A report on

“Design of experiment using Taguchhi”

Prepared By:-
Nirmal Kumar Kushwaha (P10ME140)
Vishal B. Dixit (P10ME141)

Guided By:-
B. K. Murali
Professor in Mechanical Engineering Depart
SVNIT, SURAT
1
 Acknowledegement

Every organization work has imprint of many people and this work is no
different. This work is gives me an opportunity to express deep gratitude for the same.
While preparing report we received endless help from number of people. This
report would be incomplete if we don’t convey our sincere thanks to all those who were
involved.
First and foremost we would like to thank our H.O.D Dr. H.B.Naik, Department
of Mechanical Engineering, SVNIT, Surat for giving us an opportunity to prepare this
Report and his indispensable support.
We would like to thank our respected Guide Dr.B.K.Murali for his valuable
guidance, priceless suggestions, and moral support throughout this entire work.
Finally, we wish to thank our friends and our family to being supportive for us,
without whom this work would not have seen the light of day.
Every work is an outcome of full proof planning, continuous and organized
effort. This work is combination of all the three put together sincerely.

Yours Sincerely,
Nirmal Kumar Kushwaha
Vishal B. Dixit

2
 Index

S. No Contents Page No.

1. Introduction 3

2. Taguchi Loss Function 4

3. Design of Experiment Process 4

4. Orthogonal array 7

5. Conducting Tests 8

6. Analysis and interpretation methods for experiments 8

7. Confirmation experiment 12

8. Case Study - Electrostatic Powder Coating Process 14

Optimization

9. Summary observations and conclusions 15

10. Conclusion 19

11. Reference 21

3
1. Introduction:

Taguchi addresses quality in two main areas: off-line and on-line quality control.
Both of these areas are very cost sensitive in the decisions that are made with respect to
activities in each. Off-line QC refers to improvement of quality in the product and
process development stages. On-line QC refers to monitoring of current manufacturing
process to verify quality level produced.
Products have characteristics that describe their performance relative to customer
requirements or expectations. The quality of a product is measured in terms of these
characteristics. Quality is related to loss to society caused by a product during its life
cycle. A truly high quality product will have minimal loss to society as it goes through
this life cycle. The loss a customer sustains can take many forms, but it is generally a loss
of product function or properties. Other losses are time, pollution, noise etc.
The Goalpost philosophy supports the position that a product made according to
the print, within permitted tolerance, is of high quality. This strict viewpoint embraces
only the designers and the makers. But what are missing from this philosophy are the
customer’s requirements. A product may meet print specifications, but if the print does
not meet customer requirements then true quality cannot be present. Customers want the
product close to nominal all the time and producers want to allow the product to vary to
limit of specifications. These seemingly incongruent ideas can be brought into harmony
by means of “Taguchi Loss Function”.

2 . Taguchi Loss Function:

The Taguchi loss function recognizes the customer’s desire to have products that
are more consistent, part to part, and the producer’s desire to make a low cost product.
The loss to society is composed of the costs incurred in the production process as well as
the costs encountered during use by the customer (repair, lost business, etc.). To
minimize the loss to society is the strategy that will encourage uniform products and
reduce costs at the point of production and at the point of consumption.
The Taguchi loss function quantifies the variability present in a process. If a part
reaches the end of manufacturing line with a diameter exceeding the upper or lower limit,

4
the part should be scrapped at a certain cost. The scrap cost is only one aspect of loss to
society. Presumably, the specifications are related to the reliability of the product; as the
specification limits are approached, the product is less likely to provide satisfaction to the
customer. If the product fails to perform satisfactorily, then the other losses are incurred
by the manufacturer or the customer, which makes scrap loss a conservative (low)
estimate of loss to society.
Using this cost as reference value, a loss function can be constructed for this
situation as shown in following figure.
Fig. no. 1- Taguchi loss function

Loss function
Scrap Loss

LSL O USL

Taguchi uses the mathematical equation to model this picture of cost versus process
parameter specification.
L= k(y-m) 2 ----------- (1)
In this equation,
L = the loss associated with a particular value of y
y = particular value of specification within permitted value.
m= nominal value of specification
k = constant depending on the cost at the specification limit and width
Of specification

5
3. Design of Experiment Process:
There are twelve steps in DOE process
Step1. State the problem or area of concern:
What is needed to perform this step:
Data that characterizes the problem as it occurs
How the problem is observed
When the problem occurs
How severe the problem is
Where the problem is occurs
Step 2.State the objective of the experiment
What is needed to perform this step:
A problem statement
Competitive benchmark information concerning the problem
Customer information concerning the problem
Step 3.Select the quality characteristic and measurement system:
What is needed to perform this step:
A problem statement
An experimental objective
Step 4.Select the factors that may influence the select quality:
What is needed to perform this step:
A problem statement
Product design specifications
Product control plans
Process flow diagrams
Process routings
Statistical process control chart results

Step 5. Identify control and noise factors:

What is needed to perform this step:
A list of factors to evaluate for their effect on quality characteristic of interest

6
Step 6.Select the levels for the factors:
What is needed to perform this step:
A list of control and noise factor
Product or process technical expertise
Product or process specification or operating limits
Table no.1
factors Level 1 Level 2
A cover design Production New
B gasket design Production New
Water pump leak C front bolt torque LSL USL
experiment D sealant No Yes
E pump finish Rough Smooth
F back bolt torque LSL USL
G torque sequence Front- back Back- front

Step 7.select the appropriate orthogonal array:

What is needed to perform this step:
The no. of control and noise factors
The no. of levels for the specific factors
The experimental resolution desired
Step 8.Select interactions that may influence the selected quality
characteristics or go back to step 4(iterative steps):
What is needed to perform this step:
The OA column numbers which have no factors assigned to them
Step 9. Assign factors to OA(s) and locate interactions:
What is needed to perform this step:
A list of selected control and noise factors
A list of interactions that may be of interest
The no. of levels for the factors
The existence or nonexistence of nesting factors
The existence or nonexistence of combined factors
The existence or nonexistence of idle column factors

7
Step 10. Conduct tests described by trials in OA(s):
What is needed to perform this step?
An assignment of the selected factors and levels to an OA or OAs
Values for all the levels of the factors
Trial data sheets determination of the sample size to be used
A randomization strategy
A material logistics strategy
Test equipment availability
Step 11. Analyze results of the experimental trials:
What is needed to perform this step
A complete set of test results for each and every trail of the experiment and the
test order
of the trails or repetitions
Step 12.conduct confirmation experiment:
What is needed to perform this step:
Identification of statistically significant factors and the levels those are most
desirable from the technical viewpoint.
Identification of statistically significant factors and the levels those are most
economical.

8
4. Orthogonal array:

Orthogonal arrays used to determine the combination of factors and levels which
will provide the experimenter with the desired information. One approach is to utilize a
fractional factorial approach whenever there are several factors involved, and this may be
accomplished with the aid of orthogonal arrays. Orthogonal arrays are introduced from
the viewpoint of pragmatist who is always trying to make a product or process
improvement decision with a minimum amount of test data. Using a minimal amount of
test data is not a necessarily problem in itself; however the considerations of what may
make up a valid experiment from a risk viewpoint are seldom considered by the typical
experimenter.
Before the discussion of orthogonal arrays, it would be better to review some
often used test strategies. Not being aware of efficient, proper test strategies experimenter
resort to following approaches.
1) One factor experiment:-
The one factor experiment evaluates the effect of one parameter on performance
while ostensibly holding everything else constant. For example, the effect of cutting
speed on the microfinish of machined part. Two different cutting speeds could be used.
First cutting speed is symbolized as level 1 and other as a level 2.
Table no.2- One factor experiment

Trial no. Factor level Test result

1 1 **

2 2 **
The * symbolize the value for microfinish

If there happens to be an interaction of factor studied with some other factor, then this
interaction cannot possibly observed. Also, a one factor experiment doesn’t use the test
data in an effective manner.

9
2) Several factors, one at a time:-
If the first factor chosen fails to produce the hoped results, the person usually
resorts to testing other factor, & the resultant test program would be like below.
Table no.3- Several factors, one at a time
Factor & factor level
Trial No. Test result
A B C D
1 1 1 1 1 **
2 2 1 1 1 **
3 1 2 1 1 **
4 1 1 2 1 **
5 1 1 1 2 **
Main limitation of this is that experiment will not be orthogonal. Orthogonality
means that factors can be evaluated independently. For instance, the nonorthogonality of
the data set in table can be seen. If all the data under level A1 is averaged and all the data
under level A2 is averaged and compared, this is not a fair comparison of A1 to A2. Of
the four trials under level A1, three were at level B1 and one at level B2. The one trial
under level A2 was at level B1. Therefore one can see that if factor B has an effect on
performance it will be part of the observed effect factor A and vice versa.
3) Several factors, all the same time:-
Table no.4- Several factors, all the same time
Factor and factor level
Trial no. Test results
A B C D

1 1 1 1 1 **
**
2 2 2 2 2

This situation makes separation of the main factor effects impossible. Some
factors may be making a positive contribution and others negative contribution, but no
hint of this fact will exist.

10
4) Full factorial experiment (Better Test Strategy):-
Table no.5- Full factorial experiment
Factor and factor level
Trial no. Test results
A B

1 1 1 **

2 1 2 **

3 2 1 **

4 2 2 **

The poor utilization of test data can be avoided by using full factorial method.
One can see that the full factorial experiment is orthogonal in this case. There is an equal
no. of test data points under each level of each factor. Note that under level A1 factor B
has two data points under B1 condition and under B2 condition. The same is true under
level A2. The same balanced situation is true when looking at the experiment with
respect to the two conditions of B1 and B2. Because of this balanced arrangement, factor
A does not influence on factor B and vice versa.
One can see that all possible combinations of the two factors and the two levels
are represented in the preceding test matrix. Using this information, both the factor and
interaction effect can be estimated. A full factorial experiment is acceptable when only a
few factorials are to be investigated, but not very acceptable when there are many factors.
If a full factorial experiment is used, there is a minimum of 2f possible combinations that
must be tested (f = the no. of factors each at two levels)
E.g. as we see water pump leak problem in DOE process steps and the list of the seven
factors at two levels in respective table was an actual experiment performed at an engine
plant. I f a full factorial experiment were to be used in this situation, and then total no. of
tests will be 27 = 128 tests must be conducted. This type of experiment, shown in fig.
no.2

11
 Fig.no.2 Full factorial experiment
A1 A2
B1 B2 B1 B2

C1 C2 C1 C2 C1 C2 C1 C2

D1 D2 D1 D2 D1 D2 D1 D2 D1 D2 D1 D2 D1 D2 D1 D2
G1

F1 G2
E1
G1

F2 G2

G1

F1
G2
E2
G1

F2 G2

This type of experiment estimates all the main factor effects and all the possible
interactions, all orthogonal to one another. However, usual time and financial limitations
preclude the use of full factorial experiment.
5) Fractional factorial experiments (Efficient Test Strategy):-
Statisticians have developed more efficient (economically) test plans, which are
referred as fractional factorial experiments (FFEs). FFEs are only portion of total possible
combinations to estimate the main factor effects and some not all, of interactions. There
are 1/2 FFE, a 1/4 FFE, a 1/8 FFE, a 1/16 FFE etc. Certain treatment conditions are
chosen to maintain the Orthogonality among the various factors and the interactions.
The information generated in such a small experiment is, however, substantially reduced
from that of a full-factorial experiment. For our case a 1/2 FFE would give 64 test
combinations. It is obvious that 1/16 FFE with only 8 test combinations as shown in
following figure, is much more appealing to experimenter from a time and cost
standpoint.

12
 Fig no.3-1/16 FFE

A1 A2
B1 B2 B1 B2
C1 C2 C1 C2 C1 C2 C1 C2

D1 D2 D1 D2 D1 D2 D1 D2 D1 D2 D1 D2 D1 D2 D1 D2

G1

F1 G2
E1
G1

F2 G2

G1

F1
G2
E2
G1

F2 G2

Taguchi has developed a family of FFE matrices which can be utilized in various
situations. In this situation, one possible matrix is an eight trial OA, which as labeled as
an L8 matrix. An L8, two level matrixes is shown in following table.

Table no.6- L8 OA Matrix

Column No.
Trial 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

13
 Actually this is a 1/16 FFE, equivalent to fig no.3, which has only 8 of the
possible 128 combinations represented. When water pump factors A through G are
assigned to column 1 through 7, there are eight unique pump assemblies described by all
different trials. The eight trial combinations match the eight descriptions in the 1/16 FFE
in fig no.
These are two different ways to describe exactly the same experiment; however, with the
use of OA approach it is much easier to determine the appropriate orthogonal
combinations and to perform analysis. It is easy to see that all columns in al L8 OA
provide four tests under the first level of factor and four tests under second level of the
factor. This is one of the features that provide the Orthogonality among all the columns
(factors). The real power in using an OA is the ability to evaluate several factors in a
minimum of tests. This is considered an efficient experiment since much more
information about factor is obtained from a few trials.
Selection of Orthogonal Arrays:
The selection of which OA to use predominantly depends on these items in order
of priority:
1) The number of factors and interaction of interest.
2) The number of levels for the factors of interest.
3) The desired experimental resolution or cost limitations.
The first two items determine the smallest orthogonal array that it is possible to use, but
this will automatically be the lowest resolution, lowest cost experiment. The
experimenter may choose to run a larger experiment (Larger Orthogonal array) which
will have high resolution potential but will also be more expensive to complete.
Two basic kinds of OAs are following.
Two level arrays:
L4 L8 L12 L16 L32
Three level arrays:
L9 L18 L27
The number in the array designation indicates the number of trials (different possible test
combinations) in the array; an L8 has eight trials and L27 has 27 trials. The 1s, 2s, and 3s
within the trials of the OAs designates the appropriate level of the factor assigned to that

14
column to be used for that specific trial. The number of levels used in the factor should be
used to select either two level or three level types of OAs. These are the standard OAs
have a fixed format.
The following tables are “design and analysis tables”, which lists many of the
option that is available in two level situations or three level situations. Referring to fig.
no. 4, the numbers of two level factors are listed across the top of table and the possible
down the left hand side. Looking below the number of two level factors that may be
under consideration for given experiment, the possible OAs are indicated along with
maximum resolution possible. In screening (beginning round) experiments, the
recommended strategy is to start with smallest possible OA that will accommodate
typically large number of factors under evaluation. This means that the resolution will be
low in first round experiments (a small fractional factorial) and will progress to higher
resolution experiment (a large fractional or full factorial) as a few factors are identified as
influential. This strategy will minimize the total number of tests to be conducted yet will
yield meaningful information at the same time. If tests may be vary inexpensively
conducted, then the high-resolution, full-factorial experiment may be in order. Once the
appropriate OA has been selected, the factors can be assigned to various columns of the
array and the subsequent interaction columns located.
Fig. no.4- Two level orthogonal selection
The no. of two level factors*
OA 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
L4

L8

L16

L32

L64

L128

L256

15
*Consider each three or four level factor equivalent to 3 two-level factors; for example, 1
four level factor and 5 two-level factors would have to use an L16 because that is
equivalent to a total of 8 two-level factors.
**Resolution number is measure of the amount of confounding in a column
4 = all items are in a separate column (full factorial)
3 = A and B C D E, or A B and C D E are in the same column
2 = A and B C D or A B and C D are in same column
1 = A and B C are in same column
Fig. no.5- Three level orthogonal array selection
Number of three level factors
OA 1 2 3 4 5 6 7 8
L9 4* 1 Not possible
L18 1
L27 4 2 1

* Resolution number is measure of the amount of confounding in a column

Interaction effect of factors:
Pairs of factors (A and B) in an experiment may interact with one another to
provide synergistic effect on a quality characteristic being studied. When a two-factor
experiment is conducted, there are three items that may statistically estimate:
1) Factor A’s overall effect to change the result
2) Factor B’s overall effect to change the result
3) The interaction effect of factors A and B to change the result
Statistically, these are treated as three separate items which may have their individual
strengths estimated. As a general recommendation, it is preferred to study more factor
than to study interaction.
Assignment of factors and location of interactions:
OAs has several columns available for the assignment for the factors and some
column will, subsequently estimate the effect of interaction of those factors. If one factor
is assigned to any particular column in a two-level array and second factor is assigned to
any other particular column, a specific third column will be automatically have the
interaction of those factors assigned to that column. The strength of factor A will be

16
evaluated using one column, the strength of factor B will be evaluated using second
column, and the strength of the interaction of factor A and B will be evaluated using a
specific third column. The pattern of which columns will be interaction column is known
for all of the orthogonal arrays.
Taguchi has provided two tools to aid in the assignment of factors to arrays and
location of interactions in arrays:
1) Interaction tables
2) Linear graphs
Each OA has an interaction table and a particular set of linear graphs associates with it.
The interaction tables contain all the possible interaction between factors (columns). The
linear graphs indicate various columns to which factors may be assigned and which
columns subsequently evaluate the interaction of those factors. The linear graphs are
simply visual representation of a portion of interaction table. The following table shows
the some part of two-level interaction table.

Table no.7- Two-level interaction table

Column no.

Column no. 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 - - - - - -

Let us consider simplest OA, an L4, has an interaction table, which is derived
from above table. Recall that the L4 OA has four trial and three columns. The first factor
assigned to an OA may actually be placed in any column, say column 2. The second
factor may be assigned to any column, say column 3, then the interaction table indicates
that the A×B interaction will be in column 1, as shown in two-level interaction table.

17
 Table no.8-Interaction table for L4

The interaction table shows that the three columns are mutually interactive. Any
assignment of factor A and B is mathematically and statistically equivalent. The
interaction table for an L4 is actually a small portion of a larger interaction table that will
handle up to an L32OA (except for an L12 OA which does not have specific interaction
columns).
An L4 OA has linear graph that appears like below.

The linear graph indicates that factor A may be assigned to column 1, factor B to
column 2, and the A × B interaction subsequently located in column 3. The dot represents
a column available for a two-level factor and the line represents a column which will
evaluate the interaction of the factors assigned to respective dots.
Resolution of experiments:
Resolution number is a measure of the amount of confounding (mixing) in the
column. Obviously, many interactions are confounded with the main effects. This is a
major compromise of using FFEs-to reduce the number of tests, some information must
be surrendered.
An L8 OA provides a complex enough array to demonstrate the amount of
confounding that may occur in an experiment. If two factors are assigned to an L8,
columns 1 and 2 are typically used for A and B and has the A×B interaction
automatically located in column 3. If factor C id added, the assignment of factor C should
be to column 4, which is the next truly empty column. According to interaction table for
two-level OAs, column 5 will have the A×C interaction, column 6 will have the B×C
interaction, and the column 7 will have the A×B×C interaction. In this situation, all main

18
effects and all interaction can be estimated, which results in a high resolution experiment.
Resolution power indicates the clarity with which individual effects of factors and
interactions may be evaluated in an experiment. Table no.7, indicates this experiment to
be resolution 4, which is also a full factorial experiment shown in following table.
Table No. 9- Resolution 4 experiment
Column number
1 2 3 4 5 6 7
A B A×B C A×C B×C A×B×C

If another factor D is added to this experiment, the most logical assignment is 7,

automatically confounding D with a three-factor interaction. Any other choice confounds
D two-factor interaction. The three-factor interaction is much less likely to occur, and if it
does, then it will most likely be smaller magnitude than main effect or a two-factor
interaction. Now the factors and interaction appear as in following table. The resolution
power of the experiment is subsequently much lower, resolution 2, because of greater
amount of confounding in the columns.
Table No. 10- Resolution 2 experiment
Column number
1 2 3 4 5 6 7
A B A×B C A×C B×C A×B×C
B×C×D A×C×D C×D A×B×D B×D A×D D

If a fifth factor is added, column 3, 5 and 6 are available; then the resolution
power will drop off further to a resolution 1 because of a factor being confounded with a
two-factor interaction. Since factor E was confounded with interaction A×B and C×D,
the three factor interactions A×B×E and C×D×E cannot be evaluated. This situation is
shown in following table.
Table No. 11- Resolution 1 experiment
Column number
1 2 3 4 5 6 7
A B A×B C A×C B×C A×B×C
B×C×D A×C×D C×D A×B×D B×D A×D D
B×E A×E E D×E A×D×E B×D×E C×E
B×C×E A×C×E

19
This confounding of groups of factor of two-factor interactions can be easily seen in an
L8 OA, but occurs at a much higher degree of complexity as more factors are assigned to
larger OAs. Taguchi does not place more emphasis on the confounding that exist in low-
resolution experiment; however it is good for experimenter to be aware of such situations.
Taguchi views interaction as being of minimal interest because to utilize the interactive
effect the experimenter must control two main effects. Since one or more main effects
usually need to be controlled for a product or process anyway, the interaction causes no
additional complications.
From a very practical point of view, all factors assigned to an experiment will not
be equally influential in changing the average response. Therefore what few factors are
significant will have been evaluated as if the experiment were full factorial. The
experimenter starts with what appears to be a low resolution experiment, but when the
trials are completed, the practical result will be high resolution experiment.

The following table lists the column assignment that should be used to provide the
highest resolution possible for two-level factors in standard OAs.

Table no.12-Two-level orthogonal array factor assignment

OA No. of factors Use column no. Resolution no.
L4 1-2 1,2 4 high
3 1-3 1 low
L8 1-3 1,2,4 4 high
4 1,2,4,7 2
5-7 1,2,4,7,(3,5,6) 1 low
L12 1-11 1-11 1 low
L16 1-4 1,2,4,8 4 high
5 1,2,4,8,15 3
6-8 1,2,4,7,8,(11,13,14) 2
9-15 1,2,4,7,8,11,13,14,(3,5,6,9,10,12,15) 1 low

Example: - (To choose OA and assigning factors to OA)

Water pump experiment- With 7 two-level factors in an L8 OA, there will be no empty
columns; in an L16 OA there will be eight columns that have no factors assigned. In the
L16 case, one more two level factor can be evaluated without decreasing the resolution
and more factors can be added without increasing the experiment size (resolution will
decrease, however). If additional factors are assigned to the OA, then specific interactions
20
may be identified as interesting to evaluate. With the seven recommended choice is to use
an OA to maintain a small experiment for the beginning round.
According to table no. 7 seven two-level factors in a low resolution situation
would use an L8 OA, and, in medium resolution situation, they would use an L16 OA.
The L8 OA is recommended as a beginning experiment to minimize the no. of tests, test
time and test cost. Since there are seven factors and seven columns in an L8 OA,
references to “factor assignment table” are not necessary; each factor is assigned to a
column. The column to which a factor is assigned is not important since all the
combinations of column assignments will be statistically equivalent.
Multiple-level experiments:
This is deals with modification of standard two-level arrays to handle factors of
three or four levels; for example, the steering wheel position in an automobile may be set
in left, centre, or right positions. However it is still recommended for an initial
experimental investigation to begin with two levels wherever possible. Once significant
factors are identified, multiple levels can be used for estimating nonlinear responses.
A two-level array can be converted to contain some four-level columns very
simply depends upon the concept of degrees of freedom which will be explained in
“analysis and interpretation method for experiments”. To statistically accommodate a
four-level factor in a two level OA, three two-level columns must be replaced with one
four-level column which will provide the same information potential. It is recommend
merging three columns which are mutually interactive, such as column 1, 2, and 3. The
merging of mutually interactive columns minimizes confounding of interaction as much
as possible and also maintains the Orthogonality of final array

If columns 1, 2, and 3 are merged in an L8 OA to form a single four-level column,

the eight trials need to have a particular level assigned to maintain Orthogonality of the
array. A recommended technique for accomplishing this is as follows. If any two
columns of first three columns of an L8 are studied, one will notice that there are four
combinations possible of levels 1 and 2.
11 12 21 22

21
Level 1 for the four-level corresponds to the 1 1 condition, level 2 to the 1 2 condition,
level 3 to the 2 1 condition, and the level 4 to the 2 2 condition. Using columns 1 and 2
from L8 OA, the corresponding levels would appear as follows.

Table no.13- Four level factor arrangement

Column no.
Trial no. 1 2 3 4-level factor
1 1 1 1 1
2 1 1 1 1
3 1 2 2 2
4 1 2 2 2
5 2 1 2 3
6 2 1 2 3
7 2 2 1 4
8 2 2 1 4

The three merged columns provide the necessary amount of information and any
two of those columns provide an orthogonal pattern for the four levels. The four-level
need not be in consecutive order with respect to the trial numbers. The resultant L8 array
modified to include one four-level factor and up to four two-level factors is shown in
following table.

Table no.14- L8 OA modified for a four-level factor

Column no.
Trial no. 1 2 3 4 5
1 1 1 1 1 1
2 1 2 2 2 2
3 2 1 1 2 2
4 2 2 2 1 1
5 3 1 2 1 2
6 3 2 1 2 1
7 4 1 2 2 1
8 4 2 1 1 2

The columns have been renumbered in this array. Many other combinations of the
four-level factor and the remaining two-level columns are possible. It is recommended to
use these modified OAs if the need should arise.

22
5. Conducting Tests:

Once the chosen factors are assigned to a particular column of the selected OA,
the test combinations are set and physical preparation for performing the tests can begin.
Main items to consider in the following logistics of testing are
1) A system to describe the various test combinations in operational terms (trial
description and data sheets)
2) A plan to acquire materials needed for the various test combinations
3) A plan to acquire access to test and production equipment
4) A plan to acquire access to measuring equipment
5) A definition of roles and responsibilities of those involved in conducting the
experimental trials
6) A plan for identifying the results of the experimental trials (tagging parts with
trial and repetition numbers, for example)
Statistical aspects of conducting tests:-
There are two main statistical considerations involved in conducting experiments:
1) The statistical valid sample size
2) The randomization strategy used to determine test order
1) Sample size determination:-
The type of quality characteristic under evaluation, variable or attribute, has
tremendous effect on the sample size required in the experiment. In general, variable data
will require considerably fewer experiments compared to attribute data to reach statistical
confidence that factors make a difference in the quality characteristics.
A] Variable data-
Variable data is a continuous form, which means that an infinite number of values
can occur anywhere between very low values and very high values. Examples of variable
data are temperature, pressure, power, weight, length etc. From a practical viewpoint, a
minimum of one test result for each trial is required to maintain sample size balance
(Orthogonality) of the experiment.
More than one test per trial can be used, which increases the sensitivity of the
experiment to detect the small changes in averages of populations. An economic
consideration also can be made at this time.
B] Attribute data-
Attribute data on the other hand, is a discontinuous form, which means the
experimental results can only be discrete values such as good or bad, or off and on. Two-
class attribute data provides much less discrimination than variable data. Two-class
attribute data can be of two types. One kind is a situation where the total number of tests,
the number of good results, and the number of bad results are known. Another kind of
attribute data is where the total number of tests is known and the number of one class of
occurrence may be known.

23
 To improve the discrimination power of attribute data, more classes may be used.
In this instance, the class number has some engineering and scientific meaning; the
higher the class number, the more severe effect. Multiple-class data has other advantage
over two-class data which concern the ability to detect a shift in average or to detect an
increase or decrease in variability.

2) Randomization strategy:-
Some decisions need to be made in concerning the order of testing the various
trials. The order of performing the tests of the various trials should include some form of
randomization. The randomized trial order protects the experimenter from any unknown
and uncontrolled factors that may vary during the entire experiment and which may
influence the results.
Randomization can take many forms, but the three most used approaches will be
following.
A] Complete randomization
B] Simple repetition
C] Complete randomization within blocks

A] Complete randomization-
Complete randomization means any test has an equal chance of being selected for
the first test. Of the remaining tests, each one has an equal opportunity of being selected
for the next tests, and so on. This method is used when a change of test setup is very easy
or inexpensive.

B] Simple repetition-
Simple repetition means that any trial has an equal opportunity of being selected
for the first test, but once that trial is selected; all the repetitions are tested for that trial.
This method is used if test setups are very difficult or expensive to change.

C] Complete randomization within blocks-

Complete randomization within blocks is used where one factor may be very
difficult or expensive to change the test setup for, but others are very easy. If factor A
were difficult to change, then the experiment could be completed in two halves or blocks.
All trials could be randomly selected and then all A2 trials could be randomly selected.

24
6.Analysis and interpretation methods for experiments:
Upon completion of the full array of experiments, we must analyze the results in order to
select the best design. The different designs are compared in terms of their signal-to-noise
ratio, or the S/N ratio. We shall not go into the complete details of the statistical analysis
in this course, but some insights are useful to understand how the method works.
We shall look at several methods, starting from the most basic to a somewhat analytical
method. For more complete details, you will need to study a course in design of
experiments (DOE).
I will use a few simple examples to explain these methods.
Example 1. Water Pump Design
The design of a water pump is being studied to set the optimum design parameters. The
evaluation criterion is water leakage during operation. The table below lists the factors
and the levels at which they are tested.

Table 15- Water pump design parameter levels.

factors Level 1 Level 2

A cover design Production New
B gasket design Production New
Water pump leak C front bolt torque LSL USL
experiment D sealant No Yes
E pump finish Rough Smooth
F back bolt torque LSL USL
G torque sequence Front- back Back- front

The outcome of the testing is measured in terms of the number of leaks in the pump
assembly. The tests are conducted using an L8(27) orthogonal array, and the raw results
are tabulated below. Leakage was rated on a scale from 0 Æ 5, where 0: no leaks, 5:
extremely high leaking.

25
 Table16- Experimental results on the water pump designs

Cover Gasket Front Sealant Surface Back Torque Test

Torque finish torque sequence leaking

Trial
No. A B C D E F G

1 1 1 1 1 1 1 1 4
2 1 1 1 2 2 2 2 3
3 1 2 2 1 1 2 2 1
4 1 2 2 2 2 1 1 0
5 2 1 2 1 2 1 2 2
6 2 1 2 2 1 2 1 4
7 2 2 1 1 2 2 1 0
8 2 2 1 2 1 1 2 1

Method 1. Observation method:

This is the simplest analysis one can perform. For example, looking at leakage data, it
appears design configurations 4 and 7 are the best (zero leakage).
On analysis, we see that these two designs have different levels for factor A (i.e. different
cover designs), and therefore the likelihood that cover design does not affect the
performance is increased.
Secondly, three columns are at the same level: [B: new gasket design, E: high surface
finish, and G: back-front torque sequence] in the best configurations. Therefore the
likelihood that these particular levels of these factors increase the design quality is high.
At this simple level of analysis, interaction effects are mostly ignored, and only some
basic idea of the design quality is obtained, which may be used to reduce the number of
variables in future experiments.

26
Method 2. Ranking method
We re-organize the data from the experimental runs in increasing order of the measured
statistic (Leaking intensity). Table 3 below shows this resulting data for our example.
Table 17- Ranking method
Cover Gasket Front Sealant Surface Back Torque Test
Torque finish torque sequence leaking

Trial
No. A B C D E F G

4 1 2 2 2 2 1 1 0
7 2 2 1 1 2 2 1 0
3 1 2 2 1 1 2 2 1
8 2 2 1 2 1 1 2 1
5 2 1 2 1 2 1 2 2
2 1 1 1 2 2 2 2 3
1 1 1 1 1 1 1 1 4
6 2 1 2 2 1 2 1 4

Note the following:0

(a) Parameter B, Gasket design, clearly has a strong affect on the design. The four best
designs all use the new gasket design, while the four worst use the existing gasket.

(b) If we partition the runs by gasket design, then factor E, or Surface finish, appears to
have consistent secondary affect on the performance (in the first four runs in Table ??,
Smooth surface finish pumps have lower leakage than rough; this trend is repeated in the
last four runs).

(c) From these observations, one may conclude that using a design with the new gasket
and smooth surface finish will result in a good design.

27
The observations of the ranking method can usually be identified by using first order
statistical data – statistics about each parameter, without considering interaction affects.
We call this the main effect due to a factor X, or ME( X). The easiest ME’s we can
compute are the mean values.
Let us denote ME1(X) and ME2(X) as the mean value of the output when the factor X
was at levels 1 and 2 respectively.

Thus:
ME1 (B) = (2+3+4+4)/4 = 3.25
ME2 (B) = (0+0+1+1)/4 = 0.5
For comparison, let us look at the main effects due to Factors A and E:
ME1 (A) = (4+3+1+0)/4 = 2
ME2 (A) = (2+4+0+1)/4 = 1.75
ME1 (E) = (1+1+4+4)/4 = 2.5
ME2 (E) = (0+0+2+3)/4 = 1.25
These are consistent with the conclusions above and also with the conclusions form the
observation method.

28
 Method 3. Column Effects Method
This method was suggested by Taguchi to perform a quick check on the relative
importance of each factor; the idea is to basically list out the ME1,2(X) for each parameter
(or for each column, if the column is assigned to study and interaction). In this sense, we
can look at this method as an extended form of the column ranking method.

Table 18-Column effects method (Taguchi)

Cover Gasket Front Sealant Surface Back Torque Test
Torque finish torque sequence leaking

Trial
No. A B C D E F G

4 1 2 2 2 2 1 1 0
7 2 2 1 1 2 2 1 0
3 1 2 2 1 1 2 2 1
8 2 2 1 2 1 1 2 1
5 2 1 2 1 2 1 2 2
2 1 1 1 2 2 2 2 3
1 1 1 1 1 1 1 1 4
6 2 1 2 2 1 2 1 4
ME1(X) 8 13 8 7 10 7 8

ME2(X) 7 2 7 8 5 8 7

ME2(X) - -1 -11 -1 1 -5 1 -1
ME1(X)

29
Method 4. Graphical method
The graphical method plots the ME’s. The simplest case involves plotting the ME for
each factor individually. This yields practically the same information as the previous
methods (Figure 1a). It is also possible to plot pair-wise interactions of factors. Assume
we want to plot the interaction of factors A and B, each of which can take two levels. We
compute the mean output for each combination of A and B. Then we can plot
isoparametric
lines – keep one factor constant, and draw the plot of how the output varies as
the other one changes. This is shown in Figure 1b. We get two plotted lines. If the two
lines are (nearly) parallel, then the likelihood is high that the two factors have little or no
interaction (why?).
From the table above, we have:
ME( B1E1) = (4+4)/2 = 4
ME( B1E2) = (3+2)/2 = 2.5
ME( B2E1) = (1+1)/2 = 1
ME( B2E2) = (0+0)/2 = 0

Figure 6- Graphical method for simple analysis of experimental data

30
Method 5. ANOVA (Analysis of Variations)
ANOVA is a statistical tool that allows us to analyze (or break down) the variations of
the test results into components that are contributed by the different sources. This allow
us to partition the total experiment variance to different factors, and even to combination
of different factors. Therefore ANOVA usage in Taguchi methods can be seen as a two
step procedure.
In the first stage, the different design test data points are used to compute the total
variance of the measured output. Further, the variance due to the individual factors, and
all possible combinations of factors that were studied are computed.
In the second stage, the variance due to any pair of factors (or combination) is
compare
By doing so, we can conclude which factor has more significant effect on the design
output. The measure of “relative significance” is based on a F-test, which gives us a
probabilistic evaluation, i.e. the probability that the two factors being compared have a
different enough affect on the output. When this probability is high enough (typically,
higher than 95%), we may conclude that the two factors have significantly different
affect. From this, we can then decide which design configuration is more robust.
In this section, we shall take a brief look at ANOVA basics. We begin with a simplified
example, and build up to the case for design evaluations.

Example 1. Assume that we have designed a water pump, and test it for performance
several times; the criterion is flow rate. The test results are:
Table 5. Simple example of ANOVA computations
Run 1 2 3 4 5 6 7 8
Flow rate 5 6 8 2 5 4 4 6
Let:
yi = i-th data value
N = total number of observations
T = total (sum) of all observations
T = average of all observations = T/N = y
For our case, N = 8, T = 40, T = 5.0

31
Since we are interested in variations about the mean (and in some cases, the variation of
the mean from zero), we further compute:
SST = total sum of squares = 52 + 62 + 82 + 22 + 52 +42 + 42 + 62 = 222
We denote (yi - y ) as the error, and therefore can compute the SSerror = SSe, as follows:
SSe = 02 + 12 + 32 + (-3)2 + 02 +(-1)2 + (-1)2 + 12 = 22
Further, we may write the square of the mean as T 2, and thus the sum-of-square of the
“error” due to deviation of mean from zero can be denoted as SSm = N(y ^2) = 8x 25 =
200.
Notice that SST = SSe + SSm
Further, we may think of the 8 observations to possess 7 degrees of freedom (since we
don’t really know the population mean, and therefore yield one degree of freedom to
compute the sample mean, y , form the 8 data points). You may think of this as the data
set having 7 dof, and the mean having one dof. This notation is useful mostly in
computing sample variances, which is the measure of interest for us.
We use the standard definition of variance, V2 = SS/v, where SS is the sum of squares of
the quantity of interest, and v is the degree of freedom of this quantity. We shall denote
V2 with the symbol V.
Thus, for our data, Ve = SSe/ve = 22/7 = 3.14

32
7. Confirmation experiment:

Steps in conducting a confirmation experiment are

1) Determine the preferred combination of levels of the factors and interactions
indicated to be significant by the analysis.
2) Determine the preferred levels for the factors indicated to be insignificant by the
analysis.
3) Calculate the estimated mean for the preferred combination of the levels of
significant factors and interactions.
4) Calculate the estimated standard deviation for the preferred combination of
significant factors and interactions.
5) Determine the sample size for the confirmation experiment.
6) Calculate the confidence interval value.
7) Calculate the confidence interval for the true mean around the estimated mean.
8) Conduct tests under specified conditions.
9) Compare the confirmation test avg. result to the confidence interval for the true
mean.
10) Determine the next course of action (use flow chart).

33
8 .Case Study - Electrostatic Powder Coating Process
Optimization
Project Descriptio
Process engineers and technicians in a high volume auto parts manufacturing
plant used DOE to optimize the Powder Coating Process for one of their motor housing
components.
Objective of experiment-
From the customer complaints and review of the production rejects, lack of
uniform powder thickness, voids, and bridging were considered major product
deficiencies. Powder-Coating Thickness among them was of primary concern and was
considered as the single objective of the current study. The coating thicknesses obtained
in current process range between 0.3 to 0.2 mil(1/1000 inch). A consistent higher
thickness was desired.
Identification of control and noise factors-
During the planning phase, the project team identified five factors and two
interactions from a larger list of possible control factors. In addition, the team members
also identified three uncontrollable (Noise) factors the variations of which are suspected
to adversely affect the coating process. The control factors and their levels included in the
preliminary study are as shown below.

Table 19 – Control factors

Column Factors Level 1 Level 2
no.
1 Voltage 75 KV 50 KV
2 Air box pressure 200 P S I 750 P S I
3 Interaction of 1×2 - -
4 Brushing rpm 350 rpm 750 rpm
5 Semi cure temp. 200 0 c 3000c
6 Interaction of 2×4 - -
7 Source of powder 3M Morton

34
The three noise factors were assigned two levels each and were incorporated in the study
are as shown below.
Table 20- Noise factors
Column Noise factors Level 1 Level 2
no.
1 Humidity 25 % 100%
2 Ambient Temperature 40 0 F 95 0 F
3 Ambient Pressure 28 in Hg 30 in Hg

Selection of OA and Outer Array-

An L-8 orthogonal array with seven 2-level columns was used for the experiment
design. The location of factors (columns 1, 2, 4, 5 and 7) and the columns reserved for
the two interactions (columns 3 and 6) are as indicated in the table above.
Table No. 21- an L8 array
Trial 1 2 3 4 5 6 7
no./Column no.
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

An L-4 outer array dictated that at least four separate samples be tested in each trial
condition exposing them to the combinations of the noise factors as prescribed by the
design.
Table no. 22- an L4 outer array
Trial 1 2 3
no./Column no.
1 1 1 1
2 1 2 2
3 2 1 2
4 2 2 1

35
S/N analysis-
One sample in each combination of the noise was tested under each trial
condition. This meant that there were four samples tested in each of the trial conditions.
The results from each sample and the Signal to Noise Ratios (S/N) for each trial results,
are as shown below. As there are multiple results in each trial condition, analysis using
S/N ratios was performed.

Table no. 23- S/N ratios

Trial no. R(1) R(2) R(3) R(4) S /N

1 0.21 0.23 0.19 0.22 13.5190

2 0.25 0.24 0.21 0.25 12.5549
3 0.19 0.21 0.18 0.20 14.2423
4 0.17 0.18 0.16 0.18 15.2954
5 0.24 0.22 0.23 0.25 12.6082
6 0.27 0.26 0.24 0.25 11.8943
7 0.26 0.26 0.24 0.27 11.8087
8 0.23 0.20 0.21 0.20 13.5974

In S/N analysis, all calculations are made using the single column of trial S/N values.
Thus, all numbers shown under Main Effect, ANOVA, and Optimum condition represent
values in terms of S/N ratios. Also, regardless of the quality characteristics of the original
evaluation (in this case Bigger is Better), bigger value, that is QC=Bigger, would always
be the quality characteristic when carrying out S/N analysis.
The average column effects of the control factors and the interactions included in
the study are as shown below. Notice that, because there were two columns reserved for
interactions, their influence in relation to the influences of the other control factors are
shown directly here.
Table no. 24- Main effects(S/N)
Sr. Factors Level 1 Level 2 L2 – L1
no.
1 Voltage 13.9029 12.4771 1.4257
2 Air box pressure 12.6441 13.7359 -1.0918
3 Int. of 1×2 12.8700 13.5100 -0.6400
4 Brushing rpm 13.0445 13.3354 -0.2909
5 Semi-cure temp 13.3132 13.0668 0.2464
6 Int. of 2×4 13.7549 12.6250 1.1299
7 Source of powder 13.1293 13.2507 -0.1213

36
No matter which way the analysis is performed (S/N in this case), the average effects of
the noise factors are always calculated in terms of the original units of measurements.
The noise factor effects calculated below are all in terms of the measured coating
thickness (Mil).

Table no. 25- Noise factor effects (average values)

Sr. Factors Level 1 Level 2
no.
1 Humidity 0.2262 0.2175
2 Ambient temp. 0.2175 0.2262
3 Ambient pressure 0.2274 0.2162

ANNOVA-
ANOVA of the control factor and the interactions included indicates their relative
influences (right most columns) to the variations of the results. To minimize the chance
of identifying something while it is not, it is a common practice to revise ANOVA by
pooling insignificant factors. The revised ANOVA after pooling the two insignificant
(done by comparing the confidence level with the desired value) factors are as shown
below.
Table 26- Revised ANNOVA factors
Sr. Factors f S V F S’ P%
no.
1 Voltage 1 4.0655 4.0655 53.8915 3.9901 39.3403
2 Air box 1 2.3841 2.3841 31.6035 2.3087 22.7626
pressure
3 Int. of 1×2 1 0.8193 0.8193 10.8608 0.7438 7.3344
4 Brushing rpm 1 0.1691 0.1691 2.2427 0.0937 0.9243
5 Semi-cure temp (1) (0.1214) POOLED
6 Int. of 2×4 1 2.5533 2.5533 33.8462 2.4779 24.4307
7 Source of (1) (0.0294) POOLED
powder
OTHER/ERROR 2 0.1509 0.0754 5.2075
TOTAL 7 10.1426 100%
The performance at the optimum condition is estimated by considering only the
significant factors and interactions. Since the analysis was performed using the trial S/N
ratios, the estimated value represent the S/N ratio expected from a set of samples tested at
the optimum condition.

37
 Table no.27- S/N ratio at optimum condition
Sr. Factors Level Level Contribution
no. described
1 Voltage 50 KV 2 0.7128
2 Air box pressure 200 PSI 1 0.5459
3 Int. of 1×2 1×2 1 0.3200
4 Brushing rpm 350 RPM 1 0.1454
6 Int. of 2×4 2×4 2 0.5649
TOTAL CONTRIBUTION FROM ALL FACTORS 2.2892
CURRENT GRAND AVERAGE OF PERFORMANCE 13.1900
EXPECTED RESULT AT OPTIMUM CONDITION 10.9008

Optimum condition shown above is determined from the main effects calculated above.
But because there are significant interactions, the levels of interacting factors are adjusted
by examining the test of presence of interaction plots.
The optimum factor levels are: Voltage (level 2), Air Box Pressure (level 1), and
Brushing RPM (level 1). Review of interaction between Air Box pressure and Brushing
RPM indicates (shown below) that the desirable level combination is level 1 for Air Box
and level 2 for Brushing(X1,Y2). This means that the level of Brushing needs to be set to
level 2.

38
 Similarly, review of the interaction plot between factors Voltage and Air Box
shows that level 2 of Voltage and level 1 of Air Box(X2, Y1 shown below for Bigger
S/N), which are same as originally selected from the main effects, remain unchanged.

The corrected optimum factor levels are therefore are: Voltage (level 2), Air Box
Pressure (level 1), and Brushing RPM (level 2). There is no need to recalculate the
expected performance at the optimum performance as it already includes the effect of
significant interactions as shown in the optimum screen above.
Y opt = - 10.9008 in terms of S/N, which can be back-transformed to Y (expected) = .285
(in original units)
The expected performance is also expressed with Confidence Interval (C.I.) at 80%
confidence level as:
Lower limit... S/N = -11.284 (0.273 in the original units of measurement)
Mean Value.. S/N = -10.9008 (0.285 in the original units of measurement)
Upper limit... S/N = -10.517 (0.298 in the original units of measurement)

39
9. SUMMARY OBSERVATIONS AND CONCLUSIONS

Significant Factors in Order of Influence:

• Voltage (39.34 %)
• Air Box Pressure (22.76 %)
• Brushing RPM (0.92 %)

PERFORMANCE IMPROVEMENTS:
Average performance = -13.190 [S/N RATIO of all trials]
Optimum performance = -10.901 [S/N RATIO]
Improvement expected = 17.36 % [over the average performance, S/N]

40
10 . Conclusion:

41
11. Reference:

Textbook of “Taguchi Techniques for Quality Engineering”

- Philip J. Ross

Textbook of “Robust Design using Taguchi Technique”

-Madhav Phadake

A primer on the Taguchi Method

-R. Roy

Software: Qualitek-4 for Design and Analysis of Taguchi Experiments

Websites:

www.nutekinc.com

www.iitb.ac.in/library/electrialengg.

42