Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
by
Ikjin Lee
An Abstract
August 2008
The objective of this study is to propose a new method for inverse reliability
analysis and robust design optimization using the univariate dimension reduction method
(DRM). The current research effort involves: (1) reliability-based robust design
optimization (RBRDO) method using the mean-based DRM; (2) reliability-based design
optimization (RBDO) method using the inverse reliability analysis and the most probable
point (MPP)-based DRM; (3) design sensitivity analyses for RBDO using the MPP-based
DRM; and (4) system inverse reliability analysis and RBDO using the MPP-based DRM
In the RBRDO formulation, the product quality loss function is minimized subject
to the probabilistic constraints. Since the quality loss function is expressed in terms of
the first two statistical moments, mean and variance, it is necessary to accurately estimate
developed in this study using the mean-based DRM and compared to existing methods.
This study also proposes an inverse reliability analysis method using MPP-based
probability of failure and an RBDO method using the proposed inverse reliability
analysis. Using the MPP-based DRM, a new MPP is obtained, which estimates the
probability of failure of the performance function more accurately than first order
reliability method (FORM) and more efficiently than second order reliability method
(SORM). The new MPP is then used for RBDO to obtain an accurate optimum design
Finally, the MPP-based DRM is extended to the system inverse reliability analysis
and RBDO. This study proposes to use the MPP-based DRM and Ditlevsen’s second
order upper bound for accurate system probability of failure calculation. For the system
RBDO, two efficiency strategies are proposed to save the computational cost.
Abstract Approved:
Thesis Supervisor
Date
RELIABILITY- BASED DESIGN OPTIMIZATION AND ROBUST DESIGN
by
Ikjin Lee
August 2008
CERTIFICATE OF APPROVAL
PH.D. THESIS
Ikjin Lee
Thesis Committee:
Kyung K. Choi, Thesis Supervisor
Jia Lu
Yong Chen
Shaoping Xiao
Olesya Zhupanska
TABLE OF CONTENTS
CHAPTER
I. INTRODUCTION ............................................................................................1
2.1 Introduction...............................................................................................12
2.2 Reliability Analysis ..................................................................................12
2.2.1 Transformation ...............................................................................13
2.2.2 First Order Reliability Analysis (FORM).......................................15
2.2.3 Second Order Reliability Analysis (SORM) ..................................16
2.2.4 System Reliability Analysis ...........................................................17
2.3 Inverse Reliability Analysis ......................................................................18
2.4 Reliability-Based Design Optimization (RBDO) .....................................21
2.5 Dimension Reduction Method (DRM) .....................................................22
2.5.1 Mean Value-Based Dimension Reduction Method ........................23
2.5.2 MPP-Based Dimension Reduction Method....................................25
2.5.3 Rotated Standard Normal V-Space ................................................25
2.6 Reliability-Based Robust Design Optimization........................................26
III. ROBUST DESIGN OPTIMIZATION (RDO) ...............................................30
3.1 Introduction...............................................................................................30
3.2 Mean Value-Based Dimension Reduction Method ..................................31
3.2.1 Computational Efficiency ...............................................................32
3.2.2 Sensitivity of Statistical Moments ..................................................33
3.3 Performance Moment Integration (PMI) ..................................................35
3.3.1 Derivation of PMI...........................................................................35
3.3.2 Sensitivity of Statistical Moments ..................................................38
3.4 Percentile Difference Method (PDM) ......................................................39
3.5 Comparison ...............................................................................................42
3.6 Numerical Examples .................................................................................43
3.6.1 Comparison of PMI and DRM for Computation of Moments
and Sensitivities .......................................................................................43
3.6.2 Comparison of PMI, DRM and PDM for Identification of
Robust Optimum Design .........................................................................47
ii
IV. A NEW DRM-BASED INVERSE RELIABILITY ANALYSIS ..................51
4.1 Introduction...............................................................................................51
4.2 Error in FORM-Based Reliability Analysis .............................................52
4.3 Inverse Reliability Analysis Using MPP-Based DRM .............................53
4.3.1 Probability of Failure Calculation Using Constraint Shift .............53
4.3.2 Reliability Index Update.................................................................58
4.3.3 MPP Update Method ......................................................................59
4.4 Numerical Examples .................................................................................60
4.4.1 Comparison of FORM, SORM and DRM ......................................60
4.4.2 Inverse Reliability Analysis Using DRM .......................................63
6.1 Introduction...............................................................................................85
6.2 Algorithm of DRM-Based RBDO ............................................................85
6.3 Strategy for Efficiency of DRM-Based RBDO ........................................87
6.4 Numerical Examples for DRM-Based RBDO ..........................................90
6.4.1 Effectiveness of Reduced Rotation Matrix.....................................90
6.4.2 Comparison of Various RBDO Methods .......................................91
6.4.3 RBDO for Side Impact Crashworthiness .......................................95
6.4.4 Tracked Vehicle Roadarm Problem .............................................100
6.5 Numerical Examples for RBRDO ..........................................................106
6.5.1 RBRDO for 2-D Mathematic Example ........................................106
6.5.2 RBRDO for Side Impact Crashworthiness ...................................109
7.1 Introduction.............................................................................................112
7.2 System Inverse Reliability Analysis .......................................................113
7.2.1 Component Probability of Failure Calculation.............................114
7.2.2 Joint Probability of Failure Calculation Using FORM .................115
7.2.3 System Probability of Failure Calculation ...................................118
7.3 System Reliability-Based Design Optimization .....................................119
7.3.1 Formulation of System RBDO .....................................................119
7.3.2 Sensitivity Analyses .....................................................................120
iii
7.3.3 Efficiency Strategies .....................................................................124
7.4 Numerical Examples ...............................................................................126
7.4.1 Accuracy of Sensitivity ................................................................126
7.4.2 Comparison of Critical Constraint Identification Methods ..........127
7.4.3 Comparison of System RBDO Using FORM and MPP-Based
DRM ......................................................................................................130
8.1 Conclusions.............................................................................................134
8.1.1 Reliability-Based Robust Design Optimization (RBRDO) ..........134
8.1.2 DRM-Based Inverse Reliability Analysis and RBDO .................135
8.1.3 Sensitivity Analyses for RBDO using FORM and MPP-
Based DRM ...........................................................................................136
8.1.4 System Inverse Reliability Analysis and RBDO ..........................137
8.2 Future Recommendation .........................................................................138
REFERENCES ................................................................................................................139
iv
LIST OF TABLES
Table
3.2. Sensitivity of Mean Value Using PMI and DRM for Eq. (3.29) ..............................45
3.3. Sensitivity of Variance Using PMI and DRM for Eq. (3.29). ..................................45
3.6. Sensitivity of Variance Using PMI and DRM for Eq. (3.30).. .................................47
3.7. Position and Value of Optimum Using Three Methods for Eq. (3.31) .....................49
4.6. Iterative Way of Finding DRM-Based MPP Using New MPP Search.....................64
v
6.3. Various RBDO Results with Target Probability of Failure PFTar
i
5.0% . ...............93
6.4. Updated Reliability Index at the Optimum ...............................................................94
6.6. Design History for Side Impact Example Using FORM-Based RBDO ...................97
6.7. Constraint History for Side Impact Example Using FORM-Based RBDO..............97
6.9. Design History for Side Impact Example Using DRM-Based RBDO .....................99
6.10. Constraint History for Side Impact Example Using DRM-Based RBDO ................99
6.18. Variance Estimation Using PMI at Initial and Optimum Design ...........................107
6.19. Variance Estimation Using DRM at Initial and Optimum Design .........................107
6.24. Variance Using Mean-Based DRM at Initial and Optimum Design ......................110
vi
7.4. Comparison of FORM and MPP-Based DRM ( PFall 2.275% ) ...........................132
vii
LIST OF FIGURES
Figure
2.2. Difference between Reliability Analysis and Inverse Reliability Analysis ..............19
2.3. Standard Normal U-space and Rotated Standard Normal V-space ..........................26
3.4. Accuracy of PMI and DRM with Five Quadrature Points .......................................50
4.2. Performance Function in X-space and V-space for 2-D problem ............................61
viii
1
CHAPTER I
INTRODUCTION
variables. First, a new statistical moment estimation method, which is essential for robust
design optimization (RDO), using the mean value-based dimension reduction method
(DRM), is proposed. Second, a new stochastic method to solve highly nonlinear and
presented for the inverse reliability analysis and subsequent design optimization.
Sensitivity analyses for RBDO using FORM and MPP-based DRM are also carried out.
In the last, a new system reliability analysis and system RBDO using the inverse
reliability analysis through the MPP-based DRM and Ditlevsen’s second order upper
Section 1.1 presents background and motivation of the proposed research, and
Section 1.2 provides objectives of the proposed research, and Section 1.3 describes the
thesis organization.
probabilistic level, while RDO is a method to improve the product quality by minimizing
variability of the output performance function. Since both design methods make use of
2
uncertainties in design variables and other input parameters, it is very natural for the two
approaches to integrate robust design (Su and Renaud, 1997; Kalsi et al., 2001) and
reliability-based design (Du and Chen, 2001; Youn et al., 2005a; Youn et al., 2005b)
have been proposed (Du et al., 2004; Mourelatos and Liang, 2005; Youn et al., 2005c).
The product quality in robust design can be described by use of the first two
statistical moments of a performance function: mean and variance (Chandra, 2001). Thus,
it is necessary to develop methods that estimate the first two statistical moments of the
performance function and their sensitivities accurately and efficiently. The statistical
using the multi-dimensional integral. Hence, there are various numerical methods
et al., 1989), first order Taylor series expansion (Su and Renaud, 1997; Kalsi et al., 2001;
Buranathiti et al., 2004), Monte Carlo simulation (MCS) (Rubinstein, 1981; Lin et al.,
1997), importance sampling method (Rubinstein, 1981; Schueller and Stix, 1987;
Engelund and Rackwitz, 1993; Denny, 2001), and Latin hypercube sampling method
(Mckay, 1979; Walker, 1986; Stein, 1987; Olsson and Sandberg, 2002).
MCS could be accurate for the statistical moment estimation; however it requires
applications, it is not practical to use MCS due to its expensive computational cost. The
experimental design also needs a large amount of computation when the number of
design variables is large. The first order Taylor series expansion has been widely used to
3
estimate the first and second statistical moments for robust design due to its simplicity
and easiness. However, the first order Taylor series expansion results in a large error
especially when the input random variables have large variations. This is because the first
order Taylor series expansion does not use all information of the probability density
recently proposed: the univariate DRM (Xu and Rahman, 2003; Xu and Rahman, 2004a;
Xu and Rahman, 2004b), performance moment integration (PMI) (Youn et al., 2005c),
and percentile difference method (PDM) (Du et al., 2004; Mourelatos and Liang, 2005).
Hence, a comparison study using three methods for RDO will be needed in terms of how
accurately and efficiently the methods can estimate the statistical moments of the
performance function and thus how accurately these methods can find an optimum design
In recent years, there have been various attempts to develop enhanced reliability
function. The most widely used reliability analysis methods are (1) analytical methods
and (2) simulation or sampling methods. The analytical methods include the MPP-based
method and PDF approximation method. Furthermore, the MPP-based method includes
the First Order Reliability Method (FORM) (Hasofer and Lind, 1974; Palle and Michael,
1982; Madsen et al., 1986; Haldar and Mahadevan, 2000) and the Second Order
4
Reliability Method (SORM) (Breitung, 1984; Hohenbichler and Rackwitz, 1988). FORM
G( X) using the first or second order Taylor series expansion at MPP. Since the FORM
or SORM-based method requires MPP search, the sensitivity analysis is usually used for
the methods. When the sensitivities are not available, the response surface method can be
used for the reliability analysis and design optimization (Jin et al., 2003; Youn and Choi,
2004). The PDF approximation method (Rosenblueth, 1975; Du and Huang, 2006; Youn
distribution type and then, using the approximated PDF, the method evaluates the
such as MCS (Lin et al., 1997; Rubinstein, 1981), importance sampling method
(Rubinstein, 1981; Schueller and Stix, 1987; Engelund and Rackwitz, 1993), and Latin
hypercube sampling method (Mckay, 1979; Walker, 1986; Stein, 1987; Olsson and
Sandberg, 2002), can be readily used for the probability of failure calculation since these
Among these methods, the MPP-based method including FORM and SORM is
still a common approach. However, the reliability analysis using FORM could be very
both. This is because FORM approximates the performance function using a linear
Although the reliability analysis using SORM may be accurate, it is not easy to use since
SORM requires the second-order sensitivities, which are difficult and very expensive to
obtain in practical engineering problems. The accuracy of the response surface method is
5
still challenging, especially for the highly nonlinear and multi-dimensional performance
function, even though the method could be efficient. The simulation or sampling-based
method could be accurate. However they require a very large number of function
evaluations at high computational cost. The PDF approximation method can yield
accurate results only if the output probability distribution type is known or accurately
assumed. Moreover, the method should be combined with the response surface method
for the design optimization (Youn et al., 2006), which may have accuracy problem.
the sum of lower dimensional functions. Because of its wide applicability, DRM has been
used for RDO (Lee et al., 2006; Lee et al., 2007), RBDO (Wei, 2006), and PDF
approximation (Du and Huang, 2006; Youn et al., 2006). For the robust design, the mean-
based DRM is used to calculate the statistical moments of the performance function. This
moment calculation using the mean-based DRM is also used to approximate PDF of the
performance function for reliability analysis (Du and Huang, 2006; Youn et al., 2006).
The reliability analysis using the PDF approximation method and mean-based DRM
generation, which could be limitations of the method. That is, it is inherently difficult to
accurately estimate the tail end of PDF (i.e., reliability) when PDF is approximated using
Thus, a new reliability analysis method using MPP-based DRM is needed with
greater accuracy and/or better efficiency than existing methods. Such a method could be
optimization, which requires the sensitivity of the probabilistic constraint at MPP with
respect to the design variable, which is the mean value of input random variables. Many
works have been conducted to study the sensitivities of the probabilistic constraints
(Haldar and Mahadevan, 2000; Tu and Choi, 1999; Ditlevsen and Madsen, 1996; Hou,
2004; Gumbert et al., 2003; Hohenbichler and Rackwitz, 1986; Rahman and Wei, 2007).
The sensitivity using the Reliability Index Approach (RIA) for the FORM-based RBDO
(Haldar and Mahadevan, 2000; Tu and Choi, 1999; Ditlevsen and Madsen, 1996; Hou,
2004; Gumbert et al., 2003; Hohenbichler and Rackwitz, 1986) and the DRM-based
RBDO (Rahman and Wei, 2007) is derived in detail. However, the rigorous analytical
(PMA) for the inverse reliability analysis for both FORM-based and DRM-based RBDO
has not yet been explained in detail in the literature. Thus, one of the main goals of this
obtained from the inverse reliability analysis using both FORM and MPP-based DRM.
Not only the estimation of the component probability of failure but also the
estimation of the system probability of failure has been the main concern in structural
reliability analysis for over three decades. According to the logical relationship of the
failure modes of structures, structural systems can be divided into three types: series
systems, parallel systems, and hybrid systems (Zhao et al., 2007). In this thesis, the
7
reliability analysis with the series system will be explained since it is most frequently
Since the analytic estimation of the system probability of failure involves multi-
evaluate. Hence, several approaches to resolve the numerical difficulty have been
proposed including the narrow bound estimation (Ditlevsen, 1979) and MCS. As
explained in the previous section, MCS requires very expensive computational cost,
which is not acceptable for practical engineering applications. For the narrow bound
method, Ditlevsen’s first order upper bound, which is the summation of component
failure probabilities, can be used as the system probability of failure (Ba-abbad et al.,
2006) or Ditlevsen’s second order upper bound by considering the joint probability of
failure can be used (Ang and Tang, 1984; Liang et al., 2007). However, these narrow
bound methods will only work for linear or mildly nonlinear performance functions since
they approximate performance functions using the first order Taylor series expansion,
i.e., FORM.
Thus, more accurate system reliability analysis method than the traditional
methods is needed for the system with highly nonlinear and multi-dimensional
performance functions. Furthermore, the system RBDO using the accurate system inverse
The first objective of this study is to propose RBRDO using the mean-based
DRM, which includes the accurate estimation of the statistical moments and their
carried out. For the comparison study, PMI and PDM for RBRDO are introduced and
explained. Both DRM and PMI directly estimate the statistical moments. On the other
hand, in PDM, the robustness is achieved through a design objective in which the
performance difference between the right and left tails of the performance distribution
(Du et al., 2004). Thus, three methods can be compared in terms of how accurately these
methods can find an optimum design to minimize the variance of the performance
function. Hence, in this study, three methods are evaluated by comparing the variances at
the optimum designs. PMI and DRM are also compared in terms of how accurately and
efficiently these methods can estimate the statistical moments of the performance
function. The comparison study through numerical examples illustrates that mean-based
DRM is the most accurate and efficient method when the number of design variables is
small and PMI is a better option when the number of design variables is relatively large.
The second objective of the study is to develop a new inverse reliability analysis
method and subsequent RBDO using the MPP-based DRM, which is known as a more
accurate method than the mean-based DRM for the reliability analysis (Wei, 2006). The
MPP-based DRM is used to accurately calculate the probability of failure after finding
the FORM-based MPP, which is then used to develop an enhanced inverse reliability
analysis method (i.e., MPP search) that is accurate for highly nonlinear and multi-
9
dimensional problems. Since the DRM-based inverse reliability analysis still requires
procedure is proposed in this study to carry out the inverse reliability analysis accurately
and efficiently using the MPP-based DRM: the probability of failure calculation using
constraint shift, reliability index update, and MPP update method. Using the three-step
procedure, a new DRM-based MPP is obtained, which is used for the next design
iteration in RBDO. Since the proposed RBDO using the DRM-based MPP requires more
function evaluations, the enhanced hybrid mean value (HMV+) (Youn et al., 2005a) is
used for the efficient inverse reliability analysis, and the enriched performance measure
approach (PMA+) (Youn et al., 2005b) with efficient methods, which are new tolerances
for constraint activeness and a reduced rotation matrix, are proposed for efficient design
optimization in this study. For RBDO using FORM and MPP-based DRM, it is necessary
to carry out a rigorous sensitivity analyses to obtain the optimal design efficiently.
The last objective of the study is to propose more accurate system reliability
analysis method than the traditional methods for the system with highly nonlinear and
multi-dimensional performance functions and subsequent system RBDO. For the series
second order upper bound is adapted as the system probability of failure. MPP-based
DRM will be used for the accurate estimation of the component probability of failure and
concavity of the performance functions will be used. The result is also compared with the
system reliability analysis using FORM and MCS. For the system RBDO, two efficiency
strategies, which are the Mean Value (MV) method for identification of critical
10
constraints and new design closeness concept, are proposed to save the computational
cost.
Chapter 2 presents fundamental concepts for the reliability analysis, robust design
Chapter 3 presents an RDO method using the mean-based DRM. First, the
statistical moment estimation and its sensitivity calculation using DRM are explained.
Second, the statistical moment estimation and its sensitivity calculation using PMI and
PDM are explained for the comparison purpose. Last, three methods are compared
Chapter 4 proposes new inverse reliability analysis and system reliability analysis
methods using the MPP-based DRM, which consists of three numerical procedures: the
probability of failure calculation using constraint shift, reliability index update, and MPP
update. Comparison study with FORM and SORM is carried out to evaluate how
Chapter 5 carries out rigorous sensitivity analyses for RBDO using FORM and
MPP-based DRM. For RBDO using MPP-based DRM, an approximate sensitivity is also
proposed to save the computational cost. Comparison study between analytic and
Chapter 6 proposes an RBDO and RBRDO method using the MPP-based DRM.
The proposed RBDO and RBRDO involve two efficiency strategies: new tolerances for
MPP-based DRM and Ditlevsen’s second order upper bound. Using the accurate system
inverse reliability analysis, the system RBDO is also proposed. For the system RBDO,
CHAPTER II
2.1 Introduction
Section 2.2 and 2.3 discuss the basic idea of reliability analysis and inverse reliability
analysis which are necessary for RBDO that will be explained in Section 2.4. Section 2.5
illustrates two Dimension Reduction Methods (DRM) based on the reference point used:
the mean value-based DRM and MPP-based DRM. The mean value-based DRM is an
accurate and efficient tool for the statistical moment calculation which is required for
robust design optimization explained in Section 2.6 and Chapter 3. The MPP-based DRM
will be used for more accurate reliability analysis and design optimization in Chapter 4
performance function such that G(X) 0 is defined as failure, and f X (x) is a joint
probability density function (PDF) of the random variable X . In most real engineering
applications, the exact evaluation of Eq. (2.1) is very difficult or often impossible to
13
the non-Gaussian f X (x), a transformation from the original X-space into the independent
order Taylor series expansion in the First Order Reliability Method (FORM) or second
order Taylor series expansion in the Second Order Reliability Method (SORM) if G( X)
is highly nonlinear.
2.2.1 Transformation
u1 1 FX x1
1
u2 FX x2 x1
1
T : 2 (2.2)
u N FX N xN x1 , x2 , , xN 1
1
xi
f X1 X 2 ( x1 , x2 , , xi 1 , )d
FX i xi x1 , x2 , Xi
, xi 1 (2.3)
f X1 X 2 X i1 ( x1 , x2 , , xi 1 )
1 u 1
(u )
2
exp 2 d
2
(2.4)
x1 FX11 u1
1
x2 FX21
(u2 x1 )
T : (2.5)
1
xN FX N (u N x1 , x2 , , xN 1 )
If the N-dimensional random vector X is independent, that is, the joint PDF is given by
f X (x) f X1 ( x1 ) f X 2 ( x2 ) f X N ( xN ) (2.6)
where f X i ( xi ) are the marginal PDFs, then Rosenblatt transformation and the inverse
where FX i ( xi ) are the marginal CDFs. Table 2.1 shows five representative distributions
Table 2.1. Probability Distribution and Its Transformation between X and U-space
In this study, the input random variables are assumed to be independent for the
simplicity of calculation and the study with dependent input random variables will be the
FORM and SORM, it is necessary to find the most probable point (MPP), which is
defined as the point u* on the limit state function ( g (u) 0 ) closest to the origin in the
g (u) G(x(u)) G(x) using the Rosenblatt transformation. Hence, MPP can be found
minimize u
(2.8)
subject to g(u) 0
After finding MPP, the distance from MPP to the origin is commonly called the
Hasofer-Lind reliability index (Hasofer and Lind, 1974) and denoted by HL , that is,
MPP obtained by solving Eq. (2.8) is also used for the probability of failure
U-space and the rotational transformation from U-space to V-space which will be
explained in Section 2.5.3, the probability of failure can be obtained using SORM as
(Breitung, 1984; Hohenbichler and Rackwitz, 1988; Rahman and Wei, 2006)
1
( HL ) 2
PFSORM ( HL ) I N 1 2 A N 1 (2.10)
( HL )
17
A N 1 A1N 1
where A R T HR , H is the Hessian matrix evaluated at the MPP
A N 1 A NN 2 gU
in X-space, R is the rotation matrix such that u Rv , and () is the PDF of a standard
When there are more than one performance function, the system probability of
failure is defined as
m
PFsys P Gi ( X) 0 (2.11)
i 1
as failure if Gi (X) 0 . However, since the right hand side of Eq. (2.11) is not easy to
using Ditlevsen’s first-order upper bound (Ditlevsen, 1979) by the sum of the probability
of failures as
m
PFsys PFi (2.12)
i 1
where PFi is the probability of failure for ith performance function or using Ditlevsen’s
m m
PFsys PFi max( PFij ) (2.13)
j i
i 1 i 2
where PFij is the joint probability of failure when the ith and jth failure modes occur
simultaneously.
18
The reliability analysis presented in Section 2.2 is called the reliability index
approach (RIA) (Tu and Choi, 1999) since it finds the reliability index HL using Eq.
(2.8). The advantage of RIA is that the probability of failure for the performance function
can be calculated at a given design, for example, using Eqs. (2.9) and (2.10).
However, it is well known that the inverse reliability analysis in the performance
measure approach (PMA) (Tu and Choi, 1999; Tu et al., 2001; Choi et al., 2001; Youn et
al., 2003) is much more robust and efficient than the reliability analysis in RIA. PMA
does not calculate the probability of failure directly. Instead, PMA judges whether or not
a given design satisfies the probabilistic constraint for a given target probability of failure
PFTar . Using FORM, the target reliability index t can be calculated as t 1 ( PFTar )
using Eq. (2.9), and then the feasibility of the given design can be checked by solving the
maximize g(u)
(2.14)
subject to u t
Since Eq. (2.14) is the inverse problem of Eq. (2.8), this is called the inverse reliability
analysis. The optimum point of Eq. (2.14) is also called the MPP and denoted by u* . If
the constraint function value at the MPP, g (u* ), is less than zero ( G(X) 0 is defined as
safe), then the probabilistic constraint is satisfied for the given target reliability t and
target probability of failure. The inverse reliability analysis using SORM is much more
difficult and has not been developed yet. Moreover, it requires the second order
sensitivity. We can compare the difference between the reliability analysis and inverse
To find the MPP using the inverse reliability analysis with the given target
reliability index t , some methods have been developed including the mean value (MV)
method, advanced mean value (AMV) method (Wu et al., 1990; Wu, 1994), hybrid mean
value (HMV) method (Youn et al., 2003), and enhanced hybrid mean value (HMV+)
Figure 2.2. Difference between Reliability Analysis and Inverse Reliability Analysis
N
g
g (U) g (μ U ) (U i Ui ) (2.15)
i 1 U i U=μ U
where α(μ U ) is the normalized gradient vector evaluated at the mean value and written
as
U g (μ U )
α(μ U ) (2.17)
U g (μ U )
T
where U { , , } . This MV method is a crude method to find MPP of the
U1 U N
However, since it does not require further function evaluation and sensitivity
active or not when a constraint function is far from the design point.
The MPP obtained by the MV method can be considered as the first iteration of
the AMV method. AMV uses the gradient at MPP obtained by the MV method to find the
next MPP candidate and the iteration will continue to perform until the approximate MPP
converges to the correct MPP. Hence, the AMV method can be formulated as
This AMV method is known as an efficient method when the constraint function is
reliability analysis underestimates the probability of failure and vice versa for concave.
For example, the constraint function in Figure. 2.2 is concave around the MPP since
To resolve the weakness of AMV for a concave function, the HMV method uses
AMV method when a constraint function is convex and the conjugate mean value (CMV)
21
(Youn et al., 2003) method when a constraint function is concave. HMV+ method uses an
interpolation between two previous MPP candidate points if the constraint function is
minimize Cost(d)
subject to P[Gi ( X) 0] PFTar
i
, i 1, , nc (2.19)
d L d dU , d R ndv and X R nrv
where d {di }T μ(X) is the design vector; X { X i }T is the random vector; and nc ,
ndv and nrv are the number of probabilistic constraints, design variables, and random
variables, respectively. Using the inverse reliability analysis, the ith probabilistic
minimize Cost(d)
subject to P[Gi ( X) 0] PFTar
i
( ti ), i 1, , nc (2.21)
d L d dU , d R ndv and X R nrv
where ti is the target reliability index for the ith constraint and the probabilistic
where x*FORM is the FORM-based MPP which can be obtained by solving Eq. (2.14) and
transformation T (u* ) x* in Eq. (2.7). For the simplicity, x* means the FORM-based
MPP hereafter.
The sensitivity of the probabilistic constraint with respect to the design parameter is
G(x* ) G G xi x G
N T
(2.23)
d d x x* i 1 xi x x* d xx
* d x x* x x x*
and Eq. (2.23) can be further simplified as (Gumbert et al., 2003; Hou, 2004)
G(x* ) x G G
T
(2.24)
d d xx* x x x* x x x*
The dimension reduction method (Xu and Rahman, 2004a; Xu and Rahman,
of a performance function using a function with reduced dimension. There are several
this study, the univariate DRM is used for approximation of the performance function due
to its simplicity and efficiency. Computational efficiency of the univariate DRM will be
discussed in Chapter 3. The univariate DRM can be categorized into two different DRMs
N
hk ( X) h ( X) hk ( 1 ,
k
, i 1 , xi , i 1 , , N ) ( N 1)h k ( 1 , , N ) (2.25)
i 1
where i is the mean value of a random variable X i and N is the number of design
The mean value-based univariate DRM can be used to approximate the multi-
E ({h( X)}k ) {h( X)}k f X (x)dx {hˆ( X)}k f X (x)dx (2.27)
Then, using the mean value-based univariate DRM, one N-dimensional integration in Eq.
large. This reduction of the number of function evaluations will be explained in Chapter
3.
based integration rule (MBIR) (Xu and Rahman, 2003), which is similar to Gaussian
quadrature (Atkinson, 1989). According to MBIR, the kth statistical moment of a one-
n
E ({h( X )}k ) wi h k ( xi ) (2.28)
i 1
where wi are weights, xi are quadrature points and n is the number of weights and
quadrature points. If PDF of the design variable is given, then these weights wi and
quadrature points xi can be obtained using MBIR. For the standard normal input random
variable, the weights and quadrature points are shown in Table 2.2 (Atkinson, 1989).
This mean value-based univariate DRM for the statistical moment calculation is
used for robust design optimization, which will be explained in detail in Section 2.6 and
Chapter 3.
additively decomposed into one-dimensional functions at the MPP of the random vector
X as
N
G( X) Gˆ ( X) G( x1* , , xi*1 , X i , xi*1 , , x*N ) ( N 1)G(x* ) (2.29)
i 1
where x* ={x1* , x2* , , x*N }T is the FORM-based MPP of the performance function G( X)
obtained from Eq. (2.14) and N is the number of random variables. For example, if
This MPP-based univariate DRM will be used for more accurate reliability analysis than
and whose MPP is denoted by x* ={x1* , x2* , , x*N }T . Since the reliability analysis is
u*
orthonormal matrix R R N N
whose N column is α*
th
, i.e., R [R1 α* ], where
HL
standard normal V-space with the MPP v* {0, ,0, HL }T . The orthonormal matrix R
orthonormal matrix R is not uniquely determined. Figure 2.3 shows U-space and V-
space for N 2 .
Figure 2.3. Standard Normal U-space and Rotated Standard Normal V-space
minimize h( X)
subject to Gi ( X) 0, i 1, , nc (2.31)
X L X XU , X R ndv
27
where h( X) is the cost function, Gi is the ith constraint, and X is the design variable
vector; and nc and ndv are the number of constraints and design variables, respectively.
optimum that could be sensitive to the variation of input design variables and other
parameters.
Due to the variation of the input design variables and other parameters, the output
performance function h( X) also has variation. Thus, in the robust design, the robustness
other words, the goal of the robust design is to find the most insensitive design to the
variation of the design variables and other parameters. Since the robust design is
fundamentally considering the variations of the design variables and other parameters, it
is very natural to integrate the robust design and reliability-based design in one
minimize f ( H , H2 )
subject to P(Gi ( X; d) 0) PFTar
i
, i 1, , nc (2.32)
d L d dU , d R ndv and X R nrv
where f ( H , H2 ) is the cost function, d μ(X) is the design vector, X is the random
vector, and Gi is the ith probabilistic constraint. Quantities nc , ndv , and nrv are the
design for a one-dimensional performance function. With the same variabilities of design
28
variables, the robust optimum design shows less variation of the performance function
Since the cost function in Eq. (2.32) depends on H and H2 for the robust
the bi-objective optimization depends on the weight on each term in the cost function.
However, since the main goal of this research is not focused on determination of the
weights, this topic is referred to Marler (Marler and Arora, 2004) for more details. The
engineering application types (Chandra, 2001; Youn et al., 2005c). The following are
H ht 2
f ( H , H2 ) w1 ( ) w2 ( H ) 2 (2.33)
H ht
0 0
H 0
where ht and ht0 are the target nominal value and the initial target nominal value of the
the designer. To reduce the dimensionality problem of two objectives, each term is
H 2
f ( H , H2 ) w1 sgn( H ) ( ) w2 ( H ) 2 (2.34)
H 0
H 0
the sign of H .
H 2
f ( H , H2 ) w1 sgn( H ) ( ) w2 ( H ) 2
0
(2.35)
H H 0
30
CHAPTER III
3.1 Introduction
As explained in Section 2.6, major concern of RDO is how to estimate the second
statistical moments and their sensitivities accurately and efficiently. Analytically, the kth
statistical moment of the performance function h( X) can be obtained using the following
integration
E ({h( X)}k ) {h( X)}k f X (x)dx (3.1)
where f X (x) is the joint PDF of the random parameter X . As stated before, it is
using Eq. (3.1) especially when the dimension of the problem is relatively large. The first
order Taylor series expansion has been widely used to estimate the statistical moments
due to its simplicity. Using the first order Taylor series expansion, the mean value and
variance of the performance function can be estimated (Huang and Du, 2006)
respectively as
H h(μ) (3.2)
and
2
N h
H2 X2 i (3.3)
i 1 X i
x μ
However, Taylor series expansion results in a large error especially when input
random variables have large variations. This is because the first order Taylor series
expansion does not use all information of PDF of input random variables. To overcome
the shortcoming of Taylor series expansion, three numerical methods have been recently
proposed: the mean value-based DRM, performance moment integration (PMI), and
Using Eqs. (2.25), (2.27) and (2.28), the mean value and variance of the
H E[h( X)]
N
E{ h( 1 , , i 1 , X i , i 1 , , N ) ( N 1) h( 1 , , N )} (3.4)
i 1
n N
wij h( 1 , , i 1 , xij , i 1 , , N ) ( N 1) h( 1 , , N )
j 1 i 1
and
The estimation of statistical moments using the univariate DRM involves two
approximations. As shown in Eqs (3.4) and (3.5), the univariate DRM approximates the
N
h( X) hi ( X i ) where hi ( X i ) is any function of X i only, then the approximation is
i 1
32
exact. However, if there are off-diagonal or mixed terms, then there is some error that
To reduce this error, the bivariate DRM or multivariate DRM can be used (Xu and
Rahman, 2004). The second approximation involves the numerical integration using the
weights and quadrature points. Based on Gaussian quadrature theory (Atkinson, 1989), n
quadrature points and weights for each variable are used, the numerical integration error
for a quadratic performance function will disappear. If the performance function is highly
nonlinear, then three quadrature points may not be sufficient to estimate the moments of
the performance function. In this case, the error can be reduced if the number of
Even though the accuracy is the most important concern, it is also important to
efficiently estimate the statistical moments of the performance function for large-scale
problems. In general, when the output moments are estimated using the univariate DRM
where n is the number of quadrature points and N is the number of design variables. If
the distributions of all input design variables are symmetric, e.g. normal distribution or
uniform distribution, and the number of design variables is odd, then the required number
Therefore, when the number of design variables is large, the reduction becomes
increasing proportionally to the number of design variables as shown in Eqs. (3.6) and
(3.7).
If bivariate DRM is used to estimate the first and second output moments, then
N ( N 1) 2
number of F.E. n n N 1 (3.8)
2
For example, if the number of design variables is 5 and the number of quadrature points
is 3, then the number of function evaluations by the univariate DRM is 16 from Eq. (3.7)
and the number of function evaluations by bivariate DRM is 106 from Eq. (3.8). Both
numbers are less than 35 243 , which is the required number of function evaluations for
the numerical integration of Eq (3.1) by including the mixed variable terms. However, the
number of function evaluations by the univariate DRM is significantly less than the
number of function evaluations for bivariate DRM. For this reason, the univariate DRM
To obtain a robust design, not only the values of the first and second statistical
moments but also the sensitivities of these moments are needed. Using Eq. (3.1) and
Rosenblatt transformation from the design space (X-space) to the standard Gaussian
34
sensitivities of the mean and variance of the performance function with respect to the
H (μ)
i
[ h(x) f X (x)dx]
i
i
[ h(x(u; μ))U (u) du]
(3.9)
h(x(u; μ))
i U (u)du
and
H2 (μ) H2
i
[ h 2 (x) f X ( x) dx]
i i
H2
i
[ h 2 (x(u; μ))U (u) du]
i
(3.10)
h 2 (x(u; μ))
H2
i (u ) du
i
U
where u is the standard normal variable. The input variables are assumed to be
xi (ui ; i )
To calculate in Eqs. (3.9) and (3.10), Rosenblatt transformation shown
i
in Table 2.1 is used. For example, if the input variable is normally distributed, then Table
xi (ui ; i )
independent of an input mean i , 1 is obtained. For Gumbel and uniform
i
35
xi (ui ; i )
distributions, the same result 1 is obtained from Rosenblatt transformation.
i
xi (ui ; i )
For the Lognormal and Weibull distributions, can be approximated to be 1.
i
xi (ui ; i )
1, and Eqs. (3.4) and (3.5), Eqs. (3.9) and (3.10) can be further approximated
i
as
and
Since the univariate DRM does not use sensitivities of the performance function
evaluations are needed for the sensitivity analysis using Eqs. (3.11) and (3.12).
The multi-dimensional integral in Eq. (3.1) for the statistical moments can be
E (hk ( X)) hk (x) f X (x; μ)dx hk (x(u; μ))U (u)du (3.13)
E (hk ( X)) hk (x(u; μ))U (u)du hk f H (h; μ)dh (3.14)
where f H (h) is the PDF of the output performance function h( X) . Since CDF of the
performance function can be expressed in terms of the standard normal CDF using the
E (hk ( X)) hk f H (h; μ)dh hk (t; μ) (t )dt (3.15)
where the parametric variable t is the distance from the origin in U-space to MPP as
integral. Similar to the univariate DRM, PMI makes use of three quadrature points and
between the two methods is that quadrature points of the univariate DRM lie on the xi -
axis, whereas quadrature points of PMI lie on the MPP locus (Du and Chen, 2001; Youn
and Choi, 2002). Therefore, the number of quadrature points in the univariate DRM
increases as the number of design variables increases as shown in Eqs. (3.6) and (3.7),
whereas the number of quadrature points in PMI does not change since the integration is
Since t follows the standard normal distribution, the weights and quadrature
1 4 1
E (hk ( X)) hk (t; μ) (t )dt h k (t; μ) h k (t; μ) h k (t; μ) (3.16)
6 t 3 6 t 0 6 t 3
E (h k ( X)) h k (t ; μ) (t )dt
1 4 1
{h(t ; μ)}k {h(t ; μ)}k {h(t ; μ)}k (3.17)
6 t 3 6 t 0 6 t 3
1 4 1
h k ( 3; μ) h k (0; μ) h k ( 3; μ)
6 6 6
Using FORM and MPP locus illustrated in Figure 3.1, each term in Eq. (3.17) can be
approximated as two function values at two MPPs and a function value at the design
point. The function values at MPPs can be obtained using the inverse reliability analysis
PMA to
maximize h(U)
(3.18)
subject to U 3
The optimum result of Eq. (3.18) is denoted as hmax 3 , which can be used to
t
approximated by the optimum result obtained by minimizing h(U) in Eq. (3.18), which is
denoted as hmin 3 . The term h(0; μ) in Eq. (3.17) can be approximated by h(μ X ) , which
t
is the performance function value at the design point. Hence, using these function values
and Eq. (3.17), the statistical moments of the performance function can be calculated as
1 4 1
E (hk ( X)) (hmin 3 )k hk (μ X ) (hmax 3 ) k (3.19)
6 t
6 6 t
1 4 1
H hmin 3 h(μ X ) hmax 3 (3.20)
6 t
6 6 t
and
1 4 1
H2 (hmin 3 )2 h2 (μ X ) (hmax 3 )2 H2 (3.21)
6 t
6 6 t
Thus, PMI is very efficient when the number of design variables is relatively large.
Similar to the sensitivity calculation in DRM, from Eqs. (3.15), (3.17), and (3.19),
the sensitivities of the mean and variance of the performance function with respect to a
H h(t ; μ)
i i
h(t ; μ) (t )dt
i
(t )dt
3 3
(3.22)
6 i 6 i 6 i
1 h(x) xi* 4 h(x) 1 h(x) xi*
6 xi x=x* i 6 xi x=μ 6 xi x=x* i
min X max
and
39
H2 H2 h (t ; μ )
2
H2
h (t; μ) (t )dt i i (t )dt i
2
i i
1 (ht 3 ) 4 h 2 (μ X ) 1 (ht 3 ) H2
min 2 max 2
(3.23)
6 i 6 i 6 i i
1 h 2 (x) xi* 4 h 2 (x) 1 h 2 ( x) xi* H2
6 xi x=x* i 6 xi x=μ 6 xi x=x* i i
min X max
xi* x (u ; )
By using the approximation 1 , which is a similar to i i i 1 in DRM,
i i
sensitivities of the mean and variance of the performance function with respect to i in
and
Since the sensitivities of the performance function on the right hand side of Eqs.
(3.24) and (3.25) are used during the inverse reliability analysis described in Eq. (3.18),
no additional function evaluations are required to calculate sensitivities using Eqs. (3.24)
and (3.25).
Like PMI, PDM uses results of the inverse reliability analysis. As explained in the
previous section, PMI utilizes the function values at two MPPs ( hmax 3 and hmin 3 )
t t
obtained from the inverse reliability analysis and the function value at the mean X to
40
approximate the multi-dimensional integration in Eq. (3.1), whereas PDM “uses the
difference between the function values at two MPPs to represent the variation of the
performance function” (Du et al., 2004). Hence, the RBRDO formulation using PDM is
to
When p1 0.95 and p2 0.05 (Du et al., 2004; Mourelatos and Liang, 2005), hp1 and
hp2 in Eq. (3.26) are calculated from the inverse reliability analysis with a target
As shown in Figure 3.2, the idea of PDM is simple and could be viewed as
not monotonic, it may not be possible to use hp1 hp2 as a measurement for robustness.
For a non-monotonic performance function, two MPPs obtained from the inverse
reliability analysis may not approximate the left-tail and right-tail percentile accurately
because the inverse reliability analysis searches MPPs on the surface of the hyper-sphere
1.645, then two MPPs become 1.645 and –1.645. Thus, two percentile performances hp1
and hp2 are identical. In contrast to PDM, PMI and mean-based DRM show the correct
may identify a wrong global minimum when there are several local minima, as will be
41
demonstrated in Section 3.6.2. More significantly, there is no one percentile that can be
used in PDM to identify all local optima correctly as will be shown in Section 3.6.2.
h(μ X ) h(x)
(3.27)
i xi x=μ
X
and
3.5 Comparison
and accuracy of the moment and its sensitivity estimation. In terms of computational
efficiency, both PMI and PDM will require the same number of function evaluations if
the same inverse reliability analysis method is used. In general, if the number of design
variables is large, Eqs. (3.7) and (3.8) show that the univariate DRM requires more
function evaluations than PMI and PDM. However, an advantage of using the univariate
DRM is that the univariate DRM does not require sensitivity information (i.e., no MPP
search) in estimating the moments. Hence, the univariate DRM can reduce the number of
dimensional integration in Eq. (3.1). That is, both methods attempt to transform the
PDM does not use any numerical integration; instead it uses the difference of percentile
performances. Thus, PDM may yield wrong results when the performance function is
non-monotonic. Both PMI and PDM may have a difficulty to find MPPs when the
performance function is non-monotonic and highly nonlinear. On the other hand, DRM
may accurately estimate the moments of the performance function regardless of the
better results in most cases than PMI. If the performance function is highly nonlinear,
then the mean-based DRM with three quadrature points may not accurately estimate the
second moment. In this case, the error can be reduced if more quadrature points are used
43
in the mean-based DRM. However, PMI with more quadrature points than three may not
necessarily yield more accurate results. This is because function values at quadrature
points, which are obtained using FORM and MPP search, are approximations. More
details of comparison with numerical examples are given in the following section.
In this section, four cases of comparisons are carried out using numerical
examples. In Section 3.6.1, the mean-based DRM and PMI are compared in terms of
performance function. PDM is excluded in Section 3.6.1 since it cannot estimate the
moments of the performance function. In Section 3.6.2, DRM, PMI, and PDM are
robust optimum design. In this one-dimensional problem, PMI and the mean-based DRM
X 12 X 2
h1 ( X) 1 (3.29)
20
where X i ~ N (5,1) for i 1, 2 . As shown in Table 3.1, both DRM and PMI provide good
estimation of the mean value and standard deviation in comparison with the exact
numerical integration results. The reason that DRM has a larger error in estimation of
44
standard deviation is because the performance function in Eq. (3.29) has an off-diagonal
term only. As mentioned in Section 3.2, if the performance function has off-diagonal
terms only, then the univariate additive decompositions of the moments in Eqs. (3.4) and
Table 3.1. Comparison of the First and Second Moments of Eq. (3.29)
Mean ( H ) Variance ( H2 )
PMI DRM NI* PMI DRM NI
h1 -5.6500 -5.5000 -5.5000 8.4623 7.9355 8.3175
Error, % 2.73 0 1.74 4.58
No. of F.E. 7+7** 2 2 1 7+7 2 2 1
* NI means numerical integration.
For this example, PMI yields reasonable estimation of the moments because the
design variables are normally distributed, which means that the inverse reliability
analysis does not require the nonlinear transformation from X-space to U-space, and the
performance function is monotonic at the given design. In the same token, the
sensitivities in Tables 3.2 and 3.3 have similar errors as Table 3.1.
The total number of function evaluations for PMI to evaluate the mean and
standard deviation is 7+7 as shown in Table 3.1, where the first 7 is the number of
function evaluation for MPP search and the second 7 is the number of sensitivity
calculation for the MPP search. The number of function evaluations for DRM is 5. Since
the design variables are normally distributed and the number of quadrature points is odd,
PMI does not require additional function evaluations for the sensitivity analysis
of moments because PMI uses the sensitivity information in MPP search. However, DRM
does require additional function evaluations for sensitivity analysis, thus the total number
Table 3.2. Sensitivity of Mean Value Using PMI and DRM for Eq. (3.29)
Table 3.3. Sensitivity of Variance Using PMI and DRM for Eq. (3.29)
Since the first example contains an off-diagonal term only and the design
( X1 X 2 5)2 ( X1 X 2 12) 2
h2 ( X) 1 (3.30)
30 120
46
where X i ~ Gumbel (5,1) for i 1, 2 . The performance function in Eq. (3.30) contains
both off-diagonal terms and diagonal terms, and the degree of the polynomial
performance function is 2. Therefore, it can be expected that DRM may yield better
results for this example. As expected, Tables 3.4, 3.5 and 3.6 illustrate that DRM is
On the other hand, PMI yields somewhat larger errors in estimation of the
moments and their sensitivities. This is because the design variables follow Gumbel
transformation, which makes the performance function become highly nonlinear and the
Since the Gumbel distribution is not symmetric, Eq. (3.6) is used for the total
Table 3.4. Comparison of the First and Second Moments of Eq. (3.30)
Mean ( H ) Variance ( H2 )
PMI DRM NI PMI DRM NI
h2 -1.0594 -1.1167 -1.1167 0.3357 0.3774 0.3833
Error, % 5.13 0 12.42 1.54
No. of
5+5 3 2 1 5+5 3 2 1
F.E.
Table 3.5. Sensitivity of Mean Value Using PMI and DRM for Eq. (3.30)
Table 3.6. Sensitivity of Variance Using PMI and DRM for Eq. (3.30)
In this section, three methods are compared in detail for proper identification of
the robust optimum design, using a one-dimensional example. RDO can be formulated to
minimize H2
(3.31)
subject to 0 X 5
dimensional problem, PMI and the mean-based DRM with three quadrature points can be
considered to be the same method since the design variable is normally distributed and
there is no FORM error in a one-dimensional function. Figure 3.3 (a) illustrates the shape
of the performance function and Figure 3.3 (b) illustrates the variances obtained from
As shown in Figure 3.3 (b), PMI and DRM with three quadrature points can
approximate the variance of the performance function very well. On the other hand, PDM
with various percentiles cannot estimate the moments. More significantly, the location of
the optimum point changes depending on the percentiles used. In fact, there is no one
percentile that can be used to accurately identify the location of both local minima
simultaneously in Figure 3.3 (b). Table 3.7 shows that the best percentile should be
located between 2 and 3 for the left local minimum; and the best percentile should be
located between 1.645 and 2 for the right local minimum. In Figure 3.3 (b), „Measure
for Variance‟ is used instead of variance. This is because PDM cannot estimate the
variance of the performance function and uses percentile differences as the measure for
the variance.
such as Eq. (3.31) is that PDM might not be able to identify which local minimum is the
49
global minimum when there is more than one local minimum. As shown in Table 3.7, the
results of PDM with three different percentiles indicate that the value of the cost function
at the left minimum in Figure 3.3 (b) is less than the value at the right minimum, which is
wrong.
Table 3.7 also shows that PMI and DRM with three quadrature points yields
some errors in finding the location of the optimum and estimating the value of the
three quadrature points may not be sufficient. In this case, DRM and PMI with five
quadrature points are good options to achieve accuracy. The accuracy of DRM and PMI
with five quadrature points is illustrated in Table 3.7 and Figure 3.4.
Table 3.7. Position and Value of Optimum Using Three Methods for Eq. (3.31)
Figure 3.4. Accuracy of PMI and DRM with Five Quadrature Points
51
CHAPTER IV
4.1 Introduction
FORM has been widely used for the reliability analysis and inverse reliability
analysis due to its simplicity and efficiency, and FORM shows a good approximation for
the mildly nonlinear system or linear system. However, since FORM uses a linear
could include large errors when the system is highly nonlinear and/or multi-dimensional.
These errors can be improved by SORM since SORM uses a quadratic approximation of
the performance function. However, the Hessian matrix is required to calculate the
probability of failure in Eq. (2.10) using SORM, which is very difficult or sometimes
impossible to estimate accurately in real engineering applications. For this reason, use of
reliability analysis is proposed in this chapter. Section 4.2 demonstrates the weaknesses
of FORM for a highly nonlinear and multi-dimensional system and Section 4.3 illustrates
how to obtain the DRM-based MPP using three computational steps. The DRM-based
MPP will be used for the next design iteration of RBDO and thus yield an accurate
optimum design even for highly nonlinear system. The RBDO with the MPP-based DRM
will be explained in Chapter 6. Section 4.4 compares FORM, SORM, and DRM through
numerical examples.
52
Although FORM has been widely used for the reliability analysis and inverse
reliability analysis due to its simplicity and efficiency, FORM could be erroneous if the
N 1
G( X) a X i2 X N (0.1)
i 1
x* {0, ,0, }T and the probability of failure by FORM is PFFORM ( ) regardless
PFFORM (2) 2.2750% . This probability of failure can be compared with the
probability of failure obtained using MCS for different a and N, respectively, as shown in
Tables 4.1 and 4.2. From Tables 4.1 and 4.2, it can be seen that the probability of failure
obtained using FORM has significant error when a performance function is highly
N 2 N 3 N 5 N 10
MCS
PF 1.5915% 1.1309% 0.5426% 0.0790%
DRM-based MPP, denoted by x*DRM , using the MPP from the FORM-based inverse
reliability analysis denoted by x*FORM . As stated in Section 2.3, the inverse reliability
analysis does not calculate the probability of failure directly, instead, it judges whether a
given design satisfies the probabilistic constraint by checking the performance function
value at MPP. However, the probabilistic constraint may not be satisfied even though the
constraint value at the FORM-based MPP G(x*FORM ) is less than zero, which means the
MPP is safe. This is because the probability of failure calculated by FORM may have
function. In this section, a new method is proposed to find a DRM-based MPP x*DRM .
Finding the new DRM-based MPP consists of three steps: constraint shift, reliability
Shift
In the rotated standard normal V-space, the probability of failure in Eq. (2.1) can
be rewritten as
54
where the performance function in V-space is defined as G( v) G(x( v)) . Since the
inverse reliability analysis does not calculate the probability of failure, a constraint shift
G s ( v) G( v) G( v* ) (0.3)
MPP in V-space with a given reliability index . By applying the MPP-based DRM
N
ˆs
G ( v) G ( v) Gis (vi ) ( N 1)G s ( v* )
s
(0.4)
i 1
where Gis (vi ) Gis (v1* , , vi*1 , vi , vi*1 , , v*N ) . By the definition of G s ( v) in Eq. (4.3),
N N 1
ˆ
G s ( v) G s ( v) Gis (vi ) GNs (vN ) Gis (vi ) (0.5)
i 1 i 1
Due to the rotational transformation of the coordinates as shown in Fig. 2.3, the
Nth univariate component GNs (vN ) can be linearly approximated (Wei, 2006). This linear
assumption of GNs (vN ) along vN -axis is also used for the probability of failure
calculation in SORM (Breitung, 1984; Hohenbichler, 1988). Using the linear assumption,
N 1
1 N Gi (Vi ) 0]
PF P[G s (V) 0] P[b0 bV s
(0.6)
i 1
55
Since function value and gradient at MPP are obtained during the inverse reliability
G s ( v) G( v)
GNs (vN ) b0 b1vN GNs (v*N ) (vN vN* ) (vN ) (0.7)
vN v v* vN v v*
G s ( v) G( v)
where GNs (v*N ) 0, b1 , and v*N . Inserting Eq. (4.7) into Eq.
vN v v* vN v v*
(4.6) yields
N 1 N 1
PF P[b1 (VN ) Gis (Vi ) 0] P[bV
1 N b1 Gi (Vi ) ]
s
(0.8)
i 1 i 1
inverse reliability analysis in Eq. (2.14), Eq. (4.8) can be rewritten, by dividing both sides
by b1 and using the symmetry of the standard normal distribution since V ~ N (0,1) , as
1 N 1 s
PF P[VN Gi (Vi ) ]
b1 i 1
(0.9)
1 N 1 1 N 1
P[VN Gis (Vi ) ] E[( Gis (Vi ) )]
b1 i 1 b1 i 1
where E is the expectation operator. Since Eq. (4.9) is an N−1 dimensional integration,
Eq. (4.9) can be further simplified by applying DRM to the integrand of Eq. (4.9) as
N 1 Gis (vi )
i 1
(
b1
) (vi )dvi
PFDRM (0.10)
( ) N 2
where b1 U g . Detailed derivation of Eq. (4.10) can be found in Ref. (Wei, 2006).
Using the moment-based integration rule (MBIR) (Xu and Rahman, 2003), which is
N 1 n
Gis (vij )
w j ( b )
i 1 j 1
PFDRM 1
(0.11)
( ) N 2
where vij are quadrature points, w j are weights, and n is the number of quadrature
points and weights. Since vi are standard normal random variables, quadrature points
and weights in Table 2.3 can be used to calculate Eq. (4.11) and locations of quadrature
points are illustrated in Figure 4.1. The coordinates for the quadrature points vij in Figure
4.1 are vi1,3 (0, , 3, , ( k ) ) and vi2 (0, ,0, , ( k ) ) , where 3 and 0 are at i
th
positions.
For a special case of Eq. (4.11), if n 1 , which means one quadrature point and
N 1 N 1
Gis (vi1 )
w1 ( ) ( )
b1
P DRM
i 1
i 1
( ) (0.12)
( ) N 2 ( ) N 2
F
where w1 1 and vi1 0 by Table 2.3, and Gis (vi1 ) Gis (0) 0 . Equation (4.12) is the
same as the probability of failure by FORM. Therefore, we can say that the probability of
besides the MPP search using FORM, is ( N 1) (n 1) . Hence, the total number of
Since the probability of failure calculation using DRM requires integration in Eq.
increasing the number of quadrature points and weights in Eq. (4.11). In this case, the
probability of failure by DRM requires only function values at the quadrature points,
which are Gis (vij ) in Eq. (4.11). Consequently, the accuracy of the DRM result can be
improved by increasing the number of quadrature points if necessary, which does not
require any sensitivity. The comparison of the number of function evaluations with the
Section 4.5.
Using the MPP-based DRM to use Eq. (4.11) for the probability of failure
calculation, in contrast to the mean-based DRM, is like using the importance sampling
method near the MPP in contrast to the Monte Carlo simulation (MCS) method for the
58
failure probability estimation. However, even though the importance sampling method
requires less number of samplings than MCS, it still requires quite amount of samplings
to obtain the accurate probability of failure (Denny, 2001). The MPP-based DRM
requires significantly less function evaluations than the importance sampling for the
accurate estimation of the failure probability. Hence, this proposed method can be easily
After computing the probability of failure PFDRM using the MPP-based DRM for
obtained using
It is likely that DRM from Eq. (4.14) is not the same as the target reliability index
t 1 ( PFTar ) . Hence, a new updated reliability index ( k 1) is obtained using the
where ( k ) is the reliability index at the current step, with (0) t at the initial step.
If the performance function is concave as shown in Figure 4.1 (a), then ( k 1) will
be smaller than ( k ) because DRM t , which means that a smaller reliability index
should be used to correctly update MPP using DRM for the concave performance
function, and vice versa for the convex performance function in Figure 4.1 (b). In this
59
study, a performance function is defined as concave near the MPP if the FORM-based
reliability analysis overestimates the probability of failure, and convex near the MPP if
Using this updated reliability index, we can carry out a new inverse reliability
analysis to find a better MPP which satisfies the given target probability of failure. After
finding a new MPP, constraint shift is again used to compute the probability of failure by
DRM. After iteratively doing this procedure until converged, the DRM-based MPP can
expensive if a new MPP search is carried out every time an updated reliability index is
obtained. For efficiency, the updated MPP corresponding to the probability of failure by
DRM is obtained without carrying out a new MPP search as (Ba-abbad et al., 2006)
(k+1) * (k+1) *
u*k+1 u or v *
v (0.16)
(k ) k k+1
(k ) k
That is, it is assumed that the updated MPP u*k+1 is located along the same radial direction
as the current MPP u*k in U-space. Figure 4.1 illustrates that the updated DRM-based
MPP in V-space is located along the vN-axis. The same iterative procedure explained
above can be performed until PFDRM converges to PFTar . To verify the MPP approximation
in Eq. (4.16), a comparison test to find the DRM-based MPP using two methods is
However, this iterative procedure using Eq. (4.16) still requires additional
function evaluations. To reduce the number of function evaluations, the MPP update can
be carried out only once at a given design. The updated MPP obtained through the
procedure is also called the DRM-based MPP and will be used to evaluate whether the
design satisfies the probabilistic constraint or not. Even though the MPP update is carried
out only once at the given design, the probability of failure by DRM will converge to the
target probability of failure as the design moves near the reliability-based optimum
design.
the FORM and SORM-based probabilities of failure in Section 4.4.1. For this purpose,
the probability of failure obtained using MCS is used as a benchmark data. Section 4.4.2
compares two iterative methods, a method using a new MPP search and using Eq. (4.16)
Y 0.9063 0.4226 X 1
where , X1 ~ N (4,0.3) and X 2 ~ N (3,0.3) , is used for
Z 0.4226 0.9063 X 2
the probability of failure computation. The reliability index of 1.645 is used for the
Figure 4.2 (a) shows the shifted and original performance functions and Figure
4.2 (b) shows the approximated functions by FORM, SORM, and DRM at MPP in V-
space. In Table 4.3, DRM with three and five quadrature points are used to evaluate Eq.
(4.11). From Table 4.3, it can be seen that DRM with five-quadrature points is the most
accurate method for this example. In fact, this result is even more accurate than the
SORM result, compared with the MCS result with 1 million samples, which can be
considered accurate. In terms of efficiency, FORM shows the best efficiency, which is
always true since SORM and DRM require the FORM-based MPP. However, the
additional number of function evaluations for DRM besides the MPP search does not
require sensitivity analysis. Hence, DRM can estimate the probability of failure as
efficiently as FORM without loss of accuracy - the error of the FORM result is about
DRM
FORM SORM MCS
3 pts 5 pts
PF , % 5.0000 3.4081 3.5844 3.3676 3.2791
F.E. 7* 7*+Hessian 7*+2** 7*+4**
* 7 means the number of function and sensitivity analysis for MPP search
where X i ~ N (5,0.4), i=1 ~ 4 , is used for the probability of failure computation. The
reliability index of 1.645 is used for the FORM-based inverse reliability analysis. As
described in Eq. (4.13), the total number of function evaluations for the DRM-based
reliability analysis will increase as the number of random variables increases. Since the
63
performance function in Eq. (4.20) has four random variables, (4 1) (3 1) 6 function
evaluations are required for DRM with three quadrature points; and (4 1) (5 1) 12
function evaluations are required for DRM with five quadrature points as shown in Table
4.4. Again, these function evaluations do not require sensitivity analysis. Table 4.4 shows
DRM yields the best accuracy compared with the MCS results.
DRM
FORM SORM MCS
3 pts 5 pts
PF , % 5.0000 14.0396 11.8049 11.8976 11.9064
F.E. 2 2+Hessian 2+6 2+12
The two-dimensional performance function in Eq. (4.17) is again used for the
convergence test in this section. For the given target probability of failure PFTar 2.275%,
the initial reliability index in Tables 4.5 and 4.6. This reliability index is used for the
FORM-based inverse reliability analysis to find MPP. After finding the FORM-based
MPP, the probability of failure is calculated using DRM and compared with the target
1.4109% as shown in Table 4.5, is smaller than the target probability of failure,
PFTar 2.275%, the following reliability index should be smaller than the initial reliability
where cur t 2 since it is the initial iteration and DRM 1 (0.014109) . The
updated reliability index is shown in Tables 4.5 and 4.6. Using the MPP update in Eq.
(4.16), the first DRM-based MPP candidate is obtained as (4.5009, 2.7936) in Table 4.5.
By iteratively performing this procedure, finally the DRM-based MPP can be obtained as
almost the same as the target probability of failure 2.2750%. In this example, the updated
reliability index is decreasing since the performance function is concave near the MPP,
which means that the FORM-based reliability analysis overestimates the probability of
failure.
Table 4.6 shows iterative way of finding the DRM-based MPP using the new
MPP search, which means that the new MPP search is carried out using FORM after
obtaining the updated reliability index. Since this requires the MPP search at every design
iteration, it becomes expensive to find the DRM-based MPP as shown in Tables 4.5 and
4.6. The total number of function evaluations and sensitivity analysis needed for Table
4.5 is 16 and 5, respectively. Whereas, the total number of function evaluations and
sensitivity analysis needed for Table 4.6 is 21 and 15, respectively, which is more
Table 4.6. Iterative Way of Finding DRM-Based MPP Using New MPP Search
Both methods converge very fast within three iterations. In addition, two MPPs
obtained using MPP approximation and new MPP search are close to each other as shown
in Tables 4.5 and 4.6, which means that the MPP approximation method in Eq. (4.16) can
be effectively used to find the DRM-based MPP without requiring further MPP search.
Also, the fact that the approximated MPP at the first iteration, (4.5009, 2.7936) , is very
close to the approximated MPP, (4.5029, 2.7928) , at the last iteration in Table 4.5
verifies that the MPP update using Eq. (4.16) can be carried out only once at a given
design without loss of accuracy. This reduction of the number of function evaluations
CHAPTER V
RBDO
minimize Cost(d)
subject to Gi (x* ) 0, i 1, , nc (5.1)
d L d dU , d R ndv and X R nrv
where x* is the FORM-based MPP which can be obtained by solving the inverse
reliability analysis in Eq. (2.14). In a manner similar to FORM-based PMA, the DRM-
minimize Cost(d)
subject to Gi (x*DRM ) 0, i 1, , nc (5.2)
d dd ,
L U
dR ndv
and X R nrv
For optimizations given by the formulation (5.1) for FORM-based PMA and the
formulation (5.2) for DRM-based PMA, sensitivities of the objective function and
constraints with respect to the design variables are required. In both formulations, it is
straightforward to obtain the sensitivities of the objective function with respect to design
variables since the objective is a function of the design variables, which are the mean
values of the input random variables. However, it is not straightforward to obtain the
67
sensitivities of the probabilistic constraint at MPP with respect to the design variables
since MPP of the perturbed design is involved in evaluation of the probabilistic constraint
at the perturbed design. Hence, it is required to analytically derive the sensitivities of the
G(x* ) G G xi x G
N T
(5.3)
d d x x* i 1 xi x x* d xx* d x x* x x x*
G(x*DRM ) G G xi x G
N T
(5.4)
d d xx*DRM i 1 xi x x* d x x*DRM d xx*DRM x x x*DRM
DRM
U g (u* )
u t
*
t α (5.5)
U g (u* )
derivatives on both sides of Eq. (5.5) with respect to jth design variable dj, which is the
u* U g t U g b
t 2 b1 U g 1 (5.6)
d j d j U g b1 d j d j
where b1 U g and all derivatives are evaluated at MPP. The left side of Eq. (5.6) is
u* u x
T T ej , (5.7)
d j d j d j
u=u* x x*
x j d j f (u) . (5.8)
The assumption in Eq. (5.8) works for general distributions whose contour shape
of the joint input PDF does not change when the design point moves; for example,
Weibull distribution, etc. If all input random variables are independent, T becomes a
diagonal matrix, and if the random variables are dependent, then T becomes a triangular
U g b
and 1 in Eq. (5.6) are calculated in X-space as
d j d j
U g x
TH (5.10)
d j d j
x x*
using U g TG where and
xi
b1 T GTT TH x
, (5.11)
d j b1 d j
xx *
69
respectively, where G and H are the gradient vector and Hessian matrix evaluated at
MPP in X-space. Substituting Eqs. (5.7), (5.10), and (5.11) into Eq. (5.6) yields
x t x t x T
ej TT TH TGTT TH T TG . (5.12)
d j b1 d j 3
b d j
xx
*
xx
* 1 xx
*
t T t T x
I T TH 3 T TG GT TH ej
T T
(5.13)
b1 b1 d j xx*
t T t T x
I T TH 3 T TG GT TH I.
T T
(5.14)
b1 b1 d x x*
Thus,
1
x
I t TT TH 3t TT TGTGTT TH , (5.15)
d x x* b1 b1
which are the sensitivities of MPP in X-space with respect to design variables.
From Eqs. (5.3) and (5.15), the sensitivities of the probabilistic constraint in Eq.
T
G x G
T
I t TT TH 3t TT TGTGTT TH G . (5.16)
d xx* d xx* x xx* b1 b1
To further simplify the right-hand side of Eq. (5.16), consider the following equation:
T
t T t T
I T TH 3 T TG GT TH G
T T
b1 b1
I t HTT T 3t HTT TG T GTT T G (5.17)
b1 b1
t t
G
b1
HTT TG
b13
HT TG b
T
1
2
G ,
70
since H is symmetric and b12 U g TGTT TG . Using Eq. (5.17), Eq. (5.16) can
2
be rewritten as
G x G G
T
, (5.18)
d x x* d x x* x x x* x x x*
which is the sensitivity of the probabilistic constraint with respect to the design variables
in Eq. (5.3) for the FORM-based PMA. A number of papers that use the probabilistic
constraint assumed Eq. (5.18). However, this relation is exact and there is no need of
assuming it.
When constraints are black box type, which means the design sensitivity is not
available, to evaluate the sensitivities in Eq. (5.3) using FDM, additional MPP searches at
the perturbed designs are required, which is computationally very expensive. In addition,
since the design perturbation is required for each design variable, the sensitivity
calculation using FDM will become very expensive when the number of design variables
increases. However, to evaluate the sensitivities in Eq. (5.18) using FDM, no additional
MPP search is required, and thus very efficient with the same accuracy.
The sensitivities in Eq. (5.18) can be shown in a different way. Using the
definition of MPP in Eq. (5.5) and α Tα 1 , the target reliability index is written as
t α Tu* . (5.19)
t α T * u*
u αT . (5.20)
d j d j d j
71
α T
α 0 from α α 1 and t is constant, Eq. (5.20) can be written as
T
Since
d j
T T
u* u
α α 0, (5.21)
d j d j u=u*
which yields
T
u
U g 0 . (5.22)
d j u=u*
T T
x x
T
u
U g e j T1TG e j G 0 (5.23)
d j d j
d j u=u* xx*
xx
*
Since the DRM-based MPP can be found using either a new MPP search or an
constraints at both true DRM-based MPP ( u*DRM ) and approximated DRM-based MPP (
uaDRM ) are derived in Sections 5.3.1 and 5.3.2, respectively. Section 5.3.3 illustrates that
these sensitivities converge to each other as the design point approaches the optimum
design.
72
Based MPP
The DRM-based MPP obtained from a new MPP search with the updated
U g (u*DRM )
u*DRM up upα DRM (5.25)
U g (u*DRM )
and
up αTDRMu*DRM (5.26)
,
where α DRM is the normalized gradient vector at the true DRM-based MPP. In a way
similar to that explained in Section 5.2, taking derivative of Eq. (5.26) yields
T T T
up u* u*DRM U g (u*DRM ) u*DRM U g DRM
DRM α
DRM (5.27)
d j d j d j U g (u*DRM ) d j U g DRM
73
α DRM
using the orthogonality of α DRM and (Ditlevsen and Madsen, 1996). Hence,
d j
T
u*DRM up
U g DRM U g DRM . (5.28)
d j d j
up
Since up is not constant for the DRM-based PMA, it is required to derive
d j
from the definition of the updated reliability index, as shown in Eq. (4.15). In the current
design, since cur and t are constant, the sensitivity of the updated reliability index with
up DRM 1 PFDRM
. (5.29)
d j d j ( DRM ) d j
T
u*DRM U g DRM PFDRM
U g DRM . (5.30)
d j ( DRM ) d j
Using the transformation in Eq. (5.7) and U gDRM TGDRM , the sensitivity of the
G x G g PFDRM
T
The sensitivity of PFDRM with respect to dj in Eq. (5.31) can be calculated using
Eq. (4.10). Assuming a 2-D performance function for the simplicity of calculation,
74
G s (v ) G1s (v1 )
1 1 (v1 )dv1
b1 d j 1 b
(5.32)
G1s (v1 ) G1s (v1 ) (v1 )
b
b1 G1s (v1 ) 1
2 dv1
b1 d j b1d j
n
w G s (v k ) G s (v k ) b
2k 1 1 b1 1 1 G1s (v1k ) 1 .
k 1 b1 b1 d j d j
G1s (v1k ) (G ( vk ) G( v* ))
(5.33)
d j d j
v1k
1
x k
x x
k I I A I I . (5.34)
d v1 d x x* d x x*
1
T
G1s (v1k ) x G G
A A I (5.35)
d d x x* x x=xk x x=x*
x
where is given by Eq. (5.15). Inserting Eqs. (5.35) and (5.11) into Eq. (5.32), we
d x x*
have
75
PFDRM n
w G s (v k )
2k 1 1
d k 1 b1 b1
. (5.36)
x
T
HTT TG
k x
T
b1 A A I G (x ) G G1 (v1 )
k s
d x x* d b
xx
*
1
By inserting Eq. (5.36) into Eq. (5.31), the sensitivity of the probabilistic constraint in Eq.
(5.4) at the true MPP is analytically obtained. Unlike the sensitivity of the probabilistic
constraint at the FORM-based MPP given in Eq. (5.18), the sensitivity in Eq. (5.31) of
the probabilistic constraint at the true DRM-based MPP requires the Hessian as shown in
Eq. (5.36), and thus very expensive to use for PMA of RBDO
The DRM-based MPP obtained from the approximation in Eq. (4.16) with the
up * up g g
uaDRM u = cur U up U upα. (5.37)
cur cur U g U g
uaDRM up U g U g
up . (5.38)
d j d j U g d j U g
T T
x TT TG PFDRM
ej
T
d j x=xaDRM b1 ( DRM ) d j
T
(5.39)
up x
b12 TT TH TT TG T GTT TH .
T
b d j *
3
1 x=x
76
G x G
T
, (5.40)
d x=xaDRM d x=xaDRM x x=xaDRM
analytically calculated. However, since the sensitivity requires the Hessian matrix as
shown in Eq. (5.38) like Eq. (5.36), it is also very expensive to obtain the sensitivity in
Eq. (5.40).
GDRM
a
G
Using the assumption where GDRM
a
G(xaDRM ) , the equation
GDRM
a
G
x G U g PFDRM
T
d a x * G (5.41)
x=xDRM x=x ( DRM ) d
G x G
T
d x=xaDRM d x=xaDRM x x=xaDRM
(5.42)
x
T
GDRM
a
U g DRM
a
PFDRM
G GDRM
a
d x=xaDRM G ( DRM ) d
PFDRM
where the same in Eq. (5.36) is used for Eq. (5.42). This is the sensitivity of the
d
As the design approaches the optimum design, the updated reliability index βup
converges to the current reliability index βcur. This is because Δβ is getting smaller as the
design approaches the optimum. Hence, the sensitivity of the updated reliability index
with respect to dj converges to zero as the design approaches the optimum because
77
up
cur . (5.43)
d j d j d j
up PFDRM
In conclusion, even though the analytic sensitivities for and are given in
d j d j
Eqs. (5.29) and (5.36), they can be ignored because the sensitivities are very small near
the optimum and the Hessian matrix is required for the analytic sensitivities.
Using Taylor series expansion, the gradient of the performance function at the
G( )
G( ) G( ) O( ) (5.44)
and using Eq. (5.42), the sensitivity of the performance function at the approximated
G x G x G
T T
G O ( )
d x=xaDRM d x=xaDRM x x=xaDRM d x=xaDRM
(5.45)
U g PFDRM x G
T
G O ( ) .
( DRM ) d j d x=xaDRM
As the design approaches the optimum, which means Δβ converges to zero, Eq. (5.45) is
approximated as
G
G . (5.46)
d x=xaDRM
the gradient at the FORM-based MPP as shown in Eq. (5.46), the gradient at the true
DRM-based MPP will converge to the approximated DRM-based MPP and FORM-based
78
MPP. Hence, without loss of accuracy near the optimum, the sensitivity of the
probabilistic constraint for the DRM-based RBDO in Eq. (5.4) can be approximated as
G G
(5.47)
d x=xaDRM x x=xaDRM
to save the computational cost for RBDO. However, if the initial design is far from the
RBDO optimum, there could be some error in the approximation of the sensitivities in
Eq. (5.47), especially when the performance function is highly nonlinear. To avoid this
situation, PMA+ is used for RBDO, which uses the deterministic optimum design as the
initial design for RBDO because the deterministic optimum design is usually close to the
RBDO optimum design. The RBDO examples using PMA+ and analytical sensitivities in
Numerical studies are carried out in this section to verify the analytic sensitivities
derived in Sections 5.2 and 5.3 using the FDM with various perturbation sizes. For that
Refs. (Lee et al., 2006; Lee et al., 2008a; Lee et al., 2008b), is used. Analytic sensitivities
derived for the FORM-based RBDO are compared with the FDM results in Section 5.4.1,
and analytic sensitivities derived for the DRM-based RBDO using the true and
approximated DRM-based MPP are compared with the FDM results in Section 5.4.2.
Section 5.4.3 illustrates how the sensitivity at the approximated DRM-based MPP shown
in Eq. (5.47) converges to the sensitivity at the true DRM-based MPP in Eq. (5.31) as the
Since the analytic sensitivities derived in this study do not require additional
function evaluations unlike sensitivities using FDM, the analytic calculation is always
significantly more efficient than the FDM. Hence, the efficiency comparison between the
analytic sensitivity obtained from Eq. (5.18), which was listed in the second column and
labeled “Analytic,” and the sensitivities obtained from FDM with various perturbation
sizes listed in the subsequent columns. From the table, it can be shown that sensitivities
G
-0.7082 -0.7370 -0.7141 -0.7111 -0.7088 -0.7085 -0.7081
d 2 MPP
80
As demonstrated in Table 5.3, this agreement between the FDM results and the
analytic results is also very good for a multi-dimensional function, which is one of the
where the properties of the random variables are listed in Table 5.2.
Std Distr.
Random Variable
Dev. Type dL d dU
1. B-pillar inner (mm) 0.050 Normal 0.500 0.5318 1.500
2. B-pillar reinforce (mm) 0.050 Normal 0.450 1.3500 1.350
3. Floor side inner (mm) 0.050 Normal 0.500 1.5000 1.500
4. Cross member (mm) 0.050 Normal 0.500 1.4261 1.500
5. Door beam (mm) 0.050 Normal 0.875 1.4718 2.625
6. Door belt line (mm) 0.050 Normal 0.400 1.2000 1.200
7. Roof rail (mm) 0.050 Normal 0.400 0.4000 1.200
8. Mat. B-pillar inner (GPa) 0.006 Uniform 0.192 0.3450 0.345
9. Mat. Floor side inner (GPa) 0.006 Uniform 0.192 0.1920 0.345
10. Barrier height (mm) 10.00 Uniform 10th and 11th variables are
11. Barrier hitting (mm) 10.00 Uniform not design variables
G
-0.8972 -0.8968 -0.8971 -0.8972
d 6 MPP
G
-0.5410 -0.5409 -0.5409 -0.5410
d 7 MPP
81
Consider the same performance function in Eq. (5.48), and the properties of
random variables are X1 ~ N (4,0.3) , X 2 ~ N (3,0.3) , and t 2 . Table 5.4 compares the
analytic sensitivity of the DRM-based probability of failure and the FDM results. The
second column, labeled “Analytic,” indicates the sensitivity obtained from Eq. (5.36), and
the subsequent three columns indicate the FDM results with three different perturbation
sizes. As illustrated in Table 5.4, both sensitivity results agree very well; however, the
magnitude is very small and the value will be smaller as the design approaches the
optimum, which will be shown in Section 5.4.3. This is the reason we can approximate
Tables 5.5 and 5.6 compare the FDM and analytic sensitivities of the probabilistic
constraint at the true and approximated DRM-based MPP, respectively. The second
columns of the tables indicate the sensitivities obtained from Eqs. (5.31) and (5.42),
respectively. Table 5.5 demonstrates the good agreement between the analytic and FDM
sensitivities. However, the agreement between the analytic and FDM sensitivity in Table
82
5.6 is not very good, which is attributed to the assumption that the direction of the
gradients at the FORM-based MPP and approximated DRM-based MPP is the same.
Again, the inaccuracy in Table 5.6 will disappear as the design approaches the optimum
design.
G
-0.4996 -0.5059 -0.5002 -0.4997
d 2 x=x*DRM
G
-0.4566 -0.5045 -0.4989 -0.4983
d 2 x=xaDRM
83
80
G3 ( X) 1
X 8X 2 5
1
2
d L [0, 0]T and d U [10,10]T , dinitial [4,3]T , X i ~ N (di , 0.3) for i =1,2
Table 5.7 illustrates the current design in the second column, the updated
reliability index in the third column, the probability of failure by DRM in the fourth
column, and the sensitivities at the FORM-based MPP, approximated DRM-based MPP,
and true DRM-based MPP of the second constraint in the subsequent columns,
respectively. The sensitivity at the true DRM-based MPP in the seventh column is
obtained by carrying out a new MPP search at the current design with the updated
reliability index in the third column. The last column shows the sensitivity
up U g DRM PFDRM
b1 , which is used in Eqs. (5.31) and (5.42). From Table 5.7, we
d1 ( DRM ) d1
up
can see that b1 is a very small value and converges to zero as the design approaches
d1
the optimum, which is the reason we can ignore the term. In addition, both sensitivities at
the true and approximated DRM-based MPP converge to the sensitivity at the FORM-
84
based MPP. At the optimum design, the probability of failure by DRM should be the
target probability of failure, which is PFTar (2) 2.2750 . However, due to the
These numerical results indicate that the sensitivities of the FORM-based PMA in
Eq. (5.18) and DRM-based PMA in Eq. (5.47) are suitable to use for the gradient-based
design optimization. Furthermore, the sensitivity in Eq. (5.47) for the DRM-based PMA
for RBDO is very effective because it does not require an additional MPP search and the
second-order derivatives as it shows very good accuracy near the optimum design.
85
CHAPTER VI
ANALYSIS
6.1 Introduction
The new algorithm for DRM-based RBDO will be explained in Section 6.2. As
explained in Section 4.3 using Eq. (4.13), the number of function evaluations for the
DRM-based RBDO increases as the number of design variables increases. To reduce the
number of function evaluations, two numerical strategies, which are integrated with
PMA+, will be explained in Section 6.3. The proposed DRM-based RBDO is explained
using numerical examples in Section 6.4. Furthermore, RBRDO combined with the
minimize Cost(d)
subject to Gi (x*DRM ) 0, i 1, , nc (6.1)
d L d dU , d R ndv and X R nrv
The detailed algorithm of the proposed DRM-based RBDO in Eq. (6.1) is shown in
Figure 6.1. Note that, if the DRM-based MPP identified the probabilistic constraints in
Eq. (6.1) are very close to linear so that the reliability index update in Eq. (4.15) is not
required, then the FORM-based RBDO is used during the design optimization. A
reliability analysis using the MPP-based DRM will be again used at the FORM-based
86
optimum design to verify whether the FORM-based optimum is the true optimum or not.
This procedure will reduce the computational cost and at the same time give users
probability of failure using DRM will increase as the number of design variables
increases. However, certain design variables may not affect some performance functions
even if they affect other performance functions. In that case, it is proposed to use a
reduced rotation matrix to reduce the number of required function evaluations identified
in Eq. (4.13). In addition, since the initial MPP is searched using FORM, the enhanced
hybrid mean value (HMV+) is used in this paper for efficient inverse reliability analysis,
and for the design optimization problem in Eq. (6.1), the enriched performance measure
approach (PMA+) is used for efficiency, which includes three key ideas: launching
activeness, and design closeness. In this study, new tolerances for constraint activeness
are introduced for additional numerical efficiency of DRM-based RBDO. Finally, the
deterministic optimization with shifted constraints (Wu et al., 2001) is used. This section
describes two new strategies: the reduced rotation matrix and new tolerances for
constraint activeness.
space to V-space. This rotated standard normal variable v is used to compute the
probability of failure in Eq. (4.11). If the random variable X i does not affect the
g
ui* 0 and 0 (6.2)
ui u=0
88
then Gis (vi1 ) Gis (vin ) Gis (vi* ) 0 since vi is a function of ui (i.e., function of xi )
only and xi does not affect the performance function. Hence, the integral along i th axis
can be expressed as
n
Gis (vij )
j 1
w j (
b1
) ( ) (6.3)
If there are N 0 random variables that do not affect the performance function, then Eq.
N 1 n Ne 1 n
Gis (vij ) Gis (vij )
w j (
i 1 j 1 b1
) w j (
i 1 j 1 b1
)
PFDRM (6.4)
( ) N 2
( ) Ne 2
reduced number of random variables does not change the probability of failure as shown
in Eq. (6.4), we can use Ne Ne rotation matrix, which will reduce the number of
(n 1)( Ne 1) , in addition to the FORM-based MPP search. Hence, the total number of
The reduced rotation matrix, which has full rank, can be generated using the Gram-
Schmidt orthogonalization.
This reduced rotation matrix strategy is useful when the problem is multi-
dimensional and contains a number of constraints. This is because some design variables
that affects some constraints may not affect other constraints. The computational
89
efficiency obtained by the reduced rotation matrix is demonstrated in Section 6.4 using
The PMA+ provides an efficient feasibility identification using the mean value
(MV) method as explained in Eqs. (2.15) and (2.16), which does not require the MPP
the function value at the MV-based MPP obtained using Eq. (2.16). After the feasibility
search is carried out using HMV+. However, a single tolerance f may not be effective
to identify the feasibility of all constraint functions, since the magnitude of constraint
Gi
Gi ( X*MV ) 2
f 0 (6.6)
N
Gi 2
dimensionality of the norm. However, in case that is large, the feasibility
N
identification using Eq. (6.6) may be too conservative, which makes the constraint
Gi Gi
Gi ( X*MV ) 2
f 0, if 2
1
N N (6.7)
Gi ( XMV ) f 0,
*
otherwise
One of the constraint functions from the automotive side impact problem (Youn et
al., 2004)
is used to test the effectiveness of the reduced rotation matrix. Since only six random
variables out of eleven are included in the constraint function, 6 6 rotation matrix is
used to compute the probability of failure in Eq. (4.11). This reduced rotation matrix
(3 1) (6 1) 10 when three quadrature points are used, while maintaining the
accuracy as shown in Table 6.1. The reliability index of 2 is used for the FORM-
DRM with
FORM MCS
Reduced Matrix
PF , % 2.2750 2.4317 2.4257
F.E. 3* 3*+10** 1 million
* 3 means the number of function and sensitivity analysis for MPP search.
minimize C (d)
subject to P(Gi ( X(d)) 0) PFTar
i
, i 1, , nc
d L d dU , d R ndv and X R nrv
(d1 d 2 10) 2 (d1 d 2 10) 2
where C (d)
30 120
2
X X
G1 (X) 1 1 2 (6.9)
20
G2 ( X) 1 (Y 6) 2 (Y 6)3 0.6 (Y 6) 4 Z
80
G3 (X) 1
X 8X 2 5
1
2
Y 0.9063 0.4226 X 1
where , and the target probability of failure for each
Z 0.4226 0.9063 X 2
constraint is PFTar
i
5.0%, i=1~3 . Since the given target probability of failure is 5.0%,
iteration the reliability index is updated using the DRM-based probability of failure.
Figure 6.2 (a) shows the approximated feasible region for PMA+ with the
tolerance f 0.5 and Figure 6.2 (b) shows the approximated feasible region when the
new tolerance for constraint activeness in Eq. (6.7) is used with the same f 0.5 . From
Figure 6.2 (a), we can see that even though the true MPP is far from the third constraint,
GpMV
3
(X(d)) f 0.26 0.5 0 (6.10)
92
as shown in the initial design of RBDO in Table 6.2, whereas PMA+ with the new
tolerance identified it as inactive. Table 6.2 shows the history of DRM-based RBDO with
three quadrature points when new tolerances for constraint activeness are used.
In this example, we can see that f 0.5 is rather large for the third constraint
compared to the other two constraints. In Figure 6.2 (b), since the third constraint is
G3
G3 ( X*MV ) 2
f 0 (6.11)
2
The MPP search for the third constraint is not carried out at the deterministic optimum
design with shift, which saves the number of function evaluations from 109+73 to 83+59
as shown in Table 6.3. Table 6.3 also shows that even though the optimum design of
FORM-based RBDO seems to be close to the optimum design of DRM-based RBDO, the
for the highly nonlinear second constraint. For the second constraint, three quadrature
points may not be sufficient to detect the nonlinearity of the constraint. In this case, five
quadrature points can be used to enhance the accuracy as shown in Table 6.3.
Table 6.3. Various RBDO Results with Target Probability of Failure PFTar
i
5.0%
Optimum MCS
Method Cost # of F.E.
Design PF1 , % PF2 , %
FORM* -1.77 4.580,1.863 5.8128 2.5794 103+103
3 pts -1.77 4.682,1.849 4.9857 3.8030 109+73
DRM**
5 pts -1.78 4.717,1.833 4.9616 4.5010 129+69
3 pts -1.77 4.682,1.849 4.9857 3.8030 83+59
DRM***
5 pts -1.78 4.717,1.833 4.9616 4.5010 97+57
* means FORM without PMA+.
*** means DRM with PMA+ and new tolerances for constraint activeness.
94
Since the DRM-based RBDO updates the reliability index of active constraints at
each design iteration, the reliability index at the optimum design will be different from
in Table 6.4, when DRM with five quadrature point is used, the reliability index for the
second constraint at the optimum design is 1.379, which is significantly reduced from the
initial reliability index since the second constraint is highly nonlinear and concave near
MPP, and the reliability index for the first constraint at the optimum is 1.715, which is
95
increased slightly since the first constraint is mildly nonlinear and convex near MPP as
Consider the vehicle side impact crashworthiness (Youn et al., 2004) shown in
Figure 6.4. The design objective is to minimize vehicle weight while enhancing the side
minimize Cost(d)
subject to P(Gi ( X(d)) 0) PFTar
i
, i 1, , nc (6.12)
d L d dU
Std Distr.
Random Variable
Dev. Type dL d dU
1. B-pillar inner (mm) 0.050 Normal 0.500 1.000 1.500
2. B-pillar reinforce (mm) 0.050 Normal 0.450 1.000 1.350
3. Floor side inner (mm) 0.050 Normal 0.500 1.000 1.500
4. Cross member (mm) 0.050 Normal 0.500 1.000 1.500
5. Door beam (mm) 0.050 Normal 0.875 2.000 2.625
6. Door belt line (mm) 0.050 Normal 0.400 1.000 1.200
7. Roof rail (mm) 0.050 Normal 0.400 1.000 1.200
8. Mat. B-pillar inner (GPa) 0.006 Normal 0.192 0.300 0.345
9. Mat. Floor side inner (GPa) 0.006 Normal 0.192 0.300 0.345
10. Barrier height (mm) 10.00 Normal 10th and 11th variables are
11. Barrier hitting (mm) 10.00 Normal not design variables
of 11 are regarded as design variables. Tables 6.6 and 6.7 are design and constraint
history obtained using the FORM-based RBDO and PMA+ with new tolerances for
constraint activeness. The shadowed cells in Table 6.7 show the active constraints at each
design iteration.
Table 6.6. Design History for Side Impact Example Using FORM-Based RBDO
Cost d1 d2 d3 d4 d5 d6 d7 d8 d9 NFE
Initial 30.83 1.000 1.000 1.000 1.000 2.000 1.000 1.000 0.300 0.300 0+0
D.O.* 24.23 0.500 1.226 0.500 1.207 1.238 1.200 0.400 0.345 0.300 8+8
0,1** 24.23 0.500 1.226 0.500 1.207 1.238 1.200 0.400 0.345 0.300 26+26
0,2 26.20 0.520 1.350 0.500 1.377 1.443 1.200 0.400 0.345 0.192 37+37
0,3 26.15 0.519 1.347 0.500 1.372 1.438 1.200 0.400 0.345 0.195 41+41
1,1 26.23 0.510 1.350 0.500 1.391 1.453 1.200 0.400 0.345 0.192 45+45
1,2 26.22 0.512 1.349 0.500 1.387 1.450 1.200 0.400 0.345 0.193 49+49
1,3 26.21 0.513 1.349 0.500 1.385 1.449 1.200 0.400 0.345 0.193 53+53
Opt. 26.20 0.511 1.350 0.500 1.384 1.451 1.200 0.400 0.345 0.192 57+57
* D.O. means deterministic optimum.
Table 6.7. Constraint History for Side Impact Example Using FORM-Based RBDO
G1 G2 G3 G4 G5 G6 G7 G8 G9 G10
Initial 0.04 -2.45 -1.23 -1.65 -1.13 -0.15 0.04 -0.85 -0.61 -2.23
D.O. -0.00 -1.89 -0.78 -1.23 -0.97 -0.08 -0.00 -0.82 0.00 -2.89
0,1 0.09 -0.47 -0.09 -0.87 -0.81 -0.05 0.10 -0.31 0.22 -2.30
0,2 -0.00 -1.43 -0.65 -1.04 -0.86 -0.05 0.00 -0.59 0.00 -3.09
0,3 -0.00 -1.41 -0.63 -1.04 -0.86 -0.05 0.00 -0.59 0.01 -3.07
1,1 0.00 -1.40 -0.63 -1.03 -0.86 -0.05 -0.00 -0.59 -0.00 -3.13
1,2 0.00 -1.40 -0.63 -1.03 -0.86 -0.05 -0.00 -0.59 -0.00 -3.12
1,3 0.00 -1.40 -0.63 -1.03 -0.86 -0.05 -0.00 -0.59 0.00 -3.11
2,1 0.00 -1.40 -0.63 -1.03 -0.86 -0.05 0.00 -0.59 -0.00 -3.11
Opt. Act. Inact. Inact. Inact. Inact. Inact. Act. Inact. Act. Inact.
Table 6.7 shows that three constraints, G1, G7, and G9, are active, which means
that the failure probability of three constraints using FORM is the target probability of
98
failure, PFTar
i
5.0% . However, Table 6.8 illustrates that true failure probability for each
constraint obtained using MCS is not the same as the target probability, especially, for G7
and G9, which means FORM has significant error in estimation of failure probability for
those two constraints. Hence, for more accurate RBDO results, the DRM-based RBDO is
MCS at Optimum
PFTar , %
PF1 , % PF7 , % PF9 , %
5.000 5.0075 11.4905 2.5117
Tables 6.9 and 10 demonstrates design and constraint history obtained using the
DRM-based RBDO with three quadrature points and PMA+ with new tolerances for
constraint activeness. The shadowed cells in Table 6.10 show the active constraints,
which is identified using PMA+ with new tolerances, at each design point. As illustrated
in Table 6.11, the failure probability estimation at the optimum of the DRM-based RBDO
is closer to the target probability of failure than the FORM-based RBDO optimum.
However, there exist some errors in estimation of the probability of failure using
DRM with three quadrature points. This can be resolved if five quadrature points are used
for the estimation of failure probability. Table 6.12 compares various RBDOs. From
Table 6.12, it can be seen that new tolerances for constraint activeness reduce the number
of function evaluations for both FORM and DRM-based RBDO considerably. This is
because the sixth constraint is identified as inactive using the new tolerance because of its
small sensitivity. Whereas, using PMA+ with the fixed tolerance f 0.5 , the sixth
99
DRM-based RBDO shows indeed different optimum design, especially d4 and d5,
compared with the FORM-based RBDO because of the highly nonlinear active
Table 6.9. Design History for Side Impact Example Using DRM-Based RBDO
Cost d1 d2 d3 d4 d5 d6 d7 d8 d9 NFE
Initial 30.83 1.000 1.000 1.000 1.000 2.000 1.000 1.000 0.300 0.300 0+0
D.O.* 24.23 0.500 1.226 0.500 1.207 1.238 1.200 0.400 0.345 0.300 8+8
0,1** 24.23 0.500 1.226 0.500 1.207 1.238 1.200 0.400 0.345 0.300 100+32
0,2 26.28 0.519 1.350 0.500 1.412 1.411 1.200 0.400 0.345 0.192 142+44
0,3 26.23 0.518 1.347 0.500 1.407 1.406 1.200 0.400 0.345 0.195 179+51
1,1 26.53 0.509 1.350 0.500 1.488 1.405 1.200 0.400 0.345 0.192 217+59
1,2 26.47 0.511 1.349 0.500 1.472 1.406 1.200 0.400 0.345 0.193 254+66
1,3 26.44 0.511 1.349 0.500 1.465 1.406 1.200 0.400 0.345 0.193 291+73
2,1 26.43 0.510 1.350 0.500 1.462 1.405 1.200 0.400 0.345 0.192 328+80
3,1 26.43 0.510 1.350 0.500 1.464 1.405 1.200 0.400 0.345 0.192 364+86
3,2 26.43 0.510 1.350 0.500 1.463 1.405 1.200 0.400 0.345 0.192 400+92
Opt. 26.43 0.510 1.350 0.500 1.463 1.405 1.200 0.400 0.345 0.192 436+98
Table 6.10. Constraint History for Side Impact Example Using DRM-Based RBDO
G1 G2 G3 G4 G5 G6 G7 G8 G9 G10
Initial 0.04 -2.45 -1.23 -1.65 -1.13 -0.15 0.04 -0.85 -0.61 -2.23
D.O. -0.00 -1.89 -0.78 -1.23 -0.97 -0.08 -0.00 -0.82 0.00 -2.89
0,1 0.09 -0.47 -0.09 -0.87 -0.81 -0.05 0.12 -0.31 0.19 -2.30
0,2 -0.00 -1.43 -0.65 -1.05 -0.87 -0.05 0.02 -0.61 -0.00 -3.19
0,3 -0.00 -1.40 -0.64 -1.05 -0.87 -0.05 0.03 -0.60 0.00 -3.17
1,1 0.00 -1.39 -0.63 -1.04 -0.86 -0.05 -0.01 -0.63 -0.00 -3.43
1,2 0.00 -1.39 -0.63 -1.04 -0.86 -0.05 -0.00 -0.63 -0.00 -3.37
1,3 0.00 -1.39 -0.63 -1.04 -0.86 -0.05 -0.00 -0.62 -0.00 -3.35
2,1 0.00 -1.39 -0.63 -1.04 -0.86 -0.05 -0.00 -0.62 -0.00 -3.35
3,1 0.00 -1.39 -0.63 -1.04 -0.86 -0.05 -0.00 -0.62 0.00 -3.35
3,2 0.00 -1.39 -0.63 -1.04 -0.86 -0.05 -0.00 -0.62 0.00 -3.35
3,3 0.00 -1.39 -0.63 -1.04 -0.86 -0.05 -0.00 -0.62 0.00 -3.35
Opt. Act. Inact. Inact. Inact. Inact. Inact. Act. Inact. Act. Inact.
100
MCS
PFTar , %
PF1 , % PF7 , % PF9 , %
5.000 5.0270 5.5903 5.4002
DRM-based RBDO. The roadarm is modeled using 1572 eight-node isoparametric finite
elements (SOLID45) and four beam elements (BEAM44) of Ansys (Swanson Analysis
System Inc., 1989), as shown in Figure 6.5, and is made of S4340 steel with Young’s
modulus E=3.0×107 psi and Poisson’s ratio ν=0.3. The durability analysis of the roadarm
is carried out using Durability and Reliability Analysis Workspace (DRAW) (CCAD,
101
1999a-b), to obtain the fatigue life contour as shown in Figure 6.6. The fatigue lives at
the critical nodes shown in Figure 6.6 are chosen as the design constraints of RBDO.
The shape design variables are shown in Figure 6.7. Eight shape design variables
characterize four cross sectional shapes of the roadarm. Widths ( x1 direction ) of the
cross-sectional shapes are defined by the design variables d1, d2, d5, and d6 at the
shapes are defined using the remaining four design variables. Eight shape design random
variables and six random variables for the fatigue material properties are listed in Table
6.13.
minimize Cost(d)
subject to P(Gi (d) 0) PFTar
i
, i 1, , nc (6.14)
d L d dU
where
results are shown in Tables 6.14 and 6.15. After finding the deterministic optimum design
104
first, the DRM-based RBDO is launched starting at the deterministic optimum design for
the active constraints only, which require more accurate DRM-based failure probability
estimation. The shadowed cells in Table 6.15 indicate the active constraints at each
design point. After performing the DRM-based inverse reliability analysis at the
deterministic optimum design, all fatigue life constraints in this problem turn out to be
very close to linear. Hence, the FORM-based RBDO is carried out and the DRM-based
Cost d1 d2 d3 d4 d5 d6 d7 d8 NFE
Initial 515.09 1.750 3.250 1.750 3.170 1.756 3.038 1.752 2.908 0+0
D.O. 464.56 1.588 2.650 1.922 2.570 1.476 3.292 1.630 2.508 11+11
0,1 464.56 1.588 2.650 1.922 2.570 1.476 3.292 1.630 2.508 206+24
0,2 490.33 1.789 2.650 1.993 2.570 1.587 3.480 1.871 2.508 220+38
0,3 474.79 1.669 2.650 1.951 2.570 1.521 3.367 1.727 2.508 234+52
0,4 476.11 1.679 2.650 1.954 2.570 1.527 3.377 1.739 2.508 248+66
1,1 476.28 1.681 2.650 1.954 2.570 1.541 3.365 1.730 2.508 262+80
1,2 476.27 1.681 2.650 1.954 2.570 1.541 3.366 1.730 2.508 276+94
Opt. 476.21 1.681 2.650 1.954 2.570 1.541 3.366 1.729 2.508 394+108
G1 G2 G3 G4 G5 G6 G7 G8 G9 G 10 G 11 G 12 G 13
Initial -11.2 -17.1 -12.8 -312 -836 -115 -96.8 -394 -426 -468 -542 -55.1 -45.6
D.O. 0.00 -1.03 0.00 -6325 0.00 -115 -159 0.00 -1.06 -3.92 -0.89 0.00 -0.75
0,1 0.75 0.47 0.73 -999 0.72 -23.4 -32.7 0.73 0.45 -0.22 0.50 0.82 0.71
0,2 -5.44 -5.93 -2.61 -431 -4.93 -85.4 -81.2 -2.11 -3.25 -35.7 -11.8 -9.89 -10.1
0,3 0.13 -0.48 0.23 -692 0.17 -40.0 -47.5 0.27 -0.27 -3.59 -0.75 0.04 -0.35
0,4 -0.02 -0.69 0.13 -663 0.03 -42.8 -49.8 0.17 -0.41 -4.46 -1.06 -0.19 -0.63
1,1 -0.00 -0.87 -0.01 -581 -0.00 -42.0 -47.3 -0.02 -0.70 -4.66 -1.14 -0.02 -0.36
1,2 -0.00 -0.86 -0.01 -584 -0.00 -42.0 -47.3 -0.01 -0.69 -4.65 -1.13 -0.02 -0.37
2,1 -0.00 -0.86 0.00 -586 -0.00 -42.0 -47.4 -0.00 -0.67 -4.64 -1.13 -0.00 -0.33
Opt. Act. Inact. Act. Inact. Act. Inact. Inact. Act. Inact. Inact. Inact. Act. Inact.
105
Table 6.16 shows the comparison of optimum designs obtained using two
methods. To verify the optimum result of the proposed DRM-based RBDO method, the
full DRM-based RBDO method, which does not take advantage of the FORM-based
RBDO during the optimization, is used because MCS cannot be used due to its
computational cost. As shown in Table 6.16, result of the proposed DRM-based RBDO
using the algorithm in Figure 6.1 is very close to the benchmark results since all fatigue
life constraints are linear, and yet reduced the number of function evaluations
significantly. Furthermore, the proposed method gives users confidence that the optimum
Since RDO can be naturally combined with RBDO for the constraint evaluation,
which becomes RBRDO, this section deals with two numerical examples for RBRDO.
For the cost function, which is variance of the performance function, PMI and mean-
based DRM, which are explained in Chapter 3, will be used and, for the constraint
evaluation, the FORM-based RBDO and DRM-based RBDO will be used. A 2-D
numerical example for RBRDO is used in Section 6.5.1 and side impact problem is used
in Section 6.5.2.
minimize H2
(6.16)
subject to P(Gi ( X) 0) PFTar
i
, i 1, , nc
where the performance function for robustness and constraints are, respectively,
( X 1 X 2 6) 2 ( X 1 X 2 6) 2
h( X)
3 12
2
X X
G1 ( X) 1 1 2
20 (6.17)
G2 ( X) 1 (Y 6) 2 (Y 6)3 0.6 (Y 6) 4 Z
80
G3 ( X) 1
X 8X 2 5
1
2
Y 0.9063 0.4226 X 1
where , and the target probability of failure for each
Z 0.4226 0.9063 X 2
constraint is PFTar
i
(2), i=1~3 . Two random variables are listed in Table 6.17.
107
Table 6.18. Variance Estimation Using PMI at Initial and Optimum Design
Table 6.19. Variance Estimation Using DRM at Initial and Optimum Design
Tables 6.18 and 6.19 are the cost values at the initial and optimum designs using
PMI and the mean-based DRM, respectively. When the MPP-based DRM is used for the
probabilistic constraint evaluation, the cost at the optimum is slightly increased compared
to when FORM is used. This is because the first constraint is convex near MPP at the
optimum design and thus the reliability index at the optimum design is increased as
shown in Figure 6.8. Figure 6.8 (a) and Figure 6.8 (b) show the RBRDO optimum
designs when the MPP-based DRM and FORM are used for the constraint evaluation,
respectively. Tables 6.20 and 6.21 show the optimum design and active constraints.
Because of the highly nonlinear active constraint G2 , optimum designs obtained using
108
two different constraint evaluation methods, FORM and MPP-based DRM, are indeed
different.
Variance
d1 d2 G1 G2 G3
Calculation
DRM 4.5356 2.0947 -0.00 -0.00 -0.45
PMI 4.5357 2.0941 -0.00 -0.00 -0.45
Variance
d1 d2 G1 G2 G3
Calculation
DRM 4.6552 2.0719 -0.00 -0.00 -0.42
PMI 4.6226 2.0914 -0.00 -0.04 -0.43
109
The RBRDO model of crashworthiness of the vehicle side impact which is used
minimize H2
(6.18)
subject to P(Gi ( X) 0) PFTar
i
, i 1, , nc
where nc is 10, the target probability of failure for each constraint is PFTar
i
(2) , and
the performance function h(x) for robustness is the upper rib deflection given by
h(X) 3.818 X 3 4.2 X1 X 2 0.0207 X 5 X10 6.63 X 6 X 9 7.7 X 7 X 8 0.32 X 9 X10 (6.19)
11 random variables for the side impact problem are listed in Table 6.22. Note that input
Table 6.22. Properties of Design and Random Parameters for Side Impact Problem
Std Distr.
Random Variable
Dev. Type dL d dU
1. B-pillar inner (mm) 0.100 Normal 0.500 1.000 1.500
2. B-pillar reinforce (mm) 0.100 Normal 0.450 1.000 1.350
3. Floor side inner (mm) 0.100 Normal 0.500 1.000 1.500
4. Cross member (mm) 0.100 Normal 0.500 1.000 1.500
5. Door beam (mm) 0.100 Normal 0.875 2.000 2.625
6. Door belt line (mm) 0.100 Normal 0.400 1.000 1.200
7. Roof rail (mm) 0.100 Normal 0.400 1.000 1.200
8. Mat. B-pillar inner (GPa) 0.006 Normal 0.192 0.300 0.345
9. Mat. Floor side inner (GPa) 0.006 Normal 0.192 0.300 0.345
10. Barrier height (mm) 10.000 Normal 10th and 11th random variables
are not regarded as design
11. Barrier hitting (mm) 10.000 Normal
variables
Tables 6.23 and 6.24 show the cost at the initial and optimum designs using PMI
and the mean-based DRM. As shown in Tables 6.23 and 6.24, both methods estimated
110
variance very accurately compared to MCS. In this example, when the MPP-based DRM
is used for constraint evaluation, the optimum cost is smaller than when FORM is used.
Furthermore, because of the highly nonlinear constraint G9, optimum designs of two
different constraint evaluation methods, FORM and MPP-based DRM, are indeed
Variance
Constraint
Initial Design Optimum Design
Evaluation
PMI N.I. PMI N.I.
FORM 1.4010 1.4024
2.4831 2.4852
MPP-Based DRM 1.3878 1.3871
Table 6.24. Variance Using Mean-Based DRM at Initial and Optimum Design
Variance
Constraint
Initial Design Optimum Design
Evaluation
DRM N.I. DRM N.I.
FORM 1.4008 1.4024
2.4830 2.4852
MPP-Based DRM 1.3877 1.3871
Cost
G1 G2 G3 G4 G5 G6 G7 G8 G9 G10
Evaluation
DRM* -0.001 -2.757 -1.021 -2.031 -1.080 -0.167 -0.121 -0.527 -0.000 -2.532
PMI* -0.000 -2.754 -1.019 -2.029 -1.079 -0.167 -0.124 -0.528 -0.000 -2.544
DRM** -0.000 -2.759 -1.058 -1.993 -1.082 -0.161 -0.095 -0.544 -0.000 -2.523
PMI** -0.000 -2.759 -1.058 -1.993 -1.082 -0.161 -0.095 -0.544 -0.000 -2.523
112
CHAPTER VII
7.1 Introduction
Estimations of not only the component probability of failure but also the system
probability of failure have been the main concern in structural reliability analysis for over
three decades. According to the logical relationship of the failure modes of structures,
structural systems can be divided into three types: series system, parallel system, and
hybrid system (Zhao et al., 2007). The series system is also referred to weakest link or
chain system because the system failure is caused by the failure of any one component.
The parallel system is also referred to as a redundant system because the system fails
only if all components fail. The hybrid system is a mixed system of the series and parallel
system. In this study, the reliability analysis of the series system will be discussed since it
Since the analytic estimation of the system probability of failure involves multi-
dimensional integration over the overall failure domain, it is numerically very difficult to
evaluate. Hence, several approaches to resolve the numerical difficulty have been
proposed including the narrow bound estimation (Ditlevsen, 1979). For the narrow
bound method, Ditlevsen’s first order upper bound, which is the summation of
component failure probabilities, can be used as the system probability of failure (Ba-
abbad et al., 2006) or Ditlevsen’s second order upper bound by considering the joint
probability of failure can be used (Ang and Tang, 1984; Liang et al., 2007). However, if
FORM is used, these narrow bound methods will only work for linear or very mildly
113
Thus, more accurate system reliability analysis method is needed for the system
with nonlinear and multi-dimensional performance functions. In this study, the MPP-
based dimension reduction method (DRM) and Ditlevsen’s second order upper bound are
used to propose a system reliability analysis method. In addition, using the accurate
proposed.
Section 7.2 demonstrates the accurate system reliability analysis using the
Ditlevsen’s second order upper bound and MPP-based DRM. Section 7.3 illustrates the
system RBDO. For the system RBDO, sensitivity analyses are carried out and two
efficiency strategies are proposed to save the computational burden of the system RBDO.
When there are more than one performance function, the system probability of
m
PFsys P Gi ( X) 0 (7.1)
i 1
where m is the number of performance functions and the performance function is defined
as failure if Gi (X) 0 . However, since the right side of Eq. (7.1) is not easy to compute
Ditlevsen’s first-order upper bound (Ditlevsen, 1979) by the sum of the component
probabilities of failures as
114
m
P sys
F PFi (7.2)
i 1
where PFi is the component probability of failure for the ith performance function. A
more refined method to approximate the system probability of failure using Ditlevsen’s
m m
PFsys PFi max( PFij ) (7.3)
j i
i 1 i 2
where PFij is the joint probability of failure when the ith and jth failure modes occur
simultaneously. It is noted that the error in estimating the system probability of failure is
much more significant from the component probability of failure PFi than the correction
Since it is shown in Chapter 4 that FORM is not acceptable for the component
N 1 n
Gis (vij )
w j (
i 1 j 1 b1
)
PFDRM (7.4)
( ) N 2
which is derived in Section 4.3.1, is used for Eq. (7.3). Hence, this study proposes the
conservative but accurate system probability of failure calculation using the DRM-based
component probability of failure in Eq. (7.4) and Ditlevsen’s second order upper bound
in Eq. (7.3) as
115
m m
P sys
F P DRM
Fi max( PFij ) (7.5)
j i
i 1 i 2
which will be very accurate even for a highly nonlinear and/or multi-dimensional system.
In this study, Ditlevsen’s second order upper bound is used for the system reliability
analysis since it does not require further function evaluation. The joint probability of
Based on FORM, the joint probability of failure between the ith and jth
performance function in Eq. (7.5), PFij , is approximated as (Ang and Tang, 1984;, Zhao et
i j
PFij (i , j ; ij ) ( x, y; ij )dxdy (7.6)
1 1 x 2 y 2 2 xy
( x, y; ) exp (7.7)
2 1 2 2 1 2
n n
giL (u) i ir ur and g Lj (u) j js us (7.8)
r 1 s 1
where α i and α j are normalized vector from the origin to the MPP of each constraint
and u is the standard normal variable. Let the angle between α i and α j be as shown in
n
cos ir jr (7.9)
r 1
Cov[gi , g j ] Cov[gi , g j ]
ij Cov[gi , g j ] (7.10)
g g n n
i j
r 1
2
ir
s 1
2
js
Cov[gi , g j ] E[( gi gi )( g j g j )] E[ gi g j gi g j g j gi gi g j ]
E[ gi g j ] gi g j g j gi gi g j E[ gi g j ] gi g j
n n
E[( i ir ur ) ( j js us )] gi g j (7.11)
r 1 s 1
n n
i j ir jr gi g j ir jr
r 1 r 1
117
n
ij ir jr cos . (7.12)
r 1
2
. (7.13)
xy
Then, the derivative of the bivariate standard normal CDF with respect to ρ is
x y x
y
( , ; )d d
d d
x y x
y
x 2
y
d d d d d
(7.14)
x ( , y; )
d
( x, y; )
2
and ( x, y; ) is equal to by the definition of the bivariate standard normal CDF
xy
given as
x y
( x, y; ) ( , ; )d d (7.15)
2
(7.16)
xy
Using Eq. (7.16), the joint probability of failure in Eq. (7.6) can be expressed as (Liang et
al., 2007)
ij ( i , j ; )
PFij ( i , j ; ij ) ( i , j ;0) d
0 (7.17)
ij
( i ) ( j ) ( i , j ; ) d
0
118
Since Eq. (7.17) uses the linear approximation of the performance functions at
on convexity or concavity of the performance functions near MPPs. The definition of the
convexity is the same as the definition in Section 4.3.2, that is, if PFDRM PFFORM , then the
function is “convex” around the MPP and vice versa for “concave”. One example that
FORM overestimates the true joint probability of failure is shown in Figure 7.1.
As explained in Section 7.2.1, the MPP-based DRM is used for the accurate
component probability of failure in Eq. (7.4), and the joint probability of failure in Eq.
(7.17) is conservatively obtained using the type of the performance functions as the
following cases.
Case (a). Ignore the joint probability of failure if both constraints are concave because
the FORM-based joint probability of failure will overestimate the true failure
as shown in Figure 7.2 (a). If all constraints are concave, then the system
identical with the system probability of failure using Ditlevsen’s first order
upper bound.
Case (b). For a highly correlated case shown in Figure 7.2 (b), that is, 0.95 , then
Case (c). Otherwise, use the FORM-based joint probability of failure calculation in Eq.
(7.17) because it can approximate the true joint failure reasonably as shown in
Using the system inverse reliability analysis described above, the system RBDO
is formulated to
find d, β
min Cost(d)
s.t Gi (x* (d, i )) 0, i 1, , nc (7.18)
PFsys
G0 1 0
PFall
where d μ(X) is the mean value of the input random variable X, x* is the FORM-based
MPP, PFall is the allowable system probability of failure, and PFsys is the system
probability of failure calculated from Eq. (7.5). For the formulation in Eq. (7.18), no
MPP update, which is explained in Section 4.3.3, is used since the reliability indices are
Since d and β are both design variables for the formulation in Eq. (7.18), for
constraints at MPP with respect to d and β and the sensitivity of the system probability
G (x* )
of failure with respect to d and β . is identical with the FORM-based component
d
G (x* ) G
(7.19)
d x x x*
G (x* )
can be obtained using the chain rule as
u*
From the definition of MPP in U-space in Eq. (5.5), is obtained as
U g b
b1 U g 1
u g
*
U (7.21)
b1 b12
b1
where b1 U g and b12 UT gU g . Hence, is given as
b1 1 UT g
U g (7.22)
b1
u*
By substituting Eq. (7.22) into Eq. (7.21), can be obtained as
Since Ditlevsen’s second order upper bound is used for the system inverse
respect to the design variables involve two terms: sensitivity of the component
probability of failure with respect to the design variables, and sensitivity of the joint
probability of failure with respect to the design variables. The sensitivities of the joint
probability of failure with respect to the design variables can be analytically obtained
using Eq. (7.17). Since the joint probability of failure is a function of reliability indices
only, sensitivities of the joint probability of failure with respect to d is zero and
PFij ij j
( i )( j ) ( i , j ; ) i d (7.25)
i 1 2
0
The sensitivity of the component probability of failure with respect to d is derived in Eq.
PFDRM
(5.36) and the numerical example in Section 5.4.2 shows that is very small and
d
even smaller when the design approaches the optimum and, hence, the sensitivity is
approximated as
PFDRM
0. (7.26)
d
PFDRM
To derive , let us assume a two-dimensional performance function for the ease of
G s (v1 )
PFDRM ( ) (v1 )dv1 (7.27)
b1
PFDRM
Hence, for 2-D performance function can be obtained as
G s (v1k )
in Eq. (7.28) can be obtained using the definition of the shifted performance
G ( v* ) G (v1k )
is identical with Eq. (7.24) and is given by
u k
where is given by
v1k v1k
1
u* 2 *
0
uk
u (7.31)
v1k v1k
1 2 0
using the transformation from U-space to V-space. Eq. (7.23) can be rewritten as
123
u*
where H is the Hessian matrix evaluated at MPP in X-space. Hence, is obtained as
1
u* U g b2I U gUT g b12 H HU gUT g HU g
1 I (7.33)
b1 b13 b13 b1
and by inserting Eq. (7.33) into Eq. (7.31) and Eq. (7.31) into Eq. (7.30), finally, we can
obtain the sensitivity of the component probability of failure by the MPP-based DRM
calculate the sensitivity and the Hessian matrix is very difficult and numerically
(7.28) is approximated by
PFDRM G s (v1 )
(v1 )dv1 (7.34)
b1
G s (v1 )
assuming that is very small. The verification of the assumption is shown in
b1
Section 7.4.1 using numerical examples. Using the same assumption, the sensitivity of
the component probability of failure by the MPP-based DRM with respect to for a
PFDRM N 1 N 1
f k( ) (vk )dvk f j ( ) (v j )dv j ( )
k 1
j k
(7.35)
N 1
f j ( ) (v j )dv j ( )
j 1
where f j ( ) is defined as
124
G s (v )
j j
b1
f j ( ) (7.36)
( )
As can be seen in the formulation in Eq. (7.18), since the system RBDO involves
more design variables than the component RBDO, the system RBDO will take more
iteration to converge to the optimum. Hence, two efficiency strategies for the system
Theoretically, all constraints must be considered for the calculation of the system
probability of failure calculation. However, since some constraints may not contribute to
constraints for the system probability of failure calculation, it is necessary to find out
critical constraints which will contribute to the system failure. If RIA is used for the
system RBDO, then, the reliability indices can be used to identify the critical constraints
(Ba-abbad et al., 2006; Liang et al., 2007). However, since PMA is used in this study, it is
Based on PMA+ (Youn et al., 2005b), the system RBDO directly finds the
deterministic optimum, and active constraints at the deterministic optimum will most
probably affect the system failure. But, there is possibility that certain constraints, which
are not active at the deterministic optimum, may affect the system failure because the
system RBDO optimum design is away from the deterministic optimum design. Hence,
125
the new tolerance for constraint activeness explained in Section 6.3 is reused to identify
Gi Gi
Gi ( X*MV ) 2
f 0, if 2
1
N N (7.37)
Gi ( XMV ) f 0,
*
otherwise
After identifying the number of critical constraints denoted as mc, the system probability
mc mc
P sys
F P DRM
Fi max( PFij ) (7.38)
j i
i 1 i 2
Design closeness concept was first proposed in PMA+ for the FORM-based
component RBDO (Youn et al., 2005b). The design closeness concept is that the previous
MPP in U-space will be used as the current starting MPP if the current design is very
uk(0)i *
i
u*(k-1)i u*(k-1)i (7.39)
u ( k-1)i
where uk(0)i is the 0th MPP candidate point at the kth design iteration for ith constraint and
u*(k-1)i is the MPP at the (k−1)th design for ith constraint. In this case, since the reliability
However, since the reliability indexes are changing during the system RBDO
process, it is necessary to modify Eq. (7.39) to take advantage of the design closeness
concept. The modified design closeness is similar with the MPP update in DRM-based
126
RBDO. If two designs are very close, then the current starting MPP in U-space is
obtained as
ik
u (0)
ki *
u*(k-1)i (7.40)
u ( k-1)i
where ik is the reliability index at the kth design iteration. Using this new design
closeness concept, the number of function evaluation for MPP search can be reduced.
The analytic sensitivities derived in Section 7.3.2 are compared with the
sensitivities obtained using FDM. For the comparison, consider a highly nonlinear
the derived sensitivities comparing with FDM. In the table, the analytic sensitivities are
obtained using Eq. (7.24) and Eq. (7.28), respectively. From the table, it can be shown
the analytic sensitivities in Table 7.1, hence, it is numerically very expensive and
impractical. Table 7.2 compares the analytic sensitivity in Eq. (7.28) and approximate
sensitivity in Eq. (7.35) using the same performance function in Eq. (7.41). The
127
approximate sensitivity shows very good accuracy even if the performance function is
highly nonlinear. Furthermore, the approximate sensitivity does not require additional
function evaluation, which means that it is very efficient and accurate to use the
Methods
Using the approximate sensitivity in Eq. (7.35), the system RBDO with the MPP-
based DRM is carried out. In this section, two methods to identify critical constraints are
compared. One method is to use active constraints at the deterministic optimum as critical
constraints and the other method is the proposed method in Section 7.3.3 that uses the
128
MV method to identify critical constraint. For the comparison, the system RBDO for a
( X 1 2) 2 ( X 2 0.4)
G1 ( X) 1
20
G2 ( X) 1 (0.4226 X 1 0.9063 X 2 ) (0.9063 X 1 0.4226 X 2 6) 2
, (7.43)
(0.9063 X 1 0.4226 X 2 6)3 0.6 (0.9063 X 1 0.4226 X 2 6) 4
80
G3 ( X) 1
X 8X 2 5
1
2
X i ~ N (di ,0.3) for i =1,2 , dinitial [5,5]T , βinitial [2, 2, 2]T , and the allowable system
7.3. If the active constraint at the deterministic optimum is used as a critical constraint,
then mc=1, β is reduced to β [ 2 ] and β1, β3 are unchanged during the optimization. If
the MV method is used, G1(X) and G2(X) are identified as the critical constraints, hence,
in the first column of the table means that the active constraint at the deterministic
optimum is used as the critical constraint, whereas B means that the MV method is used
to identify the critical constraints. From the table, it can be shown that the method A
gives us very unreliable design because the system probabilities of failure in the last
129
column are significantly larger than the allowable system probability of failure. This
difficulty cannot be resolved even if the accurate MPP-based DRM with 5 integration
points (DRM5) is used. This is because the method A does not identify the first constraint
as a critical constraint, even though it affects the system failure as shown in Table 7.3.
DRM3 and DRM5 mean the MPP-based DRM with 3 and 5 integration points, respectively.
130
However, if the method B is used, then FORM and MPP-based DRM show better
results than the method A. Still, FORM even with the method B shows an unreliable
optimum design. This is because FORM has error in computation of the component
probability of failure, especially for highly nonlinear constraint G2(X), as shown in Table
7.3. Since G1(X) and G2(X) are both convex around MPP, the FORM-based joint
probability of failure should be used for the system probability of failure calculation. In
this problem, the FORM-based joint probability of failure is almost zero because ρ is
close to −1. Also, the joint probability of failure by Monte Carlo Simulation (MCS) is
very small as shown in Table 7.3. Thus, it does not affect the system RBDO for Eq.
(7.43).
MPP-Based DRM
Since the FORM-based joint probability of failure is very small in the previous
example, it was hard to see the effect of the joint probability of failure. To see the effect
of the joint probability of failure, consider the following system RBDO formulation to
( X 1 0.3) 2 X 2
G1 ( X) 1
20
G2 ( X) 1 (0.4226 X 1 0.9063 X 2 ) (0.9063 X 1 0.4226 X 2 6) 2
(0.9063 X 1 0.4226 X 2 6)3 0.6 (0.9063 X 1 0.4226 X 2 6) 4 , (7.45)
80
G3 ( X) 1
X 8X 2 5
1
2
( X 1 X 2 2.5) 2 ( X 1 X 2 4.5) 2
G4 ( X) 1
30 120
X i ~ N (di ,0.3) for i =1,2 , dinitial [2,7]T , βinitial [2, 2, 2, 2]T , and the allowable system
As shown in Figure 7.4, since G2(X) and G4(X) are concave around the MPP, the
joint probability of failure PF24 is ignored. In addition, since 12 is close to −1, the
132
FORM-based joint probability of failure PF12 is also almost zero. Hence, the joint
probability of failure PF14 is the only one which affects the system probability of failure
significantly. Using the MV method, G1(X), G2(X) and G4(X) are identified as the critical
constraints. Among these constraints, G1(X) is not active at the deterministic optimum.
Table 7.4 compares FORM and MPP-based DRM for the system RBDO in Eq.
(7.44). As expected from Figure 7.4, since G2(X) is highly nonlinear and concave around
the MPP, FORM overestimates the component probability of failure as shown in Table
7.4, which affects the system probability of failure significantly. However, MPP-based
DRM with 3 and 5 integration points can accurately estimate the highly nonlinear
constraint G2(X). Hence, both yield very good estimation of the system probability of
failure as shown in the last column of Table 7.4. The joint probability of failure PF14 by
MCS is 0.0183% for DRM5 and FORM-based joint probability of failure is 0.0227%
which is close to the MCS result. Both results are very small compared with the system
probability of failure, which means that the accurate component probability of failure
calculation is more important than the joint probability of failure calculation. However, it
133
should be noted that there may be a case that the joint probability of failure dominates,
CHAPTER VIII
8.1 Conclusions
(RBRDO)
Three methods (PMI, PDM, and mean-based univariate DRM) are compared in
terms of efficiency and accuracy for computation of the statistical moments and their
performance function, two polynomial performance functions with two design variables
are employed. In this comparison, PDM is excluded since PDM cannot estimate the
moments of the performance function. The comparison shows that DRM can accurately
estimate the statistical moments of the performance function for the design variables with
both non-normal and normal distributions. On the other hand, PMI can accurately
estimate the statistical moments of the performance function for the design variables with
normal distributions. For non-normally distributed design variables, PMI shows some
errors since nonlinear transformations make the performance function become highly
nonlinear.
For RBRDO, a highly nonlinear performance function was used for comparison
purposes. Both one-dimensional and two-dimensional examples show that, in most cases,
PMI and DRM can identify the optimum design and estimate the cost function accurately,
whereas the optimum design of PDM varies depending on the percentile used, and PDM
135
has identified a wrong global minimum. To achieve better accuracy, DRM with five
PMI and PDM yield the same efficiency if the same inverse reliability analysis is
used to find MPPs. Nonlinearity of the performance function affects the total number of
function evaluations most significantly in RBRDO using PMI and PDM. In estimation of
the statistical moments using DRM, the number of design variables affects the total
Three methods of evaluating the probability of failure using FORM, SORM, and
efficiency, the probability of failure calculation by FORM is the best since the probability
of failure calculation by SORM and DRM uses the MPP of the FORM-based inverse
reliability analysis. However, as shown using the examples in this study, the probability
of failure calculation by FORM could be very erroneous in particular when the multi-
dimensional performance function is highly nonlinear. Even though SORM can evaluate
the probability of failure more accurately than FORM, SORM has limited application
since SORM requires the second-order sensitivities. On the other hand, the probability of
failure calculation by DRM is as accurate as SORM, and sometimes even better than
SORM, without requiring the second-order sensitivities. For the system probability of
failure calculation, DRM-based reliability analysis shows better results than FORM-
based one since the component probability of failure affects the system probability of
136
functions.
DRM-based inverse reliability analysis is used to find accurate MPP, which can
identify the failure region of the performance function better than the FORM-based MPP.
A three-step computational procedure is proposed to find the DRM-based MPP using the
inverse reliability analysis: the probability of failure calculation using constraint shift,
reliability index update, and MPP update. The DRM-based MPP is used in the design
evaluations, especially when the number of design variables is large, PMA+ with new
tolerances for constraint activeness and reduced rotation matrix is used to enhance the
efficiency. The design examples show that the optimum design of DRM-based RBDO is
indeed different from the optimum design of FORM-based RBDO, and the probability of
of failure by DRM.
MPP-Based DRM
for the FORM-based PMA and DRM-based PMA are analytically derived in this study.
The analytic sensitivities for the FORM-based PMA are verified using the converging
constraint at the true DRM-based MPP are also verified using the FDM results. However,
since it is computationally very expensive to find the true DRM-based MPP, the
137
probabilistic constraint at the approximated DRM-based MPP is proposed for the DRM-
based PMA. Although the analytic sensitivities of the probabilistic constraint at the
approximated DRM-based MPP yield some inaccuracy at the initial design, the
sensitivities converge to the sensitivities of the probabilistic constraint at the true DRM-
based MPP as the design approaches the optimum design. In conclusion, it is very
desirable to use the sensitivities of the probabilistic constraint at the approximated DRM-
based MPP for the DRM-based PMA and RBDO because the computational cost can be
The system probability of failure estimation based on two methods, the MPP-
based DRM and FORM, is compared through numerical examples. For the highly
significant than the estimation of the joint probability of failure. Hence, in this case, the
system reliability analysis using the MPP-based DRM yields better accuracy than the
FORM-based system reliability analysis since the MPP-based DRM can accurately
estimate the component probability of failure. Consequently, the system RBDO using
MPP-based DRM shows better results than the system RBDO using FORM. However, it
is also important to use the correct method for critical constraint identification. Numerical
examples show that the system probability of failure estimation could be wrong even if
MPP-based DRM is used for the component probability of failure calculation. Thus, it is
In the study for the DRM-based RBRDO and RBDO in the literature, all random
variables are assumed to be statistically independent to each other due to its simplicity for
variables might be correlated, especially for a fatigue problem, input material properties
RBRDO and RBDO show different results for different orders when input random
transformation makes the performance function more nonlinear, which demand the use of
MPP-based DRM.
Hence, for the future research recommendation, the DRM-based RBRDO and
RBDO with correlated random variables will be developed to reduce the order effect
caused by FORM and to enhance the accuracy of the probability of failure calculation.
139
REFERENCES
Ang, A. H-S., and Tang, W., Probability Concepts in Engineering Planning and Design.
II: Decision, Risk, and Reliability, Wiley, New York, NY, 1984.
Atkinson K. E., An Introduction to Numerical Analysis, John Wiley & Sons, Chapter 5,
New York, NY, 1989.
Ba-abbad, M. A., Nikolaidis, E., and Kapania, R. K., “New Approach For System
Reliability-Based Design Optimization,” AIAA Journal, Vol. 44, No. 5, pp. 1087-
1096, 2006.
Buranathiti, T., Cao, J., and Chen, W., “A Weighted Three-Point-Based Strategy for
Variance Estimation,” Proceedings of ASME Design Engineering Technical
Conferences (DETC), Paper # DETC2004-57363, 2004.
Center for Computer-Aided Design, College of Engineering, DRAW User Reference, The
University of Iowa, Iowa City, IA, 1999b.
Chandra, M. J., Statistical Quality Control, CRC Press, Chapter 3, Boca Raton, FL, 2001.
Choi, K. K., Tu, J., and Park, Y. H., “Extensions of Design Potential Concept for
Reliability-Based Design Optimization,” Journal of Structural and Multidisciplinary
Optimization, Vol. 22, No. 3, pp. 335-350, 2001.
Ditlevsen, O., and Madsen, H.O., Structural Reliability Methods, John Wiley & Sons
Ltd., Chichester, 1996.
Du, X., and Chen, W., “A Most Probable Point-Based Method for Efficient Uncertainty
Analysis,” Journal of Design and Manufacturing Automation, Vol. 4, No. 1, pp. 47-
66, 2001.
140
Du, X., Sudjianto, A., and Chen, W., “An Integrated Framework for Optimization Under
Uncertainty Using Inverse Reliability Strategy,” ASME Journal of Mechanical
Design, Vol. 126, No. 4, pp. 562-570, 2004.
Du, X., and Huang, B., “Uncertainty Analysis by Dimension Reduction Integration and
Saddlepoint Approximations,” Proceedings of ASME Design Engineering Technical
Conferences (DETC), Paper # DETC2005-84523, 2005.
Gumbert, C. R., Hou, G.J.-W., and Newman P. A., “Reliability Assessment of a Robust
Design under Uncertainty for a 3-D Flexible Wing,” AIAA Paper 2003-4094, 2003.
Haldar, A., and Mahadevan, S., Probability, Reliability and Statistical Methods in
Engineering Design, John Wiley & Sons, New York, NY, 2000.
Hasofer, A. M. and Lind, N. C., “An Exact and Invariant First Order Reliability Format,”
ASCE Journal of the Engineering Mechanics Division, Vol. 100, No. 1, pp. 111-121,
1974.
Hohenbichler, M., and Rackwitz, R., “Sensitivity and Important Measures in Structural
Reliability,” Civil Engineering Systems, Vol. 3, pp. 203-209, 1986.
Huang, B., and Du, X., “A Robust Design Method Using Variable Transformation and
Gauss-Hermite Integration,” International Journal for Numerical Methods in
Engineering, Vol. 66, pp. 1841-1858, 2006.
Jin, R., Du, X. and Chen, W., “The Use of Metamodeling Techniques for Design under
Uncertainty,” Structural and Multidisciplinary Optimization, Vol. 25, No. 2, pp. 99-
106, 2003.
Kalsi, M., Hacker, K., and Lewis, K., “A Comprehensive Robust Design Approach for
Decision Trade-Offs in Complex Systems Design,” ASME Journal of Mechanical
Design, Vol. 123, No. 1, pp. 1-10, 2001.
141
Lee, I. Choi, K.K., and Du, L., “Alternative Methods for Reliability-Based Robust
Design Optimization Including Dimension Reduction Method,” 32nd ASME Design
Automation Conference, September 10-13, Philadelphia, PA, 2006.
Lee, I., Choi, K.K., Du, L., and Gorsich, D., “Dimension Reduction Method for
Reliability-Based Robust Design Optimization,” Computers & Structures, Vol. 86, pp.
1550-1562, 2008a.
Lee, I., Choi, K.K., Du, L., and Gorsich, D., “Dimension Reduction Method-Based MPP
for Reliability-Based Design Optimization of Highly Nonlinear Multi-Dimensional
Systems,” Special Issue of Computer Methods in Applied Mechanics and
Engineering: Computational Methods in Optimization Considering Uncertainties, to
appear, 2008b.
Liang, J., Mourelatos, Z. P., and Nikolaidis, E., “System Reliability-Based Design Using
a Single-Loop Method,” SAE 2007 World Congress, April 16-19, Detroit, MI, 2007.
Lin, C. Y., Huang, W. H., Jeng, M. C., and Doong, J. L., “Study of an Assembly
Tolerance Allocation Model Based on Monte Carlo Simulation,” Journal of Materials
Processing Technology, Vol. 70, pp. 9-16, 1997.
Madsen, H. O., Krenk, S., and Lind, N. C., Methods of Structural Safety, Prentice-Hall
Inc., Englewood Cliffs, NJ, 1986.
Marler, R. T., and Arora, J. S., "Survey of Multi-objective Optimization Methods for
Engineering", Structural and Multidisciplinary Optimization, Vol. 26, No. 6, pp. 369-
395, 2004.
Mckay, M. D., Conover, W. J. and Beckman, R. J., “A Comparison of Three Methods for
Selecting Values of Input Variables in the Analysis of Output from a Computer
Code,” Technometrics, Vol. 21, No. 2, pp. 239-245, 1979.
Noh, Y., Choi, K.K., and Du, L., “Reliability Based Design Optimization of Problems
with Correlated Input Variables Using Copulas,” Structural and Multidisciplinary
Optimization, accepted, July 2007.
142
Olsson, A. and Sandberg, G. E., “Latin Hypercube Sampling for Stochastic Finite
Element Analysis,” ASCE Journal of Engineering Mechanics, Vol. 128, No. 1, pp.
121-125, 2002.
Palle, T. C., and Michael, J. B., Structural Reliability Theory and Its Applications,
Springer-Verlag, Berlin, Heidelberg, 1982.
Rackwitz, R. and Fiessler, B., “Structural Reliability under Combined Load Sequences,”
Computers and Structures, Vol. 9, pp. 489-494, 1978.
Rahman, S. and Wei, D., “A Univariate Approximation at Most Probable Point for Higer-
Order Reliability Analysis,” International Journal of Solids and Structures, Vol. 43,
pp. 2820-2839, 2006.
Rahman, S., and Wei, D., “Design Sensitivity and Reliability-Based Structural
Optimization by Univariate Decomposition,” Structural and Multidisciplinary
Optimization, in press, 2007.
Rubinstein, R. Y., Simulation and Monte Carlo Method, John Wiley & Sons, New York,
1981.
Stein, M., “Large Sample Properties of Simulations using Latin Hypercube Sampling,”
Technometrics, Vol. 29, No. 2, pp. 143-151, 1987.
Su, J., and Renaud, J. E., “Automatic Differentiation in Robust Optimization,” AIAA
Journal, Vol. 35, No. 6, pp. 1072-1079, 1997.
Swanson Analysis System Inc., ANSYS Engineering Analysis System User’s Manual ,
Vol. I, II, Houston, PA, 1989.
Taguchi, G., Elsayed, E., and Hsiang, T., Quality Engineering in Production Systems,
McGraw-Hill, Chapter 2 and 3, New York, NY, 1989.
Tu, J., and Choi, K. K., “A New Study on Reliability-Based Design Optimization,”
Journal of Mechanical Design, Vol. 121, No. 4, pp. 557–564, 1999.
143
Tu, J., Choi, K. K., and Park, Y. H., “Design Potential Method for Reliability-Based
System Parameter Design Using Adaptive Probabilistic Constraint Evaluation,” AIAA
Journal, Vol. 39, No. 4, pp. 667-677, 2001.
Wu, Y. T., Millwater, H.R., and Cruse, T.A., „„Advanced Probabilistic Structural
Analysis Method for Implicit Performance Functions,‟‟ AIAA Journal, Vol. 28, No. 9,
pp. 1663–1669, 1990.
Wu, Y. T., „„Computational Methods for Efficient Structural Reliability and Reliability
Sensitivity Analysis,‟‟ AIAA Journal, Vol. 32 No. 8, pp. 1717–1723, 1994.
Wu. Y. T., Shin, Y., Sues, R., and Cesare, M., “Safety-Factor Based Approach for
Probability-Based Design Optimization,” 42nd AIAA/ASME/ASCE/AHS/ASC
Structures, Structural Dynamics and Materials Conference and Exhibit, AIAA-2001-
1239, Seatle, WA, April 16-19, 2001.
Youn, B. D., and Choi, K. K., “Adaptive Probability Analysis Using Performance
Measure Approach,” 9th AIAA/ISSMO Symposium on Multidisciplinary Analysis and
Optimization, Paper # AIAA2002-5532, Atlanta, Georgia, Sep. 4-6, 2002.
Youn, B. D. and Choi, K. K., “A New Response Surface Methodology for Reliability-
Based Design Optimization,” Computers and Structures, Vol. 82, pp. 241-256, 2004.
Youn, B. D., Choi, K. K., and Du, L., “Adaptive Probability Analysis Using An
Enhanced Hybrid Mean Value (HMV+) Method,” Journal of Structural and
Multidisciplinary Optimization, Vol. 29, No. 2, pp. 134-148, 2005a.
Youn, B. D., Choi, K. K., and Du, L., “Enriched Performance Measure Approach
(PMA+) for Reliability-Based Design Optimization,” AIAA Journal, Vol. 43, No. 4,
pp. 874-884, 2005b.
Youn, B. D., Choi, K. K., and Park, Y. H., “Hybrid Analysis Method for Reliability-
Based Design Optimization,” Journal of Mechanical Design, Vol. 125, No. 2, pp.
221-232, 2003.
Youn, B. D., Choi, K. K., Yang, R. -J., and Gu, L., “Reliability-Based Design
Optimization for Crash-worthiness of Vehicle Side Impact,” Structural
Multidisciplinary Optimization, Vol. 26, No. 3-4, pp. 272-283, 2004.
144
Youn, B. D., Choi, K. K., and Yi, K., "Performance Moment Integration (PMI) Method
for Quality Assessment in Reliability-Based Robust Optimization," Mechanics Based
Design of Structures and Machines, Vol. 33, No. 2, pp. 185-213, 2005c.
Youn, B. D., Xi, Z., Wells, L. J., and Lamb, D. A., “Stochastic Response Surface Using
The Enhanced Dimension Reduction (eDR) Method For Reliability-Based Robust
Design Optimization,” III European conference on Computational Mechanics,
Lisbon, Portugal, 2006.
Xu, H., and Rahman, S., “A Moment-Based Stochastic Method for Response Moment
and Reliability Analysis,” Proceedings of 2nd MIT Conference on Computational
Fluid and Solid Mechanics, Cambridge, MA, July 17-20, 2003.
Xu, H., and Rahman, S., “A Generalized Dimension-Reduction Method for Multi-
dimensional Integration in Stochastic Mechanics,” International Journal for
Numerical Methods in Engineering, Vol. 61, No. 12, pp. 1992-2019, 2004a.
Xu, H., and Rahman, S., “A Univariate Dimension-Reduction Method for Multi-
dimensional Integration in Stochastic Mechanics,” Probabilistic Engineering
Mechanics, Vol. 19, No. 4, pp. 393-408, 2004b.
Zhao, Y-G., Zhong W-Q., and Ang A. H-S., “Estimating Joint Failure Probability of
Series Structural Systems,” Journal of Engineering Mechanics, Vol. 133, No. 5, pp.
588-596, 2007.