Basic Design of Experiments

and
Introduction to Taguchi
Methods
Flight of a Bumblebee

A helicopter built to the blue print’s

specifications will have characteristic
properties.
For example : “ when dropped from a
height of 10 feet, it will fly an average of
3.2 seconds.”

Problem : Design the longest flying

helicopter out of an 4.25” x 11” paper.
Flight of a Bumblebee
Helicopter Design Features
F

Constant
A E = 11”

A = Paper type E = Shoulder

B = Body Width F = Wing Width
C = Body Length C + D + E = 11”
D = Wing Width
Flight of a Bumblebee
Enumerate the steps you are going to take
to solve the given problem:

1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
In Need of Designed Experiments
An Integral Part of the Push Upstream

upstream

Going upstream means finding out

• what factors affect important responses

• how they do so.

In Need of Designed Experiments

How should we run

experiments ?

It is almost always
better to study
more than one factor
at a time.
In Need of Designed Experiments

Example :
Conveyor Belt Slitting Process
Problem :
Improve the slitting process to produce
belts of uniform width.

• What is the best speed of cut (S) to

use ?

• What is the best brand of knife (K) to

use ?

• Is there a need to use a lubricant ?

In Need of Designed Experiments

The belt slitting machine manufacturer

recommends :
• Speed of cut = 15 cm per second

• Knives from the Acme company

• A silicone lubricant

BUT … are these the best settings for

your installation ?

Only an experiment can answer that question.

In Need of Designed Experiments

Factors to consider :

• Speed of cut can vary from 10 cm per

second to 20 cm per second.

• The Zapum company has recently

introduced a new kind of knife.

• The vendor recommends no lubricant be

used with the Zapum knife.

Questions :
• How fast ?

• Which knife ?

• Lubricant ? Yes or no ?

Old dogma : let’s try it out ! But to eliminate

confusion - vary one factor at a time ! Hold all
other variables constant .
In Need of Designed Experiments

For example, vary the speed of cut while using

the Zapum knife without lubrication.

0.60

0.50
Width
Variation
0.40

0.30

slow fast
SPEED OF CUT

From this, we might conclude that a slow

speed of cut is better than a fast speed of cut...

But is it?
 In Need of Designed Experiments

By varying only the speed of cut, we have left

unexplored six alternative factor
combinations at the other corners of the cube.

7 8

Acme 3 4

KNIFE
5 6 yes
LUBRICATION
Zapum 1 2 no
slow fast

SPEED OF CUT

From this, we might conclude that a slow

speed of cut is better than a fast speed of cut...

But is it?
In Need of Designed Experiments

After all 8 trials, all we can really say is that it

appears to be better to use the slower speed of
cut…. If we use a Zapum knife and no
lubricant.
But if the question is asked :

Will the slow speed of cut work equally

well for the Acme knife or with the
silicone lubricant, the honest answer is ---
-
WE DON’T KNOW.

Sir Ronald A. Fisher (1920’s) :

Vary all factors simultaneously.
Factorial Design:

• Run eight experiments to test all three factors

• And get even more information than running

eight one-at-a-time experiments !

Notation :
For k factors at 2 levels each, a full design is
called a
2k factorial design with n=2k design points.
 In Need of Designed Experiments
The experiment design in geometric
notation:
Run Speed Knife Lube
1 - - -
2 + - -
3 - + -
4 + + -
5 - - +
6 + - +
7 - + +
8 + + +

7 8

Acme 3 4

KNIFE
5 6 yes
LUBRICATION
Zapum 1 2 no
slow fast

SPEED OF CUT
 In Need of Designed Experiments
The experiment design in geometric
notation:
Width
Run Speed(A) Knife(B) Lube(C)
Variation
1 slow Zapum no 0.41
2 fast Zapum no 0.51
3 slow Acme no 0.33
4 fast Acme no 0.50
5 slow Zapum yes 0.38
6 fast Zapum yes 0.44
7 slow Acme yes 0.43
8 fast Acme yes 0.54

0.43 7 8 0.54
0.33
Acme 3 4 0.50

KNIFE 0.38 0.44

5 6 yes
LUBRICATION
Zapum 1 2 no
0.41 slow fast 0.51

SPEED OF CUT
 In Need of Designed Experiments
Belt Slitting Marginal Means Response

0.55

0.5

0.45

0.4

0.35

0.3
A1 A2 B1 B2 C1 C2 (AxB)1 (AxB)2 (AxC)1 (AxC)2 (BxC)1 (BxC)2 (AxBxC)1 (AxBxC)2

Effects Level

Conclusions:

• Slower speed of cut improves belt width

uniformity even when the other factors vary.

• There is an interaction effect between brand

of knife and the use of a lubricant. It is better
to
- use a lubricant with a Zapum knife.
- not use a lubricant with an Acme knife.
This is the opposite of what the vendor
recommended !
Introduction to Factorial Experiments
Definition of Terms :
1. Factor - one of the things being studied in an
experiment. An independent variable.
2. Level - the value of the factor examined in the
experiment.
3. Treatment - a single level assigned to a single factor
during an experimental run.
4. Treatment combination - one set of levels for all
factors in a given experimental run.

5. Block - a portion of the experimental material or of

the experimental environment which is likely to be
homogeneous within itself than are other portions.

6. Experiment Design - the formal plan for conducting

the experiment. It includes the choice of factors,
levels, treatments, and use of certain tools called
planned grouping, randomization, and replication.

7. Randomization - the assignment of specimens to

treatments in a purely chance manner.

8. Replication - the repetition of an observation or

measurement. It is done to increase precision and to
provide the means for measuring precision.

9. Factorial Experiment - an experiment wherein

several factors are each investigated at several
levels by running all combinations of factors and
levels.
 Introduction to Factorial Experiments

Full Factorial Experiments

Two-level Experiment
For k factors at 2 levels each, a full design is
called a
2k factorial design with n=2k design points.

Example : How many design points are there if

there are 3 factors and 2 levels ?
n=2k = 23 = 8 design points

Three-level Experiment
For k factors at 3 levels each, a full design is
called a
3k factorial design with n=3k design points.
Example : How many design points are there if
there are 3 factors and 3 levels ?
n=3k = 33 = 27 design points
 Fractional Factorial Experiments
Full 2-level factorials can require a lot of runs.

# of factors # of runs
1 2
2 4
3 8
4 16
5 32
6 64
7 128
8 256
9 512
10 1024

With a full 27 experiment, the following effects can be

estimated :
Type of effect Number
Main 7
2-factor interactions 21
3 factor interactions 35
4 factor interactions 35
5 factor interactions 21
6 factor interactions 7
7 factor interactions 1
 Fractional Factorial Experiments
A 27 full factorial experiment : n = 128
A0 A1
B0 B1 B0 B1
C0 C1 C0 C1 C0 C1 C0 C1
D0 D1 D0 D1 D0 D1 D0 D1 D0 D1 D0 D1 D0 D1 D0 D1
G0
F0 G1
E0 G0
F1 G1
G0
F0 G1
E1 G0
F0 G1

A 1/4 fractional factorial of the 27 factorial

A0 A1
B0 B1 B0 B1
C0 C1 C0 C1 C0 C1 C0 C1
D0 D1 D0 D1 D0 D1 D0 D1 D0 D1 D0 D1 D0 D1 D0 D1
G0
F0 G1
E0 G0
F1 G1
G0
F0 G1
E1 G0
F0 G1
 ANALYSIS OF VARIANCE
for 2-factor Experiments with n replications
Source of Sum of Degrees of Variance Compute f
Variation Squares Freedom
Main Effects
A SSA a–1 s12 = SSA/(a - 1) f12 = s12/s2
B SSB b–1 s22 = SSB/(b - 1) f22 = s22/s2
Interaction
AB SS(AB) (a – 1)(b – s32 = SS(AB)/(b-1)(a – 1) f32 = s32
1)
Error SSE ab(n – 1) S2 = SSE/(ab)(n – 1)
Total SST abn – 1

TABLE OF TOTALS
B Total
A 1 2 …… b
1 T11 T12 …… T1b T1..
2 T21 T22 …… T2b T2..
. . . …… . .
. . . …… . .
. . . …… . .
a Ta1 Ta2 …… Tab Ta..
Total T.1. T.2. …… T.b. T…

a b c
T ...2
SST  
i 1 j 1 k 1
y ijk 
abn
a

 T i..2
T ...2
SSA  i 1

bn abn
b

 j 1
T .j.2
T ...2
SSB  
an abn
a b a b

 i 1 j 1
T ij.2  T i..2 
j1
T .j.2
T ...2
SS(AB)   i 1
 
n bn an abn
SSE  SST  SSA  SSB  SS(AB)
 2-factor ANOVA

Example of 2-factor Analysis of Variance

In an experiment conducted to determine which of 3 different
missile systems is preferable, the propellant burning rate for 24
static firings was measured. Four different propellant types were
used. The experiment yielded duplicate observations of burning
rates at each combination of treatments. Data are shown below:
Propellant type
Missile System b1 b2 b3 b4
a1 34.0 30.1 29.8 29.0
32.7 32.8 26.7 28.9
a2 32.0 30.2 28.7 27.6
33.2 29.8 28.1 27.8
a3 28.4 27.3 29.7 28.8
29.3 28.9 27.3 29.1
Use a 0.05 level of significance to test the following hypotheses:
(a) Ho’ : There is no difference in the mean propellant burning
rates when different missile system are used.
(b) Ho’’ : There is no difference in the mean propellant burning
rates of the 4 propellant types.
(c) Ho’’’: There is no interaction between the different missile
systems and the different propellant types.
 2-factor ANOVA
Solution:
First, a table of totals is constructed:
b1 b2 b3 b4 Total
a1 66.7 62.9 56.5 57.9 244.0
a2 65.2 60.0 56.8 55.4 237.4
a3 57.7 56.2 57.0 57.9 228.8
Total 189.6 179.1 170.3 171.2 710.2

Sum of squares are then computed:

a b c
T...2
SST   yijk 
i 1 j1 k 1 abn

= (34.0)2+(32.7)2 +…..+(29.1)2 - (710.2)2/24 = 91.68

a

T
T...2
2
i..
SSA  i 1
bn abn
(244.0) 2  (237.4) 2  (228.8) 2 (710.2) 2
   14.52
(4)( 2) 24
b

TT...2
j1
2
.j.

SSB  
an abn
(189.6) 2  (179.1) 2  (170.3) 2  (171.2) 2 (710.2) 2
   40.08
(3)( 2) 24
a b a b

 T  T  T
i 1 j1
2
ij.
2
i..
T...2 j1
2
.j.

SS(AB)    i 1

n bn an abn
(66.7)  (65.2)  ...  (57.9)
2 2 2
  21030.52  210056.08  21016.00  22.17
2
 2-factor ANOVA
Solution (cont’d):
SSE = 91.68 - 14.52 - 40.08 - 22.17 = 14.91

Now, an ANOVA table can be constructed:

Source sum of df Variance computed Table
of variation squares f f0.05
-----------------------------------------------------------------------------------
Missile sys (A) 14.52 2 7.26 5.85 3.89
Propellant (B) 40.08 3 13.36 10.77 3.49
Interaction (AB) 22.17 6 3.70 2.98 3.00
Error (E) 14.91 12 1.24
----------------------------------------------------------------------------------
Total 91.68 23

Conclusions:
a. Accept Ho’’’ and conclude that there is no interaction
between the different missile systems and the different
propellant types.
b. Reject Ho’ and conclude that different missile systems result
in different mean propellant burning rates.
c. Reject Ho’’ and conclude that the mean propellant burning
rates are not the same for the four propellant types.
 ANALYSIS OF VARIANCE
for 3-factor Experiment with n Replications

Source of Sum of Degrees of Variance Computed f

Variation Squares Freedom
Main Effects
A SSA a–1 s12 f1 = s12/s2
B SSB b–1 s22 f2 = s22/s2
C SSC c–1 s32 f3 = s3/s2
2-factor
interactions
AB SS(AB) (a – 1)(b – 1) s42 f4 = s42/s2
AC SS(AC) (a – 1)(c – 1) s52 f5 = s5/s2
BC SS(BC) (b – 1)(c – 1) s62 f6 = s62/s2
3-factor
interaction
ABC SS(ABC) (a – 1)(b – 1)(c – 1) s72 f7 = s72/s2
Error SSE abc(n – 1) s2
Total SST abcn – 1

3-way ANOVA Sum of Squares

a b c
T2 n
SST      y ijkl 2
 ....
i 1 j 1 k 1 l 1 abcn
a

T
i 1
2
i...
T....2
SSA  
bcn abcn
a

T
i 1
2
.j..
T....2
SSB  
acn abcn
a

T
i 1
2
..k.
T....2
SSC  
abn abcn
a b a b

T
i 1 j 1
2
ij..  Ti...2 T
j 1
2
.j..
T....2
i 1
SS(AB)    
cn bcn acn abcn
a c a c

T
i 1 k 1
2
.j.. T
i 1
2
i... T
k 1
2
..k.
T....2
SS(AC)    
bn bcn abn abcn
b c b c
T
j 1 k 1
2
.jk. T
j 1
2
.j.. T 2
..k.
T....2
k 1
SS(BC)    
an acn abn abcn
a b c a b a c b c a b c
T
i 1 j 1 k 1
2
ijk. T
i 1 j 1
2
ij..   T.j..2 T
j 1 k 1
2
.jk.  Ti...2 T
j 1
2
.j.. T 2
..k.
T....2
i 1 k 1 i 1 k 1
SS(ABC)        
n cn bn an bcn acn abn abcn
SSE  SST  SSA  SSB  SSC  SS(AB)  SS(AC)  SS(BC)  SS(ABC)
 3-factor ANOVA

Example of 3-factor Analysis of Variance

In the production of a particular material, three variables are of
interest: A the operator effect (three operators), B the catalyst
used in the experiment (three catalysts), and C the washing time
of the product following the cooling process (15 & 20 minutes).
Three runs were made at each combination of factors. The yield
are as follows:

Washing time, C
15 minutes 20 minutes
B B
A 1 2 3 1 2 3
10.7 10.3 11.2 10.9 10.5 12.2
1 10.8 10.2 11.6 12.1 11.1 11.7
11.3 10.5 12.0 11.5 10.3 11.0
11.4 10.2 10.7 9.8 12.6 10.8
2 11.8 10.9 10.5 11.3 7.5 10.2
11.5 10.5 10.2 10.9 9.9 11.5
13.6 12.0 11.1 10.7 10.2 11.9
3 14.1 11.6 11.0 11.7 11.5 11.6
14.5 11.5 11.5 12.7 10.9 12.2
Perform an analysis of variance to test for significant effects. Use
 = 0.05.
 3-factor ANOVA
Solution :
First, construct a table of totals:

Washing time, C
15 minutes 20 minutes
B B
A 1 2 3 1 2 3
1 32.8 31.0 34.8 34.5 31.9 34.9
2 34.7 31.6 31.4 32.0 30.0 32.5
3 42.2 35.1 33.6 34.1 32.6 35.7

B
A 1 2 3 Total
1 67.3 62.9 69.7 199.9
2 66.7 61.6 63.9 192.2
3 77.3 67.7 69.3 214.3
Total 211.3 192.2 202.9 606.4

C
A 1 2 Total
1 98.6 101.3 199.9
2 97.7 94.5 192.2
3 110.9 103.4 214.3
Total 307.2 299.2 606.4

C
B 1 2 Total
1 109.7 101.6 211.3
2 97.7 94.5 192.2
3 99.8 103.1 202.9
Total 307.2 299.2 606.4
 3-factor ANOVA
Solution (cont’d) :
From the table of totals, the sum of squares can be computed:
(606.4) 2
SST  (10.7) 2  (10.8) 2  ....(12.2) 2   6872.84  6809.65  63.19
54

(199.9) 2  (192.2) 2  (214.3) 2

SSA   6809.65  6823.63  6809.65  13.98
(3)(2)(3)

(211.3) 2  (192.2) 2  (202.9) 2

SSB   6809.65  6819.83  6809.65  10.18
(3)(2)(3)

(307.2) 2  (299.2) 2
SSC   6809.65  6810.83  6809.65  1.18
(3)(3)(3)

(67.3) 2  (66.7) 2  ....  (69.3) 2

SS(AB)   6823.63  6819.83  6809.65  4.78
(2)(3)

(98.6) 2  (97.7) 2  ....  (103.4) 2

SS(AC)   6823.63  6810.83  6809.65  2.92
(3)(3)

(109.7) 2  (97.7) 2  ....  (103.1) 2

SS(BC)   6819.83  6810.83  6809.65  3.64
(3)(3)

SS(ABC) 
32.8 2  (34.7) 2  ....  (35.7) 2  6838.59  6827.73  6824.65  6823.63
3
 6819.83  6810.83  6809.65  4.89

SSE  63.19 13.98 10.18 1.18  4.78  2.92  3.64  4.89  21.62
 3-factor ANOVA
Solution (cont’d) :
Finally, an ANOVA table is constructed:

Source sum of df Variance computed Table

of variation squares f f0.05
-----------------------------------------------------------------------------------
Operator (A) 13.98 2 6.99 11.65 3.29
Catalyst (B) 10.18 2 5.09 8.48 3.29
Washing time (C) 1.18 1 1.18 1.97 4.14
Interaction(AB) 4.78 4 1.20 2.00 2.66
Interaction(AC) 2.92 2 1.46 2.43 3.29
Interaction(BC) 3.64 2 1.82 3.03 3.29
Interaction(ABC) 4.89 4 1.22 2.03 2.66
Error (E) 21.62 36 0.60
----------------------------------------------------------------------------------
Total 63.19 53

Conclusions:
(a) Different operators produce different yields.
(b) Different catalysts have different effects on yields.
Steps in Experimentation using Taguchi Methods

Step 1. State the problem to be solved. A clear understanding of the problem by

those involved in the experiment is necessary to be able to structure the
experiment. The problem statement should be specific.
Step 2. Determine the objective of the [Link] includes the identification of
the performance characteristics (preferably measurable) and the level of
performance required when the experiment is complete.
Step 3. Determine the measurement method(s). An understanding of how the
performance characteristic(s) will be assessed after the experiment is
conducted. The measurement system may require a separate experiment to
improve the measurement system accuracy and precision.
Step 4. Identify factors which are believed to influence the performance
characteristic(s). A group of people associated with the product or process
should be utilized. Brainstorming, flowcharting, and/or fishbone charting
should be used to aid in the creation of the factors to be investigated. If this is
an initial screening experiment, many factors should be included if thought
to be influential to the response(s).
Step 5. Separate the factors into control and noise factors. This is a basic parameter
design strategy and should be used in place of a cause detection or tolerance
design approach.
Step 6. Determine the number of levels and values for all factors. The total degrees
of freedom required is a direct function of number of levels for the factors.
For initial screening experiments, the number of levels should be kept low;
two levels if possible.
Step 7. Identify control factors that may interact. These interactions use up degrees
of freedom also and may dictate the size of the experiment. A strategy may be
to select the size of the experiment based only on factors, and if there are any
columns available afterward then assign these to an interaction of interest.
Step 8. Draw the required linear graph for the control factors and interactions. The
desired factors and interactions may influence the selected orthogonal array.
Step 9. Select orthogonal arrays. The OAs, inner and outer, are a function of the
total degrees of freedom required from the factor list and the required linear
graph.
Steps in Experimentation using Taguchi Methods (continued)

Step [Link] factors and interactions to columns. The linear graph for an OA may
have to be modified to fit the required form.
Step [Link] the experiment. Randomization strategies should be considered
during the experiment.
Step [Link] the data. Several methods are used, e.g. , average graphs,
interaction graphs, ANOVA, S/N ANOVA, etc.
Step [Link] the results. Determine which factors are influential and which are
not influential to the performance characteristic(s) of influence.
Step [Link] optimum levels of most influential control factors and predict expected
results. The influential factors are the only ones necessary to have the level
set or controlled. The non-influential factors should be set at the lowest cost
level.
Step [Link] a confirmation experiment. This is to demonstrate that the factors and
levels chosen for the influential factors do provide the desired results. The
non-influential factors should be set at their economic level during the
confirmation run. If results don’t turn out as expected, then some important
factor(s) may have been left out of the experiment and more screening needs
to be done.
Step [Link] to step 4 if the objective of the experiment is not met and further
optimization is possible with the confirmed factors.
 Example Experimental Procedure: Popcorn
Experiment
Scenario : How to develop the process specifications (cooking instructions)
to go on a bag of popcorn. The owner of the popcorn company
has developed a new hybrid seed which may or may not use the
same cooking process recommended for the current seed. One of
the processes used by customers to pop corn is the hot oil
method, which is addressed in this situation.

Step 1. Statement of the Problem

The problem is to find the process factors which influence

popcorn quality characteristics relative to customer
requirements.

Step 2. Objective of the Experiment

To find the process conditions which optimize the various quality

characteristics to provide improved popping, fluffiness, color,
taste, and texture.

Step 3. Measurement Methods

Popping - count the number of unpopped kernels in a batch

Fluffiness - can be quantified by placing the popped kernels in a

measuring cup

Color - assign a color rating : good color, off color, bad color

Taste - assign a taste rating : good taste or bad taste

Texture - assign a texture rating : fine texture or coarse texture

 Example Experimental Procedure: Popcorn
Experiment
Step 4. Process Factors and Number of Levels
Process Flowchart

Heat Add Agitate Venting Popped

oil corn Corn

Fishbone Diagram

Oil Method

Type Agitation

Amount Venting
High quality
popcorn
Shape Preheat

Material Amount

Pan Heat

Step 6. Popcorn factors and levels

Factor Level 1 Level 2
A Type of oil Corn oil Peanut oil
B Amount of oil Low High
C Amount of heat Medium High
D Preheat No Yes
E Agitation No Yes
F Venting No Yes
G Pan material Aluminum Steel
H Pan shape Shallow Deep
 Orthogonal Arrays

The Meaning of “Orthogonal”

• being balanced
• not mixed
• statistically independent

L8 (27)
Column Number
Run No.\ 1 2 3 4 5 6 7
1 1 1 1 1 1 1 1
2 1 1 1 2 2 2 2
3 1 2 2 1 1 2 2
4 1 2 2 2 2 1 1
5 2 1 2 1 2 1 2
6 2 1 2 2 1 2 1
7 2 2 1 1 2 2 1
8 2 2 1 2 1 1 2

For each column: level 1 occurs 4 times, level 2 occurs 4 times.

Within each column, there is an equal number of occurrences
for each level.
Each level within any other column occurs an equal number of
times.
 Orthogonal Arrays

Standard Nomenclature
L8 (27)
Number of columns
Number of levels
Number of rows or runs
Represents Latin squares

Example

L9 (34) Number of columns = ____

Number of levels = ____
Number of runs = ____

Level Nomenclature

1 L -
2 M o
3 H +
 Orthogonal Arrays

2-Level Orthogonal Arrays

L4 (23)
Column Number
Run No.\ 1 2 3
1 1 1 1
2 1 2 2
3 2 1 2
4 2 2 1

Interaction Table

Column 1 2 3
---------------------------------------
(1) 3 2
(2) 1
(3)

The interaction table indicates that:

• the interaction between the column 1 variable & the
column 2 variable is associated with column 3.
• the interaction between the column 1 variable 2 &
the column 3 variable is associated with column 2.
• The interaction between the column 2 variable & the
column 3 variable is associated with column 1.
 Orthogonal Arrays

2-Level Orthogonal Arrays

L8 (27)
Column Number
Run No.\ 1 2 3 4 5 6 7
1 1 1 1 1 1 1 1
2 1 1 1 2 2 2 2
3 1 2 2 1 1 2 2
4 1 2 2 2 2 1 1
5 2 1 2 1 2 1 2
6 2 1 2 2 1 2 1
7 2 2 1 1 2 2 1
8 2 2 1 2 1 1 2

Interaction Table for L8 array

Column 1 2 3 4 5 6 7
----------------------------------------------------------------------------------
(1) 3 2 5 4 7 6
(2) 1 6 7 4 5
(3) 7 6 5 4
(4) 1 2 3
(5) 3 2
(6) 1
(7)
 Orthogonal Arrays

3-Level Orthogonal Array

L9 (34)
Column Number
Run No.\ 1 2 3 4
1 1 1 1 1
2 1 2 2 2
3 1 3 3 3
4 2 1 2 2
5 2 2 3 1
6 2 3 1 2
7 3 1 3 2
8 3 2 1 3
9 3 3 2 1
 Orthogonal Arrays
Uses of the L9 Orthogonal Array

For investigation of 2 factors:

Independent estimates of all effects of interest can be
obtained:
• linear effects of the 2 factors
• quadratic effects of the 2 factors
• the interaction between the linear effects of the 2 factors

For investigation of 3 factors:

Assuming almost all interaction effects are negligible,
independent estimates can be obtained for:
• linear effects of the 3 factors
• quadratic effects of the 3 factors
• two combinations of interaction effects

For investigation of 4 factors:

Assuming almost all interaction effects are negligible,
independent estimates can be obtained for:
• linear effects of the 4 factors
• quadratic effects of the 4 factors
 Degrees of Freedom

Degrees of Freedom

The number of independent comparisons that need to be

made without being redundant to derive a conclusion.

Relating to design of experiments, it is the number of

comparisons between factor or interaction levels that need to
be made to determine which level is better and specifically
how much better it is.

2 Degrees of Freedom

12 inches 8 inches

3 persons, 2 comparisons
 Degrees of Freedom
Calculating Degrees of Freedom in an Experiment

Factors or Main Effects

Degrees of freedom = number of levels - 1
2 levels = 1 degree of freedom
3 levels = 2 degrees of freedom
n levels = n - 1 degrees of freedom

Interactions
Degrees of freedom of interaction A x B =
( number of levels for factor A) - 1
x
(number of levels for factor B) - 1

Orthogonal Arrays
Degrees of freedom of an orthogonal array =
degrees of freedom per column: B - 1
x
number of columns

Example : how many degrees of freedom does an L8 have ?

Degrees of freedom per column = 2 -1 = 1
Number of columns = 7
7
How many degrees of freedom does an L9 have?
 Linear Graphs
L4(23) linear graph

3
1 2

L8(27) linear graph

1
7

3 5

2 4
6

2
3
5
1 4
6

L9(34) linear graph

3,4
1 2
 Linear Graphs
Assigning Factors to Linear Graphs
Example:
An experiment with five 2-level factors A, B, C, D, and E
is to be designed. In addition to main effects, it is also
desired to study interactions CxD and CxE.

Solution:
Select the orthogonal array.
A. Calculate the degrees of freedom required.
Factors/Interactions degree of freedom
A 1
B 1
C 1
D 1
E 1
CxD 1
CxE 1
-----------
7

B. Select the smallest array that contains the number of

degree of freedom.
From the previous example, L8 has 7 degrees of
freedom. Therefore, choose L8.
 Linear Graphs

Solution (continued):
Draw the required linear graph.
D

C
A B

Match the required linear graph to a standard linear graph.

1
7
C
3 5

2 4
D 6 E

Modify or redraw the required linear graph.

1
7 6
C
3 5 A B

2 4
D 6 E
 Linear Graphs

Solution (continued):
Assign factors and interactions to columns.

L8
Column Number
C D CxD E CxE B A
Run No.\ 1 2 3 4 5 6 7
1 1 1 1 1 1 1 1
2 1 1 1 2 2 2 2
3 1 2 2 1 1 2 2
4 1 2 2 2 2 1 1
5 2 1 2 1 2 1 2
6 2 1 2 2 1 2 1
7 2 2 1 1 2 2 1
8 2 2 1 2 1 1 2

Problem : Design an experiment with four 2-level factors: A,

B, C & D. Interactions of interest are AxB and BxC .
 The Outer Array
Control Factors - are product design and process factors that
influence product performance, whose values can be
controlled by the manufacturer.
Examples : heater temperature, cure time, pressure, conveyor
speed, % solution, dwell time, pH, acid concentration.

Noise Factors - are other variables that affect product

performance, whose values cannot be controlled by the
manufacturer or are not controlled for economic reasons
Examples : weather, ambient temperature, operating
environment, moisture content, material properties, cleanliness,
machine number, machine type.

Robust Design
• Products should be designed so that they operate even
when environmental conditions vary widely. Likewise,
processes should be designed so they are insensitive to
noise factors.
• This can be done using experimental design with inner and
outer arrays
Noise Factors
k . . . .
. . . . .
. . . . .
O 1 . . .
Control factors and Interactions N 1 . . .
Run A B C D . . i M 1 . . .
No.
1 . . . . . . .
2 . . . . . . . Experiment data
. . . . . . . .
j . . . . . . .
 The Outer Array
Example
Noise Factors
O 1 2 2 1
Control factors and Interactions N 1 2 1 2
Run A B AxB C e BxC D M 1 1 2 2
No.
1 1 1 1 1 1 1 1 2.62 2.66 10.00 10.00
2.56 2.56 8.24 10.00
2 1 1 1 2 2 2 2 2.27 2.29 5.65 6.07
2.30 2.29 10.00 10.00
3 1 2 2 1 1 2 2 2.30 2.27 3.73 4.19
2.33 2.33 8.09 10.00
4 1 2 2 2 2 1 1 2.61 2.53 6.36 8.10
2.61 2.59 10.00 8.14
5 2 1 2 1 2 1 2 2.36 2.29 2.28 4.24
2.29 2.25 4.22 2.25
6 2 1 2 2 1 2 1 2.54 2.58 6.51 6.33
2.46 2.47 10.00 6.22
7 2 2 1 1 2 2 1 2.53 2.54 4.40 4.51
2.51 2.52 6.25 8.12
8 2 2 1 2 1 1 2 2.30 2.27 10.00 4.27
2.21 2.22 2.20 8.05

Y = Switching time (milliseconds)

Control Factors Level 1 Level 2

A. Package type Type 1 Type 2
B. Burn in No Yes
C. Component Supplier Vendor X Vendor Y
Noise Factors
M. Moisture Low High
N. Drying Normal Heat Gun
O. Soldering Hand Solder Wave Solder
 Marginal Means Analysis
Digital Circuit (Larger-the-Better) Case Study
Problem : A design engineering team is working together to
develop a digital circuit that will reliably operate upon receiving
the input signal.
Objective : To maximize the mean time between failures.
Quality Characteristic, Y = mean time between failures
Control Factors Level 1 Level 2
A. Input Signal 200 nsec 400 nsec
B. Clock Frequency 10 mhz 25 mhz
C. Flip Flop LS AS
D. Synchronizers 1 2
E. Capacitive Loading 15 pf 50 pf
Interactions
Input signal x clock frequency (AxB)
Clock frequency x flip flop (BxC)

Control factors and Interactions

Run A B AxB C D BxC E Data1 Data2 Mean
No. Response
1 1 1 1 1 1 1 1 75 71
2 1 1 1 2 2 2 2 68 65
3 1 2 2 1 1 2 2 18 15
4 1 2 2 2 2 1 1 22 23
5 2 1 2 1 2 1 2 45 47
6 2 1 2 2 1 2 1 12 9
7 2 2 1 1 2 2 1 25 28
8 2 2 1 2 1 1 2 2 3
 Marginal Means Analysis
Digital Circuit Case Study (continued)
Calculations :
1. Determine the mean for each row (experimental run)
Run#1 y1 = (75 +71)/2 = 73.0
Run#2 y2 = (68 +65)/2 = 66.5
…………………………………..
Run#8 y8 = (2 + 3)/2 = 2.5

Control factors and Interactions

Run A B AxB C D BxC E Data1 Data2 Mean
No. Response
1 1 1 1 1 1 1 1 75 71 73.0
2 1 1 1 2 2 2 2 68 65 66.5
3 1 2 2 1 1 2 2 18 15 16.5
4 1 2 2 2 2 1 1 22 23 22.5
5 2 1 2 1 2 1 2 45 47 46.0
6 2 1 2 2 1 2 1 12 9 10.5
7 2 2 1 1 2 2 1 25 28 26.5
8 2 2 1 2 1 1 2 2 3 2.5

2. Calculate the mean response for each factor and interaction

level and develop a response table. This is performed by
grouping the mean responses by factor level for each column
in the array, taking the sum, and dividing by the number of
responses. The absolute difference or delta between the two
average results is the effect of the factor or interaction. The
factors and interactions are then ranked from highest to
lowest delta.
Response Table
A B AxB C D BxC E
Level 1 44.625 49.000 42.125 40.500 25.625 36.000 33.125
Level 2 21.375 17.000 23.875 25.500 40.375 30.000 32.875
Delta 23.500 32.000 18.250 15.000 14.750 6.000 0.250
Rank 2 1 3 4 5 6 7
 Marginal Means Analysis
Digital Circuit Case Study (continued)
Calculations :
3. Based on the response table, construct the means response
graph.

Marginal Means Response Graph

60.000

50.000

A
40.000 B
AxB
30.000 C
D
20.000 BxC
E

10.000

0.000
1 2 1 2 1 2 1 2 1 2 1 2 1 2

4. Analyze the means response table and means response graphs

to determine those factors that have a strong effect on the
quality characteristic of interest.
Since the case is larger-the better, select the level that
resulted in the larger average response value.
A, B, AxB, C, and D were determined to have strong effects,
recommended levels are A1, B1, A1B1, C1, and D2.
 Marginal Means Analysis
Digital Circuit Case Study (continued)
4. Compute an estimate of the predicted response based on the
selected levels of the strong effects. The prediction equation is
as follows:
 = T + (A1- T) + (B1 -T) +[(A1B1- T) - (A1- T) - (B1- T)] +
(C1- T) + (D2 - T)
 = A1B1 + C1 + D2 - 2T

Where T =  y/n
= (73.0 + 66.5 + 16.5 + 22.5 + 46.0 + 10.5 +
26.5 + 2.5)/8
= 33.0
 = 69.750 + 40.500 + 40.375 - 2 (33.000)
= 84.625
5. Conduct a confirmation run and compare the actual result to
the predicted results. If the actual results are
disappointingly low, the team will need to return to the
planning phase and consider the possible causes.
 Analysis Using Signal-to-Noise Ratio

Signal-to-Noise (S/N) ratio

• A measurement tool for determining robustness
• The signal on the numerator is interpreted as the ability of
the process to build good product or of the product to perform
correctly.
• By including the impact of the noise factors on the process or
product as the denominator, the signal-to-noise ratio can be
used as a barometer of the ability of the system to perform
well in relation to the effect of noise.

Taguchi developed several S/N equations to be applied in

different cases:
For the case of smaller-the-better

 y12  y22  ...  yn2 

S / N  10 log 
 n 

For the case of larger-the-better

 1 1 1 
 2  2  ...  2 
S / N  10 log  y1 y2 yn 
 n 
 
 
 Analysis Using Signal-to-Noise Ratio

Signal-to-Noise (S/N) ratio (continued)

For the case of Nominal-is-Best

1 (Sm  Ve)
S/N  10log
n Ve

T 2 y i
Sm   i 1
n n

( y12  y22  ...  yn2 )  Sm

Ve 
n 1
 Analysis Using Signal-to-Noise Ratio
Electronic Circuit “Time Out”(Smaller-the-Better) Case Study
Problem :For a specific control device, the time required for the
circuit to switch from “On” to “Off” when the L.E.D. drive
current is removed is critical. The specification requirement
states that the time between changing states can be no more
than 0.005 seconds. However, functional testing of current
production reveals a large percentage of units have not been in
compliance with the timing specification.
Objective : To minimize switching time.
Quality Characteristic, Y = switching time (milliseconds)
Control Factors Level 1 Level 2
A. Package type Type 1 Type 2
B. Burn in No Yes
C. Test Delay No Delay Delay
D. Component Supplier Vendor X Vendor Y
Interactions
Package type x Burn In (AxB)
Burn In x Test Delay (BxC)
Noise Factors Level 1 Level 2
M. Moisture Low High
N. Drying Normal Heat Gun
O. Soldering Hand Solder Wave Solder
 Analysis Using Signal-to-Noise Ratio
Electronic Circuit “Time Out”Case Study (continued)

Noise Factors
O 1 2 2 1
Control factors and Interactions N 1 2 1 2
Run A B AxB C e BxC D M 1 1 2 2
No.
1 1 1 1 1 1 1 1 2.59 2.61 9.12 10.00
2 1 1 1 2 2 2 2 2.30 2.29 7.82 8.00
3 1 2 2 1 1 2 2 2.31 2.30 5.91 7.10
4 1 2 2 2 2 1 1 2.61 2.56 8.18 8.12
5 2 1 2 1 2 1 2 2.32 2.27 3.25 3,24
6 2 1 2 2 1 2 1 2.50 2.52 8.25 6.28
7 2 2 1 1 2 2 1 2.52 2.53 5.32 6.32
8 2 2 1 2 1 1 2 2.26 2.25 6.10 6.16

1. Determine the S/N for each row.

S/N#1 = -10 log (2.592 +2.612 +9.122 +102)/(4) = -16.917 db
S/N#2 = -10 log (2.302 +2.292 +7.822+ 82 )/(4) = -15.305 db
……………………………………………………………………..
S/N#8 = -10 log (2.262 +2.252 +6.102+6.162)/(4) = -13.290 db
Noise Factors
O 1 2 2 1
Control factors and Interactions N 1 2 1 2
Run A B AxB C e BxC D M 1 1 2 2 S/N
No.
1 1 1 1 1 1 1 1 2.59 2.61 9.12 10.00 -16.917
2 1 1 1 2 2 2 2 2.30 2.29 7.82 8.00 -15.305
3 1 2 2 1 1 2 2 2.31 2.30 5.91 7.10 -13.800
4 1 2 2 2 2 1 1 2.61 2.56 8.18 8.12 -15.629
5 2 1 2 1 2 1 2 2.32 2.27 3.25 3.24 - 8.976
6 2 1 2 2 1 2 1 2.50 2.52 8.25 6.28 -14.775
7 2 2 1 1 2 2 1 2.52 2.53 5.32 6.32 -13.064
8 2 2 1 2 1 1 2 2.26 2.25 6.10 6.16 -13.290
 Analysis Using Signal-to-Noise Ratio
Electronic Circuit “Time Out”Case Study (continued)
2. Calculate the average S/N for each factor and interaction level
and develop a response table.
S/N Response Table
A B AxB C e BxC D
Level 1 -15.413 -13.993 -14.644 -13.189 -14.696 -13.703 -15.096
Level 2 -12.526 -13.946 -13.295 -14.750 -13.243 -14.236 -12.843
Delta 2.887 0.047 1.349 1.561 1.453 0.533 2.253
Rank 1 7 5 3 4 6 2

3. Construct the S/N response graphs.

S/N Response Graphs

-10.000
1 2 1 2 1 2 1 2 1 2 1 2 1 2

-11.000

A
-12.000
B
AxB
-13.000 C
e
BxC
-14.000
D

-15.000

-16.000
 Analysis Using Signal-to-Noise Ratio
Electronic Circuit “Time Out”Case Study (continued)
From the S/N response graph, strong effects come from A2, D2,
C1, and some level of AxB. To determine which levels of AxB
gives a strong effect, construct a S/N interaction matrix and a
S/N interaction graph of AxB.
A1B1 = ((-16.917)+(-15.305))/2 = -16.111
A1B2 = ((-13.800)+(-15.629))/2 = -14.714
A2B1 = ((- 8.976)+ (-14.775))/2 = -18.876
A2B2 = ((-13.064)+(-13.290))/2 = -13.177

Interaction Matrix
B1 B2
A1 -16.111 -14.714
A2 -18.876 -13.177

AxB Interaction Graph

-10
-11 B1 B2
-12
-13
-14
A1
-15
A2
-16
-17
-18
-19
-20

From the graph, A2B2 is recommended.

 Analysis Using Signal-to-Noise Ratio
Electronic Circuit “Time Out”Case Study (continued)
4. Determine the average response for each row.
Noise Factors
O 1 2 2 1
Control factors and Interactions N 1 2 1 2 Mean
Run A B AxB C e BxC D M 1 1 2 2 Response
No.
1 1 1 1 1 1 1 1 2.59 2.61 9.12 10.00 6.080
2 1 1 1 2 2 2 2 2.30 2.29 7.82 8.00 5.102
3 1 2 2 1 1 2 2 2.31 2.30 5.91 7.10 4.405
4 1 2 2 2 2 1 1 2.61 2.56 8.18 8.12 5.368
5 2 1 2 1 2 1 2 2.32 2.27 3.25 3.24 2.770
6 2 1 2 2 1 2 1 2.50 2.52 8.25 6.28 4.888
7 2 2 1 1 2 2 1 2.52 2.53 5.32 6.32 4.172
8 2 2 1 2 1 1 2 2.26 2.25 6.10 6.16 4.192

5. Develop an average response table.

Mean Response Table
A B AxB C E BxC D
Level 1 5.239 4.710 4.887 4.357 4.891 4.603 5.127
Level 2 4.006 4.534 4.358 4.888 4.353 4.642 4.118
Delta 1.233 0.176 0.529 0.531 0.538 0.039 1.009
Rank 1 6 5 4 3 7 2

6. Construct the mean response graphs.

Marginal Means Response Graph

6.000

A
5.000 B
AxB
C
e
4.000 BxC
D

3.000
1 2 1 2 1 2 1 2 1 2 1 2 1 2
 Analysis Using Signal-to-Noise Ratio
Electronic Circuit “Time Out”Case Study (continued)
7. Analyze the graphs and compare with previous S/N results.
The graph shows strong effects from A and D. The following
and levels are recommended: A2, D2, B2, and C1. These are
basically the same recommendations from the S/N response
graphs.

Marginal Means Response Graph

6.000

A
5.000 B
AxB
C
e
4.000 BxC
D

3.000
1 2 1 2 1 2 1 2 1 2 1 2 1 2