DoE Notes (Few Slides)

SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD
Design of Experiment
Six Sigma Black Belt 1

IN DESIGN OF EXPERIMENT
• You will find the 1st cut of important X‟s

which are very important.
• At what levels each of these X‟s has to be set

on.
• The best possible combination of levels for

all the important factors so as to optimize
the process.

STATISTICALLY DESIGNED EXPERIMENTS

 A statistically designed experiment permits simultaneous
consideration of all the possible factors that are suspected
to have bearing on the quality problem under investigation
and as such even if interactions effect exist, a valid
evaluation of the main effect can be made. Scanning a
large number of variables is one of the ready and simpler
objectives that a statistically designed experiment would
fulfill in many problem situations.
 Even a limited number of experiments would enable the

experimenter to uncover the vital factors as which further
trials would yield useful results. The approach has
number of merits, it is quick, reliable and efficient.

Objectives Of Experimentation
The following are some of the objectives of
experimentation in an industry :
 Improving efficiency or yield

 Finding optimum process settings
 Locating sources of variability
 Correlating process variables with product
characteristics
 Comparing different processes, machines, materials etc
 Designing new processes and products.

NEED FOR DESIGNED EXPERIMENT
• A design engineer is interested in studying the effect Bronze and

Synthetic material on life of Bush.
•Here, the objective is to determine the material that produces the

maximum life (minimum wear) for this particular Bush.
•The procedure thought of by the engineer is to make a number of Bush

specimens with each material and to measure the life of the specimens
after running the pump.
•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.

NEED FOR DESIGNED EXPERIMENT

• As we think about the experiment, a number of important
questions come to mind:
a) Are these two materials the only materials of potential interest?
b) Are there any other factors that might affect life that should be
investigated or controlled in this experiment?
c) How many specimens of Bush should be tested with each other
material?
d) How should the specimens be made with different materials and in
what order should the data be collected?
e) What method of data analysis should be used?
f) What difference in average observed life between the two materials
will be considered important?
• All of these questions and perhaps many others will have to be
satisfactorily answered before the experiment is performed.
PROOF OF THE NEED FOR EXPERIMENTATION

 After having selected the area for experimentation we have
to ensure that the problem is of „Break through‟ or
„Improvement‟ nature and not a problem of „control‟ nature.
For this purpose past data should be suitably analyzed and
plotted on some process control chart to check whether the
process is within statistical control or not. If the analysis
shows lack of control or statistical instability, then it is a
problem of „control‟ nature and experimentation may not be
needed.
 However if the problem is of chronic nature and there is

stability in the process, then it establishes the need for
experimentation. Before deciding to carry out
experimentation the need for experimentation must be
established.
IF CONTROL ONLY IS YOUR PROBLEM
WE SEEK ANSWERS TO THE FOLLOWING

QUESTIONS
 Which X‟s are important ?
 Are they controllable ?
 Do they change by chance or by intent (assignable) ?

IF CONTROLLABLE AND CHANGE BY CHANCE
E.g: You find that finish (Y) deteriorates with time.

You find that this is because of Tool Wear.
• Does the tool wear by chance
• Do you have some control over the above - yes, by
changing the tool
• In all such cases where „X‟ changes by chance but
some control is exercisable
USE CONTROL CHARTS ON YOUR Y

IF CONTROLLABLE AND CHANGE BY CHANCE

While using control charts
 Clearly document what is to be checked
 When is it to be checked
 If a point goes out of limits what X‟s to be
checked for
 As far as possible put all X‟s in order of priority
IN CONTROL PHASE YOU WILL SEE THIS IN DETAIL

IF CONTROLLABLE & CHANGE BY INTENT

E.g.: Supplier of sheets for vacuum forming process is
an important factor for shift of mean.
• The change in material on line is done by intent and

knowingly.
• In such cases where you know when a critical X has

changed levels. Find the optimum setting of the
process for each level of the X.
D.O.E. IS THE TOOL FOR IT

IF VARIATION IS OUR PROBLEM
• Find at what level each of the X has to be

set on i.e. the optimum process settings.
• D.O.E. is the tool for it

Design of experiments
Design of experiments (DOE) is a valuable tool to

optimize product and process designs, to accelerate
the development cycle, to reduce development costs,
to improve the transition of products from research
and development to manufacturing and to effectively
trouble shoot manufacturing problems. Today,
Design of Experiments is viewed as a quality
technology to achieve product excellence at lowest
possible overall cost.

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
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.
Speed Octane Tire Pressure MPG

55 85 30 23
60 85 30 29
60 90 30 23
60 85 35 24
How many more Combinations would you need to figure out the best combination
of variables? (3 Variables at two settings; 2x2x2 = 8 total)
How can you explain the above results? (Combination 2 is the answer)
If there were more variables, how long would it take to get a good solution?
(Multiply by another 2 for each one)
What if there’s a specific combination of two or more variables that leads to the
best mileage? (Too hard for me to figure out; What do you think?)

OFAT Studies and Interactions
• Suppose we looked at one factor at a time. We conduct a total of 8
runs
Cube Plot (data means) for Miles per Gallon
30 36
37 23
90
Octane 37 24
35
Tire Pressur
23 29
85 30
55 60
Speed
While we would have made significant improvement, we would

have missed the optimum point!
Full Factorial Experiment
Problem: Want 30 MPG
Speed Octane Tire Pressure MPG

55 85 30 23
60 85 30 29
55 90 30 37
60 90 30 23
55 85 35 37
60 85 35 24
55 90 35 30
60 90 35 36
OFAT Runs
What conclusion do you make now?

(Murphy is alive and well!)
TERMINOLOGY USED IN D.O.E.

EXPERIMENT: A planned set of operations which leads to a
corresponding set of observations. The purpose of experimentation is
to ensure that the experimenter obtains the data relevant to the task of
decision making in an economical way.
OUTCOME (RESPONSE): The numerical result of a trial based on a

given treatment combination is called Outcome or Response.
The response may be :

– Continuous or measurement type and follows a normal distribution
– Continuous or measurement type but does not follow normal distribution
– Discrete or count type and does not follow normal distribution
E.g.: diameter of a shaft, No. of rejected cylinders etc.

FACTOR (X) - The parameters of the process which are deliberately
varied from trial to trial. This could be qualitative or quantitative. e.g.
Speed, feed, coolant rate, operator skill.
LEVELS OF A FACTOR - The alternative values of a factor considered
in the experiment are called its levels.
e.g.: Speed 400 rpm, circular wheel etc.
TREATMENT COMBINATION - The set of levels of all factors
employed in a given trial is called treatment or treatment combination.
EXPERIMENTAL UNIT : It is a generic term used to denote the group
of material to which a treatment is applied in a single trial of the
experiment.
BALANCED TEST - Where number of samples in each treatment

combination is same.

EFFECT OF FACTOR :
MAIN EFFECT: The change in the average response produced by a
change in the level of the factor is called “Main Effect” of that factor.
INTERACTION EFFECT : If the effect of one factor is different at

different levels of another factor, the two factors are said to interact
(or) to have interaction.
The interaction between factors A and B, is termed as “First Order

Interaction” or “Two Factor Interaction” and is denoted by AxB.
If the interaction between two factors A and B, is different at different

levels of a third factor C, then there is said to be interaction among
three factors. This is referred to as “Second Order Interaction” or
“Three Factor Interaction” and is denoted by AxBxC.


Interactions
• Y = f (X1, X2). But if X2 = f (X1)
Then changing X1 will give other than predicted Y since X2 also
automatically changes.
• The same holds true for change of x2
e.g: leakage of dome welded components is a function of current

and electrode thickness but current also depends on electrode
thickness.
Hence there is interaction between electrode and current

An example to understand interaction
F
I
N
I
S
H
Speed X Speed Y
Changing feed from level A to level B betters finish.

But this effect is more predominant speed level Y than speed level X.
Hence there is an interaction between speed & feed
REPLICATION: Replication is a repetition of the whole
experiment in order to estimate experimental error, increase
precision (detect smaller changes).
EXPERIMENTAL
Six Sigma Black Belt ERROR: The failure of two identically 23
STEPS IN DESIGNING AND ANALYZING
1. Statement of the problem.
2. Formulation of hypothesis.
3. Planning of the experiment.
a) Choosing an appropriate experimental technique.
b) Examination of possible outcomes to make sure that the experiment
provides the required information.
c) Consideration of possible results from the point of view of statistical
analysis.
4. Collection of data, after performing the experiment according
to the plan.
5. Statistical Analysis of the data.
6. Drawing conclusions with appropriate level of significance.
7. Verification or evaluation of results (conclusions).
8. Drawing final conclusions and recommendations.

PLANNING FOR EXPERIMENTATION

The various steps to be followed in this direction are listed
below :
 Selection of area of study : Pareto analysis
 Proof of the need for experimentation
 Brain storming and Cause & Effect diagram : To list all
the possible factors
 Classification of factors
 Interactions to be studied
 Response and type of model for analysis

Classification of factors
Tools like brainstorming and cause & effect diagrams helps in
identification of factors and preparing a complete list of the factors
involved in any experiment. Factors listed can be classified into three
categories :
1. Experimental Factors
Experimental factors are those which we really experiment with by
varying them at various levels.
2. Control Factors
Control Factors are those which are kept at a constant (controlled)
level throughout experimentation.
3. Error or Noise Factors
Error or Noise factors are those which can neither be changed at our
will nor can be fixed at one particular level. Effect of these factors
causes the error component in the experiment and as such these
factors are termed as error or noise factors.
Note : At the planning stage itself all the factors viz. Experimental, Control and error should be
recognized. This will help to tackle them appropriately during experimentation.

PLANNING FOR EXPERIMENTATION

• State what do you want
• What is my response(s)
• What are my factors
• Choose the level of the factors
• Decide on the design
• Run the design and collect the data
• Analysis the data and obtain results
• Run confirming test on settings
Requisites of DOE
UNBIASEDNESS
PRECISION
INDUCTIVE SCOPE
CLEARLY DEFINED OBJECTIVES
Fulfillment of the requirements

1. RANDOMISATION
2. REPLICATION AND
3. LOCAL CONTROL OR ERROR CONTROL

Replication
Definition
•Replication means repeating all the experimental conditions (or
running a combination) two or more times.
– This does not mean measuring an experimental unit twice
– It does mean repeating a certain set of conditions and measuring
the new output
– Two replicates means that for an 8-run design you will do 16 runs
in one experiment
• Minitab will randomize all the runs (including replicates) at the
same time
• If for some reason you cannot, or choose not to, do all the runs
at the same time, you need to be concerned about blocking (a
topic we’ll discuss later in this module)
– One replicate really means no replication

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.)
Repetition v/s Replication:

• Repetition: at the same settings a number of readings are
taken. These are repetitions.
• Replication: at a setting reading is taken and next settings
done. After some time the first setting is set again &
readings taken. These readings are replicates.
When to repeat and when to replicate.

Cost low, variability high. Replicate.
Measuring variability Repeat.
greater than 10% of
tolerance.
Greater ability to generalize Replicate.
results over wide range of
conditions.
Why Randomize: Example

• Background
• Suppose that the plating thickness on printed circuit boards is the
response of interest. Notice that this value tends to decline over the
month.
Thickness vs. Day of Month
200
Plating thickness in microns

190
180
170
160
150
140
5 10 15 20 25 30
What might explain this decline?
Day of Month
Why Randomize: Example, cont.

• Suppose in an experiment to evaluate the effect
of soak temperature, the company tested 50°C
first, then 70°C . Thickness vs. Day of Month
50°C 70 °C
Plating Thickness in Microns
200
190
180
170
160
150
140
5 10 15 20 25 30
Day of Month

Why Randomize: Example, cont.

• Alternatively, what if both temperatures were tested randomly
throughout the month? Thickness vs . 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

Randomization: the Experimenter‟s Insurance
Definition
•To assign the order in which the experimental trials will
be run using a random mechanism
– It is not the standard order
– It is not running trials in an order that is convenient
– To create a random order, you can ―pull numbers from a hat‖
or have Minitab randomize the sequence of trials for you
•Why?
– Averages the effect of any lurking variables over all of the
factors in the experiment
• Prevents the effect of a lurking variable from being mistakenly
attributed to another factor
– Helps validate statistical conclusions made from the
experiment
Lurking Variables
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

An Example of a Lurking Variable

The Agricultural Sciences Department at North Carolina State University developed
a new and improved chicken feed that would supposedly promote plumper and
meatier chickens. The school contracted with a local poultry provider (Holly
Farms) and conducted a series of studies testing the new product. The NC State
Mathematics Department was asked to develop a DOE to support the above
tests.
Preliminary calculations were made and two populations of chickens were
identified, tagged (this becomes very important later),and segregated. One
population was fed the standard feed and the other fed the new and improved
feed. After feeding the two populations of chickens, statistically significant
samples from each population were slaughtered and weighed (what we in Power
Systems would refer to as 'Destructive Testing'). The outcome of the experiment
was obviously trying to prove that chickens on the new feed weighed more than
those on the old feed. In this case the Y =Weight and X = Type of feed.
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.
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.

Formulation of experimental problem
1. Process Flow chart prepared
2. Measures Established
3. Stability of the process verified
4. Cause & Effect Analysis made involving process experts and Operators
5. Technical Justification of factors and it’s level completed.
Factor Unit of Present Investigation Nature of Levels Operations

Remarks
Measure Status range Factor Implications
(E/C/N) 1 2 3
* E : Experimental factors, C : Control factors, N : Noise factors

• Suspected interactions between Factors:
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
WHY EXPERIMENTS CAN FAIL / ABORTED

• You are not clear what you want.
• Measurement systems differ at different points in time.
• Inadequate factors are considered.
• Lack of understanding of DOE tools / Strategies .
• No management support
• Need instant results
• No resources
• Lack of time
• Cost of experiment may be high

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

Design : Completely Randomized

Type of Application : Appropriate when one experimental
factor is being investigated.
Structures :
Basic : One factor is investigated by allocating
experimental units at random to treatments (Levels of a
factor)
Blocking : None
Information Sought :
1. Estimate and compare treatment effects
2. Estimate Variance.
Completely Randomized Design : Example
Example : Suppression of Bacterial Growth on Stored Meats

The shelf life of stored meats is the time a prepackaged cut
salable, safe and nutritious. There are four type of packaging
and the management is interested about the best packaging
alternatives.
The packaging types are
- Commercial plastic Wraps
- Vacuum packaged
- 1% CO, 40% O2, 59% N
-100% CO2
How to plan the experiment? We also decided to carry out the
trials in 3 samples each.

The data collected and presented below;
Psychotrophic Bacteria
Pckaging Condition Log(Count/cm2) Total Average
Commercial plastic wrap 7.66, 6.98, 7.80 22.44 7.48
Vacuum packaged 5.26, 5.44, 5.80 16.50 5.50
1% CO, 40% O2, 59% N 7.41, 7.33, 7.04 21.78 7.26
100% CO2 3.51, 2.91, 3.66 10.08 3.36
Do the following;
-Write down the test of hypothesis
-Carry out ANOVA
-Compare treatment mean
-Check the model adequacy


Ho: Null Hypothesis: Psychotropic Bacteria Growth at four different packing
material is same.
H1: Alternate Hypothesis: Psychotropic Bacteria Growth at least one packing
material is different than others.
Analysis of Variance Table
Degrees of Sum of Mean
Source Freedom Square Square F-ratio F-tab P value
Treatment 3 32.8728 10.9576 94.58 4.07 0.0000
Error 8 0.9268 0.1158
Total 11 33.7996
From ANOVA table we conclude that Ho is rejected, i.e., there

exist difference among the packaging conditions.
Now the next question is which mean(average) is different than

others. And how to do it.

The average and standard deviation is presented in the following
table along with data.
CO, O2, N 100% CO2 Commercial Vacuum
7.41 3.51 7.66 5.26
7.33 2.91 6.98 5.44
7.04 3.66 7.80 5.80
Average 7.26 3.36 7.48 5.50
Stdev i 0.1947 0.3969 0.4386 0.2750
Effect ( ) 1.36 -2.54 1.58 -0.40
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.


The two treatment mean is different when the difference between the two mean
is larger than Least Significant Difference (LSD).
The LSD can be calculated using the following formula
LSD( ) = t  /2,  s2 [ 1/ri + 1/rj]

However this is applicable when treatment are significant in ANOVA.
Commercial Vacuum CO, O2, N 100% CO2
7.48 5.50 7.26 3.36
Pckaging Condition Average
Commercial plastic wrap 7.48 - 1.98 0.22 4.12
Vacuum packaged 5.50 - -1.76 2.14
1% CO, 40% O2, 59% N 7.26 - 3.90
100% CO2 3.36 -
S = Sqrt(MSE) = SQRT(0.1158)
= 0.3403
LSD = 2.306 X SQRT(0.1158X2/3)
= 0.6407
The model is accurate when

- The variances among treatment are same
- The error follows Normal Distribution
In case it is not, then we need to transform data and analysis again

CO, O2, N 100% CO2 Commercial Vacuum
7.41 3.51 7.66 5.26
7.33 2.91 6.98 5.44
7.04 3.66 7.80 5.80
Average 7.26 3.36 7.48 5.50
Stdev 0.1947 0.3969 0.4386 0.2750
Effect ( ) 1.36 -2.54 1.58 -0.40
The test we carry out is comparing S2 max to S2 min and and F-test
The test statistics is
F0 Max = Max (si2)/Min(si2)
The test details given below.
Stdev Max = 0.4386 F cal = 5.08
Stdev Min = 0.1947 F tab = 7.18
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.

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.
1. Estimate and compare effects of treatments free of block
effects.
2. Estimate block effects.
3. Estimate variance.
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.
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.

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.
(All treatments are not estimated with equal precision)

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.
3. Certain small fractional factorial designs may not provide sufficient
information for estimating the variance.

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
1. Estimate and compare effect of several factors.
2. Estimate possible interaction effects.
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.
1. Same as factorial except certain high order interactions cannot be
estimated.

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.

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
1. Same as Latin square

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
1. Relative variation in various strata, components of
variance.

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
1. Maps illustrating the nature of the response surface .

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
1. Estimate and compare effect of several factors.
2. Estimate possible interaction effects.

Taguchi’s Design
Let us say y = f(x,z)

where y = Product Response
x’s = controllable factors
z’s = noise factors
The objective is to choose settings of x that will make
the product’s response insensitive to variability
associated with both x and z and still meet target
specifications with least variability.

Orthogonal Arrays
The symbology for an orthogonal array is La(bc) where;
L = Latin Square
a = The number of test trials
b = The number of levels for each column, and
c = The number of columns in the array.
2n Series Orthogonal Arrays
L4(23) Orthogonal Array
No. 1 2 3
1 1 1 1
2 1 2 2
3 2 1 2
4 2 2 1
FACTORIAL
EXPERIMENTS

Full Factorial Experiments

• Wear of pin is an important criteria in
affecting field life of a component.
• It is believed that hardness of pin is an
important parameter affecting it.
• Hence experiments are carried out to check
wear on :
– Pin of hardness in range of 60 - 62 RC
– Pin of hardness in range of 66 - 68 RC

Seek the answers to the following questions

• What is your response ?
• How many Factors [f] ?
• How many Levels [L] ?
• The experiment is Lf
• How many combinations/runs are possible ?
• How many runs do you plan to carry out ?

SEEK THE ANSWERS TO THE FOLLOWING QUESTIONS
• What is the response ? Wear

• How many Factors [f] ? 1
• How many Levels [L] ?
2
• The experiment is Lf 21
• How many combinations/runs are possible ? 2
• How many runs do we plan to carry out ? 2
HENCE IT IS A 21 FULL FACTORIAL.

22 Full factorial experiments
• It is believed that pin wear depends on

– Hardness
– Oil flow
• The levels of hardness are

– 60 - 62 Rc
– 66 - 68 Rc
• The levels of oil flow is

– 20 cc / min
– 120 cc / min

• Number of Factors : 2
• Number of Levels : 2
• Possible Runs : 22
• Nos. we plan to carry out: 4
Hence it is a 22 full factorial experiment.
Similarly you have 23 and 24 full factorial

experiments for 3 and 4 factors respectively.

EXAMPLE- 22 FACTORIAL EXPERIMENT

Consider a chemical process of Silicate Mfg. It is felt
that Temperature and Concentration are the
contributors to increase residue.
The factors and levels are as below

Factor -1 +1
Temp. 40ºC 80ºC
Conc. Low High
-1 signifies one level (normally lower) and +1 signifies

the other level (normally higher)

• It is now believed that residue depends on concentration of Acid and Temperature of
bath.
RUN CONC. TEMP. RESIDUE
1 Low 40 20.4
2 Low 40 19.3
3 Low 40 17.6
4 Low 40 16.3
5 Low 80 9.7
6 Low 80 16.4
7 Low 80 14.8
8 Low 80 12.3
9 High 40 17.4
10 High 40 17.7
11 High 40 23.2
12 High 40 20.4
13 High 80 15
14 High 80 24
15 High 80 15.6
16 High 80 15.2

WHAT DO WE WANT TO FIND ?

• We want to find that
– Does concentration and temperature have any effect
on residue.
– Of concentration and temperature which is more
important .
– What is the ideal and feasible level of the process
settings.
– Does any interaction exist between temperature and
concentration.
– Is there any problem with data or model adequate ?
How Do We Find This. ?

Club all values of all 4 possible combinations and represent as below
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

Calculate the average and the std.dev of each block as below
Low High
80 13.3 17.4 15.4
Temperature
(2.93) (4.37)
40
18.4 19.7 19
(1.81) (2.71)
AVG. 15.8 18.6 17.2
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
WHAT DO YOU INTERPRET. ?

Low High
80
13.3 4.1
17.4 15.4
Temp. 5.1 2.3 3.6

40
18.4 19.7 19
1.3
AVG. 15.8 2.8 18.6 17.2

The above can be graphically represented as

1. MAIN EFFECTS PLOT :
19
17
15
Low High 40 80
Concentration Temperature
2. INTERACTION PLOT :
19
18
17
15
13
Low Concentration High


Consider another setup of surface cleaning. It is felt
that Time, Temp. and Conc. are the contributors.
The factors and levels are as below

Factor -1 +1
Temp. R.T. 90ºC
Time 3 mins 10 mins
Conc. Low High
-1 signifies one level (normally lower) and +1 signifies

the other level (normally higher)

HOW MANY FACTORS? 3
HOW MANY LEVELS? 2
HOW MANY RUNS WOULD BE THERE IDEALLY? 8
HOW MANY YOU PLAN TO RUN? 8
WHICH EXPERIMENT?
23 FULL FACTORIAL EXPERIMENT

EXAMPLE: THE PROBABLE COMBINATIONS ARE
NO. TEMP. TIME CONC.

1 RT 3 mins Low
2 90 3 mins Low
3 RT 10 mins Low
4 90 10 mins Low
5 RT 3 mins High
6 90 3 mins High
7 RT 10 mins High
8 90 10 mins High
• This is called an array

• Since it contains all possible combinations. It is a full factorial array
• It is also called orthogonal array
• If columns are orthogonal we can estimate the effect of a variable independent
of the other variables

Designing the Experiment
Minitab Steps for Designing Full Factorial Experiments:
1. GO TO STAT > DOE> Factorial > Create Factorial Design > Type of Design > No of
Factors> Click Designs > Select Full Factorial > Enter no of Replications > OK>Click
Factors > Enter Factor names and Levels>OK > Click Options >Select Randomization
as required > OK>OK
Select type of Design
Input number of Factors
Click Design
Select Full Factorial
Select No. of Replicates
Click OK

1. GO TO STAT > DOE> Factorial > Create Factorial Design > Type of Design > No of Factors> Click Designs > Select Full Factorial >
Enter no of Replications > OK>Click Factors > Enter Factor names and Levels>OK > Click Options >Select Randomization as required >
OK>OK
Enter Factor name & Levels
Deactivated if you don‟t want to randomize it.

The Design out put along with the data obtained after conducting
StdOrder RunOrder CenterPt Blocks Tempareture Time Concentration Response
1 1 1 1 RT 3mins Low 65
11 2 1 1 RT 10 mins Low 43
13 3 1 1 RT 3mins High 61
12 4 1 1 90 10 mins Low 45
5 5 1 1 RT 3mins High 58
15 6 1 1 RT 10 mins High 50
3 7 1 1 RT 10 mins Low 50
7 8 1 1 RT 10 mins High 52
10 9 1 1 90 3mins Low 42
8 10 1 1 90 10 mins High 41
14 11 1 1 90 3mins High 43
9 12 1 1 RT 3mins Low 65
16 13 1 1 90 10 mins High 45
6 14 1 1 90 3mins High 45
4 15 1 1 90 10 mins Low 41
2 16 1 1 90 3mins Low 44
Note here the second column gives the run order on which the experiment
has to be conducted.

Detail of Data Analyze
• Three phases of data analysis
A: Look for B: Identify large C: View effects

problems with effects on response
the data or the
model
• Pareto chart of • Main effects plot
• Time plot of effects
• Interaction plots
response(s)
• Normal probability
• Residuals plots plot of effects • Cube plots
• p-values of effects

Look for Problems

• Background: Before you analyze your own data, we will walk through an
earlier example together, using the data in the file listed above (so we will all get
the same results). You will analyze your own MSD data (from exercise) when
the analysis demonstration is complete.
1. Make a time plot of the response:

Graph > Time Series Plot > (Select „Response‟ for Y)

Look for Problems cont.
2. Interpret the plot by looking for:
a. “Defects” in the data
– Missing values; outliers caused by typos or mistakes
– If there is no correction or explanation for an outlier, proceed with analysis (do
not ―throw it out‖)
b. Trends or cycles that indicate lurking variables associated with time
– This is not a plot of common cause variation—it (hopefully) contains special
causes induced by the factor settings of the experiment, so it probably won’t
―look‖ random
– However, this plot helps us see if the results are influenced by variables not
being tested in the experiment (that we did not anticipate or control)
– If a trend caused by another variable is identified, see a statistician for help with
the analysis (use time order as a covariate in the model)
•Conclusion
•No problems; proceed with the analysis.

Residuals
• Definition
• Residual = (Observed Y) – (Average of Ys at that experimental
condition)
• A residual is the difference between a response and

what we ―expect‖ it to be (the expected value is the
average of all replicates for a particular combination of
factor settings).
– We hope most variation in the Ys is accounted for by
deliberate changes we’re making in the factor settings
– Whatever variation is left over is residual
• The assumption is that this residual variation reflects the common
cause variation in the experiment

Residuals: An Example
• Residuals for 22 Design with 3 Replicates
Response or
Experimental Conditions Observed (Y) Average Residuals
Std. order A B (3reps)
1 – – 9, 11, 7 9.0 0.0, 2.0, –2.0
2 + – 10, 6, 8 8.0 2.0, –2.0, 0.0
3 – + 15, 19, 20 18.0 –3.0, 1.0, 2.0
4 + + 15, 18, 12 15.0 0.0, 3.0,
What is the residual

for this observation?

Assumptions of DoE Analysis
• We assume the residuals are:
– Normal: bell-shaped with a mean of 0
– Constant: do not increase as averages
of each experimental condition increase
– Stable: do not change over time
– Not related to the Xs (factors)
– Random: represent common causes of variation
– Independent

Checking Assumptions About Residuals
• Residuals plots must be checked to ensure the
assumptions hold. Otherwise conclusions may be
incorrect or misleading.
Residuals Plot Good Bad Meaning / Actions
Residual
Residual
3
3
1. Time Plot of 2
2 Any pattern visible over
Residuals* 1 1 time means another
Used to check for 0 0 factor, related to time,
stability over time -1 -1 influences Y. Try to
-2 -2 discover it.**
-3 -3
0 10 20 30 0 10 20 30
Time Order Time Order
2. Residuals vs. Residual Residual

3 3
Fitted Value 2
This fan shape means
2
(average of each 1 1
the variation increases
condition) 0
as the average of each
0
condition increases (it‟s
Used to check for -1 -1
not constant). Try a
constancy; -2 -2
square root, log, or
variation does not -3
60 65 70 75
-3
inverse transformation
50 70 90 110 130 150 170
increase as Fitted Value (Average) Fitted Value (Average) on Y.
average increases 8 conditions, 2 reps 8 conditions, 5 reps
(2 reps will appear
as a mirror image)
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

Checking Assumptions About Residuals

The three plot of residual does
not show any pattern. In the
Normality test of residual, we
get p-value is 0.717. Hence we
conclude that the residual
follows Normal distribution and
the model is adequate.

Plot of the effects
Pareto Chart of the effects shows that the

main effect of A, B and Interaction AB is
Significant

P-value of effects
Fractional Factorial Fit: Response versus Tempareture, Time, ...
Estimated Effects and Coefficients for Response (coded units)
Term Effect Coef SE Coef T P Significant

Constant 49.375 0.7512 65.73 0.000 Effects
Temparet -12.250 -6.125 0.7512 -8.15 0.000
Time -7.000 -3.500 0.7512 -4.66 0.001
Concentr 0.000 0.000 0.7512 0.00 1.000
Temparet*Time 6.500 3.250 0.7512 4.33 0.002
Temparet*Concentr 0.500 0.250 0.7512 0.33 0.747
Time*Concentr 2.250 1.125 0.7512 1.50 0.168
Analysis of Variance for Response (coded units)

Source DF Seq SS Adj SS Adj MS F P
Main Effects 3 796.25 796.25 265.417 29.40 0.000
2-Way Interactions 3 190.25 190.25 63.417 7.02 0.010
Residual Error 9 81.25 81.25 9.028
Lack of Fit 1 30.25 30.25 30.250 4.75 0.061
Pure Error 8 51.00 51.00 6.375
Total 15 1067.75

Plot of Significant effects

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

Dealing with Non-Normality

Good Bad
Residual Residual
3
3
2
2
1 1
0 0
-1 -1
-2 -2
-3 -3
0 10 20 30 0 10 20 30
Time Order Time Order
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
Fitted Value (Average) Fitted Value (Average)
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

• We assume the residuals are:
– Normal: bell-shaped with a mean of 0
– Constant: do not increase as averages
of each experimental condition increase
– Stable: do not change over time
– Not related to the Xs (factors)
– Random: represent common causes of variation
– Independent
• But in some case we found that the residual

– Does not follow Normal Distribution and
– Shows pattern when compared with residual
• The following example shows how to check the residuals and

deal with Non normality.


• The data in this exercise comes from ―tabletop hockey‖. The objective of this experiment is to learn
how to shoot a puck for distance with a flexible, 15 centimeter ruler. The templates is shown below.
Fixture
Puck (at various positions)
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 data shown in the following table
Trial No. Puck weight Stick length Windup Puck place Y

1 4 7.5 5 50 38.194
2 6 7.5 5 50 23.281
3 4 15 5 50 2.999
4 6 15 5 50 7.603
5 4 7.5 10 50 109.901
6 6 7.5 10 50 90.573
7 4 15 10 50 20.606
8 6 15 10 50 18.880
9 4 7.5 5 100 36.559
10 6 7.5 5 100 38.019
11 4 15 5 100 45.499
12 6 15 5 100 44.875
13 4 7.5 10 100 190.108
14 6 7.5 10 100 116.681
15 4 15 10 100 137.404
16 6 15 10 100 84.528
• 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 abnormal residual plot exhibit a not uncommon ―power law‖ relation ship
between the standard deviation and the mean response. Statistically, this
sy  
situation is symbolised as follows;
• 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
Time Limit 45 Mins












1. GO TO STAT > DOE> Factorial > Create Factorial

Design > Type of Design > No of Factors> Click
Designs > Select Full Factorial > Enter no of
Replications > OK>Click Factors > Enter Factor names
and Levels>OK > Click Options >Select Randomization
as required OK>OK
2. Conduct the Experiment – Ensure Measurement

System through Gage R &R as required. Make data
collection format for each experiment conducted and
write down the other parametric values which have not
been varied.


Analyzing the experimental results

1. Enter the results in the worksheet – GO TO DOE > Factorial >
Analyse Factorial Design > Enter Responses > Graph > Pareto >
Standardize >OK >Results > Unusual obs in addition to ANOVA >
Select the terms by >> OK >OK
2. Select the sources which are significant
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
4. Select Optimum , trade off if necessary .
5. Confirm , Confirm , Confirm by actual Trial.

Analyse the data and present your findings
Time Limit :- 45 Mins
2k Factorials - Center Points and Blocking

Cube Plot - Means for Yield
40
40.9 41.5
40
Time 46
39.3 40.0
30
150 160
Temp

Adding Center Points

• There is always a risk in 2-level designs of missing a curvilinear
relationship by only including two levels of the Input Variable
• The addition of Center points is an efficient way to test for
curvature without adding a large number of experimental runs and
to check the process at ―nominal‖
• Example:
– A chemical engineer wants to improve yield
– There are two inputs of interest: Reaction Time and Reaction
Temperature
– The engineer decides to conduct the experiment using a 2x2
design (Reaction Time x Reaction Temp), but will add 5
center points to estimate experimental error and curvature
– Inputs:
• Reaction Temp: 150 and 160; Center point = 155
• Reaction Time: 30 and 40; Center point = 35

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
Center Points
The experiment is carried out and the following data result:
RunOrder Blocks Temp Time Yield

1 1 150 30 39.3
2 1 160 30 40.0
3 1 150 40 40.9
4 1 160 40 41.5
5 1 155 35 40.3
6 1 155 35 40.5
7 1 155 35 40.7
8 1 155 35 40.2
9 1 155 35 40.6

Analysis
Estimated Effects and Coefficients for Yield
Term Effect Coef Std Coef t-value P

Constant 40.4444 0.06231 649.07 0.000
Temp 0.6500 0.3250 0.09347 3.48 0.018
Time 1.5500 0.7750 0.09347 8.29 0.000
Temp*Time -0.0500 -0.0250 0.09347 -0.27 0.800
Analysis of Variance for Yield

Main Effects 2 2.82500 2.82500 1.41250 40.42 0.001
2-Way Interactions 1 0.00250 0.00250 0.00250 0.07 0.800
Residual Error 5 0.17472 0.17472 0.03494
Curvature 1 0.00272 0.00272 0.00272 0.06 0.814
Pure Error 4 0.17200 0.17200 0.04300
n f nc  y f  y c 
Total 8 3.00223 2
SS Curvature  Not Important

n f  nc

Blocked Factorial Experiments
• Suppose you are going to run a 2 4

experiment. It requires 16 runs, but there is
only enough volume of business to run 8 trials
in any particular region in one month.
– What would you do?
– What if you only have enough business to run 4
trials in any particular region?

• 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.

•Definition
•A block is a carefully chosen subset of the total runs in
the experiment that are tested close together in time or
under similar conditions.
– ―Similar conditions‖ means that factors not being tested in the
experiment are the same or very nearly so (includes factors such as
staffing, physical environment, source materials, etc.).
– In the situation discussed on the previous page, you would have to
conduct the experiment in (a) 2 blocks of 8 trials or (b) 4 blocks of 4
trials, depending on how many trials you can do at the same time.
– The experimental runs within a block are performed under the same
conditions; but different blocks encounter different conditions.
– If a design is replicated, the replicates can be run in blocks instead of
totally randomized.
– If a design is not replicated, the trials can still be run in blocks; the cost
is increased confounding (the Minitab menu will select a blocked
design for you with the least confounding possible).
Situations Where Experiments May Need to Be Blocked
– Can‟t do all – Can‟t make all parts

from similar raw – Can‟t make tests under
runs at one similar conditions
time material
• Lots • Machines
• Days • Workers
• Batches
• Shifts • Customers
• Regions
• Location • Environment
Using Minitab to Design Blocked Experiment
4
• Example: 2 in 2 blocks of Size 8
Stat > DOE > Create Factorial Design > Factors “4” > Design

Interpreting Block Effects
– If any of the block effects are judged to be important, we know the blocks differ
from each other.
– This is not a surprise since the design was blocked to control known sources of
variation.
– To find how the blocks differ from each other, you must find the averages of all
blocks and compare them. This8 is best done on a plot. Here’s an example with 4
blocks. 7
6
Response 5
Average 4
3
2
1
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.)

Things to Know About Blocked Designs
– Block effects are confounded with factor effects,
usually higher-order interactions.
• Confounding ―blocks‖ with ―factors‖ does not occur if you
have a factorial that is replicated in blocks.
– An assumption made with blocked designs is that

interactions between blocking and other factors are
negligible.
• This means that main factor effects and interactions are the
same for all blocks.
• Example: If there are two shifts, treat each shift as a block; the
assumption is that the main effect for Factor A is the same in
Shift 1 and Shift 2.

Exercise
• Objective: To perform the analysis of a Four-Factor Experiment with
two blocks
• Problem: A chemical engineer is interested in maximizing Filtration
Rate of a chemical product produced in a pressure vessel. The experiment
will take 16 runs but only 8 runs can be completed in a day. Two days are
necessary for the completion of the entire experiment.
• Output: Filtration Rate (gal/h)
• Inputs:
– Temperature (T)
– Pressure (P)
– Formaldehyde Concentration (F)
– Stirring Rate (S)
• Procedure:
– Use Minitab to design the experiment with 4 Factors in 16 runs with 2
Blocks
– Analyze the data using the provided data
Exercise Data from Minitab
Worksheet Block DOE

Tempareture
Pressure Concentration
Stirring Rate
Day Filtrate
40 4 5% 10 RPM 1 71
60 3 5% 10 RPM 1 48
60 4 8% 10 RPM 1 68
40 3 8% 10 RPM 1 65
60 4 5% 15 RPM 1 43
40 3 5% 15 RPM 1 104
40 4 8% 15 RPM 1 86
60 3 8% 15 RPM 1 70
60 4 5% 10 RPM 2 45
40 3 5% 10 RPM 2 65
40 4 8% 10 RPM 2 60
60 3 8% 10 RPM 2 80
40 4 5% 15 RPM 2 100
60 3 5% 15 RPM 2 45
60 4 8% 15 RPM 2 75
40 3 8% 15 RPM 2 96
Data are in Blockfactorial.MTW. Complete the analysis

Nested Hierarchical Design
1 2 3
1 2 3 4 1 2 3 4 1 2 3 4

Nested or Hierarchical Design
• When the levels of one factor(Say B) are similar but not identical for
different levels of another factor (say A).
• Example : A company purchase raw material from 3 different
supplier. We need to determine the level of purity of each supplier. 4
batches of material available for each supplier and three samples
are taken from each batch. The situation is given below.
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.
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

Analysis of 2-stage Nested Design
MINITAB Command
2. Enter the Model as follows and Okay
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.
Design & Analysis of Nested & Crossed Factors

• In some cases some factors will be crossed and other factors will
be nested.
• An engineer want to study the effect of Layout, Fixture and
Operators on assembly time. Due to different location it is difficult
to use same operator in each layout.
• Here Operators are nested within the levels of layouts, while fixture
and layouts are crossed.
• Open the worksheet Assembly time on Nested.mpj
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


MINITAB Command
1. GO TO STAT > ANOVA> GLM and get the Window and give the commands
2. Enter the Model as follows

and Okay
Time depends on the

model which have nested
and crossed factors

The ANOVA table as follows;
Layout 1 4.083 4.083 4.083 0.34 0.581
Fixture 2 82.792 82.792 41.396 7.55 0.008
Operator(Layout) 6 71.917 71.917 11.986 2.18 0.117
Layout*Fixture 2 19.042 19.042 9.521 1.74 0.218
Fixture*Operator(Layout) 12 65.833 65.833 5.486 2.35 0.036
Error 24 56 56 2.333
Total 47 299.667
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
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.
• You are carrying out ½ the number runs, hence you

call it 25 Fractional Factorial run of order ½.

EXAMPLE-FRACTIONAL FACTORIALS
If the residue is dependent on 5 factors viz..
NO. FACTOR -1 +1
1 Conveyor Speed(m/min) 500 1500

2 Height of Acid (m) 1 3
3 Agitation rate (rpm) 100 120
4 Temperature (0C) 40 90
5 Conc. % 3 6
HOW MANY EXPERIMENTS REQUIRED? 25 = 32

THE RESULTS OF THE EXPTS ARE AS BELOW
Variables - +
1 Height of Acid (mtr) 1 3
2 Conveyor speed 500 1500
3 Agitation rate(rpm) 100 120
4 Temperature 40 90
5 Concentration (%) 3 6
RUN 1 2 3 4 5 Y
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

MAIN EFFECTS PLOT
75
R 70
e
s
65
i
d
u 60
e
50
Ht. Acid Conv.speed agitate temp conc

INTERACTION PLOTS
Residue Residue
90 87
90 78
Temperature=90ºC
80 Concentration = 3%
80 73
70 64 70
60 56 59 Concentration=6%
60
50 54
Temperature=40º 51
50
C
500 500
Conv. speed Conv. speed
150 1500
0 90 79
Concentration = 3%
80
Residue
70 64
61 Concentration=6%
60
58
50 Temperature (ºC)
40 90
THIS WAS A FULL 25 FACTORIAL EXPT.
• It requires 32 Runs.
• We would like to reduce the Runs.
• This is done by running a fractional factorial expt.

i.e. instead of 32 Runs, 16 (1/2 fraction) or 8 Runs
(1/4 fraction) are performed.
Do you lose any information capabilities by

fractionalising ?
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
TOTAL 32 Runs
THE HIGH ORDER INTERACTIONS NORMALLY

ARE NOT SIGNIFICANT
WHEN YOU USE A FRACTIONAL FACTORIAL
• The effect of High Order Interactions get mixed with

those of main effects and Low Order Interactions.
e.g. Main Effect 1  (Main Effect of 1)actual + (5 way
interaction)
• Since High Order Interactions are lower, the difference
in interpretations may not be much different.
• This combining which occurs when you use fractional
factorials is called CONFOUNDING
HIGHER THE FRACTIONAL - MORE THE

CONFOUNDING

EXAMPLE- THE FACTORS AND THEIR LEVELS ARE AS BELOW
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
• How many Factors = 5

• How many Levels = 2
• How many Runs Ideally = 32
• No. of Runs We want to Run = 8
• What Experiment do we Use?
25 Fractional factorial expt. of 1/4 fraction

THE EXPERIMENT AND ITS RESULTS

RUN VOLTAGE COAT HNG SIDE FLPR SUP DFT
1 50 0 FR 0.5 PP 46
2 50 0 BK 2 PC 65
3 50 1.5 FR 0.5 PC 20
4 50 1.5 BK 2 PP 25
5 100 1.5 FR 2 PP 100
6 100 1.5 BK 0.5 PC 72
7 100 0 FR 2 PC 100
8 100 0 BK 0.5 PP 100

ANALYSIS OF RESULTS
AT +1 LEVELS :
100 1.5 BK 2 PC
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
AVG. 39 77.75 66.5 59.5 67.75

EFFECT DIFFERENCE
54 -23.5 -1 13 -3
MAIN EFFECT PLOT

100
90
80
70
60
50
40
30
20
10
50 100 0 1.5 FR BK 0.5 2 PP PC
VOLTAGE COAT HNG SIDE FL PRESS SUPPLIERS

In 2 **5 experiment with 2 replications :
Total Degrees of Freedom = 63

Out of that Error DF = 32
Main Effects D F = 5
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.

Taguchi’s Design
Let us say y = f(x,z)

where y = Product Response
x’s = controllable factors
z’s = noise factors
The objective is to choose settings of x that will make
the product’s response insensitive to variability
associated with both x and z and still meet target
specifications with least variability.

Orthogonal Arrays
The symbology for an orthogonal array is La(bc) where;

L = Latin Square
a = The number of test trials
b = The number of levels for each column, and
c = The number of columns in the array.
L4 (23) Orthogonal Array
No. 1 2 3
1 1 1 1
2 1 2 2
3 2 1 2
4 2 2 1
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
This is a special array and no interaction is estimated.


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
L8 (41 x 24) Orthogonal Array

No. 1-2-3 4 5 6 7
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


No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
2 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2
3 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2
4 1 1 1 2 2 2 2 2 2 2 2 1 1 1 1
5 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2
6 1 2 2 1 1 2 2 2 2 1 1 2 2 1 1
7 1 2 2 2 2 1 1 1 1 2 2 2 2 1 1
8 1 2 2 2 2 1 1 2 2 1 1 1 1 2 2
9 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2
10 2 1 2 1 2 1 2 2 1 2 1 2 1 2 1
11 2 1 2 2 1 2 1 1 2 1 2 2 1 2 1
12 2 1 2 2 1 2 1 2 1 2 1 1 2 1 2
13 2 2 1 1 2 2 1 1 2 2 1 1 2 2 1
14 2 2 1 1 2 2 1 2 1 1 2 2 1 1 2
15 2 2 1 2 1 1 2 1 2 2 1 2 1 1 2
16 2 2 1 2 1 1 2 2 1 1 2 1 2 2 1


No. 1 2 3 4
1 1 1 1 1
2 1 2 2 2
3 1 3 3 3
4 2 1 2 3
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

Mixed Series Orthogonal Arrays

L18 (21 x 37) Orthogonal Array
No. 1 2 3 4 5 6 7 8
1 1 1 1 1 1 1 1 1
2 1 1 2 2 2 2 2 2
3 1 1 3 3 3 3 3 3
4 1 2 1 1 2 2 3 3
5 1 2 2 2 3 3 1 1
6 1 2 3 3 1 1 2 2
7 1 3 1 2 1 3 2 3
8 1 3 2 3 2 1 3 1
9 1 3 3 1 3 2 1 2
10 2 1 1 3 3 2 2 1
11 2 1 2 1 1 3 3 2
12 2 1 3 2 2 1 1 3
13 2 2 1 2 3 1 3 2
14 2 2 2 3 1 2 1 3
15 2 2 3 1 2 3 2 1
16 2 3 1 3 2 3 1 2
17 2 3 2 1 3 1 2 3
18 2 3 3 2 1 2 3 1

Selection Of Array - Guidelines
Array Series Effects can be estimated Remarks

L4(23) Pure 2K Main effect & Interaction effect Only 2 level Factors
7
L8(2 ) Pure 2K Main effect & Interaction effect Only 2 level Factors
15 K
L16(2 ) Pure 2 Main effect & Interaction effect Only 2 level Factors
L32(231) Pure 2K Main effect & Interaction effect Only 2 level Factors
L12(211) Pure 2K Only Main effect & no Interaction
effect Only 2 level Factors
19 K Only Main effect & no Interaction
L20(2 ) Pure 2
L24(223) Pure 2K Only Main effect & no Interaction
4 K Main effect & Interaction effect
L9(3 ) Pure 3 Only 3 level Factors
13 K
L27(3 ) Pure 3 Main effect & Interaction effect Only 3 level Factors
L18(21 X 37) Mixed Main effect & Interaction effect Only 1 interaction between 2 and 3
factor can be estimated.
For other arrays and requirements, discuss with MBB

SELECTION OF DESIGN LAYOUT
 Compute Total Degrees of Freedom Required to estimate all

Factorial Effects of Interest (TDF).
 Minimum number of Experiments (MNE) = (TDF)+1
 Choose an OA nearest to the size of Run from MINITAB.
 Let us design and analyze a experiment using Taguchi‟s method

Reaction Example
• Objective: To design and analyze a fractional factorial experiment using
Minitab [Taguchi Design]
• Output Variable: % Reacted
• Inputs:
• Feed Rate (liters/minute)[A] 10, 15
• Catalyst (%) [B] 1,2
• Agitation Rate (rpm) [C] 100, 120
• Temperature (C) [D] 140,180
• Concentration (%) [E] 3, 6
• Use Minitab to setup the Design Matrix
• All the main effects and the following interaction to be estimated.
• Feed Rate & Temperature
• Feed Rate & Agitation rate
• Temperature & Agitation rate
• Catalyst & Concentration

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
•DEGREES OF FREEDOM REQUIRED:

FACTORS DOF
A 1
B 1
C 1
D 1
E 1
AXC 1X1 = 1
AXD 1X1 = 1
DXC 1X1 = 1
BXE 1X1 = 1
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.
Designing a Taguchi‟s Experiment
Go to Stat>DOE>Taguchi>Create Taguchi Design

Reaction Example- Designing the experiment


Carry out the trials randomly and with at least 1 replicates.

Enter the data in next 2 columns and then analyse it.
Reaction Example- Analyse the experiment
Select Stat>DOE>Taguchi>Analyse Taguchi Design and click.
Enter the response


Reaction Example- Interpret the experiment

MINITAB Output
Agitation Rate and Temperature is most critical factor as the delta is

high in both cases. In case of conflict tradeoff.

Agitation Rate and Temperature is most critical factor as the slope is high.
Interaction exists between Agitation Rate and Temperature, because the lines are not parallel.

How to get ANOVA Step –1 Copy the 2nd response and stack it below
the 1st response.
Step-2 : Copy the design Matrix and paste it below.
Step – 3: Run MINITAB Command
Stat>ANOVA>General Linear Model
In Response enter the response column
In Model Enter all the main effects and selected
interaction as decided earlier
Click Okay and the ANOVA will be as below.
Check the p value and if p < 0.05 then

the effect is significant.

Reaction Example- Optimisation and Prediction
•Select Optimum , trade off if necessary .
• The significant effects are main effect of Agitation,
Temperature and interaction effect of them.
•The selected level is Agitation rate 120 and
Temperature 140
•Predict the optimum response.
•Predict the optimum response at Agitation rate of 120
and Temperature of 140.
MINITAB Command
Goto STAT>DoE>Taguhi>Predict Taguchis result and
click. And get the window

Selct the terms, Optimum levels of selected terms and click o
•Confirm , Confirm , Confirm the results by actual Trial.

Class Exercise
Time Limit : 45 mins
• Objective: To design and analyze a fractional factorial
experiment using Minitab
• Output Variable: Taste Level (Higher Is Better)
• Inputs:
– Oak Type (A) Allier, Troncais
– Stems (B) None, All
– Barrel Toast ( C) Light Medium
– Temperature (D) 75 , 92
– Age of Barrel (E) Old, New
• The interaction AB, CD is important.
• Use Minitab to setup the design matrix
Modeled After Problem 9-26 In Montgomery

Add the Data and analyse.

• The experiment has been performed and the data
is below. Add it to your worksheet.
Run 1 56 54
Run 2 53 52
Run 3 63 65
Run 4 65 63
Run 5 53 55
Run 6 55 57
Run 7 67 69
Run 8 61 59
Run 9 69 71
Run 10 45 48
Run 11 78 79
Run 12 93 94
Run 13 49 51
Run 14 60 61
Run 15 95 97
Run 16 82 85
Analyse the data and present. Time limit 45 mins

Parameter Design

Quality Engineering:
•Quality Engineering is an engineering optimization strategy

developed by a Japanese Engineer, Dr. Genichi Taguchi, in
the late 1940’s. When faced with solving an engineering
concern and armed only with scientific/statistical tools, he
formed a philosophy marrying statistics and engineering
methods to achieve rapid improvements in quality and
costs, by optimizing product/process designs. His career
has been dedicated to the improvement, evolution, and
practical implementation of these methods.
•Basically Quality Engineering deals with activities

performed for the purpose of reducing variability in
product/process function. There are two forms of Quality
Engineering: Off-line and On-line Quality Engineering.

Off-line Quality Engineering :

•Off-line Quality Engineering activities are performed upstream in
the life of a product or process design. Their intent is to optimize the
functionability of an engineered system through his design
strategies. This concept forces the engineer to focus on “how” to
divert energy into ideal functioning, rather than concentration on
specific customer related problems. This aides the engineer in
conducting more efficient industrial research to enhance
technological capability.
On-line Quality Engineering :
On-line Quality Engineering activities are utilized downstream in the
manufacturing and field service environments. Their purpose is to
control the behavior of the process/product through control
mechanisms, inspection or preventive maintenance. It’s aim is to
balance the relationship between monitoring quality and the cost
associated with this monitoring.

8 stages of Parameter Design :
The basic procedures in parameter design are as follows :
1. Define project scope/objectives

- Define project objectives
-Identify the system or subsystems
-Select team leader and members
-Establish overall strategies
2. Identify Ideal Function

-Establish Intent, Desired results
-Define Input Signal and Output responses
-Define Input Function
-Determine measurements feasibility
3. Develop Signal and Noise Factor Strategies

-Define Signal levels and ranges
-Identify all noise factors
-Select critical noise factors and set levels
-Determine noise strategies

8 stages of Parameter Design :

4. Establish Control Factors and Levels
Identify all control factors
Select critical control factors and set levels
Select orthogonal array
Assign control factors to orthogonal array
5. Conduct experiments
Plan/prepare for experiment
Conduct experiment
Collect data
6. Conduct Data Analysis

Calculate S/N ratios and 's
Complete/interpret response tables/graphs
Perform 2-step optimization
Make predictions
7. Conduct Confirmation Run
8. Implement and Document Results

An Example of a Parameter Design Experiment Plan

Design Parameter Matrix Noise Factor Matrix Performance Performance
(Inner Array) (Outer Array) Characteristic Statistic
Control Factors Noise Factors Response S/N Ratio
Test Design Parameters

Runs 1 2 3 4 W1 W2 W3
1 1 1
1 1 1 1 1
Y1
1 2 2
2
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
In tile manufacturing process, the dimension of tiles varying

Beyond acceptance limit.
Raw Material
Crushing & Moulding Calcining Glazing Calcining
Mixing
The Tile Manufacturing Process

Step-2 : Identify Ideal Function
This was a Nominal-the-Best type of response and related to the energy

transformation of the heat treatment process.
Y = m with m is the nominal dimension set at the process.
Step-3 : Develop the Noise Strategy
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.
Noise Factors : P: Position of tile on cart

P1: Outside Top P2: Outside Middle P3: Outside Bottom
P4: Inside Top P5: Inside Middle P6: Inside Bottom

Step-4 : Establish Control Factors and Levels

TABLE: FACTOR LISTINGS
A: Amount of Lime Stone A1: Existing A2: 50% or more

B: Type of Additive A B1: Existing B2: New
C: Fitness of Additive C1: Existing C2: Coarser
D: Amount of Agalmatolite D1: Existing D2: 10% more
E: Charging Quantity E1: Existing E2: 100 kg less
F: Percent Waste Return F1: Existing F2: 4%
G: Amount of Feldspar G1: Existing G2: 0%
H: Type of Agalmatolite H1: Existing H2: New
I: Chamotte I1: Existing I2: 20% more
J: Type of Additive Z J1: Existing J2: New
K: particle size of additives K1: Existing K2: Smaller

Planning for the Trial
In traditional Design of Experiments, there is a technique called fractional
factorial design. Instead of running all possible combinations, you can run a
fraction of the full factorial so that you can study certain selected interactions.
Example: To study five 2-level factors and all of the two factor interactions
between them, it takes only 16 runs instead of 32 runs using the full factorial.
Orthogonal Arrays
Another method of experimentation makes use of orthogonal arrays.
No. 1 2 3 4 5 6 7 8
1 1 1 1 1 1 1 1 1
2 1 1 2 2 2 2 2 2
3 1 1 3 3 3 3 3 3
4 1 2 1 1 2 2 3 3
5 1 2 2 2 3 3 1 1
6 1 2 3 3 1 1 2 2
7 1 3 1 2 1 3 2 3
8 1 3 2 3 2 1 3 1
9 1 3 3 1 3 2 1 2
10 2 1 1 3 3 2 2 1
11 2 1 2 1 1 3 3 2
12 2 1 3 2 2 1 1 3
13 2 2 1 2 3 1 3 2
14 2 2 2 3 1 2 1 3
15 2 2 3 1 2 3 2 1
16 2 3 1 3 2 3 1 2
17 2 3 2 1 3 1 2 3
18 2 3 3 2 1 2 3 1

Orthogonality
The term "Orthogonal" means "balanced" or "separable"
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?

Step-5 : Conduct Experiment and collect data
ORTHOGONAL ARRAY, TILE DIMENSION AND SIGNAL-TO-NOISE RATIO

A B C D E F G H I J K
L12 P1 P2 P3 P4 P5 P6 Avg. S/N
1 2 3 4 5 6 7 8 9 10 11
1 1 1 1 1 1 1 1 1 1 1 1 151.9 151.4 150.4 150.2 149.6 149.5 150.5 43.8
2 1 1 1 1 1 2 2 2 2 2 2 157.0 156.8 156.0 153.0 152.7 152.0 154.6 36.7
3 1 1 2 2 2 1 1 1 2 2 2 150.5 150.2 149.8 147.1 146.6 145.5 148.3 36.8
4 1 2 1 2 2 1 2 2 1 1 2 152.2 151.3 151.1 150.6 150.1 150.0 150.9 45.2
5 1 2 2 1 2 2 1 2 1 2 1 149.7 149.5 149.4 147.2 145.0 144.1 147.5 35.5
6 1 2 2 2 1 2 2 1 2 1 1 154.3 154.0 153.6 152.8 152.2 151.9 153.1 43.8
7 2 1 2 2 1 1 2 2 1 2 1 156.5 152.1 150.3 148.5 146.3 144.6 149.7 30.9
8 2 1 2 1 2 2 2 1 1 1 2 154.5 153.3 151.8 150.9 150.4 149.6 151.8 38.3
9 2 1 1 2 2 2 1 2 2 1 1 153.0 152.5 152.0 151.9 150.9 149.0 151.6 40.5
10 2 2 2 1 1 1 1 2 2 1 2 151.5 150.8 150.0 149.8 149.4 149.1 150.1 44.5
11 2 2 1 2 1 2 1 1 1 2 2 153.1 151.8 151.8 151.4 150.6 150.3 151.5 43.6
12 2 2 1 1 2 1 2 1 2 2 1 151.5 150.8 150.6 150.2 149.9 149.7 150.5 47.2

Step-6 : Conduct Data Analysis
Calculating the Signal-to-Noise Ratio (S/N Ratio)

The S/N Ratio most commonly used for the nominal-the- best
characteristic, and one used in this case study, is given below. In theory,
the larger-the S/N ratio, the higher the quality. Therefore, emphasis is
placed on always maximizing the S/N ratio.
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

Response Analysis
RESPONSE TABLE FOR SIGNAL-TO-NOISE RATIOS
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
Example : The average S/N Ratio when factor A was at its

level 1 condition would be calculated as follows.
43 .8  36 .7  36 .8  45 .2  35 .5  43 .8
A1   40 .3 ( dB )
6
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

Response Analysis
RESPONSE TABLE FOR AVERAGE
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

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.
Prediction of the Optimized Response (S/N Ratio):
  ( 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.

Prediction of the Optimized Response (Process Average):


  ( C 1  T )  ( E 2  T )  ( F 1  T )  ( G1  T )  ( I 2  T )  ( J 1  T )  T
( T  average of 12 Averages)

  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.

Step-7 : Conduct Confirmation Run

The Confirmation Trial
In order to test the reproducibility of experimental conclusions, a
confirmation trial was conducted using the optimum condition for
maximizing robust ness:
A2 B2 C1 D1 E2 F1 G1 H1 I2 J1 K2
Process centering was achieved by changing the mold dimension.
EXPERIMENTAL RESULTS
Prediction Confirmation
Existing 43.8 (dB) 44.1 (dB)
Optimal 51.3 (dB) 51.7 (dB)
Gain 7.5 (dB) 7.6 (dB)
Step-8 : Implement and document Result

Parameter Design-Discussions:
Interpretation of the confirmation results
In this experiment, 43.8 db improvement was predicted and 44.1 db
improvement was confirmed. This confirms excellent additivity, or excellent
reproducibility or experimental conclusions. We are confident in the
factorial effects we have observed. We are able to reproduce these
factorial effects in an additive fashion.
However, what if it does not confirm? Consider the following situations.
Why conclusions may not confirm ?
SITUATION-1 Prediction Confirmation
Existing Design 43.8 (dB) 44.1 (dB)
Optimal Design 51.3 (dB) 51.7 (dB)
Gain 7.5 (dB) 7.6 (dB)

Situation-1:
This is the same situation as what we have covered in the previous
example. This situation shows excellent additivity and reproducibility. Both
S/N values and gain are confirmed. We have generated reliable knowledge
from this experiment.
Parameter Design-Discussion:
Gain 7.5 (dB) 6.5 (dB)

Situation-2:
6.5 db gain has confirmed when 7.5 db gain was predicted. Although this is not
perfect, it is not bad at all. Most of the factorial effects on variability have been
reproduced.
Situation-3. Gain 7.5 (dB) 3.5 (dB)

Things are getting shaky. Only 3.5 db has confirmed. Only half of what we
expected has been reproduced. Additivity is poor. Notice that the confirmed S/N
for the optimum is closer to existing than to the prediction. The question is, what
can we conclude from this? We will get to this question momentarily.
Gain 7.5 (dB) 1.5 (dB)

Situation-4:
Only 1.5 db improvement. S/N for the optimum is almost the same as the
experimental average S/N. Just about none of those effects were additive, hence
not reproducible. It is telling us something. We will discuss this momentarily.
Gain 7.5 (dB) 7.4 (dB)

Situation-5:
The gain has confirmed, but the absolute S/N values have not. This situation
shows that the optimum design is predictably better then the initial design by 11.1
db, but an unknown noise factor is consistently affecting both designs.

Reason-1
Strong Interaction between control factors
Reason-2
Noise Factor(s)
Reason - 3
Pure Experimental Error.

DoE Notes (Few Slides)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

DoE Notes (Few Slides)

Uploaded by

Copyright:

Available Formats

SQC & OR UNIT, INDIAN STATISTICAL INSTITUTE, HYDERABAD

Six Sigma Black Belt 1

• You will find the 1st cut of important X‟s

• At what levels each of these X‟s has to be set

• The best possible combination of levels for

Six Sigma Black Belt 2

STATISTICALLY DESIGNED EXPERIMENTS

 Even a limited number of experiments would enable the

Six Sigma Black Belt 3

 Improving efficiency or yield

Six Sigma Black Belt 4

NEED FOR DESIGNED EXPERIMENT

• A design engineer is interested in studying the effect Bronze and

•Here, the objective is to determine the material that produces the

•The procedure thought of by the engineer is to make a number of Bush

Six Sigma Black Belt 5

NEED FOR DESIGNED EXPERIMENT

PROOF OF THE NEED FOR EXPERIMENTATION

 However if the problem is of chronic nature and there is

IF CONTROL ONLY IS YOUR PROBLEM

WE SEEK ANSWERS TO THE FOLLOWING

 Which X‟s are important ?

 Are they controllable ?

 Do they change by chance or by intent (assignable) ?

Six Sigma Black Belt 8

IF CONTROLLABLE AND CHANGE BY CHANCE

E.g: You find that finish (Y) deteriorates with time.

USE CONTROL CHARTS ON YOUR Y

Six Sigma Black Belt 9

IF CONTROLLABLE AND CHANGE BY CHANCE

IN CONTROL PHASE YOU WILL SEE THIS IN DETAIL

Six Sigma Black Belt 10

IF CONTROLLABLE & CHANGE BY INTENT

• The change in material on line is done by intent and

• In such cases where you know when a critical X has

D.O.E. IS THE TOOL FOR IT

IF VARIATION IS OUR PROBLEM

• Find at what level each of the X has to be

• D.O.E. is the tool for it

Six Sigma Black Belt 12

Design of experiments (DOE) is a valuable tool to

Six Sigma Black Belt 13

Six Sigma Black Belt 14

Speed Octane Tire Pressure MPG

Six Sigma Black Belt 16

While we would have made significant improvement, we would

Speed Octane Tire Pressure MPG

What conclusion do you make now?

TERMINOLOGY USED IN D.O.E.

OUTCOME (RESPONSE): The numerical result of a trial based on a

The response may be :

E.g.: diameter of a shaft, No. of rejected cylinders etc.

Six Sigma Black Belt 19

BALANCED TEST - Where number of samples in each treatment

Six Sigma Black Belt 20

INTERACTION EFFECT : If the effect of one factor is different at

The interaction between factors A and B, is termed as “First Order

If the interaction between two factors A and B, is different at different

Six Sigma Black Belt 21

TERMINOLOGY USED IN D.O.E.

e.g: leakage of dome welded components is a function of current

Hence there is interaction between electrode and current

Six Sigma Black Belt 22