a u.s. army center of excellence for modeling and...
TRANSCRIPT
REC 2014; Chicago, IL 1
Recent Advances in Reliability Estimation of Time-Dependent Problems Using the Concept of
Composite Limit State
Automotive Research CenterA U.S. Army Center of Excellence for Modeling and Simulation of Ground Vehicles
Zissimos P. Mourelatos
Monica Majcher
Vijitashwa Pandey
Mechanical Engineering Department
Oakland University
REC 2014; Chicago, IL 2
Background information
Definition of time-dependent Pf
Out-crossing rate approach
Proposed approach to estimate time-dependent Pf using
Composite Limit State (CLS)
Identification of CLS
Calculation of time-dependent Pf
Implementation points
Summary and future work
Overview
REC 2014; Chicago, IL 3
SystemInput Output
Uncertainty
(Quantified)Uncertainty
(Calculated)
Propagation
Design
• Random Variables (Time-Independent)
• Random Processes (Time-Dependent)
Challenges:
• Quantification of a Random Processes
• Estimation of time-dependent reliability
Background
REC 2014; Chicago, IL 4
Time-Dependent Probability of Failure
2t Ttn 3t01 t
timet
n
ii
ti
tgPTfP
1
0,,,0 YX
0,,:,0,0 ttgTtPTf
P YX
Series System Reliability Problem
REC 2014; Chicago, IL 5
Time-Dependent Probability of Failure
T
ttif
PTf
P0
dexp011,0
Failure rateOut-crossing rate approach
0),(,0),(,
0lim
ttgttgPt
YXYX
tt
Up-crossing rate
Accurate ONLY if up-crossings are
statistically independent
REC 2014; Chicago, IL 6
K-L expansion is used to represent random process Y(t) as :
: mean of random process Y(t)
: eigenvalue and eigenvector of covariance matrix of Y(t)
: standard normal random variable
tY
tΦ,
Z
Y(t) is assumed Gaussian
p
i
iTiiY ZtttY
1
Φ
Characterization of Input Random Process
REC 2014; Chicago, IL 7
The Composite Limit State defines a convex
domain representing the intersection of safe
regions of all instantaneous limit states
Concept of Composite Limit State is used
Calculation of Time-Dependent
Probability of Failure (Pf)
0,,:,0,0 ttgTtPTf
P YX
REC 2014; Chicago, IL 8
Identification of Composite Limit State
z1
z2
Composite limit state
*iz
*kz
cosik
Correlation
Coefficient
Two-step approach to identify composite limit state
REC 2014; Chicago, IL 9
Identification of Composite Limit State
z1
z2
Composite limit state
*iz
*kz
Step 1: Delete highly correlated
instantaneous limit states (almost parallel) 99.0ij
REC 2014; Chicago, IL 10
Identification of Composite Limit State
z1
z2
Composite limit state
*iz
*kz
Step 2:
Delete instantaneous limit
states that are not part of
the composite
Check by solving a series of LPs
jii
itg
jtg
,1
0,,0,: ZZZIf set is null,
the jth limit state is deleted
REC 2014; Chicago, IL 11
Calculation of Time-Dependent Pf
z1
z2
Composite limit state
*iz
*kz
dzztt
ttP
ij
ji
jif
ij
0
;,
i
Bivariate standard normal vector
REC 2014; Chicago, IL 12
Calculation of Time-Dependent Pf
Our approach calculates Pf exactly as
by eliminating ALL other terms using the convex
polyhedron of the safe domain
This is a substantial contribution in
both Time-Dependent Reliability and
System Reliability
REC 2014; Chicago, IL 13
Calculation of Time-Dependent Pf
G
F
H
I
z1
z2
g3
Composite limit state
g6 g4 g
5
B
B A
C
D
E
z2*
z3*
z1
*
0,,*** kjikkjj zzz
*2zIf is a positive linear
combination of and ,
( )
can be eliminated
enlarging the safe domain
(SD) so that :
*1z
2g
*3z
fABGGCDEFGABCDEFA PPPFP 11
Original SDEnlarged
SD
ffffABG PPPgggPP 1323122321 0,0,0
REC 2014; Chicago, IL 14
Calculation of Time-Dependent Pf
G
F
H
I
z1
z2
g3
Composite limit state
g6 g4 g
5
B
B A
C
D
E
z2*
z3*
z1
*
f
FHIf
DHEf
ABGGCIf
DHEf
ABG
GCHFGf
ABGGCDEFGABCDEFA
PPPPPP
PPPPFP
1
111
fffGCI PPPP 1341341431
Finally:
This is an EXACT
calculation of Pf
involving the convex
polyhedron of the SD
REC 2014; Chicago, IL 15
Example : Hydrokinetic Turbine Blade under Time-dependent
River Flow Loading*
EI
tCtv
EI
tMtg m
allow
flap
allow2
1
21
: Gaussian random process with
autocorrelation coefficient function
tv
1221 2cos, ttttv
*Hu, Z. and Du, X., (2012), “Time-dependent Reliability Analysis by a Sampling Approach
to Extreme Values of Stochastic Processes,” Proceedings of the ASME 2012 IDETC/CIE
is calculated from 0 to 12 monthsFP
Illustration of Composite Limit State
: Allowable strain
: Random variables
: ConstantsECm ,,
allowItallow ,, 1
REC 2014; Chicago, IL 16
t = 8.2 months
Composite Limit State
Instantaneous Limit State
Safe
Region
Illustration of Composite Limit State
REC 2014; Chicago, IL 17
Time, months
Rel
iab
ilit
y I
nd
ex t
t = 8.2 months
Illustration of Composite Limit State
0.2 month discretization
REC 2014; Chicago, IL 18
t = 5.2 months
Composite Limit State
Instantaneous Limit State
Safe
Region
Illustration of Composite Limit State
REC 2014; Chicago, IL 19
Time, months
Rel
iab
ilit
y I
nd
ex t
t = 5.2 months
Illustration of Composite Limit State
0.2 month discretization
REC 2014; Chicago, IL 20
Illustration of Composite Limit State
0.1 month discretization
-6 -4 -2 0 2 4 6-6
-4
-2
0
2
4
6
z1
z2
REC 2014; Chicago, IL 21
Illustration of Composite Limit State
0.05 month discretization
-6 -4 -2 0 2 4 6-6
-4
-2
0
2
4
6
z1
z2
REC 2014; Chicago, IL 22
Illustration of Composite Limit State
0.01 month discretization
-6 -4 -2 0 2 4 6-6
-4
-2
0
2
4
6
z1
z2
REC 2014; Chicago, IL 23
Probability of Failure Calculation
REC 2014; Chicago, IL 24
Implementation Points
Observations:
• Proposed approach requires a time-
independent analysis (beta and MPP) at ALL
time steps.
• Only low-beta limit states contribute to
composite limit state.
To avoid calculating beta and MPP at ALL time
steps:
• Build a surrogate of beta curve using Kriging
• Build composite limit state (CLS) progressively
starting with times where beta is low.
• Stop if CLS does not change further.
REC 2014; Chicago, IL 25
Reliability Index Estimation using Kriging
REC 2014; Chicago, IL 26
Reliability Index Estimation using Kriging
REC 2014; Chicago, IL 27
Reliability Index Estimation using Kriging
REC 2014; Chicago, IL 28
Reliability Index Estimation using Kriging
REC 2014; Chicago, IL 29
Reliability Index Estimation using Kriging
REC 2014; Chicago, IL 30
Reliability Index Estimation using Kriging
REC 2014; Chicago, IL 31
Reliability Index Estimation using Kriging
REC 2014; Chicago, IL 32
Reliability Index Estimation using Kriging
REC 2014; Chicago, IL 33
Reliability Index Estimation using Kriging
Only 14 Evaluations
REC 2014; Chicago, IL 34
β(t
)
Time t
5Δt 10Δt
15Δt
20Δt
Progressive Estimation of Composite
REC 2014; Chicago, IL 35
Total Probability Theorem
Proposed approach is based on FORM.
FORM’s accuracy deteriorates if
• Limit state is nonlinear
• Random variables are non-normal and / or
correlated.
The Total Probability Theorem can increase
accuracy
wwWW
dfFPFPTfP
,0
Mean value of conditional probability
REC 2014; Chicago, IL 36
Total Probability Theorem
032,, 15
2
2124
23
221
31 tYXtYXtXtXXXXttg YX
tYtY 21 , : Gaussian Processes
542 ,, XXX : Non-normal R.V.s
Example :
tYxtyxtxtXxxxtg 15
2
2124
23
221
31 32,
tYXXXX 25421W W
FPIf , then is calculated using
Realization of X1Normal R.V. Function of
Normal R.V.s
Linear function of Normal R.V.s
REC 2014; Chicago, IL 37
Key Points of Composite Limit State
Method
Efficient estimation of reliability index
Kriging metamodel using prediction variance estimation
Identification of Composite Limit State using
Convex domain formed by instantaneous limit states with
“smallest” betas
Calculation of using Composite Limit State
Exact calculation using the convex polyhedron of the safe
domain
fP
t
t
REC 2014; Chicago, IL 38
Developed a new time-dependent reliability
estimation method using the concept of Composite
Limit State (CLS). It
Identifies the CLS automatically
Calculates time-dependent probability of
failure exactly for convex linear safe sets
Summary and Future Work
Improve computational efficiency (e.g. use Kriging
to estimate MPP locus)
Apply method to estimating remaining life due to
fatigue failure
Future Work
REC 2014; Chicago, IL 39
Thank you for your
attention
REC 2014; Chicago, IL 40
A novel MC-based method to calculate the
time-dependent reliability (cumulative
probability of failure) based on :
short-duration data and an
exponential extrapolation using MCS
or Importance Sampling (Infant
Mortality)
Poisson’s assumption (Useful Life)
Our Approach
REC 2014; Chicago, IL 41
btet 0)(ˆ
00
ˆ1
tdt
db
MCS
Exponential
Extrapolation
0
Poisson’s Assumptiontint
Efficient MC Simulation Approach
],[,))(1(1
],0[,1)(
int
)(
int
int
ˆ
int
0
f
ttc
T
dtt
c
T
tttetF
ttetF
m
t
REC 2014; Chicago, IL 42
Objectives and Scientific Contributions
• Develop a methodology to estimate reliability and remaining life of a vehicle system using time-dependent reliability / durability principles.
• Use the methodology to improve existing accelerated Life Testing (ALT) methods by
– Shortening testing time, and
– Using realistic testing conditions
• Implement all developments in TARDEC’s Physical Simulation Lab
Objectives for Project 5.3
• Developed advanced statistical methods to calibrate a math model using a limited number of tests.
• Developed a novel time-dependent reliability method using the concept of composite limit state. The method also advances state-of-the-art in system reliability
• Developed a new paradigm for Accelerated Life Testing
Fundamental Scientific Contributions
REC 2014; Chicago, IL 43
Relates reliability measured under high
stress conditions to expected reliability
under normal conditions
Advantage:
Shortens testing time
Disadvantage:
Uses unrealistic testing conditions
Our Goal: Shorten
testing time and use
realistic testing
conditions
Accelerated Life Testing (ALT)
REC 2014; Chicago, IL 44
Vehicle speed : 20 mph; Mission distance : 100 miles
Simulation can be practically performed for a
short-duration time
0 200 400 600 800 1000-3
-2
-1
0
1
2
3
4
Roa
d H
eigh
t, in
Longitudinal Distance, ft
20 40 60 80 100
-2
-1
0
1
2
3
Time
Ver
tica
l A
ccel
erat
ion
in
G, S
(d,X
,t)
Durability/Performance
Measures in Time
Terrain, Engine
Load, etc.
Random Variables
Problem Description
REC 2014; Chicago, IL 45
0 200 400 600 800 1000-3
-2
-1
0
1
2
3
4
Roa
d H
eigh
t, in
Longitudinal Distance, ft
20 40 60 80 100
-2
-1
0
1
2
3
Time
Ver
tica
l A
ccel
erat
ion
in
G, S
(d,X
,t)
Durability/Performance
Measures in Time
Terrain, Engine
Load, etc.
Random Variables
Model input random processes (terrain, engine load)
Develop detailed, and yet simple and accurate vehicle
math models
Run math models for long time
Major Challenges:
Observations / Challenges
REC 2014; Chicago, IL 46
Model input random processes (terrain, engine load)
Time series and spectral decomposition methods
Develop detailed, and yet simple and accurate vehicle math
models
Use available math models
Calibrate math models using tests to improve their
accuracy
(Model V&V approach)
Run math models for long time
Run calibrated math models for a short duration
Our Approach to Address Challenges
REC 2014; Chicago, IL 47
Random process
characterized by
time series
Available math model
calibrated using tests
years
tC
Degraded vehicle
parameter (e.g. or )sk sb
t years
Response
years
Proposed Approach for ALT
REC 2014; Chicago, IL 48 t years
Response
months
Reliability at time t is the probability that the system
has not failed before time t.
Calculate reliability through time
Must calculate
time-dependent
probability of
failure
Main Task in Proposed ALT
REC 2014; Chicago, IL 497/23/2014
Definitions / Observations
0 Lt time
L
c
T tF
Time-Variant
Reliability0 Lt time
L
i
T tF
Time-Invariant Reliability
Reliability: Ability of a system to carry out a function in a time period
[0, tL]
Cumulative Prob.
of Failure 0,,,0 ttgthatsuchttPtF LL
c
T X
Instantaneous Prob. of Failure 0, LLL
i
T ttgPtF X
L
c
TL
c
f tFttPp Prob. of Time to Failure
REC 2014; Chicago, IL 507/23/2014
Definitions / Observations
0 Lt time
L
c
T tF
Time-Variant Reliability
0 Lt time
L
i
T tF
Time-Invariant Reliability
Reliability: Ability of a system to carry out a function in a time period [0, tL]
Cumulative
Prob. of Failure 0,,,0 ttgthatsuchttPtF LL
c
T X
Instantaneous Prob. of Failure 0, LLL
i
T ttgPtF X
L
c
TL
c
f tFttPp Prob. of Time to Failure
REC 2014; Chicago, IL 517/23/2014
Calculation of Cumulative
Probability of Failure
1t Ft2Kt 1Kt2t0
timet
tKtF
• State-of-the-art Approaches
PHI2 method (Andrieu-Renaud, et al., RESS, 2004)
Set-Based approach (Son and Savage, Quality & Rel.
Engin., 2007)
• State-of-the-art approaches are in general, inaccurate due to:
Choice of
Not including contribution of all discrete times
t
REC 2014; Chicago, IL 52
What is Reliability?Cumulative Probability of Failure
Cumulative
Prob. of Failure 0,,,0 ttgthatsuchttPtF LL
c
T X
Instantaneous Prob. of Failure 0, LLL
i
T ttgPtF X
Reliability at time t is the probability that the system
has not failed before time t.
Maximum Response Method
Niching GA & Lazy Learning Local Metamodeling
MCS / Importance sampling
Analytical
Simulation-based
Calculation Methods for tF c
T
t
c
T dtttF0
]exp[1
REC 2014; Chicago, IL 53
HMMWV Control arm / Spring
STM Table
Control
Arm
Actuation
point for
road input
Output : Stress or
Strain at different
locations on control
arm
Test Fixture at TARDEC
REC 2014; Chicago, IL 54MotionView Math Model Physical Model
HMMWV Lower Control Arm Fixture
REC 2014; Chicago, IL 55
Characterization of random processes
Calculation of time-dependent
probability of failure (or reliability)
Model validation through calibration
Time series and Spectral
Decomposition
Composite Limit State
Approach
Companion
project
Prediction bias Zero mean random error
,,,,, xxcxx mt yy
Key Points of Proposed Approach