Professional Documents
Culture Documents
Design of Experiment
IN DESIGN OF EXPERIMENT
Objectives Of Experimentation
The following are some of the objectives of
experimentation in an industry :
•The results thus obtained will be averaged and the average life of the
specimens made with each material will be used to determine which
material is best.
Design of experiments
Traditional approach
One-factor-At-A-Time
This is a traditional method of experimentation which tests, then changes, one factor
at a time to allow for observation and comparison. Note on the example below, all 8
factors are varied one-at-a-time . It is efficient because it takes only 16 runs.
•A1 and A2 are evaluated by comparing Result - 1 and Result - 2
B1, B2 and B3 are evaluated by comparing Result-2, Result-3 and Result-4.
C1, C2, and C3 are evaluated by comparing Result-4, Result-5 and Result-6, etc.
Run No. A B C D E F G H Re sult
1 1 1 1 1 1 1 1 1 Result 1
2 2 1 1 1 1 1 1 1 Result 2
3 2 2 1 1 1 1 1 1 Result 3
4 2 3 1 1 1 1 1 1 Result 4
5 2 3 2 1 1 1 1 1 Result 5
6 2 3 3 1 1 1 1 1 Result 6
7 2 3 3 2 1 1 1 1 Result 7
8 2 3 3 3 1 1 1 1 Result 8
9 2 3 3 3 2 1 1 1 Result 9
10 2 3 3 3 3 1 1 1 Result 10
11 2 3 3 3 3 2 1 1 Result 11
12 2 3 3 3 3 3 1 1 Result 12
13 2 3 3 3 3 3 2 1 Result 13
14 2 3 3 3 3 3 3 1 Result 14
15 2 3 3 3 3 3 3 2 Result 15
16 2 3 3 3 3 3 3 3 Result 16
Traditional approach
• Problem: Current Car gas mileage is 20 mpg.
Would like to get 30 mpg.
• We might try:
• Change brand of gas
• Change octane rating
• Drive Slower
• Tune-up Car
• Wash and wax car
• Buy new tires
• Change Tire Pressure
• What if it works?
• What if it doesn’t?
“Survey Says” These variables greatly effect MPG
Six Sigma Black Belt 15
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
One-Factor-At-A-Time
Problem: Fuel economy we want is 30 MPG
Try changing each input variable at two settings believed to be
associated with dramatically changing fuel economy. See what
happens.
30 36
37 23
90
Octane 37 24
35
Tire Pressur
23 29
85 30
55 60
Speed
Requisites of DOE
UNBIASEDNESS
PRECISION
INDUCTIVE SCOPE
CLEARLY DEFINED OBJECTIVES
Why Do Replicates?
– To measure pure error: the amount of variability
among runs performed at the same experimental
conditions (this represents common cause variation)
– To see more clearly whether or not a factor is
important—
is the difference between responses due to a change in
factor conditions (an induced special cause) or is it
due to common cause variability?
– To see the effect of changing factor conditions not
only on the average response, but also on response
variability, if desired (two responses can be analyzed:
the mean and the st. dev.)
Six Sigma Black Belt 30
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
200
180
170
160
150
140
5 10 15 20 25 30
What might explain this decline?
Day of Month
Six Sigma Black Belt 32
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
200
190
180
170
160
150
140
5 10 15 20 25 30
Day of Month
50°C
70 °C
200
190
180
Plating Thickness in Microns
170
160
150
140
5 10 15 20 25 30
Day of Month
Definition*
•A lurking variable is one that has an important effect and yet is not included
among the factors under consideration because:
– Its existence is unknown
– Its influence is thought to be negligible
– Data on it are unavailable
•Safeguard
– Randomize the order of the experimental trials to protect against the
effect of lurking variables
– If a lurking variable creates a trend, it can be compensated for in the
numerical analysis
– Valid conclusions can still be drawn about the factors in the experiment
in spite of the presence of lurking variables
After reviewing the data, the scientists were surprised to learn that there was no
statistical difference between the two populations. The average weight (from the
samples) was actually slightly higher (although not statistically higher - p-value >
0.05) for those chickens fed the standard feed. Obviously, this baffled the
scientists involved in the experiment.
Six Sigma Black Belt 37
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
An Example of a Lurking Variable (Contd..)
After a few weeks of evaluating the experiment and the data, one of the grad students asked
Holly Farms for a map of their facility.
After reviewing the map, the student noticed that some of the chicken houses were located
immediately next to the slaughter house. This raised a question in the student's mind and
he decided to drive out to the farm for some first-hand observations. He was escorted to
the slaughter house area and immediately noticed that the chickens located in the houses
next to the slaughter house demonstrated significantly higher levels of activity - i.e.
clucking, pecking, and running around like... well... like chickens. After another review of
the experimental data (by tag number), it was discovered that all of the chickens on the
new feed had been located in the house immediately adjacent to the slaughter house - a
lurking variable had been identified. (NOTE: without the tag numbers being recorded, this
lurking variable may have never been discovered once again illustrating the importance of
proper, thorough data collection). After reviewing his findings with the team, it was decided
to introduce a new variable into the experiment - chicken house location. The experiment
was re-run with the new variable included (i.e. - chicken locations were randomized) and
the results analyzed. On the second attempt, the results validated the scientists original
hypothesis - the new feed produced plumper, meatier chickens. Evidently, those chickens
located next to the slaughter house experienced higher stress levels and subsequent
weight loss. As a result, Holly Farms opted to use the new feed AND also relocated all
chicken houses AWAY from their slaughter houses.
Moral of the Story: Don't keep your chickens too close to the slaughter house.
Local control
• By local control is meant blocking, grouping or
balancing the experimental units. Balancing is done
by replicating all the treatment combination, the
same number of times under different conditions.
Local control makes the test more sensitive and
powerful, by reducing the experimental error.
A WORD OF ADVICE
• It is observed that only 2 to 6 variables end up being
vital few.
• Try to keep the design simple by utilizing your
experience to decide which are the most likely factors
unless you know nothing of the process.
• The above calls for judgement which sometimes can
be wrong.
REMEMBER:
The Experiment is Run to Understand
Reality, Not the Data
Six Sigma Black Belt 41
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
Classification of Design
•Completely Randomized
•Randomised Block
•Balanced incomplete block
•Partially balanced incomplete block
•Latin Square
•Factorial
•Blocked factorial
•Fractional factorial
•Youden square
•Nested
•Response surface
•Mixture designs
•Taguchi’s Design
•EVOP
Classification of Design
Do the following;
-Write down the test of hypothesis
-Carry out ANOVA
-Compare treatment mean
-Check the model adequacy
The overall average is 5.9. Hence the effect of each condition is calculated as
(Treatment mean – Overall mean)
The model followed here is
yij = + i + ij
where yij is the ith treatment and jth observation
is the overall mean
i is the effect of ith treatment
ij is the error follows ND.
As the Fcal is lower than Ftab we conclude that the variation within treatment are
same.
The NORMALIRTY of error and transformation will be discussed in the class.
Classification of Design
Design :Randomised Block
Type of Application : Appropriate when one factor is being
investigated and experimental material or environment can be
divided into blocks or homogeneous groups.
Structures :
Basic : Each treatment or level of factor is run in each block.
Blocking : Usually with respect to one variable.
Information Sought :
1. Estimate and compare effects of treatments free of block
effects.
2. Estimate block effects.
3. Estimate variance.
Six Sigma Black Belt 51
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
Classification of Design
Design :Balanced incomplete block
Type of Application : Appropriate when all the treatment cannot be
accommodated in a block.
Structures :
Basic : Prescribed assignment of treatments to blocks are made.
Every treatments will appear at least once in the experimental
design, but each block will contain only a subset of pairs.
Information Sought :
1. Estimate and compare effects of several factors.
2. Estimate certain interaction effect (some may not be possible).
3. Certain small fractional factorial designs may not provide
sufficient information for estimating the variance.
Classification of Design
Design : Partially balanced incomplete block
Type of Application : Appropriate if a balanced incomplete block
requires a larger number of blocks than is practical.
Structures :
Basic : Prescribed assignment of treatments to blocks are made.
Information Sought :
1. Estimate and compare effects of several factors.
2. Estimate certain interaction effect (some may not be possible).
3. Certain small fractional factorial designs may not provide
sufficient information for estimating the variance.
(All treatments are not estimated with equal precision)
Classification of Design
Design :Latin Square
Type of Application : Appropriate when one primary factor is under
investigation and results may be affected by two other experimental
variables or by two sources of nonhomogenity. It is assumed that no
interaction exists.
Structures :
Basic : Two cross grouping of the experimental units are made
corresponding to the square and row of the square. Each treatment occurs
once in every row and every column. Number of treatments must equal
number of rows and number of columns.
Information Sought :
1. Estimate and compare effects of several factors.
2. Estimate certain interaction effect (some may not be possible).
3. Certain small fractional factorial designs may not provide sufficient
information for estimating the variance.
Classification of Design
Design : Factorial
Type of Application : Appropriate several experimental
factor are to be investigated at two or more levels and
interaction of factors may be important.
Structures :
Basic : Several factors are investigated at several levels
by running all combinations of factors and levels.
Blocking : None
Information Sought :
1. Estimate and compare effect of several factors.
2. Estimate possible interaction effects.
3. Estimate Variance.
Six Sigma Black Belt 55
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
Classification of Design
Design : Blocked factorial
Type of Application : Appropriate when number of runs required for
factorial is too large to be carried out under homogeneous
conditions.
Structures :
Basic : Full set of combinations of factors and levels is divided into
subsets so that high order interactions are equated into blocks.
Each subsets constitutes a block. All blocks are run.
Blocking : Block are usually units in space of time. Estimate of
certain interaction are sacrificed to provide blocking.
Information Sought :
1. Same as factorial except certain high order interactions cannot be
estimated.
Classification of Design
Design :Fractional factorial
Type of Application : Appropriate when there are many factors and
many levels and it is impractical to run all combinations.
Structures :
Basic : Several factors are investigated at several levels but only a
subset of the full factorial is run.
Blocking : Sometimes possible.
Information Sought :
1. Estimate and compare effects of several factors.
2. Estimate certain interaction effect (some may not be possible).
3. Certain small fractional factorial designs may not provide
sufficient information for estimating the variance.
Classification of Design
Design : Youden square
Type of Application : Same as Latin square but number of rows,
columns and treatments need not be same
Structures :
Basic: Each treatment occurs once in every row. Number of
treatments must equal number of columns.
Blocking: With respect to other variables is a two-way layout
Information Sought :
1. Same as Latin square
Classification of Design
Design : Nested
Type of Application : Appropriate when objective is to
study relative variability instead of mean effect of
sources of tests on the same sample and variance of
different samples
Structures :
Basic: Factors are strata in some hierarchical structure,
units are tested from each stratum
Information Sought :
1. Relative variation in various strata, components of
variance.
Classification of Design
Design : Response surface
Type of Application : Appropriate several experimental factor
are to be investigated at two or more levels and interaction of
factors may be important.
Structures :
Factor settings are viewed as defining points in the factor space
(may be multidimensional) at which the response will be
recorded
Information Sought :
1. Maps illustrating the nature of the response surface .
Classification of Design
Design : Mixture designs
Type of Application : Objective is to provide empirical maps contour
diagrams) illustrative of how factors under the experimenter’s
control influence the response
Structures :
May unique arrays, Factor settings are constrained. Factor levels
are often percentages that must sum to 100%. Other factor level
constraints are possible
Information Sought :
1. Estimate and compare effect of several factors.
2. Estimate possible interaction effects.
3. Estimate Variance.
Classification of Design
Taguchi’s Design
No. 1 2 3
1 1 1 1
2 1 2 2
3 2 1 2
4 2 2 1
Six Sigma Black Belt 63
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
FACTORIAL
EXPERIMENTS
9.7 15
16.4 24
80
Temperature.
14.8 15.6
12.3 15.2
20.4 17.4
40 19.3 17.7
17.6 23.2
16.3 20.4
Low High
Concentration
Six Sigma Black Belt 73
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
(2.93) (4.37)
40
18.4 19.7 19
(1.81) (2.71)
The value in bkt. is the std.dev. while the other value is the average
S.D. (pooled), SP = {( 2.932 + 4.372 + 1.812 +2.712 ) / 4 } = 3.09
Six Sigma Black Belt 74
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
80
13.3 4.1
17.4 15.4
17
15
Low High 40 80
Concentration Temperature
Six Sigma Black Belt 76
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
2. INTERACTION PLOT :
19
18
17
15
13
WHICH EXPERIMENT?
23 FULL FACTORIAL EXPERIMENT
Click Design
Click OK
Note here the second column gives the run order on which the experiment
has to be conducted.
•Conclusion
•No problems; proceed with the analysis.
Residuals: An Example
Response or
Experimental Conditions Observed (Y) Average Residuals
Std. order A B (3reps)
Nscore Nscore
3. Normal 3 3
The residuals are not
Probability Plot 2 2
normal. Try a
of Residuals 1 1
transformation on Y.
0
Used to check that 0
-1 -1
residuals are
-2 -2
Normal -3 -3
-3 -2 -1 0 1 2 -5 0 5 10 15
Residual Residual
EXERCISE
An experiment was run to determine the effect of the type of tool, the bevel angle, and the type
of cut on the power consumption for ceramic-tool cutting. Other variables such as cutting
speed and depth of cut were constant during this experiment. The following data were
recorded on power consumption.
Tool A B
Type Bevel Angle Bevel Angle
15 0 30 0 15 0 30 0
32 31 30 32
Interrupted Continuous
27 31 31 38
Type of cut
35 34 30 32
28 39 24 30
30 28 23 29
24 32 24 30
27 29 30 28
27 29 26 26
1 1
0 0
-1 -1
-2 -2
-3 -3
0 10 20 30 0 10 20 30
Residual Residual
3 3
2 2
1 1
0 0
-1 -1
-2 -2
-3 -3
60 65 70 75 50 70 90 110 130 150 170
Nscore Nscore
3 3
2 2
1 1
0 0
-1 -1
-2 -2
-3 -3
-3 -2 -1 0 1 2 -5 0 5 10 15
Residual Residual
Fixture
7.5 11.25 15
5
7.5
• After brain storming, the team decided the following factors and the corresponding levels.
10
Template for tabletop hockey
Factor Name Units Level-1 Level-2
A Puck weight Quarters 4 6
B Stick length Centimeters 7.5 15
C Windup Centimeters 5 10
D Puck place Percent 50 100
• The analysis, residual plot and Normality check of residual is shown in the next
slide.
• The Normal probability plot shows that the residual is normal ( p value = 0.083)
and the Residual plot shows a pattern in data.
• The Normal probability plot shows that the residual is normal ( p value = 0.624)
and the Residual plot shows a random pattern in data.
• Hence we conclude that the model is adequate.
• The following table shows just a few of the possibilities for this power law
along with the appropriate transformation.
Power () Transformation Comment
0 None Normal
0.5 Square Root Counts
1 Logarithm constant percent error
Exercise of MSD
3. GO TO DOE > Factorial > Factorial Plots > Select Main Effects
Plot > GO TO SET UP > Enter Responses > Select Factors by >>
OK , Do same steps for Interaction and Cube Plot as necessary
40
40.9 41.5
40
Time 46
39.3 40.0
30
150 160
Temp
Design Matrix
Let’s design the matrix in Minitab utilizing 5 center points
Design Matrix
StdOrder RunOrder Blocks Temp Time
1 1 1 150 30
2 2 1 160 30
3 3 1 150 40
4 4 1 160 40
5 5 1 155 35
6 6 1 155 35
7 7 1 155 35
Center Points
8 8 1 155 35
9 9 1 155 35
Six Sigma Black Belt 121
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
Center Points
The experiment is carried out and the following data result:
n f nc y f y c
Total 8 3.00223 2
• Blocking
– Any time you cannot perform all trials at
approximately the same time, using the same raw
materials, staff, etc., you have introduced a new
source of variability into the experiment.
– You need a way to separate this unavoidable—and
uninteresting—source of variability from the more
interesting effects of the experimental factors.
– The solution is to treat the unavoidable source of
variation as another factor in the experimental design,
and perform the experiment in blocks.
1 Block 2
» If none of the block effects are judged to be important, you can drop the blocking factor
from the analysis without biasing the conclusions about other factors. (We assume that
the interactions confounded with the block effects are also negligible.)
1 2 3
1 2 3 4 1 2 3 4 1 2 3 4
Supplier 1 2 3
Batches 1 2 3 4 1 2 3 4 1 2 3 4
Observations
This is a two stage nested design and the batches nested with
supplier, because the batch 1 of supplier 1 is not same of any other
supplier.
Six Sigma Black Belt 134
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
Analysis of 2-stage Nested Design
• Open the file Nested.Mpj
• The Purity data is given below
Supplier 1 2 3
Batch 1 2 3 4 1 2 3 4 1 2 3 4
93.45 92.7 92.7 93.45 93.45 93.2 92.95 93.2 93.7 92.7 93.45 93.95
Purity 92.95 92.45 93.2 94.2 92.7 94.2 93.2 93.95 94.2 93.2 92.95 93.7
93.2 92.2 93.45 93.2 92.45 93.7 92.7 93.7 93.2 93.7 93.7 93.45
MINITAB Command
1. GO TO STAT > ANOVA> Fully Nested ANOVA and get the Window
and give the commands
Purity depends on
Supplier and
Batch within
Supplier. The
model entered in
The ANOVA table
this way.
Source DF SS MS F P
Supplier 2 0.941 0.4705 0.969 0.416
Batches(Supplier) 9 4.3698 0.4855 2.944 0.017
Error 24 3.9583 0.1649
Total 35 9.2691
Conclusion :Batches (Supplier) differs significantly.
Six Sigma Black Belt 136
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
Layout-1 Layout - 2
1 2 3 4 1 2 3 4
22 23 28 25 26 27 28 24
Fixture 1
24 24 29 23 28 25 25 23
30 29 30 27 29 30 24 28
Fixture 2
27 28 32 25 28 27 23 30
25 24 27 26 27 26 24 28
Fixture 3
21 22 25 23 25 24 27 27
Exercise:
Consider the following three stage nested design shown below.
Open the Worksheet Hardness.MTW.
Analyse the data and conclude.
Alloy formulation 1 1
Heats 1 2 3 1 2 3
Ingots 1 2 1 2 1 2 1 2 1 2 1 2
Observations
Six Sigma Black Belt 139
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
FRACTIONAL FACTORIAL
EXPERIMENTS USING
TAGUCHI METHODS
FRACTIONAL FACTORIALS
• If you have 5 factors, each at two levels, number of
runs will be 25 i.e. 32 runs for a full factorial.
• If you want to carry out the same in less runs (say 16)
you call it fractional factorial.
EXAMPLE-FRACTIONAL FACTORIALS
NO. FACTOR -1 +1
1 - - - - - 6.1
*2 + - - - - 5.3
*3 - + - - - 6.3
4 + + - - - 6.1
*5 - - + - - 5.3
6 + - + - - 5.6
7 - + + - - 5.4
*8 + + + - - 6.1
*9 - - - + - 6..9
10 + - - + - 6.1
11 - + - + - 9.4
*12 + + - + - 9.3
13 - - + + - 6.6
*14 + - + + - 6.0
*15 - + + + - 9.5
16 + + + + - 9.8
*17 - - - - + 5.6
18 + - - - + 6.3
19 - + - - + 7.0
*20 + + - - + 6.5
*21 - - + - + 5.9
*22 + - + - + 5.5
*23 - + + - + 6.7
24 + + + - + 6.5
25 - - - + + 4.4
*26 + - - + + 4.5
*27 - + - + + 7.8
28 + + - + + 7.7
*29 - - + + + 4.9
30 + - + + + 4.2
31 - + + + + 8.1
*32 + + + + + 8..2
75
R 70
e
s
65
i
d
u 60
e
50
Ht. Acid Conv.speed agitate temp conc
61 Concentration=6%
60
58
50 Temperature (ºC)
40 90
Six Sigma Black Belt 145
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
THIS WAS A FULL 25 FACTORIAL EXPT.
• It requires 32 Runs.
YES
We use fractionalizing when there are more (say 5) factors.
Effects that are estimated
Overall average 1
Main Effect 5
2 way Interactions 10
3 way Interactions 10
4 way Interactions 5
5 way Interactions 1
TOTAL 32 Runs
FACTOR -1 +1
Voltage 50 kV 100 kV
Coating Hanger 0 mm <1.5 mm
Side Front (FR) Back (BK)
Fluoridation pressure 0.5 Kg/cm2 2 kg/cm2
Supplier Ployplast (PP) Ploycoat (PC)
EXAMPLE-FRACTIONAL FACTORIALS
100 20 65 65 65
72 25 25 25 20
100 100 72 100 72
100 100 100 100 100
AVG. 93 54.25 65.5 72.5 64.25
AT -1 LEVELS :
50 0 FR 0.5 PP
46 46 46 46 46
65 65 20 20 25
20 100 100 72 100
25 100 100 100 100
2 Factor Interactions : ( AB,AC, AD, AE, BC, BD, BE, CD, CE, DE ) : DF = 10
3 Factor Interaction : DF = 10
4 Factor Interaction : DF = 5
5 Factor Interaction : DF =1
For estimating the main effects only, we require 5 degrees of freedom and we
can get 5 degrees of freedom from 6 observations. Thus minimum no of
experiments = RDF(required degree of freedom) + 1.
In an experiment, if we have 7 factors each at 2 levels. If we conduct full
factorial experiments it will be 27 = 128 experimental combinations called as trials
or runs. Suppose we intend to carry out experiments only to estimate Main Effects,
we need 7 DF. Then MNE(min no of expr.)= 7 + 1 = 8. We sacrifice here the
knowledge of interactions.
This is done through Taguchi Designs.
Classification of Design
Taguchi’s Design
1 2 3 4 5 6 7 8 9 10 11
1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 2 2 2 2 2 2
1 1 2 2 2 1 1 1 2 2 2
1 2 1 2 2 1 2 2 1 1 2
1 2 2 1 2 2 1 2 1 2 1
1 2 2 2 1 2 2 1 2 1 1
2 1 2 2 1 1 2 2 1 2 1
2 1 2 1 2 2 2 1 1 1 2
2 1 1 2 2 2 1 2 2 1 1
2 2 2 1 1 1 1 2 2 1 2
2 2 1 2 1 2 1 1 1 2 2
2 2 1 1 2 1 2 1 2 2 1
Reaction Example
•NO. OF TRIALS FOR FULL FACTORIAL : 2 5 = 32 NOS.
•STUDY THE EFFECT OF :
A, B, C, D, E : MAIN EFFECTS
AXD, AXC, DXC & BXE : INTERACTIONS
TOTAL 9
Minimum number of experiment = 9 + 1 = 10
The nearest OA table to select is L16(2 15), as we need to assess the interaction.
Six Sigma Black Belt 165
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
Designing a Taguchi‟s Experiment
Go to Stat>DOE>Taguchi>Create Taguchi Design
Agitation Rate and Temperature is most critical factor as the slope is high.
Six Sigma Black Belt 173
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
Reaction Example- Analyse the experiment
Interaction exists between Agitation Rate and Temperature, because the lines are not parallel.
Parameter Design
Quality Engineering:
5. Conduct experiments
Plan/prepare for experiment
Conduct experiment
Collect data
3
1
1
2
3
2
3
2
3
2 1 2
Y2 1
2 2 1 Y3
4 2 1 2 3
. . . Y4
5 2 2 3 1
. . .
6 2 3 1 2
1 1 1
7 3 1 3 2
1 2 2 Y33
8
9
3
3
2
3
1
2
3
1
2 1 2 Y34 9
2 2 1 Y35
Y36
Parameter Design-Example:
Step-1 : Define the project Scope
Raw Material
Crushing & Moulding Calcining Glazing Calcining
Mixing
It was apparent that the uneven distribution of heat causing the problem.
Hence it is necessary that we have to check dimension at each position.
Parameter Design-Example:
Orthogonal arrays may not sound familiar, but they have been around for
over 2000
years. Without knowing it, you are very familiar with orthogonality in your
everyday
life. Take football games, for example. In the first half, one team kicks off
and in the
second half, the other team gets to kick off. Every quarter, they exchange
the field.
Why do they do that? In a tournament, half of the games are played at
home and
the other half away. Why?
1
n ( S m Ve )
S / N 10 Log (in decibles)
Ve
where, n sample size
S m sum of squares for mean
Ve sample variance
Parameter Design-Example:
Response Analysis
RESPONSE TABLE FOR SIGNAL-TO-NOISE RATIOS
A B C D E F G H I J K
1 40.3 37.8 42.8 41.0 40.5 41.4 40.8 42.3 39.5 42.7 40.3
2 40.8 43.3 38.3 40.1 40.6 39.7 40.3 38.8 41.6 38.4 40.8
Diff 0.5 5.5 4.5 0.9 0.1 1.7 0.5 3.5 2.1 4.3 0.5
Parameter Design-Example:
Optimum Combination:
Response table for Signal-to-Noise Ratios illustrates that
factors B, C, F, H, I and J have the strongest impact on
robustness. Their preferred levels are B2, C1, F1, H1, I2 and J1.
The other factors have a weak affect on reducing variability.
From this response table, the optimum condition for reducing
variability is:
A2 B2 C1 D1 E2 F1 G1 H1 I2 J1 K2
Parameter Design-Example:
Response Analysis
A B C D E F G H I J K
1 150.8 151.0 151.6 150.8 151.5 149.9 149.9 150.9 150.3 151.3 150.4
2 150.8 150.6 150.0 150.8 150.1 151.7 151.7 150.7 151.3 150.3 151.2
0.0 0.4 1.6 0.0 1.4 1.8 1.8 0.2 1.0 1.0 0.8
Parameter Design-Example:
Optimum Combination:
Response table for Average reveals that factors C, E, F, G, I
and J have the strongest impact on adjusting or tuning the
average tile dimension to the target value. However, since
factors C, F, I and J also affect the variability in the process,
only factors E and G are considered as adjustment factors.
Notice here that factors A, D and K are relatively
insignificant with respect to reducing variability and
adjusting the process average. This being the case, these
factors become potential cost reduction factors. In other
words, level settings for these factors could be determined by
considering the lowest cost alternative.
Six Sigma Black Belt 196
SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
Parameter Design-Example:
( B2 T ) ( C 1 T ) ( F 1 T ) ( H 1 T ) ( I 2 T ) ( J 1 T ) T
( T average of 12 S/N Ratios)
B 2 C 1 F 1 H 1 I 2 J 1 5 T
43.3 42.8 41.4 42.3 41.6 42.7 - 5(40.56)
51.3 (dB)
Note : The existing condition had a S/N Ratio of 43.8 (dB), therefore
a 7.5 (dB) gain is anticipate d.
C 1 E 2 F 1 G 1 I 2 J 1 5 T
151.6 150.1 149.9 149.9 151.3 151.3 - 5 (40.56)
150.1
Note : Factors E and G cannot adjust the process average to the target
of 150. Consequent ly, the average tile dimension would be adjusted by
changing the mold dimension for the tiles.
Prediction Confirmation
Parameter Design-Discussion:
Why conclusions may not confirm ?
Reason-1
Strong Interaction between control factors
Reason-2
Noise Factor(s)
Reason - 3
Pure Experimental Error.