Beruflich Dokumente
Kultur Dokumente
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
1.
INTRODUCTION
Temperature
(Celsius)
Humidity
(%)
25
120
95
25
85
85
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
STATISTICAL BACKGROUND
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
SST = Yi Y
i =1
)2
(2)
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)
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
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
F0 =
(9)
Response
Treatment
-1
-1
-1
-1
30
-1
25
ab
35
20
(13)
30 + 35 20 + 25
= 10
2
2
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
AB
AC
BC
-1
-1
-1
-1
ABC
1
-1
-1
-1
-1
-1
-1
-1
-1
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
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
(%)
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
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
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
DF
Sum of
Squares
Mean
Squares
F
Ratio
P Value
5747.25
1436.8125
560.7073
1.25E-12
4.
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
(13)
4.8
254
85
39.35
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
Point
Type
A:Temperature
(F)
B:Reaction Time
(min)
Yield
(%)
345
155
89.75
355
155
90.2
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
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
25.088
25.088
Residual
0.477
0.1193
Total
38.12
Curvature
210.38
1E-04
(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
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)
A:Temperature
(F)
345
175
Point
Type
1
155
345
(0, -)
Std. Order
355
(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
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
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):
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)
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.
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
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
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+
-1
-1
-1
(8,10)
(10, 12)
-1
-1
(18,20)
20+
20+
-1
-1
-1
20+
20+
-1
-1
(12,14)
20+
-1
(16,18)
20+
-1
-1
(12,14)
(14, 16)
CONCLUSION
6.
Outline
Design of Experiments
(DOE) and Data Analysis
Why DOE
What DOE Can Do
Common Design Types
General Guidelines for Conducting DOE
Statistic Background
Introduction
1
2010 RAMS Tutorial DOE Guo and Mettas
Why DOE
Two Level
Factorial Design
Introduction
Response
Surface Method
Example 2
Reliability DOE
Summary
Num of Units
25
25
Temperature
(Celsius)
120
85
For comparison
Humidity (%)
95
85
Comparisons
Variable screening
Transfer function exploration
System optimization
System robustness
Y = 0 + 1 X 1 + ... + p X p +
Assumptions
Model
Var (Y ) = 2
11
Statistical Background
Linear regression
Statistical
Background
Two Level
Factorial Design
Statistical Background
Response
Surface Method
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
Within-run Variations
Y = 0 + 1 X 1 + 2 X 2 +
Source of Degrees of
Variation Freedom
Model
2
X1
1
X2
1
Residual
3
Total
5
Between-run Variations
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
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
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
MST =
17
SS T
1 n
=
Yi Y
n 1 n 1 i =1
)2
MS R =
SS R
p
MS E =
n: Number of observations
13
SS E
n 1 p
SS R = SS X1 + SS X 2 + ... + SS X p
15
F0 =
MS X i
MS E
MS R
MS E
14
16
Two Level
Factorial Design
For example,
H0: 1 = 0;
The test statistic is:
T0 =
Response
Surface Method
H1: 1 0
Reliability DOE
Summary
se( 1 )
21
18
T-test Results
Term
Intercept
X1
X2
Coefficient
-2135
7
18
Standard
Error
588.7624
0.8248
5.7735
T Value
-3.6263
8.487
3.1177
Mean
Squares
3.07E+04
6.00E+04
8100
833.3333
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
23
Note: The P values from the T-test are the same as in the F-test.
2010 RAMS Tutorial DOE Guo and Mettas
19
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
Answer
Example 1
Usage voltage: V=10v
Test all the 50 units at V=25v to predict the reliability at V=10v.
22
Order
2
3
5
8
(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.
2010 RAMS Tutorial DOE Guo and Mettas
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
Y = 0 + 1 X 1 + 2 X 2 + 12 X 1 X 2 +
29
25
30 + 35 20 + 25
= 10
2
2
Effect of A = 2 1
26
Temperature
(Celsius)
120 (+1)
85 (-1)
Humidity (%)
95 (+1)
85 (-1)
Y = 0 + 1 X 1 + 2 X 2 +
2010 RAMS Tutorial DOE Guo and Mettas
31
If effect ABC can be ignored in the study, then we can select the
treatments with ABC=1 for a Fractional Factorial Design.
2010 RAMS Tutorial DOE Guo and Mettas
28
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
33
30
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
2010 RAMS Tutorial DOE Guo and Mettas
35
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
Montgomery, D. C, 2001
2010 RAMS Tutorial DOE Guo and Mettas
32
2010 RAMS Tutorial DOE Guo and Mettas
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
Yield
10
21
45
16
52
60
30
15
9
34
8
50
44
6
22
63
34
2010 RAMS Tutorial DOE Guo and Mettas
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.
2010 RAMS Tutorial DOE Guo and Mettas
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
39
41
Statistical
Background
Two Level
Factorial Design
Response
Surface Method
Reliability DOE
Summary
The final linear regression model (in coded values) with only
significant terms:
38
RSM: Example
Reaction Temperature
Reaction Time
Step
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
2010 RAMS Tutorial DOE Guo and Mettas
45
42
Point Type
1
1
1
1
0
0
0
0
0
A:Temperature (F)
345
355
345
355
350
350
350
350
350
Quadratic model
Y = 0 + 1 X 1 + 2 X 2 + 11 X 12 + 22 X 22 + 12 X 1 X 2 +
At the optimal region, the curvature is significant. The simple linear model is not enough any more.
2010 RAMS Tutorial DOE Guo and Mettas
47
2010 RAMS Tutorial DOE Guo and Mettas
48
(-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
49
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
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
46
2010 RAMS Tutorial DOE Guo and Mettas
51
Model Diagnostic
99.000
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
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
.
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
Factor vs Response
Continuous Function
93.733
ns
Optimal Solution 1
95.326
348.586
345.757
Optimization
Yield
(Maximize)
Y = 95.3264
54
B:Reaction Time
X =168.9973
55
ANOVA Table
52
Data Types
Complete data
Censored data
Right Censored
Interval Censored
57
2010 RAMS Tutorial DOE Guo and Mettas
58
59
2010 RAMS Tutorial DOE Guo and Mettas
Statistical
Background
t
f (t ) = e
Two Level
Factorial Design
Response
Surface Method
Reliability DOE
Summary
60
1 ln( t )
1
e 2
t 2
f (t ) =
2010 RAMS Tutorial DOE Guo and Mettas
56
1 m
e
m
61
Life-Factor Relationship
where:
Lognormal:
xj :
Exponential: m
' = ln( )
or
'=
or
' = ln(m)
63
2010 RAMS Tutorial DOE Guo and Mettas
Life-Factor Relationship
L f = f (Ti ; i, )
Suspension Data
'
i
Life-factor relationship:
i =1
M
LS = R ( S j ; i, )
j =1
=
LI
Interval Data
[ F (I
l =1
Ul
0 , 1 , 2 ,...
; i, ) F ( I Ll ; i, ) ]
MLE
L = L f Ls LI
and
64
LR (effect k ) = 2 ln
L(effect k removed )
L( full Model )
for lognormal
If
2010 RAMS Tutorial DOE Guo and Mettas
65
2010 RAMS Tutorial DOE Guo and Mettas
66
62
R-DOE: Example
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
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
67
2010 RAMS Tutorial DOE Guo and Mettas
68
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
69
70
For example,
H0: 1 = 0; H1: 1 <> 0
se( 1 )
se( 1 ) is the standard error.
71
2010 RAMS Tutorial DOE Guo and Mettas
72
Z-test Results
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
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
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
75
2010 RAMS Tutorial DOE Guo and Mettas
76
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.
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
2010 RAMS Tutorial DOE Guo and Mettas
77
2010 RAMS Tutorial DOE Guo and Mettas
78
More Information
http://www.weibull.com/doewebcontents.htm
http://www.itl.nist.gov/div898/handbook/
83
79
Two Level
Factorial Design
Summary
Response
Surface Method
Reliability DOE
Summary
80
81
Blocking
Power and sample size
RSM with multiple responses
RSM: Box-Behnken design
D-optimal design
Taguchi robust design
Taguchi orthogonal array (OA)
Mixture design
Random and mixed effect model
more
2010 RAMS Tutorial DOE Guo and Mettas
82
The End
84