2010 RAMS Doe and Data Analysis

Copyright 2010 IEEE.
Reprinted from 2010 Reliability and

Maintainability Symposium, San Jose, CA, USA, January 2528, 2010.
This material is posted here with permission of the IEEE. Such
permission of the IEEE does not in any way imply IEEE
endorsement of any of ReliaSoft Corporation's products or
services. Internal or personal use of this material is permitted.
However, permission to reprint/republish this material for
advertising or promotional purposes or for creating new
collective works for resale or redistribution must be obtained
from the IEEE by writing to pubs-permissions@ieee.org.
By choosing to view this document, you agree to all provisions
of the copyright laws protecting it.
2010 Annual RELIABILITY and MAINTAINABILITY Symposium
Design of Experiments and Data Analysis
Huairui Guo, Ph. D. & Adamantios Mettas
Huairui Guo, Ph.D., CPR.

ReliaSoft Corporation
1450 S. Eastside Loop
Tucson, AZ 85710 USA
e-mail: Harry.Guo@ReliaSoft.com
Adamantios Mettas, CPR

ReliaSoft Corporation
1450 S. Eastside Loop
Tucson, AZ 85710 USA
e-mail: Adam.Mettas@ReliaSoft.com
Tutorial Notes 2010 AR&MS
SUMMARY & PURPOSE

Design of Experiments (DOE) is one of the most useful statistical tools in product design and testing. While many
organizations benefit from designed experiments, others are getting data with little useful information and wasting resources
because of experiments that have not been carefully designed. Design of Experiments can be applied in many areas including
but not limited to: design comparisons, variable identification, design optimization, process control and product performance
prediction. Different design types in DOE have been developed for different purposes. Many engineers are confused or even
intimidated by so many options.
This tutorial will focus on how to plan experiments effectively and how to analyze data correctly. Practical and correct
methods for analyzing data from life testing will also be provided.
Huairui Guo, Ph.D., CRP

Huairui Guo is the Director of Theoretical Development at ReliaSoft Corporation. He received his Ph.D. in Systems and
Industrial Engineering from the University of Arizona. He has published numerous papers in the areas of quality engineering
including SPC, ANOVA and DOE and reliability engineering. His current research interests include repairable system
modeling, accelerated life/degradation Testing, warranty data analysis and robust optimization. Dr. Guo is a member of SRE,
IIE and IEEE. He is a Certified Reliability Professional (CRP).
Adamantios Mettas, CRP

Mr. Mettas is the Vice President of product development at ReliaSoft Corporation. He fills a critical role in the advancement of
ReliaSoft's theoretical research efforts and formulations in the subjects of Life Data Analysis, Accelerated Life Testing, and
System Reliability and Maintainability. He has played a key role in the development of ReliaSoft's software including,
Weibull++, ALTA and BlockSim, and has published numerous papers on various reliability methods. Mr. Mettas holds a B.S
degree in Mechanical Engineering and an M.S. degree in Reliability Engineering from the University of Arizona. He is a
Certified Reliability Professional (CRP).
Table of Contents
1.
2.
3.
4.
5.
6.
7.
8.
Introduction ..........................................................................................................................................................................1
Statistical Background ..........................................................................................................................................................2
Two Level Factorial Design .................................................................................................................................................3
Response Surface Methods (RSM) ......................................................................................................................................6
DOE for Life Testing ...........................................................................................................................................................9
Conclusions ........................................................................................................................................................................10
References ..........................................................................................................................................................................11
Tutorial Visuals.. .................................12
ii Guo & Mettas
2010 AR&MS Tutorial Notes
1.
INTRODUCTION
The most effective way to improve product quality and

reliability is to integrate them in the design and manufacturing
process. Design of Experiments (DOE) is a useful tool that can
be integrated into the early stages of the development cycle. It
has been successfully adopted by many industries, including
automotive, semiconductor, medical devices, chemical
products, etc. The application of DOE is not limited to
engineering. Many successful stories can be found in other
areas. For example, it has been used to reduce administration
costs, improve the efficiency of surgery processes, and
establish better advertisement strategies.
1.1 Why DOE
DOE will make your life easier. For many engineers,
applying DOE knowledge in their daily work will reduce lots
of trouble. Here are two examples of bad experiments that will
cause trouble.
Example 1: Assume the reliability of a product is affected by
voltage. The usage level voltage is 10. In order to predict the
reliability at the usage level, fifty units are available for
accelerated life testing. An engineer tested all fifty units at a
voltage of 25. Is this a good test?
Example 2: Assume the reliability of a product is affected by
temperature and humidity. The usage level is 40 degrees
Celsius and 50% relative humidity. In order to predict the
reliability at the usage level, fifty units are available for
accelerated life testing. The design is conducted in the
following way:
Number of
Units
Temperature
(Celsius)
Humidity
(%)
25
120
95
25
85
85
Table 1 Two Stress Accelerated Life Test

Will the engineer be able to predict the reliability at the usage
level with the failure data from this test?
1.2 What DOE Can Do
DOE can help you design better tests than the above two
examples. Based on the objectives of the experiments, DOE
can be used for the following purposes [1, 2]:
1. Comparisons. When you have multiple design options,
several materials or suppliers are available, you can design an
experiment to choose the best one. For example, in the
comparison of six different suppliers that provide connectors,
will the components have the same expected life? If they are
different, how are they different and which is the best?
2. Variable Screening. If there are a large number of
variables that can affect the performance of a product or a
system, but only a relatively small number of them are
important, a screening experiment can be conducted to
identify the important variables. For example, the warranty
return is abnormally high after a new product is launched.
Variables that may affect the life are temperature, voltage,

duty cycle, humidity and several other factors. DOE can be
used to quickly identify the troublemakers and a follow-up
experiment can provide the guidelines for design modification
to improve the reliability.
3. Transfer Function Exploration. Once a small number of
variables have been identified as important, their effects on the
system performance or response can be further explored. The
relationship between the input variables and output response is
called the transfer function. DOE can be applied to design
efficient experiments to study the linear and quadratic effects
of the variables and some of the interactions between the
variables.
4. System Optimization. The goal of system design is to
improve the system performance, such as to improve the
efficiency, quality, and reliability. If the transfer function
between variables and responses has been identified, the
transfer function can be used for design optimization. DOE
provides an intelligent sequential strategy to quickly move the
experiment to a region containing the optimum settings of the
variables.
5. System Robustness. In addition to optimizing the
response, it is important to make the system robust against
noise, such as environmental factors and uncontrolled
factors. Robust design, one of the DOE techniques, can be
used to achieve this goal.
1.3 Common Design Types
Different designs have been used for different experiment
purposes. The following list gives the commonly used design
types.
1. For comparison
One factor design
2. For variable screening
2 level factorial design
Taguchi orthogonal array
Plackett-Burman design
3. For transfer function identification and optimization
Central composite design
Box-Behnken design
4. For system robustness
Taguchi robust design
The designs used for transfer function identification and
optimization are called Response Surface Method designs. In
this tutorial, we will focus on 2 level factorial design and
response surface method designs. They are the two most
popular and basic designs.
1.4 General Guidelines for Conducting DOE
DOE is not only a collection of statistical techniques that
enable an engineer to conduct better experiments and analyze
data efficiently; it is also a philosophy. In this section, general
guidelines for planning efficient experiments will be given.
The following seven-step procedure should be followed [1, 2].
1. Clarify and State Objective. The objective of the
experiment should be clearly stated. It is helpful to prepare a
Guo & Mettas 1
list of specific problems that are to be addressed by the

experiment.
2. Choose Responses. Responses are the experimental
outcomes. An experiment may have multiple responses based
on the stated objectives. The responses that have been chosen
should be measurable.
3. Choose Factors and Levels. A factor is a variable that is
going to be studied through the experiment in order to
understand its effect on the responses. Once a factor has been
selected, the value range of the factor that will be used in the
experiment should be determined. Two or more values within
the range need to be used. These values are referred to as
levels or settings. Practical constraints of treatments must be
considered, especially when safety is involved. A cause-andeffect diagram or a fishbone diagram can be utilized to help
identify factors and determine factor levels.
4. Choose Experimental design. According to the objective
of the experiments, the analysts will need to select the number
of factors, the number of level of factors, and an appropriate
design type. For example, if the objective is to identify
important factors from many potential factors, a screening
design should be used. If the objective is to optimize the
response, designs used to establish the factor-response
function should be planned. In selecting design types, the
available number of test samples should also be considered.
5. Perform the Experiment. A design matrix should be used
as a guide for the experiment. This matrix describes the
experiment in terms of the actual values of factors and the test
sequence of factor combinations. For a hard-to-set factor, its
value should be set first. Within each of this factors settings,
the combinations of other factors should be tested.
6. Analyze the Data. Statistical methods such as regression
analysis and ANOVA (Analysis of Variance) are the tools for
data analysis. Engineering knowledge should be integrated
into the analysis process. Statistical methods cannot prove that
a factor has a particular effect. They only provide guidelines
for making decisions. Statistical techniques together with good
engineering knowledge and common sense will usually lead to
sound conclusions. Without common sense, pure statistical
models may be misleading. For example, models created by
smart Wall Street scientists did not avoid, and probably
contributed to, the economic crisis in 2008.
7. Draw Conclusions and Make Recommendations. Once the
data have been analyzed, practical conclusions and
recommendations should be made. Graphical methods are
often useful, particularly in presenting the results to others.
Confirmation testing must be performed to validate the
conclusion and recommendations.
The above seven steps are the general guidelines for
performing an experiment. A successful experiment requires
knowledge of the factors, the ranges of these factors and the
appropriate number of levels to use. Generally, this
information is not perfectly known before the experiment.
Therefore, it is suggested to perform experiments iteratively
and sequentially. It is usually a major mistake to design a
single, large, comprehensive experiment at the start of a study.
2 Guo & Mettas
As a general rule, no more than 25 percent of the available

resources should be invested in the first experiment.
2.
STATISTICAL BACKGROUND
Linear regression and ANOVA are the statistical methods

used in DOE data analysis. Knowing them will help you have
a better understand of DOE.
2.1 Linear Regression[2]
A general linear model or a multiple regression model is:
(1)
Y = 0 + 1 X 1 + ... + p X p +
Where: Y is the response also called output or dependent
variable. X i is the predictor also called input or independent
variable. is the random error or noise, which is assumed to
be normally distributed with mean 0 and variance 2 , usually
noted as ~ N (0, 2 ) . Because is normally distributed,
then for a given value of X, Y is also normally distributed and
Var (Y ) = 2 .
From the model, it can be seen that the variation or the
difference of Y consists of two parts. One is the random part of
. The other is the difference caused by the difference of the
X values. For example, consider the data in Table 2:
Observation
X1
X2
120
90
300
120
90
350
85
95
150
85
95
190
120
95
400
120
95
430
Y
Mean
325
170
415
Table 2 Example Data for Linear Regression

Table 2 has three different combinations of X1 and X2,
showing at different colors. For each combination, there are
two observations. Because of the randomness caused by ,
these two observations are different although they have the
same X values. This difference usually is called within-run
variation. The mean values of Y at the three combinations are
different too. This difference is caused by the difference of X1
and X2 and usually is called between-run variation.
If the between-run variation is significantly larger than the
within-run variation, it means most of the variation of Y is
caused by the difference of X settings. In other words, Xs
significantly affect Y. The difference of Ys caused by the Xs
are much more significant than the difference caused by the
noise.
From Table 2, we have the feeling that the between-run
variation is larger than the within-run variation. To confirm
this, statistical methods should be applied. The amount of the
total variation of Y is defined by the sum of squares:
n
SST = Yi Y
i =1
)2
(2)
Where Yi is the ith observed value and Y is the mean of all
the observations. However, since SST is affected by the

number of observations, to eliminate this effect, another
metric called mean squares is used to measure the normalized
variability of Y.
SS
1 n
2
(3)
MST = T =
Yi Y
n 1 n 1 i =1
Equation (3) is also the unbiased estimator of Var (Y ) .

As mentioned before, the total sum of squares can be
partitioned into two parts: within-run variation caused by
random noise (called sum of squares of error SS E ) and the
between-run variation caused by different values of Xs (called
sum of squares of regression SS R ).
n
i =1
i =1
SS T = SS R + SS E = (Yi Y ) 2 + (Yi Yi ) 2
(4)
Where: Yi is the predicted value for the ith test. For tests with
the same X values, the predicted values are the same.
Similar to equation (3), the mean squares of regression
and the mean squares of error are calculated by:
2
SS
1 n
(5)
MS R = R = Yi Y
p
p i =1
MS E =
n
SS E
1
=
Yi Yi
n 1 p n 1 p i =1
)2
(6)
Where: p is the number of Xs.

When there is more than one input variable, SS R can be
further divided into the variation caused by each variable, such
as:
(7)
SS R = SS X1 + SS X 2 + ... + SS X p
The mean squares of each input variable MS X i is compared
with MS E to test if the effect of X i is significantly greater
than the effect of noise.
The mean squares of regression MS R is used to measure
the between-run variance that is caused by predictor Xs. The
mean squares of error MS E represents the within-run variance
caused by noise. By comparing these two values, we can find
out if the variance contributed by Xs is significantly greater
than the variance caused by noise. ANOVA is the method
used for the comparison in a statistical way.
2.2 ANOVA (Analysis of Variance)
can easily verify that:

SST = SS R + SS E = SS X1 + SS X 2 + SS E
The fifth column shows the F ratio of each source. All the
values are much bigger than 1. The last column is the P value.
The smaller the P value is, the larger the difference between
the variance caused by the corresponding source and the
variance caused by noise. Usually, a significance level ,
such as 0.05 or 0.1 is used to compare with the P values. If a
P value is less than , the corresponding source is said to be
significant at the significance level of . From Table 3, we
can see that both variables X1 and X2 are significant to the
response Y at the significance level of 0.1.
MS R
(8)
MS E
is used to test the following two hypotheses:
H0: There is no difference between the variance caused
by Xs and the variance caused by noise.
H1: The variance caused by Xs is larger than the variance
caused by noise.
Under the null hypothesis, the ratio follows the F distribution
with degree of freedoms of p and n 1 p . By applying
ANOVA to the data for this example, we get the following
ANOVA table.
The third column shows the values for the sum of squares. We
Degrees
of
Freedom
Source of
Variation
Sum of
Squares
Mean
Squares
F
Ratio
P
Value
Model
6.14E+04
3.07E+04
36.86
0.0077
X1
6.00E+04
6.00E+04
72.03
0.0034
X2
9.72
0.0526
8100
8100
Residual
2500
833.3333
Total
6.39E+04
Table 3 ANOVA Table for the Linear Regression Example

Another way to test whether or not a variable is
significant is to test whether or not its coefficient is 0 in the
regression model. For this example, the linear regression
model is:
(10)
Y = 0 + 1 X 1 + 1 X 2 +
If we want to test whether or not 1 = 0 , we can use the
following hypothesis:
H0: 1 = 0
Under this null hypothesis, the statistic is a t distribution:
1
(11)
se( 1 )
Se( 1 ) is the standard error of 1 that is estimated from the
data. The t test results are given in Table 4.
T0 =
Term
Coefficient
Standard Error
T Value
P Value
Intercept
-2135
588.7624
-3.6263
0.0361
X1
0.8248
8.487
0.0034
X2
18
5.7735
3.1177
0.0526
The following ratio
F0 =
(9)
Table 4 Coefficients for the Linear Regression Example

Table 3 and Table 4 give the same P values.
With linear regression and ANOVA in mind, we can start
discussing DOE now.
3.
TWO LEVEL FACTORIAL DESIGNS
Two level factorial designs are used for factor screening.

In order to study the effect of a factor, at least two different
settings for this factor are required. This also can be explained
from the viewpoint of linear regression. To fit a line, two
points are the minimal requirement. Therefore, the engineer
Guo & Mettas 3
who tested all the units at a voltage of 25 will not be able to

predict the life at the usage level of 10 volts. With only one
voltage value, the effect of voltage cannot be evaluated.
3.1 Two Level Full Factorial Design
When the number of factors is small and you have the
resources, a full factorial design should be conducted. Here
we will use a simple example to introduce some basic
concepts in DOE.
For an experiment with two factors, the factors usually are
called A and B. Uppercase letters are used for factors. The
first level or the low level is represented by -1, while the
second level or the high level is represented by 1. There are
four combinations of a 2 level 2 factorial design. Each
combination is called a treatment. Treatments are represented
by lowercase letters. The number of test units for each
treatment is called the number of replicates. For example, if
you test two samples at each treatment, the number of
replicates is 2. Since the number of replicates for each factor
combination is the same, this design is also balanced. A two
level factorial design with k factors usually is written as 2 k
design and read as 2 to the power of 3 design or 2 to the 3
design. For a 2 2 design, the design matrix is:
Factors
Response
Treatment
-1
-1
-1
-1
30
-1
25
ab
35
20
Table 5 Treatments for 2 Level Factorial Design

This design is orthogonal. This is because the sum of the
product
of
A
and
B
is
zero,
which
is
An
orthogonal
(1 1) + (1 1) + (1 1) + (1 1) = 0 .
design will reduce the estimation uncertainty of the model
coefficients.
The following linear regression model is used for the
analysis:
(12)
Y = 0 + 1 X 1 + 2 X 2 + 12 X 1 X 2 +
Where: X 1 is for factor A; X 2 is for factor B; and their
interaction is represented by X 1 X 2 . The effects of A and B
are called main effects. The effects of their interaction are
called two-way interaction effects. These three effects are the
three sources for the variation of Y. Since equation (12) is a
linear regression model, the ANOVA method and the t-test
given in Section 2 can be used to test whether or not one effect
is significant.
For a balanced design, a simple way to calculate the effect
of a factor is to calculate the difference of the mean values of
the response at its high and low setting. For example, the
effect of A can be calculated by:
4 Guo & Mettas
Effet of A = Avg. at A high - Avg. at A low

=
(13)
30 + 35 20 + 25
= 10
2
2
3.2 Two Level Fractional Factorial Design

When you increase the number of factors, the number of
test units will increase quickly. For example, to study 7
factors, 128 units are needed. In reality, responses are affected
by a small number of main effects and lower order
interactions. Higher order interactions are relatively
unimportant. This statement is called the sparsity of effects
principle. According to this principle, fractional factorial
designs are developed. These designs use fewer samples to
estimate main effects and lower order interactions, while the
higher order interactions are considered to have negligible
effects.
Consider a 2 3 design. 8 test units are required for a full
factorial design. Assume only 4 test units are available
because of the cost of the test. Which 4 of the 8 treatments
should you choose? A full design matrix with all the effects
for a 2 3 design is:
Order
1
A
-1
B
-1
AB
1
C
-1
AC
1
BC
1
ABC
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
Table 6 Design Matrix for a 2 3 Design

If the effect of ABC can be ignored, the following 4
treatments can be used in the experiment.
Order
AB
AC
BC
-1
-1
-1
-1
ABC
1
-1
-1
-1
-1
-1
-1
-1
-1
Table 7 Fraction of the Design Matrix for a 2 3 Design

In Table 7, the effect of ABC cannot be estimated from the
experiment because it is always at the same level of 1. Since
Table 7 uses only half of the treatments from the full factorial
design in Table 6, it is represented by 2 31 and read as 2 to

the power of 3 minus 1 design or 2 to the 3 minus 1 design.
From Table 7, you will also notice that some columns
have the same values. For example, column AB and C are the
same. Using equation (13) to calculate the effect of AB and C,
we will end up with the same procedure and result. Therefore,
from this experiment, the effect of AB and C cannot be
distinguished because they change with the same pattern. In

DOE, effects that cannot be separated from an experiment are
called confounded effects or aliased effects. A list of aliased
effects is called the alias structure. For the design of Table 7,
the alias structure is:
[I]=I+ABC; [A]=A+AC; [B]=B+BC; [C]=C+AB
Where: I is the effect of the intercept in the regression model,
which represents the mean value of the response. The alias for
I is called the defining relation. For this example, the defining
relation is written as I = ABC. In a design, I may be aliased
with several effects. The order of the shortest effects that
aliased with I is the resolution of this design.
From the alias structure, we can see that main effects are
confounded with 2-way interactions. For example, the
estimated effect for C in fact is the combination effect of C
and AB.
Checking Table 7, we can see it is a full factorial design if
we have only factor A and B. Therefore, A and B usually are
called basic factors and the full factorial design for them is
called the basic design. A fractional factorial design is
generated from its basic design and basic factor. For this
example, the values for factor C are generated from the values
of the basic factors A and B using the relation C=AB. Usually
AB is called the factor generator for C.
By now, it should be clear that the design given in Table 1
at the beginning of this tutorial is not a good design. If you
check the design in terms of coded value, the answer is
obvious. Table 8 shows the design again.
Number of
Units
Temperature
(Celsius)
25
120 (1)
95 (1)
25
85 (-1)
85 (-1)
Aperture Setting
small
large
Exposure Time
minutes
20
40
Develop Time
seconds
30
45
Mask Dimension
small
large
Etch Time
14.5
15.5
Table 9 Factor Settings for a Five Factor Experiment

With five factors the total number of runs required for a
full factorial is 2 5 = 32 . Running all of the 32 combinations is
too expensive for the manufacturer. At the initial
investigation, only main effects and two factor interactions are
of interest, while higher order interactions are considered to be
unimportant. It is decided to carry out the investigation using
the 2 51 design, which requires 16 runs. The defining relation
is I=ABCDE, or in other words, the generator for factor E is
E=ABCD. Table 10 gives the experiment data.
Run Order
Yield
Large
20
30
Large
15.5
10
Large
20
45
Large
14.5
21
Small
40
45
Small
15.5
45
Small
20
45
Small
14.5
16
Large
40
30
Small
15.5
52
Large
40
45
Small
14.5
60
Small
40
30
Large
15.5
30
Small
20
45
Large
15.5
15
Large
20
30
Small
14.5
10
Small
40
30
Small
14.5
34
11
Small
20
30
Small
15.5
12
Large
40
30
Large
14.5
50
13
Small
40
45
Large
14.5
44
14
Small
20
30
Large
14.5
15
Large
20
45
Small
15.5
22
16
Large
40
45
Large
15.5
63
Humidity
(%)
Table 8 Two Stress Accelerated Life Test (Coded Value)

In this design, temperature and humidity are confounded. In
fact, to study the effect of two factors, at least three different
settings are required. From the linear regression point of view,
at least three unique settings are needed to solve three
parameters: the intercept, the effect of factor A and the effect
of factor B. If their interaction is also to be estimated, four
different settings should be used. Many DOE software
packages can generate a design matrix for you according to
the number of factors and the level of factors. It will help you
avoid bad designs such as the one given in Table 8.
Table 10 Design Matrix and Results

Since the design has only 16 unique factor combinations, it
can be used to estimate only 16 parameters in the linear
regression model. If we include all main and 2-way
interactions in the model, we get the following ANOVA table.
3.3 An Example of a Fractional Factorial Design

Assume an engineer wants to identify the factors that
affect the yield of a manufacturing process for integrated
circuits. By following the DOE guidelines, five factors are
brought up and a two level fractional factorial design is
decided to be used [1]. The five factors and their levels are
given in Table 9.
Factor
Name
Unit
Low
(-1)
minutes
High
(1)
Source of
Variation
DF
Sum of
Squares
Mean
Squares
P
Value
Model
15
5775.4375
385.0292
495.0625
495.0625
4590.0625
4590.0625
473.0625
473.0625
3.0625
3.0625
1.5625
1.5625
AB
189.0625
189.0625
AC
0.5625
0.5625
AD
5.0625
5.0625
Guo & Mettas 5
AE
5.0625
5.0625
BC
1.5625
1.5625
BD
0.0625
0.0625
BE
0.0625
0.0625
CD
3.0625
3.0625
CE
0.5625
0.5625
DE
7.5625
7.5625
Residual
Total
15
495.0625
495.0625
193.1951
2.53E-08
4590.0625
4590.0625
1791.244
1.56E-13
473.0625
473.0625
184.6098
3.21E-08
73.7805
3.30E-06
1.3964
0.3128
189.0625
189.0625
Residual
11
28.1875
2.5625
Lack of Fit
9.6875
3.2292
Pure Error
18.5
2.3125
15
5775.4375
AB
Total
5775.4375
Table 11 ANOVA Table with All Effects
Table 12 ANOVA Table with Significant Effects
There are no F ratio and P values in the above table. This is

because there are no replicates in this experiment when all the
effects are considered. Therefore, there is no way to estimate
the error term in the regression model. This is why the SSE
(Sum of Squares of Error), labeled as Residual in Table 11, is
0. Without SSE, the estimation of the random error, how can
we test whether or not an effect is significant compared to
random error? Dont panic. Statisticians have already
developed methods to deal with this situation. When there is
no error in a screening experiment, Lenths method can be
used to identify significant effects. Lenths method assumes
that all the effects should be normally distributed with a mean
of 0, given the hypothesis that they are not significant. If any
effects are significantly different from 0, they should be
considered significant. So we can check the normal probability
plot for effects.
From Table 12, we can see that effects A, B, C and AB are

indeed significant because their P values are close to 0.
Once the important factors have been identified, followup experiments can be conducted to optimize the process.
Response Surface Methods are developed for this purpose.
ReliaSoft DOE++ - www.ReliaSoft.com
Normal ProbabilityPlot of Effect
99.000
Effect Probability
B:Exposure Time
A:Aperture Setting
Yield
Y' = Y
Non-Significant Effects
Significant Effects
Distribution Line
C:Develop Time
Probability
AB
50.000
10.000
5.000
1.000
-40.000
-24.000
-8.000
8.000
24.000
QA
Reliasoft
7/10/2009
3:03:18 PM
40.000
Effect
Alpha = 0.1; Lenth's PSE = 0.9375
Figure 1-Effect Probability Plot Using Lenths Method

From Figure 1, the main effect A, B, C and the 2-way
interaction AB are identified as significant at a significance
level of 0.1. Since the rest of the effects are not significant,
they can be treated as noise and used to estimate the sum of
squares of error. In DOE, it is a common practice to pool nonsignificant effects into error. With only A, B, C and AB in the
model, we get the following ANOVA table.
Source of
Variation
Model
DF
Sum of
Squares
Mean
Squares
F
Ratio
P Value
5747.25
1436.8125
560.7073
1.25E-12
6 Guo & Mettas
4.
RESPONSE SURFACE METHODS (RSM)
Response surface methods (RSM) are used to estimate the

transfer functions at the optimal region. The estimated
function is then used to optimize the responses. The quadratic
model is the model used in RSM. Similar to the factorial
design, linear regression and ANOVA are the tools for data
analysis in RSM. Lets use a simple example to illustrate this
type of design.
4.1 Initial Investigation
Assume the yield from a chemical process has been found
to be affected by two factors [1]:
Reaction Temperature
Reaction Time
The current operating conditions of 230 Fahrenheit and 65
minutes give a yield of about 35%. The engineers decide to
explore the current conditions in the range [L=225, H=235]
Fahrenheit and [L=55, H=75] minutes to see how the
temperature and time affect the yield. The design matrix is:
Std.
Order
1
Point
Type
1
A:
Temperature
-1
B:
Reaction Time
-1
-1
-1
35
37.25
35.45
35.75
36.05
35.3
35.9
Yield (%)
33.95
36.35
Table 13-Design Matrix for the Initial Experiment

Table 13 is the design matrix in coded values where -1 is the
lower level and 1 is the higher level. For a given actual value
for a numerical factor, its corresponding coded value can be
calculated by:
Actual Value - Middle Value of High and Low

Half of the Range Between High and Low
(13)
4.8
254
85
39.35
In Table 13, there are several replicated runs at the setting

of (0, 0). These runs are called center points. Center points
have the following two uses.
To estimate random error.
To check whether or not the curvature is significant.
At least five center points are suggested in an experiment.
When we analyze the data in Table 13, we get the
ANOVA table below.
From Table 14, we know curvature is not significant.
Therefore, the linear model is sufficient in the current
experiment space. The linear model that includes only
significant effects is:
(14)
Y = 35.6375 + 1.1625 X 1 + 0.4875 X 2
Equation (14) is in terms of coded value. We can see both
factors have a positive effect to the yield since their
coefficients are positive. To improve the yield, we should
explore the region with factor values larger than the current
operation condition. There are many directions we can move
to increase the yield, but which one is the fastest lane to
approach the optimal region? By checking the coefficient of
each factor in equation (14), we know the direction should be
(1.1625, 0.4875). This also can be seen from the contour plot
in Figure 2.
7.2
266
95
45.65
9.6
278
105
49.55
Coded value =
Source of
Variation
DF
Sum of
Squares
Mean
Squares
F Ratio
P Value
6.368
1.592
16.4548
0.0095
5.4056
5.4056
55.8721
0.0017
0.9506
0.9506
9.8256
0.035
AB
0.0056
0.0056
0.0581
0.8213
Curvature
0.0061
0.0061
0.0633
0.8137
0.0967
Model
Residual
0.387
Total
6.755
12
290
115
55.7
14.4
302
125
64.25
16.8
314
135
72.5
19.2
326
145
80.6
21.6
338
155
91.4
10
24
10
350
165
95.45
11
26.4
11
362
175
89.3
12
28.8
12
374
185
87.65
Table 15-Path of Steepest Ascent

From Figure 2, we know the fastest lane to increase yield
is to move along the direction that is perpendicular to the
contour lines. This direction is (1.1625, 0.4875), or about (2.4,
1) in terms of normalized scale. Therefore, if 1 unit of X2 is
increased, 2.39 units of X1 should be used in order to keep
moving on the steepest ascent direction. To convert the code
values to the actual values, we should use the step size of (12
degree, 10 mins) for factor A and B. The table above gives the
results for the experiments conducted along the path of
steepest ascent.
From Table 15, it can be seen that at step 10, the factor
setting is close to the optimal region. This is because the yield
decreases on either side of this step. The region around setting
of (350, 165) requires further investigation. Therefore, the
analysts will conduct a 2 2 factorial design with the center
point of (350, 165) and the range of [L=345, H=355] for factor
A (temperature) and [L=155, H=175] for factor B (reaction
time). The design matrix is given in Table 16.
Std.
Order
Point
Type
A:Temperature
(F)
B:Reaction Time
(min)
Yield
(%)
345
155
89.75
355
155
90.2
Table 14-ANOVA Table for the Initial Experiment
345
175
92
355
175
94.25
350
165
94.85
350
165
95.45
350
165
95
350
165
94.55
350
165
94.7
Table 16-Factorial Design around the Optimal Region

The ANOVA table for this data is:
Figure 2--Contour Plot of the Initial Experiment
Factor Levels
Step
Coded
Actual
Yield
(%)
Current Operation
230
65
35
2.4
242
75
36.5
Source of
Variation
DF
Sum of
Squares
Mean
Squares
F Ratio
P
Value
37.643
9.4107
78.916
5E-04
1.8225
1.8225
15.283
0.017
9.9225
9.9225
83.208
8E-04
AB
0.81
0.81
6.7925
0.06
Model
Guo & Mettas 7
25.088
25.088
Residual
0.477
0.1193
Total
38.12
Curvature
210.38
1E-04
Table 17-ANOVA for the Experiment at the Optimal Region

Table 17 shows curvature is significant at this experiment
region. Therefore, the linear model is not enough for the
relationship between factors and response. A quadratic model
should be used instead. An experiment design that is good for
the quadratic model should be used for further investigation.
Central Composite Design (CCD) is one of these designs.
4.2 Optimization Using RSM
Table 17 is the 2-level factorial design at the optimal
region. CCD is build based on this factorial design and used to
estimate the parameters for a quadratic model such as:
Y = 0 + 1 X 1 + 2 X 2 + 11 X 12 + 22 X 22 + 12 X 1 X 2 + (15)
In fact, a CCD can be directly augmented from a regular
factorial design. The augmentation process is illustrated in
Figure 3.
(0, )
(-1,1)
(1,1)
(-, 0)
(1,-1)
(-1,-1)
(0, 0)
(, 0)
(0, -)
(-, 0)
(1,1)
(0, 0)
1/ 4
2 k f (n f )
(16)
=
ns
Where: nf is the number of replicates of the runs in the

original factorial design. ns is the number of replicates of the
runs at the axial points. 2k-f represents the original factorial or
fractional factorial design.
8 Guo & Mettas
92
355
175
94.25
350
165
94.85
350
165
95.45
350
165
95
350
165
94.55
B:Reaction
Time (min)
155
Yield (%)
89.75
90.2
350
165
94.7
10
-1
342.93
165
90.5
11
-1
357.07
165
92.75
12
-1
350
150.86
88.4
13
-1
350
179.14
92.6
DF
Sum of
Squares
Mean
Squares
65.0867
4.3247
AB
AA
F Ratio
P Value
13.0173
91.3426
3.22E-06
4.3247
30.3465
0.0009
18.7263
18.7263
131.4022
8.64E-06
0.81
0.81
5.6838
0.0486
16.0856
16.0856
112.8724
1.43E-05
BB
30.1872
30.1872
211.8234
1.73E-06
Residual
1.4551
0.3527
Model
(, 0)
Points outside the rectangle in Figure 3 are called axial points

or start points. By adding several center points and axial
points, a regular factorial design is augmented to a CCD. In
Figure 3, we can see there are five different values for each
factor. So CCD can be used to estimate the quadratic model in
equation (15).
Several methods have been developed to calculate to
make the CCD have special properties such that the designed
experiment can better estimate model parameters or can better
explore the optimal region. The commonly used method to
calculate is:
A:Temperature
(F)
345
175
factorial design at the optimal region, we only need to add

start points to have a CCD. The complete design matrix for
the CCD is:shown in Table 18. The last 5 runs in the above
table are added to the previous factorial design. The fitted
quadratic model is:
Y = 94.91 + 0.74 X 1 + 1.53 X 2 1.52 X 12 2.08 X 22 + 0.45 X 1 X 2 (16)
The ANOVA table for the quadratic model is:
Figure 3-Augmnent a Factorial Design to CCD
Point
Type
1
155
345
We use equation (15) to calculate for our example. So

= 1.414. Since we already have five center points in the
(0, -)
Std. Order
355
Table 18-CCD around the Optimal Region
(1,-1)
(-1,-1)
Source of
Variation
(0, )
(-1,1)
0.9976
0.1425
Lack of Fit
0.5206
0.1735
Pure Error
4
12
0.477
66.0842
0.1193
Total
Table 19-ANOVA Table for CCD at the Optimal Region

As mentioned before, the model for CCD is used to
optimize the process. Therefore, the accuracy of the model is
very important. From Table 19, the Lack of Fit test, we see the
P value is relatively large. It means the model can fit the data
well. The Lack of Fit residual is the estimation of the
variations of the terms that are not included in the model. If its
amount is close to the pure error, which is the within-run
variation, it can be treated as part of the noise. Another way to
check the model accuracy is to check the residual plots.
Through the above diagnostic, we found the model is
adequate and we can use it to identify the optimal factor
settings. The optimal settings can be found easily by taking the
derivative of each factor from equation (16) and setting them
to 0. Many software packages can do optimization. The
following results are from DOE++ from ReliaSoft [3].
ReliaSoft DOE++- www.ReliaSoft.com
Optimal Solution 1
Central Composite Design
95.326
Factor vs Response
Continuous Function
Factor Value
Response Value
Yield
(Maximize)
Y = 95.3264
93.733
92.141
90.548
88.955
QA
Reliasoft
7/11/2009
1:16:17 PM
A:Temperature
X =351.5045
179.142
173.485
167.828
162.172
156.515
150.858
357.071
354.243
351.414
348.586
345.757
342.929
87.362
B:Reaction Time
X =168.9973
Figure 4-Optimal Solution for the Chemical Process

Figure 4 shows how the response changes with each
factor. The red dashed line points to the optimal value of each
factor. we can see that the optimal settings are 351.5 degrees
Fahrenheit for temperature and 169 minutes for the reaction
time. At this setting, the predicted yield is 95.3%, which is
much better than the yield at the current condition (about
35%).
5.
data is to use the center point of the interval data as the failure
time, and treat the suspension units as failed.
Even with the modification of the original data, another
issue may still exist. In the using of linear regression and
ANOVA, the response is assumed to be normally distributed.
The F and T tests are established based on this normal
distribution assumption. However, life time usually is not
normally distributed.
Given the above reasons, correct analysis methods for
data from life testing are needed.
5.2 Maximum Likelihood Estimation and Likelihood Ratio
Test [2]
Maximum Likelihood Estimation (MLE) can estimate
model parameters to maximize the probability of the
occurrence of an observed data set. It has been used
successfully to handle different data types, such as complete
data, suspensions and interval data. Therefore, we will use
MLE to estimate the model parameters for life data from
DOE.
Many distributions are used to describe lifetimes. The
three most commonly used are [4]:
Weibull distribution with probability density function
(pdf):
DOE FOR LIFE TESTING
When DOE is used for life testing, the response is the life
or failure time. However, because of the cost, time or other
constraints, you may not have observed values of life for some
test units. They are still functioning at the time when the test
ends. The end time of the test is called the suspension time for
the units that are not failed. Obviously, this time is not their
life. Should the suspension time be treated as the life time in
order to analyze the data? In this section, we will discuss the
correct statistical method for analyzing the data for life testing.
First, lets explain some basic concepts in life data analysis.
5.1 Data Type for Life Test
When the response is life, there are two types of data
Complete Data
Censored Data
o Right Censored (Suspended)
o Interval Censored
If a test unit is failed during the test and the exact failure
time is known, the failure time is called complete data.
If a test unit is failed and you dont know the exact failure
time -- instead, you know the failure occurs within a time
range -- this time range is called interval data.
If a unit does not fail in the test, the end time of the test of
the unit is called right censored data or suspension data.
Obviously, ignoring the censored data or treating them as
failure times will underestimate the system reliability.
However, in the use of the linear regression and ANOVA, an
exact value for each observation is required. Therefore,
engineers have to tweak the censored data in order to use
linear regression and ANOVA. A simple way to tweak the
t

Lognormal distribution with pdf:
f (t ) =
f (t ) =
t 2
1 ln(t )

e 2
(17)
(18)
Exponential distribution with pdf:

t
1
(19)
f (t ) = e m
m
Assume there is only one factor (in the language of DOE), or
stress (in the language of accelerated life testing) that affects
the lifetime of the product. The life distribution and factor
relationship can be described using the following graph.
Figure 5-pdf at Different Stress/Factor Levels

Figure 5 shows that life decreases when a factor is changed
Guo & Mettas 9
from the low level to the high level. The pdf curves have the
same shape while only the scale of the curve changes. The
scale of the pdf is compressed at the high level. It means the
failure mode remains the same, only the time of occurrence
decreases at the high level. Instead of considering the entire
scale of the pdf, a life characteristic can be chosen to represent
the curve and used to investigate the effect of potential factors
on life. The life characteristics for the three commonly used
distributions are:
Weibull distribution:
Lognormal distribution:
Exponential distribution: m
The life-factor relationship is studied to see how factors affect
life characteristic. For example, a linear model can be used as
the initial investigation for the relationship:
(20)
' = 0 + 1 X 1 + 1 X 2 + ... + 12 X 1 X 2 + ...
Where: ' = ln( ) for Weibull; ' = for lognormal and
' = ln(m) for exponential.
Please note that in equation (20) a logarithmic
transformation is applied to the life characteristics of the
Weibull and exponential distributions. One of the reasons is
because and m can take only positive values.
To test whether or not one effect in equation (20) is
significant, the likelihood ratio test is used:
L(effect k removed)
LR (effect k ) = 2 ln
(21)
L(full model)
Where L() is the likelihood value. LR follows a chi-squared
distribution if effect k is not significant.
If effect k is not significant, whether or not it is removed
from the full model of equation (20) will not affect the
likelihood value much. It means the value of LR will be close
to 0. Otherwise, if the LR value is very large, it means effect k
is significant.
5.3 Life Test Example
Consider an experiment to improve the reliability of
fluorescent lights [2]. Five factors A-E are investigated in the
experiment. A 2 52 design with factor generators D=AC and
E=BC is conducted. The objective is to identify the significant
effects that affect the reliability. Two replicates are used for
each treatment. The test ends at the 20th day. Inspections are
conducted every two days. The experiment results are given in
Table 20.
Failure Time
Table 20- Data for the Life Test Example

20+ means that the test unit was still working at the end of the
test. So this experiment has suspension data. (14, 16) means
that failure occurred at a time between the 14th and the 16th
day. So this experiment also has interval data. The Weibull
model is used as the distribution for the life of the fluorescent
lights. The likelihood ratio test table is given below.
LR
P
Value
-20.7181
3.1411
0.0763
-24.6436
10.9922
0.0009
-19.2794
0.2638
0.6076
-25.7594
13.2237
0.0003
-21.0727
3.8504
0.0497
-19.1475
Model
Effect
DF
Ln(LKV)
Reduced
Full
Table 21- LR Test Table for the Life Test Example

Table 21 has a layout that is similar to the ANOVA table. This
makes it easy to read for engineers who are familiar with
ANOVA. From the P value column, we can see factor A, B, D
and E are important to the product life. The estimated factorlife relationship is:
ln( ) = 2.9959 + 0.1052 A 0.2256 B 0.0294C
(22)
0.2477 D + 0.1166 E
The estimated shape parameter for the Weibull distribution
is 7.27.
For comparison, the data was also analyzed using
traditional linear regression and the ANOVA method. To
apply linear regression and ANOVA, the data set was
modified by using the center points as the failure time for
interval data and treating the suspensions as failures. Results
are given below.
Source of
Variation
Model
DF
Sum of
Squares
Mean
Squares
F Ratio
P Value
143.3125
28.6625
4.2384
0.025
1.5625
1.5625
0.2311
0.6411
33.0625
33.0625
4.8891
0.0515
3.0625
3.0625
0.4529
0.5162
95.0625
95.0625
14.0573
0.0038
1.5619
0.2398
10.5625
10.5625
Residual
10
67.625
6.7625
Total
15
210.9375
Run
-1
-1
-1
-1
-1
(14,16)
-1
-1
(18,20)
20+
Table 23- ANOVA Table for the Life Test Example
-1
-1
-1
(8,10)
(10, 12)
-1
-1
(18,20)
20+
In Table 23, only effects B and D are showing to be

significant. The estimated linear regression model is:
Y = 16.9375 + 0.3125 A 1.4375 B + 0.4375C
(23)
2.4375 D + 0.8127 E
Comparing the results in Tables 22 and 23, we can see
that they are quite different. The linear regression and
ANOVA method failed to identify A and E as important
20+
-1
-1
-1
20+
20+
-1
-1
(12,14)
20+
-1
(16,18)
20+
-1
-1
(12,14)
(14, 16)
10 Guo & Mettas
factors at the significance level of 0.1.

6.
CONCLUSION
In this tutorial, simple examples were used to illustrate the

basic concepts in DOE. Guidelines for conducting DOE were
given. Three major topics were discussed in detail: 2-level
factorial design, RSM and DOE for life tests.
Linear regression and ANOVA are the important tools in
DOE data analysis. So, they are emphasized. For DOE
involving censored data, the better method of MLE and
likelihood ratio test should be used.
DOE involves many different statistical methods. Many
useful techniques, such as blocking and randomization,
random and mixed effect model, model diagnostic, power and
sample size, measurement system study, RSM with multiple
responses, D-optimal designs, Taguchi orthogonal array,
Taguchi robust designs, mixture designs and so on are not
covered in this tutorial [1, 2, 5, 6]. However, with the basic
knowledge of this tutorial, readers should be able to learn most

of them easily.
REFERENCES
1.
2.
3.
4.
5.
6.
D. C. Montgomery, Design and Analysis of Experiments,

5th edition, John Wiley and Sons, Inc. New York, 2001.
C. F. Wu and M. Hamad, Experiments: Planning,
Analysis, and Parameter Design Optimization, John
Wiley and Sons, Inc. New York, 2000.
ReliaSoft, http://www.ReliaSoft.com/doe/index.htm
ReliaSoft, Life Data Analysis Reference, ReliaSoft
Corporation, Tucson, 2007.
R. H. Myers and D. C. Montgomery, Response Surface
Methodology: Process and Product Optimization Using
Designed Experiments, 2nd edition, John Wiley and Sons,
Inc. New York, 2002.
ReliaSoft, Experiment Design and Analysis Reference,
ReliaSoft Corporation, Tucson, 2008.
Guo & Mettas 11
San Jose, CA USA

January 25-28, 2010
Outline
Design of Experiments
(DOE) and Data Analysis
Huairui (Harry) Guo, Ph.D.

Adamantios Mettas
Why DOE
What DOE Can Do
Common Design Types
General Guidelines for Conducting DOE
Statistic Background
2010 RAMS Tutorial DOE Guo and Mettas
Introduction
Linear Regression and Analysis of Variance (ANOVA)
Two Level Factorial Design

Response Surface Method
Reliability DOE
1
Tutorial: Design of Experiments (DOE) and Data Analysis

Introduction
Statistical
Background
Why DOE
Two Level
Factorial Design
DOE Makes Your Life Easier

Example 1
Introduction
Response
Surface Method
Usage voltage: V=10v

Test all the 50 units at V=25v to predict the reliability at V=10v
Example 2
Reliability DOE
Usage temperature T=40 (Celsius) and humidity H=50%

Accelerated life testing to predict reliability at usage level
Summary
Num of Units
25
25
Temperature
(Celsius)
120
85
What DOE Can Do
For comparison
2 level factorial design

Plackett-Burman design
Taguchi orthogonal array design
For transfer function identification and optimization
One factor design
For variable screening
Central composite design

Box-Behnken design
For system robustness
12 Guo & Mettas
Humidity (%)
95
85
Common Design Types
Comparisons
Variable screening
Transfer function exploration
System optimization
System robustness
Linear Regression (contd)
Y = 0 + 1 X 1 + ... + p X p +
General Guidelines for Conducting DOE
Assumptions
The random error or noise. ~ N (0, )

It is assumed to be normally distributed with mean of 0
and variance of 2
Clarify and state objective

Choose responses
Choose factors and levels
Choose experimental design
Perform the experiment
Analyze the data
Draw conclusions and make recommendations
Model
Y : For a given model, the response is also normally

distributed and
Var (Y ) = 2
Parameter Estimation: Least Squares Estimation

11

Introduction
Statistical Background
Linear regression
Statistical
Background
Two Level
Factorial Design
It is the foundation for DOE data analysis.
ANOVA is a way to present the findings from the

linear regression model.
ANOVA can tell you if there is a strong relationship
between the independent variables, Xs, and the
response variable, Y.
ANOVA can test whether or not an individual X can
affect Y significantly.
Statistical Background
Response
Surface Method
ANOVA (analysis of variance)
Reliability DOE
Summary
Linear Regression
Regression analysis is a
statistical technique that
attempts to explore and
model the relationship
between two or more
variables.
A linear regression
model attempts to explain
the relationship between
two or more variables
using a straight line.
Temperature Anomoly ( F)
1.5
1.0
0.5
0.0
-0.5
-1.0
1880
1900
1920
1940
Year
1960
1980
2000
10
Guo & Mettas 13
Example: Linear Regression
Within-run Variations
ANOVA Table for the Example
Y = 0 + 1 X 1 + 2 X 2 +
For the same values of X1

and X2, the observed Y
values are different.
This difference of Y is
caused by noise.
Source of Degrees of
Variation Freedom
Model
2
X1
1
X2
1
Residual
3
Total
5
Between-run Variations
For different values of X1

and X2, the observed mean
values of Y are different.
This difference of Y is
caused by the different
values of the Xs.
Sum of
Squares
6.14E+04
6.00E+04
8100
2500
6.39E+04
Mean
Squares F Ratio
3.07E+04 36.86
6.00E+04 72.03
8100
9.72
833.3333
P
Value
0.0077
0.0034
0.0526
* Residual is the estimated value for error or noise.
P Value: The smaller the P value is, the more significant the corresponding source.
The P value is compared with a given significance level, alpha. If a P value is less than
the alpha, the corresponding source is significant at the significance level of alpha. The
commonly used alpha values are 0.01, 0.05, 0.1.
12
The Model and X1 are significant at Level of 0.05.

The Model, X1 and X2 are significant at Level of 0.1.
Measure the Variation
Partition the Total Variation
Sum of Squares:
SST = Yi Y
i =1
)2
Mean Squares:
Between-run Variation
Within-run Variation
SST = SS R + SS E
n
i =1
i =1
= (Yi Y ) 2 + (Yi Yi ) 2
SSR: Sum of Squares of Regression. The amount of variation explained by Xs.
MST =
The total variation of Y includes two parts
n: Number of observations. (It is 6 for this example.)
17
SS T
1 n
=
Yi Y
n 1 n 1 i =1
SSE: Sum of Squares of Error. The amount of variation caused by noise.
)2
Mean Squares is the normalized variation of a data set.

It is also the unbiased estimation of the variance of Y.
Mean squares of regression and error
MS R =
SS R
p
MS E =
n: Number of observations
13
SS E
n 1 p
p: Number of Xs. It is 2 for the previous example.
Partition the Sum of Squares of Regression
Analysis of Variance (ANOVA)
The SSR can be divided into the variation caused by each

variable:
SS R = SS X1 + SS X 2 + ... + SS X p
The mean squares of each source is compared with MSE
If MSXi is significantly larger than MSE, Xi significantly affects Y.

If it is not significantly larger than MSE, Xis effect is close to the
effect of noise. Xi doesnt significantly affect Y.
F test for the regression model: F0 =
15
F test for each variable xi:
F0 =
MS X i
MS E
H0: There is no difference between the variation caused by Xi and

the variation caused by noise.
H1: The variation caused by Xi is larger than the variation
caused by noise.
14 Guo & Mettas
MS R
MS E
H0: There is no difference between the variation caused by Xs

and the variation caused by noise.
H1: The variation caused by Xs is larger than the variation
caused by noise.
Each X has its own mean squares

MS X 1 , MS X 2 ,..., MS X P
14
16

Introduction
Statistical
Background
T-test for the Coefficients
Two Level
Factorial Design
T-test is an alternative to the F-test.

T-test is used to test whether or not a coefficient is 0.
For example,
H0: 1 = 0;
The test statistic is:
T0 =
Response
Surface Method
H1: 1 0
Reliability DOE
Summary
se( 1 )
1 is the estimated value for 1.
se( 1 ) is the standard error.

21
Under the Null-hypothesis, T0 follows the t distribution with the degree of

freedom (df) that is the same as the df of the error term.
18
Some Basic Concepts in DOE
T-test Results
Term
Intercept
X1
X2
Coefficient
-2135
7
18
Standard
Error
588.7624
0.8248
5.7735
Source of Degrees of Sum of

Variation Freedom
Squares
Model
2
6.14E+04
X1
1
6.00E+04
X2
1
8100
Residual
3
2500
Total
5
6.39E+04
T Value
-3.6263
8.487
3.1177
Mean
Squares
3.07E+04
6.00E+04
8100
833.3333
Treatment: A unique combination of all the factors.

Replicate: The number of test units under each
treatment.
P Value
0.0361
0.0034
0.0526
Treatment = 4
P
F Ratio Value
36.86 0.0077
72.03 0.0034
9.72
0.0526
Replicate = 1
Design notation:
2k
k is the number of factors.
23
Note: The P values from the T-test are the same as in the F-test.
19
Pareto Chart for the T-Values

Pareto Chart
Pareto Chart
Critical Valu
Significant
Term
X1
X2
0.000
2.353
3.600
5.400
7.200
QA
Reliasoft
7/28/2009
4:57:54 PM
9.000
Effect
Alpha = 0.1; Threshold = 2.3534
20
Guo & Mettas 15
Two Level Fractional Factorial Design
Why two levels?
To study the effect of a factor or a variable, at least two

different settings are needed. High = 1, Low = -1.
From a linear regression viewpoint, you need at least
two points to fit a straight line.
Answer
to the first example
Example 1
Usage voltage: V=10v
Test all the 50 units at V=25v to predict the reliability at V=10v.
Why fractional factorial design
Why fractional factorial design works
22
Reduce the sample size. For a 7 factor design, you need

27 = 128 runs.
Reduce cost and time by only using part of the full
design.
Sparsity of effects principle: Most of the time,
responses are affected by only a small number of main
effects and lower order interactions.
Fractional factorial design: Focus on main effects and
lower order interactions.
Balanced and Orthogonal
Half Fractional Factorial Design 23-1
Balanced: The number
Order
2
3
5
8
of test units is the same

for all the treatments.
Orthogonal: The sum
of the product of each

factor column is 0.
(1 1) + (1 1) + (1 1) + (1 1) = 0
A balanced and orthogonal design can evaluate the effect of each factor more
accurately.
If you add one more sample for any one of the treatments, the design will be
unbalanced and non-orthogonal.
27
24
A
1
-1
-1
1
B
-1
1
-1
1
AB
-1
-1
1
1
C
-1
-1
1
1
AC
-1
1
-1
1
BC ABC
1
1
-1
1
-1
1
1
1
From this experiment, the effect of AB and C, AC and B, and BC

and A cannot be distinguished. Each pair of columns (AB and C,
A and BC, B and AC) have the same pattern.
Effects that cannot be distinguished within a design are called
Confounded Effects or Aliased Effects.
The summary of the aliased effects for the design is called Alias
Structure.
[I]=I+ABC; [A]=A+AC; [B]=B+BC; [C]=C+AB
I=ABC is called the Defining Relation.
Linear Model for 2 Factorial Design
A Simple Way to Estimate the Coefficients
Y = 0 + 1 X 1 + 2 X 2 + 12 X 1 X 2 +
X1: Factor A, X2: Factor B
Least Square Estimation is used for any design

Special case: only valid for a balanced design
Main effect: effect A(X1) and B(X2)

Two-way interaction effect: AB(X1X2)
29
The number of factors in a interaction effect is the

order of the interaction.
ANOVA is used to test whether or not an effect is

significant.
25
Effet of A = Avg. at A high - Avg. at A low

=
30 + 35 20 + 25
= 10
2
2
Effect of A = 2 1
16 Guo & Mettas
26
Review the Second Example

Example 2
Design Matrix for a 23 Design
Usage temperature T=40 (C) and humidity H=50%.

Accelerated life test to predict reliability at usage level.
Num of Units
25
25
Temperature
(Celsius)
120 (+1)
85 (-1)
Humidity (%)
95 (+1)
85 (-1)
It is a bad design: Main effects Temperature and Humidity are

Confounded.
From the viewpoint of Linear Regression, at least 3 unique
combinations are needed why?
Y = 0 + 1 X 1 + 2 X 2 +
31
If effect ABC can be ignored in the study, then we can select the
treatments with ABC=1 for a Fractional Factorial Design.
28
Fractional Factorial Design Example (contd)
Generate a Fractional Factorial Design

Order
2
3
5
8
A
1
-1
-1
1
B
-1
1
-1
1
C
-1
-1
1
1
AB
-1
-1
1
1
AC
-1
1
-1
1
BC ABC
1
1
-1
1
-1
1
1
1
With five factors the total number of runs required for a

full factorial is 25 = 32. These runs are too expensive for
the manufacturer.
Only the main effects and the two factor interactions are
assumed to be important.
It is decided to carry out the investigation using the 25-1
design (requiring 16 runs) with the factor generator
E= ABCD.
The design is a Full Factorial Design for A and B.

A and B are called Basic Factors.
The Full Factorial Design for A and B is called the Basic
Design.
Factor C can be generated from C = AB.
C = AB is called the Factor Generator for C.
33
30
ANOVA Table with All Effects

Source of
Variation
DF
Model
A
B
C
D
E
AB
AC
AD
AE
BC
BD
BE
CD
CE
DE
Residual
Total
15
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
15
Sum of
Squares
Mean
Squares
P Value
5775.4375
495.0625
4590.0625
473.0625
3.0625
1.5625
189.0625
0.5625
5.0625
5.0625
1.5625
0.0625
0.0625
3.0625
0.5625
7.5625
385.0292
495.0625
4590.0625
473.0625
3.0625
1.5625
189.0625
0.5625
5.0625
5.0625
1.5625
0.0625
0.0625
3.0625
0.5625
7.5625
5775.4375
35
Guo & Mettas 17
Fractional Factorial Design Example
Pool Non-Significant Effects to Error
The yield in the manufacture of

integrated circuits is thought to be
affected by the following five
factors*:
Factor
Name
A
B
C
D
E
Aperture Setting
Exposure Time
Develop Time
Mask Dimension
Etch Time
Unit
minutes
seconds
minutes
Low
High
small
20
30
small
14.5
large
40
45
large
15.5
Non-significant effects are treated as noise and are

used to estimate the error.
The objective of the

experimentation is to increase the
yield by finding the significant
effects.
Montgomery, D. C, 2001
32
Design Matrix and Observations

Run Order
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
A
1
1
-1
-1
1
1
-1
-1
1
-1
-1
1
-1
-1
1
1
B
-1
-1
1
-1
1
1
1
-1
-1
1
-1
1
1
-1
-1
1
C
-1
1
1
1
-1
1
-1
1
-1
-1
-1
-1
1
-1
1
1
D
1
1
-1
-1
-1
-1
1
1
-1
-1
-1
1
1
1
-1
1
E
1
-1
1
-1
1
-1
1
1
-1
-1
1
-1
-1
-1
1
1
37
Response Surface Methods (RSM)
Yield
10
21
45
16
52
60
30
15
9
34
8
50
44
6
22
63
RSM is used to estimate the transfer function at

the optimal region.
The Quadratic model is used in RSM.
The estimated function is used for optimization.
Linear regression and ANOVA are the tools for
data analysis.
Run Order: The test sequence of the treatments. The experiment is

conducted in a Random Sequence of the factor combinations.
34
Identify Significant Effects
Design Matrix for the Initial Investigation

Experiment
Normal ProbabilityPlot of Effect
99.000
Effect Probability
B:Exposure Time
A:Aperture Setting
Yield
Y' = Y
Non-Significant Effects
Significant Effects
Distribution Line
The initial design is a modified factorial design. The Design Matrix is:
C:Develop Time
AB
Probability
Std. Order
1
2
3
4
5
6
7
8
9
50.000
10.000
5.000
1.000
-40.000
-24.000
8.000
-8.000
24.000
QA
Reliasoft
7/10/2009
3:03:18 PM
40.000
Effect
Alpha = 0.1; Lenth's PSE = 0.9375
Lenths method assumes all the effects should be normally distributed with a mean of
0, given the hypothesis that they are not significant.
Point Type
1
1
1
1
0
0
0
0
0
A:
Temperature
-1
1
-1
1
0
0
0
0
0
B:
Reaction Time
-1
-1
1
1
0
0
0
0
0
Yield (%)
33.95
36.35
35
37.25
35.45
35.75
36.05
35.3
35.9
The initial design at the current operation condition is used to search

36
the direction for locating the optimal region for optimization.

18 Guo & Mettas
39
41

Introduction
Results for the Initial Experiment

ANOVA Table
Statistical
Background
Two Level
Factorial Design
Response Surface Method
Response
Surface Method
Reliability DOE
* Curvature and AB are not significant.
Summary
The final linear regression model (in coded values) with only
significant terms:
Y = 35.6375 + 1.1625 X 1 + 0.4875 X 2
38
RSM: Example
Path of Steepest Ascent
The yield from a chemical process has been

found to be affected by two factors:
Reaction Temperature
Reaction Time
Step
The current operating conditions of

230 Fahrenheit and 65 minutes give a yield
of about 35%.
The engineer decides to explore the current
conditions in the range (225, 235)
Fahrenheit and (55, 75) minutes so that
steps can be taken to achieve the maximum
yield.
43
Current
Operation
1
2
3
4
5
6
7
8
9
10
11
12
Coded
Factor Levels
Actual
Yield (%)
230
65
35
2.4
4.8
7.2
9.6
12
14.4
16.8
19.2
21.6
24
26.4
28.8
1
2
3
4
5
6
7
8
9
10
11
12
242
254
266
278
290
302
314
326
338
350
362
374
75
85
95
105
115
125
135
145
155
165
175
185
36.5
39.35
45.65
49.55
55.7
64.25
72.5
80.6
91.4
95.45
89.3
87.65
40
45
Coded Values and Center Points
Coded value (-1, 0, 1)

Coded value =
Actual Value - Middle Value of High and Low

Half of the Range Between High and Low
For example, the temperature is (Low = 225, High = 235). So when

Temperature = 230, the corresponding coded value is 0.
Reasons for using Center Points
Replicates at Center Points can be used to estimate the random error.

Replicates at Center Points will keep the Orthogonality of the design.
Can be used to check whether or not the Linear Model is sufficient or
whether the curvature is significant.
42
Guo & Mettas 19
Factorial Design Around the Optimal Region

Std. Order
1
2
3
4
5
6
7
8
9
Point Type
1
1
1
1
0
0
0
0
0
A:Temperature (F)
345
355
345
355
350
350
350
350
350
B:Reaction Time (min) Yield (%)

155
89.75
155
90.2
175
92
175
94.25
165
94.85
165
95.45
165
95
165
94.55
165
94.7
Designs for Quadratic Model
Quadratic model
Y = 0 + 1 X 1 + 2 X 2 + 11 X 12 + 22 X 22 + 12 X 1 X 2 +
Special designs are used for the Quadratic model.

Central Composite Design (CCD) is one of them.
CCD can be directly augmented from a regular

factorial design.
At the optimal region, the curvature is significant. The simple linear model is not enough any more.
47
Direction of Steepest Ascent
48
Augment a Factorial Design to a Central Composite Design
Y = 35.6375 + 1.1625 X 1 + 0.4875 X 2
The coefficients show that the direction is (1.1626, 0.4875) or
can be normalized to (2.4, 1).

The direction of steepest ascent is perpendicular to the contour lines.
Using a 2 factor factorial design as example:

(0, )
(-1,1)
(1,1)
(-, 0)
(1,-1)
(-1,-1)
(0, 0)
(, 0)
(0, -)
(0, )
(1,1)
(-1,1)
(-, 0)
(, 0)
(0, 0)
(1,-1)
(-1,-1)
(0, -)
44
The points outside the rectangle are called axial points or

star points.
Further Investigation at the Optimal Region
From the sequential tests, we found the setting at

step 10 (Temp = 350, Time = 165) is close to the
optimal region.
The region around the setting at step 10
(Temp = 350, Time = 165) should be investigated.
A Factorial Design with a center point of (350, 165)
was conducted.
49
The Complete CCD Matrix for the Example

Std. Order
1
2
3
4
5
6
7
8
9
10
11
12
13
Point Type
1
1
1
1
0
0
0
0
0
-1
-1
-1
-1
A:Temperature (F)
345
355
345
355
350
350
350
350
350
342.93
357.07
350
350
B:Reaction Time (min)

155
155
175
175
165
165
165
165
165
165
165
150.86
179.14
Yield (%)
89.75
90.2
92
94.25
94.85
95.45
95
94.55
94.7
90.5
92.75
88.4
92.6
Four Star Points are added to the previous factorial design.

46
20 Guo & Mettas
51
Model Diagnostic
Model Diagnostic (contd)
Check the residual plots
Normal Probability Plot of Residual
99.000
Residual vs Run Order
0.700
Residual Probability
Residual vs Order
0.621
Yield
Y' = Y
Data Points
Residual Line
Yield
Y' = Y
Data Points
Upper Critical Value
Lower Critical Value
0.420
Residual
Probability
Zero Value Line

0.140
50.000
0.000
-0.140
10.000
5.000
1.000
-2.000
-1.200
-0.400
0.400
1.200
QA
Reliasoft
7/11/2009
1:07:07 PM
2.000
-0.420
-0.700
0.000
QA
Reliasoft
7/11/2009
1:08:33 PM
15.000
-0.621
Residual
Anderson-Darling =0.3202; p-value =0.4906
53
3.000
6.000
9.000
Run Order
12.000
Alpha =0.1; Upper Critical Value =0.6209; Lower Critical Value =-0.6209
How to Decide
Several methods have been developed for deciding
.
Special values of can make the design have special

properties, such as a design that can better estimate model
parameters, or that can better explore the optimal region.
Central Composite Design
1/ 4
Factor Value
Response Value
92.141
90.548
88.955
QA
Reliasoft
7/11/2009
1:16:17 PM
A:Temperature
X =351.5045
50
179.142
173.485
167.828
162.172
156.515
357.071
150.858
354.243
342.929
351.414
87.362
nf is the number of replicates of the runs in the original factorial design.

ns is the number of replicates of the runs at the axial points.
2k-f represents the original factorial or fractional factorial design.
Factor vs Response
Continuous Function
93.733
A commonly used method is:

2 k f (n f )
ns
Optimal Solution 1
95.326
348.586
Y = 94.91 + 0.74 X 1 + 1.53 X 2 1.52 X 12 2.08 X 22 + 0.45 X 1 X 2
345.757
Optimization
Yield
(Maximize)
Y = 95.3264
54
B:Reaction Time
X =168.9973
Optimal Setting (Temperature = 351.5, Time = 169.0)

Predicted Yield at the Optimal Setting is 95.3%
55
Results from the Experiment
ANOVA Table
Model in terms of coded value:
Y = 94.91 + 0.74 X 1 + 1.53 X 2 1.52 X 12 2.08 X 22 + 0.45 X 1 X 2

52
Guo & Mettas 21
DOE For Life Testing
Data Types
The response is the failure time.

The failure time usually is not normally distributed, so linear
regression and ANOVA are not good for life data.
If a unit is still running at the time when the test ends, you only
have suspension time, not failure time for this unit.
The correct analysis method is needed for reliability

DOE.
Complete data
Censored data
A method that can handle lifetime distributions such as Weibull,

lognormal and exponential.
A method that can handle suspension times correctly.
A method that can evaluate the significance of each effect,
similar to ANOVA.
Right Censored
Interval Censored
57
Right Censored (Suspended) Data: Example
Interval Censored Data: Example
Imagine that we tested five units and three failed. In this

scenario, our data set is composed of the times-to-failure
of the three units that failed and the running time of the
other two units that did not fail.
58
Imagine that we are running a test on five units and inspecting

them every 100 hours. If a unit failed between inspections, we
do not know exactly when it failed, but rather that it failed
between inspections. This is also called inspection data.
59

Introduction
Distributions Commonly Used in Reliability
Statistical
Background
Weibull distribution pdf:

1
t
f (t ) = e

Two Level
Factorial Design
Response
Surface Method
Reliability DOE (R-DOE)
Lognormal distribution pdf:

f (t ) =
Reliability DOE
Summary
60
1 ln( t )

1
e 2
t 2
Exponential distribution pdf:

t
f (t ) =
56
1 m
e
m
22 Guo & Mettas
61
Life-Factor Relationship Simplify: Infer a Characteristic
Life-Factor Relationship
Using the life characteristic, the model to investigate the effect of

factors on life can be expressed as:
' = 0 + 1 x1 + 2 x2 + ... + 12 x1 x2 + ...
where:
Instead of considering the entire scale of the pdf, the life

characteristic can be chosen to investigate the effect of potential
factors on life.
The life characteristics for the three commonly used distributions
are:
Weibull:
Lognormal:
xj :
Exponential: m
' = ln( )
or
'=
or
' = ln(m)
jth factor value
Note that a logarithmic transformation is applied to the life

characteristics of the Weibull and exponential distributions.
This is because and m can take only positive values.
63
The Likelihood Function Based on LifeFactor Relationship
Life-Factor Relationship
= 0 + 1 xi1 + 2 xi 2 + ... + 12 xi1 xi 2 + ...
Failure Time Data
L f = f (Ti ; i, )
Suspension Data
Testing Effect Significance: Likelihood

Ratio Test
'
i
Life-factor relationship:
i =1
M
i' = 0 + 1 xi1 + 2 xi 2 + ... + 12 xi1 xi 2 + ...
LS = R ( S j ; i, )
j =1
=
LI
Interval Data
[ F (I
l =1
Ul
0 , 1 , 2 ,...
; i, ) F ( I Ll ; i, ) ]
Likelihood ratio test:
MLE
L = L f Ls LI
and
64
LR (effect k ) = 2 ln
L(effect k removed )
L( full Model )
for lognormal
If
LR (effect k ) > 1,2
then effect k is significant.
65
66
Combining Reliability and DOE: Life-Factor

Relationship
The graphic shows an example where life decreases when a factor is

changed from the low level to the high level.
It is seen that the pdf changes in scale only. The scale of the pdf is
compressed at the high level.
The failure mode remains the same. Only the time of occurrence
decreases at the high level.
62
Guo & Mettas 23
R-DOE: Example
R-DOE: Example (contd)
Consider an experiment to
improve the reliability of
fluorescent lights. Five
factors A-E are investigated
in the experiment. A 25-2
design with factor generators
D=AC and E=BC was
conducted.*
Objective: To identify the
significant factors and adjust
them to improve life.
A
-1
-1
-1
-1
1
1
1
1
*Taguchi, 1987, p. 930.
B
-1
-1
1
1
-1
-1
1
1
C
-1
1
-1
1
-1
1
-1
1
D
1
-1
1
-1
-1
1
-1
1
E
1
-1
-1
1
1
-1
-1
1
Failure Time
14~16
20+
18~20
20+
8~10
10~12
18~20
20+
20+
20+
12~14
20+
16~18
20+
12~14
14~16
Two replicates at each treatment.

Inspections were conducted every two hours.
Results have interval data and suspensions.
67
Traditional DOE Approach
Assumes that the response (life) is normally

distributed.
Treats suspensions as failures.
Uses the middle point of the interval data as the
failure time.
Problem: The above assumptions and adjustments
are incorrect.
68
Fluorescent Lights R-DOE: Life-Factor

Relationship
For the ith observation
i ' = 0 + 1 Ai + 2 Bi + 3Ci + 4 Di + 5 Ei + ...

A
-1
-1
-1
-1
1
1
1
1
B
-1
-1
1
1
-1
-1
1
1
C
-1
1
-1
1
-1
1
-1
1
D
1
-1
1
-1
-1
1
-1
1
E
1
-1
-1
1
1
-1
-1
1
Failure Time
14~16
20+
18~20
20+
8~10
10~12
18~20
20+
20+
20+
12~14
20+
16~18
20+
12~14
14~16
For example, for the first run, the equation is:

1' = 0 + 1 (1) + 2 (1) + 3 (1) + 4 (+1) + 5 (+1)
assuming that the interactions are absent.
69
MLE and Likelihood Ratio Test
Z-test for the Coefficients
Life-factor relation (Weibull distribution)

ln( ) = 2.9959 + 0.1052 A 0.2256 B 0.0294C
0.2477 D + 0.1166 E
70
Z-test is used to test whether or not a coefficient is 0.
For example,
H0: 1 = 0; H1: 1 <> 0
Likelihood Ratio (LR) Test Table
The test statistic is:

Z0 =
1 is the estimated value for 1.
se( 1 )
se( 1 ) is the standard error.
Under the Null-hypothesis, Z0 is assumed to be standard normally distributed.

71
24 Guo & Mettas
72
Z-test Results
Pareto Chart for the Z-values

Z-values are the normalized coefficients.
Pareto Chart
Pareto Chart
Critical Value
Non-Significant
Significant
D:D
Term
B:B
Note: The P values from the Z-test are slightly different from the Likelihood Ratio test.
A:A
When the sample size is large, they are very close.

When the sample size is small, the likelihood ratio test is more
accurate than the Z-test.
73
E:E
C:C
0.000
1.000
1.645 2.000
3.000
Standardized Effect (Z Value)
4.000
QA
Reliasoft
7/11/2009
9:06:30 PM
5.000
From both LR test and the Z-test, A, B, D and E are significant.

Model Diagnostic
When using the Weibull distribution for life, the residuals from the lifefactor relationship should follow the extreme value distribution with a
mean of zero.
74
Model Diagnostic (contd)

Residuals against run order plot
75
76
Fluorescent Lights R-DOE: Interpreting the Results
From the results, factors A,B, D and E are significant at the risk level of
0.10. Therefore, attention should be paid to these factors.
Fluorescent Lights Example: Traditional DOE

Approach
Suspensions are treated as failures.

Mid-points are used as failure times for interval data.
Life is assumed to follow the normal distribution.
In order to improve the life, factors A and E should be set to the high
level; while factors B and D should be set to the low level.
MLE Information
Term
Coefficient
A:A
0.1052
B:B
-0.2256
C:C
-0.0294
D:D
-0.2477
E:E
0.1166
77
78
Guo & Mettas 25
More Information
On-line textbook and examples
Traditional DOE Approach: Fluorescent Lights Example - Results
http://www.weibull.com/doewebcontents.htm
http://www.itl.nist.gov/div898/handbook/
83
B and D come out to be significant using traditional DOE approach.

A, B, D and E were found to be significant using R-DOE.
Traditional DOE fails to identify A and E as important factors at a
significance level of 0.1.
79

Introduction
Statistical
Background
Summary: Topics Covered
Why DOE, what DOE can do and common design

types
General guidelines for conducting DOE
Linear regression and ANOVA
2-level factorial and fractional factorial design
Response surface method
Reliability DOE
Two Level
Factorial Design
Summary
Response
Surface Method
Reliability DOE
Summary
80
81
Topics Not Covered
Blocking
Power and sample size
RSM with multiple responses
RSM: Box-Behnken design
D-optimal design
Taguchi orthogonal array (OA)
Mixture design
Random and mixed effect model
more
26 Guo & Mettas
82
The End
84
Guo & Mettas 27

2010 RAMS Doe and Data Analysis

Transféré par

Informations du document

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

2010 RAMS Doe and Data Analysis

Transféré par

Droits d'auteur :

Formats disponibles

Copyright 2010 IEEE.

Reprinted from 2010 Reliability and

2010 Annual RELIABILITY and MAINTAINABILITY Symposium

Design of Experiments and Data Analysis

Huairui Guo, Ph. D. & Adamantios Mettas

Huairui Guo, Ph.D., CPR.

Adamantios Mettas, CPR

Tutorial Notes 2010 AR&MS

SUMMARY & PURPOSE

Huairui Guo, Ph.D., CRP

Adamantios Mettas, CRP

ii Guo & Mettas

2010 AR&MS Tutorial Notes

The most effective way to improve product quality and

Table 1 Two Stress Accelerated Life Test

Variables that may affect the life are temperature, voltage,

2010 Annual RELIABILITY and MAINTAINABILITY Symposium

Guo & Mettas 1

list of specific problems that are to be addressed by the

As a general rule, no more than 25 percent of the available

Linear regression and ANOVA are the statistical methods

Table 2 Example Data for Linear Regression

Where Yi is the ith observed value and Y is the mean of all

2010 AR&MS Tutorial Notes

the observations. However, since SST is affected by the

Equation (3) is also the unbiased estimator of Var (Y ) .

Where: p is the number of Xs.

can easily verify that:

Table 3 ANOVA Table for the Linear Regression Example

The following ratio

Table 4 Coefficients for the Linear Regression Example

TWO LEVEL FACTORIAL DESIGNS

Two level factorial designs are used for factor screening.

2010 Annual RELIABILITY and MAINTAINABILITY Symposium

Guo & Mettas 3

who tested all the units at a voltage of 25 will not be able to

Table 5 Treatments for 2 Level Factorial Design

4 Guo & Mettas

Effet of A = Avg. at A high - Avg. at A low

3.2 Two Level Fractional Factorial Design

Table 6 Design Matrix for a 2 3 Design

Table 7 Fraction of the Design Matrix for a 2 3 Design

design in Table 6, it is represented by 2 31 and read as 2 to

2010 AR&MS Tutorial Notes

distinguished because they change with the same pattern. In

Table 9 Factor Settings for a Five Factor Experiment

Table 8 Two Stress Accelerated Life Test (Coded Value)

Table 10 Design Matrix and Results

3.3 An Example of a Fractional Factorial Design

2010 Annual RELIABILITY and MAINTAINABILITY Symposium

Guo & Mettas 5

Table 11 ANOVA Table with All Effects

Table 12 ANOVA Table with Significant Effects

There are no F ratio and P values in the above table. This is

From Table 12, we can see that effects A, B, C and AB are

ReliaSoft DOE++ - www.ReliaSoft.com

Normal ProbabilityPlot of Effect

Figure 1-Effect Probability Plot Using Lenths Method

6 Guo & Mettas

RESPONSE SURFACE METHODS (RSM)

Response surface methods (RSM) are used to estimate the

Table 13-Design Matrix for the Initial Experiment

2010 AR&MS Tutorial Notes

Actual Value - Middle Value of High and Low