6.4 Project of 2k design 6-62 10.

3 Determination of the best design 10-19

6.5 Single replicate of 2k design 6-68 10.4 Main steps for Taguchi parameter design 10-39
6.6 Center points in 2k design 6-79
6.7 The advantages of factorial designs 6-91 11. Other Experimental Methods (cancelled)
7. Fractional Experiments
7.1 Introduction 7-1
7.2 One half fraction 7-3
7.3 Smaller fraction 7-43
8. A Case Study
9. Response Surface Methods
9.1 Introduction 9-1
9.2 First-order model 9-10
9.3 Second-order model 9-30
9.4 Choice of an experiment design 9-34
9.5 Main steps for RSM search 9-39
10. Taguchi’s Methods
10.1 Introduction 10-1
10.2 Parameter design 10-6
1-5 1-6

1 INTRODUCTION Example 1.1: Two factors: Machine, Feed rate

Response: Process capability
A designed experiment is a test or series of tests, in which, the
experimenter chooses certain factors for study. He purposely varies Table 1.1
those factors in a controlled fashion, and then observes the effect of Process capability Cp
such factors on the response function. Feed rate (mm/sec.)
Machine 0.2 0.4 0.6
USA 2.1, 2.2 1.7, 1.6 1.2, 1.3
Japan 0.9, 1.0 0.7, 0.6 0.4, 0.3

Conclusion: using USA machine at lowest feed rate will result in

max. process capability.

1-7 1-8
Factor (input variable) (3) Run: A run is one set of levels for all factors in an experiment.
In above experiment, there are three factors: (A) temperature,
(1) Types (B) machine and (C) switch. The total number of runs is
Quantitative: Temperature, time: They can be set at any value.
R=4 3 2 = 24
Qualitative: Machines, operators: They have different categories
Switch: It has different status For example, [1] A = 10, B = Japan-made, C = on
..........
(2) Levels
[24] A = 40, B = UK-made, C = off
Temperature (oC): 10, 20, 30, 40 [4 levels]
Machine: made by Japan, by USA, by UK [3 levels] For single factor experiments, run is equivalent to level of the
factor. The number R of runs is equal to the number a of levels.
Switch: on, or off [2 levels]
(4) Replicates (n): number of observations at each run. A larger
value of n increases the accuracy of the experimental results, but
makes the experiment more expensive.
1-9 1-10

Run A B C Run A B C Response (or output, observation)

No Temp. Machine Switch No Temp. Machine Switch
1 10 Japan On 13 30 Japan On
Response is dependent on the values of the factors.
2 10 Japan Off 14 30 Japan Off
3 10 USA On 15 30 USA On Types of response:
4 10 USA Off 16 30 USA Off
5 10 UK On 17 30 UK On Larger the better: product yield, process capability, student’s
6 10 UK Off 18 30 UK Off
mark
7 20 Japan On 19 40 Japan On
Smaller the better: product cost, number of errors
8 20 Japan Off 20 40 Japan Off
9 20 USA On 21 40 USA On Nominal the best: diameter of a shaft
10 20 USA Off 22 40 USA Off
11 20 UK On 23 40 UK On By DOE, we can find out which factors have effects on the response,
12 20 UK Off 24 40 UK Off
and the direction of the influence (i.e., whether the increase of a
factor will increase or decrease the response ?)
1-11 1-12
Example 1.2: Single factor: A text book (a = 2, n = 3, R = 2) Example 1.3: Two factors: A text book, B studying hours
Response: Students’ marks Response: Students’ marks

Table 1.2 (a = 2, b = 3, n = 2, R = 6)
Students’ marks
Table 1.3
Text 1 76, 74, 78 Ave = 76
Students’ marks
Text 2 48, 50, 52 Ave = 50 1 hr 2 hr 3 hr
Text 1 45, 47 59, 58 76, 75
Text 2 32, 33 51, 54 70, 71
Text book (a qualitative factor)
Text 1: book by Montgomery
Text 2: book by Taguchi

1-13 1-14

The objectives of an experiment may include What Makes Designed Experiment (DOE) Special?

(1) Determining which factors xs are most influential on the response, (1) It conducts the experiments in a systematic and efficient way.
y. Reduce design and development time, as well as the cost.

(2) It presents experiment results in the simplest and clearest way.

(2) Determining where to set the influential x's, so that y is near the
desired requirement. (3) It extracts the maximum amount of information from a given set
of experiments.
(3) Determining where to set the influential x's, so that variability in y
is small. (4) It draws the right conclusions, despite the variability presented in
the data.

Examples: Increase process capability

Improve students’ study
Reduce output variability.

1-15 1-16
Example 1.4 Reducing Defect Level in a Process In this application, DOE
A manufacturing engineer is going to apply DOE to a process for
(1) Determines which factors affect the occurrence of defects
soldering electronic components to printed circuit boards, in order to
reduce defect levels even further.
(2) Determines the direction of adjustment for the effective factors to
Factors to be investigated: further reduce the number of defects per unit.
Solder temperature
Preheat temperature (2) Decides whether changing the factors together produces different
Conveyor speed results than just adjusting individual factor separately?
Flux type
Flux specific gravity
Solder wave depth
Conveyor angle
Thickness of the printed circuit board
Types of components used on the board

1-17 1-18
3 SINGLE FACTOR EXPERIMENTS Table 3.1 Tensile strength of paper
___________________________________________________________________________________________________________

Hardwood Observations
3.1 Introduction _____________________________________________________________________________________________

Concen. (%) 1 2 3 4 5 6 Total Aver.

_________________________________________________________________________________________________________________________________
Single factor experiment can be used in the situation where the effect
(influence) of a single factor on the response is dominant. The effects 5 7 8 15 11 9 10 60 10.00
of all other factors are negligible. 10 12 17 13 18 19 15 94 15.67

The procedures and formulae developed for the single factor 15 14 18 19 17 16 18 102 17.00
experiments can be easily modified, and then used for the several 20 19 25 22 23 18 20 127 21.17
factors experiments. __________________________________________________
383 15.96

Study the effect (influence) of hardwood concentration on the tensile

strength of paper.

3-1 3-2

(1) One factor: hardwood concentration.

Paper strength
(2) Four levels: 5%, 10%, 15%, 20%: a = 4.

(3) Response is the tensile strength of paper.

(4) Each level has six observations or replicates: n = 6.

(5) Total number of observations: N = an = 24.

It is extremely important to apply the Randomization in taking Temperature

observations from different levels of hardwood concentration, in
order to balance out the effects of any nuisance variables (e.g., warm 10 20 30
up effect) that may influence the response.
Figure 3.1 Paper strength vs Temperature

3-3 3-4
 Table 3.2 Randomization in taking observations Table 3.3 General data table for a single factor experiment
___________________________________________________________________________________________________________

Hardwood Observations Level Observation Total Average

_____________________________________________________________________________________________

Concen. (%) 1 2 3 4 5 6 Total Aver.

_________________________________________________________________________________________________________________________________ 1 y11 y12 . . y1n y1. y1.
5 712 84 157 112 915 1021 60 10.00 2 y21 y22 . . y2n y2. y2.
10 1219 171 1313 1814 1917 156 94 15.67 . . . . . .
15 1410 1820 1911 1724 1622 185 102 17.00 . . . . . .
20 199 2516 2223 233 188 2018 127 21.17 a ya1 ya2 . . yan ya. ya.
__________________________________________________
y.. y..
383 15.96

The subscript attached to each observation indicates the run order yij the jth observation taken under the ith level of the factor.
(order of taking the observation).
3-5 3-6

yi. the total of the observations at the ith level 3.2 The Analysis of Variance (ANOVA)
n ANOVA is the most important statistical tool used in DOE. Its task
y i. yij (3.1) is to decide whether and which factors or interactions have
j 1
significant effect (influence) on the response.
yi. the average of the observations at the ith level
Model of ANOVA
y i. = y i. / n (3.2)
Yij i ij i ij (3.5)
y.. the grand total of all observations
a n a the overall mean. y.. (mean average)
y.. yij yi . (3.3)
i 1 j 1 i 1 the ith level mean.
i i yi.
y.. the grand average of all observations. the ith level effect.
i i i y i. y..
y.. = y.. / N (3.4)
ij the random error. ij yij i yij yi.
N = an is the total number of observations
3-7 3-8
Machining Example (1) y1. 2.1 2.0 1.9 6.0, y1. y1. / 3 2.0
Study the effect of machine on the process capability y 2. 0.9 1.0 1.1 3.0, y 2. y 2. / 3 1.0
Table 3.4 y.. (2.1 2.0 1.9) (0.9 1.0 1.1) 9.0
Machine Process capability Cp
y.. 9.0 / 6 1.5
USA 2.1, 2.0, 1.9
Japan 0.9, 1.0, 1.1 1 y1. 2.0, 2 y2. 1.0
y.. 1.5
(1) One factor: machine (qualitative).
1 1 y1. y.. 2.0 1.5 0.5
(2) Two levels: USA-made, Japan-made, a = 2.
2 2 y2. y.. 1.0 1.5 0.5
(3) Response is the process capability.
(4) Each level has 3 observations or replicates: n = 3.
(5) Total number of observations: N = an = 2 3 = 6.
3-9 3-10

Evaluation of Errors
y
11 = 0.1
For instance, if i = 1, j = 1, yij = y11 = 2.1
1 =2
ij yij i
1 = 0.5
11 y11 1 y11 y1. 2.1 2.0 0.1 = 1.5
y11 = 2.1 2 = -0.5
Errors for the machining example (1) 2 =1
Table 3.5
Machine Errors
USA 0.1, 0.0, -0.1 y11 = + 1 + 11
Japan -0.1, 0.0, 0.1 = 1.5 + 0.5 + 0.1 = 2.1

Figure 3.2

3-11 3-12
 1 2
1

1 2
1

1 2 1 2 3
1 2 3
3 4 2 3 4
2 3 4
3 4
4

3 4
4

Figure 3.3
Figure 3.4

3-13 3-14

Hypothesis Test: to test if a factor has effect on the response. Null: H o: the factor has no effect on the response

If a factor has no effect on the response, the response is always the Alternative: H1: the factor has effect on the response
same, regardless the change of the factor,
Ho : 1 2 a 0 (| i| is small, no effect)
1 2 a
H1 : i o for at lease one i (| i| is larger, has effect) (3.6)
( 1 ) ( 2 ) ( a ) 0
1 2 a 0 The larger the value of any | i|, the more effective the factor will be.
Here, the magnitude of i, rather than its sign, makes sense.
If a factor has effect, the value of the response will be different, along
with the change of the factor. The more the response differs, the Type I error: the null hypothesis Ho is rejected when it is actually
greater the effect of the factor is. true (The factor is concluded effective, when it is not)

i o for at least one i Type II error: the null hypothesis is accepted when it is actually
false (The factor is concluded ineffective, when it is effective).

3-15 3-16
Machining Example (2) y
Table 3.6
Machine Process capability Cp 1 = 2 =0
USA 1.6, 1.5, 1.4 = 1 = 2 =1.5
Japan 1.4, 1.5, 1.6
Minor variation is caused by
errors, rather than the machines
1 1.5, 2 1.5 1.5
1 1 15
. 15
. 0
2 2 15
. 15
. 0

Conclusion: in this example, the factor (machine) has no effect on

the process capability. Because, when using different Figure 3.5
machines, the response is always the same, on average.
3-17 3-18

Analysis of Variance (ANOVA) (1) Calculate sum of squares (SS)

Major Steps Three types of differences:

[1] Difference between an individual observation and the grand
(1) Calculate sum of squares (SS). average:
yij - y..
(2) Determine degrees of freedom (D.O.F.).
[2] Difference between a level average and the grand average:
(3) Calculate mean square (MS).
yi . - y.. i

(4) Calculate the test ratio F0. [3] Difference between an individual observation and the
corresponding level average:
(5) Make conclusion. yij - yi. ij

Note: yij - y.. yi. - y.. yij - yi.

3-19 3-20
a n a n a n
Sum of squares is the sum of squares of the above three differences.
yij - y.. yi. - y.. yij - yi.
i 1 j 1 i 1 j 1 i 1 j 1
SST The total sum of squares, which is a measure of total
a n a n a n variability in the data.
2 2 2
yij - y.. yi . - y.. yij - yi .
i 1 j 1 i 1 j 1 i 1 j 1 a n
SS T ( yij - y.. ) 2 (3.7)
i 1 j 1
SST + SSF = SSE
SSF the sum of squares due to the factor, that is the sum of squares
of differences between factor level averages and the grand
average (difference between levels).
a n a a
2
SS F ( yi . y.. ) 2 n ( yi . y.. ) 2 n i (3.8)
i 1 j 1 i 1 i 1

3-21 3-22

SSE the sum of squares due to error, that is the sum of squares Alternative Formulae of Calculating the Sum of Squares
of differences between observations and their level
a n
averages (difference within levels). y . .2
SS T = yij 2 - (3.11)
a n a n i 1 j 1 an
2
SS E ( yij yi . ) 2 ij (3.9)
i 1 j 1 i 1 j 1 a
yi . 2 y. .2
SS F = - (3.12)
i 1 n an

SS T = SS F + SS E (3.10) SS E = SS T - SS F (3.13)

Formulae (3.7) to (3.9) are used to describe the underlying ideas.

Formulae (3.11) to (3.13) are used for actually computation.

3-23 3-24
Machining Example (1) If SSF is large, it is due to differences among the means at the
different factor levels. See equation (3.8), a large SSF means that 2i
Table 3.7
(or | i|) is great. It is an indication that the factor has significant effect
Machine Process capability Cp
on the response.
USA 2.1, 2.0, 1.9
Usually, SSF is standardized by taking SSF / SSE . By comparing SSF
Japan 0.9, 1.0, 1.1
to SSE, we can see how much variability is due to changing factor
levels and how much is due to error.
a = 2, n = 3
y1 6 y2 3 y 9 Machining Examples
y1 2 y2 1 y 15
. Table 3.8
Effective ? SSF/SSE
SST 2.12 2.0 2 1.9 2 0.9 2 1.0 2 1.12 9 2 / 6 1.54 Example 1 Yes 37.5
SS F 6 2 32 / 3 9 2 / 6 1.50 Example 2 No 0
SS E 1.54 1.50 0.04
3-25 3-26

(2) Determine degrees of freedom (D.O.F.). Machining Example (1)

Table 3.9
total number of observations N: an
Machine Process capability Cp
degrees of freedom of SST: N - 1 = an - 1 (3.14) USA 2.1, 2.0, 1.9

degrees of freedom of SSF: a-1 (3.15) Japan 0.9, 1.0, 1.1

degrees of freedom of SSE: R (n - 1) = a (n – 1) (3.16) a = 2, n = 3

where, (n - 1) is the degrees of freedom of errors in a run. total number of observations N: 2 3=6
degrees of freedom of SST: N–1 =6–1=5
Equation of DOF
degrees of freedom of SSF: a–1 =2–1=1
an - 1 = (a - 1) + a (n - 1) (3.17)
degrees of freedom of SSE: a (n-1) = 2 (3 - 1) = 4
DOF of SST = DOF of SSF + DOF of SSE
Check equation (3.17): 5=1+4
3-27 3-28
(3) Calculate mean square (MS). (4) Calculate the test ratio F0.

The ratio between a sum of squares and its DOF. MS F

Fo = (3.20)
MS E

SS F
MS F = (3.18)
a 1 If H0 is true, F0 follows a theoretical F distribution, which is
completely determined by the two parameters 1 and 2.
SS E
MS E = (3.19) 1 =a-1 the D.O.F. of MSF (numerator)
a ( n - 1)
2 = a(n - 1) the D.O.F. of MSE (denominator)

There is no need to calculate MS for the SST

3-29 3-30

f(x) (5) Make conclusion.

1 =1
2 = 4
(1) Decide the curve from
SS F
1 and 2 .
MS F a (n 1) SS F SS F
(2) Decide F , 1, 2 from Fo = a 1 Q
MS E SS E a 1 SS E SS E
a ( n 1)

a (n 1)
Where, Q is a constant.
a 1

SS F
F , 1,
F0 is equivalent to (the standardized SSF) except for a constant Q.
2
SS E
Figure 3.6 SS
The density function curve of an example of the F distribution However, while F0 follows the theoretical F distribution, F does
SS E
not.
3-31 3-32
 SS F
If F0 is large ----> is large ---> SSF is large ----> 2i is large --->
SS E
| i| is large ---> the factor has significant effect on the response.

Control limit: F , 1, 2

p
1 = a - 1 and 2 = a(n - 1) are the two parameters of an F
distribution
is the probability of Type I error, which means concluding that F , 1, 2
the factor has significant effect on the response when it in fact has
no effect. Usually, set = 1% or 5% Sometimes, p value is used to indicate the result of ANOVA. p is
equal to the area under the F probability curve and beyond F0.
If Fo < F , 1, 2
, conclude that the factor has no effect.
If p is small ----> F0 is large ---> the factor has significant effect on
If Fo > F , 1, 2
, conclude that the factor has significant effect. the response
3-33 3-34

Table 3.10 Control Limit F , 1, 2

found from the F Table Table 3.11

= 0.05

1 1 2 3 4 5 .
2
1 161.4 199.5
2 18.51 19.00
3 10.13 9.55
Examples:
4 7.71 6.94
F0.05,2,1 = 199.5
.
F0.05,1,4 = 7.71

3-35 3-36
 Table 3.12 Table 3.13
Analysis of Variance for a Single-Factor Experiment
___________________________________________________

Source of Sum of Degrees of Mean Fo

Variation Squares Freedom Square
___________________________________________________

Factor SSF a-1 MSF MS F

MS E

Error SSE a(n-1) MSE

Total SST an-1

___________________________________________________

3-37 3-38

Machining Example (1) (a = 2, n = 3, N = 6), from Table 3.7 Specify = 0.05, Control limit F , 1, = F0.05,1, 4 = 7.71
2

Table 3.14 Since F0 F , 1, 2

, machines have significant effect on the process
___________________________________________________
capability.
Source of Sum of Degrees of Mean Fo
Variation Squares Freedom Square Sum of squares: 1.50 + 0.04 = 1.54
___________________________________________________
Degrees of freedom 1 + 4 = 5
Machine 1.50 1 (= 1) 1.50 150
Error 0.04 4 (= 2) 0.01
Total 1.54 5
___________________________________________________

3-39 3-40
Machining Example (2) (a = 2, n = 3, N = 6), Table 3.16 Example: Tensile strength of paper
___________________________________________________________________________________________________________
Table 3.15 Hardwood Observations
_____________________________________________________________________________________________
Source of Sum of Degrees of Mean Fo Concen. (%) 1 2 3 4 5 6 Total Aver.
_________________________________________________________________________________________________________________________________
Variation Squares Freedom Square
5 7 8 15 11 9 10 60 10.00
Machines 0 1 (= 1) 0 0 10 12 17 13 18 19 15 94 15.67
Error 0.04 4 (= 2) 0.01 15 14 18 19 17 16 18 102 17.00
Total 0.04 5 20 19 25 22 23 18 20 127 21.17
___________________________________________________ __________________________________________________
Specify = 0.05, Control limit F , 1, = F0.05,1, 4 = 7.71 383 15.96
2

Since F0 F , 1, 2
, machines don’t have significant effect on the a = 4, n = 6, N = an = 24.
process capability.
3-41 3-42

4 6
2 y.. 2 Table 3.17 ANOVA for tensile strength of paper
SS T = yij - ___________________________________________________________________________________________________________________________________

i 1 j 1 an Source of Sum of Degrees of Mean Fo

(383) 2 Variation Squares Freedom Square
__________________________________________________________________________________________________________________________________
= (7) 2 + (8) 2 + + (20) 2 - = 512.96
24 Hardwood 382.79 3 (= 127.60 19.61
1)
4
yi .2 y..2 concentration
SS F = -
i 1 n an Error 130.17 20 (= 6.51
2)
2 2 2 2 2
(60) (94) (102) (127 ) (383)
= - = 382.79 Total 512.96 23
______________________________________________________________________________________________________________________________________
6 24
SS E = SS T - SS F Specify = 0.01, Control limit F , 1, = F0.01, 3, 20 = 4.94
2
= 512.96 - 382.79 = 130.17
Since F0 F , 1, 2
, hardwood concentration has significant effect on
the tensile strength of paper.

3-43 3-44
SUMMARY: ANOVA for Single Factor Experiment 5. Calculate the degrees of freedom for each of SST, SSF, SSE (3.14,
3.15, 3.16).
1. Decide the factor to be investigated and the following two
parameters: 6. Calculate the mean squares MSF and MSE (3.18, 3.19).
1) number of levels a; 7. Calculate the ratio F0 between MSF and MSE (3.20).
2) replicate n.
8. Specify a type I error , and find the control value F , 1, 2
from
2. Carry out the experiments, obtain N (= an) observations yij by a
random manner. the F distribution Tables ( 1 = a - 1, 2 = a (n – 1) ).

3. Check if the residuals (estimates of errors) satisfy the 9. Conclude if the factor has significant effect on the response.
requirements.

4. Calculate sum of squares SST, SSF and SSE (3.11, 3.12, 3.13).

3-45 3-46

3.3 Test on Individual Level Means

ANOVA results in a single parameter F0, which can tell whether a

factor has effect on the response, or whether the mean response
values will be different at different levels of the factor.

However, ANOVA can neither decide the direction of influence of

the factors, nor identify which factor level (or levels) will result in
significant difference.

Orthogonal Contrast is the most useful tool to test the individual

level means.

Figure 3.7 Box plots of paper strength example

3-47 3-48
The plots indicate that changing the hardwood concentration has a Example (1)
strong effect on the paper strength; specifically, higher hardwood
concentrations produce higher observed paper strength. It might be suspected that hardwood concentrations 5% and 20%
produce the same paper strength.
Box plots show the variability of the observations within a factor
level and the variability between factor levels. Hypothesis (in terms of level mean):

However, less subjective approach (contrast method) is required to Ho : 1 = 4 (no difference between levels 1 and 4)
test the individual level means. (3.21)

H1 : 1 = 4 (different) (3.22)

3-49 3-50

yi. Example (2)

Since i yi. and y i . nyi . yi.
n It is suspected that the average paper strength at hardwood
(3.21) can be rewritten as: concentrations 5% and 10% does not differ from the average paper
y1. y4 . strength at hardwood concentrations 15% and 20%.
ny1. ny4. Hypothesis (in terms of level mean):
y1. y4 .
1 + 2 3 + 4
y1. y4 . 0 Ho : =
2 2
(3.25)
1 + 2 3 + 4
Hypothesis (in terms of level total): H1 : =
2 2
H o : y1. - y4. 0 (3.23) It can be simplified as:
H1 : y1. - y4. 0 (3.24) Ho : 1 + 2 = 3 + 4 (3.26)
H1 : 1 + 2 = 3 + 4 (3.27)
3-51 3-52
(3.26) can be rewritten as: General Form of Contrast
Contrast is a linear combination of run totals (or level totals for single
y1. y 2. y 3. y 4.
factor experiments):
ny1. ny 2. ny3. ny4. a

y1. y 2. y 3. y 4. C = ci y i. (3.30)
i 1
y1. y 2. y 3. y4. 0
With the restriction on the coefficients ci:
a
Hypothesis (in terms of level total): ci = 0. (3.31)
i 1
H o : y1. + y 2. - y3. - y 4. 0 (3.28)

H1 : y1. + y 2. - y 3. - y 4. 0 (3.29) For example, if a = 4,

Contrast: C c1 y1. c2 y 2. c3 y 3. c4 y 4 .
The Hypothesis is transformed by replacing the level mean i by the
Restriction: c1 c2 c3 c4 0
corresponding level total yi..

3-53 3-54

The right-hand side of (3.23) or (3.24) is a contrast, The right-hand side of (3.28) or (3.29) is also a contrast,
C = y1. - y 4. C = y1. + y 2. - y 3. - y 4.
1 y1. + 0 y 2. + 0 y 3. + (-1) y 4. 1 y1. + 1 y 2. + (-1) y 3. + (-1) y 4.
where, where,
c1 = 1, c2 = 0, c3 = 0, c4 = -1 c1 = 1, c2 = 1, c3 = -1, c4 = -1
c1 + c2 + c3 + c4 = 0 c1 + c2 + c3 + c4 = 0

So, hypothesis (3.23) and (3.24) can be rewritten in terms of the So, hypothesis (3.28) and (3.29) can also be rewritten in terms of the
contrast: contrast:
Ho : C 0 (no difference between level means) (3.32) Ho : C 0 (no difference between level means) (3.34)
H1 : C 0 (great difference between level means) (3.33) H1 : C 0 (great difference between level means) (3.35)

3-55 3-56
Procedure to Test the Hypothesis for Contrasts (3) The mean square of the contrast:
(similar to ANOVA)
SS C
MS C SS C (3.38)
(1) The sum of squares of a contract C: 1

a 2 (4) F0 is the ratio between MSc and the mean square error MSE. It
c i y i. follows an F distribution with 1 = 1 and 2 = a(n-1).
i 1 C2
SS c = a
(3.36) MS C SS C
2 Q F0 (3.39)
n ci
i 1
MS E MS E

a SSC
Q is a constant Q n ci2 (3.37) If F0 is large ----> is large ---> SSC is large
i 1
MS E
----> C2 (or C ) is large ---> C 0
(2) The degrees of freedom of SSC is always equal to one.
---> significant difference between level means.
3-57 3-58

(5) Making conclusion (type I error should be specified): The Example of Tensile Strength of Paper
Table 3.18
If Fo < F , H0 is true, no difference between the level ___________________________________________________________________________________________________________
, 1, 2
Hardwood Observations
means. _____________________________________________________________________________________________

Concen. (%) 1 2 3 4 5 6 Total Aver.

_________________________________________________________________________________________________________________________________

If Fo > F , 1, 2
, H1 is true, significant difference between the
5 7 8 15 11 9 10 60 10.00
level means.
10 12 17 13 18 19 15 94 15.67
15 14 18 19 17 16 18 102 17.00
20 19 25 22 23 18 20 127 21.17
__________________________________________________
a = 4, n = 6
y1. = 60, y2. = 94, y3. = 102, y4. = 127

3-59 3-60
Problem: if hardwood concentrations 5% and 20% produce the same Five Steps:
paper strength ?
(1)
a
Hypothesis:
Q n ci2 6 12 02 0 2 ( 1)2 12
i 1
Ho : 1 = 4 Ho : y1. - y 4. 0 Ho : C 0
H1 : = 0 C2 ( 67 ) 2
1 4 H1 : y1. - y 4. H1 : C 0 SS c = 374
Q 12

Contrast: C y1. - y4. 60 127 67 (2) 1 = 1, 2 = 4 (6 - 1) = 20

c1 = 1, c4 = -1, and c2 = c3 = 0 (3) The calculation of MSC can be omitted, since it is equal to SSC.

3-61 3-62

(4) General Steps to Test the Individual Level Means

SS C 374
F0 57.46 (1) Decide the level means to be tested, and set up the hypothesis in
MS E 6.51 terms of the level means.
(Note: MSE = 6.51 was calculated during ANOVA, which is (2) Convert the hypothesis to a contrast form in terms of the level
always carried out before the test on individual level means) totals.

(5) Specify = 0.01, control limit: F .

F0.01,1, 20 = 810 (3) Calculate the sum of squares of the contrast SSC (3.36).
, 1, 2

(4) Determine the degrees of freedom: 1 = 1, 2 = a(n-1).

Since Fo > F , 1, 2 (5) Calculate the ratio F0 (3.39).
Conclusion: There is significant difference in paper strength (6) Specify , and find the control limit F , 1, 2 .
between using hardwood concentrations 5% and
20%. (7) Conclude if there is significant difference between the level
means.

3-63 3-64
3.4 Guidelines for Designing Experiments (3) Selection of the response
(1) Recognition and statement of the problem. Most often, the average or standard deviation (or both) of the
measured characteristic will be the response variable.
Fully develop all ideas about the problem and the specific
objectives of the experiment; (4) Choice of experimental design.
Solicit input from all concerned parties -- engineering, quality,
Decide the factorial fraction experiment, select the replicate
marketing, customer, management and operators.
(sample size) and run order for the experimental trials.
(2) Choice of factors and levels. Decide whether or not blocking or other randomization methods are
Choose the factors, their ranges and levels. Process knowledge is involved.
required.
(5) Performing the experiment.
Investigate all factors that may be of importance and avoid being
overly influenced by past experience. Ensure that everything is being done accordingly to plan.
Keep the number of factor levels low (Most often two levels are
used.) in the early stage
3-65 3-66

(6) Data analysis

Analyze the data so that results and conclusions are objective rather
than judgmental.
Use software packages and simple graphical methods
Carry out residual analysis.

(7) Conclusions and recommendations.

(8) Experimentation is an important part of the learning process.

New hypotheses are formulated based on the investigation of the
tentative hypotheses. Don’t design a single, large comprehensive
experiment at the start of a study. As a rule of thumb, the first
experiment should spend no more than 25% of the total budget of
the study.

3-67
5 FACTORIAL EXPERIMENTS Case one
Two factors: A: Machine, B: Feed rate
5.1 Introduction
Response: Process capability
Factorial experiments study the effects of several factors on the
response. Table 5.1
Process capability Cp
In factorial experiments, all possible runs (combinations of the levels Feed rate (mm/sec.)
of the factors) are investigated. Machine 0.2 0.4 0.6
For example, for two factors A: temperature (with a levels) and B: USA 2.1, 2.2 1.7, 1.6 1.2, 1.3
humidity (with b levels), the factorial experiment has ab possible Japan 0.9, 1.0 0.7, 0.6 0.4, 0.3
runs.
a = 2, b = 3, n = 2, R =ab = 6
If for each run, n observations are taken, then, the total number N of N = Rn = abn = 2 3 2 = 12
the observations is equal to abn.

5-1 5-2

Case two 5.2 ANOVA for Two-Factor Factorial Experiments

Three factors: A, B and C.
Mathematical model
a = 2, b = 3, c = 4, n = 2, R = abc = 24
Yijk i j ij ijk (5.1)
N = Rn = abcn = 2 3 4 2 = 48
the overall mean

i the effect incurred at the ith level of factor A

Number of runs is equal to the product of the levels of all factors.
j the effect incurred at the jth level of factor B
( )ij the effect incurred at the interaction between the ith
level of A and the jth level of B
ijk the random error (normally and independently
distributed, zero mean and constant variance)

5-3 5-4
Interaction indicates the influence of one factor on the effect of Table 5.2 Data Arrangement for a Two-Factor Factorial Design
another factor, and vice versa. Factor B
1 2 b Total Average
Factors A: temperature 1 y111, y112, y121, y122, y1b1, y1b2, y1.. y1..
Factor A , y11n , y12n , y1bn
B: pressure
2 y211, y212, y221, y222, Y2b1, y2b2, y2.. y 2..
Response Process yield , y21n , y22n , y2bn

If: (1) Increasing temperature alone cannot increase the yield; and
a ya11, ya12, ya21, ya22, yab1, yab2, ya.. y a..
(2) Increasing pressure alone cannot increase the yield; and , ya1n , ya2n , yabn
(3) Increasing temperature and pressure at the same time does Total y.1. y.2. y.b. y…
increase the yield Average y . 1. y . 2. y.b. y ...

Then: There is interaction between temperature (factor A) and yijk is the kth observation taken at the ith level of factor A and the jth
pressure (factor B) level of factor B.
The total number of observations: N = abn
5-5 5-6

Total and average at the ith level of factor A: Hypotheses

b n
yi ..
yi .. yijk y i .. i = 1,2,...,a Factor A:
j 1 k 1 bn (5.2)
Ho : 1 2 ... a 0 Factor A has no effect on y
Total and average at the jth level of factor B:
a n y. j . H 1 : at least one i 0 Factor A has effect on y
y. j . yijk y.j. j = 1,2,...,b
i 1 k 1 an (5.3) Factor B:

Total and average when A at the ith level and B at the jth level: H0 : 1 2 b 0 Factor B has no effect on y
n yij .
yij . yijk y ij . H 1 : at least one j 0 Factor B has effect on y
k 1 n (5.4)
Interaction AB:
Grand total and average
a b n Ho : ( ) ij 0 for all i, j Interaction AB has no effect y
y...
y...= yijk y...= (5.5)
i 1 j 1k 1 abn H1 : at least one ( ) ij 0 Interaction AB has effect on y
5-7 5-8
Equation of the Sum of Squares Sum of squares for interaction AB:
a b
SST SSA SSB SSAB SSE (5.6) 2
SSAB n yij . yi.. y. j . y...
i 1 j 1
Total sum of squares: (5.10)
a b yij2. y...2
a b n a b n
2 y...2 SSA SSB
SS yijk y... yijk2 (5.7) i 1 j 1
n abn
i 1 j 1 k 1 i 1 j 1 k 1 abn

Sums of squares for main effects: Sum of squares for error:

a a
2 yi2.. y...2 SSE SS SSAB SSA SSB (5.11)
SS A bn yi .. y... (5.8)
i 1 i 1 bn abn
b b
2 y.2j . y...2
SS B an y. j . y... (5.9)
j 1 j 1 an abn

5-9 5-10

Degrees of freedom Mean Square

SST abn - 1 (or N - 1) MS = SS / D.O.F. (5.13)

SSA a-1
F0 = MS / MSE (5.14)
SSB b-1
SSA MSA
Mean square of A: MSA F0
SSAB (a - 1)(b - 1) a 1 MSE

SSE ab(n - 1) SSB MSB

Mean square of B: MSB F0
b 1 MSE
Equation of D.O.F.
SSAB MSAB
DOFSST DOFSS A DOFSSB DOFSS AB DOFSSE Mean square of AB: MSAB F0
a 1 b 1 MSE
abn 1 ( a 1) (b 1) ( a 1)(b 1) ab( n 1) (5.12)

5-11 5-12
Mean square of error: MSE
SSE Table 5.3 ANOVA for a two-factor factorial design
ab n 1
Source of Sum of Degrees of Mean square F0
variation squares freedom

The denominator of the F0 ratio is always equal to MSE A SSA a-1 SS A MS A

MS A
a 1 MS E
Control limit: F , 1, 2 B SSB b-1 SS B MS B
MS B
b 1 MS E
is the probability of Type I error.
AB SSAB (a - 1)( b- 1) SS AB MS AB
1 = D.O.F. of the numerator of the F0 ratio. MS AB
( a 1)( b 1) MS E

2 = ab(n - 1), D.O.F. of MSE Error SSE ab(n - 1) SS E

MS E
ab(n 1)
F , 1 , 2 may differ for different effects or interactions as 1 may
Total SST abn - 1
not be the same.
5-13 5-14

How to find whether a factor (or an interaction) has effect on the Example 5.1
response? A: machine B: temperature Response: process capability Cp
(1) Calculate the sum of squares SS and then the mean square MS a = 3, b = 2, n = 3, N = 3 2 3 = 18
of the factor (or interaction) to be studied. Table 5.4
1 10o C 2 20o C yi..
(2) Calculate the ratio F0 between this mean square MS and the
mean square error MSE. 1 USA 4.0, 4.5, 4.3 12.8 5.4, 4.9, 5.6 15.9 28.7
2 Japan 5.6, 4.9, 5.4 15.9 5.8, 6.1, 6.3 18.2 34.1
(3) Decide the control limit F , 1 , 2 , where, is the type I error
probability, 1 is the D.O.F. of the factor (or the interaction), 3 UK 3.8, 3.7, 4.0 11.5 5.5, 5.0, 5.0 15.5 27.0
2 is the D.O.F. of the error. y.j. 40.2 49.6 89.8 = y…

(4) If F0 is larger than F , 1 , 2 , the factor (or interaction) has Number inside the rectangle is the total observations in a run.
significant effect on the response; otherwise it has no effect.
5-15 5-16
 a b n a b
2 y...2 yij2. y...2
SS yijk SS AB SS A SS B
i 1 j 1k 1
abn i 1 j 1 n abn

89.82 12.82 15.92 15.52 89.82

2 2 2 4.58 4.91 0.24
4.0 4.5 5.0 10.72 3 3 3 3 2 3
3 2 3
a SS E SS SS AB SS A SS B
yi2.. y...2
SS A
i 1 bn abn 10.72 0.24 4.58 4.91 0.99
28.7 2 341
. 2 27.0 2 89.82
4.58
2 3 2 3 2 3 3 2 3
b y.2j. y...2
SS B
j 1 an abn
40.2 2 49.6 2 89.8 2
4.91
3 3 3 3 3 2 3
5-17 5-18

Table 5.5 ANOVA Table Specify = 0.05

(1) Factor A (machine)
Source of Sum of Degrees of Mean F0
variation squares freedom square F , 1, 2 = F0.05,2,12 = 3.89

machine 4.58 2 2.29 27.758 F0 > F , 1, 2 , so, machine has effect on the Cp.

temp. 4.91 1 4.91 59.515 (2) Factor B (temperature)

F , 1, 2 = F0.05,1,12 = 4.75
AB 0.24 2 0.12 1.455
F0 > F , 1, 2
, so, temperature has effect on the Cp.
Error 0.99 12 0.0825
(3) Interaction AB (between machine and temperature)
Total 10.72 17 F = F0.05,2,12 = 3.89
, 1, 2

F0 < F , 1, 2 , so, interaction almost has no effect on the Cp.

5-19 5-20
6 2K EXPERIMENTS Advantages:
6.1 Introduction (1) 2k experiment can substantially reduce the total number of runs in
2k experiment (or called 2k design) is a special and simplest, but the an experiment for given number of factors, therefore, more factors
most widely used, factorial experiment with k factors, each at two can be considered at the same time. 2k design is especially useful
levels. to the screening experiments in the early stage when we are not
certain which factor (or factors) may have significant effect on the
The value of 2k by itself is equal to the total number of runs in the response.
experiment.
For example, due to the time limit, we can only conduct 65 runs in
For example, if there are three factors A (temperature), B (humidity) an experiment.
and C (feed rate) in an experiment, i.e., k = 3.
If each factor is to be investigated in 3 levels, only 3 factors can be
Number of levels for each factor: a = b = c = 2
studied, because 33 < 65 < 34
Number of runs: abc = 2 2 2 = 23 = 8
If each factor is to be investigated in 2 levels, 6 factors can be
The replicate n is the number of observations at each run. studied, because 26 < 65 < 27
6-1 6-2

(2) In 2k experiments, the formulae for calculating the sum of squares In 2k experiments, each factor has only two levels. They are called
are much simpler than for other types of factorial experiments. low (-) level and high (+) level, respectively.

(3) In 2k experiments, we can easily use a simple statistics effect to For examples (response is Cp)
estimate the effect (or influence) of a factor on the response. The
complicated ANOVA procedure may be saved under certain Table 6.1
circumstances. Factor Low (-) High (+)
A temperature 10 oC 30 oC
B machine USA-made Japan-made

The definition of low and high levels for the qualitative factor
machine (B) is arbitrary. If they are defined in another way (Japan-
made is low level and USA-made is high level), the conclusions
resulting from a DOE will not be changed

6-3 6-4
Special Notations Used in the 2k Designs:
High b ab B
(1) A run is represented by a series of lower case letters. (Japan) (-) (+)
(2) If a letter is present, the corresponding factor is set at the high (-) b
(1)
level in that run. first diagonal
B A
(3) If a letter is absent, the corresponding factor is set at the low level. a ab
(+)

Low A run is indicated

(USA) (1) a by a point
Low A High
(10oC) (30oC)

Figure 6.1
6-5 6-6

The Four Runs in a 22 Design: In addition, each of the notations (1), a, b, and ab has a value, which
is equal to the sum of all the n observations taken at the
(1) factor A at low level; factor B at low level. corresponding run.
a factor A at high level; factor B at low level. Table 6.3
Run A B Response
b factor A at low level; factor B at high level. temp. machine Cp
(1) 10 USA 1.2 1.3 (1) = 2.5
ab factor A at high level; factor B at high level. a 30 USA 2.1 2.2 a = 4.3
Table 6.2 b 10 Japan 1.7 1.8 b = 3.5
ab 30 Japan 2.7 2.8 ab = 5.5
Run A B
temp. machine Two usages of the lower-case-letter notations:
(1) - -
a + - (1) represent a run;
b - + (2)indicate the total of the observed values of the response in that run.
ab + +

6-7 6-8
The special notation is used throughout 2k experiments (designs). 6.2 22 Design

Example: A 23 design with three factors A, B, C y+ is the response value when a factor is at its high level

Table 6.4 y- is the response value when a factor is at its low level
No Run A B C
1 (1) - - - The difference (y+ - y-) is called the effect of that factor. It is defined
2 a + - - as the change in response produced by a change in the level of the
3 b - + - factor. The larger the effect (the difference (y+ - y-)), the more
4 ab + + - influential the factor is on the response.
5 c - - +
6 ac + - +
7 bc - + +
8 abc + + +

6-9 6-10

For example,
Y(Cp) Y(Cp)
(1) If temperature is the factor, and Cp is the response,
Y+
y+ = 2.5 when temperature is at the high level of 30oC; 2.5

1.7 Y+
y- = 0.8 when temperature is at the low level of 10oC; 1.25
Y-

Then, the effect of the temperature is: (y+ - y-) = 1.7. This means Y-
0.8
temperature has significant effect on the Cp. Increasing the T machine
temperature will increase Cp.
low high low high
10 30 USA Japan

Figure 6.2

6-11 6-12
(2) If machine is the factor, and Cp is the response, Usually, n (n 1) observations are taken at the low level or high level
of a factor, so effect should be generalized as
y+ = 1.25 when machine is made by Japan (high level);
Effect = y y (6.1)
y- = 1.24 when machine is made by USA (low level);
y : Average of all the observations taken when the factor is at the
Then, the effect of the machine: (y+ - y-) = 0.01. This means high level.
machine has very minor effect on the Cp. The value of Cp is
almost constant no matter which machine is used. y : Average of all the observations taken when the factor is at the
low level.

6-13 6-14

Example 6.1 The main effect (or simply the effect) of A is denoted by A.

Response is the process capability Cp A yA yA

a ab b (1) 1
Table 6.5 [ a ab b (1)] (6.2)
2n 2n 2n
Factor A Factor B (temperature)
(machine) 1
8 6 2 4 2
B- (100 C) B+ (200 C) 2 2
A- (USA) (1) 1.8, 2.2 b 0.8, 1.2
The main effect of B is denoted by B.
A+ (Japan) a 3.8, 4.2 ab 2.8, 3.2
B yB yB
(1) = 1.8 + 2.2 = 4
b ab a (1) 1 (6.3)
[b ab a (1)]
a = 3.8 + 4.2 = 8 2n 2n 2n
b = 0.8 + 1.2 = 2 1
2 6 8 4 1
2 2
ab = 2.8 + 3.2 = 6
6-15 6-16
A positive effect means that the response will be greater when the
factor is at the high level. Cp Cp

A = 2 > 0, means Cp is greater if Japanese machine (A+) is

y
used. y
-1
A negative effect means that the response will be greater when the 2
y
factor is at the low level.
y
B = -1 < 0, means Cp is greater if temperature is low. A B
USA machine temp
Japan 10 20

When the magnitude of an effect is large, its influence on response is

significant, no matter the effect is positive or negative.
Figure 6.3

6-17 6-18

For Example 6.1, if we set A+ as USA machine and A- as Japanese From (6.2)
machine, the observed values are as follows:
1
A [ a ab b (1)]
Table 6.6 2n
Factor A Factor B (temperature) 1
(machine) 4 2 6 8 2
0 0 2 2
B- (10 C) B+ (20 C)
A- (Japan) (1) 3.8, 4.2 b 2.8, 3.2
A+ (USA) a 1.8, 2.2 ab 0.8, 1.2 Since A = -2 < 0,
(1) = 3.8 + 4.2 = 8 Cp will be greater if Japanese machine (A-) is used.
a = 1.8 + 2.2 = 4
This conclusion is the same as obtained before.
b = 2.8 + 3.2 = 6
ab = 0.8 + 1.2 = 2
6-19 6-20
Interaction Example 6.2 (without interaction)
If the effects of one factor are different at different levels of another Response is the process capability Cp
factor, there is interaction between these two factors.
For example, if the effects of A are different at different levels of
factor B, there is interaction between A and B. This interaction is Table 6.7
denoted as AB. Factor A Factor B (temperature)
(machine)
For example, a particular medicine has effect on children, but no B- (100 C) B+ (200 C)
effect on adults. This means the age influence the effect of the A- (USA) (1) 2 b 1
medicine. Or in other words, there is interaction between the factor A+ (Japan) a 4 ab 3
medicine and the factor age.

6-21 6-22

The A effect (with the condition that B is low), Example 6.3 (with interaction)
Output: the process capability Cp
A(B-) = 4 – 2 = 2

The A effect (with the condition that B is high),

Table 6.8
A(B+) = 3 – 1 = 2 Factor A Factor B (temperature)
(machine)
The effect of A is identical, no matter B is at the low or high lever. B- (100 C) B+ (200 C)
A- (USA) (1) 3 b 2
Conclusion: there is no interaction between factors A and B. No A+ (Japan) a 1 ab 4
matter temperature is low or high, Japanese machine is always better.

6-23 6-24
The A effect (with the condition that B is low), The greater the difference between A(B+) and A(B-), the more
significant the interaction AB is . So, A(B+) - A(B-) is a measure of the
A(B-) = 1 – 3 = -2 interaction AB (refer to Figure 6.1).

The A effect (with the condition that B is high), 1

AB A( B ) A( B )
2
A(B+) = 4 – 2 = 2 1
yA yA B yA yA B
2
The effects of A are different at different levels of factor B.
1 ab b a (1)
Conclusion: there is interaction between factors A and B. When 2 n n n n
temperature is low, USA machine is better; when temperature is high, 1
Japanese machine is better. ab b a (1)
2n (6.4)

6-25 6-26

When the interaction AB is positive, the response will be augmented Alternative Approach (equivalent)
if both A and B are increased.
1
AB B( A ) B( A )
When the interaction AB is negative, the response will be augmented 2
if A or B is increased and another is decreased. 1
ab b a (1)
2n
Only the magnitude (not the sign) of AB is important. Interaction AB is mutual and equivalent to factors A and B.

Diagonal method (see Figure 6.1):

The expression inside the bracket of (6.4) is equal to the difference
between the total of observations at the first diagonal and the total of
observations at the second diagonal.
The first diagonal includes the run of (1) and ab (for interaction AB)
6-27 6-28
For example 6.2 (without interaction) When interaction is large, the main effects have little meaning.
In Example 6.3 (with interaction)
1
AB ab b a (1)
2n Table 6.9
1 Factor A Factor B (temperature)
3 1 4 2 0
2 1 (machine)
B- (100 C) B+ (200 C)
For example 6.3 (with interaction) A- (USA) (1) 3 b 2
A+ (Japan) a 1 ab 4
1
AB ab b a (1)
2n (1) = 3, a = 1, b = 2, ab = 4
1
4 2 1 3 2
2 1

6-29 6-30

From (6.2), the main effect of factor A is Guidelines for Handling the Interaction
1
A [ a ab b (1)] (1) First calculate the interaction.
2n
1 (2) If the interaction is equal or close to zero, the effect of each
1 4 2 3 0
2 1 factor can be studied by simply calculating its main effect.
Can we conclude that factor A (machine) has no effect ??? NO!
(3) Otherwise, the effect of a factor should be studied separately at
o different levels of another factor.
In fact, at the low temperature (10 C), USA machine is better
A (B-) = 1 - 3 = -2
At the high temperature (30oC), Japanese machine is better
A (B+) = 4 - 2 = 2
That is, factor A (machine) does have significant effect on Cp. But,
this effect differs when temperature is different.
6-31 6-32
 In (6.5) to (6.7)
1 CA
A (1) a b ab (6.5)
2n 2n CA (1) a b ab
1 CB CB (1) a b ab
B (1) a b ab (6.6)
2n 2n
1 C AB C AB (1) a b ab
AB (1) a b ab (6.7)
2n 2n
CA, CB and CA are contrasts, because each of them is a linear
Evaluating effects is a simple way to estimate the influence of the combination of run totals and the sum of the coefficients is equal to
factors and interactions on the response. But, it sometime fails to give zero.
a clear-cut conclusion and cannot provide a confidence level for the
conclusion. Therefore, more complicated ANOVA may have to be
carried out.

6-33 6-34

The contrast of a main effect or an interaction can be easily obtained (1) The signs under column A or B are determined by the run
from the plus and minus sign Table. notation.

Table 6.10 Plus and Minus Signs Table for the 22 Design For example, for run (1), both factors A and B are at the low level,
so, both columns A and B have a minus sign.
Factor
For run (b), factor A is at the low level and column A has a minus
Run I A B AB sign; factor B is at the high level and column B has a plus sign.
(1) + - - +
(2) The signs under column AB are decided by the products of the
a + + - - signs under columns A and B.
b + - + -
ab + + + +

6-35 6-36
(3) The contrast of a factor or an interaction can be obtained by The Sum of Squares for the Contrast
summing up the products of each run and the sign under the 2
R
corresponding column.
ci yi.
i 1 C2
CA = (-)(1) + (+)(a) + (-)(b) + (+)(ab) SS = R
, R 2k (6.12)
2 Q
= -(1) + a - b + ab (6.8) n ci
i 1

CB = (-)(1) + (-)(a) + (+)(b) + (+)(ab) k

For two-factor case, k = 2. The coefficients ci of the contrasts in a 2
2
= -(1) - a + b + ab (6.9) design is equal to either 1 or (-1), therefore, (ci) 1. So,
CAB = (+)(1) + (-)(a) + (-)(b) + (+)(ab) R R
Q n ci2 n 1 nR N (6.13)
= (1) - a - b + ab (6.10) i 1 i 1

I is the grand total (D.O.F. of the contrast ) 1

I = + (1) + a + b + ab (6.11)
6-37 6-38

(6.12) can be simplified as: Example 6.4 A 22 Design (n = 4)

A: Spindle speed (A- = 1000 rpm, A+ = 3500 rpm)
C2 C2
SS (6.14) B: Feed rate (B- = 0.1 m/s, B+ = 0.15 m/s)
N 4n
Response: Thickness of a component
C A2 [ (1) a b ab]2 Table 6.11
SS A
4n 4n Factor
2
CB [ (1) a b ab]2 Run A B AB Thickness ( m) Total Average
SS B (6.15)
4n 4n (1) - - + 14.037 14.165 13.972 13.907 56.081 14.020
2
C AB [(1) a b ab]2 a + - - 14.821 14.757 14.843 14.878 59.299 14.825
SS AB
4n 4n b - + - 13.880 13.860 14.032 13.914 55.686 13.922
ab + + + 14.888 14.921 14.415 14.932 59.156 14.789
(6.15) is based on the concept of contrast, and can be applied to the 2k
designs only. If any factor has more than two levels, general y… = 230.222
formulae should be used.
(1) = 56.081 a = 59.299 b = 55.686 ab = 59.156
6-39 6-40
Use contrast formulae: Use conventional formulae

a b n
C A2 [ (1) a b ab]2 2 y...2
SS A SS y ijk
4n 4n i 1 j 1k 1 abn
(6.16)
2 2 2 230.2222
56.081 59.299 55.686 59156
. 14.037 ... 14.932 3.0672
SS A 2.7956 2 2 4
4 4
2
56.081 59.299 55.686 59156
. SS E SST SS A SS B SS AB
SS B 0.0181 (6.17)
4 4 30672
. 2.7956 0.0181 0.0040 0.2495
2
56.081 59.299 55.686 59156
.
SS AB 0.0040
4 4

6-41 6-42

Table 6.12 Table of ANOVA Steps to Calculate the Sum of Squares for 2k Design
Source of Sum of Degrees of Mean square F0
variation squares freedom (1) Establish the plus and minus signs table.

A 2.7956 1 2.7956 134.40 (2) Decide the contrast (6.8), (6.9), (6.10).

B 0.0181 1 0.0181 0.87 (3) Calculate the sum of squares of the main effects and
interactions from the contrasts (6.15).
AB 0.0040 1 0.0040 0.19
(4) Calculate the total sum of squares by the conventional formula
Error 0.2495 12 0.0208 (6.16).

Total 3.0672 15 (5) Calculate the sum of squares of the error by subtraction (6.17).

(6) Complete the ANOVA.

[Link] = N – 1 = 15 [Link] = 2k(n – 1) = 4 (4 – 1) = 12
6-43 6-44
6.3 2k Design for k 3 Factors
bc abc
For example, 23 design
c ac
+
k = 3, 3 factors, each at two levels

C
8 runs: (1) a b c ab ac bc abc b ab
- (1)
3 main effects: A, B, C a +
3 two-factor interactions: AB, AC, BC - B
- A +
1 three-factor interaction: ABC
Figure 6.4 The 23 factorial design
6-45 6-46

23 Design Main effect of C

Main effect of A c ac bc abc (1) a b ab
C yC yC
4n 4n
a ab ac abc (1) b c bc (6.20)
A yA yA 1
4n 4n (1) a b ab c ac bc abc
(6.18) 4n
1
(1) a b ab c ac bc abc
4n

Main effect of B
b ab bc abc (1) a c ac
B yB yB
4n 4n
(6.19)
1
(1) a b ab c ac bc abc
4n

6-47 6-48
When calculating the effects, the denominator is the total number of Interaction AB
observations at the low (or high) level of a factor. It is always equal 1
to half of the total observations N in a 2k design. Because, for each AB = [ A (B ) - A (B-) ]
4n
factor, one half of the observations are taken at its high level and
another half at its low level. 1
[(ab abc bc b ) ( a ac (1) c)] (6.22)
4n
N 2k n 1
2k 1 n (6.21) [ab abc (1) c bc b a ac]
2 2 4n
c ac bc abc
k R=2k N denominator 2k-1n B low B high
2 4 4n 2n
3 8 8n 4n Figure 6.5
4 16 16n 8n
(1) a
5 32 32n 16n b ab
6-49 6-50

Interaction AC Interaction BC

1 1
AC = [ A (C ) - A (C-) ] BC = [ B (C ) - B (C-) ]
4n 4n
(6.23) (6.24)
1 1
[( ac abc bc c ) (a ab b (1))] [(bc abc c ac) (b ab (1) a )]
4n 4n

bc
C high abc
c ac

b ab
C low Figure 6.6
(1)
a

6-51 6-52
Interaction ABC (refer to Figure 6.6) Table 6.13 Plus and Minus Signs Table for the 23 design
Factor effect
Let T AB
Run I A B C AB AC BC ABC
ABC TC (1) + - - - + + + -
1 a + + - - - - + +
T(C ) T(C ) b + - + - - + - +
4n
ab + + + - + - - -
1 c + - - + + - - +
AB(C ) AB(C )
4n ac + + - + - + - -
1 bc + - + + - - + -
abc c) ( ac bc) ab (1) a b
4n abc + + + + + + + +
1 (6.25) A (1) a [ Standard Order ]
abc - bc - ac + c - ab + b + a - (1)
4n
B b ab
C c ac bc abc
6-53 6-54

Systematic way to write the standard run order: Table 6.14

Decimal run number Binary run number Run notation
(1) Convert the run number (from 0 to 7) from a decimal number into
a binary number. 0 000 (1)
1 001 a
(2) If the first digit (from right) is 1, a is present; if it is 0, a is absent. 2 010 b
3 011 ab
(3) If the second digit is 1, b is present; if it is 0, b is absent. 4 100 c
5 101 ac
(4) If the third digit is 1, c is present; if it is 0, c is absent. 6 110 bc
7 111 abc

Standard run order is required only for a few methods, for example
the Yates’ Method

6-55 6-56
(1) The signs under columns A, B and C are determined by the run General Formulae to Calculate effect and sum of squares from
notation. the contrast
(2) The signs under column AB are decided by the products of the
signs under columns A and B. The signs under columns AC and
BC can be decided similarly. Contrast
Effect = (6.26)
(3) The signs under column ABC can be decided by several different N /2
ways, e.g., the products of the signs under columns A, B and C; or
the products of the signs under column A and column BC. (Contrast) 2
SS = (6.27)
N
(4) The effect of a factor or an interaction can be obtained by
summing up the products of each run and the sign under the The total number of observations in a 2k design:
corresponding column. For example,
CAB = (+)(1) + (-)(a) + (-)(b) + (+)(ab) + (+)(c) + (-)(ac) + (-)(bc) N Rn 2k n (6.28)
+ (+)(abc)
= (1) - a - b + ab + c – ac – bc + abc

6-57 6-58

Example 6.5 Response: Surface roughness (n = 2) CA = -(1) + a - b + ab - c + ac - bc + abc

A: Spindle speed (A- = 1000 rpm, A+ = 3500 rpm) = -16 + 22 - 20 + 27 -21 + 23 - 18 + 30 = 27
B: Feed rate (B- = 0.1 m/s, B+ = 0.15 m/s)
N = 2kn = 23 2 = 16
C: Machine (C- = USA machine, C+ = UK machine)
Table 6.14
Factors Surface CA 27
A 3.375
Run A B C roughness Total N / 2 16 / 2
(1) - - - 9, 7 16
a + - - 10, 12 22 C A2 27 2
b - + - 9, 11 20 SS A = 45.563
N 16
ab + + - 12, 15 27
c - - + 11, 10 21
ac + - + 10, 13 23
bc - + + 10, 8 18
abc + + + 16, 14 30
6-59 6-60
Values of all main effects and interactions Table 6.15 Table of ANOVA
Source of Sum of Degrees of Mean F0
A = 3.375 variation squares freedom square

B = 1.625 A 45.563 1 45.563 18.69

C = 0.875 B 10.563 1 10.563 4.33

C 3.063 1 3.063 1.26
AB = 1.375
AB 7.563 1 7.563 3.10
AC = 0.125 AC 0.063 1 0.063 0.03

BC = -0.625 BC 1.563 1 1.563 0.64

ABC 5.063 1 5.063 2.08
ABC = 1.125
Error 19.500 8 2.438

The highlighted values are the significant effects or interactions. A, B Total 92.938 15
and AB are dominant.
6-61 6-62

Specify = 0.01, control limit F , 1, 2

F0.01,1,8 11.26 6.4 Project of 2k Design

Only factor A (spindle speed) is considered by ANOVA having Any 2k design can be projected into a smaller 2r design (k r), if one
significant influence on the response roughness or more of the original factors are negligible and can be dropped.

23 22 21
Usually, the conclusions from the values of effects are similar to that
from ANOVA.
Projecting reduces the number of factors, so that it can simplify
The calculation of effects is easier than ANOVA. However, the ANOVA and provide more insight into the remaining important
information from the effects only has relative or comparative factors.
meaning, and cannot give the confidence level as in ANOVA.

6-63 6-64
 C Use Example 6.5 again Response: Surface roughness (n = 2)
bc abc
A: Spindle speed (A- = 1000 rpm, A+ = 3500 rpm)
c ac B: Feed rate (B- = 0.1 m/s, B+ = 0.15 m/s)
C: Machine (C- = USA machine, C+ = UK machine)

B Table 6.16
Factors Surface Run after
b ab Run A B C roughness Total dropping C
(1) - - - 9, 7 16 (1)
(1)
A a + - - 10, 12 22 a
a
b - + - 9, 11 20 b
Figure 6.7
ab + + - 12, 15 27 ab
c - - + 11, 10 21 (1)
Projecting the 23 design (in terms of factors A, B and C) into a 22 ac + - + 10, 13 23 a
design (in terms of A and B), suppose factor C is negligible and can bc - + + 10, 8 18 b
be dropped. abc + + + 16, 14 30 ab
6-65 6-66

Projecting this 23 design (in terms of A, B and C) into a 22 design (in Further projecting the 22 design (in terms of A and B) into a 21 design
terms of A and B). Because, effect C is very minor. (in terms of A). Since, effect B is also relatively smaller than A.
Table 6.17
Run A B Surface Run after Table 6.18
roughness dropping B Run A Surface
(1) - - 9, 7, 11, 10 (1) roughness
a + - 10, 12, 10, 13 a (1) - 9, 7, 11, 10, 9, 11, 10, 8
b - + 9, 11, 10, 8 (1) a + 10, 12, 10, 13, 12, 15, 16, 14
ab + + 12, 15, 16, 14 a
(1) Start from the table for the 23 design. Erase the column C.
(2) In column Run, for every run notation including letter c, drop
letter c (i.e., change c to (1), ac to a, bc to b, abc to ab).
(3) Work out a table for the 22 design in terms of A and B.
(4) Merge the rows that have identical run notation.
6-67 6-68
(1) Start from the table for the 22 design. Erase the column B. 6.5 Single Replicate of 2k Design

(2) In column Run, for every run notation including letter b, drop Single replicate: n = 1
letter b (i.e., change b to (1), ab to a).
In single replicate design, the degrees of freedom for error is equal to
1
(3) Work out a table for the 2 design in terms of A. zero.

(4) Merge the rows that have identical run notation. DOFE = R (n - 1) = R (1 -1) = 0

SS E SS E
MS E
After projection, the number of factors and number of runs will be DOFE 0
reduced, but the total number of observations is always the same.

In projecting, no extra experiments are needed. Only the available MS F

observations have to be re-organized. F0 ANOVA cannot be carried out !!
MS E

6-69 6-70

For a 24 experiment: Interactions having three or more factors (p 3) are called high
order interactions. The magnitudes of the high order interactions
Total number of runs = 24 = 16
are usually negligible and their physical meaning is hard to explain.
Let p indicate the number of factors involved in an interaction.
The conclusions of an experiment are usually dominated by the main
Table 6.19 effects and 2-factor interactions.
p p-factor interaction list of interactions
1 1-factor interaction (main effect) A, B, C, D
2 2-factor interaction AB, AC, AD, BC, BD, CD
3 3-factor interaction ABC, ABD, ACD, BCD
4 4-factor interaction ABCD

Number of p-factor interactions in a 2k design can be calculated by

k!
C pk k p 1 k p 2 k (6.29)
p! k p !

6-71 6-72
Guideline for dealing with the single replicate experiments (3) A check into the high-order interaction effects is necessary,
because, occasionally one or two high order interactions may
(1) When the number of factors is larger than 3 or 4, the number have relatively high values. If such non-trivial high order
of runs (=2k) is very large. In order to minimize the number of interactions have been found, they should be excluded from
observations (=2kn), single replicate design is widely used. the sums in (6.30) and (6.31).

(2) Under this circumstance, in order to carry out the ANOVA, a

common practice is to pool the higher order interactions as an
estimate of error.

SSE = SSABC + SSABD + SSACD + SSBCD + SSABCD (6.30)

DOFE = DOFABC+DOFABD+DOFACD+DOFBCD+DOFABCD
(6.31)
Then, ANOVA can be carried out as usual.

6-73 6-74

Example of Single Replicate 24 Experiment Table 6.20

Run A B C D Observation
A: Gap dimension (A- = 0.01 m, A+ = 0.02 m)
(1) - - - - 550
a + - - - 669
B: Pressure (B- = 300 mpa, B+ = 700 mpa) b - + - - 604
ab + + - - 650
C: Flow rate (C- = 1.2 m2/min, C- = 3.2 m2/min) c - - + - 633
ac + - + - 642
bc - + + - 601
D: Power (D- = 1000 kw, D+ = 3500 kw) abc + + + - 635
d - - - + 1037
Response: Etch rate (Å / min) ad + - - + 749
bd - + - + 1052
abd + + - + 868
cd - - + + 1075
acd + - + + 860
bcd - + + + 1063
abcd + + + + 729

6-75 6-76
CA = -(1) + a - b + + abcd Main and 2-factor interactions

= -550 + 669 - 604 + + 729 = -813 A = -101.625

B = -1.625
N 2k n 2 4 1 16
AB = -7.875
CA 813 C = 7.375
A 101.625
N /2 16 / 2
AC = -24.875
2
C A2 813 BC = -43.875
SS A 41311
N 16 D = 306.125
AD = -153. 625
BD = -0.625
CD = -2.125

6-77 6-78

High order interactions (D.O.F. = 1. for all of them): Table 6.21 ANOVA Table
Source of Sum of Degrees of Mean F0
ABC = -15.625 SSABC = 977 variation squares freedom square
ABD = 4.125 SSABD = 68 A 41311 1 41311 20.28
B 11 1 11 0.01
ACD = 5.625 SSACD = 127 C 218 1 218 0.11
BCD = -25.375 SSBCD = 2576 D 374850 1 374850 183.99
AB 248 1 248 0.12
ABCD = -40.125 SSABCD = 6440 AC 2475 1 2475 1.21
AD 94403 1 94403 46.34
All of the high order interactions are small compared to the most BC 7700 1 7700 3.78
dominant main effects A and D, and 2-factor interaction AD. BD 2 1 2 0.00
CD 18 1 18 0.01
SSE = 977 + 68 + 127 + 2576 + 6440 = 10187 Error 10187 5 2037
Total 531421 15
DOFE = 1+ 1 + 1 + 1 + 1 = 5

6-79 6-80
6.6 Center Points in 2k Design AC is the center point between A- and A+:
Assumption in 2k design: The response y is a linear function of the AC = 0.5 [(A-) + (A+)]
factor, say A. The curve y versus A is a straight line (no-curvature)
yC is the observed response value at point AC.
which can be completely decided by two points (y-, A-) and (y+, A+).
Therefore, the 2k design is adequate. yF is calculated by
y y
y yc yF
y+ 2
yF (1) If the distance yC yF is close to zero, the response function is
y- likely to be a linear function of the factor, and the 2k design is
adequate.

Figure 6.8 (2) If this distance is great, the response function is likely to be a
nonlinear function of the factor, and the 2k design is inadequate.
A- AC A+ More levels for the factor may have to be investigated
6-81 6-82

Example 6.6 Response: yield per hour nc number of observations at the center point.

nF = 2kn, number of observations in the ordinary factorial design.

160 +1 40.0 41.5
yc the average of the nc observations in the center point.
40.3
B 40.5
0 40.7 yF the average of the nF observations in the ordinary factorial
temperature 155
40.2 design.
40.6
150 -1 39.3 40.9
0
-1 +1

Figure 6.9 30 35 40
A reaction time
6-83 6-84
Hypothesis: SS curvature
MS curvature 1 SS curvature
k F0 (6.33)
H 0: yF yc 0 response is linear, non-curvature, 2 design is MS E MS E MS E
adequate.
H 1: yF yc 0 response is nonlinear, curvature, 2k design is 2
If F0 is large ----> SScurvature is large ----> yF yc is large --->
inadequate
yF yc is large ---> response is nonlinear, 2k design is inadequate

Sum of square
nF nc 2 2
SS curvature yF yc R yF yc (6.32)
nF nc

R is a constant.
D.O.F. = 1
6-85 6-86

2
See Example 6.6 SS curvature R yF yc
20 2
nc = 5, nF = 4 40.425 40.460 0.0027
9
40.3 40.5 40.7 40.2 40.6
yc 40.460 SS curvature 0.0027
5 F0 0.06
MS E 0.043
39.3 40.9 40.0 415
.
yF 40.425
4

nF nc (4)(5) 20
R
nF nc 4 5 9

6-87 6-88
 Table 6.22 Table of ANOVA Specify = 0.01, the control limit F ,1,4 = 21.2.
Source of Sum of Degrees of Mean F0
variation squares freedom square Since (for curvature) F0 < F ,1,4

A (time) 2.4025 1 2.4025 55.87

Conclusion: H0 is true. Response is linear, 2k design is adequate.
B (temp.) 0.4225 1 0.4225 9.83
AB 0.0025 1 0.0025 0.06
Curvature 0.0027 1 0.0027 0.06 In another example, if F0 > F ,1,4 (for curvature), H1 is true.
Response is nonlinear. Under such situation, we may have to carry
Error 0.1720 4 0.0430 out the more complicated experiment, in which each factor has more
Total 3.0022 8 than 2 levels.

[Link] = N – 1 = (nc + nF) – 1 = (5 + 4) – 1 = 8

[Link] = nc – 1 = 5 – 1 = 4
6-89 6-90

Another Application of the Center Points For example 6.6

If single replicate is used (as in example 6.6) in a 2k design, we may 2 2
have to pool the high order interactions to estimate the error. 40.3 40.46 ... 40.6 40.46
MS E 0.043
However, in example 6.6, there are only main effects A, B and 2- 5 1
factor interaction AB.
D.O.F .E 5 1 4
Under such circumstance, we may use the nc center observations to
estimate MSE for the error. MSE is actually the sample variance 2 in
probability and statistics. Two usages of the center points
nc
2 (1) Check if the response function is linear, i.e., if the 2k design is
yi yc
2 i 1
adequate?
MS E (6.34)
nc 1 (2) Estimate the errors.
And D.O.F .E D.O.F . 2 nc 1 (6.35)

6-91 6-92
6.7 The Advantages of Factorial Designs For experiments with same degree of accuracy,
(1) They are more efficient than one-factor-at-a-time experiments. N one factor at a time k 1 2k n / 2 k 1
R (6.36)
N factorial 2k
2k n 2
Table 6.23
B- B+ Where, 2kn/2 is the number of pairs for each factor. It is a
measure of the accuracy of experiments.
A- A-B- A-B+
A+ A+B- The relative efficiency increases along with the increase of the
number k of factors.

6-93 6-94

(2) Factorial design is necessary to avoid misleading conclusions

when interactions may be present.

(3) Factorial design allows the effect of a factor to be estimated at

several levels of the other factors, yielding conclusions that are
valid over a range of experimental conditions.

6-95
7 FRACTIONAL EXPERIMENTS Fractional factorial design allows us to conduct only a fraction (1/2,
1/4, 1/8, and so on) of the complete set of 2k runs in order to save the
7.1 Introduction number of runs.
A full 2k design can provide information of all the main effects and
interactions. For a 25 design: However, when using fractional design, information for some main
effects or interactions may be lost.
5 main effects A, B, C, D, E
Good fractional design:
10 two-factor interactions.
(1) Important information for main effects is reserved;
16 high-order interactions, which are usually negligible and little (2) Insignificant information for high order interactions is lost.
meaningful.
Poor fractional design:
But full 2k design requires a large number of runs, which may be
impossible to carry out due to the limits in time and resource. For (1) Important information for main effects is lost;
example, a full 25 design requires 52 = 32 runs. Even if n = 1, we (2) Insignificant information for high order interactions is
have to take 32 observations. reserved.
7-1 7-2

7.2 One Half Fraction A 23 design

A one-half fraction of a 2k design is called a 2k-1 fractional factorial A: Spindle speed (A- = 1000 rpm, A+ = 3500 rpm)
design. It reduces the number of runs by half, from 2k to 2k-1.
B: Feed rate (B- = 0.1 m/s, B+ = 0.15 m/s)
2k k 1
2 (7.1)
2 C: Machine (C- = USA machine, C+ = UK machine)
Table 7.1
k Full design Original One-half Reduced y: Response: Surface roughness (n = 1)
number fractional number
of runs design of runs
3 23 8 23-1 4
4 24 16 24-1 8
5 25 32 25-1 16
In fractional design, single replicate (n = 1) is usually used.
7-3 7-4
 Table 7.2 Plus and minus signs for the full 23 design Now, a one-half 23-1 fractional design is to be constructed from the
full 23 design. The key problem is how to select the 4 runs from the
Factor effect y original 8 runs.
Run I A B C AB AC BC ABC
a + + - - - - + + 16 If we select the top four runs that have a plus sign under column
b + - + - - + - + 22 ABC, the information for the highest interaction ABC will be lost.
c + - - + + - - + 20
abc + + + + + + + + 27 It is what we prefer to. Because, by sacrificing ABC, we may
ab + + + - + - - - 21 hopefully preserve the information for the more important main
ac + + - + - + - - 23 effects.
bc + - + + - - + - 18
(1) + - - - + + + - 30
Dividing the contrast from column A by N/2
a b c abc ab ac bc (1)
A (7.2)
N /2
7-5 7-6

3-1
Table 7.3 Plus and minus signs for the fractional 23-1 design In the 2 design, we may be tempted to calculate the main effects
and interactions by (7.2) as for the full 2k design, i.e., “Dividing the
Factor effect y contrast from a column by N/2”.
Run I A B C AB AC BC ABC
The main effects.
a + + - - - - + + 16
b + - + - - + - + 22 A a b c abc / 2 0.5
c + - - + + - - + 20 B a b c abc / 2 6.5 (7.3)
abc + + + + + + + + 27
C a b c abc / 2 4.5
Here, ABC is called the generator. In the selected fraction, the signs
The two-factor interactions
are identical under the column of the generator.
BC a b c abc / 2 0.5
In the selected fraction: I ˆ ABC (equal by sign)
AC a b c abc / 2 6.5 (7.4)
Equation I ˆ ABC is called the defining relation. AB a b c abc / 2 4.5
7-7 7-8
Unfortunately, neither (7.3) nor (7.4) is correct. The formulation Applying (7.2) (dividing the contrast from a column by N/2) to a full
(7.2) used by the full 2k design cannot be applied to fractional design. 2k design (Table 7.2), we obtain the corresponding effect, say
From (7.3) and (7.4), it is found:
a b c abc ab ac bc (1)
A
A ˆ BC , B ˆ AC , C ˆ AB (7.5) N /2
Aliases: A and BC B and AC C and AB
Applying (7.2) to a fractional 2k-1 design (Table 7.3), we obtain the
Equal by Sign corresponding joint effect, say lA.

Example: A ˆ BCD a b c abc a b c abc

lA = (7.6)
Meaning: The signs in columns A and BDC are the same, in N /2 2
the plus and minus signs table for the fractional
design. It does not mean that the value of the main lA is neither the pure effect of A nor the pure effect of BC. Instead, lA
effect A is equal to the value of the interaction BCD. is the joint effect of A and BC. More generally speaking, lA is the
joint effect of A and all of its aliases

7-9 7-10

A: Spindle speed (rpm) Go back to Table 7.3

B: Feed rate ( m/s) lA = A + BC = a b c abc / 2 (7.7)

y: yield (kg) lB = B + AC = a b c abc / 2 (7.8)

Table 7.4 lC = C + AB = a b c abc / 2 (7.9)

No A B y
In fractional design, there is no way to calculate the pure value of the
1 1000 0.1 2000
main effects, such as A, B, C. What we can calculate are just the joint
2 3500 0.15 3200
effects lA, lB, lA.
The difference of 1200kg in the yield indicates the joint effect of A In fractional design, we usually use the joint effect to estimate the
(spindle speed) and B (feed rate) main effect, say
A B
A lA, B lB, C lC (7.10)
A + B = 3200 – 2000 = 1200 600 600
1000 200 A good fractional design will ensure that the error of the estimation is
A 1200, B 1200 1200 0 very small.
7-11 7-12
Rules for interaction The aliases of a factor can be found from the defining relation.

(1) IQ ˆ Q (7.11) Defining relation I ˆ ABC

(2) QQ ˆ Q 2 ˆ I (7.12) A ˆ AI ˆ A ABC ˆ A2 BC ˆ BC

They can be easily verified from Table 7.2 B ˆ BI ˆ B ABC ˆ AB 2C ˆ AC

Q is any effect or interaction
C ˆ CI ˆ C ABC ˆ ABC 2 ˆ AB
Example
They are exactly the same as (7.5)
IA ˆ A IAB ˆ AB A2 ˆ I

AB ABC ˆ A2 B 2C ˆ IIC ˆ IC ˆ C

7-13 7-14

Example 7.1 A 24 design Table 7.5 (1) a b abcd 2 99 80 178

Run A B C D AB ABCD y A 99.875
8 8
(1) - - - - + + 2
a + - - - - - 99 (7.13)
b - + - - - - 80
ab + + - - + + 181
c - - + - + - -2
ac + - + - - + 98
bc - + + - - + 82
abc + + + - + - 179
d - - - + + - -1
ad + - - + - + 102
bd - + - + - + 81
abd + + - + + - 182
cd - - + + + + 1
acd + - + + - - 102
bcd - + + + - - 79
abcd + + + + + + 178

7-15 7-16
(1) Good fractional design: ABCD is used as the generator. Defining relation is: I ˆ ABCD

Table 7.6 A ˆ AI ˆ A ABCD ˆ A2 BCD ˆ BCD

Run A B C D ABCD y B ˆ BI ˆ B ABCD ˆ AB 2CD ˆ ACD

(1) - - - - + 2
ab + + - - + 181 C ˆ CI ˆ C ABCD ˆ ABC 2 D ˆ ABD
ac + - + - + 98
bc - + + - + 82
D ˆ DI ˆ D ABCD ˆ ABCD 2 ˆ ABC
ad + - - + + 102
bd - + - + + 81
cd - - + + + 1
abcd + + + + + 178 All main effects are aliased with high-order interactions.

From Table 7.5, select all the runs with a plus sign under column
ABCD

7-17 7-18

Since we are quite sure that the high order interactions are negligible Similarly
( 0), so
A lA = A + BCD
lA = A + BCD A+0=A
B lB = B + ACD
From Table 7.6
C lC = C + ABD
A A D lD = D + ABC (7.15)
(1) ab ac bc ad bd cd abcd
(7.14)
4
2 181 98 82 102 81 1 178
98.25
4

The estimated A resulting from (7.14) is very close to the accurate A

resulting from (7.13)

7-19 7-20
(2) Poor fractional design: AB is used as the generator Defining relation is: I ˆ AB

Table 7.7 A ˆ AI ˆ A AB ˆ A2 B ˆ B

Run A B C D AB y B ˆ BI ˆ B AB ˆ AB 2 ˆ A
(1) - - - - + 2
ab + + - - + 181 C ˆ CI ˆ C AB ˆ ABC
c - - + - + -2
abc + + + - + 179
d - - - + + -1
D ˆ DI ˆ D AB ˆ ABD
abd + + - + + 182
cd - - + + + 1
abcd + + + + + 178 Some main effects (i.e., A and B) are aliased with each other.

From Table 7.5, select all the runs with a plus sign under column AB

7-21 7-22

Since we are not sure that a particular main effect is always Summary
negligible, we may not be able to use
(1) If the main effects are aliased with high order interactions, we can
A lA = A + B use
From Table 7.7
A lA, B lB, C lC
(1) ab c abc d abd cd abcd
A to get very accurate estimates of the main effects. The higher the
4 order of the interactions, the more likely they are close to zero,
(7.16)
2 181 ( 2) 179 ( 1) 182 1 178 and the more accurate the estimates are.
180
4
(2) The order of the interactions aliased with the main effects in a
lA resulting from (7.16) is quite different from the accurate A fractional design is completely determined by the generator. So,
resulting from (7.13) to develop a good fractional design, the key problem is to select
the generator.

7-23 7-24
(3) For one-half fractional design, we always select the highest order A 2k-1 design can be constructed by first working out a full factorial
interaction as the generator. It guarantees the best 2k-1 design. design (the Basic Design) in terms of (k-1) basic factors, and then
adding the last factor.
Table 7.8
The number of basic factors is (3 – 1) = 2. The two basic factors are
Fractional design Generator A and B.
For example, for a 23-1 design, defining relation: I ˆ ABC
23-1 ABC
The basic design is a full 22 design in terms of A and B:
4-1
2 ABCD
Table 7.9
5-1 Run A B AB
2 ABCDE
(1) - - +
26-1 ABCDEF a + - -
b - + -
ab + + +
7-25 7-26

C ˆ CI ˆ C ABC ˆ ABC 2 ˆ AB Design of a 2k-1 One-Half Fractional Experiment

Replace AB by its alias C, we obtain the required 23-1 design. (1) Decide the (k-1) basic factors.

Table 7.10 (2) Work out the basic design in terms of the basic factors.
Run A B C y
c - - + 3 (3) Decide the alias of the last factor from the generator, which is the
a + - - 4 highest order interaction.
b - + - 2
abc + + + 1 (4) Replace the last factor for the alias in the basic design. The table
for the 2k-1 design is completed.
This is the same 23-1 design as given in Table 7.3
(5) The entire aliases structure of the 2k-1 design can be determined
from the defining relation.

7-27 7-28
(6) Evaluate the joint effects from the plus and minus signs table of Example: Conduct a 24-1 design with factors A, B, C, D (k = 4)
the 2k-1 design. Response is the number of defects per lot
(7) Since the high-order interactions are usually negligible, these joint (1) Decide the 3 basic factors: A, B, C.
effects can be used as the estimates of the main effects and the (2) Work out the basic design in terms of the basic factors.
two-factor interactions. Table 7.11
Runs for 23 A B C ABC
(1) - - - -
a + - - +
b - + - +
ab + + - -
c - - + +
ac + - + -
bc - + + -
abc + + + +

7-29 7-30

(3) The generator is the highest order interaction ABCD, the defining (4) Replace D for ABC in the basic design, the table for the 24-1
relation is I ˆ ABCD design is completed.

D ˆ D I ˆ D ABCD ˆ ABCD 2 ˆ ABC Table 7.12

Runs No of
for 24-1 A B C D ˆ ABC AB defects
(1) - - - - + 550
ad + - - + - 749
bd - + - + - 1052
ab + + - - + 650
cd - - + + + 1075
ac + - + - - 642
bc - + + - - 601
abcd + + + + + 729

7-31 7-32
(5) The entire aliase structure of the 24-1 design can be determined (6) Evaluate the joint effects. For examples
from the defining relation I ˆ ABCD
A A BCD
Examples 1
(1) ad bd ab cd ac bc abcd
2 4
A ˆ A I ˆ A ABCD ˆ A BCD ˆ BCD
1
AB ˆ AB I ˆ AB ABCD ˆ A2 B 2CD ˆ CD ( 550 749 1052 650 1075 642 601 729)
4
ABC ˆ ABC I ˆ ABC ABCD ˆ A2 B 2C 2 D ˆ D 127.00

Entire aliases structure B B ACD 4.00

A ˆ BCD , B ˆ ACD , C ˆ ABD , D ˆ ABC , C C ABD 11.50
D D ABC 290.51
AB ˆ CD , AC ˆ BD , AD ˆ BC ,

7-33 7-34

AB AB CD (7) Use the joint effects to estimate the main effects and the two-
1 factor interactions.
(1) ad bd ab cd ac bc abcd
4
Since all of the main effects are aliased with the high order
1
( 550 749 1052 650 1075 642 601 729) interactions which are usually negligible, we have
4
10.00 A lA, B lB, C lC, D lD

AC AC BD 25.50 However, 2-factor interactions may not be negligible, usually,

AD AD BC 197.50
AB AB and CD AB

We can only assert that the joint effect of AB and CD is (-10)

7-35 7-36
Project of the 2k-1 Design ab
B b
After a one-half fractional 2k-1 experiment has been completed, if one
factor and all of its interactions are found negligible, this factor can
be dropped. That is, the design will be projected into a full factorial b (1) a
b
design in terms of the remaining (k - 1) factors.
bc abc
A
(1)
a
c c
(1) a Figure 7.1
C c ac

7-37 7-38

A 23-1 design A A a b c abc / 2 0.5

A: Spindle speed (A- = 1000 rpm, A+ = 3500 rpm) B B a b c abc / 2 6.5
C C a b c abc / 2 4.5
B: Feed rate (B- = 0.1 m/s, B+ = 0.15 m/s)
C: Machine (C- = USA machine, C+ = UK machine) Since A is much smaller than B and C, we can drop it and project the
y: Response: Surface roughness (n = 1) 23-1 fractional design into a full 22 design in terms of B and C.

Table 7.13 (1) Erase the column A.

Run A B C y
(2) In column Run, for every run notation including letter a, drop
letter a (i.e., change a to (1), abc to bc).
a + - - 16
b - + - 22
c - - + 20
abc + + + 27

7-39 7-40
 Table 7.14 Design Resolution
A design is resolution R if q-factor effect is only aliased with the
Run B C y
effects containing (R – q) or more factors (Note: 1-factor effect is the
(1) - - 16 main effect).
b + - 22 In a one-half design, R is equal to the number k of the factors.
c - + 20
bc + + 27 Designs of resolution III: 2III3-1 with I ˆ ABC
Designs of resolution V: 2V5-1 with I ˆ ABCDE

Resolution R is an indicator of the effectiveness (goodness) of the

fractional design. The greater the R, the better the fractional design. If
R is greater, the main effects will be aliased with higher order
interactions. Since the higher order interactions tend to be more
negligible, the estimate of the main effects from the joint effects (e.g.,
A lA) will be more accurate.
7-41 7-42

7.3 Smaller Fraction (e.g., 1/4 fraction, 1/8 fraction) An experiment with six factors:
A 2k design may run in a 1/2p fraction called a 2k-p fractional factorial
Full design: 26 64 runs
design.

k p 2k 1/4 fraction: 26-2 16 runs

2 (7.17)
2p
1/8 fraction: 26-3 8 runs
k k
1 2 2
p = 1, 1/2 Fraction 2k
21 2
The larger the p value, the smaller the fraction design, the more
k k
2 2 2 number of runs can be saved, but, more information will be lost.
p = 2, 1/4 Fraction 2k
22 4
3 2k 2k
p = 3, 1/8 Fraction 2k
23 8

7-43 7-44
26-2 Design (p = 2, 1/4 Fraction) In general 2k-p fractional design,

Number of runs can be reduced from 64 to 16. (1) Number of basic generators = p

(1) Two basic generators: ABCE, BCDF (2) Number of induced generators = 2p - 1 - p
I ˆ ABCE I ˆ BCDF (7.18)
(3) Total number of generators = p + (2p - 1 - p) = 2p - 1
(2) One induced generator: ADEF
(4) Each generator corresponds to an alias, so the number of the
ABCE ( BCDF ) ˆ AB 2C 2 DEF ˆ ADEF aliases for each effect is equal to the total number of
ABCE ( BCDF ) ˆ I I ˆ I generators, i.e., 2p - 1

I ˆ ADEF (7.19)

(3) Total number of generators is 3 (for one-quarter fraction)

7-45 7-46

In the 26-2 fractional design, the total number of generators is three, Table 7.15 The Complete Alias Structure for the 26-2 Design:
so each effect has 3 aliases, each of them can be obtained from a
particular generator. For example, for main effect A A ˆ BCE ˆ DEF ˆ ABCDF AB ˆ CE ˆ ACDF ˆ BDEF

B ˆ ACE ˆ CDF ˆ ABDEF AC ˆ BE ˆ ABDF ˆ CDEF

From the first generator ABCE A ˆ AI ˆ A( ABCE ) ˆ BCE C ˆ ABE ˆ BDF ˆ ACDEF AD ˆ EF ˆ BCDE ˆ ABCF
From the second generator BCDF A ˆ AI ˆ A( BCDF ) ˆ ABCDF D ˆ BCF ˆ AEF ˆ ABCDE AE ˆ BC ˆ DF ˆ ABCDEF

From the third generator ADEF A ˆ AI ˆ A( ADEF ) ˆ DEF E ˆ ABC ˆ ADF ˆ BCDEF AF ˆ DE ˆ BCEF ˆ ABCD

F ˆ BCD ˆ ADE ˆ ABCEF BD ˆ CF ˆ ACDF ˆ ABEF

Number of aliases for an effect = number of generators ABD ˆ CDE ˆ ACF ˆ BEF BF ˆ CD ˆ ACEF ˆ ABDE

ACD ˆ BDE ˆ ABF ˆ CEF

7-47 7-48
(1) The main effects are aliased with three-factor or higher order CREATION OF A 26-2 DESIGN
interactions.
Six factors: A, B, C, D, E, F
A = A + BCE + DEF + ABCDF A+0+0+0=A
(1) Decide the basic generators ABCE and BCDF, and work out
(2) The two-factor interactions are aliased with each other. the induced generator ADEF.

(2) From the three generators, work out the alias structure table –
Even though only one quarter of the original runs have been carried Table 7.15.
out, this design would provide very good estimates on the main
effects and give some idea about the two-factor interactions. (3) Decide (6-2) = 4 basic factors A, B, C, D, and work out the
basic design, i.e., the full 24 design in terms of A, B, C, D (for
general 2k-p design, the number of basic factors is equal to
(k - p)).

7-49 7-50

Table 7.16 Basic design. (4) From Table 7.15,

No Run 24 A B C D E ˆ ABC F ˆ BCD
E ˆ ABC ˆ ADF ˆ BCDEF
1 (1) - - - - - -
2 a + - - - + - The interaction ABC is composed of basic factors only. So, the
3 b - + - - + + signs of this interaction can be determined from the signs of A,
4 ab + + - - - + B and C in Table 7.16. Then, replace ABC by E, the signs for
5 c - - + - + +
6 ac + - + - - +
E is determined.
7 bc - + + - - -
(5) Similarly, from Table 7.15,
8 abc + + + - + -
9 d - - - + - + F ˆ BCD ˆ ADE ˆ ABCEF
10 ad + - - + + +
11 bd - + - + + - The interaction BCD is composed of basic factors only. So,
12 abd + + - + - - the signs of this interaction can be determined from the signs
13 cd - - + + + - of B, C and D in Table 7.16. Then, replace BCD by F, the
14 acd + - + + - -
signs for F is determined.
15 bcd - + + + - +
16 abcd + + + + + + (6) Carry out the experiment and record all of the observations.
7-51 7-52
 Table 7.17 26-2 design (7) The joint effects can be evaluated from the plus and minus
No Run 26-2 A B C D E F No of defects signs table. For example,
1 (1) - - - - - - 6
2 ae + - - - + - 10 A = A + BCE + DEF + ABCDF
3 bef - + - - + + 32 1
4 abf + + - - - + 60 = ( -(1) + ae – bef + abf – cef + + abcdef )
5 cef - - + - + + 4 8
6 acf + - + - - + 15 1
7 bc - + + - - - 26 = (-6 + 10 - 32 + 60 - 4 + +52)
8 abce + + + - + - 60
8
9 df - - - + - + 8 = 13.88
10 adef + - - + + + 12
11 bde - + - + + - 34
12 abd + + - + - - 60
(8) Since the high-order interactions are thought negligible, these
13 cde - - + + + - 16 joint effects can be used as the estimates of the main effects.
14 acd + - + + - - 5 For example,
15 bcdf - + + + - + 37
16 abcdef + + + + + + 52 A = 13.88
A
7-53 7-54

Table 7.18 Selected 2k-p fractional factorial designs Advantages of fractional designs

Input Output 1. It can significantly reduce the number of runs

Number of Number of Fraction Basic
factors k Runs design generators 2. It can eliminate the negligible factors by projecting a fractional
3 4 2III3-1 ABC factorial design into a full factorial design. It results in a stronger
4 8 2IV4-1 ABCD experiment in the active factors that remain.
5 16 2V5-1 ABCDE
8 2III5-2 ABD, ACE 3. By combining a sequence of small fractional factorial designs, it
6 32 2VI6-1 ABCDEF can isolate both the main effects and interactions, and also takes
16 2IV6-2 ABCE, BCDF the advantage of learning about the process as it goes along.
8 2III6-3 ABD, ACE, BCF

Fractional design based on this table is always the best one.

7-55 7-56
8 A CASE STUDY (1) Decide the factors to be investigated and the ranges of their
values
Problem:
At the early stage, it is desired to include as many factors as
possible. So, 2k experiment is used.
For the diameter (a key dimension) of a shaft, the current process
capability Cp is only 0.8 Table 8.1
No. Factors Low High
The QA engineers decide to conduct a designed experiment in hope 1 A feed rate (liters/min) 10 15
to increase Cp substantially (an application of Larger the better). 2 B carbon element (%) 1 2
3 C cooling rate (m3/min) 100 120
4 D machine Japan USA
5 E speed (m/min) 10 20
6 F temperature (o C) 15 25
The central point values over the ranges are the current values for
the factors.
8-1 8-2

(2) Decide a fractional factorial design (3) Decide the generators

Due to the limitations of time and resources, only 16 observations, Basic generators: ABCE and BCDF
at most, can be taken.
Table 8.2 Selected 2k-p fractional factorial designs Induced generator: ( ABCE )( BCDF ) ˆ ADEF
Input Output
Number of Number of Fraction Basic
factors k Runs design generators
3 4 2III3-1 ABC
4 8 2IV4-1 ABCD
5 16 2V5-1 ABCDE
8 2III5-2 ABD, ACE
6 32 2VI6-1 ABCDEF
16 2IV6-2 ABCE, BCDF
8 2III6-3 ABD, ACE, BCF
The 26-2 fractional design is selected, basic generators: ABCE and
BCDF
8-3 8-4
(4) determine the alias structure (5) Decide the basic factors and basic design (Table 8.3)
A ˆ AI ˆ A( ABCE ) ˆ BCE (6-2) = 4 basic factors: A, B, C, D. Basic design is a 24 design
A ˆ AI ˆ A( BCDF ) ˆ ABCDF 24 Run A B C D ABC BCD
(1) - - - - - -
A ˆ AI ˆ A( ADEF ) ˆ DEF a + - - - + -
b - + - - + +
A ˆ BCE ˆ DEF ˆ ABCDF ab + + - - - +
c - - + - + +
B ˆ ACE ˆ CDF ˆ ABDEF ac + - + - - +
bc - + + - - -
C ˆ ABE ˆ BDF ˆ ACDEF abc + + + - + -
d - - - + - +
D ˆ BCF ˆ AEF ˆ ABCDE
ad + - - + + +
E ˆ ABC ˆ ADF ˆ BCDEF bd - + - + + -
abd + + - + - -
F ˆ BCD ˆ ADE ˆ ABCEF cd - - + + + -
acd + - + + - -
All main effects are aliased with the high order interactions. bcd - + + + - +
abcd + + + + + +
8-5 8-6

(6) Decide the signs for columns E and F Table 8.4

Run order 26-2 Run A B C D E ˆ ABC F ˆ BCD Cp
From the alias structure, select 4 (1) - - - - - - 0.6
10 ae + - - - + - 1.0
E ˆ ABC , F ˆ BCD 14 bef - + - - + + 3.2
1 abf + + - - - + 6.0
The signs of ABC and BCD can be determined from that of the 13 cef - - + - + + 0.4
16 acf + - + - - + 1.5
basic factors. And then, columns ABC is replaced by E, and 7 bc - + + - - - 2.6
column BCD by F. 6 abce + + + - + - 6.0
2 df - - - + - + 0.8
9 adef + - - + + + 1.2
3 bde - + - + + - 3.4
15 abd + + - + - - 6.0
11 cde - - + + + - 1.6
5 acd + - + + - - 0.5
12 bcdf - + + + - + 3.7
8 abcdef + + + + + + 5.2

8-7 8-8
(7) Carry out the experiment and take the observations (response (8) Calculate the joint effects
C p)
1
Note, run order (the order of taking the observations) must be A (1) ae bef abf abcdef
8
random (see the first column).
1
0.6 1.0 3.2 6.0 5.2 1.39
8
B 3.56

C 0.09

D 0.14

E 0.04

F 0.04

8-9 8-10

(9) Estimate the main effects (10) Decide the effective factors and project the 26-2 design
From the alias structure, it is found that all of the main effects are Compare the estimated main effects, it is found that only factors
aliased only with high order interactions which are negligible. So, A and B are significant. So, the data of the 26-2 experiment can be
projected into a 22 design in terms of A and B.
A A 139
.
[1] In Table 8.4, erase columns C, D, E and F.
B B 356
.
[2] Drop c, d, e and f from all of the run notations, e.g.,
C C 0.09
cef --> (1)
D D 014
. abcdef --> ab.
E E 0.04 [3] In the 22 design in terms of A and B, merge all of the
F 0.04 observations under a same run notation.
F

8-11 8-12
 Table 8.5 (11) Calculate the totals of the runs
Run A B AB Cp For the remaining factors A and B, ANOVA is to be carried out.
(1) - - + 0.6, 0.4, 0.8, 1.6 (1) = 0.6, + 0.4 + 0.8 + 1.6 = 3.4
a + - - 1.0, 1.5, 1.2, 0.5 a = 1.0 + 1.5 + 1.2 + 0.5 = 4.2
b - + - 3.2, 2.6, 3.4, 3.7 b = 3.2 + 2.6 + 3.4 + 3.7 = 12.9
ab + + + 6.0, 6.0, 6.0, 5.2 ab = 6.0, + 6.0 + 6.0 + 5.2 = 23.2
(12) Calculate the contrasts
k = 2, a = 2, b = 2, n = 4, N = 16 CA = -(1) + a - b + ab = -3.4 + 4.2 - 12.9 + 23.2 = 11.1
CB = -(1) - a + b + ab = -3.4 - 4.2 + 12.9 + 23.2 = 28.5
CAB = (1) - a - b + ab = 3.4 - 4.2 - 12.9 + 23.2 = 9.5
y… = I = (1) + a + b + ab = 3.4 + 4.2 + 12.9 + 23.2 = 43.7
8-13 8-14

(13) Calculate the sum of squares (14) Conduct the analysis of variance

C A2 11.12 Table 8.6

SS A 7.7
N 16 Source of Sum of DOF Mean F0
variation squares square
CB2 28.5 2
SS B 50.8
N 16 A 7.7 1 7.7 37.0
2
C AB 9.52 B 50.8 1 50.8 243.8
SS AB 5.6
N 16
2 2 4 AB 5.6 1 5.6 26.9
2 y...2 2 2 2 43.72
SST y ijk 0.6 0.4 5.2 66.6
i 1 j 1 k 1 N 16 Error 2.5 12 0.208
SS E SS T SS A SS B SS AB 66.6 7.7 50.8 5.6 2.5
Total 66.6 15

8-15 8-16
(15) Decide control limit F , 1,
(17) Check the residuals
2

Specify type I error = 0.01. For main effects A, B and Residuals eij should satisfy following requirements:
interaction AB, 1 = 1, 2 = 12. So (1) has normal distribution
F , 1,
= 9.33 (2) has independent distribution
2

(3) has zero mean value

(16) Decide the significance of the main effects and interaction 2
(4) has constant variance for all runs
A: F0 = 37.0 > F ,
, factor A has significant effect.
1, 2

B: F0 = 243.8 > F ,
, factor B has significant effect.
1, 2

AB: F0 = 26.9 > F ,

, interaction AB has significant effect.
1, 2

8-17 8-18

(18) Checking curvature n F nc 2 16 5 2

SS curvature yF yc 2.73 2.70 0.0034
n F nc 16 5
Take five additional observations at the center point, i.e.,
Feed rate A: 12.5 liter/min SS curvature 0.0034
F0 0.0165
Carbon element B: 1.5 % MS E 0.208

The observed process capability: Specify = 0.01. Also, 1 = 1, 2 = 12. F , 1,

= 9.33
2

2.70, 2.74, 2.71, 2.68, 2.69

Since F0 < F , 1 , 2 , there is no obvious curvature, the response is
nc = 5, nF = 16 linear function of the factors, 22 design is adequate.
yc ( 2.70 2.74 2.71 2.68 2.69 ) / 5 2.70
yF y... / 16 437
. / 16 2.73

8-19 8-20
(19) Illustrate the interaction (20) Conclusions
A1B1 = (1) / 4 = 0.85, A2B1 = a / 4 = 1.05 [1] Factors C (cooling rate), D (machine), E (speed) and F
A1B2 = b / 4 = 3.23, A2B2 = ab / 4 = 5.80 (temperature) have negligible effect on the process capability.
[2] Both factor A (feed rate) and B (carbon element) have
6 significant, positive effect on the process capability.
B2
5 A2B2 [3] At the low level of B, factor A has little influence on the
4 process capability.
3 B2
A1B2 [4] If we set B at the high level (2%) and A also at the high level
(15 liter/min), the average process capability will be as high as
2 A2B1 5.80 on average (compared to the original value of 0.8).
A1B1 B1
1
B1 Figure 8.1
0
A1 A2
8-21 8-22
9 RESPONSE SURFACE METHODS (RSM)
9.1 Introduction x2 (B)
Optimization + Design of Experiments

Factors (Input variables) x 1 , x 2 , . . . , x k

2K
Response (output function) Y start point
*
The purpose of RSM is to optimize (maximize or minimize) Y.
optimum
CCD

2K
x1 (A)

9-1 9-2

Y f x1, x2 , , xk (9.1)

f is called the response surface function. It is an approximate of the

function relationship between the factors and the response Y.
represents the noise or error.

In RSM, the response surface function f is determined from the data

observed during the experiments, and using the least-squares method
(regression).

Figure 9.1
9-3 9-4
 Each contour line corresponds to a particular height of the response
surface. All the points (x1, x2) on a same contour line result in the
same value of the response.

After the response surface function f for Y has been established, we

can use it to study the relationship between the input
x 1 , x 2 , . . . , x k and the response Y. The goal is to optimize Y.

(1) Larger the Better: If Y is the production yield, Y should be

maximized.
(2) Smaller the Better: If Y is the production cost, Y should be
minimized.
(3) Nominal the Best: If Y is the diameter of a shaft, Y should be
Figure 9.2 made as close as possible to the nominal diameter.
(Factor x1: temperature Factor x2: pressure Response Y: yield)
9-5 9-6

Y = f(x) + f(x) (9.2) First-order model (when f is a linear function)

A low-order polynomial is usually employed to approximate the true Y f 0 x
1 1 2 2x k xk
(9.3)
functional relationship between Y and the set of independent Yˆ f x x
0 1 1 2 2 k xk
variables (factors) xi. The approximation is very accurate in a small
region. The response surface analysis is then carried out in terms of Second-order model (when f is a quadratic function)
the fitted polynomial. k k
2
Table 9.1 Y 0 i i x ii ix ij ix xj
i 1 i 1 i j
X1 X2 Y k k
(9.4)
2
80 170 76.5 Yˆ 0 i i x ii ix ij ix xj
i 1 i 1 i j
80 180 77.0
90 170 78.0 For k = 2
90 180 79.5
85 175 79.9 2 2
Y 0 1 x1 2 x2 11 x1 22 x 2 12 x1 x 2 (9.5)
85 175 80.3
85 175 80.0
9-7 9-8
RSM Is a Sequential Procedure
(1) Set up and conduct a series of experiments that will give
sufficient and reliable data on the process.
(2) Fit a response surface function f that describes the relationship
between the factors x1 , x 2 ,..., x k and the response Y.
(3) Use the response surface function to determine the optimal values
of the input variables that optimize Y.
[1] When the design point is remote from the optimum. Use the
first-order model to move the design point rapidly to the
vicinity of the optimum.
[2] Once the optimum region has been found, the second-order Figure 9.3 A contour plot of the yield response surface
model is used to locate the optimum point accurately.

9-9 9-10

9.2 First-Order Model (The Method of Steepest Ascent) Yˆ

x1 1
Y 0 1 x1 2 x2 k xk (9.6)
Gradient vector (9.8)
This method is used when the design point is remote from the Yˆ
k
optimum. The purpose is to move the design point rapidly along the xk
steepest ascent direction to the vicinity of the optimum.
The steepest ascent direction is used if the maximum response is For k =2,
expected (Larger the Better). This direction is indicated by the Yˆ
gradient vector. It starts from the center of the interest region and
x1 1
normal to the contour line. (9.9)
Yˆ 2

Yˆ x2
From (9.6) i (9.7)
xi
The steepest direction is completely determined by the signs of 1
and 2 and their ratio. The magnitudes of 1 and 2 are immaterial.
9-11 9-12
 If the minimum response is expected (Smaller the Better), the
steepest descent direction should be used. This direction is indicated
by the negative gradient vector.

Negative gradient vector

Yˆ
x1 1

(9.10)
Yˆ
k
xk

Figure 9.4 First order response surface and

path of steepest ascent
9-13 9-14

Example 9.1 Response: yield per hour Table 9.2

Factor min center max
41.5
160 +1 40.0 A: time coded var. x1 -1 0 +1
X 1 35 40 35 natural var. X1 30 35 40
B x1 0 1 0
temperature 155 0 B: temp. coded var. x2 -1 0 +1
X 1 5x1 35
1 natural var. X2 150 155 160
x1 X1 7
40.9 5
150 -1 39.3
0 x1 Coded variables are used for the factorial experiment, where (-1) and
-1 +1
(+1) correspond to low and high levels in the 2k design.
Figure 9.5 30 35 X1 40
Natural variables are used for actual experiment setting and final
A reaction time reports.

9-15 9-16
Conversion (From Fig 9.5): A 22 experiment (Figure 9.5) is used to obtain the data for fitting the
response surface function.
X 1 35 40 35
x1 0 1 0 Table 9.3
Natural Coded
X1 5 x1 35 dX 1 5dx1 (9.11) variables variables Response
X1 X2 x1 x2 y
1 1 30 150 (1) -1 -1 39.3
x1 X1 7 dx1 dX 1
5 5 30 160 b -1 1 40.0
40 150 a 1 -1 40.9
X 2 155 160 155 40 160 ab 1 1 41.5
x2 0 1 0
X2 5 x2 155 dX 2 5dx2 (9.12)
1 1
x2 X2 31 dx2 dX 2
5 5
9-17 9-18

In coded variable system (xi = 1)

Yˆ
Yˆ x x (9.17) x1 1 0.755
0 1 1 2 2
0.325
0 is the grand average of all observations Yˆ 2 0.325
is one-half of the A effect. x2 0.755
1

2 is one-half of the B effect. In order to make the movement along the steepest ascent direction,
Y = (41.5+40.9+40.0+39.3) / 4 = 40.425 we must have

A = (41.5 + 40.9)/2 – (40.0 + 39.3) / 2 = 1.55 A/2 = 0.775 x1 Yˆ / x1 1 0.755

(9.19)
B = (41.5 + 40.0)/2 – (40.9 + 39.3) / 2 = 0.65 B/2 = 0.325 x2 Yˆ / x2 2 0.325

Yˆ 40.425 0.775 x1 0.325 x2 (9.18) x1 step size (increment in a step) for x1 variable in coded system.
x2 step size (increment in a step) for x2 variable in coded system.
9-19 9-20
The step size is determined by the experimenter based on process Now, if the step size for A (time) is specified as: X1 = 5 minute
knowledge or other practical considerations.
From (9.11) and (9.12)
(1) One, and only one, variable has to be specified.
X1 5 x1 5 x1 0.755
1 2.323
(2) Specification is made in natural system. X2 5 x2 5 x2 0.325

(3) The step size of the specified variable must be suitable from a X1 5
X2 = 2.152 2.0
practical viewpoint. 2.323 2.323

(4) The step sizes of other variables are determined so that the move So, X1 = 5 minute, X2 = 2 oC are the desired step sizes in natural
will follow the steepest ascent direcgtion. variable system.

During the derivation of X2, we have made use of (9.19). It ensures

that the movement of the design point will follow the steepest ascent
direction.
9-21 9-22

If the step size for B (temperature) is specified as: X2 = 2 oC. Starting point (X1 = 35, X2 = 155), see Figure 9.5

Again from (9.11) and (9.12) Move the first step:

X1 5 x1 5 x1 0.755 X1 = 35 + X1 = 35 + 5 = 40
1 2.323
X2 5 x2 5 x2 0.325 X2 = 155 + X2 = 155 + 2 = 157

X1 = 2.323 × X2 = 2.323 × 2 = 4.646 5 Move the second step:

X1 = 40 + X1 = 40 + 5 = 45
X2 = 157 + X2 = 157 + 2 = 159

9-23 9-24
 Table 9.4 Steepest ascent experiment
Natural Coded
variables variables Response
Steps X1 X2 x1 x2 y
origin 35 155 0 0 40.44
5 2 1 0.42
origin + 40 157 1 0.42 41.0
origin + 2 45 159 2 0.84 42.9
origin + 3 50 161 3 1.26 47.1
origin + 4 55 163 4 1.68 49.7
origin + 5 60 165 5 2.10 53.8
origin + 6 65 167 6 2.52 59.9
origin + 7 70 169 7 2.94 65.0
origin + 8 75 171 8 3.36 70.4
origin + 9 80 173 9 3.78 77.6
origin + 10 85 175 10 4.20 80.3
origin + 11 90 177 11 4.62 76.2
origin + 12 95 179 12 5.04 75.1

Figure 9.6
9-25 9-26

At each design point along the steepest ascent direction, conduct an (1) If either 1 , or 2 , or both are large, we have to decide a new
experiment with the particular values of (X1, X2). If the observed steepest ascent direction from the current values of 1 and 2 ,
response function value Y is increased, move the design to the next and carry out a new linear search.
point along the same direction and using the same step sizes X1 and
X 2. (2) If both 1 and 2 are very close to zero, the design point is
already close to the optimum. Therefore, we can turn to the
The movement will be continued until the response function cannot second-order search.
be further increased.
Yˆ
Then, we have to decide if the design point is close to the optimum. It 0
x1 1
is done by carrying out a new 22 experiment, from which, the new
values of 1 and 2 can be evaluated ( 1 = A/2 and 2 = B/2). Yˆ 2 0
x2

Gradient being very close to zero indicates that the response

function is close to maximum (or minimum).
9-27 9-28
 9.3 Second-Order Model

k k
2
Y 0 i xi ii x i ij x i x j (9.22)
i 1 i 1 i j

This method is used to locate the optimum point more accurately

when the design is close to the optimum (or when all of the gradient
elements i are very small).

If k = 2
2 2
Y 0 1 x1 2 x2 11 x1 22 x 2 12 x1 x 2 (9.23)

Figure 9.7 The coefficients are obtained by least squares method.

9-29 9-30

The maximum (or minimum) of Y can be obtained by General mathematical solution

(1) Contour Plots (see Figure 9.2). 1
Xs = - B-1b (9.25)
By using the contour plots in the software for response surface 2
analysis, the experimenter can usually characterize the shape of
the surface and locate the optimum with reasonable precision. x1s 11 , 12 / 2, 1k / 2, 1
x2 s 22 , 2k / 2, 2
(2) Calculus Xs = B= b= (9.26)

Y xks sym kk , k
1 2 11 x1 12 x 2 0
x1
(9.24)
Y Xs is the stationary point. Furthermore, we can find the response
2 2 22 x 2 12 x1 0 function value at Xs:
x2
Two simultaneous equations with two unknowns x1 and x2. 1
ys 0 x’sb (9.27)
2
9-31 9-32
Characterize the response surface 9.4 Choice of an Experimental Design
A stationary point may be a For the first-order designs, we use the 2k factorial design.
(1) Maximum point; (2) Minimum point; (3) Saddle point

It can be checked by
40.0 41.5
(1) Examine a contour plot (Figure 9.2)

(2) Canonical analysis.

39.3 40.9
Figure 9.8

9-33 9-34

For the second-order design:

(1) The Central Composite Design (CCD)
In order to use least squares method to fit the response function
(9.23), a special experiment, Central Composite Design, is carried
out to take sufficient number of observations.
(1) 2k factorial observations.
(2) 2k axial observations.
(3) c center points (usually 3 to 5).
Total number of observations is (2k + 2k + c)

Axial spacing = F1/4

F = 2k is the number of runs in the factorial part of the design. Figure 9.9
2 1/4 Coded coordinates should be converted to natural coordinates (9.12)
For 2-factor experiment, F = 2 = 4, =4 = 1.414
for actually experiment
9-35 9-36
 Table 9.5 Central Composite Design (CCD) Coordinate conversion:
180 +1
Natural Coded For example of X1:
variables variables Response
X1 X2 x1 x2 y1 y2 y3 X 1 85 90 85
(yield) (viscosity) (mole. w) 175 0
80 170 -1 -1 76.5 62 2940 x1 0 1 0
80 180 -1 1 77.0 60 3470
90 170 1 -1 78.0 66 3680 X1 5 x1 85
90 180 1 1 79.5 59 3890 170 -1
85 175 0 0 79.9 72 3480 For x1 = 1.414 0
85 175 0 0 80.3 69 3200 -1 +1
85 175 0 0 80.0 68 3410 X1 5 1.414 85 92.07
85 175 0 0 79.7 70 3290
85 175 0 0 79.8 71 3500
80 85
92.07 175 1.414 0 78.4 68 3360
For x1 = -1.414
77.93 175 -1.414 0 75.6 71 3020 X1 5 ( 1.414) 85 77.93
85 182.07 0 1.414 78.5 58 3630
85 167.93 0 -1.414 77.0 57 3150

9-37 9-38

9.5 Main Steps for RSM Search (6)Move the design point one step along the steepest ascent direction.

First Order Method (Steepest Ascent Method) (7) Carry out experiment in the new design point. Check if the
response function has been improved.
(1) Find the relation functions (9.11) and (9.12) between the natural
variables and the coded variables. (8) Steps (6) and (7) should be repeated until the response function
cannot be further improved.
(2) Carry out the 2k experiment, calculate the factor effects and use
them to determine k (= half of the effect). (9) Carrying out a new 2k experiment. If either 1 , or 2 , or both are
large, a new linear search is started (repeat steps (5) to (8)). If
(3) Determine the steepest ascent direction by equation (9.8). both 1 and 2 are very close to zero, we can turn to the second-
(4) Specify the step size in terms of one of the natural variable. order search.

(5) Determine the step sizes of other natural variables through the
coded variables.

9-39 9-40
Second Order Method

(10) Carry out the central composite design experiment.

(11) Use the observed data to fit a second order model (9.22) using
least squares method.

(12) Use calculus or contour plot to find the optimum design.

9-41
10 TAGUCHI'S METHOD Parameter Design

10.1 Introduction Controllable factor: It can be set at any value or level that is desired
by the user during the real field operation. Examples: feed rate of a
Three-stage designs machine, status of a switch (on or off).
(1) system design, scientific and engineering principles and
Uncontrollable factor (noise): The user is unable to control the
experience are used to create a prototype of the product that will
value or level of an uncontrollable factor during the real field
meet the functional requirements and also to create the process
operation. Examples: environmental temperature, number of
that will build it.
participants for a seminar.
(2) Parameter design involves finding the optimal settings of the
But, the experimenter must be able to control the uncontrollable
product and process parameters in order to optimize performance
factors during the experiment (i.e., be able to set them in certain
characteristics.
given levels), so that their influence on the response can be observed
(3) Tolerance design and studied in a predetermined way.

10-1 10-2

Purpose for using the parameter designs: Three types of responses

(1) Designing products or processes so that they perform consistently (1) Nominal the best search for a response value that is closest to
on target and are robust (insensitive) to environmental conditions. the target value (e.g., diameter of a shaft).
It means to reduce the variation caused by the uncontrollable
factors. (2) Smaller the better search for a small response value (e.g.,
roughness of a surface).
Uncontrollable factors cause the deviation from the target. But, to
eliminate them is impractical and often impossible. Taguchi (3) Larger the better search for a large response value (e.g., yield
method tries to determine the optimal level of the controllable of a product).
factors in order to minimize the effects of uncontrollable factors.
Compare Taguchi parameter design with the classic DOE
(2) Designing/developing products so that they are robust to the
variation due to the controllable factors. (1) Taguchi emphasizes the reduction of variation.

(3) Optimizing the response function. (2) Taguchi separate the controllable and uncontrollable factors
explicitly.

10-3 10-4
For a Larger the better example (yield) 10.2 Parameter Design
Example 10.1: A Parameter Design
y
1st design Output is the yield, y. (the larger the better).
20 Table 10.1
Controllable factor Levels
A. interference Low Medium High
15 2nd design B. wall thickness Thin Medium Thick
C. insertion depth Shallow Medium Deep
10 D. percent adhesion Low Medium High
Temperature (uncontrollable)
Uncontrollable factor Levels
Figure 10.1
The first design gives yield not only higher, but also more consistent. E. conditioning time 24h 120 h
F. working temperature 72o C 150o C
G. relative humidity 25% 75%
10-5 10-6

The purpose of this parameter design is to find the best run Table 10.2 Inner Array and Outer Array
(combination of the levels of the controllable factors), so that: L9 orthogonal array for L8 orthogonal array
controllable factors for uncontrollable
(1) The value of the yield (response) will be very stable, i.e., least factors
influenced by the uncontrollable factors. Inner array Outer array
(2) The value of the yield will be maximized. Run A B C D Run E F G
1 1 1 1 1 1 1 1 1
2 1 2 2 2 2 1 1 2
3 1 3 3 3 3 1 2 1
4 2 1 2 3 4 1 2 2
5 2 2 3 1 5 2 1 1
6 2 3 1 2 6 2 1 2
7 3 1 3 2 7 2 2 1
8 3 2 1 3 8 2 2 2
9 3 3 2 1

10-7 10-8
The inner and outer orthogonal arrays can be selected from the text Table 10.3 for Parameter Design
book according to the following parameter values:
Outer E 1 1 1 1 2 2 2 2
Number of factors Array F 1 1 2 2 1 1 2 2
G 1 2 1 2 1 2 1 2
Number of levels for each factor Inner Array
Run A B C D Response y y SNL
Number of runs 1 1 1 1 1 15.6 9.5 16.9 19.9 19.6 19.6 20.0 19.1 17.5 24.0
2 1 2 2 2 15.0 16.2 19.4 19.2 19.7 19.8 24.2 21.9 19.5 25.5
3 1 3 3 3 16.3 16.7 19.1 15.6 22.6 18.2 23.3 20.2 19.0 25.3
Most of the orthogonal arrays are the fractional factorial designs. 4 2 1 2 3 18.3 17.4 18.9 18.6 21.0 18.9 23.2 24.7 20.1 25.9
5 2 2 3 1 19.7 18.6 19.4 25.1 25.6 21.4 27.5 25.3 22.8 26.9
For examples, the inner L9 orthogonal array is a 34-2 fractional design 6 2 3 1 2 16.2 16.3 20.2 19.8 14.7 19.6 22.5 24.7 19.2 25.3
7 3 1 3 2 16.4 19.1 18.4 23.6 16.8 18.6 24.3 21.6 19.9 25.7
2 34 81 8 3 2 1 3 14.2 15.6 15.1 16.8 17.8 19.6 23.2 24.2 18.3 24.9
Number of runs 34 9
32 9 9 3 3 2 1 16.1 19.9 19.3 17.3 23.1 22.7 22.6 28.6 21.2 26.1

The outer L8 orthogonal array is simply a 23 factorial design.

Number of runs = 9, n = 8, N = 72
Number of runs 23 = 8

10-9 10-10

Each row indicates a run determined by a set of the values of the An illustrating parameter design
controllable factors.
Response: performance index of a car.
For each run, n = 8 observations are taken under different
environmental (noise) conditions, which are generated by the outer Controllable factors
array of the uncontrollable (noise) factors. It implies that each run A: Engine of the car
can be investigated under different environmental conditions.
B: Material of the car

Uncontrollable factor
T: Temperature

10-11 10-12
(1) Suppose the uncontrollable factor T cannot be controlled during (2) If the uncontrollable factor T can be controlled (set at different
the experiment. It always takes the ambient temperature value, values) during the experiment, the performance of the car can be
fixed at 30 (as in Singapore). As a result, we only know the studied under different temperature. As a result, a complete
performance of the car under high temperature. If we conclude picture of the overall performance of the car is available. If we
Run 2 is the best run, the conclusion is valid only in tropical conclude Run 4 is the best run, the conclusion is valid in
countries. different countries all around the world.

Table 10.4 Table 10.5

Temperature T 30 Temperature T -30 -20 -10 0 10 20 30 40
Run A B Response y y Run A B Response y y
1 1 1 15.6 9.5 16.9 19.9 19.6 19.6 20.0 19.1 17.5 1 1 1 15.6 9.5 16.9 19.9 19.6 19.6 20.0 19.1 17.5
2 2 1 18.3 17.4 18.9 18.6 21.0 18.9 23.2 24.7 20.1 2 2 1 15.0 16.2 19.4 19.2 19.7 19.8 24.2 21.9 19.5
3 1 2 16.3 16.7 19.1 15.6 22.6 18.2 23.3 20.2 19.0 3 1 2 16.3 16.7 19.1 15.6 22.6 18.2 23.3 20.2 19.0
4 2 2 15.0 16.2 19.4 19.2 19.7 19.8 24.2 21.9 19.5 4 2 2 18.3 17.4 18.9 18.6 21.0 18.9 23.2 24.7 20.1
Number of runs = 4, n = 8, N = 32 Number of runs = 4, n = 8, N = 32

10-13 10-14

On Table 10.3, y is the mean (average) response over the n SN is the signal-to-noise ratio, measuring the variation of the
observations on a run observations in a run.

n There are three different formulae for calculating the SN ratio,

yi depending on the type of the response.
i 1
y (10.1)
n (1) Nominal the best

yi is the ith observation in this run. SN N 10 log10 ( y 2 / S 2 ) (10.2)

n
2
On Table 10.3, for the first run, yi y
i 1
S2
15.6 9.5 19.1 n 1
y 17.5
8

10-15 10-16
(2) Smaller the better To calculate the SN ratio for each run (row), we first have to decide
the type of the response, and then select an appropriate formula.
1 n 2
SN S 10 log10 yi (10.3)
ni 1 Even though there are three different cases, in any case, a larger SN
always indicates a smaller variation, and therefore the design is
(3) Larger the better more robust. We are always seeking a design that results in a large
SN ratio.
1 n 1
SN L 10 log10 (10.4)
n i 1 y i2

For the 1st run on Table 10.3,

1 1 1 1
SN L 10 log10 24.0
8 15.6 2 9.5 2 19.12

10-17 10-18

10.3 Determination of the Best Design

25 y
(1) Primitive Method to Find the Best Design
5th run
Identify the best run (which produces largest SN and optimum mean
response y ) directly from the parameter design table (Table 10.3). 20
1st run
Optimum y means:
15
a y closest to the nominal value (for “nominal the best”); or
a minimum y (for “smaller the better”); or 10
a maximum y (for “larger the better”)
Figure 10.2
On Table 10.3, the 5th run results in maximum SN and maximum y .
Optimum run: A 2, B 2, C 3, D1
By using this run, we will obtain largest yield on average, and the Factor A B C D
variation of yield caused by the variation of the uncontrollable factors
Level medium medium deep low
will be minimum. Thus, the 5th run is the optimum design.
10-19 10-20
(2) Marginal Graph Method Each controllable factor has a marginal graph on mean response.
Figure 10.3 Marginal Graph on Mean Response Each marginal graph has k (= 3) dots (k is the number of levels of the
factors).
21 21
y 20 y 20
[1] The height of the L dot is equal to the average of the y values of
19 19
all the runs in which the factor is at the low level.
18 18
L M H L M H
A [2] The height of the M dot is equal to the average of the y values of
B
all the runs in which the factor is at the medium level.
21 21
y 20 y 20 [3] The height of the H dot is equal to the average of the y values of
19 19 all the runs in which the factor is at the high level.
18 18
L M H L M H
C D
10-21 10-22

Factor A on Table 10.3 Factor B on Table 10.3

Height of L dot = yrun1 yrun 2 yrun 3 / 3 Height of L dot = yrun1 yrun 4 yrun 7 / 3
(17.5 + 19.5 + 19.0) / 3 = 18.7 (17.5 + 20.1 + 19.9) / 3 = 19.2

Height of M dot = yrun 4 yrun 5 yrun 6 / 3 Height of M dot = yrun 2 yrun 5 yrun 8 / 3
(20.1 + 22.8 + 19.2) / 3 = 20.7 (19.5 + 22.8 + 18.3) / 3 = 20.2

Height of H dot = yrun 7 yrun8 yrun 9 / 3 Height of H dot = yrun 3 yrun 6 yrun 9 / 3
(19.9 + 18.3 + 21.2) / 3 = 19.8 (19.0 + 19.2 + 21.2) / 3 = 19.8

The positions of the dots on the marginal graphs of other factors can
be determined similarly.

10-23 10-24
For a controllable factor, if its marginal graph on mean response is Figure 10.5 Marginal Graph on SN
almost horizontal, it has negligible effect on the response. Otherwise,
this factor has significant effect on the response. 27 27
SNL 26 26
21 21
SNL
25 25
y 20 y 20 24 24
19 19
effective L M H L M H
18 18 not effective
A B
L M H L M H
A B 27 27
26 26
21 21
SNL SNL
25 25
y 20 y 20 24 24
19 19
effective effective L M H L M H
18 18 D
C
L M H L M H Figure 10.4
C D
10-25 10-26

Similarly, each controllable factor has a marginal graph on SN ratio. Now, from the marginal graph on mean response (Figure 10.3), in
order to obtain the maximum y , the factors should be set as follows:
[1] The height of the L dot is equal to the average of the SN values of
all the runs in which the factor is at the low level. A 2, B 2, C 3, D 1
[2] The height of the M dot is equal to the average of the SN values of Similarly, from the marginal graph on SN ratio (Figure 10.5), in
all the runs in which the factor is at the medium level. order to obtain the maximum SN (or the minimum variation), the
[3] The height of the H dot is equal to the average of the SN values of factors should be set as follows:
all the runs in which the factor is at the high level. A 2, B 2, C 3, D 1
The procedure is similar to that for the mean response, except that y Both settings are the same. It is the 5th run in Table 10.3. This
values should be replaced by SN values. combination of the controllable factor levels will result in the
maximum SN and maximum y simultaneously.
For a controllable factor, if its marginal graph on SN ratio is almost
horizontal, it has negligible effect on the SN ratio. Otherwise, this Factor A B C D
factor has significant effect on the SN ratio. Level medium medium deep low

10-27 10-28
Example 10.2 Another set of the marginal graphs are as follows: Figure 10.7 Marginal Graph on SN
Figure 10.6 Marginal Graph on Mean Response 27 27
SNL 26 26
21 21 SNL
25 25
y 20 y 20 24 24
19 19
L M H L M H
18 18 A B
L M H L M H
A B 27 27
26 26
21 21 SNL SNL
25 25
y 20 y 20 24 24
19 19
L M H L M H
18 18
C D
L M H L M H
C D
10-29 10-30

The best run, that will result in the maximum SN and maximum y Table 10.6 for Parameter Design
simultaneously, is as follows:
Outer E 1 1 1 1 2 2 2 2
Array F 1 1 2 2 1 1 2 2
A 3, B 2, C 3, D 1 G 1 2 1 2 1 2 1 2
Inner Array
Factor A B C D Run A B C D Response y y SNL
Level high medium deep low 1 1 1 1 1 15.6 9.5 16.9 19.9 19.6 19.6 20.0 19.1 17.5 24.0
2 1 2 2 2 15.0 16.2 19.4 19.2 19.7 19.8 24.2 21.9 19.5 25.5
However, this run cannot be found in Table 10.3. Because the 3 1 3 3 3 16.3 16.7 19.1 15.6 22.6 18.2 23.3 20.2 19.0 25.3
4 2 1 2 3 18.3 17.4 18.9 18.6 21.0 18.9 23.2 24.7 20.1 25.9
orthogonal arrays are usually the fractional designs. Only a fraction 5 2 2 3 1 19.7 18.6 19.4 25.1 25.6 21.4 27.5 25.3 22.8 26.9
of all possible runs have been included. 6 2 3 1 2 16.2 16.3 20.2 19.8 14.7 19.6 22.5 24.7 19.2 25.3
7 3 1 3 2 16.4 19.1 18.4 23.6 16.8 18.6 24.3 21.6 19.9 25.7
Under such circumstance, we should take 8 additional observations, 8 3 2 1 3 14.2 15.6 15.1 16.8 17.8 19.6 23.2 24.2 18.3 24.9
under the above predicted optimum run. This is called the 9 3 3 2 1 16.1 19.9 19.3 17.3 23.1 22.7 22.6 28.6 21.2 26.1
Conf.
confirmation test. The purpose is to use real experimental data to test 3 2 3 1 24.1 23.9 24.1 24.0 24.0 24.1 23.9 24.1 24.0 27.6
verify the optimum run obtained by the marginal graph approach.

10-31 10-32
(3) Empirical Optimization Approach (Example 10.3) Figure 10.9 Marginal Graph on SN
Figure 10.8 Marginal Graph on Mean Response 27 27
SNL 26 26
21 21 SNL
25 25
y 20 y 20 24 24
19 19
L M H L M H
18 18 A B
L M H L M H
A B 27 27
26 26
21 21 SNL SNL
25 25
y 20 y 20 24 24
19 19
L M H L M H
18 18
C D
L M H L M H
C D
10-33 10-34

From the marginal graph on mean response, in order to obtain the Three types of the controllable factors
maximum y , the factors should be set as follows:
Control factor significantly affects SN (disregard the effect on
A 2, B 3, C 3, D 1 mean response)
From the marginal graph on SN ratio, in order to obtain the Factors A and C (from Figure 10.9)
maximum SN, the factors should be set as follows:
Signal factor has little influence on the SN ratio, but significant
A 2, B 1, C 1, D 3 effect on the mean response
There is a contradiction ! Factor D (from Figure 10.8)

Other factor almost has neither effect on the mean response nor
effect on the SN ratio
Factor B (from Figures 10.8 and 10.9)

10-35 10-36
Steps for the Empirical Optimization Approach [3] Other factors can be set arbitrarily, or based on some other
considerations, such as cost reduction, ease of manufacture.
[1] Identify the control factors, and set them to the levels that
maximize SN For example, if the manufacturing cost is lowest when factor
B is set at the high level, the final optimal design is
A 2, B ?, C 1, D ? (from Figure 10.9)
A 2, B 3, C 1, D 1
SN ratio is considered first, because Taguchi believes that
minimizing variation is most important. Since factor B has little effect on both SN and mean response,
when it is set to any level, the SN value remains maximum as
[2] Identify the signal factors, and use them to bring the mean obtained in step [1] and the y value also remains maximum as
response to optimum. obtained in step [2].

A 2, B ?, C 1, D 1 (from Figure 10.8)

Since factor D has little effect on SN, when it is set to any
level, the SN value remains maximum as obtained in step [1]
10-37 10-38

10.4 Main Steps for Taguchi’s Parameter Design (5) Use one of the following methods to decide a best run
(combination of the levels of the controllable factors), so that
(1) Decide the controllable and uncontrollable factors and their the SN ratio is maximum and the mean response is optimum.
levels to be tested.
Primitive method
(2) Select an appropriate inner array for the controllable factors, Marginal graph method
and an outer array for the uncontrollable factor.
Empirical optimization
(3) Set up the table for the parameter design, and carry out the
experiment accordingly. (6) If the best run has not been included in the table, carry out n
additional confirmation tests.
(4) From the observations, calculate the mean response and SN
ratio for each run (row) on the table.

10-39 10-40