6th SIAA Structures, Dynamics and Materials Confere Austin, TX, April 18-21, 2005 Application of the Generalized Conditional Expectation Method for Enhancing a Probabilistic Design Fatigue Code Faiyazmehadi Momin, Harry Millwater, R. Wes Osborn Department of Mechanical Engineering University of Texas at San Antonio Michael P. Enright Southwest Research Institute
43
Embed
46th SIAA Structures, Dynamics and Materials Conference Austin, TX, April 18-21, 2005 Application of the Generalized Conditional Expectation Method for.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
46th SIAA Structures, Dynamics and Materials ConferenceAustin, TX, April 18-21, 2005
Application of the Generalized Conditional Expectation Method for Enhancing a Probabilistic Design Fatigue Code
Faiyazmehadi Momin, Harry Millwater, R. Wes Osborn
Department of Mechanical EngineeringUniversity of Texas at San Antonio
Michael P. EnrightSouthwest Research Institute
University of Texas at San Antonio
Motivation
Probabilistic design codes are specialized for particular application Highly optimized for particular application
Specific mechanics modelSpecific random variablesSpecific probabilistic methods
Prominent codes in industry include PROF, DARWIN, VISA
Codes may need to be enhanced Add more random variables Source code may not be available
University of Texas at San Antonio
Objective
Present the methodology of GCE to enhance a probabilistic design code by considering additional random variables
Compute the sensitivities of the probability-of-failure to ALL random variables
Demonstrate the methodology using a probabilistic fatigue code DARWIN
University of Texas at San Antonio
Approach
Basic idea - discrete distribution
40% 60%
Speed()
CPOF1
CPOF2POFTOTAL = CPOF1 * 0.4 +
CPOF2 * 0.6 = E[CPOFi]
1 2CPOF - Conditional
Probability of Failure
University of Texas at San Antonio
Approach
Generalized Conditional Expectation
€
POFTotal = CPOF(xInternal | xExternal ) fX (External )
dx∫
= E[CPOF(xInternal )]
≅CPOF(xInternal )∑
N
Multiple runs of the probabilistic design
code needed to compute expected
value
University of Texas at San Antonio
Methodology
Generalized Conditional Expectation (GCE) methodology is implemented without modifying the source code Random variables are partitioned as “internal” and
“external” variablesInternal - random variables already considered in
probabilistic design code (called control variables in GCE vernacular)
External - additional random variables to be considered (called conditional variables in GCE vernacular)
University of Texas at San Antonio
Variance Reduction Approach
Traditional use for GCE is variance reduction with sampling methods - reduce the sampling variance by eliminating the variance due to the control variables
Ayyub, B. M., Haldar, A., “Practical Structural Reliability Techniques,” Journal of Structural Engineering, Vol. 110, No. 8, August 1984, pp. 1707-1725.
Ayyub, B. M., Chia, “ Generalized Conditional Expectation for Structural Reliability Assessment,” Structural Safety, Vol.11, 1992, pp. 131-146
University of Texas at San Antonio
Generalized Conditional Expectation
The conditional expected value can be approximated by
Pfi is the conditional POF of ith realization is conditional variables
The variance and coefficient of variation are given by
∑=
≈N
ifif P
NP
1
1
)1(
)()var( 1
2
−
−=∑=
NN
PPP
N
i
ffi
ff
ff
P
PPCOV
)var()( =
University of Texas at San Antonio
Implementation
1. Partition random variables into two categories: • Internal - random variables within the probabilistic design code• External - additional random variables to be considered not
within the probabilistic design code
2. Generate a realization of external variables using Monte Carlo sampling
3. Determine the conditional probability-of-fracture (CPOFi) given this realization of external random variables
4. Compute the expected value (average) of the CPOF results using MC sampling
5. Compute the sensitivities of the POF to the parameters of the internal and external random variables
University of Texas at San Antonio
Implementation - Response Surface Option
1. Partition random variables into two categories: • Internal - random variables within the probabilistic design code• External - additional random variables to be considered not
within the probabilistic design code
2. Generate a realization of external variables using response surface designs
3. Determine the conditional probability-of-fracture (CPOFi) given this realization of external random variables
4. Build a response surface relating the external variables to the CPOF results.
5. Compute the expected value (average) of the CPOF results using MC sampling of the Response Surface
6. Compute the sensitivities of the POF to the parameters of the internal and external random variables
University of Texas at San Antonio
Response Surface Option
Build a response surface representing the relationship between the conditional POF and the external random variables
Use classical design of experiments and goodness of fit tests
€
CPOF( ˜ X 1, ˜ X 2 ) = A0 + A1˜ X 1 + A2
˜ X 2 + A3˜ X 1
2 + A4˜ X 2
2 + A5˜ X 1 ˜ X 2
The response surface is not use to approximate the limit state
University of Texas at San Antonio
Sensitivities
Methodology developed to compute the sensitivities of the POF to the parameters of the internal and external random variables Compare the effects of internal and external variables
€
dP fd ˜ θ i
= E Pf i( ˜ X , ˆ X )
∂fX j ( ˜ x )
∂ ˜ θ j
1
fX j ( ˜ x )
⎡
⎣ ⎢ ⎢
⎤
⎦ ⎥ ⎥
dP fd ˆ θ i
= E∂Pfi( ˜ X , ˆ X )
∂ ˆ θ j
⎡
⎣ ⎢ ⎢
⎤
⎦ ⎥ ⎥
Internal
External
No additional limit state analyses needed
University of Texas at San Antonio
DARWIN®
University of Texas at San Antonio
Implementation with DARWIN
1. Partition random variables into two sets • Internal - DARWIN variables (crack size, life scatter, stress
scatter)• External - non-DARWIN (geometry, loading, structural and
thermal material properties, etc.)2. Generate a realization of external variables using Monte Carlo
sampling3. Run the finite element solver to obtain updated stresses 4. Execute DARWIN given this realization of external random
variables and associated stresses to determine CPOFi 5. Compute the expected value (average) of the DARWIN CPOF
results using Monte Carlo sampling6. Compute the sensitivities of the POF to the parameters of the
internal and external random variables
University of Texas at San Antonio
Implementation - Response Surface Option
1. Partition random variables into two sets • Internal - DARWIN variables (crack size, life scatter, stress
scatter)• External - non-DARWIN (geometry, loading, structural and
thermal material properties, etc.)2. Generate a realization of external variables using response
surface design points3. Run the finite element solver to obtain updated stresses 4. Execute DARWIN given this realization of external random
variables and associated stresses to determine CPOFi 5. Build a response surface relating conditional variables to
DARWIN CPOF6. Compute the expected value (average) of the DARWIN CPOF
results using Monte Carlo sampling of the response surface 7. Compute the sensitivities of the POF to the parameters of the
internal and external random variables
University of Texas at San Antonio
FLOW CHARTstart
Parametric Deterministic Model
Enter RV
Results2NEU .UIF/.UOF
DARWINInput file
Darwin ResultsDarwin CPOF
i = K
Expected CPOFSensitivities
No Yes
Results file
Control Software
Finite Element Solver
Generate samples, Build RS,
Compute Expected CPOF,
Sensitivities
Design Point Loop
University of Texas at San Antonio
Implementation
Ansys probabilistic design system used to control analysis
Ansys FE solver used to compute stressesANS2NEU used to extract stresses for DARWINDARWIN used to compute the CPOFText utility used to extract DARWIN CPOF results and
return to AnsysSensitivity equations programmed within AnsysPOF determined by computing the expected value of the
CPOF using Monte Carlo or Monte Carlo with Response Surface
University of Texas at San Antonio
UTSA FLOW CHARTstart
Parametric Deterministic Model
Enter RV
ANS2NEU .UIF .UOF file
DARWINInput file
DARWIN ResultsDARWIN POF
i = K
Expected POFSensitivities
No Yes
Results file
Ansys PDS
Ansys Solver
Design Point Loop
Generate samples, Build RS,
Compute Expected CPOF,
Sensitivities
University of Texas at San Antonio
Application Example
FA Advisory Circular 33.14 test case Internal variables: initial crack size(a)External variables: rotational speed(RPM), external
pressure(Po), inner radius(Ri)Surface crack on inner boreConsider POF (assuming a defect is present) at 20,000
cycles
Solve using GCE with Darwin and AnsysCompare to independent Monte Carlo solution
University of Texas at San Antonio
FEM MODEL
R2
R1L
6800 rpm
t
Speed
Por
x
Surface Crack
Element type - Plane42
1444 elements
University of Texas at San Antonio
Initial Crack Size
aMIN aMAX
Exceedance Curve
University of Texas at San Antonio
Probabilistic Model
Procedure not limited to Normal distributions
No Name TypeParameter1
(mean)
Parameter2
(COV)
2 Pressure Normal 7250 psi 0.1
3 Speed Normal 712.35 rad/sec 0.05
4 Inner radius Normal 11.81 inches 0.02
No. Name Type
Parameter 1
amin (mils2 )
Parameter 2
amax (mils2 )
1Initial crack
sizeExceedance
Curve3.5236 111060.0
Inte
rnal
Ext
ern
al
University of Texas at San Antonio
Independent Benchmark Solution Developed
MethodRandom variables
No. of Samples
POF
Monte Carlo (Benchmark)
Initial crack size (ai)
10,000 0.1702
Monte Carlo
(DARWIN)Initial crack size
(ai)10,000 0.1703
Approximate analytical fatigue algorithm developed for verification Uses standard Monte Carlo sampling (no GCE) of all variables
University of Texas at San Antonio
Probabilistic Results Using GCE Method
MethodRandom variables
No. of Samples
POF
Monte Carlo
(Benchmark)
ai
Omega
Pressure
Radius
1000 0.2040
GCE
(Ansys (MC) and Darwin)
ai
Omega
Pressure
Radius
1000
DARWIN0.2074
GCE
(Ansys (RS) and Darwin)
ai
Omega
Pressure
Radius
15 DARWIN &100,000
RS0.2079
MC-Monte Carlo
RS-Response Surface simulations
University of Texas at San Antonio
Effects of Response Surface Transformations
Transformation No. of SamplesExpected
POF
Monte Carlo
(Comparison)1000 Darwin 0.2040
Linear15 Darwin &
100,0000.2832
Quadratic
with cross-terms15 Darwin &
100,0000.2077
Exponential15 Darwin &
100,0000.2073
Logarithmic15 Darwin &
100,0000.2119
Power15 Darwin &
100,0000.2077
Box-Cox15 Darwin &
100,0000.2077
All quadratic
University of Texas at San Antonio
Goodness of Fit
TransformationError sum of
Squares
(Close to Zero)
Coefficient of Determination (R2)
(Close to one)
Maximum Absolute Residual
(Close to Zero)
Linear 8.634E-2 0.8908 0.14078
Non-linear quadratic
2.450E-3 0.9969 0.01784
Exponential 2.474E-3 0.9910 0.01813
Logarithmic 7.840E-3 0.9393 0.06510
Power 1.014E-4 0.9996 0.00514
Box-Cox 9.196E-5 0.9996 0.00464
University of Texas at San Antonio
Response Surface Implementation
Note: Response Surface is only used to compute the expected value of a function
This is completely different from the traditional use of RS in probabilistic analysis, i.e., to approximate the limit state and estimate an often very small probability
Curse-of-Dimensionality is still present if a quadratic model is used; however, only the external random variables enter the equation
University of Texas at San Antonio
Sensitivity Results (Mean)
Parameter Monte Carlo(Comparison)
GCE
ANSYS-DARWIN
GCE
Ansys-Darwin(Finite difference)
Rotational speed
0.00316 0.00334 0.00351
External pressure
0.00008 0.000074 0.000087
Inner radius 0.1325 0.1156 0.1618
Sensitivity of POF with respect to mean value€
∂Pf
∂μ
Sensitivities have units
University of Texas at San Antonio
Sensitivity Results (Mean)
Parameter Monte Carlo(Comparison)
GCE
ANSYS-DARWIN
GCE
Ansys-Darwin(Finite difference)
Rotational speed
10.8 11.5 12.1
External pressure
2.79 2.59 3.04
Inner radius 7.54 6.58 9.21
Sensitivity of POF with respect to mean value€
∂Pf
∂μ i
*μ i
Pf
University of Texas at San Antonio
Sensitivity Results (Std Dev)
Parameter Monte Carlo(Comparison)
GCE
ANSYS-DARWINGCE
Ansys-Darwin(Finite difference)
Rotational speed
0.123E-2 0.143E-2 0.124E-2
External pressure
0.500E-4 0.609E-4 0.551E-4
Inner radius 0.0287 0.0244 0.0258
Sensitivity of POF with respect to standard deviations€
∂Pf
∂σ
Sensitivities have units
University of Texas at San Antonio
Sensitivity Results (Std Dev)
Parameter Monte Carlo(Comparison)
GCE
ANSYS-DARWINGCE
Ansys-Darwin(Finite difference)
Rotational speed
0.21 0.25 0.21
External pressure
0.17 0.21 0.19
Inner radius 0.03 0.03 0.03
Sensitivity of POF with respect to standard deviations€
∂Pf
∂σ i
*σ i
Pf
University of Texas at San Antonio
Sensitivities of Internal RV
Parameter
Monte Carlo sampling
(1000 Monte Carlo MATLAB runs)
GCE ANSYS-DARWIN (1000 DARWIN runs)
GCE
ANSYS-DARWIN (15 RS and 100,000
MC)
Finite Difference
ANSYS-DARWIN (15 DARWIN & 100,000
MC)
(mils-2) 0.03494 0.03726 0.0371 0.0340
(mils-2) 6.977E-11 7.441E-11 7.398E-011 ***
mina
Pf
∂∂
maxa
Pf
∂∂
*** The sensitivity of CPOF with respect to amax is too small to be computed using finite difference method
Sensitivities have units
University of Texas at San Antonio
Sensitivities of Internal RV
Parameter Monte Carlo sampling
(1000 Monte Carlo MATLAB runs)
GCE ANSYS-DARWIN (1000 DARWIN runs)
GCE
ANSYS-DARWIN (15 RS and 100,000
MC)
Finite Difference
ANSYS-DARWIN (15 DARWIN & 100,000
MC)
(mils-2) 0.59 0.63 0.63 0.58
(mils-2) 4E-5 4E-5 4E-5 ***
mina
Pf
∂∂
maxa
Pf
∂∂
€
∂Pf
∂amin/ max
*amin/ max
Pf
Non-dimensionalized sensitivities are significantly smaller than other random variables
Application Problem
Zone 14
Zone 13
Zone 12
Zone 11
Zone 10
Zone 9 Zone 2
Zone 3
Zone 4
Zone 5
Zone 6
Zone 7
Zone 1Zone 8
University of Texas at San Antonio
POF per Flight results
Method Random Variables No. of SamplesExpected POF per
Flight
Monte Carlo (within DARWIN)
Initial crack size 10,000 / Zone 1.330E-9
GCE (ANSYS MC and DARWIN)
Initial crack size ai
Pressure Po
Rotational Speed Inner Radius ri
1000 DARWIN runs
1.917E-9
GCE ANSYS RS and DARWIN
(Power Transformation)
Initial crack size ai
Pressure Po
Rotational Speed Inner Radius ri
15 DARWIN and 100,000 RS
1.935E-9
University of Texas at San Antonio
Response Surface Transformations
Transformations No. of Samples Mean POF
Monte-Carlo 1000 DARWIN 1.957E-9
None - Linear15 DARWIN & 100,000 MC
2.637E-9
None - Quadratic
with Cross Terms15 DARWIN & 100,000 MC
1.945E-9
Logarithmic15 DARWIN & 100,000 MC
1.932E-9
Square Root15 DARWIN & 100,000 MC
1.930E-9
Power (0.45)15 DARWIN & 100,000 MC
1.935E-9
Box-Cox (0.45)15 DARWIN & 100,000 MC
1.966E-9
All quadratic
University of Texas at San Antonio
Goodness of Fit Measures
TransformationError sum of
Squares
(Close to Zero)
Coefficient of Determination
(Close to Zero)
Maximum Absolute Residual
(Close to Zero)
Linear 3.077E-17 0.7184 3.083E-9
Quadratic
with Cross Terms
4.322E-19 0.9565 6.238E-10
Logarithmic 3.991E-19 0.9598 1.162E-9
Square root 2.090E-20 0.9978 2.006E-10
Power (0.45) 7.716E-20 0.9922 4.354E-10
University of Texas at San Antonio
Sensitivity Results (Mean) -
ParameterGCE
ANSYS-DARWIN (1000 DARWIN runs)
GCE
ANSYS-DARWIN (15 RS and 100,000 MC)
Finite Difference
ANSYS-DARWIN (15 DARWIN & 100,000
MC)
Rotational speed
4.092E-11 4.479E-11 4.727E-11
External pressure
9.974E-13 1.079E-12 1.035E-12
Inner radius
1.609E-09 1.562E-09 1.660E-09
€
∂Pf
∂μ i
Sensitivities have units
University of Texas at San Antonio
Sensitivity Results (Mean)
ParameterGCE
ANSYS-DARWIN (1000 DARWIN runs)
GCE
ANSYS-DARWIN (15 RS and 100,000 MC)
Finite Difference
ANSYS-DARWIN (15 DARWIN & 100,000
MC)
Rotational speed
15.2 16.7 17.6
External pressure
3.77 4.06 3.92
Inner radius
9.91 9.61 10.3
€
∂Pf
∂μ i
*μ i
Pf
University of Texas at San Antonio
Sensitivity Results (Std Dev)
ParameterGCE
ANSYS-DARWIN (1000 DARWIN runs)
GCE
ANSYS-DARWIN (15 RS and 100,000 MC)
Finite DifferenceANSYS-DARWIN (15 DARWIN & 100,000
MC)
Rotational speed
2.930E-11 5.515E-11 3.4665E-11
External pressure
9.659E-13 5.172E-13 3.0758E-13
Inner radius
1.634E-09 1.684E-09 6.740E-10
Sensitivities have units
€
∂Pf
∂σ i
University of Texas at San Antonio
Sensitivity Results (Std Dev) -
ParameterGCE
ANSYS-DARWIN (1000 DARWIN runs)
GCE
ANSYS-DARWIN (15 RS and 100,000 MC)
Finite DifferenceANSYS-DARWIN (15 DARWIN & 100,000
MC)
Rotational speed
0.55 1.03 0.65
External pressure
0.36 0.19 0.12
Inner radius
0.201 0.207 0.083
€
∂Pf
∂σ i
*σ i
Pf
University of Texas at San Antonio
Conclusions
Methodology to consider affects of additional random variables is developed and demonstrated on a probabilistic fatigue analysis
POF from MC sampling and GCE method are in good agreement
Response surface method used to reduce signicantly the computational time with good accuracy
The sensitivities obtained from MC simulations, GCE formulae and finite difference method are in good agreement and indicate the importance of internal and external random variables on the POF
University of Texas at San Antonio
Conclusions
Enables the user to consider additional random variables without modifying the source code
Enables the developer to consider the importance of implementing additional random variables