0% found this document useful (0 votes)
62 viewsChapter 7 First Order and Second Reliability Methods
Uploaded by
Chris RodopoulosCopyright
© © All Rights Reserved
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
62 viewsChapter 7 First Order and Second Reliability Methods
Uploaded by
Chris RodopoulosCopyright
© © All Rights Reserved
Available Formats
Download as PDF or read online on Scribd
/ 34
Probabilistic Engineering Design
Chapter Seven
First Order and Second Reliability Methods
Xiaoping Du
University of Missouri — Rolla
| Berd)
sy Go a0, 10
Contour of joint pa Safe region oe
og
I ehh 0
Failure region
September 2005
Copyright Agreement: All the materials are for the use of individual private study only. They may not otherwise be copled or
stint in any way in whole or part without the pormasion ofthe author Chapter 7 Reliability Analysis Method
Chapter 7
First Order and Second Reliability Methods
7A
7.2
Introduction
As discussed in Chapter 6, reliability is defined as the probability of a performance
function g(X) greater than zero, ie. Pig(X)>0}. In other words, reliability is the
probability that the random variables 3 yey) are in the safe region that is
defined by g(X)> 0. The probability of failure is defined as the probability P{g(X)<0}
Or it is the probability that the random variables X=(,.X,,-.Y,) are in the failure
region that is defined by g(X)< 0. If the joint pdfof Xis /,(x), the probability of failure
is evaluated with the integral
PePigX)<0= J Lode 7.)
ae
The reliability is computed by
pp-Pig(X)>0}= [filo a2)
380
In this chapter, wo of the most commonly used reliability analysis methods, the First
Order Reliability Method (FORM) and the Second Order Reliability Method (SORM)
will be presented. The basic idea of the methods is to ease the computational difficulties
through simplifying the integrand f(x) and approximating the performance function
g(X). With the simplification and approximation, solutions to Eqs. 7.1 and 7.2 will be
easily obtained
All the random variables X herein are assumed mutually independent. The methods
discussed can be extended to problems with comelated random variables after those
variables are converted to independent variables.
First Order Reliability Method
The name of First Order Reliability Method (FORM) comes from the fact that the
performance function g(X) is approximated by the first order Taylor expansion
(linearization),
The probability integrations in Eqs. 7.1 and 7.2 are visualized with a two-dimensional
case in Fig. 7.1. The figure shows the joint pdf f(x) and its contours, which are Probabilistic Engineering Design,
projections of the surface of f(x) on X1 - Xs plane. All the points on the contours have
the same values of f(x) or the same probability density. The integration boundary
g(X)= 0 is also plotted on X) -%3 plane
hep
(4 80
Failure region _
ity Integration
The probability integration in Eqs. 7.1 or 7.2 is the volume undemeath the surface
(hyper-sucface for the higher than 2-D problems) of the joint pdf f(x) in the failure
region g(X)< or the safe region ¢(X)> 0. Imagine that the surface of the integrand
f(x) forms a “bill”. Ifthe hill were cut by a knife that has a blade shaped with the curve
g(X)=0, the hill would be divided into two parts. If the part on the side of g(X)< 0
were removed, the part left would be on the side of g(X)> 0 as shown in Fig. 7.1. The
volume left is the probability integration in Eq. 7.2, which represents the reliability. In
other words, the reliability is the volume underneath f(x) on the side of safe region
g(X)> 0. Of course, the probability of failure will be the volume undemeath f(x) on
the side of failure region g(X)< 0, the removed part,
To show the integration region more clearly, the contours of integrand f(x) and the
integration boundary g(X)=0 are plotted again in the random variable space (X-space) Chapter 7 Reliability Analysis Method
in Fig. 7.2, which is Xj - 1X3 plane. The integration for the reliability is performed in the
region where g(X)> 0 while the integration for the probability of failure is performed in
the region where g(X)< 0.
paf contour
\
aX, XP0
Figure 7.2 Probability Integration in X-Space
The direct evaluation of the probability integration in Eqs 7.1 and 7.2 is extremely
difficult. The reasons are multifold. First, since a number of random variables X are
involved, the probability integration is multidimensional. The dimensionality is typically
high for engineering applications. Second, the integrand f,(x) is the joint pdf of X and is
generally a nonlinear multidimensional function. Third, the integration boundary
o(X is also multidimensional and usually a nonlinear function. In many engineering
applications, g(X) is a black-box model (or simulation model), and the evaluation of
g(X) is computationally expensive. Examples of black-box models include finite
element analysis, dynamic simulation, and computational fluid dynamics. Because of the
complexities, there is seldom an analytical solution to the probability integration, except
for very special cases. It is also unpractical using aumerical integration to find the
solution due to the high dimensionality in most engineering applications. To this end,
approximation methods, such as the First Order Reliability Method (FORM) and Second
Order Reliability Method (SORM) have been developed in the area of structural
reliability, Probabilistic Engineering Design,
Two steps are involved in these approximation methods to make the probability
integration easy to be computed. The frst step is to simplify the integrand /,(x) so that
its contours become more regular and symmetric, and the second step is to approximate
the integration boundary g(X)=0. After the two steps, an analytical solution to the
probability integration will be easily found.
The way to approximate the probability integration divides the methods into two types:
the First Order Reliability Method (FORM) and the Second Order Reliability Method
(SORM). We will discuss FORM first. The procedure of the First Order Reliability
Method (FORM) is described below.
Step One — Simplify the integrand
The simplification is achieved through transforming the random variables from their
original random space into a standard normal space. The space that contains the original
x, (,) is called X-space. To make the shape of the
(x,
from X-space to a standard normal space, where the transformed random variables
.U,) follow the standard normal distribution, The transformed space is
termed as U-space. Recall that a standard normal variable has a mean of 0 and a standard
deviation of 1.0.
integrand f,(x) regular, all the random variables 3 ye X,) are transformed
The transformation from X to U is based on the condition that the cdfs of the random
variables remain the same before and after the transformation This type of
transformation is called Rosenblatt transformation [1], which is expressed by
F(x) =u) 73)
inwhich ©(-) is the cdf of the standard normal distribution.
The transformed standard normal variable is then given by
x)| (ay
For example, for a normally distributed random variable X ~V(«x,6 ), Eq. 7.4 yields
[FX] = oo
(7.6) Chapter 7 Reliability Analysis Method
It should be noted that in the above example the transformation fiom a normal variable to
a standard normal variable is linear. The general transformation from a non-normal
variable to a standard normal variable, however, is nonlinear,
After the transformation, the performance function becomes
¥=g(U) (7.7)
It is worthwhile to mention that after the transformation, the mathematical formulation of
the performance function g(X) will change. Without confusion, we will still use g(U) to
denote the transformed performance function in U-space in order to avoid introducing an
additional symbol for the performance function.
After the transformation, the probability integration becomes
ig(U)<0
J ge(w)da (7.8)
where 9,(u) is the joint pdf of U, Since all the random variables are independent, the
Joint pdf is the product of the individual pdf of standard normal distribution and is then
given by
ge(u) =] J eexp| -4 (79)
Ie \ 2
‘Therefore, the probability integration becomes
pe ff lee fy Van du, (7.10)
alien ao EV OTT -
It should be noted that after the transformation, the integration in Eq. 7.10 in U-space is
identical to that in Eq. 7.1 in X-space without any loss of accuracy, but the contours of
the integrand q, become concentric circles (or hyperspheres for a higher dimensional
problem). The circular contours are shown in Figs. 7.3 and 7.4. It is obvious that the
integrand 9, is easier to be integrated. Probabilistic Engineering Design,
aU, Uypo
Figure 7.4 Probabili
y Integration in U - Space Chapter 7 Reliability Analysis Method
Step Two — Approximate the integration boundary
In order to further make the probability integration easier to be evaluated, in addition to
simplifying the shape of the integrand, the integration boundary g(U)=0 will also be
approximated. FORM uses a liner approximation (the first order Taylor expansion) as
shown below.
g(U)=L(U) = g(a") +¥g(u KU -w (7.11)
where (U) is the linearized performance function, w
tathe--su,) is the expansion
point, T stands for a transpose, and Vg(u') is the gradient of g(U) at u’. Vgiu') is
given by
(7.12)
To minimize the accuracy loss, it is natural to expand the performance function g(U) at
a point that has the highest contribution to the probability integration. In other word, itis
preferable to expand the function at the point that has the highest value of the integrand,
namely, the highest probability density. With the integration going away from the
expansion point, the integrand function values will quickly diminish. The point that has
the highest probability density on the performance g(U)=0 is termed as the Most
Probable Point (MPP). Therefore, the performance function will be approximated at the
MPP. Maximizing the joint pdf 9,(u) at the limit state of g(U) =0 gives the location of
the MPP. The mathematical model for locating the MPP is then given by
{os lye 3 | (713)
Foubject to g(u)=0
Since
vin
(7.14)
Loup
Yan Pa
q
ft | L
maximizing [] p—exp| —5
MPP search can be rewritten as
Jis equivalent to minimizing QYve the model forthe Probabilistic Engineering Design,
sal fl
« (7.15)
subject to g(u)=0
where {| stands for the norm (length or magnitude) of a vector, namely,
\ul=Jigeae+ oe a (7.16)
The solution to the model given in Eq, 7.15 is the MPP and is denoted by
w’ =(2,24,---1j,). As shown graphically in Figs. 7.5 and 7.6, the MPP is the shortest
distance point from the limit state_g(U)=0 to the origin O in Uspace. The minimum
distance 8 =[u'| is called reliability index.
paf contour
Figure 7.5 Probability Integration in FORM Chapter 7 Reliability Analysis Method
Highest probability density
at the MPP
ure 7.6 Highest Probability Densi
Since at the MPP w’, g(U)=0, Eq, 7.11 becomes
(7.17)
where
(7.18)
and
(7.19)
Eq. 7.17 indicates that £(U) is a linear function of standard normal variables. Therefore,
L(W) is also normally distributed. Its meanis given by Probabilistic Engineering Design,
(7.20)
(7.21)
(7.22)
where
(7.23)
Let the vector of a, be
(7.24)
The probability of failure can also be written as.
(7.25)
in which au’ is the inner (dot) product of the unit vector a and the vector of the MPP
r
0 Chapter 7 Reliability Analysis Method
As shown in Fig. 7.7, since the MPP u’ is the shortest distance point from the origin to
the performance fimetion curve g(U)=0, the MPP is the tangent point of the curve
g(U)=0 and the circle with the radius of B . Therefore, the MPP vector u’ is
perpendicular to the curve g(U)=0. The direction of the MPP can be represented by the
unit vector w’ /fu'||=u'/B . On the other hand, the direction of the
dient is also
perpendicular to the curve at the MPP, and its direction can be represented by the unit
vector a (see Eq. 7.24). Therefore
(7.26)
or,
(7.27)
u’ — a tangent point
Figure 7.7 The MPP is A Tangent Point
Therefore, the probability of failure is evaluated by
pb, = PIL(U)<0} = (aw
‘Note that in the above derivation, aa”
The reliability is then given by
R
1- p,=1-®(-B)=0(6) (7.29)
un Probabilistic Engineering Design,
7.3
The procedure of the FORM is briefly summarized below.
1) Transform the original random variables from Xspace to U-space by Rosenblatt
transformation
2) Search the MPP in U-space and calculate the reliability index
3) Calculate reliability R= (8)
MPP Search
From the above discussion, it is noted that the key to calculating the probability of failure
or reliability is to locate the MPP in U-space. Since it is very difficult or even impossible
to solve the MPP search model in Eq, 7.15 analytically, many numerical methods have
been developed for the MPP search. Next, we will introduce a commonly used MPP
search algorithm.
The MPP search algorithm uses a recursive formula and is based on the linearization of
the performance function, The procedure is demonstrated in Fig. 7.8
gu’)
g(a") +Vg(ut yuh 0
g(a") + Te(u' (ua)?
uy
Figure 7.8 MPP Search
Let the MPP in Ath iteration be uw’. The performance function is linearized at u' as
shown by the lower line in Fig. 7.8. The linearized function is given by
g(u) = g(u')+ Vg(au uP (7.30)
Let the linearized function be zero, then the MPP u*~ in the next iteration will be on the
line, namely, Chapter 7 Reliability Analysis Method
g(a’ g(u')+Vg(u' Que! —ub) =0. (7.31)
The line is shown as the upper line in Fig. 7.8.
From Eq. 7.27,
—ptat (7.32)
‘As shown in Fig. 7.8, since u" is the shortest distance point from the origin to the line,
vector u’' is perpendicular to the line and is directed from the origin O to u™'. a’,
which is the unit vector of the gradient, is also perpendicular to the line and in the
opposite direction to wu". Because the magnitude of u* is the distance from the origin
to u™ (celiability index), then
(7.33)
Substituting u* in Eq. 7.32 and u‘“ in Eq. 7.33 into Eq. 7.31 yields
atu’) + Ug(u'ya')"(BF— 8) =g(u') +P g(u)|(B -B =o (7.34)
Rearranging Eq, 7.34 gives
Ry (7.35)
Therefore, the updated point is given by
ey
‘ett om
To use the recursive formula in Eqs. 7.35 and 7.36, a starting point u’is required.
Usually, the origin uw’ =0 is set as the starting point. Three convergence criteria may be
used to terminate the MPP search process.
1) If jut" -u'], Ps, Ps, Ps and Ps are applied at six different locations along the beam, Ly,
Ls, LL Ls, and Le. The allowable tensile stress is S.
P) P. Ps Ps Ps Ps
4
«|
Oo B
D
iS]
M~M cross-section,
Figure 7.13 A Composite Beam.
In this problem, the twenty random variables are
X=[X-.Xy] =[4 BCD, Ly Ly Ly, Ly Ls, Ly, L By, Py, Py Py PsP Eg ES)
Details of the random variables are given in Table 7.2.
Table 7.2 Random Variables of the Beam Problem
ariable No. Variable Mean value | Standard deviation | Distribution
T a 100 mm 02mm Normal
2 B 200 mm 0.2 mm ‘Normal
3 Cc 30 mm, 02 mm Normal
4 D 20 mm 0.2 mm Normal
3 Zi 200 mm Timm Normal
6 Lb 400 mm Imm Normal Chapter 7 Reliability Analysis Method
Variable No. ‘Variable Mean value | Standard deviation | Distribution
7 i 600 mm Timm Nomal
8 Le 300 mm Tmm Normal
9 Zs 1000 mm Imm Nonmal
10. Zs 1200 mm Tmm ‘Nonmal
Tl L 1400 mum 2am ‘Nonmal
12 Pi DEN TSN ‘Nommal
B Pp 15EN SKN ‘Normal
14 Ps ISN TSN ‘Normal
15 Ps 1.5kN Normal
16 Ps T.5kN ‘Nonmal
7 Pe. SKN ‘Normal
18 Ee 7GPa Normal
19 Ew 0.875 GPa Normal
20 Ss 3.825MPa_ Normal
‘The maximum stress occurs in the middle cross-section M-M and is given by
osas’+Expeve+D)
E,
Laem
1-1) RU -L)
fo sBooweo]
eos
RE
The maximum stress 6 should be less than the allowable stress (strength) S. The
probability of failure of the beam is defined by
{g(X)< 0} = P{S—o <0}
The MPP found by FORM is shown in Table 7.3. The reliability index B =3.1317 and
the probability of failure is
Pp, = 0(-B)=9(-3.1317) =8.6908 - 10*
Table 7.3 The MPP in U-Space and X-space
Variable No. Variable MPP in Uspace | MPP in Xspace
T A ~1.8148E-02 9.9996E-02
2 B =1.9192E-02 2.0000E-01
3 Cc ~4.9159E-03 T.9999E-02 Probabilistic Engineering Design,
7.4
‘Variable No. Variable MPP in U-space ‘MPP in X-space
4 D -2.8892E-02 1.9994E-02
5 TZ; SS311E-03 2.0001E-01
6 Ze =5.5885E-03 3.9999E-01
7 Za 9.2576E-03 6.0001E-01
8g Ly ~S.AT3TE-03 7.9999E-01
9 Is ~5.4163E-03 9,9999E-01
10 Zs T.9859E-02 1.2000E=00
1 L =1.0603E-02 1-4000E=00
2 Py 3.2153E-01 1.3482E +04
1B P: 2.1435E-01 S322E+04
4 Ps 3.2154E-01 5482E+04
15 Ps 2.1437E-01 1,.5322E+04
16 Ps 1.07I9E-01 LSIGIE+O4
7 Ps =L.07ISE-O1 TA839E=04
18 Ez; =2.0064E-01 6.8596E+10
19 Ew 1.9291E-01 S.9188E+09
20 S =3.0669E+00 1.3769E=07
Second Order Reliability Method
As its name implies, the Second Order Reliability Method (SORM) uses the second order
Taylor expansion to approximate the performance function at the MPP uw’. The
approximation is given by
a(U)=G(U) = g(a) + Vw (Uw (Uw aa (Uw)? (737)
where H(u’)is the Hessian matrix at the MPP, namely,
Hu’) = eu, (7.38) Chapter 7 Reliability Analysis Method
After a set of linear transformation, such as coordinate rotation and orthogonal
diagonalization, the performance function is further simplified as
a)
v-| B +4v"pw | (7.39)
where D is a (1-1) -(1-1) diagonal matrix whose elements are determined by the
+ U3}
Hessian matrix H(u'), and U ={U,, Uy
When 8 is large enough, an asymptotic solution of the probability of failure can be then
derived as
P= PiatX)<0}= BC B)L LICH B 5) (7.40)
in which x,denotes the i-th main curvature of the performance function g(U) at the
MPP.
Since the approximation of the performance function in SORM is better than that in
FORM (see Fig. 7.14), SORM is generally more accurate than FORM. However, since
SORM requires the second order derivative, it is mt as efficient as FORM when the
derivatives are evaluated numerically. If we use the number of performance function
evaluations to measure the efficiency, SORM needs more function evaluations than
FORM.
Figure 7.14 Comparison of FORM and SORM Probabilistic Engineering Design,
Example 7.4
Use SORM to solve problem 7.2,
The same MPP is used for SORM. The probability of failure is found to be
P, =0.04098. For the nonlinear performance function in this problem, no analytical
solution exists. To compare the accuracy of the results, Monte Carlo Simulation (MCS) is
used to solve the problem. As we will discuss later in this book, the higher the number of
simulations is used, the higher the result is. For this problem, one million simulations are
conducted. The result from MCS is considered as the accurate solution for the
comparison, The results of the probability of failure are displayed in Table 7.3. The
results indicate that SORM is more accurate than FORM.
Table 7.3 The Probability of Failure from Different Methods
‘Method FORM SORM MCS
Ps 0.04054 0.04098 0.04092
As in Example 7.2, if P. and P, follow lognormal distributions with the same mean and
standard deviation, the probability of faifure calculated from FORM is p, = 0.0538. The
comparison between FORM and SORM is given in Table 7.4. MCS with 10° simulations
is again used as the reference. The result shows that SORM is more accurate than FORM.
‘Table 7.4 The Probability of Failure When P, and P, Follow Lognormal
Distributions
Method FORM SORM MCS
Py 0.0531 0.0538 0.0541
Example 7.5
Use SORM to solve problem 7.3
The result from SORM is given in Table 5. The results from FORM and MCS are also
listed in the table for the comparison of the accuracy. MCS uses 10° simulations and its
result considered as the reference. For this problem, the first order and second order
derivatives of the performance function g(X) are evaluated by difference finite method.
The number of evaluating g(X) is used to measure the efficiency. Chapter 7 Reliability Analysis Method
75
Table 7.5 Comparison of Accuracy and Efficiency
Method FORM SORM MCS
Py 8.6908 -10* | 8.7813 - 10% 8,870 - 10°
‘Number of function evaluations 38 350 10
‘The results show that SORM is more accurate than FORM. However, SORM is much
less efficient than FORM. SORM calls the performance function 550 times while FORM
calls the performance function only 88 times,
Inverse Reliability Analysis
As we will see in reliablity-based design later in this book, the use of the percentile value
of a performance function corresponding to a given reliability is more efficient. The
evaluation of the percentile value of the performance function is an inverse problem of
the reliability analysis. The problem can be stated as: Find the p — percentile value g?
give the probability
Pig(X)
