Sie sind auf Seite 1von 9

Computers and Geotechnics 38 (2011) 196204

Contents lists available at ScienceDirect

Computers and Geotechnics


journal homepage: www.elsevier.com/locate/compgeo

Efcient system reliability analysis illustrated for a retaining wall and a soil slope
B.K. Low a,, J. Zhang b, Wilson H. Tang c
a
School of Civil and Environmental Engineering, Nanyang Technological University, Singapore
b
Key Laboratory of Geotechnical and Underground Engineering of Ministry of Education, Department of Geotechnical Engineering, Tongji University, Shanghai, China
c
Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong

a r t i c l e i n f o a b s t r a c t

Article history: Although rst-order reliability method is a common procedure for estimating failure probability, the for-
Received 15 July 2010 mulas derived for bimodal bounds of system failure probability have not been widely used as expected in
Accepted 19 November 2010 present reliability analyses. The reluctance for applying these formulas in practice may be partly due to the
Available online 23 December 2010
impression that the procedures to implement the system reliability theory are tedious. Among the meth-
ods for system reliability analysis, the approach suggested in Ditlevsen 1979 is considered here because it
Keywords: is a natural extension of the rst-order reliability method commonly used for failure probability estima-
System reliability
tion corresponding to a single failure mode, and it can often provide reasonably narrow failure probability
Probability
Slope
bounds. To facilitate wider practical application, this paper provides a short program code in the ubiqui-
Retaining wall tous Excel spreadsheet platform for efciently calculating the bounds for system failure probability. The
Multiple failure modes procedure is illustrated for a semi-gravity retaining wall with two failure modes, a soil slope with two
and eight failure modes, and a loaded beam with three failure modes. In addition, simple equations are
provided to relate the correlated but unrotated equivalent standard normals of the Low and Tang 2007
FORM procedure with the uncorrelated but rotated equivalent standard normals of the classical FORM
procedure. Also demonstrated are the need for investigating different permutations of failure modes in
order to get the narrowest bounds for system failure probability, and the use of SORM reliability index
for system reliability bounds in a case where the curvature of the limit state surface cannot be neglected.
2010 Elsevier Ltd. All rights reserved.

1. Introduction estimating the system failure probability are generally tedious


for practicing engineers.
Performance of engineering systems often involve multiple Among the techniques for estimating probability bounds, the
failure modes. For instance, the geotechnical failure modes to be method for calculating the bimodal bounds suggested by Ditlevsen
considered in the design of a semi-gravity retaining wall may in- [8] is considered here for two reasons. First, it is a natural exten-
clude rotation about the toe of the wall, horizontal sliding along sion of the widely used rst-order reliability method (FORM) for
the base of the wall, and bearing capacity failure of the soil beneath estimating the failure probability of a single failure mode. Second,
the wall. Procedures for estimating failure probability of a system it can often provide reasonably narrow probability bounds. The
have been summarized in [16], for example. While it is usually objective of this paper is to provide a convenient procedure for
difcult to calculate system reliability exactly, methods for esti- assessing the bounds of system failure probability based on the
mating the bounds of system reliability have been suggested, for method suggested by Ditlevsen [8]. The structure of this paper is
example [713]. The unimodal bounds (e.g., [7]) are often too wide as follows. First, simple equations are provided to relate the corre-
for practical use. Bimodal (e.g., [8]) or higher order bounds (e.g., lated equivalent standard normal variates of Low and Tang [14]
[10]) would be needed if a more accurate estimate is desirable. FORM approach to the uncorrelated but rotated dimensionless
Yet, the formulas derived for bimodal or high order probability variates of the classical FORM approach; the reliability indices for
bounds are still not widely used as expected. There is an apparent individual modes and associated design points obtained by either
gap between research in system reliability and their applications in FORM approach are required in subsequent system reliability anal-
practice. One possible reason for this is that procedures for ysis. Next, the established expressions for estimating the bounds of
system failure probability are reviewed, and simple and short
spreadsheet function codes are created in a ubiquitous spreadsheet
Corresponding author. Address: School of Civil and Environmental Engineering,
platform to enable efcient evaluation of the lower and upper
Nanyang Technological University, Singapore 639798, Singapore. Tel.: +65 67905270;
fax: +65 67910676. bounds on system failure probability. Finally, three examples are
E-mail address: bklow@alum.mit.edu (B.K. Low). used to illustrate the procedure for assessing the bounds of system

0266-352X/$ - see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compgeo.2010.11.005
B.K. Low et al. / Computers and Geotechnics 38 (2011) 196204 197

failure probability, with insights, discussions, and comparisons sign point n = (1.2939, 0.8934, 2.2930), which translates to
with Monte Carlo simulations. u = L1n = (1.2939, 0.4101, 2.2930), practically the same as
the design point u (1.2943, 0.4098, 2.2926) obtained via Rosen-
2. Relationship between two computational approaches for blatt transformation in Appendix II of Low and Tang [14]. (The Excel
FORM VBA code for obtaining the lower triangular matrix L of the correla-
tion matrix R is simple and available in the public domain, as
This section relates the unrotated equivalent standard normal described in the Appendix.)
vector n of the Low and Tang [14] FORM approach with the ro- Note that the uncorrelated u vector of the classical FORM ap-
tated equivalent standard normal vector u of the classical FORM proach is not unique and may differ somewhat depending on
approach. Either FORM approach can be used to obtain the in- whether in the classical approach one is using the eigenvector
puts for system reliability analysis. The results will be virtually and eigenvalue transformation approach or the Rosenblatt trans-
identical. Practice-inclined readers may focus on Eqs. (1), (7a), formation approach. In the former case, the components ui of the
(7b) and (8a), (8b) below which are used in subsequent sections. u vector may also emerge in different sequence, depending on
Eqs. (2a), (2b), (3), (4), (5), (6a), (6b) provide the basis for Eqs. the assembling of the eigenvectors in forming the coordinate rota-
(7a), (7b) and (8a), (8b), but are not themselves used in later tion matrix T.
sections. The direction cosine vector a, required in system reliability
The Low and Tang 2007 FORM procedure [14] is an alternative analysis, can be obtained via the design point u or n, as follows:
to the Low and Tang 2004 FORM procedure [15], and uses the fol-
u L1 n
lowing equation for the reliability index b of a single failure mode: a 4
b b
s
 T  
xi  lNi p
1 xi  li
N
It may be noted that when the random variables are uncorrelated,
b min R min nT R1 n 1
x2F N
ri ri N x2F u = n by Eq. (3), because then L1 = I (the identity matrix).
For system reliability evaluation, the correlation coefcients qij
where x is the vector of random variables, F the failure domain, lNi
between failure mode i and failure mode j are needed and can be
and rNi the equivalent normal means and standard deviations, and R
estimated from the following established equation (e.g., Ang and
the correlation matrix of the random variables. The reliability index
Tang [1]):
b and the numerical values of the dimensionless equivalent stan-
dard normal vector n at the design point (denoted as n) are ob- qij aTi aj uT 
i uj =bi bj 5
tained via spreadsheet-automated constrained optimization. The
original random variables xi (on which the performance functions which can also be evaluated in terms of the n vectors if the Low
are formulated) are computed automatically from the probabilistic and Tang 2007 FORM procedure [14] is used, as follows:
connections between ni and xi. (If desired, the original correlation  T  
L1 ni L1 nj nT T 1 1 
matrix R can be modied in line with the equivalent normal trans- uT
i uj i L L nj
formation, as suggested in Der Kiureghian and Liu [16]. For the qij
bi bj bi bj bi bj
cases illustrated herein, the correlation matrix thus modied differs 1 
only very slightly from the original correlation matrix R. Hence, for nT
i LU nj
6a
simplicity, the examples of this study retain the original unmodied bi bj
correlation matrices.)
The next few paragraphs provide the basis of some useful equa- nT 1 
i R nj
tions which enable users to relate the vector n of Eq. (1) with the i:e:; qij 6b
bi bj
vector u of the classical FORM approach.
It is long established that the correlation matrix R can be The correlation matrix q of a system of m failure modes can be
decomposed into a lower triangular matrix L and an upper triangu- assembled efciently from the vectors ni and the reliability indices
lar matrix U, i.e., R = LU, in which U = LT. Hence, based on the basic bi of the m failure modes, as follows:
properties of the inverse and transpose of matrices, namely  
(AT)1 = (A1)T, (AB)T = BTAT, (AB)1 = B1A1, e.g., Anton and Ror- n1 n2 n
A ; ;...; m 7a
res [17], one can rewrite Eq. (1) as follows: b1 b2 bm
p q
b min nT R1 n min nT LU1 n q AT R1 A 7b
x2F x2F
q q
It can be shown that the element in the ith row and jth column of
min n L L n min L1 nT L1 n
T T 1 1
2a
x2F x2F the q matrix is exactly the correlation coefcient between failure
mode i and failure mode j implied by Eq. (6b).
Alternatively, if the uncorrelated standard normal vectors ui in
p
i:e:; b min uT u 2b the rotated space of the classical FORM approach have been used to
x2F obtain the reliability indices of the different failure modes, the sys-
where tem correlation matrix q can be obtained as follows:
 
u L1 n 3 u1 u2 u
B ; ;...; m ; 8a
b1 b2 bm
in which u is the uncorrelated standard equivalent normal vector in
the rotated space of the classical mathematical approach of FORM. q BT B 8b
The equation u = L1n provides a very useful relationship between
the vector n of one FORM approach and the vector u of the other The vector of FORM reliability indices bi obtained using either
FORM approach. For example, for the case with correlated nonnor- the Low and Tang 2007 FORM procedure [14] or the classical
mals and performance function = YZ  M in Ang and Tang [1], the FORM procedureand the correlation matrix q of the failure
Low and Tang 2007 computational approach [14] obtained the de- modes (by Eqs. (7a), (7b) or (8a), (8b)) are the inputs required in
198 B.K. Low et al. / Computers and Geotechnics 38 (2011) 196204

( " #)
the practical spreadsheet algorithm for system reliability bounds X
m X
i1

explained next. PE1 max PEi  PEi Ej ; 0  PF;sys


i2 j1
( ( ))
X
m
 
3. Practical spreadsheet algorithm for system reliability bounds 6 PE1 PEi  max PEi Ej 9
i2 j<i
For a series system with m limit state functions (failure modes),
violation of any limit state function would result in system failure. where P(Ei) = the failure probability corresponding to the ith failure
Let Ei denotes the event that the ith limit state is violated, and PF,sys mode; and P(EiEj) = the probability that the i and jth limit state func-
denotes the probability of system failure. Exact calculation of PF,sys tions are violated simultaneously.
is often difcult. The following bimodal bounds which account for In Eq. (9), P(Ei) can be assessed via a variety of methods, such as
the correlation between pairs of potential failure modes are sum- point estimate methods (e.g., Rosenblueth [22]), FORM (Hasofer
marized, for example, in Ang and Tang [1], Madsen et al. [18], Mel- and Lind [23]; Rackwitz and Fiessler [24]), Monte Carlo simulations
chers [2], Haldar and Mahadevan [19], and Tung et al. [6], based on (e.g., Melchers [2]), or SORM (e.g., Haldar and Mahadevan [19],
Kounias [20], Hunter [21], and Ditlevsen [8]: Melchers [2]). Calculating P(EiEj), which is the intersection of two

Fig. 1. Excel function code for automatically computing the KouniasDitlevsen bi-modal bounds for systems with multiple failure modes.
B.K. Low et al. / Computers and Geotechnics 38 (2011) 196204 199

failure events, is generally not easy. A weakened version of Eq. (9) in which bi and bj = reliability indices corresponding to failure
was due to Ditlevsen [8], who suggested that the lower and upper modes i and j, respectively; and qij = correlation coefcient between
bounds of P(EiEj) can be calculated as follows: failure modes i and j. Substituting the bounds of Eqs. (10a) and
(10b) into Eq. (9), the upper and lower bounds for PF,sys can be esti-
maxa; b  PEi Ej  a b; for qij  0 10a
mated. In this way, the complex problem of estimating system fail-
ure probability is reduced to problems of estimating failure
0  PEi Ej  mina; b; for qij < 0 10b
probabilities corresponding to various single limit states as well
where a and b are dened as: as the correlation coefcients among various failure modes.
0 1 Although Eqs. (9), (10a), (10b), (11a), (11b) for calculating prob-
b  q b ability bounds are well established, direct calculations based on
B j ij i C
a Ubi U@ qA 11a them could be tedious for practicing engineers when there are
1  q2ij three or more failure modes. To develop a convenient procedure
for system reliability assessment, Eqs. (9), (10a), (10b), (11a),
0 1
(11b) have been coded in the Visual Basic programming environ-
B b i  qij bj C ment of spreadsheet software Microsoft Excel, as shown in Fig. 1.
b Ubj U@ qA 11b
1  q2ij One only needs to paste the function code in a module sheet in
the Visual Basic Editor environment of Microsoft Excel, and uses

Fig. 2. System reliability analysis of a semi-gravity retaining wall.


200 B.K. Low et al. / Computers and Geotechnics 38 (2011) 196204

it like other Excel functions. For program codes like Fig. 1 to work, of 3  3 correlation matrix among the random variables u,
the Macro Security level has to be set to medium in Excel 2003 and d and ca shown at the top of Fig. 2. (Array formulas have to
earlier, or the Enable Contents option activated in Excel 2007. The be entered while pressing the Ctrl + Shift keys.) The results
VBA code of Fig. 1 can be converted readily for use in other plat- are shown in the 2  2 matrix q below the retaining wall
forms (e.g., MatLab) if desired. in Fig. 2, in which the correlation coefcient between the
In Fig. 1, the rst function code SysProbLowerBound evaluates rotation mode and sliding mode is 0.5364. Had the rotated
the left-hand term of Eq. (9), the second function code SysProbUp- and uncorrelated u vectors been obtained via the classical
perBound evaluates the right-hand term of Eq. (9), and the third FORM computations, Eqs. (8a), (8b) can be used (instead of
function code ProbFiFjinvoked by both the rst two function Eqs. (7a) and (7b)) to obtain the same intermodal correlation
codesevaluates the lower and upper bounds of P(EiEj) according matrix q.
to Eqs. (10a), (10b), (11a), and (11b). The upper and lower bounds (iii) Assuming that the function code shown in Fig. 1 has been
of P(EiEj) are used in the left and right terms of Eq. (9), respectively, inserted in a module sheet in the Visual Basic Editor of
thereby yielding the lower and upper bounds of system failure Microsoft Excel, and that the macro is enabled (as described
probability PF,sys. in the previous section), one obtains the system lower bound
The use of the algorithm in Fig. 1 to facilitate system reliability of the PF,sys in Eq. (9) by simply entering the formula = Sys-
evaluation is illustrated in the three examples to follow. ProbLowerBound(b vector, q matrix) in a cell. In the for-
mula, b vector and q matrix are the selected two cells
4. System reliability bounds for a semi-gravity retaining wall of b1 and b2, and the 2  2 matrix q, respectively. The lower
bound is 0.705% as shown. Likewise, one obtains the system
Low [25] analyzed a semi-gravity retaining wall for FORM reli- upper bound of the PF,sys in Eq. (9) by simply entering the for-
ability indices with respect to rotation and sliding modes, and also mula = SysProbUpperBound(b vector, q matrix) in a cell,
obtained unimodal system reliability bounds. The case is revisited yielding 0.717% as the upper bound of system failure
in this section to illustrate bimodal system reliability bounds, rst probability.
for the case with correlated normal random variables, then with The system failure probability bounds of 0.7050.717% is
correlated lognormals. The steps for evaluating system reliability remarkable when compared with the failure probability of
are as follows: 0.715% from Monte Carlo simulation (Latin Hypercube sampling
of size 800,000, using @RISK program, http://www.palisade.com).
(i) Obtain the reliability index and the design point for each The bimodal bounds are also much sharper than the unimodal
failure mode. In Fig. 2, for the case with correlated normals, bounds of 0.6370.732% obtained in Low [25].
the reliability indices are 2.491 for overturning mode and In this case when there are only two failure modes, Eq. (9) re-
3.102 for sliding mode. The design point information is con- duces to the following:
tained in the standard normal vectors n1 (2.4492, 2.2329, fPE1 maxPE2  PE1 E2 ; 0g  PF;sys  fPE1 PE2  PE1
0.0) and n2 (1.6752, 1.3963, 2.6092) based on the Low E2 g, and a manual check is possible and instructive: with
and Tang 2007 spreadsheet FORM procedure [14]. (The clas- P(E1)  U(b1) = NormSDist(b1) = 0.6366%, P(E2)  U(b2) = Nor-
sical FORM procedure can also be used, to obtain the uncor- mSDist(b2) = 0.0961%, lower bound of P(E1E2) = ProbFiFj(b1, b2,
related u1 and u2 vector instead of the n1 and n2 vectors.) 0.5364, 0) = 0.0157%, upper bound of P(E1E2) = ProbFiFj(b1, b2,
(ii) Obtain the 3  2 matrix A according to Eq. (7a), and then the 0.5364, 1) = 0.0273%, where typing ProbFiFj(. . .) in a cell invokes
2  2 matrix of intermodal correlation matrix q according to the third function code in Fig. 1 to execute Eqs. (10a), (10b),
Eq. (7b). The latter is computed easily by selecting a 2  2 (11a), and (11b). Then, using 0.0273% and 0.0157% for the P(E1E2)
block of cells, and typing the following array formula for in the left and right terms, respectively, of the above expression
Eq. (7b): = mmult(transpose(MatrixA),mmult(minverse for PF,sys, one obtains 0.705% 6 PF,sys 6 0.717%, identical with the re-
(RMatrix),MatrixA)), where RMatrix is the selected block sults obtained much more efciently by directly entering in two

Fig. 3. FORM results for two reliability-based critical slip circles followed by system reliability analysis.
B.K. Low et al. / Computers and Geotechnics 38 (2011) 196204 201

spreadsheet cells the formulas SysProbLowerBound (. . .) and


SysProbUpperBound (. . .) thereby engaging the codes shown in
Fig. 1. When there are three or more failure modes to be considered
(as in the next section) and the sharpest bounds on PF,sys is to be
obtained by trying all permutations of the failure modes, manual
computations of Eq. (9) are virtually ruled out and the merit of
the function code in Fig. 1 increases.
The lower right of Fig. 2 also shows the system reliability anal-
ysis for the case with lognormally distributed random variables
which have the same mean and covariance values as the normally
distributed random variables analyzed above. The system failure
probability bounds of 0.2620.264% is remarkably close to that
(0.2560.260%) from Monte Carlo simulations. That the Monte
Carlo average (0.258%) did not quite lie within the system reliabil-
ity bounds (0.2620.264%) is in this case due to the approximation
involved in transforming the original lognormals to equivalent
standard normals in FORM computations. The error introduced
by this approximation is very slight in this case. 30 = 2.795
= 2.893
20
5. System reliability bounds for a slope with two equally likely
failure modes
10
Clay layer 1

The limit equilibrium single-mode slope reliability analysis of


Low [26] is extended to a system reliability analysis accounting 0 Clay layer 2
for multiple failure modes. Other approaches and issues of deter- -10 0 10 20 30 40 50
ministic and probabilistic slope analysis have been studied by -10
Wu and Kraft [27], Vanmarcke [2829], Chen and Morgenstern
Fig. 4. Variation of reliability indices with the x-coordinate of the lower exit point
[30], Li and Lumb [31], Christian et al. [32], Morgenstern [33],
of slip circles, and the two reliability-based critical slip circles with lower exit
Low and Tang [34], Low et al. [35], Au and Beck [36], Nadim points at xexit of 1.5 m and +9 m, respectively.
et al. [37], and Low et al. [38], among others.
Ching et al. [39] analyzed several slopes using Monte Carlo sim-
ulation (MCS) and importance sampling (IS) methods. The results interesting to note that although cu1 and cu2 are uncorrelated,
were compared with single-mode FORM analysis. It was concluded there is correlation between the two failure modes (q12 =
that FORM is computationally less demanding and hence a pre- 0.4535), because cu1 affects both slip circles. The bounds on sys-
ferred method over the MCS and IS methods for problems with a tem failure probability, computed in two cells in Fig. 3 by enter-
single failure mode. On the other hand, for a slope in two clayey ing the functions of Fig. 1 which implement Eqs. (9), (10a), (10b),
soil layers, Ching et al. [39]s single-mode FORM analysis yielded (11a), (11b), are 0.4320.441%, compared with the MCS estimated
a probability of failure smaller than half the failure probabilities one-sigma range of 0.370.51% (from Ching et al. [39]s reported
obtained by MCS and IS methods. It was then concluded that sin- MCS mean of 0.44% and c.o.v. of 15.04%).
gle-mode FORM analysis signicantly underestimated the failure The two reliability-based critical slip circles in Fig. 3 have the
probability, and that both the IS and MCS methods provided unbi- smallest b values among all possible slip circles tangent to the bot-
ased estimates for the failure probability. toms of the upper and lower clay layers, respectively. One can
Ching et al. [39]s slope in two clayey soil layers is analyzed in search for more reliability-based critical slip circles corresponding
this section, using system reliability bounds for multiple failure to different trial tangent depths. Alternatively, a series of b values
modes. It will be shown that when two (rather than one as was can be obtained as a function of the x-coordinate values of the low-
done in [39]) reliability-based critical slip surfaces are considered, er exit end of critical slip circles, as shown in Fig. 4, where the exis-
the system failure probability bounds are consistent with the MCS tence of two stationary values (troughs) of b is obvious. It would
and IS results, and hence the perceived inadequacy of FORM (rela- be interesting to investigate the effect on the bounds of system
tive to MCS and IS) was due entirely to performing single-mode failure probability when more reliability-based modes are consid-
analysis when multiple mode system reliability could have been ered. This is done in Fig. 5, which, in contrast to Fig. 3, has three
done. The latter part of this section also investigates the effect of additional modes (b3, b6, b7) adjacent to the mode corresponding
increasing the number of failure modes to eight. to the local minimum b1, and three additional modes (b4, b5, b8)
As shown in Fig. 3, the upper clay layer is 18 m thick, with un- adjacent to the mode corresponding to the local minimum b2.
drained shear strength cu1; the lower clay layer is 10 m thick, with It was noted, for example in Ang and Tang [1], that the bimodal
undrained shear strength cu2. The undrained shear strengths are bounds on failure probability of systems with multiple failure
lognormally distributed and independent. A hard layer exists be- modes will depend on the ordering of the individual failure modes.
low the second clay layer. It was suggested, for example in Madsen et al. [18], Melchers [2]
Since the shear strengths are characterized by cu1 and cu2, with and Haldar and Mahadevan [19], that ordering the failure modes
uu = 0, Bishops simplied method and the ordinary method of in decreasing probabilities of failure (i.e., increasing b values) will
slices yield the same results, and either can be used. Also, in this lead to closer bounds. This has been done in Fig. 5, yielding
case where the upper clay layer is weaker than the lower clay 0.416% 6 PF,sys 6 0.441%, practically the same range as that in
layer, it is logical to locate two reliability-based critical slip cir- Fig. 3 when only the two local minimum modes were considered.
cles, as shown in Fig. 3, one entirely in the upper clay layer and A simple VBA code was also created to investigate the effects of
the other passing through both layers. The FORM reliability indi- all possible permutations (8!) of the eight failure modes on the sys-
ces for the two modes are 2.795 and 2.893, respectively. It is tem failure probability bounds. The narrowest bounds were
202 B.K. Low et al. / Computers and Geotechnics 38 (2011) 196204

Fig. 5. System reliability analysis considering eight failure modes, including the two reliability-based critical modes of Fig. 3.

0.432% 6 PF,sys 6 0.441% (the same as the two-critical-mode case), normal distributions and are statistically independent. The statis-
obtained when the eight b values (and associated vectors n) are tics of these three variables are shown in Fig. 6. Let x = (w, M0,
in the order b2, b1, b4, b5, b8, b6, b3, b7. V0). The three performance functions for the beam are (Ang and
That the system reliability bounds of the eight modes in Fig. 5 Tang [1]):
differ little from the bounds of the two local minimum modes of
Fig. 3 can be attributed to the strong correlations among modes g 1 x M 0  wL2 =8 12
1, 3, 6 and 7, and among modes 2, 4, 5 and 8, as seen from the very
high (1.0) intermodal correlation coefcients of q13, q16, q17, and g 2 x V 0  wL=2 13
of q24, q25, q28. Physically this means that the direction vectors
!
(linking the mean value point and design points of the failure wL2 wM 0
modes) are nearly parallel for modes 1, 3, 6 and 7, and for modes g 3 x 1  14
8M 0 2V 20
2, 4, 5 and 8. The implied overlapping of the failure probability con-
tents for modes 1, 3, 6 and 7, and also for modes 2, 4, 5 and 8 means where L denotes beam length and is equal to 6.1 m.
that it is sufciently accurate to calculate the bounds for the sys- Based on the FORM reliability index of each failure mode as well
tem failure probability by considering only the two stationary val- as the correlation coefcients between various failure modes ob-
ues of reliability index, namely b1 and b2 of Fig. 3. tained from Eqs. (7a) and (7b), the bounds of the system failure
probability are 5.775.84%, as shown in Fig. 6. For comparison,
6. Use of SORM reliability index in system reliability evaluation the lower and upper bounds in Ang and Tang [1] are 5.82% and
5.87%, respectively, which were obtained based on manual calcula-
Example 7.2 in Ang and Tang [1] is used to illustrate how the tion with round-off.
bounds for system failure probability can be rendered more accu- Six Monte Carlo simulations each with Latin Hypercube sam-
rate (by using reliability index based on SORM rather than FORM) pling of size 100,000 yielded probabilities of failure varying from
when one of the three limit state surfaces has signicant curvature. 6.21% to 6.29%, with an average of 6.25%, which is outside the
In this example, a simply supported beam subjected to uniformly 5.775.84% bounds shown in Fig. 6. This is because the reliability
distributed load w with three possible failure modes is investi- index bi for each failure mode is calculated based on FORM, and
gated. The three failure modes are shear failure, bending failure, the relationship P(Ei) = U(bi) used in the code of Fig. 1 for Eqs.
and failure under the combined effect of bending and shear. The (9), (10a), (10b), (11a), (11b) is exact only if the random variables
uncertain variables involved in estimating the performance of the are normally distributed and the limit state surface is a hyperplane.
beam are the load w, the shear capacity V0, and the bending capac- In this case the random variables w, M0 and V0 are indeed normally
ity M0. These three random variables are assumed to follow the distributed, and two of the three limit state surfacesEqs. (12) and
B.K. Low et al. / Computers and Geotechnics 38 (2011) 196204 203

Fig. 6. System reliability analysis of a simply supported beam, and use of SORM to account for the curvature of the limit state surface of mode 3.

(13)are linear and are hence hyperplanes in the 3-D space of w, 6.21% 6 PF,sys 6 6.29% mentioned at the beginning of the previous
M0 and V0. However, the third limit state surface, Eq. (14), is non- paragraph.
linear. It would be interesting to see whether reliability indices in-
ferred from SORM (which accounts to some extent for curvatures
of limit state surfaces near the FORM design points) can lead to 7. Summary and conclusions
more accurate bounds on the system failure probability. This is
done below. This paper began by providing a simple equation (Eq. (3)) and
For the limit state surface of Eq. (14), a SORM analysis yielded an example to relate the correlated but unrotated equivalent stan-
probability of failure equal to 6.20%. The equivalent FORM b3 for dard normal vector n of the Low and Tang 2007 efcient spread-
use in Eqs. (9), (10a), (10b), (11a), (11b) (and in their Excel codes sheet FORM procedure [14] to the uncorrelated but rotated
of Fig. 1) should therefore be b3 = U1(10.062) = Norm- vector u of the classical mathematical FORM procedure. The
SInv(0.938) = 1.538, denoted as b3,SORM at the bottom of Fig. 6, in component reliability indices and intermodal correlation matrix
contrast to the original FORM b3 of 1.573. (The values of b1 and (Eqs. (7b) or (8b)), which are required for system reliability evalu-
b2 remain unchanged whether one uses FORM or SORM, because ation, can be obtained by either FORM procedure. The paper next
their corresponding limit state surfaces are hyperplanes with zero presented a simple program code (Fig. 1) for automatic and ef-
curvature.) Another assumption is needed, and is logical: the direc- cient implementation (in a ubiquitous spreadsheet platform) of
tion cosine vector of b3,SORM is the same as that of the original b3, the KouniasDitlevsen bimodal bounds (Eqs. (9), (10a), (10b),
that is, n3,SORM/b3,SORM  n3/b3. With these, and arranging the (11a), (11b)) for systems with multiple failure modes. The short
modes in increasing values of b (i.e., decreasing values of failure program code can be rewritten readily in other programming lan-
probabilities), namely b3,SORM, b1, b2, the bounds on system failure guages for other platforms (e.g., Matlab), if desired. The system
probability, 6.20% 6 PF,sys 6 6.23%, computed at the bottom of reliability evaluation procedure was then illustrated using three
Fig. 6, are much more consistent with the Monte Carlo bounds of examples, namely a semi-gravity retaining wall with two failure
204 B.K. Low et al. / Computers and Geotechnics 38 (2011) 196204

modes, a slope with two and eight failure modes, and a loaded References
beam with three failure modes. The following may be noted:
[1] Ang HS, Tang WH. Probability concepts in engineering planning and design.
Decision, risk, and reliability, vol. 2. New York: John Wiley; 1984.
(i) The code of Fig. 1 renders it virtually effortless to compute [2] Melchers RE. Structural reliability analysis and prediction. 2nd ed. New
the lower and upper bounds of system failure probability York: John Wiley; 1999.
once the component reliability indices and the intermodal [3] Zhao GF, Jin WL, Gong JX. Structural reliability theory. China Architecture &
Beijing Building Press; 2000 [in Chinese].
correlation matrix have been obtained, as explained in the [4] Ramachandran K. System reliability bounds: a new look with improvements.
three steps of the retaining wall example. Civ Eng Environ Syst 2004;21(4):26578.
(ii) For the example of the slope with two clay layers, system [5] Ditlevsen O, Madsen HO. Structural reliability methods. Internet edition 2.2.5;
2005. <http://www.mek.dtu.dk/staff/od/books.htm>.
FORM reliability analysis is an attractive and efcient
[6] Tung YK, Yen BC, Melching CS. Hydrosystems engineering reliability
alternative to Monte Carlo simulations and importance assessment and risk analysis. New York: McGraw-Hill; 2006.
sampling method. It is sufcient to use only the stationary [7] Cornell CA. Bounds on the reliability of structural systems. J Struct Div, ASCE
values (i.e., local minimums, or troughs) of component 1966;93(1):171200.
[8] Ditlevsen O. Narrow reliability bounds for structural systems. J Struct Mech
reliability indices and their respective design points in sys- 1979;7(4):45372.
tem reliability evaluation. For the case in hand, the two [9] Moses F. System reliability developments in structural engineering. Struct Saf
stationary b values of 2.795 and 2.893 (Fig. 3) and their 1982;1(1):313.
[10] Zhang YC. High-order reliability bounds for series systems and application to
design points (vectors n1 and n2 obtained using the Low structural systems. Comput Struct 1993;46(2):3816.
and Tang 2007 FORM approach [14], or vectors u1 and u2 [11] Song J, Der Kiureghian A. Bounds on systems reliability by linear
of the classical FORM approach) lead to practically the programming. J Eng Mech 2003;129(6):62736.
[12] Zhao YG, Zhong WQ, Ang HS. Estimating joint failure probability of series
same narrow bounds on system failure probability as the structural systems. J Eng Mech 2007;133(5):58896.
eight modes shown in Fig. 5, of which six modes are adja- [13] Song J, Kang WH. System reliability and sensitivity under statistical
cent to the two local minimums, but are not local mini- dependence by matrix-based system reliability method. Struct Saf
2009;31(2):14856.
mums themselves. [14] Low BK, Tang WH. Efcient spreadsheet algorithm for rst-order reliability
(iii) The two-layered slope example has two equally likely failure method. J Eng Mech ASCE 2007;133(12):137887.
modes. If one recognizes only one failure mode, the failure [15] Low BK, Tang WH. Reliability analysis using object-oriented constrained
optimization. Struct Saf 2004;26(1):6989.
probability would have been underestimated, whether by
[16] Der Kiureghian A, Liu PL. Structural reliability under incomplete probability
FORM, importance sampling, or MC simulations. On the information. J Eng Mech ASCE 1986;112(1):85104.
other hand, provided one accounts for the two most critical [17] Anton H, Rorres C. Elementary linear algebra: applications version. 9th
failure modes, the computed system failure probabilities are ed. New York: Wiley; 2005.
[18] Madsen HO, Krenk S, Lind NC. Methods of structural safety. Englewood Cliffs,
similar whether one uses FORM or MC simulations, with New Jersey: Prentice Hall; 1986.
FORM (in combination with bimodal bounds) being more [19] Haldar A, Mahadevan S. Probability, reliability and statistical methods in
efcient. engineering design. New York: John Wiley; 1999.
[20] Kounias EG. Bounds for the probability of a union with applications. Ann Math
(iv) In principle, one needs to consider all permutations of failure Stat 1968;39(6):21548.
modes in order to get the narrowest bounds. For examples 2 [21] Hunter D. An upper bound for the probability of a union. J Appl Probab
and 3, the established suggestion of ranking failure modes in 1976;3(3):597603.
[22] Rosenblueth E. Point estimates for probability moments. Proc Natl Acad Sci
decreasing failure probabilities is sufcient to produce rea- 1975;72(10):38124.
sonably sharp failure bounds. [23] Hasofer AM, Lind NC. Exact and invariant second-moment code format. J Eng
(v) As shown in the last example, SORM-based reliability indices Mech ASCE 1974;100(1):11121.
[24] Rackwitz R, Fiessler B. Structural reliability under combined random load
may improve the accuracy of bounds on system failure prob- sequences. Comput Struct 1978;9:48494.
ability if some of the more-probable failure modes have limit [25] Low BK. Reliability-based design applied to retaining walls. Geotechnique
state surfaces with signicant curvatures. In many geotech- 2005;55(1):6375.
[26] Low BK. Practical probabilistic slope stability analysis. In: Proceedings, soil and
nical problems (including those with complicated or implicit
rock America, vol. 2, MIT, Massachusetts, June 2003, Verlag Glckauf GmbH
performance functions such as slope stability) the limit state Essen; 2003. p. 277784.
surfaces encountered are almost hyperplanes in the vicinity [27] Wu TH, Kraft LM. Safety analysis of slopes. J Soil Mech Found Div ASCE
of the design points, and system reliability evaluation based 1970;96(2):60930.
[28] Vanmarcke EH. Reliability of earth slopes. J Geotech Eng ASCE 1977;103(11):
on FORM results are sufciently accurate for practical 124766.
purposes. [29] Vanmarcke EH. Probabilistic stability analysis of earth slopes. Eng Geol
1980;16:2950.
[30] Chen ZY, Morgenstern NR. Extensions to the generalized method of slices for
stability analysis. Can Geotech J 1983;20(1):10419.
[31] Li KS, Lumb P. Probabilistic design of slopes. Can Geotech J 1987;24(4):52035.
[32] Christian JT, Ladd CC, Baecher GB. Reliability applied to slope stability analysis.
Appendix J Geotech Eng ASCE 1994;120(12):2180207.
[33] Morgenstern NR. Toward landslide risk assessment in practice. In: Cruden DM,
The lower triangular matrix L obtained from the LU decomposi- Fell R, editors. Proceedings of the international workshop on landslide risk
assessment, vol. 1524. Rotterdam: Balkema; 1997.
tion of the correlation matrix R is useful in relating the correlated [34] Low BK, Tang WH. Probabilistic slope analysis using Janbus generalized
standard normal vector n of the Low and Tang 2007 FORM ap- procedure of slices. Comput Geotech, vol. 21. U.K.: Elsevier; 1997 (2): 121-142.
proach [14] to the uncorrelated (but rotated) standard normal vec- [35] Low BK, Gilbert RB, Wright SG. Slope reliability analysis using generalized
method of slices. J Geotech Geoenviron Eng ASCE 1998;124(4):35062.
tor u of the classical FORM approach, via Eq. (3). The very short [36] Au SK, Beck JL. Estimation of small failure probabilities in high dimensions by
Excel VBA code (about 25 lines) for Cholesky LU factorization is subset simulation. Probab Eng Mech 2001;16(4):26377.
available in the public-domain and free, for example from http:// [37] Nadim F, Einstein HH, Roberds W. State of the art paper 3 probabilistic stability
analysis for individual slopes in soil and rock. In: International conference on
www.vbnum.com/math.
landslide risk management, Vancouver, Canada; 31 May2 June, 2005. p. 6398.
(The Cholesky LU decomposition is not required for the system [38] Low BK, Lacasse S, Nadim F. Slope reliability analysis accounting for spatial
reliability approach of this paper, since Eq. (7a) and (7b), instead of variation. Georisk: assessment and management of risk for engineered systems
and geohazards, vol. 1, no. 4. London: Taylor & Francis; 2007. p. 17789.
Eqs. (3) and (8a), (8b), can be used to obtain the intermodal corre-
[39] Ching JY, Phoon KK, Hu YG. Efcient evaluation of reliability for slopes with
lation matrix had the Low and Tang 2007 approach [14] been used circular slip surfaces using importance sampling. J Geotech Geoenviron Eng,
to obtain the component bi and ni.) ASCE 2009;135(6):76877.

Das könnte Ihnen auch gefallen