
nitride (SiNx) with a specified thickness. In the past, the uniformity of the output
was unstable. The reason for this low uniformity was unknown. Therefore,
engineers were unaware of how to adjust the settings of the process parameters
374 when the quality of the wafer did not match the required specifications. In this
study, an experiment was performed to determine the effects of process
parameters on the silicon nitride deposition process. Optimal settings can,
hopefully, be found such that a high uniformity (i.e. low variability) for the
response can be achieved.
The two responses (in order of importance) are:
(1) RI: refractive index, in which the target value is 2;
(2) DT: deposition thickness, in which the target value is 1,000Å.
The priority of the RI response is higher than that of the DT-response.
Following discussion with the engineers, the weight ratio was set to 2:1. In the
experiment, eight controllable factors were selected for optimization. These
factors and their alternative levels are listed in Table I. The standard array L18
was selected for the experiment. The data for 18 experiments are summarized
in Table II.
For experiment number 1, the MRSN ratio was computed by using equations
(6), (7) and (8):
L11 0.00156
Step 1 : For the RI- response C11 = = = 0.19024.
L*1 0.00820
L 0.00732
For the DT- response C21 = 21 = = 0.15234.
*
L2 0.04805
Step 2 : TNQL1 = w1C11 + w2C21 = 2 × C11 + 1 × C21 = 0.5328.
Step 3 : MRSN 1 = –10 log( TNQL1 ) = –10 log( 0.5328 ) = 2.734.
Following the same procedure described in the previous section, the observed
data were computed and those results are displayed in Table II. The main effects
on MRSN, the average of RI-response and the average of DT-response are
summarized in Tables III, IV and V respectively. Their corresponding factor
effects are also plotted in Figures 2, 3 and 4 respectively.
The controllable factors on MRSN (in order of significance) are: F, E, H, C, G,
D, A and B. The larger the MRSN ratio would imply the better the quality.
Consequently, the tentative optimal condition can be set as A1B3C2D2E2F2G2H3.
In this experiment, factors C, E, F and H are chosen as the optimal control
factors. Factor D has only a slight effect on MRSN and the average of DT-
response, but a more significant effect on the average of RI-response. Factor G
has a slight effect on MRSN and the average of RI-response, but a more
Factor Level 1 Level 2 Level 3
Multi-response
problems in the
A. Cleaning method No Yesa – Taguchi method
B. The chamber temperature 100°Ca 200°C 300°C
C. Number of runs after the
chamber has been cleaned 1st 2nda 3rd
D. The flow rate of SiH4 6% 7%a 8%
375
E. The flow rate of N2 30% 35%a 40%
F. The chamber pressure 160 mtorr 190 mtorra 220 mtorr
G. R.F. power 30 watt 35 watta 40 watt
H. Deposition time 11.5 min 12.5 min 13.5 min
Table I.
Note: Control factors and
a Starting levels their levels

Experiment Deposition thickness Refractive index


no. Factors (DT) (RI) Average
A B C D E FG H 1 2 3 4 5 1 2 3 4 5 DT RI MRSN

1 1 1 1 1 1 11 1 694 839 728 688 704 2.118 1.919 1.985 2.085 2.056 730.6 2.033 2.734
2 1 1 2 2 2 22 2 918 867 861 874 851 2.205 2.240 2.234 2.165 2.275 874.2 2.224 9.910
3 1 1 3 3 3 33 3 936 954 930 1,058 958 2.677 2.643 2.714 2.456 2.565 967.2 2.611 3.593
4 1 2 1 1 2 23 3 765 828 842 768 801 2.096 1.997 1.949 2.046 2.000 800.8 2.018 6.487
5 1 2 2 2 3 31 1 709 743 753 752 989 2.032 2.007 1.943 2.003 1.845 789.2 1.966 1.058
6 1 2 3 3 1 12 2 795 785 846 722 833 1.860 1.838 1.842 1.999 1.858 796.2 1.879 4.059
7 1 3 1 2 1 32 3 711 816 1,085 787 1,150 2.012 1.909 1.797 1.930 1.819 909.8 1.893 1.681
8 1 3 2 3 2 13 1 580 644 602 607 811 1.834 1.760 1.760 1.782 1.744 648.8 1.776 2.773
9 1 3 3 1 3 21 2 590 812 627 595 609 1.719 1.707 1.676 1.704 1.675 646.6 1.696 3.290
10 2 1 1 3 3 22 1 917 1,142 1,126 916 966 2.097 1.911 1.889 2.014 1.960 1,013.4 1.974 1.583
11 2 1 2 1 1 33 2 1,389 1,405 1,219 2,063 1,392 1.927 1.860 1.945 1.539 1.867 1,293.6 1.828 –4.771
12 2 1 3 2 2 11 3 865 914 993 838 893 1.963 1.881 1.812 1.923 1.899 900.6 1.896 5.203
13 2 2 1 2 3 13 2 827 884 884 851 1,066 1.903 1.829 1.788 1.863 1.767 902.4 1.830 3.470
14 2 2 2 3 1 21 3 787 805 780 776 976 2.103 2.020 2.011 2.107 1.968 824.8 2.042 3.568
15 2 2 3 1 2 32 1 739 779 745 724 976 2.182 2.080 2.071 2.179 1.968 792.6 2.096 0.971
16 2 3 1 3 2 31 2 724 721 690 1023 915 2.274 2.166 2.215 2.103 2.203 814.6 2.192 0.570 Table II.
17 2 3 2 1 3 12 3 771 806 785 869 859 1.942 1.905 1.909 1.916 1.900 818.0 1.914 11.069 Data summary by
18 2 3 3 2 1 23 1 712 781 749 692 760 2.077 1.961 1.985 2.101 1.980 738.8 2.021 5.367 experiment

significant effect on the average of DT-response. Factors D and G can be chosen


as adjustment factors for RI-response and DT-response respectively.
The optimal condition is verified by a confirmation experiment. The results
under the optimum condition A 1B 3C 2D 2E 2F 2G 2H 3 and under the starting
condition A2B1C2D2E2F2G2H2 are tabulated in Table VI. According to this table,
IJQRM Maximum-
14,4 Factor Level 1 Level 2 Level 3 minimum

A 3.581 3.003 – 0.578


B 3.043 3.269 3.565 0.522
C 2.194 3.934 3.747 1.740
376
D 3.297 3.888 2.691 1.197
E 1.546 4.319 4.010 2.733
F 4.885 5.034 –0.043 5.077
Table III. G 2.738 4.319 2.820 1.581
Main effects on MRSN H 2.415 2.754 4.706 2.291

Maximum-
Factor Level 1 Level 2 Level 3 minimum

A 2.011 1.977 – 0.034


B 2.094 1.972 1.916 0.179
C 1.990 1.958 2.033 0.075
D 1.931 1.972 2.079 0.148
E 1.949 2.034 1.999 0.084
F 1.888 1.996 2.098 0.210
Table IV.
Main effects on the G 1.971 1.997 2.014 0.043
average of RI-response H 1.978 1.942 2.062 0.121

Maximum-
Factor Level 1 Level 2 Level 3 minimum

A 795.993 922.089 – 125.156


B 996.600 817.667 762.767 233.833
C 861.933 908.100 807.000 101.100
D 880.367 852.500 844.167 36.200
E 915.633 805.267 856.133 110.367
Table V. F 799.433 816.433 961.167 161.733
Main effects on the G 784.400 867.367 925.267 140.867
average of DT-response H 785.567 921.2647 870.200 135.700

an improvement in refractive index is 5.84dB (the variance dropped to 24 per


cent) and the deposition thickness can be made by 3.77dB (the variance dropped
to 36 per cent). Furthermore, adjustment factors D and G can be used to adjust
the averages of these two responses to their targets.
Case study 2 Multi-response
Phadke[3] considered a case study to improve a polysilicon deposition process. problems in the
This study was conducted by Peter Hey in 1984. Six controllable factors were Taguchi method
identified. These were deposition temperature(A), deposition pressure(B),
Nitrogen flow(C), Silane flow(D), settling time(E) and cleaning method(F). All
factors were studied at three levels each. The L18 orthogonal array was used
and factors (A)-(F) were assigned to columns 2, 3, 4, 5, 6 and 8 respectively. The 377
quality characteristics of interest were the surface defects, the thickness and the
deposition rate. The target value in the study for the thickness of the polysilicon

Figure 2.
Factor effects on MRSN
(case study 1)

Figure 3.
Factor effects on the
average of RI-response

Figure 4.
Factor effects on the
average of DT-response
IJQRM layer was 3,600Å. Nine observations were taken for each trial run. The starting
14,4 condition was set as A2B2C1D3E1F1. The optimum condition chosen from the
experimental data by Phadke was A1B2C1D3E2F2.
The above case is analysed again by the proposed optimization procedure.
Two different weight ratios for these three responses (the surface defects, the
thickness and the deposition rate) are set as 1:1:1 and 2:1:1. The MRSN ratios
378 are computed on the basis of the observed data. The factor effects on MRSN are
plotted in Figure 5. Therefore, the optimal conditions can be set as
A2B1C3D2E3F2 and A1B1C3D2E3F2 for the cases of weight ratios 1:1:1 and 2:1:1
respectively.
To predict the anticipated improvements under the chosen optimum
conditions, the SN ratios for surface defects, thickness and deposition rate are
predicted using the additive model. These computations for Phadke’s[3] study

Figure 5.
Factors effects on
MRSN (case study 2)
and proposed optimization analyses are displayed in Table VII. According to Multi-response
the table, an improvement in surface defects for the proposed optimization problems in the
procedure analysis (when the weight ratio is set to 2:1:1) is equal to [1.49 – Taguchi method
(–56.69)] = 58.18dB, which is larger than the improvement from Phadke’s study
36.85dB. Similarly, the improvement in thickness uniformity for the proposed
optimization procedure analysis is better than that of Phadke’s study. A slight
increase occurs in the deposition rate for the proposed procedure when the 379
weight ratio is set to 1:1:1, but a slight reduction when the weight ratio is set to
2:1:1. As a result, the proposed procedure can be considered as a more effective
approach than Phadke’s[3] study (based on an engineer’s judgement) in the
multi-response case.

Starting Optimum
condition condition Improvement

Refractive SN 32.09 37.93 5.84dB


index Average 2.0216 1.939
Variance 0.00198 0.00048
Deposition SN 21.617 25.44 3.77dB Table VI.
thickness Average 1,043.267 976.6 Results of confirmation
Variance 6,000.8 2,179.84 experiment

Optimum condition (dB) Anticipated condition (dB)


Proposed procedure Proposed procedure
Starting Phadke’s weight weight Phadke’s weight weight
condition study[3] 1:1:1 2:1:1 study[3] 1:1:1 2:1:1

Surface
defects –56.69 –19.84 –24.22 1.49 36.85 32.47 58.18

Thickness 29.95 36.79 40.24 41.23 6.84 10.29 11.28 Table VII.
Prediction using the
Deposition additive model (case
Rate 34.97 29.60 32.44 27.08 –5.37 –2.53 –7.89 study 2)

Conclusions
A procedure has been proposed in this study to achieve the optimization of
multi-response problems in the Taguchi method. The procedure includes four
phases, i.e. computation of quality loss, determination of the multi-response
signal to noise ratio, determination of the optimal factor/level combination, and
performing the confirmation experiment. Four significant contributions are
achieved in this procedure. First, the procedure is a universal approach which
can be used in any type of multi-response problem. Second, only one
performance measurement (MRSN) is required for the multiple responses at
IJQRM each experimental trial. Third, the proposed method can simultaneously deal
14,4 with the multi-response problem involving continuous and discrete data types.
Fourth, the importance of each response can be adjusted easily by altering the
weight of the response. Additionally, an experiment of plasma-enhanced
chemical vapour deposition process in the IC manufacturing has been
performed and a case study from Phadke’s[3] textbook has also been evaluated.
380 Those results indicate that the proposed procedure can achieve the optimization
of multi-response problems in the Taguchi method.
References
1. Taguchi, G., Elsayed, E.A. and Hsiang, T., Quality Engineering in Production Systems,
McGraw-Hill, Singapore, 1989.
2. Kackar, R.N., “Off-line quality control, parameter design and the Taguchi method”, Journal
of Quality Technology, Vol. 17 No. 4, 1985, pp. 176-88.
3. Phadke, M.S., Quality Engineering Using Robust Design, Prentice-Hall, Englewood Cliffs,
NJ, 1989.
4. Logothetis, N. and Haigh, A., “Characterizing and optimizing multi-response processes by
the Taguchi method”, Quality and Reliability Engineering International, Vol. 4, 1988,
pp. 159-69.
5. Hung, C.H., “A cost-effective multi-response off-line quality control for semiconductor
manufacturing”, Master thesis, National Chiao Tung University, Taiwan, 1990.
6. Shiau, G. H., “A study of the sintering properties of iron ores using the Taguchi’s parameter
design”, Journal of the Chinese Statistical Association, Vol. 28 No. 2, 1990, pp. 253-75.
7. Tai, C.Y., Chen, T.S. and Wu, M.C., “An enhanced Taguchi method for optimizing SMT
processes”, Journal of Electronics Manufacturing, Vol. 2, 1992, pp. 91-100.
8. Maghsoodloo, S., “The exact relation of Taguchi’s signal-to-noise ratio to his quality loss
function”, Journal of Quality Technology, Vol. 22 No. 1, 1990, pp. 57-67.

You might also like