You are on page 1of 13

Low, B. K. (2005). Géotechnique 55, No.

1, 63–75

Reliability-based design applied to retaining walls


B. K . L OW *

Geotechnical engineers have long used the lumped factor Les géotechniciens utilisent depuis longtemps la méthode
of safety approach in the design of foundations and à facteurs regroupés de sécurité pour la conception de
retaining walls. A more recent alternative is the limit fondations et de murs de soutènement. Il existe une
state approach using partial factors, for example as alternative plus récente, la méthode d’état limite utilisant
suggested in Eurocode 7. Yet another approach is per- des facteurs partiaux, comme ce qui est suggéré dans
haps more flexible and rational: design based on a target Eurocode 7 par exemple. Pourtant, il existe une autre
reliability index that reflects the uncertainty of the para- méthode qui est peut-être plus flexible et plus ration-
meters and their correlation structure. Among the var- nelle : une conception basée sur un indice de fiabilité qui
ious versions of reliability indices, the Hasofer–Lind reflète l’incertitude des paramètres et leur structure de
index and first-order reliability method (FORM) are corrélation. Parmi les diverses versions d’indices de fiabi-
more consistent. This paper illustrates practical reliabil- lité, l’indice Hasofer-Lind et FORM (méthode de fiabilité
ity-based design procedures for retaining walls based on de premier ordre) sont les plus constants. Cet exposé
the Hasofer–Lind index and FORM. Correlated normal illustre les procédures pratiques de conception basées sur
and non-normal random variables are considered. The la fiabilité pour des murs de soutènement, procédures
efficient spreadsheet-based probabilistic approach, based basées sur l’indice Hasofer-Lind et FORM. Nous consid-
on the work of Low & Tang, achieves the same result as érons les variables aléatoires corrélées normales et non
the Hasofer–Lind method and FORM, but uses an in- normales. La méthode probabiliste à tableaux, basée sur
tuitive expanding dispersion ellipsoid perspective that les travaux de Low & Tang, arrive au même résultat que
greatly simplifies the computations and interpretations. la méthode Hasofer-Lind et FORM mais utilise une per-
The differences between a reliability-based design and spective ellipsoı̈de intuitive de dispersion croissante qui
one based on partial factors will be discussed. Sensitivity simplifie énormément les calculs et interprétations. Nous
information as conveyed in a reliability analysis will be discutons des différences entre une conception basée sur
studied. The probabilities of failure inferred from relia- la fiabilité et une conception basée sur des facteurs
bility indices will be compared with Monte Carlo simula- partiels. Nous étudions l’information de sensitivité telle
tions. This paper deals only with certain aspects of qu’elle est donnée dans une analyse de fiabilité. Nous
reliability, namely methodology and concepts, and not comparons les probabilités d’une défaillance dérivée des
reliability in its widest sense. indices de fiabilité à des simulations de Monte Carlo. Cet
exposé ne traite que de certains aspects de fiabilité,
nommément de méthodologie et de concepts et non pas
KEYWORDS: design; retaining walls; statistical analysis de fiabilité dans son sens le plus large.

INTRODUCTION earth thrust Pa , with levers Armah and Armav respectively;


The objectives of this paper are to describe an alternative and 9 and  are the internal friction angle of the retained
interpretation of the Hasofer–Lind reliability index (Hasofer soil and the soil–wall interface friction angle respectively.
& Lind, 1974) in the original space of the random variables, With the notations as defined in Fig. 1, if H ¼ 6 m, ªwall ¼
and to illustrate practical reliability-based geotechnical de- 24 kN/m3 , º ¼ 108, Æ ¼ 908, ªsoil ¼ 18 kN/m3 , a ¼ 0.4 m,
sign of retaining walls using object-oriented constrained b ¼ 1.8 m, 9 ¼ 358 and  ¼ 208, then the factor of safety
optimisation in the ubiquitous spreadsheet platform (Low & against overturning is Fs  1.65, by equation (1). In the
Tang, 1997a, 2004). The difference between reliability-based two-dimensional space of 9 and  one can also plot the Fs
design and one based on partial factors will be discussed. contours for different combinations of 9 and , as shown in
Sensitivity information conveyed by the reliability analysis Fig. 2, where the average point (9 ¼ 358 and  ¼ 208) is
will be studied. situated on the contour (not plotted) of 1.65. Design is
Consider the semi-gravity retaining wall shown in Fig. 1. considered satisfactory with respect to overturning if the
A long-established deterministic approach evaluates the factor of safety by equation (1) is not smaller than a certain
lumped factor of safety (Fs ) against rotational failure as value (e.g. when Fs > 1.5).
A more recent and logical approach (e.g. Eurocode 7;
W 1 3 Arm1 þ W 2 3 Arm2
Fs ¼ ¼ f ð9, , . . .Þ (1) CEN, 1994) applies partial factors to the parameters in the
Pah 3 Armah  Pav 3 Armav evaluation of resisting and overturning moments. Design is
where W1 and W2 are the component weights of the semi- acceptable if
gravity wall, with horizontal lever distances Arm1 and Arm2 X
respectively, measured from the toe of the wall; Pah and Pav ðResisting moments, factoredÞ
are the horizontal and vertical components of the active X
> ðOverturning moments, factoredÞ (2)
Manuscript received 30 April 2004; revised manuscript accepted 21
October 2004.
Discussion on this paper closes 1 August 2005, for further details A third approach is reliability-based design, in which the
see p. ii. uncertainties and correlation structure of the parameters are
* School of Civil and Environmental Engineering, Nanyang represented by a one-standard-deviation dispersion ellipsoid
Technological University, Singapore. (Fig. 2) centred at the mean-value point, and safety is

63
64 LOW
ì ó
x* mean StDev nx
H ãwall ë á ö¢ ä ãsoil a b
6 24 10 90 (x*, in radians) 18 0.4 1.8 ö¢ 26.428 35 3.5 22.449

0.1745 1.5708 0.4613 0.2711 ä 15.534 20 2 22.233

ca 100 100 15 0.000


Boxed cells contain equations

Correlation matrix
ö¢ 1 0.8 0
Ka Pa
1 K ãH 2 ä 0.8 1 0
0.3978 128.9 Pa 5 a (x* 2 ì)
2 nx 5
ó
ca 0 0 1

ö ä ca

Force Arm Moment


=SQRT(MMULT(TRANSPOSE(nx),MMULT(MINVERSE(crmat),nx)))
Pav 34.52 1.8 62.136
Ctrl + Shift, then Enter
W1 100.8 0.9333 94.08

W2 57.6 1.6 92.16 PerFn1 â PerFn2

192.9 248.4 ÓM 3 3 1028 2.491 55.812

Overturning mode Sliding mode

Pah 124.2 2 248.4

ì ó
a x* mean StDev nx
Soil : Unit weight ã ö¢ 29.136 35 3.5 21.6756
Angle of friction ö¢
ä 17.207 20 2 21.3967

ca 60.866 100 15 22.6089


1
Pa 5 K ãH 2
2 a
H
W2
â PerFn2
W1 ä 1 á 2 90°
3.102 0.00

Sliding mode

b Adhesion ca Stiff clay

Fig. 1. Reliability analysis of overturning failure mode and sliding mode, for correlated normal random variables, using
Microsoft Excel spreadsheet

gauged by a reliability index that is the shortest distance soil shear strength parameters c9 and 9) show the same
(measured in units of directional standard deviations, R/r) values of lumped factor of safety, yet case A is clearly safer
from the safe mean-value point to the most probable failure than case B. The higher reliability of case A over case B
combination of parameters (the ‘design point’) on the limit will be correctly revealed when the reliability indices are
state surface (defined by Fs ¼ 1.0). Furthermore, the prob- computed. On the other hand, a slope may have a computed
ability of failure (Pf ) can be estimated from the reliability lumped factor of safety of 1.5, and a particular foundation
index  using the established equation Pf ¼ 1  Ö() ¼ (with certain geometry and loadings) in the same soil may
Ö(), where Ö is the cumulative distribution (CDF) of the have a computed lumped factor of safety of 2.5, as in case
standard normal variate. The relationship is exact when the C of Fig. 3(b). Yet a reliability analysis may show that they
limit state surface is planar and the parameters follow both have similar levels of reliability.
normal distributions, and approximate otherwise. As will be made clear in the following section, the design
The merits of a reliability-based approach over the lumped point (Fig. 2) is the most probable failure combination of
factor-of-safety approach are illustrated in Fig. 3(a), in parametric values. The ratios of the respective parametric
which case A and case B (with different average values of values at the centre of the dispersion ellipsoid (correspond-
RELIABILITY-BASED DESIGN APPLIED TO RETAINING WALLS 65
Fs 5 1.75 One-standard-deviation parisons are made with Monte Carlo simulations, and simila-
45 dispersion ellipse rities and differences between reliability-based design and
1.50
limit state design using partial factors are discussed. Reason-
40 able statistical properties are assumed for the hypothetical
1.25
cases presented, although actual determination of the statis-
ìö¢ â-ellipse tical properties is not covered. Only parametric uncertainty
35
r is considered, and model uncertainty is not dealt with. Hence
this paper is about reliability method and perspectives, and
30 not reliability in its widest sense. The focus is on introdu-
R
cing an efficient and rational design approach using the
Design point ubiquitous spreadsheet platform.
25
ö¢: degrees

Safe
domain
20 Limit state surface
(Fs 5 1·0, or PerFn1 5 0) RELIABILITY ANALYSIS OF A SEMI-GRAVITY
RETAINING WALL USING A SPREADSHEET
15
Three geotechnical failure modes need to be considered in
Failure the design of a semi-gravity retaining wall: rotation about
10 â 5 R/r 5 2·49 the toe of the wall, horizontal sliding along the base of the
domain
wall, and bearing capacity failure of the soil beneath the
5
wall, under the inclined and eccentric resultant load derived
µä from the weight of the wall and the active earth thrust Pa
acting on the back of the wall.
0 The reliability analysis of the retaining wall in Fig. 1 is
0 10 20 30
performed using Microsoft Excel software and its built-in
ä: degrees optimisation program Solver. For simplicity, it is assumed
that the in-situ stiff clay offers ample reliability against
Fig. 2. Design point and normal dispersion ellipsoids illustrated bearing capacity failure: hence only the rotation and sliding
in the space of ö9 and ä. The correlation coefficient r is 0.8, as
shown in Fig. 1
modes will be considered. (Reliability analysis with respect
to bearing capacity limit state was illustrated in Low &
Phoon, 2002).
ing to the mean values) to those at the design point are
similar to the partial factors in limit state design, except that
these factored values at the design point are arrived at Limit state functions with respect to rotation and sliding
automatically (and as by-products) via spreadsheet-based The Coulomb active earth pressure coefficient will be
constrained optimisation. The reliability-based approach is used, which is based on the assumption of a plane slip
thus able to reflect varying parametric sensitivities from case surface in the soil. The value is practically the same as the
to case in the same design problem (e.g. Fig. 3(a)) and more rigorous Caquot and Kerisel active earth pressure coef-
across different design realms (e.g. Fig. 3(b)). ficient which assumes a logarithmic spiral slip surface, as
In the next section a practical spreadsheet-automated given in BS 8002 (BSI, 1994), for example. The Coulomb
reliability analysis is first illustrated and discussed for a active coefficient Ka is
simple retaining wall with only two random variables. The
intuitive expanding dispersion ellipsoid perspective and the Ka ¼
meaning of reliability index are explained. This is followed " #2
by the reliability-based design of an anchored wall involving sinðÆ  9Þ=sin Æ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
nine random variables, which are first treated as correlated sinðÆ þ Þ þ sinð9 þ Þsinð9  ºÞ=sinðÆ  ºÞ
normal variates, and then as correlated non-normals. Com-
(3)

One-standard-deviation Slope
Fs ⫽ 3·0
Fs ⫽ 1·4 dispersion ellipsoid Foundation

Fs ⫽ 1·2 A 1·5
2·0

1·2 C
Safe
c′

c′

Fs ⫽ 1·0 B Fs ⫽ 1·0
(Foundation)

Unsafe, Fs ⬍ 1·0 Fs ⫽ 1·0


Unsafe
(Slope)

φ′ φ′
(a) (b)
Fig. 3. Schematic scenarios showing possible limitations of lumped factor of safety. (a) Cases A
and B have the same lumped Fs 1.4, but Case A is clearly more reliable than Case B. (b)
Case C may have Fs 1.5 for a slope and Fs 2.5 for a foundation, and yet have a similar
level of reliability
66 LOW
where Æ and º are the inclinations (Fig. 1) of the back of minverse are Microsoft Excel’s built-in functions, each being
the wall and the retained fill surface with respect to the a container of program codes for matrix operations.
horizontal, 9 is the angle of internal friction of the soil, Initially the column labelled ‘x*’ is assigned the mean
and  is the interface friction angle between the concrete values. The spreadsheet’s built-in Solver optimisation routine
wall and the soil. The water table is below the base of the is then invoked, to ‘Minimize’ , ‘By Changing’ the three
retaining wall. The active earth thrust Pa (kN/m) is taken to ‘x*’ values, ‘Subject To’ the constraints PerFn1 , ¼ 0 and
act at a height of H/3 above the base of the wall and at an the x* value of 9 . ¼ º. The Solver option ‘Use automatic
angle  with the normal to the back of the wall: that is, at scaling’ can also be activated. (The constraint PerFn2 , ¼
an angle ( + Æ  908) with the horizontal. In Fig. 1, the 0 would be specified instead when performing reliability
following equations have been set up, for a wall with a analysis with respect to the sliding mode.)
vertical back (i.e. Æ ¼ 908): The solution ( ¼ 2.491) obtained by Solver is shown in
Fig. 1 for the overturning mode. The solution for the sliding
Pav ¼ Pa sin , Armav ¼ b, W 1 ¼ 0:5ªwall ð b  aÞ H, mode ( ¼ 3.102) is also shown in the insert. The spread-
Arm1 ¼ 23ð b  aÞ sheet approach is simple and intuitive because it works in
the original space of the variables. It does not involve the
a H
W 2 ¼ ªwall aH, Arm2 ¼ b  , Pah ¼ Pa cos , Armah ¼ orthogonal transformation of the correlation matrix, and
2 3 iterative numerical partial derivatives are done automatically
The performance functions (PerFn1 and PerFn2) with re- on spreadsheet objects that may be implicit or contain
spect to rotational mode and sliding mode are respectively codes.
The x* values obtained in the top right of Fig. 1 represent
PerFn1 :¼ W 1 Arm1 þ W 2 Arm2 þ Pav Armav  Pah Armah the most probable failure point on the limit state surface. It
(4) is the point of tangency (Fig. 2) of the expanding dispersion
PerFn2 :¼ b 3 ca  Pah (5) ellipsoid with the overturning limit state surface and, in the
bottom insert, with the sliding limit state surface. The
If the base resistance to sliding has a frictional component following may be noted:
(W1 + W2 + Pav )tan a , it can be added to the adhesion (a) For the overturning limit state, the x* values shown in
component b 3 ca without affecting the solution procedure the top right of Fig. 1 render equation (4) (PerFn1)
described below. equal to zero. Hence the point represented by these x*
The matrix formulation (Ditlevsen, 1981) of the Hasofer– values lies on the overturning limit state surface. That
Lind index  is the nx value of ca turns out to be zero (by virtue of the
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi x* value of ca not deviating at all from the mean ca )
 ¼ min x2 F ð x  ÞT C 1 ð x  Þ (6a) implies that the overturning mode is insensitive to the
base adhesion ca , as one would expect when the
or, equivalently, overturning mode does not depend on ca at all. This
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
    means that PerFn1 and its associated  index can be
xi   i T 1 xi   i plotted in the two-dimensional space of 9 and , as
 ¼ min x2 F ½ R (6b)
i i shown in Fig. 2, to scale. The limit state surface
separating the safe domain from the unsafe domain is
where x is a vector representing the set of random variables described by equation (4). The one-standard-deviation
xi ,  is the vector of mean values  i , C is the covariance ellipse and the  ellipse are tilted because the
matrix, R is the correlation matrix, i is the standard correlation coefficient between 9 and  is 0.8. The
deviation, and F is the failure domain. Low & Tang (1997b, design point is where the expanding dispersion ellipse
2004) used equation (6b) in preference to equation (6a) touches the limit state surface, at the point represented
because the correlation matrix R is easier to set up, and by the x* values of Fig. 1.
conveys the correlation structure more explicitly than the (b) As a multivariate normal dispersion ellipsoid expands,
covariance matrix C. its expanding surfaces are contours of decreasing
For the reliability example in Fig. 1, the random variables probability values, according to the established prob-
are soil friction angle 9, the interface friction angle , and ability density function of the multivariate normal
the base adhesion ca . It is expected that 9 and  are distribution:
positively correlated, and a correlation coefficient of 0.8 is 1
adopted, as shown by the correlation matrix in Fig. 1. The f ð xÞ ¼ n exp[12ð x  ÞT C 1 ð x  Þ] (9a)
parameters 9,  and ca in the above equations read their ð2Þ jCj0:5
2

values from the column labelled ‘x*’. The x* values, and the 1
functions dependent on them, change during the optimisation ¼ n exp[122 ] (9b)
search for the most probable failure point. ð2Þ 2 jCj0:5
The column labelled ‘nx’ in Fig. 1 contains the equation where  is defined by equation (6a) or (6b), without the
x   i ‘min’. Hence, to minimise  (or 2 in the above
nxi ¼ i (7)
i multivariate normal distribution) is to maximise the
value of the multivariate normal probability density
where  i stands for the mean value and i stands for the function, and to find the smallest ellipsoid tangent to
standard deviation. Equation (6b) for the reliability index is the limit state surface is equivalent to finding the most
entered in the cell labelled  as an array formula: probable failure point (the design point). This intuitive
¼ sqrt(mmult(transpose(nx), mmult(minverse(crmatrix), nx))) and visual understanding of the design point is
consistent with the more mathematical approach in
(8) Shinozuka (1983), equations 4, 41, and Fig. 2, in which
all variables were transformed into their standardised
followed by ‘Enter’ while holding down the ‘Ctrl’ and forms and the limit state equation had also to be
‘Shift’ keys. In the above formula, mmult, transpose and written in terms of the standardised variables. The
RELIABILITY-BASED DESIGN APPLIED TO RETAINING WALLS 67
differences between the present original space and ρ ⫽ ⫺0·8 ρ ⫽ 0·8
Shinozuka’s standardised space of variables will be ρ⫽0
39 ρ ⫽ ⫺0·99
further discussed in ( j) below. ρ ⫽ 0·99
(c) Therefore the design point, being the first point of
contact between the expanding ellipsoid and the limit
state surface in Fig. 2, is the most probable failure 37 σφ′
point with respect to the safe mean-value point at the
centre of the expanding ellipsoid, where Fs ¼ 1.654
35 µφ′
against overturning. The reliability index  is the axis

φ′: degrees
ratio (R/r) of the ellipse that touches the limit state
surface and the one-standard-deviation dispersion
33
ellipse. By geometrical properties of ellipses, this co-
directional axis ratio is the same along any ‘radial’
direction. 31
(d) For each parameter, the ratio of the mean value to the σδ
x* value is similar in nature to the partial factors in
limit state design (e.g. Eurocode 7). However, in a 29
reliability-based design one does not specify the partial
factors. The design point values (x*) are determined µδ
automatically and reflect sensitivities, standard devia- 27
tions, correlation structure, and probability distributions 14 16 18 20 22 24 26
in a way that prescribed partial factors cannot. δ: degrees
(e) One may investigate the significance of assuming
uncorrelated 9 and , by replacing the two 0.8 values Fig. 5. Rotation of one-standard-deviation dispersion ellipsoid as
in the correlation matrix (Fig. 1) with 0, reinitialising correlation coefficient r changes
x* values to mean values, and re-invoking Solver. A
higher reliability index of  ¼ 3.13 is obtained, as
shown in Fig. 4, to scale. The two ellipses in Fig. 4 are ( f ) Figure 5 provides a graphical appreciation of positive
canonical (i.e. non-tilted) because the correlation coef- and negative correlation for 9 and , corresponding to
ficient is zero. Thus in this case when 9 and  are the mean and standard deviation values shown in Fig.
both resistance parameters, ignoring their positive 1. These one-standard-deviation dispersion ellipses were
correlation leads to an unconservative estimate of plotted based on equation (6b), with  ¼ 1.0, for
reliability against overturning mode:  ¼ 2.49 in Fig. different values of the correlation coefficient r. Since in
2 compared with  ¼ 3.13 in Fig. 4. (On the other this case the internal friction angle and interface
hand, had the two random variables been 9 and soil friction angle could only be positively correlated, the
unit weight ª, which logically should also be positively two negatively tilted ellipses are shown as dashed
correlated, ignoring the correlation would result in an curves.
underestimation of the reliability index  against the (g) The results of reliability analysis for the sliding limit
overturning limit state.) state given by equation (5) are shown in the bottom-
right insert in Fig. 1. A reliability index of 3.102 is
obtained. The x* values and the three non-zero values
One-standard-deviation of nx imply that the sliding limit state is sensitive to all
50 dispersion ellipse three parameters, 9,  and ca , with ca being highest on
Fs 5 1·75 â-ellipse the sensitivity scale. The mean value point, at (35, 20,
45 100), is safe against sliding; but sliding failure occurs
1·50 when the 9,  and ca values are decreased to the
40 values shown: (29.136, 17.207, 60.866). The distance
1·25
from the safe mean-value point to this most probable
35 ìö¢ failure combination of parameters, in units of direc-
r
tional standard deviations, is the reliability index ,
30
equal to 3.102 in this case. One may further note that
the ratios of /x* (akin to the partial factors of
ö¢: degrees

R Eurocode 7) are not pre-defined. This ability of the


25 Limit state surface reliability analysis to seek the most probable x* values
Safe
(Fs 5 1·0) domain (corresponding to a target reliability level) without
20 presuming any partial factors is an important and
desirable feature.
15 Failure (h) As mentioned in the introduction, the probability of
Design point domain
failure (Pf ) can be estimated from the reliability index
10 . Microsoft Excel’s built-in function NormSDist(.) can
â 5 R/r 5 3·13 be used to compute Ö(.) and hence Pf. Thus for the
overturning mode of Fig. 1, Pf ¼ NormSDist(2.491)
5
ìä ¼ 0.006 37, and for the sliding mode Pf ¼
NormSDist(3.102) ¼ 0.000 961. These values com-
0 pare remarkably well with the values 0.006 33 and
0 10 20 30
0.001 08 in Table 1, obtained from Monte Carlo
ä: degrees
simulation with 800 000 trials using the commercial
Fig. 4. Design point and normal dispersion ellipsoids illustrated simulation software @RISK (www.palisade.com). The
in the space of ö9 and ä. Correlation coefficient r is 0 correlation matrix was accounted for in the simulation.
68 LOW
Table 1. Probabilities of failure of Fig. 1’s semi-gravity retaining wall, using Monte
Carlo simulations
Monte Carlo Overturning Sliding Simultaneous Total failure
simulations failure failure overturning probability
(1) (2) and sliding (3) (1) + (2)  (3)
200 000 trials .
0 006 05 .
0 000 99 0.000 21 0.006 83
500 000 trials 0.006 29 0.001 10 0.000 28 0.007 11
800 000 trials 0.006 33 0.001 08 0.000 26 0.007 15

The excellent agreement between 0.006 37 from the shown in Fig. 1, translates to
reliability index and 0.006 33 from Monte Carlo
simulation is hardly surprising given the almost linear 0:006 37 < P(failure) < 0:007 32 (10b)
limit state surface and normal variates shown in Fig. 2.
However, for the anchored wall shown in the next
section, where nine random variables are involved and ( j) Figure 2 defines the reliability index  as the
non-normal distributions are used, the nine-dimensional dimensionless ratio R/r, in the direction from the
equivalent hyperellipsoid and the limit state hypersur- mean-value point to the design point. This is the axis
face can only be perceived in the mind’s eye. Never- ratio of the  ellipsoid (tangential to the limit state
theless, as shown in Table 2, the probabilities of failure surface) to the one-standard-deviation dispersion ellip-
inferred from reliability indices are again in close soid. This axis ratio is dimensionless and independent
agreement with Monte Carlo simulations. Computing of orientation, when R and r are co-directional. For
the reliability index and Pf ¼ Ö() by the present instance, a vertical line through the mean-value point
approach takes only a few seconds. In contrast, the (centre of the ellipses in Fig. 2) will intersect the upper
time needed to obtain the probability of failure by part of the ellipses at 9 ¼ 37.100 and 40.2316
Monte Carlo simulation is several orders of magnitude respectively, and the dimensionless ratio (40.2316 
longer, particularly when the probability of failure is 35)/(37.1000  35) is equal to 2.491, identical to the
small and many trials are needed. It is also a simple R/r ratio from the mean-value point to the design point,
matter to investigate sensitivities by re-computing the and in fact identical to the co-directional R/r ratios in
reliability index  (and Pf ) for different mean values any other direction. This axis-ratio interpretation in the
and standard deviations in numerous what-if scenarios; original space of the variables overcomes a drawback in
not so with Monte Carlo simulations. (Note that the Shinozuka’s (1983) standardised variable space that ‘the
probability of failure as used here means the probability interpretation of  as the shortest distance between the
that, in the presence of parametric uncertainties in 9 origin (of the standardised space) and the (transformed)
and , the factor of safety, equation (1), will be < 1.0, limit state surface is no longer valid’ if the random
or, equivalently, the probability that the performance variables are correlated. A further advantage of the
function, equation (4), will be < 0.) original space, apart from its intuitive transparency, is
(i) Monte Carlo simulations provide additional information that it renders feasible and efficient the computational
on the number of simultaneous overturning-and-sliding approach involving non-normals as presented in Low &
failures, from which the probability of simultaneous Tang (2004).
failure and the total failure probability (Columns 4 and (k) The soil–wall interface friction angle  is often
5, Table 1) can be inferred. The value 0.007 15 for empirically given as a fraction  of the internal friction
800 000 trials is likely to be more accurate than the angle 9 of the soil,  ¼ 9, where  may depend on
corresponding values (0.007 11 and 0.006 83) for several factors and cannot be quantified precisely.
500 000 and 200 000 trials. The two-mode total failure Instead of treating 9 and  as two highly correlated
probability of 0.007 15 is within the range indicated by random variables as in Fig. 1, one may instead choose
first-order series bounds for different failure modes Fi to model the uncertainties of  and 9, for example
(Cornell, 1967): with a mean of  equal to 0.5 and a standard deviation
of  equal to 0.1. If this is done,  will be calculated
Y by using the formula  ¼ (9)*, where ‘*’ means the
Max[P(Fi )] < P(failure) < 1  [1  P(Fi )] (10a) values of  and 9 under the x* column in Fig. 1. In
this way the uncertainty of  will be due to those of 
which, for the reliability indices of 2.491 and 3.102 and 9.

Table 2. Comparison of probabilities of failure inferred from reliability indices with probabilities of failure from Monte Carlo
simulations
Case Probability of rotational Probability of rotational
failure inferred from  index failure, from Monte Carlo
¼ Ö() ¼ NormSDist() simulation, 200 000 trials
Semi-gravity retaining wall of Fig. 1  ¼ 2.491 200 000 trials: 0.0060
Ö() ¼ 0.0064 (800 000 trials: 0.0063)
Anchored wall of Figs 6 and 7, L ¼ 12.2 m, and all normal  ¼ 2.50
variates Ö() ¼ 0.0062 0.0065
Anchored wall of Figs 6 and 7, L ¼ 12.2 m, and all symmetric  ¼ 2.535
beta distribution variates with (max  min) ¼ 6 Ö() ¼ 0.0056 0.0042
Anchored wall of Figs 6 and 7, with L ¼ 12.2 m, and all  ¼ 2.592
lognormal variates Ö() ¼ 0.0048 0.0050
RELIABILITY-BASED DESIGN APPLIED TO RETAINING WALLS 69
RELIABILITY-BASED DESIGN OF AN ANCHORED (z2  z1 )/(2L  z1  z3 ). The net pore water pressure
WALL diagram shown in Fig. 6 has been obtained following an
In this section the embedment depth d of an anchored established assumption (e.g. Craig, 1997) that the total head
wall (Fig. 6) will be determined so as to achieve a reliability is dissipated approximately uniformly along the back and
index of 2.5 against rotational failure. Analysis involving front wall surfaces between the water table level behind the
correlated normals and non-normals will be illustrated, and wall and the dredge level: that is, over a flow distance in the
comparisons made with Monte Carlo simulations. This is soil of 2(L  z3 ) + (z3  z1 ).
followed by reliability-based design of the required mean The active pressure distribution on the right side of the
strength of the tie rod so as to achieve a reliability index of wall in Fig. 6 is divided into rectangular and triangular
2.5. blocks for computing resultant forces and lever arms with
The anchored wall in Fig. 6 has a total height equal to L, respect to anchor point A. In the spreadsheet of Fig. 7, the
of which d is the depth of embedment. The bulk unit weight first four cells under the column labelled ‘Forces’ contain
of the soil is ª (kN/m3 ) above the water table and ªsat below equations for the resultants (kN/m) of the active pressure
the water table. A surcharge pressure q acts at the top of the blocks on the right side of the wall, and the fifth cell
retained soil. The tie rod acts horizontally at a depth a contains the equation computing the resultant of the triangu-
below the top of the wall. Relevant soil properties are the lar passive pressure block on the left side of the wall. All
effective angle of friction 9 and the interface friction angle these equations refer, where relevant, to the parametric
 between the soil and the wall. The cohesion of the soil is values under the column labelled ‘x*’, not those under the
assumed to be zero in this example. The water table level ‘mean’ column. The Coulomb active and passive earth
behind the wall is at a depth z1 from the top of the wall; the pressure coefficients have been used, justified by the fact
tidal level in front is at depth z2 , and the dredge level at that the active coefficient is practically the same as the
depth z3 . In the reliability analysis the depths z1 , z2 and z3 Caquot and Kerisel coefficient based on a logarithmic spiral
will be treated as random variables, each with a mean value surface, whereas the Coulomb passive coefficient is reason-
and a standard deviation. The deterministic formulation is ably close to the Caquot and Kerisel passive coefficient
described next as it underlies the limit state function in the when 9 and  are near the x* values shown in Fig. 7. (The
reliability analysis. Coulomb Kp value overestimates passive resistance for high-
er values of 9.)

Deterministic set-up based on free earth support method


The deterministic model is set up in Microsoft Excel Uncertainties and correlations in anchored wall design
based on the x* column values in Fig. 7. Initially these x* For the case in hand, the assumed mean values and
values are set equal to the respective mean values. For standard deviations (StDev) of the nine random variables are
steady seepage conditions, the effective unit weight of the as shown in Fig. 7. The two columns labelled ‘ N ’ and ‘ N ’
soil below the water table is ªea ¼ ª9 + iªw behind the wall contain the formulae ‘¼EqvN(. . ., 1)’ and ‘¼EqvN(. . ., 2)’
and ªep ¼ ª9  iªw in front of the wall, where ª9 is the respectively, which invoke the user-created functions shown
buoyant unit weight and i the hydraulic gradient, equal to in Fig. 8 to perform the equivalent normal transformation

Surcharge q 5 10 kN/m2 No surcharge


above anchorage

a
T da
b
A
z
z2
ã, ö¢,ä

L5? 2
Water table z1
z1 p¢h1
1
z2 6
pw1
ãsat
3
z3 2 z 2 7 8

z3 pw2

ãep 9
d ãea Net water
pressure
5 4

p¢hb

Fig. 6. Notations for reliability-based design of anchored wall with steady-state seepage. Uncertainties in water levels (z1 and z2 ),
dredge level (z3 ) and soil properties will be reflected in the analysis
70 LOW
Mean StDev ìN
x* óN nx ã ãsat q ö¢ ä Tyld z1 z2 z3
normal ã 17 0·4 16·58 17 0·4 21·048 ã 1 0·5 0 0·5 0·4 0 0 0 0

normal ãsat 20 0·5 19·38 20 0·5 21·233 ãsat 0·5 1 0 0·5 0·4 0 0 0 0

normal q 10 2 10·17 10 2 0·0829 q 0 0 1 0 0 0 0 0 0

normal ö¢ 36 3 29·82 36 3 22·059 ö¢ 0·5 0·5 0 1 0·8 0 0 0 0

normal ä 18 1·5 15·32 18 1·5 21·788 ä 0·4 0·4 0 0·8 1 0 0 0 0


normal Tyield 200 15 200 200 15 21E-06 Tyld 0 0 0 0 0 1 0 0 0
normal z1 4·5 0·23 4·508 4·5 0·23 0·0368 z1 0 0 0 0 0 0 1 0·7 0
normal z2 6 0·3 6·068 6 0·3 0·2269 z2 0 0 0 0 0 0 0·7 0·1 0
normal z3 8 0·4 8·541 8 0·4 1·3526 z3 0 0 0 0 0 0 0 0 1

crmatrix (Correlation matrix)


Boxed cells contain equations· The formulae in the boxed
cells below refer to the x* values, not the mean values·

Ka Kah Kp Kph ãw L d i ãea ãep p¢h1 p¢hb pw1 pw2 a â

0·303 0·292 4·991 4·814 9·8 12·20 3·66 0·137 10·9 8·24 24·83 49·4 13·2 9·85 1·50 2·50

Forces Lever arm Moments Forces Lever arm Moments

(kN/m) (m) (kN-m/m) (kN/m) (m) (kN-m/m) 5 sqrt(mmult(transpose(nx),


mmult(minverse(crmat),nx)))
1 236·3 4·600 2167 6 210·3 4·05 241·6
Ctrl+Shift, then Enter
2 249·3 1·506 274·2 7 224·4 5·8 2141 This encapsulates equation (6b)
in a single cell, for constrained
3 2168 6·854 21152 8 24·12 5·39 222·2
optimization by Excel's built-in
4 294·5 8·136 2769 9 218·0 8·26 2149 Solver program·

5 265·4 9·48 2516 PerFn1

5 sum(Moments) 2 3 1026

All normal variates: L 512·20 m, for â 5 2·50, as shown, against rotational failure

All symmetric beta distribution variates: L 5 12·17 m, for â 5 2·50· (If L 512·20 m, â 5 2·53)

All lognormal variates: L 5 12·12 m, for â 5 2·50· (If L 512·20 m, â 5 2·59)

Fig. 7. Reliability-based design of wall height L for target reliability index of 2.50. The required average embedment
is (L 2 mean z3 )

(when variates are non-normal) based on the two equations as explained in Low & Tang (2004).
in Rackwitz & Fiessler (1978). Reliability analysis can be For the case in Fig. 7, the uncertainties in parameters z1 ,
performed with various non-normal distributions merely by z2 and z3 could arise due to uncertain or fluctuating water
entering ‘BetaDist’ or ‘Lognormal’ . . . (instead of ‘normal’) table elevations and likely deviations of actual dredged level
in the first column of Fig. 7, and distribution parameters in from desired dredged level. Since a normal variate lies
the columns to the left of the x* column. For correlated within  three standard deviations from the mean with a
non-normals, the ellipsoid perspective (Fig. 2) and the con- probability of 0.997, one may regard the mean value of 8 m
strained optimisation approach still apply in the original and standard deviation of 0.4 m for the dredge level z3 as
coordinate system, except that the non-normal distributions implying that z3 is almost certain to be within the range
are replaced by an equivalent normal ellipsoid, centred not 6.8 m to 9.2 m, with 8.0 m as the most likely value. Some
at the original mean of the non-normal distributions, but at correlations among parameters are assumed, as shown in the
an equivalent normal mean N : correlation matrix (crmatrix). It is judged logical that the
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi unit weights ª and ªsat should be positively correlated, and
u" # " #
u x  N T that each is also positively correlated to the angle of friction
1 xi   i
N
 ¼ min x2 F t i i
½ R (11) 9. Also, the angle of friction 9 and the wall–soil interface
 iN  Ni friction angle  are deemed to be positively correlated with
RELIABILITY-BASED DESIGN APPLIED TO RETAINING WALLS 71

Fig. 8. User-created Excel function codes for automatic equivalent normal transformation when performing non-normal
reliability analysis (more distribution options available in Low & Tang, 2004)

a correlation coefficient r ¼ 0.8, as in the previous illustra- Given the uncertainties and correlation structure in Fig. 7,
tive semi-gravity wall example. and with the x* column initially given the mean values, it is
desired to find the required total wall height L so as to
achieve a reliability index of 2.5 with respect to rotation
Reliability-based design of total wall height L for a desired failure about the anchor level ‘A’. The solution, obtained
 index value of 2.5 using Excel’s built-in constrained optimisation program Sol-
The analytical formulations based on force and moment ver (to minimise the  cell, by changing the x* column
equilibrium in the deterministic analysis are also required in automatically, subject to the constraint that the cell PerFn1
a reliability analysis, but are expressed as limit state func- be equal to zero) is as shown in Fig. 7. The Solver option
tions or performance functions. ‘Use automatic scaling’ can also be activated. A few L
72 LOW
values were tried, each followed by Solver minimisation to 5·50
obtain the reliability index promptly within seconds. A wall
height of L ¼ 12.2 m, with expected embedment depth ¼ L
5·00
  z3 ¼ 4.2 m, was found to give a reliability index 

Required average embedment


of 2.50. Note that in this case the performance function Normal variates

5 (Required L 2 ìz3)
(¼ sum(Moments)) is a non-linear and lengthy function of
4·50
the nine random variables under the column labelled ‘x*’. It
is interesting to note that, at the point where the nine-
dimensional dispersion ellipsoid touches the limit state sur- 4·00
face, both unit weights ª and ªsat (16.58 and 19.38) are Symmetric beta distribution
lower than their corresponding mean values, contrary to the variates
expectation that higher unit weights will cause higher active 3·50
Lognormal variates
pressure and hence greater instability. This apparent paradox
is resolved if one notes that smaller ªsat will (via smaller
ªep ) reduce passive resistance, that smaller 9 will cause 3·00
greater active pressure and smaller passive pressure, and that 0 0·2 0·4 0·6 0·8 1·0
Standard deviation of z3
ª, ªsat and 9 are positively correlated. Note also that at
failure (x* values being on the limit state) by rotation about Fig. 9. Variation of required average embedment with the
A, active and passive pressures are justifiably fully mobi- standard deviation of the passive soil surface level, for different
lised. assumed distributions of the nine random variables in Fig. 7.
The values in the x* column (following Solver’s search Target â index is 2.50.
for the critical ellipsoid) denote the most probable failure
point: that is, the most likely combination of parametric
values that would cause a particular failure mode to occur. 0.541 m below its average value of 8.0 m. However, this
The reliability index  is the distance from the mean-value difference between the design value and the mean value is a
point (the centre of the nine-dimensional dispersion ellip- function of uncertainty: if the standard deviation of z3 is
soid) to this most probable failure point on the limit state 0.2 m instead of 0.4 m, the reliability index will be 2.81 for
hypersurface (analogous to Fig. 2), in units of directional length L ¼ 12.2 m, and the failure combination of para-
standard deviations. One notes that the x* value of Tyield (the meters shows the design value of the passive soil surface to
yield force of the tie rod) is unchanged at its mean value of be only 0.16 m below its average value of 8.0 m.
200, indicating the insensitivity of the reliability index to The two-parameter normal distribution is symmetrical and,
Tyield when rotational failure about anchor point A is being theoretically, has a range from 1 to +1. For a parameter
considered. (Tyield is important in the next failure mode to be that admits only positive values, the probability of encroach-
considered.) Another parameter that is a lightweight in terms ing into the negative realm is extremely remote if the coef-
of sensitivity is the surcharge q, owing to its relatively small ficient of variation (standard deviation/mean) of the
mean value. However, in a different scenario where the mean parameter is 0.25 or smaller. The lognormal distribution has
value of q is of considerable relative magnitude (and for a often been suggested in lieu of the normal distribution, as it
similar /  ratio), its sensitivity scale could conceivably be excludes negative values and affords some convenience in
more significant. The sensitivity measures of parameters may mathematical derivations. A more versatile distribution is the
not always be obvious from a priori reasoning. A case in four-parameter (a1 , a2 , min, max) beta distribution, which
point is the strut with complex supports analysed in Low & can be symmetrical (if a1 ¼ a2 ) or non-symmetrical (a1 6¼
Tang (2004, p. 85), where the mid-span spring stiffness k3 a2 ). The mean  and standard deviation  of a beta
and the rotational stiffness º1 at the other end both turn out distribution with parameters a1 , a2 , min and max are (e.g.
to have surprisingly negligible sensitivity weights: this sensi- Evans et al., 2000)
tivity conclusion was confirmed by previous elaborate deter- a1
ministic parametric plots. In contrast, reliability analysis  ¼ min þ ðmax  minÞ 3 ,
a1 þ a2
achieved the same conclusion relatively effortlessly.
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Sensitivities of the design variable or reliability index to ðmax  minÞ a1 a2
different mean values, standard deviations, assumed prob-  ¼ (12)
ð a1 þ a2 Þ a1 þ a2 þ 1
ability distributions or correlation structure can be investi-
gated efficiently and promptly. For example, Fig. 9 shows Hence, if a1 ¼ a2 ¼ 4, the mean is at the mid-point between
that, for a target reliability index of 2.50, the required min and max, and the standard deviation is equal to 1/6 of
wall height L and the required average embedment length the range (max  min). The mode of the beta distribution is
(¼ Required L  Average passive soil surface elevation  z3 )
ð a1  1Þ
increase with increasing standard deviation of z3 , and to mode ¼ min þ ðmax  minÞ 3 ;
some extent also vary with the assumed probability distribu- ð a1 þ a2  2Þ
tion of the variates. a1 . 1, a2 . 1 (13)
As noted in Simpson & Driscoll (1998, pp. 81, 158),
clause 8.3.2.1 of Eurocode 7 requires that an ‘overdig’ In Fig. 7, when performing reliability analysis assuming
allowance shall be made for walls that rely on passive beta-distribution variates, the parameters used are (4, 4,
resistance. This is an allowance ‘for the unforeseen activities mean  3*StDev, mean + 3*StDev). For example, the
of nature or humans who have no technical appreciation of surcharge q would be given beta distribution parameters (4,
the stability requirements of the wall’. The reliability analy- 4, 4, 16), which is the symmetric BetaDist curve shown in
sis in Fig. 7 accounts for uncertainty in the ground level (z3 ) Fig. 10. As summarised in the bottom insert of Fig. 7, for a
of the passive soil, requiring only the mean value and the target reliability index of 2.50, the required total wall height
standard deviation of z3 (and its distribution type, if not L is 12.20 m, 12.17 m and 12.12 m when all nine variates
normal) to be specified. For the case in hand, the failure follow the normal distribution, beta distribution and lognor-
combination of parametric values under the x* column mal distribution respectively. Since  z3 ¼ 8.0 m, these
shows the design value of the passive soil surface to be required L values correspond to expected embedment depths
RELIABILITY-BASED DESIGN APPLIED TO RETAINING WALLS 73
0·25
Lognormal (ì 5 10, ó 5 2) Normal (ì 5 10, ó 5 2)

BetaDist (a1, a2, min, max) (4, 4, 4, 16)


0·20
(1·5, 4, 4, 16)
(4, 1·5, 4, 16)

Probability density
0·15

(2·5, 4, 4, 16) (4, 2·5, 4, 16)


0·10

0·05

0
0 5 10 15 20
4 16
q

Fig. 10. Comparing the symmetrical normal distribution (21 < x < 1) with beta
distributions bounded between 4 and 16. The parameters a1 and a2 determine the shape
of the beta distribution

of 4.20 m, 4.17 m and 4.12 m respectively, as shown in Fig. the approximation of using an equivalent normal ellipsoid
9, at z3 ¼ 0.4 m, which is the standard deviation value (equation (11)) in the evaluation of the reliability index
assumed in Fig. 7. The difference in the required wall height when the random variables follow beta distributions.
L reaches a maximum value of 0.17 m at z3 ¼ 0.8 m, when
the required L is 13.21 m assuming lognormal variates, and
13.04 m assuming symmetric BetaDist variates. From a Reliability-based design for the required mean strength of
practical perspective, the difference between 13.21 m and anchor tie rod
13.04 m is small, and hence the required L may be deemed For the case in hand, the yield strength (Tyield, in kN/m) of
not sensitive to the three assumed probability distributions, the tie rod is r 2 yield /(tie rod horizontal spacing), where
for the case in hand. Such insensitivity of the design to the r 2 is the cross-sectional area of the steel rod, and yield is
underlying probability distributions may not always be ex- the yield stress of steel. The performance function for
pected—even when the mean and standard deviations remain yielding of the tie rod is PerFn2 ¼ Tyield + sum(Forces 1 to
unchanged—particularly when the coefficient of variation 9), where Tyield refers to its value under the x* value column
(standard deviation/mean) exceeds about 20%, or when of Fig. 7, and forces 1 to 9 are computed by the formulae in
skewness varies significantly from one distribution to an- the nine cells under the two columns labelled ‘Forces’ in
other. Fig. 7. For L ¼ 12.2 m, with Tyield having a coefficient of
variation (/  ratio) of 7.5%, and in the presence of other
uncertainties as shown in Fig. 7, reliability analysis indicates
Failure probabilities inferred from reliability indices and that a mean tie rod strength of 140 kN/m (with standard
from Monte Carlo simulations deviation ¼ 10.5 kN/m) would be sufficient to secure a
Table 2 compares the probabilities of rotational failure reliability index of 2.5 with respect to tie rod yield failure.
inferred from reliability indices with those from Monte Had a mean tie rod strength of 200 kN/m (with standard
Carlo simulations (with correlation) using the @RISK com- deviation 15 kN/m) been provided, as in Fig. 7, the reliabil-
mercial software, for L ¼ 12.2 m. The same restrictive sense ity index against tie rod yielding would be 3.01. These
of ‘probability of failure’ mentioned earlier (in connection solutions were obtained within a minute.
with Fig. 1) applies here also, in that it means the prob- Other limit states that need to be considered include pull-
ability that the resisting moment will be less than the out failure due to insufficient anchorage, and structural limit
overturning moment, based on the analytical model adopted, states of the wall. These are not pursued here.
and considering only parameter uncertainty. Because the
probability of failure is smaller than 1%, many trials are
needed in the Monte Carlo simulation, as illustrated in Table Distinguishing negative from positive reliability indices
1 earlier. Unlike Table 1, which has only two normal In Fig. 7, if a trial L value of 10 m is used, and the entire
variates, the case in Table 2 with nine correlated beta ‘x*’ column is given values equal to the ‘mean’ column
distribution variates took close to 2 h to complete 200 000 values, the performance function PerFn1 exhibits a value
trials, yielding a probability of failure of 0.0042, compared 631.1, meaning that the mean value point is already inside
with the 0.0056 obtained from a few seconds’ computation the unsafe domain (e.g. to the left of the limit state surface
of the  index and Ö(). (The author sees much value in in Fig. 2, but in the nine-dimensional space). Upon Solver
using Monte Carlo simulation for verification and compari- optimisation with constraint PerFn1 ¼ 0, a  index of 1.88
son, but feels that it would be time-consuming to carry out is obtained, which should be regarded as a negative index,
reliability-based design and sensitivity investigations—such i.e. 1.88, meaning that the unsafe mean value point is at
as Fig. 9—using Monte Carlo simulations.) some distance from the nearest safe point on the limit state
In Table 2, for the anchored wall with BetaDistribution surface that separates the safe and unsafe domains. In other
variates, the larger relative difference between the probabil- words, the computed  index can be regarded as positive
ities 0.0056 and 0.0042, inferred from reliability index and only if the PerFn1 value is positive at the mean value point.
Monte Carlo simulations respectively, may be attributed to For the case in hand, the mean value point (prior to Solver
74 LOW
optimisation) yields a positive PerFn1 for L . 10.2 m. The the established method of equivalent normal transformation,
computed  index increases from about 0 (equivalent to a as in FORM, but without involving the concepts of eigenva-
factor of safety equal to 1.0) when L is 10.2 m to 2.5 when lues, eigenvectors and transformed space, which are used in
L is 12.2 m for the assumed normal variates of Fig. 7. the classical approach. The latter mathematical concepts,
though elegant and useful, tend to inhibit wider application
of the classical method of reliability computation.
CORRELATION MATRIX NEEDS TO BE POSITIVE- Reliability-based design was illustrated for an anchored
DEFINITE wall involving nine correlated random variables (Figs 6 and
Although the correlation coefficient between two random 7). The correlation structure of the nine variables was
variables has a range 1 , r ij , 1, one is not totally free defined in a correlation matrix. A normal distribution,
in assigning any values within this range for the correlation symmetrical beta distribution and skewed lognormal distribu-
matrix. The correlation matrix has to be positive-definite. tion were assigned in turn to all nine variables, to investigate
For example, consider the correlation coefficients r ij among the implication of different probability distributions (Figs 7,
three random variables X1 , X2 , and X3 : if r12 ¼ r13 ¼ 0.8, 9 and 10). The procedure is able to incorporate and reflect
the correlation coefficient r23 has to be greater than 0.3 in the uncertainty of the passive soil surface elevation. The
order for R to be positive-definite. It can be appreciated sensitivity of the design variable to correlation structure,
intuitively that, when both the random variables X2 and X3 probability distributions, and standard deviations was also
are highly and positively correlated with X1 (via r12 ¼ r13 discussed.
¼ 0.8), then X2 and X3 should also be positively correlated For the retaining wall examples considered herein, the
with each other to a certain extent (. 0.3). More discussions probabilities of failure inferred from reliability indices were
on the need for a positive-definite correlation matrix are found to be in very good agreement with those from Monte
given in Ditlevsen (1981) and Low & Tang (2004, pp. 81– Carlo simulations (Tables 1 and 2). It is emphasised that
82), with a pragmatic checking procedure in the latter. sensitivity analysis and what-if scenarios of reliability-based
design by the present approach can be accomplished effi-
ciently and promptly. In contrast, to do so by Monte Carlo
COUPLING RESPONSE SURFACE METHOD WITH simulations would be time-consuming.
CONSTRAINED OPTIMISATION The differences between reliability-based design and de-
Although programs can be written in spreadsheet form to sign by Eurocode 7 were briefly discussed. The merits of
handle implicit and complicated limit state functions (e.g. reliability-based design are thought to lie in its ability to
Low et al., 1998, 2001; Low, 2003; Low & Tang, 2004, explicitly reflect correlation structure, standard deviations,
p. 87), there are many situations in which serviceability probability distributions and sensitivities, and to automati-
limit states can only be evaluated using stand-alone finite cally seek the most probable failure combination of para-
element or finite difference programs. In these circum- metric values case by case without relying on fixed partial
stances, reliability analysis and reliability-based design by factors. Corresponding to each desired value of reliability
the present spreadsheet-based constrained optimisation ap- index there is also a reasonably accurate simple estimate of
proach can still be performed, provided one first obtains a the probability of failure.
response surface function via the established response sur- The spreadsheet-based reliability approach can be coupled
face methodology, somewhat similar to curve-fitting via with stand-alone finite element or other numerical packages,
deterministic parametric studies. Such response surface func- via the established response surface methodology.
tions are closed-form expressions, usually polynomials, The partial factors of Eurocode 7 are the outcome of
which closely approximate the outcome of the stand-alone many years of deliberation by multinational committees. The
finite element or finite difference programs. Once the closed- reliability methodology presented in this paper is thought to
form response functions have been obtained, performing be in line with the spirit of Eurocode 7, in that both
reliability-based design for a target reliability index is endeavour to enhance the rationality of design procedures,
straightforward and fast. Examples of coupled response sur- which are not meant to be static and final. The method in
face method and spreadsheet-based reliability analysis were this paper can play a supplementary verification and com-
illustrated in Li (2000) and Vipman et al. (2000). parison role to a design based on Eurocode 7, particularly
when the designer or code developer would like to explicitly
incorporate his or her estimates of parametric uncertainty
SUMMARY AND CONCLUSIONS and correlation into the design so as to achieve a target
There are conceptual and computational barriers in probability of non-failure.
reliability analysis by the Hasofer–Lind method and the The Excel files for the reliability analysis shown in Fig. 1
first-order reliability method (FORM). This is because the and for the reliability-based design shown in Figures 6 and
classical approaches of these methods (well-documented in 7 are available for download at Low (2005).
Ditlevsen, 1981; Ang & Tang, 1984; Madsen et al., 1986;
Haldar & Mahadevan, 1999; USACE, 1999; Baecher &
Christian, 2003) require frame-of-reference rotation and co- NOTATION
ordinate transformation, and iterative numerical derivatives C covariance matrix
using less ubiquitous special-purpose programs. This paper ca adhesion between wall base and soil
provides graphical extensions (Figs 2–5) of the intuitive d embedment depth of anchored wall
expanding dispersion ellipsoid perspective of Low & Tang Fs factor of safety
(1997a, 2004), in the context of a two-random-variable i hydraulic gradient
gravity retaining wall. The ellipsoidal perspective in the Ka , Kp active and passive pressure coefficients
Kah , Kph horizontal component of Ka and Kp
original coordinate space of the variables leads to an effi-
L total height of anchored wall
cient automatic constrained optimisation approach of relia- Pa active earth thrust
bility-based design in a spreadsheet. The approach obtains Pah, Pav horizontal and vertical components of active earth
the same result as the Hasofer–Lind method and FORM, but thrust
is operationally more direct and transparent, and the plat- pw1, pw2 net water pressures acting on anchored wall
form is ubiquitous. For non-normals, the approach adopts q surcharge pressure on retained fill
RELIABILITY-BASED DESIGN APPLIED TO RETAINING WALLS 75
R correlation matrix Hasofer, A. M. & Lind, N. C. (1974). Exact and invariant second-
Tyield yield strength of tie rod moment code format. J. Engng Mech. 100, 111–121.
xi , i , i a random variable, its mean value, and its standard Li, G. J. (2000). Soft clay consolidation under reclamation fill and
deviation reliability analysis. PhD thesis. Nanyang Technological Univer-
x, ,  vectors of random variables, mean values, and standard sity, School of Civil & Environmental Engineering, Singapore.
deviations Low, B. K. (2003). Practical probabilistic slope stability analysis.
 Hasofer-Lind reliability index Proceedings of Soil and Rock America, Vol. 2, pp. 2777–2784.
ª, ªsat unit weight and saturated unit weight of soil Essen: Verlag Glückauf GmbH.
ªw , ªwall Unit weight of water, and unit weight of wall material Low, B. K. (2005). Excel files for ‘Reliability-based design applied
ªea , ªep effective unit weights (in active and passive zones) to retaining walls’, http://www.ntu.edu.sg/home/cbklow/
accounting for seepage pressure Low, B. K. & Phoon, K. K. (2002). Practical first-order reliability
 interface friction angle (between wall and retained soil) computations using spreadsheet. Proceedings of the Conference
9 effective angle of shearing resistance on Probabilistics in Geotechnics: Technical And Economic Risk
 mean value Estimation, pp. 39–46. Essen: Verlag Glückauf GmbH.
N equivalent normal mean of nonnormal distribution Low, B. K. & Tang, W. H. (1997a). Efficient reliability evaluation
r correlation coefficient using spreadsheet. J. Engng Mech., ASCE 123, No. 7, 749–752.
 standard deviation Low, B. K. & Tang, W. H. (1997b). Reliability analysis of
N equivalent normal standard deviation of nonnormal reinforced embankments on soft ground. Can. Geotech. J. 34,
distribution No. 5, 672–685.
Low, B. K. & Tang, W. H. (2004). Reliability analysis using object-
oriented constrained optimization. Structural Safety 26, No. 1,
REFERENCES 69–89.
Ang, H. S. & Tang, W. H. (1984). Probability concepts in engineer- Low, B. K., Gilbert, R. B. & Wright, S. G. (1998). Slope reliability
ing planning and design, vol. 2: Decision, risk, and reliability. analysis using generalized method of slices. J. Geotech. Geoen-
New York: John Wiley. viron. Engng, ASCE 124, No. 4, 350–362.
Baecher, G. B. & Christian, J. T. (2003). Reliability and statistics in Low, B. K., Teh, C. I. & Tang, W. H. (2001). Stochastic nonlinear
geotechnical engineering. Chichester: John Wiley. p-y analysis of laterally loaded piles. Proc. 8th Int. Conf. on
BSI (1994). Code of practice for earth retaining structures, BS Structural Safety and Reliability, Newport Beach.
8002. Milton Keynes: British Standards Institution. Madsen, H. O., Krenk, S. & Lind, N. C. (1986). Methods of
CEN (European Committee for Standardisation) (1994). Eurocode structural safety. Englewood Cliffs, NJ: Prentice Hall.
7: Geotechnical design. Part 1: General rules, ENV 1997-1. Rackwitz, R. & Fiessler, B. (1978). Structural reliability under
Brussels: CEN. combined random load sequences. Comput. Struct. 9, 484–94.
Cornell, C. A. (1967). Bounds on the reliability of structural Shinozuka, M. (1983). Basic analysis of structural safety. J. Struct.
systems. J. Struct. Div., ASCE 93, No. 1, 171–200. Engng, ASCE 109, No. 3, 721–740.
Craig, R. F. (1997). Soil Mechanics, 6th edn. London: Chapman & Simpson, B. & Driscoll, R. (1998). Eurocode 7, a commentary.
Hall. London: ARUP/BRE, Construction Research Communications
Ditlevsen, O. (1981). Uncertainty modeling: With applications to Ltd.
multidimensional civil engineering systems. New York: McGraw- USACE (1999). Risk-based analysis in geotechnical engineering for
Hill. support of planning studies, ETL 1110-2-556, Appendix A, pp.
Evans, M., Hastings, N. & Peacock, B. (2000). Statistical distribu- A11–A12. US Army Corps of Engineers (http://www.usace.ar-
tions, 3rd edn. New York: Wiley. my.mil/publications/eng-tech-ltrs/etl1110–2–556/toc.html).
Haldar, A. & Mahadevan, S. (1999). Probability, reliability and Vipman, T., Teh, C. I. & Low, B. K. (2000). Reliability analysis of
statistical methods in engineering design. New York: John laterally loaded piles using response surface methods. Structural
Wiley. Safety 22, No. 4, 335–355.

You might also like