You are on page 1of 15

Computers and Geotechnics 37 (2010) 359373

Contents lists available at ScienceDirect

Computers and Geotechnics


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

Probabilistic analysis of retaining walls


Ioannis E. Zevgolis a,*, Philippe L. Bourdeau b
a
Edafomichaniki S.A., Athens, Greece
b
School of Civil Engineering, Purdue University, USA

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

Article history: A methodology for the probabilistic analysis of reinforced concrete cantilever walls is developed and
Received 12 October 2008 described in this paper. The walls external stability under static conditions is addressed and modeled
Received in revised form 7 November 2009 as a series system with correlated failure modes. Computations of reliability are performed using Monte
Accepted 8 December 2009
Carlo simulations for assumed probability distributions of the backll and foundation material engineer-
Available online 11 February 2010
ing properties. A case example is analyzed based on the described methodology. The results indicated
that risk, measured by the system probability of failure, is not a linear function of safety ratios. All three
Keywords:
safety ratios in question were positively correlated, with bearing capacity being subjected to higher
Retaining structures
Reliability analysis
degree of uncertainty. The degree of correlation was found to have an important effect on the system
Stochastic modeling probability of failure. Considering the width of the base as varying design parameter, the study also
System reliability showed that rst-order reliability bounds, which are often applied in practice, may lead to a noteworthy
Monte Carlo over- or under-estimation of the design.
Optimization 2009 Elsevier Ltd. All rights reserved.

1. Introduction Besides the questions in principles, for example the accuracy of


the earth pressure theories, the above approach presents two
Reinforced concrete cantilever walls are among the most com- inconveniences. The rst one is that the uncertainty on the design
mon earth-retaining structures. Their design must satisfy two ma- loads and the soil resistance is not considered explicitly, but
jor requirements: internal stability, which is ensured by sufcient implicitly by using a FS based on experience. This approach is
resistance against bending moments and shear forces, and external known as allowable or working stress design method (ASD or
stability, which means that, except for small movements necessary WSD, respectively). For years, there has been a lot of effort to de-
to mobilize the earth pressures, the wall must be in equilibrium velop new reliability-based design methods that will be founded
with respect to external forces [51]. Traditionally, conventional de- on rigorous analysis and statistical treatment of all included uncer-
sign of external stability has been based on deterministic methods tainties. The greatest challenge in this effort is to identify and
and on the concept of Factors of Safety (FS). Stability of the wall is quantify the uncertainty on the soil resistance, i.e., the uncertainty
examined as a system consisting of different modes of failure. Typ- of design soil parameters (see for example [42,70,64,35,37,52]).
ically, these modes are due to overturning of the wall about its toe, Nowadays there is a trend towards design methods that explicitly
due to sliding along its base, and due to bearing capacity inade- take the geotechnical variability into account. These methods have
quacy of the foundation soil [62,15]. Analysis consists of examining become known as Load and Resistance Factor Design (LRFD) in the
separately every mode and modifying the design until the respec- USA, a term borrowed from structural design codes, and Limit
tive FS is larger than or equal to a predened value. In addition, the States Design (LSD) in Canada and Europe (see for example
eccentricity of the loads resultant with respect to the centerline of [17,49,16,53,26]). The second inconvenience of the conventional
the base shall intersect the base within its middle third. When the approach is that the analysis is not based on an overall estimate
wall is constructed on slopped ground or in presence of complex of the walls safety, but only on partial estimates of the safety rel-
stratication, deep-seated stability shall be addressed as well. Fi- ative to the individual modes of failure [4,14]. So, the analysis does
nally, an appropriate design must ensure that the anticipated total not provide an overall integrated measure of safety, but several
and differential settlements will not be excessive. measures (as many as the failure modes under consideration).
Probabilistic analysis offers the framework to encounter the
above shortcomings. In this type of analysis, sources of uncertainty
* Corresponding author. Address: 6 Gavriilidou Street, Athens 11141, Greece. Tel.:
+30 210 2236467, mobile: +30 6944 926540; fax: +30 210 3303625.
are characterized and explicitly accounted for in the computation
E-mail addresses: i.zevgolis@alumni.purdue.edu (I.E. Zevgolis), bourdeau@pur of the reliability. Each mode of failure can be analyzed separately
due.edu (P.L. Bourdeau). and then corresponding modal reliabilities can be computed. But

0266-352X/$ - see front matter 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compgeo.2009.12.003
360 I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373

Nomenclature

ASD allowable stress design qb,L ultimate bearing capacity pressure


B footing width qmax maximum soil pressure
c1 cohesive intercept of backll material qmin minimum soil pressure
C1 unit cost of consequences of a failure R probability of success (e.g., reliability)
c2 cohesive intercept of foundation material S shear resistance along the interface between base and
c2 cohesive intercept along the interface of the wallsoil foundation soil
system SR safety ratio
CFS Central Factor of Safety V unit volume of the concrete
C0 unit price of construction WCONC weight of the concrete components of the wall (footing
Cov covariance and stem)
COV coefcient of variation WSOIL weight of the backll and the soil above the toe of the
CT total expected cost of the wall wall
D embedment depth WSD working stress design
e eccentricity of the loads resultant with respect to the c1 unit weight of backll
centerline of the base c2 unit weight of foundation material
E event el mean standard error
FS safety factor(s) ka/2 lower critical value of a standard normal PDF
H0 total height of the wall (base to top) k(1a/2) upper critical value of a standard normal PDF
KA coefcient of active earth pressure l mean value
KP coefcient of passive earth pressure hli1a condence interval
LH length of heel (design variable) q coefcient of linear correlation
LRFD Load and Resistance Factor Design r standard deviation
LSD Limit States Design RFR summation of resisting forces
N total number of trials in Monte Carlo simulations RFSL summation of sliding forces
nFi number of times that the safety ratio of the ith mode is RMOT summation of overturning moments of the forces
less than one RMR summation of resisting moments of the forces
PA active (lateral) earth pressure RV sum of the vertical forces acting on the wall
P[. . .] probability of occurrence of the event within the brack- u1 friction angle of backll material
ets u2 friction angle of foundation material
PF probability of failure u2 interface friction angle along the walls base and the soil
PP passive (lateral) earth pressure

more importantly, and provided some modeling simplications are fect of correlations on the system reliability decreases as the
accepted, an overall measure of reliability of the retaining wall can width of the footing increases. Using a simplied cost function, they
be assessed. The key advantage of the system modeling approach is also indicated that the correlations may have an impact on the
that it provides a single index for quantifying the structures over- geometry of the wall, and consequently on the total expected cost
all reliability, instead of partial and unrelated modal reliabilities. of construction. Unless the reliability bounds are close to each
This greatly facilitates the use of reliability as a criterion of design other, none of the above simplied solutions is very satisfactory.
optimization and decision support. This is so, because in the case of non-correlation of the failure
For gravity or semi-gravity earth-retaining structures, the model modes, the reliability is underestimated, while in the case of perfect
in question is a series system where failure of at least one compo- correlation of the failure modes the reliability is overestimated. The
nent results in the systems ruin. Among important characteristics degree of under- or over-estimation is not always clear, especially
of the series system in the case of a retaining wall is that its compo- in the absence of a dominant failure mode. A pioneering work pre-
nents (i.e., the resistances to failure modes) are not independent; senting a methodology of system reliability analysis that avoids the
they all depend to some extent on the same design parameters or above shortcomings is the work of Biernatowski and Pua [10].
material and soil properties. However, a commonly accepted sim- Their study provides a probabilistic procedure for analyzing the sta-
plication consists in ignoring this dependency [31,4,7,60]. Another bility of massive bridge abutments, based on simulation tech-
simplication consists in computing system reliability bounds cor- niques. Recently, Low [40] presented an approach for calculating
responding to complete independence and/or perfect correlation. the system reliability of a bi-modal system (overturning and slid-
For example, Ang and Tang [2] and Tang [61] evaluated the limit ing) based on the method of Low et al. [41].
values of the overall reliability of a reinforced concrete cantilever One of the objectives of this paper is to revisit the issue of
wall for two extreme hypotheses of correlation between the failure dependency between failure modes that had been observed by
modes: non-correlated and perfectly correlated. For a bi-modal sys- Biernatowski and Pua [10]. A methodology is presented for the
tem (overturning and sliding) they showed that the domain of var- reliability analysis of retaining walls that takes into account the
iation of the reliability in function of the degree of modal dependency between different failure modes. This dependency is
correlation is quite narrow when there is a dominant mode of fail- investigated and quantied through their correlation, and its inu-
ure. Same conclusion was drawn by Blazquez and Der Kiureghian ence on the system reliability is addressed. Computations are per-
[11] for three modes of failure of a gravity wall subjected to seismic formed using a Monte Carlo simulation algorithm for assumed
loading. Bourdeau and Gutierrez [14], also considering the two lim- probability distributions of the random variables. The impact of
iting cases of perfect correlation and non-correlation, studied the failure mode correlation on the system reliability and, in the light
ranges of wall width that inuence these correlations. In the ab- of these data, the relevance of common simplications is dis-
sence of a dominant failure mode, their study showed that the ef- cussed. Finally, considering the width of the base as varying design
I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373 361

Fig. 1. Geometry of the analyzed wall.

parameter, the study investigates the signicance of the modal cor- not necessary. Moreover, settlement analysis is not included in
relations in terms of optimal design of the retaining wall. The the model in its current formulation.
methodology presented in this study is envisioned as a supplemen- An overall reliability analysis of a retaining wall shall include
tary tool to recently developed LRFD and Eurocode. analysis of the structural failure modes. However, these are not ad-
dressed in the present study, because the authors concentrate on
geotechnical sources of uncertainty and consider, as a reasonable
2. System reliability simplication, that the internal probability of failure would be
much smaller than the external modes. This is so because the resis-
2.1. Stability considerations tance of the reinforced concrete has lesser variability than soil
shear strength. In other words, in geotechnical modes, both load
The type of retaining wall being considered herein for reliability and resistance are affected by soil parameters randomness,
analysis is shown in Fig. 1. It is representative of backlled rein- whereas in structural reliability only the load is affected by soil
forced concrete walls frequently used in practice. Only the walls variability. So, the approximation that is made here is that the
external stability under static conditions is considered through structural mode reliability is close to one (relatively to geotechni-
analysis of overturning around the toe, base sliding, foundation cal modes), and this component is neutral in the system reliability.
bearing capacity, and location of the resultant of applied loads Of course, a comprehensive model should include it specically,
(eccentricity). These modes of failure or instability are illustrated but the goal here is to investigate the geotechnical design process.
in Fig. 2. Note that in cases of special structures, such as retaining Besides, a complete analysis of the structural failure modes would
walls that serve as bridge abutments, walls constructed on slopped have to include at least control against bending moment and shear
ground and/or walls constructed in the presence of complex strat- at stem, at toe, and at heel (i.e., six more safety factors). In addition,
ication, deep-seated stability analysis (also called rotational or in order for the study to be accurate, elements of reinforced con-
overall stability) would be necessary and should be included in crete design would have to be included. Such an analysis falls be-
the model. However, in the present work, assuming the wall is yond the scope of the present study, whose focus is the
built on leveled ground without a complex stratication, this is geotechnical design process.
362 I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373

in which KA is the coefcient of active earth pressure, c1 the


unit weight of the backll, H0 the total height of the wall
(base to top), and c1 is the cohesive intercept of the backlled
material. For horizontal backll and vertical wall stem, KA is
given by:
1  sin u1
KA 2
1 sin u1
in which u1 is the friction angle of the backlled material.

The lateral earth pressure acting on the front of the wallsoil


unit. This is usually taken as the passive earth pressure PP
that, assuming zero friction at the interface, is given by:

1 p
PP K P c2 D2 2c2 K P D 3
2
in which KP is the coefcient of passive earth pressure, c2 the
unit weight of the foundation material, D the embedment
depth, and c2 is the cohesive intercept of the foundation
material. In this study the material below, in front, and above
the toe of the wall is assumed to be the same. The coefcient
of passive earth pressure KP is given by:
1 sin u2
KP 4
1  sin u2
in which u2 is the friction angle of the foundation material.

The soil pressure q that acts vertically on the base of the wall.
Its maximum and minimum values are given by:
P  
V 6e
qmax;min 1 5
B B
P
in which V is the summation of the vertical forces acting on
the wall, B the width of the base of the wall, and e is the
eccentricity of the loads resultant with respect to the center-
line of the base given by:
P P
B MR  M OT
e  P 6
2 V
P P
where MR and MOT are the summations of overturning
resisting and driving moments, with respect to the toe of
the base. Typically, the resultant is required to intersect the
base of the wall within the middle third; hence the entire
area beneath the base is theoretically subjected to compres-
sion [51]. Numerically this means that the eccentricity must
be smaller than or equal to one sixth of the base length. If the
resultant falls on the right side of the base centerline, then
based on Eq. (6) the eccentricity becomes negative. So the
Fig. 2. Modes of instability: (a) overturning, (b) sliding, (c) bearing capacity, and (d)
condition that must be met is:
excessive eccentricity.
B
jej 6 7
6
According to conventional geotechnical design procedures for The shear resistance S acting along the interface between the
this type of wall, the mass of backll overlaying the stem is as- base of the wall and the foundation soil. Typically, this is
sumed to form a resisting block attached to the structure [51]. Ac- given by:
tive earth thrust on the ctive vertical interface between this
X
resisting mass and the retained backll is computed according to S Bc2 V  tan u2 8
Rankines theory. For dry backll conditions, and in addition to
the self-weight of the wall (WCONC) and the weight of the soil above in which c2 is the cohesive intercept along the interface, and
the base (WSOIL), the following pressures are applied on the wall: u2 is the interface friction angle along the wall and the soil at
the base.
The lateral (active) earth pressure PA acting on the back of the In principle, the wall is safe when the loads that tend to activate
wallsoil unit: a mechanism of instability, are smaller than or equal to the loads
that tend to resist to this mechanism (capacitydemand model).
1 p In a traditional deterministic analysis following ASD, a safety factor
PA K A c1 H02  2c1 K A H0 1
2 would be computed for each of the postulated modes of failure
I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373 363

using nominal (characteristic) values of the controlling parameters.


Each of the resulting safety factors would then be required to be
greater than a recommended empirical value, specic to the partic-
ular failure mode. Required safety factors are typically in the order
of 1.52 for overturning, 1.252 for base sliding, and 23 for bear-
ing capacity [62,15]. The fact that these required values are larger
than one is an acknowledgement of important uncertainty being
present in the design process, though this is not explicitly ana-
lyzed. Additionally, the eccentricity is required to be less than B/
6, according to Eq. (7). In fact, some references recommend the
eccentricity criterion as the criterion against overturning [63]. This
approach was considered more appropriate by the authors and it is
followed in the present study. Further reasoning for this decision is
provided in the Appendix A.
It is acknowledged that design analyses are simplications of
reality and these may produce biased prediction of actual perfor-
mance. Model uncertainty, in the context of reliability analysis, is
still a domain largely unexplored. Its assessment would require
statistical databases of prediction vs. performance be assembled
and analyzed. In absence of such assessment, it is important to rec-
ognize that computed reliability or probability of failure have only
a relative meaning. This can be used in comparative analyses, such
as optimization studies, provided these comparisons are based on
the same models.

2.2. Modal probabilities of failure

In contrast to a deterministic approach, in a probabilistic analy-


sis of the wall stability, sources of uncertainty are characterized and
explicitly accounted for in the computation of the reliability. In or-
der to represent limit states of equilibrium, it is convenient to de-
ne performance functions by analogy with safety factors, as
safety ratios (SR). Safety ratios with respect to sliding and bearing
capacity inadequacy, SRSL and SRBC, respectively, are expressed by:
P
FR PP S
SRSL P 9 Fig. 3. Schematic representation of a system in series.
FSL PA
qb;L
SRBC 10
qmax 2.3. Assessment of system reliability
P P
where FR and FSL are the summations of the base sliding resist-
ing and driving forces, and qb,L and qmax are the foundation ultimate The global stability of the wall is modeled as a system in series
bearing pressure and maximal applied pressure. Note that the with zero redundancy. This means that if at least one of the postu-
above mathematical expressions are similar to those of safety fac- lated modes of instability occurs, then the system fails. On a prob-
tors but, in contrast with the latter, the safety ratios are functions abilistic context, the global failure is the event in which any of
of soil and material parameters affected by uncertainty. As these [SRe < 1], [SRSL < 1], [SRBC < 1] occurs, and its probability of occur-
parameters are modeled as random variables, SRSL and SRBC are rence, PF, is given by the union of these events:
functions of random variables and therefore, are random quantities,  
too. In this study a performance function, SRe, is also dened for the PF P SRe < 1 [ SRSL < 1 [ SRBC < 1 14
eccentricity. Based on Eq. (7), this can be dened as:
where P[] denotes the probability of the event indicated within the
B=6
SRe 11 brackets. The global stability is the event in which [SRe P 1],
jej [SRSL P 1], and [SRBC P 1] occur simultaneously. The probability
For all three postulated mechanisms of instability, failure is dened of occurrence of this event is the reliability of the event given by:
as the case where the corresponding SR is less than one. It shall be  
emphasized that the term failure herein does not necessarily imply R P SRe P 1 \ SRSL P 1 \ SRBC P 1 15
a collapse of the wall or a catastrophic failure, but it refers to any
unacceptable difference between expected and observed perfor- An illustration of the series system concept is given in Fig. 3. The
mance [39]. Another option, instead of using safety ratios, would global stability of the wall is shown at the end of the chain whose
have been to use safety margins (expressed as the difference be- links (components) are the three modes of instability. If any of
tween resistance and load). The probability PFi that [SRi] < 1 for the links breaks, then the global stability is lost.
any given mechanism i is given by the following expression: In general, computation of PF or R by exact integration of the in-
volved multivariate functions is impractical. This difculty can be
PFi PSRi < 1 12 overcome by using approximate methods that have been devel-
while the reliability is given by: oped specically for system reliability analysis. Among these are
the Hasofer and Lind [30] procedure or the system reliability
Ri 1  P Fi PSRi P 1 13 bound methods [1,2]. Other methods include the rst- and
364 I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373

second-order reliability methods, either on their classical format or


using concepts of asymptotic analysis [32].
The simplest in the present case would consist in determining
the rst-order reliability bounds of the system. If the safety ratios
are all positively correlated, then the system probability of failure
and the system reliability can be, respectively, written as:

Y
k
maxfPFi g 6 P F 6 1  1  PFi 16
i1
Fig. 4. First-order reliability bounds for positive correlations.
Y
k
Ri 6 R 6 minfRi g 17
i1
chastic technique that generates a great number of repeated simu-
where PFi and Ri are the modal probabilities of failure and modal lation processes (realizations). Each simulation is based on the
reliabilities, respectively, relative to each of the three individual generation of a series of values of one or more random variables.
failure modes and k = 3 (since three failure modes are examined The procedure requires complete denition of the random variable
in the present study). If the safety ratios are negatively correlated, probability distributions, but, through simple computations, it pro-
then: vides empirical outcomes of numerically simulated random real-
izations of the safety ratios. Then, statistical estimates can be
Y
k
obtained for the system reliability, safety ratios, their coefcients
06R6 Ri 18
i1
of correlations, or other quantities of interest. The Monte Carlo
Q Q technique is extensively documented in geotechnical engineering
The terms 1  ki1 1  PFi and ki1 Ri are the exact system proba- literature. Among other applications, it has been used for probabi-
bility of failure and system reliability in the case of modal indepen- listic analysis on earth pressure problems, retaining structures,
dence. The terms max {PFi} and min{Ri} correspond to the exact deep foundations, slope stability problems, and other (see for
system probability of failure and system reliability when the safety example [59,4,10,68,27,43]). The recent evolution of personal com-
ratios are perfectly and positively correlated. Because, as shown in puters has lead to more accurate and sophisticated use of the tech-
Fig. 4, the interval dened by Eq. (18) is usually very large, the nique than three decades ago, and often the method is combined
rst-order reliability bounds solution is useful mainly in the case with nite elements analyses [23,25]. It is noted here that due to
of positive correlations between safety ratios (i.e., safety ratios are the complexity of the studied problem and the correlations be-
either simultaneously increasing or decreasing functions of the tween the failure modes, analytical solution through rst or sec-
engineering properties of the soils). This is usually, but not always, ond-order reliability methods (both based on the Taylor Series
the case in practice. For instance, increases in the friction angle val- expansion) may be cumbersome and difcult to achieve. However,
ues have generally positive impact on Eqs. (9)(11), while increases rst or second-order reliability methods improved by concepts of
in the unit weights may, in some cases, have negative impact. So, asymptotic analysis, such as the ones presented by Hohenbichler
due to the complexity of the above equations, a safe conclusion et al. [32], could be alternatively used in order to solve the problem
regarding their behavior upon a simultaneous variation of these without signicant difculties. On the other hand, the Point Esti-
variables cannot be drawn. Therefore being free from the constrain- mate Method would be adequate for computing the reliability, pro-
ing assumption of positive correlations is an advantage [72]. The vided that assumptions are made on the distributions of the safety
rst-order bounds described above may be improved by taking into ratios, but it would not provide a direct assessment of their corre-
account the correlation between pairs of the potential modes of fail- lation. In the case of Monte Carlo simulations, and assuming that
ure. The resulting improved bounds necessarily require the proba- the modes of instability are adequately represented by Eqs. (9)
bilities of joint events and thus may be called bi-modal or second (11), the dependency between these modes can be evaluated by
order bounds [2]. For k failure modes (in this study k = 3), the sec- means of coefcients of linear correlation (q). For any pair of ij
ond order lower and upper bound failure probabilities are given modes, q is given by [58]:
by the following correlation [36,33]:  
" ( )# Cov SRi ; SRj
X
k X
i1 qSRi SRj 20
PF1 max 0; PFi  PEi Ej 6 PF
rSRi rSRj
i2 j1
in which rSRi ; rSRj are the standard deviations of the variables SRi,
X
k X
k
SRj, and Cov is their covariance. The covariance is dened as the ex-
6 P Fi  maxPEi Ej  19
j<i pected value E of the products of SRi  SRi and SRj  SRj :
i1 i2
h  i
where Ei and Ej are the individual failure events (with E1 being the CovSRi ; SRj  E SRi  SRi SRj  SRj 21
largest set and PF1 being its probability of occurence). A similar ver-
sion of the above second order bounds has been proposed by Ditlev- where SRi ESRi  and SRj ESRj .
sen [22]. The bi-modal bounds depend on the ordering of the
individual failure modes. This means that different orderings of 2.4. Design optimization
the individual failure modes may yield different values in the above
equation, so the bounds corresponding to different orderings may Reliability analysis offers the framework for optimization of the
have to be evaluated to determine the narrowest bounds [2]. In design process. In simple risk-decision problems, an objective
addition, in cases of correlated failure modes (such as in the exam- function, often expressed in monetary units, is dened in terms
ined study), the calculations of the joint probabilities P(EiEj) remain of one or more decision variables. The optimal design is deter-
difcult. mined by the values of decision variables that maximize (in case
The approach used in the present study is a still approximate, of benet) or minimize (in case of cost) the objective function
but more direct computation of the system reliability using a [2]. Previous literature includes optimization analyses for retaining
Monte Carlo simulation algorithm. Monte Carlo simulation is a sto- walls based on the rst-order reliability bounds [31,14,7].
I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373 365

Optimization in the current study is based on the actual system Table 1


reliability given by Eq. (15). Qualitatively, the objective function Geometric characteristics of the analyzed wall.

is expressed in terms of cost by the following expression: Characteristic Notation Value (m)

Total Expected Cost Initial Costof Construction Risk 22 Free height H 6


Depth of embedment D 1.5
In engineering context, risk is often dened as the product of prob- Base thickness HT 0.8
ability of occurrence of an event by the consequence of this event. Toe length LT 0.8
Stem thickness LS 0.8
Based on this, the objective function is expressed by the following
Heel length LH 1.53.5
equation [14]:
C T C 0 V PF C 0 V C 1 V 23
in which CT is the total expected cost of the wall, C0 the unit price of
construction, V the unit volume of concrete, PF the probability of soil resistance in limit equilibrium problems, and may result in
failure of the wall, and C1 is the unit cost of consequences of a fail- smaller probabilities of failure, as compared to long autocorrela-
ure. The terms V and PF are the decision variables of Eq. (23). In the tion distances (e.g., [24,25,28,46,5456]). However, even when
second part of the equation, the rst term, C0V, represents the initial large arrays of data are available for a site, empirical autocorrela-
cost of construction of the wall and the second term, PF(C0V + C1V), tion functions are inaccurate and difcult to estimate with con-
represents the cost of failure. For simplicity, the cost of reconstruc- dence [12,13]. It should also be noted that two-dimensional
tion of the wall after a potential failure is assumed to be the same modeling of a geotechnical problem is incompatible with a correct
with the initial cost of construction. Moreover, the term V refers random eld description of the site because, considering a two-
only to the volume of concrete, i.e., it does not include the volume dimensional cross-section as representative of the site implicitly
of excavated material. Finally, the consequences of failures by the assumes all other parallel cross-sections are identical and their
three modes of instability are assumed to be the same and all to- properties perfectly autocorrelated in the transverse direction
gether expressed by the term C1. A more detailed analysis would [6,5]. In the present case example, no such detailed geostatistical
have considered the consequences of each mode separately. How- description of the site is considered. Subgrade soil and backll
ever, this would require modeling deformation (such as foundation properties are assumed to be estimated based on a limited number
settlement and tilting), which in any case is not an option with limit of data or expert opinion. This situation is likely to be the case in
equilibrium models. So, such an analysis falls beyond the scope of practice when rather simple retaining earth systems are to be de-
this study. Besides, the cost of consequences of a failure is difcult signed. These properties are modeled here as an estimated eld
to estimate, because intangible as well as tangible factors must be using random variables that are assumed homogeneous (or per-
considered [31]. By normalizing with C0, Eq. (23) can be written fectly autocorrelated) over the volume of interest. This is similar
as following: to the approach used for instance by Harr [29] or Low [40]. It is
 
acknowledged that this approach likely leads to conservative esti-
CT C1 mates of failure probabilities, but this should not affect the gener-
V 1 PF 1 24
C0 C0 ality of the discussion on system modal correlations and design
So, the normalized total expected cost is expressed as a function of optimization.
the unit volume of the concrete, i.e., as a function of the geometry of
the wall. The ratio C1/C0 can be viewed as a risk factor, assigned on a 2.5.2. Cross-correlation
case-by-case basis. Based on Table 1 and Fig. 1, the volume of inter- Regarding the cross-correlation between the involved random
est is given by: variables, and particularly between friction angle and cohesive
intercept, literature has not been clarifying [23]. Values varying
V HT LT HT LH H0 LS  1 linear meter 25
in both negative and positive range have been quoted by several
Optimal design is the one for which CT/C0 becomes minimum. The authors (see for example [67,18]). Fenton and Grifths [23], inves-
length of heel for which this situation happens is called the optimal tigating a bearing capacity problem, have found that even correla-
length of heel, LH(OPT). tion extremes have only a minor inuence on the stochastic
behavior of the bearing capacity. Zevgolis [71], examining the
2.4.1. Case example external stability of a reinforced soil wall, has shown that positive
In the case example being analyzed, geometric characteristics of cross-correlations overestimate the probability of failure, while
the wall are considered deterministic quantities. These are indi- negative cross-correlations underestimate it. Therefore, in the
cated in Table 1, with reference to notations in Fig. 1. One of these, absence of reliable data that would indicate the true value of
the heel of the wall, LH, is used as variable design parameter. cross-correlations, one should be cautious with using any particu-
lar value. Arbitrary use of positive values as a safety net is too con-
2.5. Random variables servative, taking into account that real values are usually expected
to be negative. So, the model in its current formulation does not
2.5.1. Spatial variability take into account cross-correlations between variables, i.e., the
When extensive site investigations are performed and soil prop- random variables are considered independent to each other. In
erties are measured at a large number of points of a particular pro- addition, the paper assumes statistical independence between
ject site, it may be possible to analyze these data statistically and to the foundation and backll soil properties. This is so because the
describe their spatial variability, within homogeneous sub-regions, backll would be borrowed from the site (thus its properties be
as a continuous random eld [65]. The main elements of such a similar or strongly correlated to those of the foundation soil), only
descriptive random eld are the expected value, variance and auto- if it is adequate with respect to drainage, compaction and frost sen-
covariance function (or the autocorrelation function) of the ana- sitivity, etc. In this study it was assumed this was not the case and
lyzed soil property. A number of examples have been published therefore these two materials are unrelated. Overall, the above
where the random elds were numerically simulated. Sensitivity simplifying assumptions about the correlation structure of the data
analyses have shown, for instance, that short autocorrelation dis- set should slightly affect the system reliability analysis numerical
tances for strength properties tend to reduce the variance of the results but without loss of general validity.
366 I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373

2.5.3. Probability distribution function proach would have been to prescribe the targeted accuracy and
Each one of the random variables is represented by its rst two let the number of simulations be selected in order to achieve this
order moments (mean value l and standard deviation r, or coef- accuracy. For each realization, a different group of the ve soil
cient of variation COV), and its minimum and maximum value. parameters (random variables) is generated following a beta distri-
Therefore, based on the principle of maximum entropy [34], the bution with characteristics listed in Table 2. As a result, each group
random variables are modeled using Beta (i.e., Pearson type I) dis- corresponds to a different outcome regarding the three modes of
tributions, with characteristics listed in Table 2. For the use of beta instability. The probability of failure is then given by:
distribution in civil engineering applications see Oboni and Bour- nFi
deau [47] and Harr [29]. It is noted here that other type of distribu- P Fi 26
N
tions, such as normal, lognormal, or gamma, could have been
another option for modeling the random variables. Nevertheless, in which nFi is the number of times that the safety ratio of the ith
at least one bound of these distributions goes to innity (i.e., the mode is less than one and N is the total number of trials that are per-
upper bound in the case of lognormal and gamma distributions, formed. The generation of the values of the random variables is per-
and both lower and upper bounds in the case of normal distribu- formed using the random number generator (RNG) of the risk
tion), which is not physically consistent with the physical param- analysis software Crystal Ball v.7.2, which operates as an add-into
eters examined in this study. Practically, the unbounded extreme Microsoft Excel [20]. Crystal Ball allows for the denition of design
implies that the variate may take on values many times greater variables as random variables that follow a probability mass or den-
than its expected value. Although theoretically this is possible, sity function (pmf or pdf, respectively). Using the dened random
the probability of such an occurrence is fairly small [29]. variables, the program performs Monte Carlo simulations and pro-
The mean values are selected as representative of typical soils vide outcome in terms of a wide range of elements, such as probabil-
encountered on site (i.e., coarse-grained backll material, more ity distribution functions (PDFs), cumulative distribution functions
ne-grained foundation material). The coefcients of variation, (CDFs), statistical parameters, and other. A great advantage of the
COV, are consistent with representative data reported in literature program is that allows for full extraction of data in Microsoft Excel,
(e.g., [29,37,52]). In the absence of extensive site-specic data base, including both input and output based on which the simulation is
minimal and maximal values are selected based on an arbitrary performed. This gives the user the opportunity not only to perform
rule of 4 standard deviations from the mean. From a statistical his/her own statistical analysis, but also to examine carefully and
point of view, such a range guarantees a high probability of occur- validate the results that are provided by the program. Crystal Ball
rence: 93.75% using Chebyshevs inequality, and 97.22% using has been increasingly used in the recent past for risk analysis in civil
Gauss inequality for symmetrical cases. At the same time, the and geotechnical engineering applications [44,66,9,50]. Based on
physical meaning and the likely range of uctuation is consistent comparative evaluation of six available software programs that offer
with these values. Note that for l  4r, the minimum value of c2 Monte Carlo capabilities, Metzger et al. [45] recommended Crystal
becomes negative, and therefore in this example the minimum va- Ball as the best one, for risk assessment problems that can be
lue was taken equal to zero. The interface properties along the base implemented on a spreadsheet. As far as the (pseudo-)random num-
of the wall, u2 and c2 , are factored down with respect to the ran- ber generator is concerned, Crystal Ball uses a multiplicative linear
dom variables u2 and c2, respectively, by 0.7, which is a value con- congruential generator (LCG). The recursive formula that is being
sistent with common practice [69]. As a result, these properties are used has the following format [20,21]:
also random variables, modeled as linear functions of u2 and c2, fol- Z n1 62; 089; 911  Z n mod231  1 27
lowing a beta distribution. The passive pressure, PP, is neglected in
the calculations of all safety ratios, which is often the case in design The above generator has a full period of length, equal to
[51,62], and all bearing capacity factors, including depth and incli- (232  1)  1 = 2147,483,646, i.e., the cycle of random numbers re-
nation factors, are calculated based on Meyerhofs formulae. peats after more than two billion trials. Several researchers have
tested and evaluated the above recursive formula and have classi-
2.6. Monte Carlo realizations ed it as one of the best available for regular risk assessment appli-
cations [48,57,8,38].
Analyses are performed for a wide range of heels length, from
LH = 1.5 m to LH = 3.5 m, in increments of 0.1 m (21 cases in total). 3. Results and discussion
As a deterministic point of reference, Central Factors of Safety (CFS)
is also computed. The CFS is dened as the Factor of Safety com- 3.1. Reference to deterministic analysis
puted with the mean values of the parameters [29], in contrast
with actual design Factors of Safety that would result from using Fig. 5 is provided as a reference to deterministic analysis, and it
conservative estimates of the parameters expected values. For shows the variation of Central Factor of Safety (CFSi) computed for
the probabilistic simulations thirty thousand realizations (trials) each instability mode in function of heels length (LH). As shown in
are performed for each examined case. This was a reasonable num- the Figure, the rates of increase of the CFS are quite different from
ber for keeping errors in the estimated probabilities within tolera- one mode to another: CFSSL and CFSBC increase linearly with
ble limits. Analyses were also performed for LH > 3.5 m, however increasing LH, while CFSe increases non-linearly. The type of infor-
the mean standard errors in these cases were high. Another ap- mation shown in Fig. 5 presents three inconveniences. First, it does
not provide any output on the probability of occurrence of any
Table 2 instability mode and it does not scale safety. This means that for
Probabilistic parameters (input) of soil properties. instance, a design decision for increasing LH does not reect the in-
Soil property Unit l COV Min Max crease in the reliability of the structure. Second, it does not provide
c1 kN/m3 20 0.07 14.4 25.6
any information related to the dependency of the instability modes
u1 32 0.12 16.6 47.4 with each other, and as a result it does not provide any information
c1 kN/m2 0 0 0 0 on the global stability of the wall. Third, this information in its cur-
c2 kN/m3 18 0.07 13 23 rent format cannot be easily used on an optimization analysis.
u2 25 0.20 5 45
These inconveniences can be eliminated when, in conjunction to
c2 kN/m2 30 0.30 0 66
conventional design, a probabilistic analysis is performed.
I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373 367

Fig. 5. Central Factor of Safety vs. length of heel.


Fig. 7. Typical histogram for CFSBC.

3.2. Modal probabilities of failure

Typical obtained histograms for Central Factor of Safety with re-


spect to sliding, bearing capacity and eccentricity are provided in
Figs. 68, respectively. Table 3 tabulates the basic statistical
parameters obtained from the analysis. Statistical errors associated
with these parameters can be estimated using the mean standard
error (el) and the corresponding condence intervals (hli1a).
These two are expressed as [3]:
r
el p 28
N
 
r r
hli1a l ka=2 p ; l k1a=2 p 29
N N

Fig. 8. Typical histogram for CFSe.

Table 3
Statistical parameters of the resulting safety ratios (for LH = 3 m).

Statistics Notation SRSL SRBC SRe


Realizations n 30,000 30,000 30,000
Mean l 1.74 3.99 1.89
Median lm 1.70 3.35 1.79
Variance Var 0.18 6.17 0.28
Standard deviation r 0.42 2.48 0.53
Coefcient of variation COV 0.24263 0.62325 0.28104
Coefcient of skewness b1 0.56365 2.18 1.24
Coefcient of kurtosis b2 3.47 11.51 5.64
Minimum min 0.61 0.54 0.90
Maximum max 4.12 32.39 6.93
Range width 3.51 31.85 6.03
Fig. 6. Typical histogram for CFSSL.
368 I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373

Table 4
Error estimates (for LH = 3 m).

Statistics Notation SRSL SRBC SRe


3 3
Mean standard error el 2.434  10 14.344  10 3.061  103
Conf. interval 95% hli0.95 4.771  103 28.115  103 5.999  103
Conf. interval 99% hli0.99 6.280  103 37.008  103 7.897  103

in which (1  a) is a specied probability and ka/2 and k(1a/2) are


the lower and upper critical values of a standard normal PDF (see
[3] for more details). The values of the above are provided in
Table 4 for LH = 3 m. Based on the above, it is indicated that bear-
ing capacity computations are subjected to higher degree of uncer-
tainty COVCFSBC 0:6232, compared to sliding and eccentricity
(COVCFSSL 0:2426 and COVCFSe 0:2810, respectively). This is con-
sistent with the conclusions drawn from the study of Biernatowski
and Pua [10], as well. This is so because bearing capacity primar-
ily depends on the cohesion and friction angle of the foundation
soil, which are both expected to demonstrate higher dispersion
compared to the friction angle of the backlled material, and be-
cause the bearing capacity coefcients depends in non-linear
way on friction angle (particularly for higher values). This has an
impact on the corresponding errors and condence intervals, too Fig. 10. Central Factor of Safety vs. modal probabilities of failure (log).
(Table 4). For instance, the sampling error of the mean of CFSBC
is approximately six times as big as the equivalent error of CFSSL
and almost ve times as big as that of CFSe. As far as the symme- Similar conclusions are drawn for sliding and eccentricity. On the
try of the distributions is concerned, all three appear to be skewed other hand, as expected, the same level of safety for the three pos-
with the long tail of the distribution being on the right side of the tulated mechanisms of instability is achieved for different values of
mean (b1 > 0). This is well illustrated in the histograms, especially LH and CFS. Another observation from the above Figures is the sen-
in the case of CFSBC, which shows the higher degree of skewness. sitivity of PFe with varying LH. Specically, the probability drasti-
Moreover, CFSBC shows the higher degree of kurtosis. cally changes within a short range of LH. It is worth mentioning
Figs. 9 and 10 graphically present the modal probabilities of that for CFSe = 1, which is the requirement in conventional design,
failure (PFi) plotted in logarithmic scale, with respect to LH and CFSi, the probability that e will exceed B/6 is approximately 50%. Note
respectively. It is shown that the modal probabilities of failure are that for the example analyzed here, CFSe = 1 corresponded to
not linear functions of LH and CFS. For instance, CFSBC = 2.4 corre- CFSOT = 2.4 (CFS against overturning), which means that even a safe
sponds to P FBC 4:95%, CFSBC = 3.0 to PFBC 1:48% and CFSBC = 4 design with respect to overturning does not necessarily result in
to P FBC 0:29% (Fig. 10). So, CFSBC alone does not provide any safety with respect to the location of the resultant of the forces.
information as to the likelihood of a failure occurrence, nor as to Opposite to eccentricity, PFSL and P FBC demonstrate a smoother var-
the achieved improvement of stability when design is modied. iation with varying length of heel.

Fig. 9. Length of heel vs. modal probabilities of failure (log). Fig. 11. Joint distribution of CFSeCFSSL.
I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373 369

Fig. 12. Joint distribution of CFSBCCFSSL.


Fig. 14. Coefcients of linear correlation vs. length of heel.

3.3. Dependency of modes of instability


vice versa. High degree correlation was also found between CFSSL
and CFSe, while CFSBC and CFSe were correlated on a lower degree.
Figs. 1113 graphically present the joint distributions of CFS ta-
The trend of correlations appears to be gradually decreasing with
ken two by two for the case of LH = 3 m. It is noticed that CFSBC
increasing LH. Overall, it is clear that the three failure modes are
demonstrates some very high values (up to 33). Opposite to this,
correlated, to a higher or lower degree, to each other.
the ranges of CFSe and CFSSL are much lower, with their upper
bounds at around seven and four, respectively. This is so because
the equation providing CFSBC is very sensitive with respect to vary- 3.4. System reliability
ing parameters and can result in high values if random variables
values are close to their upper bound. Results in terms of coef- Fig. 15 graphically presents the system probability of failure
cient of linear correlation with varying LH are shown in Fig. 14. P FSYS as calculated by Eq. (14), with varying length of heel. As ex-
The highest degree of correlation was found between CFSSL and pected, PFSYS falls within the range of the rst-order bounds, which
CFSBC. In terms of Monte Carlo realizations, this meant that for are calculated according to Eq. (16). However, the question is
the analyzed example most of the times where failure by sliding whether or not these bounds are appropriate to use, especially in
occurred, failure by bearing capacity was also taking place, and the absence of a dominant failure mode. In regards to the latter

Fig. 13. Joint distribution of CFSBCCFSe. Fig. 15. System probability of failure (log) vs. length of heel.
370 I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373

Fig. 16. Computed system probability of failure and modal probabilities of failure. Fig. 18. Optimum designs for different C1/C0.

comment, Fig. 16 illustrates PFSYS together with the three modal correlation). For instance, for design value of LH = 3 m, the above
probabilities of failure. It is shown that the system is controlled ratios are 1.407 and 0.911, respectively.
by the eccentricity for LH up to 2.7 m, and by sliding for LH up to
3.5 m. In other words, for different values of LH different modes
of instability control the design and therefore a safe conclusion 3.5. Design optimization
regarding the appropriate of use of the rst-order bounds cannot
be made in advance. For instance, in this case example and for a In this study, optimization analysis was performed for a wide
target P FSYS 0:011%, LH shall be 3.3 and 3.5 m based on the low- range of risk factors C1/C0 and representative results are provided
er and upper bounds, respectively (Fig. 15). Using Eq. (14), LH shall in Fig. 18. As shown in the Figure, LH(OPT) increases as the risk factor
be 3.5 m. So, assuming a modal independence leads to over-design increases. For instance, when the unit cost of consequences of a
of the heel by 0.2 m. On an attempt to quantify the degree of over- failure is equal to the unit cost of initial construction (C1/C0 = 1),
or under-estimation, Fig. 17 illustrates the rst-order bounds prob- the optimum length of heel is 3 m. Another interesting observation
abilities of failure normalized by P FSYS (Eq. (16)). So, axis Y shows is made when comparing optimization curves for rst-order reli-
Q3
1 1P Fi maxfPFi g ability bounds. In this context, Fig. 19 shows the curves obtained
the ratios i1
PF
(complete independence) and PF
(perfect
SYS SYS for a risk factor equal to 1. The lower and upper curves represent

Fig. 17. Ratios of P FSYS with rst order probabilities of failure. Fig. 19. Optimum designs for P FSYS and rst-order bounds.
I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373 371

Table 5 Overall, reliability assessment is not a substitute to the conven-


Optimum heel length. tional methods of design, but it is a complementary to them tool,
C1/C0 LH(OPT) necessary on a risk analysis context. Nowadays, the latter one be-
Perfect correlation P FSYS (Eq. (17)) Complete independence comes more and more an essential part of civil engineering pro-
jects. The present methodology can provide a framework for
1 3 3.1 3.2
2 3.1 3.2 3.3
more complex systems analysis in geotechnical engineering and
3 3.3 3.3 3.4 help optimizing designs in presence of uncertainty.
4 3.3 3.4 >3.5
5 3.4 3.4 >3.5
Appendix A

Overturning rarely governs the design of retaining walls


perfect correlation and complete independence, respectively, be- [62,15]. Some authors even suggest that this type of failure is
tween instability modes. As shown in the Figure, LH(OPT) shows a indirectly accounted for, if the eccentricity is kept smaller than
tendency to increase from perfectly correlated to totally indepen- B/6 and the wall is safely designed with respect to sliding [19].
dent failure modes. Table 5 tabulates the different LH(OPT) for vary- When overturning is addressed separately, through a safety factor
ing risk factors. So, the assumption of complete independence dealing with overturning resisting and driving moments, it is usu-
(which is mostly used in the literature) may overestimate LH(OPT) ally evaluated based on the assumption that rotation takes place
and therefore lead to a conservative design decision. On the other around the bottom point of the walls toe. However, this is not al-
side, assuming perfect correlation between the different modes of ways a kinematically realistic assumption. The location of the ac-
instability modes usually results in slightly lower LH(OPT). tual point of rotation depends on the foundation soil
characteristics. On a compressible material the wall would start
plunging before the overturning mechanism is mobilized, so the
4. Summary and conclusions point of rotation would move towards the right side of the base.
In addition to these qualitative comments, it can be shown that a
In this study, the stability of reinforced concrete cantilever wall control against overturning is redundant in the design process, if
was analyzed following a probabilistic approach. Only the walls the resultant of forces intersects the base of the wall within the
external stability under static conditions was considered through middle third, i.e., if e < B/6. Following is the proof of this
analysis of base sliding, foundation soil bearing capacity, and statement.
excessive eccentricity. In order to represent limit states of equilib- Supposedly a safety factor (FS) applied in the ratio of overturn-
P P
rium, performance functions were dened as safety ratios. Overall ing driving and resisting moments, MOT and MR, respectively,
stability was addressed as a system in series. Engineering proper- is used as the criterion against overturning of the wall about its
ties of the backlled and foundation soil were considered random toe. This is dened as:
variables following a beta distribution. As such, they were each P
MR
represented by their mean, coefcient of variation, minimum, FSOT P
and maximum value. Spatial variability and cross-correlations MOT
among the random variables were not included in the model. Geo- If FSOT < 1, then design is performed again until FSOT > 1. If FSOT > 1,
metric characteristics were considered deterministic quantities, then:
and the heels length was used as variable design parameter. Anal- P
M X X X X
yses were performed for a wide range of this variable, using a P R >1) MR > MOT ) MR  M OT > 0
Monte Carlo simulation algorithm. MOT
P P
A case example was analyzed in order to illustrate the above M R  M OT
) P >0
methodology, and the following conclusions (for this example) V
were drawn from the analysis: Computations of bearing capacity
In this case, based on Eq. (6), the eccentricity may be either
are subjected to higher degree of uncertainty, compared to sliding
smaller or larger than B/6. So, in addition to control with respect
and eccentricity (a conclusions consistent with the observations
to FSOT, a separate check addressing the eccentricity shall be per-
made by Biernatowski and Pua [10]). The same level of safety
formed in order to determine its value.
for the three postulated mechanisms of instability is achieved
Supposedly, on the other hand, that control with respect to FSOT
for different values of the design variable LH. In addition, the mod-
is not performed and one directly determines the eccentricity
al probabilities of failure are not linear functions of the corre-
according to Eq. (6). Then, if e > B6, design is performed again until
sponding safety ratios (nor of the design variable LH). This is
e < B6. If e < 6B, then the following hold true.
why safety ratios do not scale safety. Safety ratios are all posi-
Assuming the resultant falls on the left side of the base center-
tively correlated. Their correlation decreases as LH increases. Par-
line, the eccentricity is dened as:
ticularly high degree of correlation was found between CFS pairs
P P
of sliding bearing capacity and sliding eccentricity. The system B MR  MOT
e  P
probability of failure fell within the rst-order bounds. However, 2 V
depending on the target reliability and the presence or not of a
dominant failure mode, these bounds may lead to a noteworthy In this case we have:
over- or under-estimation of the system probability of failure. P P P P
Obtaining a unique value for the system probability of failure al-
B B MR  M OT B B MR  M OT
e< )  P < ) < P
lows for the optimization of the design process. The results of a 6 2 V 6 3 V
simple optimization analysis, performed for illustrative purposes, But B3 > 0 always holds true, therefore:
indicated that the optimum design variable may signicantly in- P P P
crease with increasing risk factors. In addition, it was shown that MR  MOT X X MR
P >0) MR  MOT > 0 ) P > 1 ) FSOT
the assumption of complete independence or perfect correlation V MOT
between the failure modes, may have an impact on the design >1
variable.
372 I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373

Assuming that the resultant falls on the right side of the base [19] Coduto DP. Foundation design: principles and practices. 2nd ed. Prentice Hall;
2001.
centerline, and in order to dene the eccentricity as a positive
[20] Decisioneering Inc. Crystal Ball v.7.2, Premium Edition. Denver, CO, USA; 2006.
number, the eccentricity is dened as: <www.decisioneering.com>.
P P [21] Decisioneering Inc. Personal communication (on the random number
MR  MOT B generator of Crystal Ball). Denver, CO, USA; 2007.
e P 
V 2 [22] Ditlevsen O. Narrow reliability bounds for structural systems. J Struct Mech
1979;7(4):45372.
In this case we have: [23] Fenton G, Grifths DV. Bearing-capacity prediction of spatially random cu
P P P P soils. Can Geotech J 2003;40(1):5465.
MR  M OT B MR  MOT B [24] Fenton GA, Vanmarcke E. Simulation of random elds via local average
e>0) P  >0) P > subdivision. J Eng Mech 1990;116(8):173349.
V 2 V 2
[25] Fenton GA, Grifths DV, Williams MB. Reliability of traditional retaining wall
design. Geotechnique 2005;55(1):5562.
But B2 > 0 always holds true, therefore:
[26] Gilbert RB, Najjar SS, Choi YJ. Incorporating lower-bound capacities into LRFD
P P X X P codes for pile foundations. In: Anderson JB, Phoon KK, Smith E, Loehr JE,
MR  MOT MR
P >0) MR  MOT > 0 ) P > 1 ) FSOT editors. Geo-Frontiers 2005: contemporary issues in foundation engineering
V M OT (GSP 131), ASCE, January 2426, 2005, Austin, T{, USA; 2005. p. 36177.
[27] Greco VR. Efcient Monte Carlo technique for locating critical slip surface. J
>1 Geotech Eng 1996;122(7):51725.
[28] Grifths DV, Fenton GA. Probabilistic slope stability analysis by nite
So, whenever the eccentricity is smaller than B/6, the safety factor elements. J Geotech Geoenviron Eng 2004;130(5):50718.
against overturning is always larger than one. Therefore, a separate [29] Harr ME. Reliability-based design in civil engineering. McGraw-Hill; 1987.
[30] Hasofer AM, Lind NC. Exact and invariant second-moment code format. J Eng
control against overturning is not necessary.
Mech Div 1974;100(1):11121.
In summary, using a safety factor against overturning requires a [31] Hoeg K, Murarka RP. Probabilistic analysis and design of a retaining wall. J
separate calculation of the eccentricity. On the other hand, using Geotech Eng Div 1974;100(3):34966.
the eccentricity as the criterion for overturning, makes the calcula- [32] Hohenbichler M, Gollwitzer S, Kruse W, Rackwitz R. New light on rst- and
second-order reliability methods. Struct Safety 1987;4(4):26784.
tion of the safety factor a redundant step. These statements are [33] Hunter D. An upper bound for the probability of a union. J Appl Probab
true for safety ratios and eccentricity dened as in the present 1976;3(3):597603.
study. [34] Jaynes ET. Information theory and statistical mechanics. II. Phys Rev
1957;108(2):17190.
[35] Kulhawy FH. On the evaluation of static soil properties. Stability and
References performance of slopes and embankments II In: Seed RB, Boulanger RW,
editors. Geotechnical special publication 31, ASCE, Berkeley, California, June
[1] Ang AH-S, Amin M. Reliability of structures and structural systems. J Eng Mech 29July 1; 1992, p. 95115.
Div 1968;94(2):67191. [36] Kounias EG. Bounds for the probability of a union with applications. Ann. Math
[2] Ang AH-S, Tang WH. Probability concepts in engineering planning and design. Stat 1968;39(6):21548.
Decision, risk and reliability, vol. II. John Wiley & Sons; 1984. [37] Lacasse S, Nadim F. Uncertainties in characterising soil properties. In:
[3] Ang AH-S, Tang WH. Probability concepts in engineering emphasis on Shackelford CD, Nelson PP, Roth MJS, editors. Uncertainty 96: uncertainty in
applications in civil and environmental engineering. 2nd ed. John Wiley & the geologic environment from theory to practice (geotechnical special
Sons; 2007. publication 58), ASCE, July 31August 3 1996, Madison, WI; 1996. p. 4975.
[4] Athanasiou-Grivas D. Reliability analysis of retaining structures. In: Ingles OG, [38] Law AM, Kelton WD. Simulation modeling and analysis. 3rd ed. McGraw-Hill;
editor. 3rd International conference on application of statistics and probability 2000.
on soil and structural engineering (ICASP 3), Kensington NSW, 29 January2 [39] Leonards GA. Investigation of failures. J Geotech Eng Div
February 1979, Sydney, Australia; 1979. p. 63241. 1982;108(2):185246.
[5] Auvinet G. Probabilistic approach in geotechnical engineering practice. In: [40] Low BK. Reliability-based design applied to retaining walls. Geotechnique
Proceedings, 4eme conference nationale sur la abilite des materiaux et 2005;55(1):6375.
structures (JNF05), Clermont-Ferrand, Universite Blaise Pascal, France, [41] Low BK, Wilson H, Tang F. Efcient reliability evaluation using spreadsheet. J
Keynote Lecture; 2005. 9 p [in French]. Eng Mech 1997;123(7):74952.
[6] Auvinet G, Gonzalez JL. Three-dimensional reliability analysis of earth slopes. [42] Lumb P. The variability of natural soils. Can Geotech J 1966;3(2):7497.
Comput Geotech 2000;26(34):24761. [43] Malkawi AIH, Hassan WF, Sarma SK. An efcient search method for nding the
[7] Babu GLS, Mythily M, Rao DB. Probabilistic design of retaining walls. In: critical circular slip surface using the Monte Carlo technique. Can Geotech J
Lacasse S, Singh SK, editors. International conference on offshore and 2001;38(5):10819.
nearshore geotechnical engineering (GEOShore), 23 December 1999, [44] McKone TE. Alternative modeling approaches for contaminant fate in soils:
Mumbai, India; 2000. p. 3836. uncertainty, variability, and reliability. Reliab Eng Syst Safety 1996;54(2
[8] Barry TM. Recommendations on the testing and use of pseudo-random number 3):16581.
generators used in Monte Carlo analysis for risk assessment. Risk Anal [45] Metzger JN, Fjeld RA, Hammonds JS, Hoffman FO. Evaluation of software for
1996;16(1):93105. propagating uncertainty through risk assessment models. Hum Ecol Risk
[9] Benardos AG, Kaliampakos DC. A methodology for assessing geotechnical Assess 1998;4(2):26390.
hazards for TBM tunnelling illustrated by the Athens Metro, Greece. Int J [46] Mostyn GR, Soo S. The effect of autocorrelation on the probability of failure of
Rock Mech Min Sci 2004;41(6):98799. slopes. In: Proceedings, 6th Australia, New Zealand conference on
[10] Biernatowski K, Pua W. Probabilistic analysis of the stability of massive bridge geomechanics: geotechnical risk; 1992. p. 5426.
abutments using simulation methods. Struct Safety 1988;5(1):115. [47] Oboni F, Bourdeau PL. Simplied use of the beta distribution and sensitivity to
[11] Blazquez R, Der Kiureghian A. Seismic reliability of retaining walls. In: Lind NC, the bound locations. Struct Safety 1985;3(1):636.
editor. 5th International conference on application of statistics and probability [48] Onghena P. A theoretical and empirical comparison of mainframe,
in soil and structural engineering (ICASP 5), 2529 May 1987, BC (Canada): microcomputer, and pocket calculator pseudorandom number generators.
Vancouver; 1987. p. 114956. Behav Res Methods, Instrum, Comput 1993;25(3):38495.
[12] Bolle A. Investigation and allowance for spatial variability. Rev. Franaise de [49] Orr TLL. Selection of characteristic values and partial factors in geotechnical
Geotech 2000;93:5566. designs to Eurocode 7. Comput Geotech 2000;26(3):26379.
[13] Bourdeau PL, Amundaray JI. Non-parametric simulation of geotechnical [50] Oztas A, Okmen O. Risk analysis in xed-priced design-build construction
variability. Geotechnique 2005;55(2):95108. projects. Build Environ 2004;39(2):22937.
[14] Bourdeau PL, Gutierrez A. Inuence de la corrlation entre modes de [51] Peck RB, Hanson WE, Thornburn TH. Foundation engineering. 2nd ed. John
dfaillance sur le dimensionnement des soutnements massifs. In: Bourdeau Wiley & Sons; 1974.
PL, editor. Symposium on reliability-based design in civil engineering, 79 July [52] Phoon K-K, Kulhawy FH. Characterization of geotechnical variability. Can
1988, EPFL, Lausanne, Switzerland; 1988. p. 34350. Geotech J 1999;36(4):61224.
[15] Bowles JE. Foundation analysis and design. 5th ed. McGraw-Hill; 1997. [53] Phoon K-K, Kulhawy FH, Grigoriu MD. Multiple resistance factor design for
[16] Cardoso AS, Fernandes MM. Characteristic values of ground parameters and shallow transmission line structure foundations. J Geotech Geoenviron Eng
probability of failure in design according to Eurocode 7. Geotechnique 2003;129(9):80718.
2001;51(6):51931. [54] Pua W. 2002. On spatial averaging in reliability computations of shallow
[17] CEN. Eurocode 7: geotechnical design. Part 1: general rules, ENV 1997-1. foundations. In: Pande GN, Pietruszczak S, editors. Numerical models in
Comit Europen de Normalisation European Committee for geomechanics: proceedings of the 8th international symposium (NUMOG VIII),
Standardization, Brussels; 1994. 1012 April 2002, Rome, Italy; 2002.
[18] Cherubini C. Reliability evaluation of shallow foundation bearing capacity on [55] Pua W, Shahrour I. Inuence of vertical and horizontal variability of soil
c0 , u0 soils. Can Geotech J 2000;37(1):2649. strength on the reliability of shallow foundations. In: Symposium
I.E. Zevgolis, P.L. Bourdeau / Computers and Geotechnics 37 (2010) 359373 373

international sur les foundations supercielles (FONDSUP 2003), 57 [63] US Army Corps of Engineers. Retaining and ood walls. EM 1110-2-2502,
November 2003, Paris, France, vol. 1; 2003. p. 42332. Department of the Army, Washington, DC, USA; 1989.
[56] Rackwitz R. Reviewing probabilistic soils modeling. Comput Geotech [64] Vanmarcke EH. Probabilistic modeling of soil proles. J Geotech Eng Div
2000;26:199223. 1977;103(11):122746.
[57] Snchez-Bruno A, Luis-Costas CS. A statistical analysis of seven multipliers for [65] Vanmarcke E. Random elds. Massachusetts (USA): The MIT Press, Cambridge;
linear congruential random number generators with modulus 231  1. Qual 1983.
Quant 1995;29(3):3317. [66] Warith M, Fernandes L, Gaudet N. Design of in-situ microbial lter for the
[58] Shooman ML. Probabilistic reliability: an engineering approach. 2nd ed. RE remediation of naphthalene. Waste Manage 1999;19(1):925.
Krieger Publishing Company; 1990. [67] Wolf TF. Analysis and design of embankment dam slopes: a probabilistic
[59] Singh G. How reliable is the factor of safety in foundation engineering? In: approach. PhD dissertation, School of Civil Engineering, Purdue University,
Lumb P, editor. 1st International conference on application of statistics and West Lafayette, IN, USA; 1985.
probability to soil and structural engineering (ICASP 1), 1316 September [68] Wolff TF. Probabilistic assessment of pile interference. J Geotech Eng
1971, Hong Kong; 1972. p. 390424. 1993;119(3):52542.
[60] Szwed A, Nowak AS, Withiam JL. Reliability models for bridge substructures. [69] Wolff TF. Spreadsheets for geotechnical engineering. PWS Publications; 1995.
In: 9th Joint specialty conference on probabilistic mechanics and structural [70] Wu TH. Uncertainty, safety, and decision in soil engineering. J Geotech Eng Div
reliability (PMC 2004), ASCE, July 2628 2004, Albuquerque, NM, USA, Paper 1974;100(3):32948.
3.03; 2004. [71] Zevgolis IE. Numerical and probabilistic analysis of reinforced soil structures.
[61] Tang WH. Correlation, multiple random variables, and system reliability. In: PhD Dissertation, School of Civil Engineering, Purdue University, West
Fenton GA, editor. Workshop presented at GeoLogan 97 conference: Lafayette, IN, USA; 2007.
probabilistic methods in geotechnical engineering, July 15 1997, Logan, [72] Zevgolis IE, Bourdeau PL. System reliability of cantilever retaining walls with
Utah, USA; 1996. p. 3950. correlated failure modes. In: DeGroot DJ, DeJong JT, Frost D, Baise LG, editors.
[62] Terzaghi K, Peck RB, Mesri G. Soil mechanics in engineering practice. 3rd GeoCongress 2006: geotechnical engineering in the information technology
ed. John Wiley & Sons; 1996. age, ASCE, February 26 March 1, 2006, Atlanta, GA, USA; 2006.

You might also like