Probabilistic Fatigue

1
Introduction
Professor Darrell F. Socie
Department of Mechanical and
Industrial Engineering
2003-2005 Darrell Socie, All Rights Reserved
Probabilistic Aspects of Fatigue
Probabilistic Fatigue 2003-2005 Darrell Socie, University of Illinois at Urbana-Champaign, All Rights Reserved 1 of 352
Contact Information
Darrell Socie
Mechanical Engineering
1206 West Green
Urbana, Illinois 61801
Office: 3015 Mechanical Engineering Laboratory
dsocie@uiuc.edu
Tel: 217 333 7630
Fax: 217 333 5634
2
Fatigue Calculations
Who really believes
these numbers ?
SAE Specimen
Suspension
Transmission
Bracket
Fatigue Under Complex Loading: Analysis and Experiments, SAE AE6, 1977
3
Analysis Results
1 10 100
48 Data Points
COV 1.27
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Analytical Life / Experimental Life
Strain-Life analysis of all test data
Non conservative
Conservative
Material Variability
1 10 100
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Material Analysis
Strain-Life back calculation of specimen lives
4
Probabilistic Models
Probabilistic models are no better than the
underlying deterministic models
They require more work to implement
Why use them?
Quality and Cost
Taguchi
Identify factors that influence performance
Robust design reduce sensitivity to noise
Assess economic impact of variation
Risk / Reliability
What is the increased risk from reduced testing ?
5
Risk
10
-9
10
-8
10
-7
10
-6
10
-5
10
-4
10
-3
R
i
s
k
Time, Flights etc
Acceptable risk
Reliability
99 %
80 %
50 %
10 %
1 %
0.1 %
E
x
p
e
c
t
e
d

F
a
i
l
u
r
e
s
10
6
Fatigue Life
10
5
10
4
10
3
6
Risk Contribution Factors
Operating
Temperature
Analysis Uncertainty
Speed
Material
Properties
Manufacturing
Flaws
Uncertainty and Variability
customers
materials manufacturing
usage
10
7
Fatigue Life, 2N
f
1
10 10
2
10
3
10
4
10
5
10
6
0.1
10
-2
1
10
-3
10
-4
S
t
r
a
i
n

A
m
p
li
t
u
d
e
time
50%
100 %
F
a
i
l
u
r
e
s
Strength
Stress
7
Deterministic versus Random
Deterministic from past measurements the future position
of a satellite can be predicted with reasonable accuracy
Random from past measurements the future position of
a car can only be described in terms of probability and
statistical averages
Deterministic Design
Stress Strength
Safety
Factor
Variability and uncertainty is accommodated by introducing
safety factors. Larger safety factors are better, but how much
better and at what cost?
8
Probabilistic Design
Stress Strength
Reliability = 1 P( Stress > Strength )
3 Approach
3 contains 99.87% of the data
If we use 3 on both stress and strength
The probability of the part with the lowest strength
having the highest stress is very small
P( s < S ) = 2.3 10
-3
= =

5 . 4 10 3 . 5 ) S s s ( P ) failure ( P
6
I
For 3 variables, each at 3 :
=

7 . 5 10 2 . 1 ) failure ( P
8
9
Benefits
Reduces conservatism (cost) compared to
assuming the worst case for every design
variable
Quantifies life drivers what are the most
important variables and how well are they
known or controlled ?
Quantifies risk
Introduction
Basic Probability and Statistics
Statistical Techniques
Analysis Methods
Characterizing Variability
Case Studies
FatigueCalculator.com
GlyphWorks
10
Introduction
Analysis Methods
Case Studies
GlyphWorks
11
Deterministic versus Random
Deterministic from past measurements the future position
of a satellite can be predicted with reasonable accuracy
Random from past measurements the future position of
a car can only be described in terms of probability and
statistical averages
Random Variables
Discrete - fixed number of outcomes
Colors
Continuous - may have any value in the
sample space
Strength
12
Descriptive Statistics
Mean or Expected Value
Variance / Standard Deviation
Coefficient of Variation
Skewness
Kurtosis
Correlation Coefficient
Central tendency of the data
( )
N
x
X E x Mean
N
1 i
i
x
=
= = = =
13
Variance / Standard Deviation
Dispersion of the data
( )
N
) x x (
X Var
N
1 i
2
i
=

=
) X ( Var
x
=
Standard deviation
Coefficient of Variation
x
x
COV
=
Useful to compare different dispersions
= 10
= 1
COV = 0.1
= 100
= 10
COV = 0.1
14
Skewness
Skewness is a measure of the asymmetry of the data
around the sample mean. If skewness is negative, the
data are spread out more to the left of the mean than
to the right. If skewness is positive, the data are spread
out more to the right. The skewness of the normal
distribution (or any perfectly symmetric distribution) is zero.
( )
3
N
1 i
3
i
N
) x x (
X Skewness
=

=
Kurtosis
Kurtosis is a measure of how outlier-prone a distribution is.
The kurtosis of the normal distribution is 3. Distributions that
are more outlier-prone than the normal distribution have
kurtosis greater than 3; distributions that are less
outlier-prone have kurtosis less than 3.
( )
4
N
1 i
4
i
N
) x x (
X Kurtosis
=

=
15
Covariance
A measure of the linear association between
random variables
| | ) Y ( ) X ( E ) Y , X ( COV
Y X xy
= =
Correlation Coefficient
0
X
Y
0
X
Y
0
X
Y
0 X
Y
0 X
Y
0 X
Y
= 1 = 0 = -1
0 < < 1 ~0 ~ 0
| |
Y X
Y X
) Y ( ) X ( E

=
16
Probability
Basic probability
Conditional probability
Reliability
Basic Probability
( ) 1 A P 0
( ) certain 1 A P =
( ) e imposssibl 0 A P =
The probability of event A occurring:
17
Conditional Basic Probability
( ) ( ) ( ) ( )
B and A B or A
B A P B P A P B A P
I U
+ =
( )
( )
( ) A P
B A P
A B P
I
=
( )
( )
( ) B P
B A P
B A P
I
=
P(B) given A has occurred
union intersection
Independent Events
1 2
If A and B are unrelated
( ) ( ) B P A B P =
( ) ( ) A P B A P =
( ) ( ) ( ) B P A P B A P = I
Suppose the probability of bar 1 failing is 0.03
and the probability of bar 2 failing is 0.04.
What is the probability of the structure failing?
18
Failure Probability
Bar 1 or bar 2 fails
( ) 03 . 0 A P =
( ) 04 . 0 B P =
( ) ( ) ( ) 0012 . 0 B P A P B A P = = I
P( failure) = 0.03 + 0.04 - 0.0012 = 0.0688
( ) ( ) ( ) ( ) B A P B P A P B A P I U + =
Reliability
P( no failure ) = Reliability
( ) A not of y probabilit A P Define
( ) ( ) A 1 P A P =
( ) B A P liability Re I =
( ) ( ) ( ) B P A P B A P = I
( ) 9312 . 0 96 . 0 97 . 0 B A P = = I
For the 2 bar structure
( ) 0688 . 0 liability Re 1 failure P = =
19
Structural Reliability
Collapse occurs if any member fails
( ) ( )
n 4 3 2 1
A A A A A P A P = I I I I
( ) ( ) ( ) ( ) ( ) ( )
n 4 3 2 1
A P A P A P A P A P A P =
( ) link each for 01 . 0 A P Suppose =
( ) ( ) 932 . 0 99 . 0 A P
7
= =
Reliability
6
5
4
3
2
1
0
10
-9
10
-8
10
-7
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
1
S
t
a
n
d
a
r
d

D
e
v
i
a
t
i
o
n
s
Probability
20
6 Sigma
6 sigma is 1 in a billion ( 0.999999999 Reliability )
Suppose a structure has 1000 bolted joints:
( ) ( ) 999999 . 0 999999999 . 0 A P
1000
= =
1 in a million
3 sigma is ( 0.99865 Reliability )
( ) ( ) 26 . 0 99865 . 0 A P
1000
= =
74 % failures
Statistical Distributions
Uniform
Normal
LogNormal
Gumble
Weibull
http://www.itl.nist.gov/div898/handbook/
A useful on-line reference:
21
Cumulative Distribution Function
( ) ( ) x X P X F
x
=
( ) 0 F
x
=
( ) 1 F
x
=
0
1
a b
F
x
(X)
Probability Density Function
( ) ( )dx X f b X a P
b
a
x
= <
( ) ( ) =

d f X F
x
x x
0
1
a b
f
x
(X)
a b
1
22
Probability
( ) ( )dx X f x X x P
x
x
x
+

= + <
0
1
a b
f
x
(X)
a b
1
x
2
Survival Function
0
1
a b
F
x
(X)
( ) X F 1 ) X ( S
x x
=
23
Hazard Function
a b
h
x
(X)
( ) X F 1
) X ( f
) X ( h
x
x
x
=
Instantaneous failure rate
0
1
a b
f
x
(X)
( )

= = dx ) X ( f x X E
x x
24
Variance
( ) dx ) X ( f ) X ( X Var
x
2
x

=
( ) ( )
2
x
2
X E X Var =
Probability Density
x
) X ( Var
x
=
Normal Distribution
0.1
0.2
0.3
0.4
-3 -2 -1 0 1 2 3
Quantile x
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y
-3 -2 -1 0 1 2 3
Quantile x
0.5
1.0
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
(

=
2
2
x
2
) x (
exp
2
1
) x ( f
tables
25
Normal Probability Plot
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
0
1
2
3
-1
-2
-3
S
t
a
n
d
a
r
d

D
e
v
i
a
t
i
o
n
s
Linear Scale
1 N
m
P
+
=
Normal Plot
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
0 200 400 600 800 1000
Mean 432.67
COV 0.34
Maximum force from 42 drivers
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
26
LogNormal Distribution
0.5
1.0
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
1 3 2
Quantile x
0.2
0.4
0.6
0.8
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y
1 3 2
Quantile x
= 0.6
= 1.0
= 0.6
= 1.0
(
(
(
(
|
.
|
\
|

=
2
2
x
2
)
m
x
ln (
exp
2 x
1
) x ( f
tables
( ) = exp m
Useful Relationships
( )
2
x ln x ln x
5 . 0 exp + =
|
|
.
|
\
|
+
=
2
x
x
x ln
COV 1
ln
( )
2
x
2
x ln
COV 1 ln + =
1 ) exp( COV
2
x ln x
=
2
x
x
x ln x
COV 1
) exp( X
+
= =
27
Median / Mean ratio
COV
X
0 1 2 3 4 5
0
0.2
0.4
0.6
0.8
1.0
X
X
lnX
- COV
X
COV
X
0 0.5 1.0 1.5 2.0
0.5
1.0
1.5
l
n
X
28
LogNormal Probability Plot
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
0
1
2
3
-1
-2
-3
S
t
a
n
d
a
r
d

D
e
v
i
a
t
i
o
n
s
Log Scale
LogNormal Plot
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
100
1000
Median 431
COV 0.34
200 500
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
29
Gumbel Extreme Value
a-3b a+4b a a-3b a+4b a
1.0
0.5
C
u
m
u
l
a
t
i
v
e
P
r
o
b
a
b
i
l
i
t
y
Quantile x Quantile x
0.5
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y
)
`

=
b
) a x (
exp exp ) x ( F
x
)
`

=
b
) a x (
exp exp
b
) a x (
exp
b
1
) x ( f
x
Gumble Probability Plot
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
99.9 %
99 %
80 %
50 %
10 %
1 %
99.9 %
99 %
80 %
50 %
10 %
1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Linear Scale
-2
-1
0
1
2
3
4
5
6
7
-
l
n
(

-
l
n
(

F
x
(
X
)

)

)
30
Gumble Plot
200 400 600 800
Location 362.94
Scale 124.61
Maximum force from 42 drivers 99.9 %
99 %
80 %
50 %
10 %
1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Weibull
1 3 2
Quantile x
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y
1.0
0.8
0.2
0.6
0.4
1 3 2
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Quantile x
c=0.5
c=1.0
c=3.0
1.0
0.8
0.2
0.6
0.4
c=0.5
c=1.0 c=3.0
(
(
|
.
|
\
|
=
c
x
b
x
exp 1 ) x ( F
(
(
|
.
|
\
|
|
|
.
|
\
|
=

c
c
1 c
x
b
x
exp
b
x c
) x ( f
31
Weibull Probability Plot
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
-7
-6
-5
-4
-3
-2
-1
0
1
2
l
n
(

-
l
n
(

F
x
(
X
)

)

)
99.9 %
99 %
80 %
50 %
10 %
1 %
0.1 %
Log Scale
Weibull Plot
200 1000
99.9 %
99 %
80 %
50 %
10 %
1 %
0.1 %
Scale Factor: 473.56
Slope: 3.8
500
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
32
99.9 % Probability
785 Weibull
1223 Gumbel
1125 LogNormal
893 Normal
Estimated maximum force from the distributions
Comparison
0 200 400 600 800 1000 1200
Maximum Load
0.001
0.002
0.003
0.004
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y
Normal
LogNormal
Gumble
Weibull c = 3.5
All distributions are approximately normal around the median
33
Joint Probability Density
x
y
f
xy
(x,y)
Y
Strength
Stress
( )
( ) (
(
|
|
.
|
\
|

+
|
|
.
|
\
|

|
|
.
|
\
|

|
|
.
|
\
|

=
2
y
y
y
y
x
x
2
x
x
2
2
y x
xy
y y
x
2
x
1 2
1
exp
1 2
1
y , x f
correlation coefficient
Normal distributions
34
2D - Joint Probability Density
Contours of
equal probability
y
x
Y
-3
x
3
Y
1
2
3
4
Stress
Strength
Acceptable Risk
Safety
P(failure) ~ 10
-6
10
-9
Economic
P(failure) ~ 10
-2
10
-4
35
Extreme Values
For most durability problems, we are not interested in
the large extremes of stress or strength. Failure is much
more likely to come from moderately high stresses
combined with moderately low strengths.
Key Points
Basic Nomenclature
Random Variables
Statistical Distribution
Cumulative distribution function
Mean
Variance
Probability
Marginal
Conditional
Joint
36
Introduction
Analysis Methods
Case Studies
GlyphWorks
37
Normal Distributions
LogNormal Distributions
Monte Carlo
Sampling
Distribution Fitting
Failure Probability
) S s s ( P I
Let be the stress and S the fatigue strength
Given the distributions of and S find the
probability of failure
Stress, Strength, S
s
38
Normal Variables
Linear Response Function
=
+ =
n
1 i
i i o
X a a Z
X
i
~ N(
i
, C
i
)
=
+ =
n
1 i
i i o z
a a
=
=
n
1 i
2
i
2
i z
a
Calculations
Z = S -
z
=
S
-
2 2
S z
+ =
~ N( 100 , 0.2 )
= 20
z
= 200 -100 = 100
S ~ N( 200 , 0.1 )
S
= 20
Stress, Strength, S
Safety factor of 2
2 . 28 20 20
2 2
z
= + =
Let Z be a random variable:
39
Failure Probability
Failure will occur whenever Z <= 0
Z = S -
Z =
z
z
z
= 0
z = 3.54 standard deviations
P(failure) = 2 x 10
-4
2 . 28
100
z
Z
Z
=
=
For this case only, a safety factor of 2 means a probability
of failure of 2 x 10
-4
. Other situations will require different
safety factors to achieve the same reliability.
Failure Distribution
Stress, Strength, S
What is the expected distribution in fatigue lives?
40
Fatigue Data
b
f
'
f
) N 2 ( 2
S
=
10
2
10
3
10
4
10
5
10
6
10
7
Fatigue Life
10 1
1000
100
10
1
S
t
r
e
s
s

A
m
p
l
i
t
u
d
e
'
f
b
1
'
f
f
2
S
N 2
|
|
.
|
\
|
=
b
1
LogNormal Variables
=
=
n
1 i
a
i o
i
X a Z
as are constant and X
i
~ LN( x
i
, C
i
)
=
=
n
1 i
a
i o
i
X a Z median
( )
=
+ =
n
1 i
a
2
X Z
1 C 1 C COV
2
i
i
41
Calculations
Stress, Strength, S
~ LN( 250 , 0.2 ) = 50
~ LN( 1000 , 0.1 ) = 100
b
1
'
f
f
2
S
N 2
|
|
.
|
\
|
=
'
f
b = -0.125
2
S
8
'
f
8
f
2
S
N 2 Z |
.
|
\
|
= =

8
'
f
8
f
2
S
N 2 Z
|
|
.
|
\
|

= =

( ) ( ) 1 COV 1 COV 1 COV
2
f
2
8
2
8
2
S Z
+ + =

Results
'
f
S/2 2N
f
Percentile Life
x
250 1000 355,368 99.9 17,706,069
COV
x
0.2 0.1 4.72 99 4,566,613
95 1,363,200
lnx
5.50 6.90 11.21 90 715,589
X 245 995 73,676 50 73,676
x
50 100 1,676,831 10 7,586
lnx
0.198 0.100 1.774 5 3,982
1 1,189
b = -0.125 0.1 307
42
Monte Carlo Methods
( ) ( )
|
|
.
|
\
|
+
+c b
f
'
f
'
f
b 2
f
2
'
f f
N 2 N 2
E
E
2
S K
Given random variables for K
f
, S,
Find the distribution of 2N
f
'
f
'
f
and
Z = 2N
f
= ?
Simple Example
Probability of rolling a 3 on a die
1 2 3 4 5 6
f
x
(x) 1/6
Uniform discrete distribution
43
Computer Simulation
1. Generate random numbers between 1 and 6,
all integers
2. Count the number of 3s
Let X
i
= 1 if 3
0 otherwise
( )

=
=
n
1 i
i
X
n
1
3 P
EXCEL
=ROUNDUP( 6 * RAND() , 0 )
=IF( A1 = 3 , 1 , 0 )
=SUM($B$1:B1)/ROW(B1)
5 0 0
3 1 0.5
4 0 0.333333
4 0 0.25
5 0 0.2
6 0 0.166667
1 0 0.142857
3 1 0.25
3 1 0.333333
6 0 0.3
44
Results
0.1
0.2
0.3
0.4
1 100 200 300 400 500
trials
p
r
o
b
a
b
i
l
i
t
y
Evaluate
x
y
P( inside circle )
4
r
P
2
=
= 4 P
x = 2 * RAND() - 1
y = 2 * RAND() - 1
IF( x
2
+ y
2
< 1 , 1 , 0 )
1
-1
1
-1
45
0
1
2
3
4
1 100 200 300 400 500
trials
Monte Carlo Simulation
Stress, Strength, S
~ LN( 250 , 0.2 ) = 50
~ LN( 1000 , 0.1 ) = 100
'
f
b = -0.125
2
S
b
1
'
f
f
2
S
N 2
|
|
.
|
\
|
=
Randomly choose values of S and
from their distributions
'
f
Repeat many times

46
Generating Distributions
0
1
x
F
x
(X)
Randomly choose a value between 0 and 1
x = F
x
-1
( RAND )
RAND
Generating Distributions in EXCEL
=LOGINV(RAND(),ln,ln)
=NORMINV(RAND(),,)
Normal
Log Normal
47
EXCEL
893 204 134,677
1102 301 32,180
852 285 6,355
963 173 929,249
1050 283 35,565
1080 265 77,057
965 313 8,227
1073 213 420,456
1052 226 224,000
954 322 5,878
965 240 68,671
993 207 277,192
1191 368 11,967
831 210 59,473
'
f
2
S
2N
f
Simulation Results
10
3
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
10
4
10
5
10
6
10
7
10
8
Monte Carlo Analytical
Mean 11.25 11.21
Std 1.79 1.77
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Life
48
Summary
Simulation is relatively straightforward and simple
Obtaining the necessary input data is difficult
Nomenclature
Population: the totality of the observations
Sample: a subset of the population
Population mean:
x
Population variance:
x
2
Sample mean: X
Sample variance: s
2
( )
x
X E =
( )
2
x
2
s E =
49
Sample Mean and Variance
in EXCEL =AVERAGE(A1:A
n
)
in EXCEL =STDEV(A1:A
n
)
n
x
X
n
1
=
1 n
) X x (
s
n
1
2
=

Mean
Variance
Confidence Intervals
( )
x
X E =
What is the probability that a sample X is greater than
x
? 50%
P( L <=
x
<= U ) = 1 -
There is a 1 chance of selecting a sample in the interval
between L and U that contains the true mean of the population
50
Translation
90% confidence
If we sampled a population many times to
estimate the mean, 90% of the time the true
population mean would lie between the
computed upper and lower limit.
Confidence Interval - mean
Lower limit of
x
x
1 n ,
n
s
t X

Upper limit of
x
x
1 n ,
n
s
t X +

For a normal distribution:
51
Sample Size
0 5 10 15 20 25 30
Sample size, n
0
0.5
1.0
1.5
2.0
x
x
1 n ,
n
s
t X

n
t
1 n ,
95%
90%
5 Samples from Normal Distribution
0
20
40
60
80
100
120
140
0 10 20 30 40 50 60 70 80 90 100
95 % confidence 5 samples will be outside confidence limit
Lower limit above mean
Upper limit below mean
Sample Number
x
= 100
x
= 10
52
Confidence Interval - variance
Upper limit of
2
x
1 n , 1
2
2
x
s ) 1 n (

For a normal distribution:
Sample Size
0 5 10 15 20 25 30
0
2
4
6
8
10
Sample size, n
2
x
1 n , 1
2
2
x
s ) 1 n (

1 n , 1
2
) 1 n (

95%
50%
90%
53
Maximum Load Data
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
100
1000
Median 431
COV 0.34
200 500
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
90% confidence COV 0.41
Maximum Load Data
Maximum Load
0 200 400 600 800 1000 1200
0.001
0.002
0.003
0.004
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y
Normal COV = 0.34
LogNormal
Normal COV = 0.41
Uncertainty in Variance is just as important,
perhaps more important than the choice of the distribution
54
Choose the Best Distribution
0 200 400 600 800
1000
Maximum Load
Normal
LogNormal
Gumble
Weibull
0.001
0.002
0.003
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y
15 samples from a Normal Distribution
Goodness of Fit Tests
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
0 200 400 600 800 1000
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
0 200 400 600 800 1000
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
100 1000 200 500
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
100 1000 200 500
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
99.9 %
99 %
80 %
50 %
10 %
1 %
200 400 600 800
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
99.9 %
99 %
80 %
50 %
10 %
1 %
200 400 600 800
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Normal
LogNormal
Gumble
200
99.9 %
99 %
80 %
50 %
10 %
1 %
0.1 %
1000 500
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
200
99.9 %
99 %
80 %
50 %
10 %
1 %
0.1 %
1000 500
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Weibull
55
Analytical Tests
Stephens, M. A. (1974). EDF Statistics for Goodness of Fit and Some Comparisons,
Journal of the American Statistical Association, Vol. 69, pp. 730-737.
Kolmogorov-Smirnov
Chakravarti, Laha, and Roy, (1967). Handbook of Methods of Applied Statistics, Volume I,
John Wiley and Sons, pp. 392-394.
Anderson-Darling
Snedecor, George W. and Cochran, William G. (1989), Statistical Methods, Eighth Edition,
Iowa State University Press.
Chi-square
any univariate distribution
tends to be more sensitive near the center of the distribution
gives more weight to the tails
Distributions
Normal
Strength
Dimensions
LogNormal
Fatigue Lives
Large variance in properties or loads
Gumble
Maximums in a population
Weibull
Fatigue Lives
56
Statistics Software
www.palisade.com
BestFit Minitab
www.minitab.com
Central Limit Theorem
as n is the standard normal distribution
If X
1
, X
2
, X
3
.. X
n
is a random sample from the population,
with sample mean X, then the limiting form of
n /
X
Z
X

=
57
Translation
When there are many variables affecting the outcome,
The final result will be normally distributed even if the
individual variable distributions are not.
Example
Probability of rolling a die
1 2 3 4 5 6
f
x
(x) 1/6
Uniform discrete distribution
Let Z be the summation of six dice
Z = X
1
+ X
2
+ X
3
+ X
4
+ X
5
+ X
6
58
Results
0
10
20
30
40
50
60
6 12 18 24 30 36
Z
F
r
e
q
u
e
n
c
y
Central limit theorem states that the result should be
normal for large n
500 trials
C
Z
= 0.20
X
Z
= 21.12
Central Limit Theorem
Sums: Z = X
1
X
2
X
3
X
4
.. X
n
Z Normal as n increases
Products: Z = X
1
X
2
X
3
X
4
.. X
n
Z LogNormal as n increases
Normal and LogNormal distributions are often employed
for analysis even though the underlying population
distribution is unknown.
59
Key Points
All variables are random and can be characterized
by a statistical distribution with a mean and variance.
The final result will be normally distributed even if
the individual variable distributions are not.
Analysis Methods
60
Introduction
Analysis Methods
Case Studies
GlyphWorks
Reliability Analysis
1 3 2
0.2
0.4
0.6
0.8
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y
0.2
0.4
0.6
0.8
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y
1 3 2
0.1
0.2
0.3
0.4
-3 -2 -1 0 1 2 3
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y
Stressing Variables
Strength Variables
0.2
0.4
0.6
0.8
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y
1 3 2
P( Failure )
Analysis
?
61
Probabilistic Analysis Methods
Monte Carlo
Simple
Hypercube sampling
Importance sampling
Analytical
First order reliability method FORM
Second order reliability method SORM
Limit States
Limit States
Equilibrium
Strength
Deformation
Wear
Functional
62
Limit State Problems
Z(X) = Z( Q, L, b, h )
Response function
Limit State function
g = Z(X) - Z
o
= 0
Same as the failure function
Limit States
Stress
Strength
g < 0
g > 0
failures
g = Strength - Stress
63
Monte Carlo
Stress
Strength
g < 0
g > 0
failures
Monte Carlo is simple but what if each of these calculations
required a separate FEM model?
Low Probabilities of Failure
Stress
Strength
failures
rare event
Need about 10
5
simulations for P(Failure) = 10
-4
64
Transform Variables
u
1
u
2
x
x
1
x
u
u1
= 0
u1
= 1
Standard Monte Carlo
-4 -3 -2 -1 0 1 2 3 4
-4
-3
-2
-1
0
1
2
3
4
u
1
u
2
1000 trials
g = 3 ( u
1
+ u
2
)
018 . 0
N
N
) f ( P
f
= =
u
1
and u
2
normally distributed
65
Multiple Simulations
1 25
2 15
3 19
4 21
5 16
6 22
7 16
8 12
9 15
10 11
11 19
12 15
13 20
14 17
15 9
16 23
17 15
18 20
19 14
20 21
21 16
22 18
23 22
24 13
25 20
1000 trials
25 simulations
= 0.017
= 0.0040
25,000 calculations
Stratified Sampling Methods
-4 -3 -2 -1 0 1 2 3 4
Divide into regions of constant
probability and then sample by region
66
Stratified Sample
-4 -3 -2 -1 0 1 2 3 4
-4
-3
-2
-1
0
1
2
3
4
u
1
u
2
Each box should have
one sample drawn at random
from the underlying pdf
100 trials
25 simulations
= 0.02
= 0.0093
2,500 calculations
Latin Hypercube Sample
-4 -3 -2 -1 0 1 2 3 4
-4
-3
-2
-1
0
1
2
3
4
u
1
u
2
10 trials
25 simulations
= 0.017
= 0.038
250 calculations
67
Importance Sampling (continued)
-4 -3 -2 -1 0 1 2 3 4
-4
-3
-2
-1
0
1
2
3
4
u
1
u
2
= 2
The probability of a sample
outside the circle is P
S
= 0.14
( 2 standard deviations )
0183 . 0 P
N
N
) f ( P
S
S
f
= =
138 simulations
Analytical Methods Strategy
Develop a response function
Transform the set of N variables to a set of
uncorrelated u variables with =0 and =1
Locate the most likely failure point
Approximate the integral of the joint
probability distribution over the failure region
68
Analytical Methods Outline
Response Surface
Transform Variables
Limit State Concept
Most Probable Point
Probability Integration
FORM
SORM
Response Surface Methodology
Response surface methodology is a well
established collection of mathematical and
statistical techniques for applications where the
response of interest is influenced by several
variables.
+ = ) x , x , x , x ( f ) X ( Z
n 3 2 1
69
Response Surface
X
1
X
2
Z
Fit a surface through the points
Solution points
Mathematical Representation
+ + +
+ + +
+ + + + =
3 2 3 3 1 2 2 1 1
2
3 3
2
2 2
2
1 1
3 3 2 2 1 1 o
x x C x x C x x C
x B x B x B
x A x A x A A ) X ( Z Linear
Incomplete Quadradic
Complete Quadradic
Solutions
Linear N + 1
Complete Quadradic N(N + 1)/2
Incomplete Quadradic 2N + 1
70
Evaluation of Response Surface
Factorial Design
Suppose stress is affected by speed and temperature
high low
h
i
g
h
l
o
w
X
X
X
X
temperature
s
p
e
e
d
4 deterministic solutions needed
2
n
solutions
Evaluation of Response Surface
Factorial Design
high low
h
i
g
h
l
o
w
X
X
X
X
temperature
s
p
e
e
d9 deterministic solutions needed
3
n
solutions
X X X
X
X
71
Graphical Representation
X
2
X
1
Z
x
y
f
xy
(x,y)
Stress
Strength
72
2D - Joint Probability Density
Contours of
equal probability
y
Stress
Strength
Transform Variables
x
x
1
x
u
u
= 0
u
= 1
Any distribution can be mapped into a normal distribution
u
1
u
2
d
( )
2
d 5 . 0 exp P
The mathematics and behavior of normal distributions
are well understood
73
X to u mapping
f(X)
F(X) F(X)
f(X)
uniform
normal
Limit States
Limit states define probability problems
g = Strength(x
1
) Stress(x
2
)
For example:
Prob( g<= 0 ) = Probability of failure = P
f
Analysis focuses on g = 0
74
Failure Probability
g = 0
u
1
u
2
<
=
0 g
2 1 2 1 u u
du du ) u , u ( f ) Failure ( P
2 1
Thought Experiment
Suppose stress and strength had the same distribution
What is the probability of failure?
-3 -2 -1 0 1 2 3
x
0.1
0.2
0.3
0.4
P
r
o
b
a
b
i
l
i
t
y

D
e
n
s
i
t
y
75
g function
u
1
u
2
g = Strength(x
1
) Stress(x
2
)
P
f
= 50%
Probable Combinations of u
1
and u
2
g = 0
u
1
u
2
Most likely failure point
g functions are not straight lines in N dimensional space
76
Most Probable Point
u
1
g = 0
u
2
Probability is related to the minimum distance point

Safety Index
Sometimes called reliability index
Unlike safety factors, failure probability is
directly related to the safety index
P
f
= (-)
is a number measured in standard deviations
77
Determining the MPP
Requires an efficient numerical search to find the
tangent point of a hypersphere (-sphere) and the limit
state function in u space
Probability Integration
FORM
SORM
78
First Order Reliability Model
g = 0
u
2
exact
Linear approximation
u
1
( ) ( )
=
+ =
n
1 i
i i i o
* u u a a u g
FORM
MPP
FORM
u
1
P
f
(-)
u
2
MPP
u
2
*
u
1
*
The joint probability density is rotationally symmetric
so it is possible to rotate the coordinate system
79
Second Order Reliability Model
g = 0
u
2
exact
Quadradic approximation
u
1
( ) ( ) ( ) ( )( )

= = = =
+ + + =
n
1 i
n
1 i
n
1 i
j j i i
n
1 j
ij
2
i i i i i i o
* u u * u u c * u u b * u u a a u g
SORM
SORM
u
2
MPP
u
2
*
u
1
*
( ) ( )
=
=
n
1 i
i f
1 P
i
surface curvature
80
Sensitivity Factors
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Design
point
Change in response
with a change in mean
Change in probability
with a change in mean
Sensitivity Factors
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Design
point
Change in response
with a change in variance
Change in probability
with a change in variance
81
Deterministic Sensitivity Factors
i
X
) X ( Z
Change in response mean with respect to a

change in the input variable mean
Frequently normalized by the means to compare
relative importance of variables
X / X
Z / ) X ( Z
i
Probabilistic Sensitivity Coefficient
i i
i
/
P / P
S

i i
i
/
P / P
S

Standard deviation sensitivity

Mean deviation sensitivity
82
Probabilistic Sensitivity Factors
1
X
) X ( Z
2
i i
i
i
=
|
|
.
|
\
|
i
- probabilistic sensitivity
Z(X) - response
X
i
- input variable
i
- standard deviation of X
i
Software
General Purpose
Durability
83
Variability
Introduction
Analysis Methods
Case Studies
GlyphWorks
84
Sources of Variability
customers
materials manufacturing
usage
10
7
Fatigue Life, 2N
f
1
10 10
2
10
3
10
4
10
5
10
6
0.1
10
-2
1
10
-3
10
-4
S
t
r
a
i
n

A
m
p
li
t
u
d
e
Strength
Stress
Stress, Strength, S
Variability and Uncertainty
Variability: Every apple on a tree has a different mass.
Uncertainty: The variety of the apple is unknown.
Variability: Fracture toughness of a material
Uncertainty: The correct stress intensity factor solution
85
Sources of Variability
Stress Variables
Loading
Customer Usage
Environment
Strength Variables
Material
Processing
Manufacturing Tolerance
Environment
Sources of Uncertainty
Statistical Uncertainty
Incomplete data (small sample sizes)
Modeling Error
Analysis assumptions
Human Error
Calculation errors
Judgment errors
86
Modeling Variability
Products: Z = X
1
X
2
X
3
X
4
.. X
n
Central Limit Theorem:
COV
X
0 0.5 1.0 1.5 2.0
0.5
1.0
1.5
l
n
X
lnX
~ COV
X
COV
X
is a good measure of variability
Standard Deviation, lnx
COV
X
1 2 3
68.3% 95.4% 99.7%
0.05
0.1
0.25
0.5
1
1.05
1.10
1.28
1.60
2.30
1.11
1.23
1.66
2.64
5.53
1.16
1.33
2.04
3.92
11.1
99.7% of the data is within a factor of 1.33 of the mean for a COV = 0.1
COV and LogNormal Distributions
87
Variability in Service Loading
Quantifying Loading Variability
Maximum Load
Load Range
Equivalent Stress
Maximum Force
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
100
1000
Median 431
COV 0.34
200 500
Maximum force from 42 automobile drivers
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
100
1000
200 500
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Force
88
Maximum Load Correlation
0 200 400 600 800 1000
10
8
10
7
10
6
10
5
10
4
10
3
Maximum Load
F
a
t
i
g
u
e

L
i
v
e
s
Loading Variability
1 10 10
2
10
3
10
4
10
5
Cumulative Cycles
0
16
32
48
L
o
a
d

R
a
n
g
e
54 Tractors / Drivers
89
Variability in Loading
1 10
54 Tractors/Drivers
COV 0.53
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
n
n
eq
S S

=
Equivalent Load
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Mechanisms and Slopes
100
10
3
10
4
S
t
r
e
s
s

A
m
p
l
i
t
u
d
e
,

M
P
a
10
0
Cycles
10
1
10
2
10
3
10
4
10
5
10
6
10
7
1
10
10
-12
10
-11
10
-10
10
-9
10
-8
10
-7
10
-6
Crack Growth Rate, m/cycle
1
10
100
m
M
P
a
,
K
1
3
Crack Nucleation
Crack Growth
10
100
10
3
10
4
10
5
10
6
10
7
10
8
Total Fatigue Life, Cycles
1
5
1
E
q
u
i
v
a
l
e
n
t

L
o
a
d
,

k
N
Structures
A combination of nucleation and growth
n = 3
n = 5
n = 10
90
Effect of Slope on Variability
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
n = 10 5 3
Equivalent Load
0.1 1 10 0.1 1 10
n COV
3 0.53
5 0.43
10 0.38
Loading History Variability
Test Track
Customer Service
91
Test Track Variability
0.1
1 10
COV 0.12
40 test track laps of a motorhome
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Equivalent Load
Customer Usage Variability
0.1
1 10
COV 0.32
42 drivers / cars
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Equivalent Load
92
Variability in Environment
Inclusions
Pit depth
Inclusions That Initiated Cracks
Barter, S. A., Sharp, P. K., Holden, G. & Clark, G. Initiation and early growth of fatigue cracks in an aerospace
aluminium alloy, Fatigue & Fracture of Engineering Materials & Structures 25 (2), 111-125.
COV = 0.27
93
Pits That Initiated Cracks
7010-T7651
Pre-corroded specimens
300 specimens
246 failed from pits
Crawford et.al.The EIFS Distribution for Anodized and Pre-corroded 7010-T7651 under Constant Amplitude Loading
Fatigue and Fracture of Engineering Materials and Structures, Vol. 28, No. 9 2005, 795-808
Pit Size Distribution
40
30
20
10
100 200 400 500 300
F
r
e
q
u
e
n
c
y
area m
Mean = 230
COV = 0.32
94
Pit Depth Variability
Pits
12 Data Points
Median 24.37
100
COV 0.33
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
10
Pit Depth, m
Dolly, Lee, Wei, The Effect of Pitting Corrosion on Fatigue Life
Fatigue and Fracture of Engineering Materials and Structures, Vol. 23, 2000, 555-560
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Variability in Materials
Tensile Strength
Fracture Toughness
Fatigue
Fatigue Strength
Fatigue Life
Strain-Life
Crack Growth
95
Tensile Strength - 1035 Steel
25
50
75
100
500 550 600 650 700
N
u
m
b
e
r

o
f

h
e
a
t
s
Tensile Strength, MPa
Metals Handbook, 8
th
Edition, Vol. 1, p64
Mean = 602
COV = 0.045
Fracture Toughness
100
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
Median 76.7
COV 0.06
60 70 90 80
26 Data Points
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
K
Ic
, Ksiin
Kies, J.A., Smith, H.L., Romine, H.E. and Bernstein, H, Fracture Testing of Weldments, ASTM STP 381, 1965, 328-356
Mar-M 250 Steel
96
Fatigue Variability
10
2
10
3
10
4
10
5
10
6
10
7
Fatigue Life
10 1
1000
100
10
1
S
t
r
e
s
s

A
m
p
l
i
t
u
d
e
Fatigue life
F
a
t
i
g
u
e

s
t
r
e
n
g
t
h
Fatigue Life Variability
50 100 150 200 250
10
20
30
40
10
20
N
u
m
b
e
r

o
f

T
e
s
t
s
Life x10
3
50 100 150 200 250
10
20
30
40
10
20
N
u
m
b
e
r

o
f

T
e
s
t
s
Life x10
3
Production torsion bars
5160H steel One lot, 71 parts
25 lots, 300 parts
Metals Handbook, 8
th
Edition, Vol. 1, p219
x
= 123,000 cycles
COV = 0.25
x
= 134,000 cycles
COV = 0.27
97
Statistical Variability of Fatigue Life
50
10
2
90
98
10
4
10
5
10
6
10
7
10
8
Cycles to Failure
P
e
r
c
e
n
t

S
u
r
v
i
v
a
l
s=210 s=245 s=315
s=280 s=440
Sinclair and Dolan, Effect of Stress Amplitude on the Variability in Fatigue Life of 7075-T6 Aluminum Alloy
Transactions ASME, 1953
7075-T6 Specimens
COV vs Fatigue Life
440 14,000 0.12
315 25,000 0.38
280 220,000 0.70
245 1,200,000 0.67
210 12,000,000 1.39
S COV X
98
Variability in Fatigue Strength
( )
=
+ =
n
1 i
a
2
X
1 C 1 C COV
2
i
i
085 . 0 b ) N ( S
2
S
b
f
'
f
=
( ) 088 . 0 1 39 . 1 1 C
2
'
f
) 085 . (
2
S
= + =

Strain Life Data for 950X Steel
10
2
10
3
10
4
10
5
10
6
10
7
Fatigue Life
10 1
1
0.1
10
-2
10
-3
10
-4
S
t
r
a
i
n

A
m
p
l
i
t
u
d
e
378 Fatigue Tests
29 Data Sets
99
29 Individual Data Sets
'
f
Median 820
300 1000
COV 0.25
2000
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
b
-0.12 -0.08 -0.04
Mean -0.09
COV 0.25
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
29 Individual Data Sets (continued)
0.1 1 10
Median 0.57
COV 1.15
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
'
f
-0.8 -0.6 -0.4 -0.2

c
Mean -0.62
COV 0.23
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
100
Input Data Simulation
( )
( )
( )
( )( )
( )
b b
f f
b b f f
, , N c
f
'
f
, , N b
f
'
f
N 2 , , L N 2
E
, , L
2

+

=

Simulation Results
0.0001
0.001
0.01
0.1
1
10
100
1 10 100 10
3
10
4
10
5
10
6
10
7
S
t
r
a
i
n

A
m
p
l
i
t
u
d
e
Fatigue Life
101
Correlation
0
0.25
0.5
0.75
1.0
0.01 0.1 1 10
'
f
c
0
0.05
0.1
0.15
0.2
10
2
10
3
10
4
'
f
b
= -0.828 = -0.976
Generating Correlated Data
z
1
= ( rand() )
z
2
= ( rand() )
2
2 1 3
1 z z z + =
) z exp(
1 ln ln
'
f
'
f
'
f

+ =
3 b b
z b + =
z
1
= N(0,1)
102
Correlated Properties
0.0001
0.001
0.01
0.1
1
10
100
1 10 100 10
3
10
4
10
5
10
6
10
7
Fatigue Life
S
t
r
a
i
n

A
m
p
l
i
t
u
d
e
Curve Fitting
10
2
10
3
10
4
10
5
10
6
10
7
Fatigue Life
10 1
1
0.1
10
-2
10
-3
10
-4
S
t
r
a
i
n

A
m
p
l
i
t
u
d
e
c
f
'
f
b
f
'
f
) N 2 ( ) N 2 (
E 2
+
=

Assume a constant slope to get
a distribution of properties
'
f
103
Property Distribution
0.1 1
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
365 Data Points
Median 0.26
COV 0.42
'
f
378 Data Points

Median 802 MPa
COV 0.12
'
f
100 1000
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
b = -0.086 c = -0.51
Correlation
0
500
1000
1500
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
'
f
'
f
104
Simulation
0.001
0.01
0.1
1
10
0.0001
1 10 100 10
3
10
4
10
5
10
6
10
7
S
t
r
a
i
n

A
m
p
l
i
t
u
d
e
Fatigue Life
Strength Coefficient
5000
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
500
0
500
1000
1500
0 500 1000 1500
2000 2500
365 Data Points
Median 1002 MPa
COV 0.14
'
f
' K
' K
= 0.863
105
Crack Growth Data
0
10
20
30
40
50
C
r
a
c
k

L
e
n
g
t
h
,

m
m
0 50 100 150 200 250 300 350
Cycles x10
3
Virkler, Hillberry and Goel, The Statistical Nature of Fatigue Crack Propagation, Journal of Engineering Materials
and Technology, Vol. 101, 1979, 148-153
Crack Growth Rate Data
300,000
68 Data Points
Median 257,000
COV 0.07
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Fatigue Lives Crack Growth Rate
5 10 50
10
-5
10
-4
10
-3
10
-2
C
r
a
c
k

g
r
o
w
t
h

r
a
t
e
,

m
m
/
c
y
c
l
e
Stress intensity, MPam
106
Crack Growth Properties
Median 5 x 10
-8
COV 0.44
10
-7
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
10
-6
Median 3.13
COV 0.06
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
2 5 4
m
K C
dN
da
=
C m
Beware of Correlated Variables
( ) 2 / m 1 S C
a a
N
2
m
m
2 / m 1
i
2 / m 1
f
f

=

N
f
and C are linearly related and should have
the same variability, but
07 . 0 COV
f
N
=
44 . 0 COV
C
=
because C and m are correlated.
107
Correlation
2.0
2.5
3.0
3.5
4.0
10
-8
10
-6
10
-7
C
m
Correlated Variables
2
Y Y X
2
X Z
2 + + =
-1 -0.5 0 0.5 1
0.5
1
1.5
2
X
=
Y
= 1
108
Calculated Lives
10
5
10
6
Computed from
C and m pairs
experimental
10
5
10
6
0.1 % 0.1 %
99.9 %
99 %
90 %
50 %
10 %
1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
99.9 %
99 %
90 %
50 %
10 %
1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
experimental
( )

=
f
o
a
a
m
f
K C
da
N
Manufacturing/Processing Variability
Bolt Forces
Surface Finish
Drilled Holes
109
Variability in Bolt Force
100 1000
Force
200 Data Points
Median 130
COV 0.14
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
Bolt Force, kN
Preload force in bolts tightened to 350 Nm
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Surface Roughness Variability
100
125 Data Points
Median 43
COV 0.10
Machined aluminum casting
10
Surface Finish, in
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Surface Finish
110
Drilled Holes
Fighter Spectrum
154 Data Points
Median 126,750
COV 0.22 in life
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
10
5
Cycles
From: J.P. Butler and D.A. Rees, "Development of Statistical Fatigue Failure Characteristics of 0.125-inch
2024-T3 Aluminum Under Simulated Flight-by-Flight Loading," ADA-002310 (NTIS no.), July 1974.
180 drilled holes in a single plate
COV 0.07 in strength
Miners Linear Damage rule
Strain Life Analysis
111
50
90
99.9
99
10
1
0.01
0.01 0.1 1 10 100
Computed Life / Experimental
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y

o
f

F
a
i
l
u
r
e964 Tests
COV = 1.02
Miners Rule
From Erwin Haibach Betriebsfestigkeit, Springer-Verlag, 2002
A safety factor of 10 in life would result in a 10% chance of failure
SAE Specimen
Suspension
Transmission
Bracket
Fatigue Under Complex Loading: Analysis and Experiments, SAE AE6, 1977
112
Analysis Results
1 10 100
48 Data Points
COV 1.27
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Strain-Life analysis of all test data
Material Variability
1 10 100
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
Material Analysis
Strain-Life back calculation of specimen lives
113
Modeling Uncertainty
( )
=
+ =
n
1 i
a
2
X
1 C 1 C COV
2
i
i
C
U
= 1.09
Analysis Uncertainty C
U
= ?
The variability in reproducing the original strain life data
from the material constants is C
M
~ 0.44
2
M
2
N 2
U
C 1
C 1
C 1
f
+
+
= +
90% of the time the analysis is within a factor of 3 !
99% of the time the analysis is within a factor of 10 !
Variability from Multiple Sources
( )
=
+ =
n
1 i
a
2
X
1 C 1 C COV
2
i
i
Suppose we have 4 variables each with a COV = 0.1
The combined variability is COV = 0.29
Suppose we reduce the variability of one of the variables to 0.05
The combined variability is now COV = 0.27
If all of the COVs are the same, it doesnt do any good to
reduce only one of them, you must reduce all of them !
114
Variability from Multiple Sources
( )
=
+ =
n
1 i
a
2
X
1 C 1 C COV
2
i
i
Suppose we have 3 variables each with a COV = 0.1
and one with COV = 0.4
The combined variability is COV = 0.65
Suppose we reduce the variability of these variables to 0.05
The combined variability is now COV = 0.60
If one of the COVs is large, it doesnt do any good to
reduce the others, you must reduce the largest one !
Variability Summary
Source
COV
Service Loading
0.5
Materials
0.1
Manufacturing
0.1
Surface Finish
0.1
Environment
0.3
Strength
Stress
Fatigue Lives
1.0
1.0
5
Stress
Strength
life Fatigue |
.
|
\
|
115
Variability: Every apple on a tree has a different mass.
Uncertainty: The variety of the apple is unknown.
Variability: Multiple samples of the same material
Uncertainty: What is the material
Strain Life Data for 93 Steels
1 10 10
2
10
3
10
6
10
7
10
5
10
4
10
-4
10
-2
10
-3
0.1
10
1
S
t
r
a
i
n

A
m
p
l
i
t
u
d
e
1 10 10
2
10
3
10
6
10
7
10
5
10
4
10
-4
10
-2
10
-3
0.1
10
1
S
t
r
a
i
n

A
m
p
l
i
t
u
d
e
116
Uncertainty for all Steels
100 1000 10000
55 Data Points
COV 0.48
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
'
f
Uncertainty for Structural Steels
100 1000 10000
COV 0.12
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
'
f
S
u
600 900 MPa
19 Data Points
117
Fatigue Strength Coefficient
Variability Uncertainty Combined
All Steels 0.12 0.48 0.75
Structural Steel 0.12 0.12 0.24
Quiz
What is the expected variability ?
At my last seminar everyone hit a golf ball and we
recorded the maximum acceleration.
118
Results
12.9
7.7
5.88
11.1
15.5
10.3
18.1
11.64
4.26
0.37
COV
Case Studies
119
Introduction
Analysis Methods
Case Studies
GlyphWorks
Case Studies
DARWIN
Southwest Research
Bicycle
Loading Histories
120
A Software Framework for
Probabilistic Fatigue Life Assessment
ASTM Symposium on
Probabilistic Aspects of Life Prediction
Miami Beach, Florida
November 6-7, 2002
R. C. McClung, M. P. Enright, H. R. Millwater
*
,
G. R. Leverant, and S. J. Hudak, Jr.
Southwest Research
Slides 6 27 used with permission of of Craig McClung
Motivation
121
UAL Flight 232
July 19,1989
Turbine Disk Failure
Anomalies in titanium engine disks
Hard Alpha
Very rare
Can cause failure
Not addressed by safe life methods
Enhanced life management process
Requested by FAA
Developed by engine industry
Probabilistic damage tolerance methods
Supplement to safe life approach
SwRI and engine industry developed DARWIN with FAA funding
122
Probabilistic Damage Tolerance
Probabilistic Fracture Mechanics
Probability of Detection
Anomaly Distribution
Finite Element Stress Analysis
Material Crack Growth Data
NDE Inspection Schedule
Pf vs. Cycles
Risk Contribution Factors
Zone-Based Risk Assessment
Define zones based on similar stress,
inspection, anomaly distribution, lifetime
Total probability of fracture for zone:
(probability of having a defect) x (POF given
a defect)
Defect probability determined by anomaly
distribution, zone volume
POF assuming a defect computed with
Monte Carlo sampling or advanced
methods
POF for disk obtained by summing zone
probabilities
As individual zones become smaller (number of
zones increases), risk converges down to
exact answer
1
2 3 4
m
5 6 7
123
Fracture Mechanics Model of Zone
m
7
Retrieve stresses
along line
Fracture Mechanics Model
h
x
hy
x
Y
g
r
a
d
i
e
n
t

d
i
r
e
c
t
i
o
n
1
2
3
4
5
Defect
Finite Element Model
(Not to Scale)
Stress Processing
FE Stresses and plate definition
stress
gradient
Stress gradient extraction
FE Analysis
0.0 0.2 0.4 0.6 0.8 1.0
Normalized distance fromthe notch tip, x/r
-0.8
-0.4
0.0
0.4
0.8
1.2
1.6
2.0
/
o
(z
)
relax
(z
)
residual
(z
)
elastic
Shakedown module
Computed relaxed stress
elastic - residual
0/
Residual stress analysis
3 4 5 6 7 0 1 2 3
Load Step
0
10
20
30
40
50
60
70
80
H
o
o
p

S
t
r
e
s
s

(
k
s
i
)
Rainflow stress pairing
124
Anomaly Distribution
# of anomalies per volume of material as function of defect size
Library of default anomaly distributions for HA (developed by RISC)
Probability of Detection Curves
Define probability of NDE flaw detection as function of flaw size
Can specify different PODs for different zones, schedules
Built-in POD library or user-defined POD
125
Random Inspection Time
Opportunity Inspections during on-condition maintenance
Inspection time modeled with Normal distribution or CDF table
Output: Risk vs. Flight Cycles
126
Output: Risk Contribution Factors
Identify regions of component with highest risk
Implementation in Industry
FAA Advisory Circular 33.14 requests risk
assessment be performed for all new
titanium rotor designs
Designs must pass design target risk for rotors
Components
Risk
Maximum
Allowable
Risk
10
-9
Risk
Reduction
Required
C A B
127
Probabilistic Fracture Mechanics
Material Crack Growth Data
Finite Element Stress Analysis
Anomaly Distribution NDE Inspection
Pf vs. Flights
RPM-Stress Transform Crack Initiation
0
4000
8000
12000
0 1000 2000 3000 4000
Load Spectrum Editing
Code Enhancements
Sensor (RPM) Input
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0 5 10 15 20 25 30 35 40
THRESHOLD STRESS RANGE (KSI)
P
R
O
B
A
B
IL
IT
Y
O
F
F
A
IL
U
R
E
A
T
20,000 C
Y
C
L
E
S
0
20
40
60
80
100
120
N
U
M
B
E
R
O
F
C
Y
C
L
E
S
P
E
R
M
IS
S
IO
N
Pf
Cycles Per Mission
DARWIN for Prognosis Studies
Three Sources of Variability
Anomaly size (initial crack size)
FCG properties (life scatter)
Mission histories (stress scatter)
128
Hard Alpha Defects in Titanium
Initial DARWIN focus on
Hard Alpha
Small brittle zone in
microstructure
Alpha phase stabilized by N
accidentally introduced
during melting
Cracks initiate quickly
Extensive industry effort
to develop HA
distribution
Resulting Anomaly Distributions
Post 1995 Triple Melt/Cold Hearth + Vacuum Arc Remelt
1.00E-02
1.00E-01
1.00E+00
1.00E+01
1.00E+02
1.00E+02 1.00E+03 1.00E+04
DEFECT INSPECTION AREA (sq mils)
E
X
C
E
E
D
E
N
C
E
(
p
e
r

m
i
l
l
i
o
n

p
o
u
n
d
s
)
MZ(5-10in Billet)/#1 FBH
MZ(5-10in Billet)/#3 FBH
MZ (12-13in Billet)/#1 FBH
#2/#1 FBH
#2/#2 FBH
#3/#1 FBH
#3/#2 FBH
#3/#3 FBH
129
FCG Simulations for AGARD Data
Use individual fits to
generate set of a vs.
N curves for identical
conditions
Characterize resulting
scatter in total
propagation life
Lognormal distribution
appropriate in most
cases
N, cycle
0 2000 4000 6000 8000
a
,

c
m
0.0
0.1
0.2
0.3
0.4
0.5
Corner Crack Specimen
K
i
=18.7 MPam, K
f
=56.9 MPam
AGARD
Engine Usage Variability
Stress/Speed:
(RPM)
2
Total Cyclic Life (LCF):
N
f
= N
i
+ N
p
N
i

3-5
N
p
3-4
Life/Speed:
N
f
(RPM)
6
Component life is very sensitive
to actual usage
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
0 2500 5000 7500 10000 12500
TIME (SEC)
R
P
M
(
L
o
w
S
p
e
e
d
S
p
o
o
l)
Air Combat Tactics
Air to Ground & Gunnery
Basic Fighter Maneuvers
Intercept
Peace Keeping
Surface to Air Tactics ( hi alt)
Surface to Air Tactics ( lo alt)
Suppression of Enemy Defenses
Cross Country
130
Usage Variability
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
0 2500 5000 7500 10000 12500
TIME (SEC)
R
P
M

(
L
o
w

S
p
e
e
d

S
p
o
o
l
)
Air Combat Tactics
Air to Ground & Gunnery
Basic Fighter Maneuvers
Intercept
Peace Keeping
Surface to Air Tactics ( hi alt)
Suppression of Enemy Defenses
Cross Country
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
0 2500 5000 7500 10000 12500
TIME (SEC)
R
P
M
(
L
o
w
S
p
e
e
d
S
p
o
o
l)
Peace Keeping
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
0 2500 5000 7500 10000 12500
TIME (SEC)
R
P
M
(L
o
w
S
p
e
e
d
S
p
o
o
l)
Components of Usage Variability:
Mission type
Mission-to-mission variability
Mission mixing variability
0.001
0.01
0.1
1
0 1000 2000 3000 4000 5000 6000 7000 8000
Number of Flight Cycles
N
o
r
m
a
l
i
z
e
d

P
O
F
Air Combat Tactics
Combat Air Patrol
Air-Ground Weapons Delivery
Air-Air Weapons Delivery
Instrument/Ferry
Initiation and Propagation
No Inspection
Variability in Mission Type
Air-Air Weapons Delivery
Time
R
P
M
Air-Ground Weapons Delivery
Time
R
P
M
131
Web Site: www.darwin.swri.org
Fatigue Design and Reliability in the
Automotive Industry
Fatigue Design and Reliability
ESIS Publication 23
J-J. Thomas, G. Perroud, A. Bignonnet and D. Monnet
PSA Peugeot Citron
132
Fatigue Design at PSA
Fatigue design at PSA is done with a probabilistic
approach that includes analysis of customer usage,
production scatter, definition of the appropriate design
loads and an acceptance testing criterion.
Stress-Strength Interference
Stress Strength
How do you really get these distributions ?
Design Load
How do you establish a design load ?
How do you validate the design ?
133
Stress Distribution
Variability in loading has two components,
how it is used and how it is driven.
Car Usage
Highway, city, fully loaded, empty etc.
Driving Style
passive, aggressive etc.
The usage of a car is independent of the owners driving style
so that the distributions of car usage and driving style can
be obtained separately.
Car Usage
Customer surveys
k l c
k
% r
kl
%
1 Unloaded 27
1 Highway 10
2 Good Road 25
3 Mountain 40
4 City 25
2 Half Load 58
1 Highway 5
2 Good Road 30
3 Mountain 30
4 City 35
3 Fully Loaded 15
1 Highway 15
2 Good Road 25
3 Mountain 40
4 City 20
12 Customer Usage
Categories
134
Owner Behavior
Instrumented Vehicles
-3
-2
-1
0
1
2
3
0 -1 -2 -3 3 2 1
To Load, kN
0 -1 -2 -3 3 2 1
To Load, kN
F
r
o
m

L
o
a
d
,
k
N
Extensive field testing for
each customer usage
category produces a large
number of histograms.
Let the usage histogram
be denoted U
kl
Extrapolation
-3
-2
-1
0
1
2
3
0 -1 -2 -3 3 2 1
To Load, kN
0 -1 -2 -3 3 2 1
To Load, kN
F
r
o
m

L
o
a
d
,
k
N
-3
-2
-1
0
1
2
3
0 -1 -2 -3 3 2 1
To Load, kN
F
r
o
m

L
o
a
d
,
k
N
-3
-2
-1
0
1
2
3
0 -1 -2 -3 3 2 1
To Load, kN
F
r
o
m

L
o
a
d
,
k
N
Measured data, e.g. 1,000 km Extrapolated data, 300,000 km
U
kl
U*
kl
135
Virtual Customers
Thousands of virtual customers can now be generated
by combining customer usage with driving style.
| | | |
=
*
kl kl k
i
U r c N U
| |
i
U rainflow histogram for an individual car
N kilometers
c
k
car loading, %
r
kl
car usage, %
| |
*
kl
U
distributions of rainflow histograms for
different car usage classifications
Design Loads
Initial design is done on the basis of a single
constant amplitude load, F
eq
Find a constant amplitude load and number of cycles
that will produce the same fatigue damage as the customer
operating a car for the design life.
136
Equivalent Fatigue Loading
10
1
10
2
10
3
10
4
10
5
10
6
F
eq
m
1
6
m
eq
10
F
F
|
|
.
|
\
|
=

1
N
n
f
=
Distribution of F
eq
F
F
n
F
n
=
F
+
F
Design Customer:
137
Choosing
Contours of
equal probability
y
x
x
Y
-3
x
3
Y
1
2
3
4
Stress
Strength
If we use 3 on both stress and strength
3 P( s < S ) = 2.3 10
-3
= =

5 . 4 10 3 . 5 ) S s s ( P ) failure ( P
6
I
Distribution of Strength
10
6
S
exp
S
138
What Strength is Needed?
F
F
n

S
S
Some Statistics
Z = S - F
Suppose we want a probability of failure of 1 in 50,000
5
f
10 x 2 P

=
( )
2
F
2
S
F S
Z
Z
f
1
1 . 4 P
+

=
= =
F
n
=
S
( 1 COV
S
)
F
n
=
F
( 1 + COV
F
)
139
Mean Component Strength
( )
2
F
2
S
F S
Z
Z
f
1
1 . 4 P
+

=
= =
( )
2
F
F
2
S
n
S
F n
S
f
1
COV 1
COV
COV
F
COV 1
1
F
P
|
|
.
|
\
|
+
+
|
|
.
|
\
|
+
Reliability
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
10
-1
10
-2
10
-3
10
-4
10
-5
10
-6
10
-7
10
-8
10
-9
R
e
l
i
a
b
i
l
i
t
y
n
S
F
= 3
COV
F
= 0.2
COV
S
= 0.1
140
Mean Strength
F

F
F
n

S
S
= 1.4 F
n
for a reliability of 2 x 10
-5
85 . 2
COV
F
1
S
S
n
=
=
Design
Design calculations are made with loads F
n
Large database relates F
n
to vehicle parameters
so that a new vehicle can be designed from
historical measurements
Fatigue calculations are made with material
properties standard deviations from the
mean
141
Component Testing
Test load is frequently higher than F
n
to get
failures near the design life
Component strength not life !
Interpreting the Test Data
Component tests are done with small sample sizes
2
x
1 N , 1
2
2
x
s ) 1 N (

Confidence limits
( )
2
F
F
2
1 N , 1
2
S
n
S
F n
S
f
1
COV 1
COV 1 N
COV
F
COV 1
1
F
P
|
|
.
|
\
|
+
+

|
|
.
|
\
|
+
142
Validation
Full scale vehicle simulation done at the end for
design final validation
Extrapolation
-3
-2
-1
0
1
2
3
0 -1 -2 -3 3 2 1
To Load, kN
0 -1 -2 -3 3 2 1
To Load, kN
F
r
o
m

L
o
a
d
,
k
N
-3
-2
-1
0
1
2
3
0 -1 -2 -3 3 2 1
To Load, kN
F
r
o
m

L
o
a
d
,
k
N
-3
-2
-1
0
1
2
3
0 -1 -2 -3 3 2 1
To Load, kN
F
r
o
m

L
o
a
d
,
k
N
Measured data, e.g. 1,000 km Extrapolated data, 300,000 km
U
kl
U*
kl
How does this process work?
143
Service Loading Spectra
-5
5
0
L
a
t
e
r
a
l

F
o
r
c
e
,
k
N
Time history
0 1 2 3 -3 -2 -1
-3
-2
-1
0
1
2
3
To Load, kN
F
r
o
m

L
o
a
d
,
k
N
0
1
2
3
4
1 10 100 1000
Cumulative Cycles
L
o
a
d

R
a
n
g
e
,
k
N
Rainflow
Histogram
Exceedance
Diagram
Problem Statement
Given a rainflow histogram for a single user,
extrapolate to longer times
Given rainflow histograms for multiple users,
extrapolate to more users
144
Probability Density
-3
-2
-1
0
1
2
3
0 -1 -2 -3
3 2 1
To Load, kN
F
r
o
m

L
o
a
d
,
k
N
Kernel Smoothing
-3
-2
-1
0
1
2
3
F
r
o
m

L
o
a
d
,
k
N
0 -1 -2 -3
3 2 1
To Load, kN
145
Sparse Data
-3
-2
-1
0
1
2
3
F
r
o
m

L
o
a
d
,
k
N
0 -1 -2 -3
3 2 1
To Load, kN
Exceedance Plot of 1 Lap
0
1
2
3
4
1 10 100 1000
Cumulative Cycles
L
o
a
d

R
a
n
g
e
,
k
N
weibull distribution
146
10X Extrapolation
0
1
2
3
4
1 10 100 1000
Cumulative Cycles
L
o
a
d

R
a
n
g
e
,
k
N
6
5
-3
-2
-1
0
1
2
3
F
r
o
m

L
o
a
d
,
k
N
Probability Density
0 -1 -2 -3
3 2 1
To Load, kN
147
Results
-3
-2
-1
0
1
2
3
0 -1 -2 -3 3 2 1
To Load, kN
F
r
o
m

L
o
a
d
,
k
N
-3
-2
-1
0
1
2
3
F
r
o
m

L
o
a
d
,
k
N
Simulation Test Data
0 -1 -2 -3 3 2 1
To Load, kN
Exceedance Diagram
0
1
2
3
4
5
6
1 10 100 10
3
10
4
Cycles
1 Lap (Input Data)
Model Prediction
Actual 10 Laps
L
o
a
d

R
a
n
g
e
,
k
N
148
Problem Statement
Given a rainflow histogram for a single
user, extrapolate to longer times
Given rainflow histograms for multiple
users, extropolate to more users
Extrapolated Data Sets
0.1 1 10 10
2
10
3
10
4
Normalized Life
Airplane
Tractor
ATV
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
149
Issues
In the first problem the number of cycles is
known but the variability is unknown and
must be estimated
In the second problem the variability is known
but the number and location of cycles is
unknown and must be estimated
Assumption
0
16
32
48
1 10 10
2
10
3
10
4
10
5
Cumulative Cycles
L
o
a
d

R
a
n
g
e
0
16
32
48
0
16
32
48
1 10 10
2
10
3
10
4
10
5
On average, more severe users tend
to have more higher amplitude cycles
and fewer low amplitude cycles
150
Translation
-3
-2
-1
0
1
2
3
F
r
o
m

L
o
a
d
,
k
N
0 -1 -2 -3
3 2 1
To Load, kN
Damage Regions
-3
-2
-1
0
1
2
3
F
r
o
m

L
o
a
d
,
k
N
0 -1 -2 -3
3 2 1
To Load, kN
151
ATV Data - Most Damaging in 19
-3
-2
-1
0
1
2
3
F
r
o
m

L
o
a
d
,
k
N
-3
-2
-1
0
1
2
3
F
r
o
m

L
o
a
d
,
k
N
0 -1 -2 -3 3 2 1
To Load, kN
0 -1 -2 -3 3 2 1
To Load, kN
ATV Exceedance
0
1
2
3
4
5
6
1 10 100 10
3
10
4
Cycles
L
o
a
d

R
a
n
g
e
,
k
N
Actual Data
Simulation
152
Airplane Data - Most Damaging in 334
-30
-20
-10
0
10
20
30
0 -10 -20 -30 30 20 10
To Stress, ksi
F
r
o
m

S
t
r
e
s
s
,
k
s
i
-30
-20
-10
0
10
20
30
F
r
o
m

S
t
r
e
s
s
,
k
s
i
0 -10 -20 -30 30 20 10
To Stress, ksi
Airplane Exceedance
0
10
20
30
40
1 10 100 1000
Cycles
S
t
r
e
s
s

R
a
n
g
e
,
k
s
i
Simulation
Actual Data
153
Tractor Data - Most Damaging in 54
-1.5
-1.0
-0.5
0
0.5
1.0
1.5
0 -0.5 -1.0 -2.0 1.5 1.0 0.5
To Strain x 10
-3
F
r
o
m
S
t
a
r
i
n
x
1
0
-
3
-1.5
-1.0
-0.5
0
0.5
1.0
1.5
F
r
o
m

S
t
r
a
i
n

x

1
0
-
3
0 -0.5 -1.0 -2.0 1.5 1.0 0.5
To Strain x 10
-3
Tractor Exceedance
0
1.0
2.0
3.0
1 10 100 10
3
10
4
10
5
Cycles
S
t
r
a
i
n

R
a
n
g
e

x

1
0
-
3
Simulation
Actual Data
154
www.FatigueCalculator.com
Introduction
Analysis Methods
Case Studies
GlyphWorks
155
www.FatigueCalculator.com
Probabilistic Fatigue Analysis
156
Modeling Variability
Products: Z = X
1
X
2
X
3
X
4
.. X
n
Central Limit Theorem:
COV
X
0 0.5 1.0 1.5 2.0
0.5
1.0
1.5
l
n
X
lnX
~ COV
X
COV
X
is a good measure of variability
Standard Deviation, lnx
COV
X
1 2 3
68.3% 95.4% 99.7%
0.05
0.1
0.25
0.5
1
1.05
1.10
1.28
1.60
2.30
1.11
1.23
1.66
2.64
5.53
1.16
1.33
2.04
3.92
11.1
99.7% of the data is within a factor of 1.33 of the mean for a COV = 0.1
COV and LogNormal Distributions
157
Strain Life Analysis
Material Properties
158
Stress Concentration
Results
159
Results (continued)
Results (continued)
160
Results (continued)
Ten Simulations
Life COV
6470 0.959
6930 0.898
6710 0.688
6640 0.908
6580 0.869
6470 0.959
7010 0.723
6690 0.908
6170 0.791
6560 0.971
Mean 6623 0.8674
COV 0.038 0.114
161
29 Individual Data Sets
'
f
Median 820
300 1000
COV 0.25
2000
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
b
-0.12 -0.08 -0.04
Mean -0.09
COV 0.25
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
29 Individual Data Sets (continued)
0.1 1 10
Median 0.57
COV 1.15
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
'
f
-0.8 -0.6 -0.4 -0.2

c
Mean -0.62
COV 0.23
99.9 %
99 %
90 %
50 %
10 %
1 %
0.1 %
C
u
m
u
l
a
t
i
v
e

P
r
o
b
a
b
i
l
i
t
y
162
Correlation
0
0.25
0.5
0.75
1.0
0.01 0.1 1 10
'
f
c
0
0.05
0.1
0.15
0.2
10
2
10
3
10
4
'
f
b
= -0.828 = -0.976
Correlated Variables
163
Results (continued)
Results (continued)
164
Uncorrelated Variables
Results (continued)
165
Strain Amplitude
1000 250
Deterministic Analysis
166
Deterministic Analysis (continued)
Deterministic Analysis (continued)
167
Deterministic Analysis Results
Probabilistic Analysis
168
Probabilistic Analysis (continued)
Probabilistic Analysis (continued)
169
Probabilistic Analysis Results
GlyphWorks
2005 Darrell Socie, All Rights Reserved
170
Introduction
Analysis Methods
Case Studies
GlyphWorks
GlyphWorks / Rainflow
171
StatStrain.flo
StatStrain Glyph
172
StatOutput.xls
10000 100000 1000000
99.9 %
99 %
90 %
50 %
10 %
1 %
0 1 %
Life Distribution,
Channel 6
StatOutput.xls (continued)
Probabilistic Sensitivity, Channel 6
Load History Variables
57%
Stress Concentrators
3%
Material Properties
16%
Analysis Variables
24%
173
Channel # 6 Total Channels 1
Variable Probabilistic Deterministic
NumCases 11 Load History Variables 0.89 -5.42
Median 45634 ScaleFactor 0.89 -5.42
COV 1.91 Offset 0.00 0.00
Stress Concentrators 0.05 -5.51
Probability (%) Life Kf 0.05 -5.51
99 817179 Material Properties 0.25 -14.54
90 223648 E 0.00 -4.37
50 45634 Sfp 0.23 3.42
10 9311 b 0.00 -4.39
1 2548 efp 0.00 1.42
c 0.00 -9.86
np 0.00 -2.32
Kp 0.10 1.56
Analysis Variables 0.37 1.00
Uncertainty 0.37 1.00
Life Distribution Sensitivity
Variable Distribution Value
Scale
Parameter Value
Scale
Parameter
ScaleFactor Normal 100 0.25 103 0.18
Offset Normal 0 0.10 -0.01 0.08
Kf Uniform 3 0.05 3.0 0.04
E None 208000 0.00 208000 0.00
Sfp Log-Normal 1000 0.10 1054 0.11
b None -0.1 0.00 -0.10 0.00
efp None 1 0.20 1.0 0.00
c None -0.5 0.00 -0.50 0.00
np None 0.2 0.00 0.20 0.00
Kp Log-Normal 1200 0.10 1194 0.10
Uncertainty Log-Normal 1 0.50 0.93 0.48
Inputs Outputs
174
StatStress.flo
Rainflow Extrapolation
175
Duration Extrapolation
Results Files
176
Rainflow Duration
Percentile Extrapolation
177
Results

Probabilistic Fatigue

Diunggah oleh

Informasi Dokumen

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Probabilistic Fatigue

Diunggah oleh

Hak Cipta:

Format Tersedia

1

Repeat many times

Probability is related to the minimum distance point

Change in response mean with respect to a

Standard deviation sensitivity

-0.8 -0.6 -0.4 -0.2

378 Data Points

-0.8 -0.6 -0.4 -0.2

Anda mungkin juga menyukai