You are on page 1of 14

672

Reliability analysis of reinforced embankments


on soft ground
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15

B.K. Low and Wilson H. Tang

Abstract: A limit equilibrium stability model of reinforced embankments on soft ground is formulated and a practical
reliability evaluation procedure is proposed. The stability model allows for a tension crack in the embankment, tensile
reinforcement at the base of embankment, and a nonlinear undrained shear strength profile in the soft ground. The proposed
reliability evaluation procedure derives from an intuitively obvious ellipsoidal perspective. It obtains the same reliability index
as the current first-order reliability method (FORM), but is more practical and transparent because the spreadsheet-automated
search is in the original space of the variables. The proposed method deals with correlations without orthogonal
transformation of the covariance matrix. The versatility of the proposed method is demonstrated in an example reinforced
embankment that has 12 random variables, including six autocorrelated strength values, and a performance function that is
nonexplicit due to the need to search for the critical slip surface and to integrate numerically the depth-dependent shear
strength. The probability of failure compares well with Monte Carlo simulation. Reliability-based design is also discussed,
and comparisons with case histories are made. The ease of implementation, conceptual clarity, and versatility of the proposed
method should enhance the wider use of the more rational reliability-based analysis and design over the conventional
factor-of-safety approach.
Key words: embankment, reinforcement, stability, reliability, probability, spreadsheet.

Résumé : L’on a formulé un modèle de stabilité à l’équilibre limite de remblais sur sols mous, et proposé une procédure
For personal use only.

pratique d’évaluation de la fiabilité. Le modèle peut tenir compte d’une fissure de traction dans le remblai, d’armature en
tension à la base du remblai, et d’un profil non linéaire de résistance au ciaillement non drainé dans le sol mou. La procédure
d’évaluation de la fiabilité proposée découle d’une perspective ellipsoïdale intuitivement évidente. Elle génère le même indice
de fiabilité que la méthode courante de fiabilité de premier ordre (FORM), mais elle est plus pratique et plus transparente
parce que la recherche automatisée sur le chiffrier est dans l’espace originale des variables. La méthode proposée manipule
des corrélations sans transformation orthogonale de la matrice de covariance. La versatilité de la méthode proposée est
démontrée dans un exemple de remblai armé qui a 12 variables aléatoires, incluant 6 valeurs de résistance autocorrélées, et
une fonction de performance qui est non explicite à cause de la nécessité de rechercher la surface critique de glissement et
d’intégrer numériquement la résistance au cisaillement en fonction de la profondeur. La probabilité de rupture se compare
bien avec la simulation de Monte-Carlo. L’on discute également de la conception basée sur la fiabilité, et l’on fait des
comparaisons avec des histoires de cas. La facilité d’exécution, la clarté du concept et la versatilité de la méthode proposée
devraient favoriser un usage plus étendu de l’analyse et de la conception basées sur la fiabilité, plus rationnelles que
l’approche conventionnelle du coefficient de sécurité.
Mots clés: remblai, armature, stabilité, fiabilité, probabilité, chiffrier.
[Traduit par la rédaction]

Introduction Tang (1993), Christian et al. (1994), Chowdhury and Xu


The conventional factor of safety in geotechnical engineering (1995), and Kitch et al. (1995), among others.
cannot reflect the uncertainty of its underlying parameters. A The reliability index β as defined by Hasofer and Lind
better approach is to compute a reliability index that depends (1974) can be interpreted in the original space of the random
not only on the mean values of the parameters but also on their variables, as shown in Fig. 1, where m1, σ1, m2, and σ2 are the
inherent scatter. Applications of reliability concepts in slopes mean values and standard deviations of the parameters X1 and
have been reported by Vanmarcke (1977), Whitman (1984), Li X2, respectively. One may regard β as the minimum distance
and Lumb (1987), Luckman et al. (1987), Oka and Wu (1990), from m to the boundary of the failure region, in units of direc-
tional standard deviation.
An established elegant procedure for computing the β index
Received November 15, 1996. Accepted May 6, 1997. is to transform the failure surface into the space of reduced
B.K. Low. School of Civil and Structural Engineering,
variates, whereby the shortest distance from the transformed
Nanyang Technological University, Nanyang Avenue, failure surface to the origin of the reduced variates is the reliabil-
Singapore 639798. ity index β. The procedure involves first rotating the frame of
W.H. Tang. Department of Civil and Structural reference to the new Y1 and Y2 axes in Fig. 1, which are each
Engineering, Hong Kong University of Science and parallel to a principal axis of the original dispersion ellipse (or
Technology, Clear Water Bay, Kowloon, Hong Kong. ellipsoid for higher dimensions). This is followed by translation

Can. Geotech. J. 34: 672–685 (1997). © 1997 NRC Canada


Low and Tang 673

Fig. 1. Reliability index in original space and classical mathematical approach in transformed space.
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15
For personal use only.

and normalization so that the one-sigma (1σ) dispersion ellip- need for a more comprehensive risk assessment approach
soid in original space becomes a unit sphere centered at the (e.g., Morgenstern 1995).
origin of the transformed space. The solution process involves The stability model and the conventional deterministic factor-
orthogonal transformation of the covariance matrix, by finding of-safety analysis will be presented before the reliability analy-
eigenvalues and eigenvectors, and iterative partial differentia- sis, in order to show the link between the two methods of
tion. It is well explained in Ditlevsen (1981), Ang and Tang analysis, the latter being an extension of the former and more
(1984), Madsen et al. (1986), Melchers (1987), Dai and Wang rational than the former.
(1992), Tichy (1993), and Haldar and Mahadevan (1995), Other methods of limit equilibrium analysis of slopes are
among others. reviewed in Duncan (1996).
An alternative interpretation of the Hasofer-Lind β index
was described in Low and Tang (1997), where the perspective Stability model of reinforced embankment
of an expanding ellipsoid in the original space of the variables
on soft ground
led to a simple method of computing the Hasofer-Lind index
using an optimization tool available in most spreadsheet soft- The stability model presented below extends the work of Low
ware packages. The method achieves the same result as the (1989) and Low et al. (1990) to cases involving tension cracks
classical approach, but without involving the concepts of or- and water pressure in cracks. Exact equations for arc lengths—in
thogonal transformation and transformed space. contrast to the expedient but quasi-analytic expressions in the
In this paper the spreadsheet method of reliability analysis earlier two papers—are used, made possible because the search
is applied to reinforced embankments on soft ground. A sig- for the critical slip surface is now automatic using a spread-
nificant difference between this reliability analysis and those sheet’s optimization tool.
found in Low and Tang (1997) is that the problem being ad- The factor of safety against rotational failure can be defined
dressed has no closed form performance functions, due to the in terms of overall moment equilibrium. In Fig. 2, the origin of
presence of a user-defined macro function to perform numeri- the x–z coordinate system is at the toe of the embankment. The
cal integration of shear strength and the need to search differ- coordinate values of x and z represent the center of an arbitrary
ent slip circles. Nevertheless, the procedure as implemented slip circle, of radius equal to (z + D), where D is an arbitrary
by the user is simple, transparent, and intuitive. The ellipsoid “tangent depth” in the soft ground. The overturning moment (Mo)
approach via spreadsheet may contribute to reducing the lan- derives from block ABJKEA. When a tension crack intersects the
guage barrier in geotechnical reliability noted by Whitman slip surface, the overturning moment due to the water pressure in
(1984), the National Research Council (1995), and Hoek et al. the tension crack has to be added, and that due to block JKQ
(1995), all of which contained lucid explanations of reli- subtracted. By integration, the exact overturning moment (Mo) of
ability concepts. The example calculations below aim to block ABJKEA and the net effect (∆Mo) of water pressure in
demonstrate the versatility—and to suggest possibilities yet crack and block JKQ are as shown by eqs. d and e of Fig. 3,
unexplored—of the spreadsheet approach, without excluding respectively, in which the symbols are as defined in Fig. 2.
the possibilities of refinement in the analytical model, in The available soil resisting moment MR along IGEK
the treatment of the uncertain statistical parameters, or the (Fig. 2) consists of the two components given in eq. f: the first

© 1997 NRC Canada


674 Can. Geotech. J. Vol. 34, 1997

Fig. 2. Notations for reinforced embankment on soft ground.


Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15

due to the equivalent undrained shear strength ca along arc IGE A feature of the stability model described above is that the
(subtending an angle 2θ, where θ is given by eq. b); the second overturning moment (Mo + ∆Mo) is obtained exactly by con-
due to embankment cohesion (cm) and friction (φm) along arc sidering the blocks ABJKEA and JKQ. It may also be men-
EK (subtending an angle θh, eq. c). The equivalent undrained tioned that for a soft foundation with horizontal top surface
shear strength ca is determined numerically by a user-defined and with shear strength characterized by cu only, the unit
For personal use only.

spreadsheet function that averages the depth-dependent un- weight of the soft foundation has no effect on the factor of
drained shear strength cu over the arc length within the depth safety against the initiation of rotational failure. The γ that
D, as explained in eq. g. The λ, which also contributes to MR, appears in the equations of Fig. 3 and in the dimensionless
is calculated using eq. h, which is an approximation. The ra- parameters (ca/γH and cm/γH) of Low’s (1989) stability charts
tionale is described in Low (1989). It affects the frictional refers solely to the unit weight of the embankment material.
resistance along arc EK in the embankment. For embankments To erroneously take γ to be the weighted average of the em-
on soft ground, the contribution of embankment friction angle bankment unit weight and the soft foundation unit weight is
(φm) to the resisting moment MR is usually minor; the approxi- likely to overestimate the safety, since the unit weight of a soft
mate nature of λ is therefore likely to be inconsequential in foundation clay is usually lighter than that of a compacted
most cases. (In cases where φm = 0 in the embankment and φu = 0 embankment.
in the foundation soil—i.e., when only cm in embankment and
cu in foundation contributes to shear strength along the slip Deterministic analysis: nonlinear
surface IGEK in Fig. 2—the equation for MR is exact and the
optimization using spreadsheet
calculated minimum factor of safety will be identical to that
from the ordinary method of slices and the Bishop’s simplified An optimization tool exists (under different names) in most
method.) spreadsheet software packages, for example the Solver in Microsoft
Other than the available resisting moment MR that derives Excel and Lotus 123, the Optimizer in Quattro Pro. The tool can
from the shearing resistance of soil, the available resisting mo- be used to identify among the infinite slip circles (defined by
ment (MT) due to tensile reinforcement of strength T also en- x, z, and D) the one that produces the smallest factor of safety
hances the factor of safety. This component is given by eq. i, for the stability model described above. An example spread-
where the lever arm can be assumed equal to the average be- sheet setup (using Microsoft Excel version 5 or 7) is shown
tween two scenarios, namely that (1) T acts horizontally, and in Fig. 4. The input parameters are in the unframed cells. Values
(2) T acts tangential to the slip surface at limiting conditions. within boxes are calculated by entering the formulas shown in
Low et al. (1990) discussed the effects of the different assump- Fig. 3. The equivalent undrained shear strength ca(D, z, cu pro-
tions of the lever arm of T on the factor of safety. file) is evaluated numerically by a user-defined spreadsheet
The equations in Fig. 3 are for an arbitrary slip circle of function (also shown in Fig. 4) that integrates the undrained
center (x, z) and extending to depth D in the soft foundation. shear strength profile over the arc length in the foundation soil.
The minimum factor of safety has to be determined by trying (In this example the undrained shear strength profile is defined
different combinations of x, z, and D in eq. j (or eq. k). This, by six cu values at different depths.) Initial values need to be
together with the use of numerical integration to determine the assigned to x, z, and D. For embankments on soft ground,
equivalent undrained shear strength ca (eq. g), makes the sta- suitable initial values are x = l/2 (i.e., midpoint of slope),
bility model nonexplicit. Nevertheless, a widely available z = 2H, D = the mid-depth of soft foundation. For the case in
spreadsheet optimization tool greatly facilitates the search for Fig. 4, initially x = 4, z = 8, D = 6; other arbitrary but reason-
the minimum factor of safety and the associated critical slip able initial values can be used. (Examples of unreasonable
surface, as described in the next section. initial values of (x, z, D) are: z smaller than H; D outside the

© 1997 NRC Canada


Low and Tang 675

Fig. 3. Equations for stability analysis of reinforced embankments on soft ground. H, Ω, x, z, D, γ, γw, hw, hc, cm, φm, T, and lever as defined in
Fig. 2 and related text.
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15
For personal use only.

defined domain of the soft foundation; or initial (x, z, D) tool (e.g., Microsoft Excel’s Solver) is that the target and con-
such that the slip surface fails to intersect the embankment straint cells could be calling complicated user-defined macro
at all.) functions or could contain nested IF statements. The optimi-
The Solver tool in Excel is then invoked to “Minimize” Fs, zation tool need not know the algorithms by which the cell
“By Changing” x, z, and D, “Subject To” the constraint that D values are determined because the gradients required for it-
be within the foundation thickness for which undrained shear erative searching are calculated automatically by the optimi-
strength profile has been defined. The critical slip surface (x, zation tool via numerical partial differentiation. The general
z, D) and the minimum Fs (1.289) obtained by Solver are as reduced gradient (GRG) method is used by Solver for non-
shown in Fig. 4. linear optimization problems. Additional information on
The ease of searching with the spreadsheet optimization Solver options, algorithms, and completion messages can be
tool greatly facilitates experimenting with different initial values found in the Microsoft Excel Solver’s help file, and at the
should the need arise. For example, a different initial depth D website www.frontsys.com.
should be tried if the soft foundation shear strength profile Since explicit equations are not required, and numerous
varies with depth cyclically (e.g., due to a sequence of alter- constraints and changing cells can be specified, the spread-
nating hard and soft layers) in order to identify the over- sheet optimization method is likely to be more versatile and
all minimum Fs from among the local minimums. flexible than mathematical methods that work on closed
An attractive feature of using spreadsheet’s optimization form equations.

© 1997 NRC Canada


676 Can. Geotech. J. Vol. 34, 1997

Fig. 4. Example deterministic analysis of reinforced embankment on soft ground (units: m, kN/m, kN⋅m/m, kN/m2, kN/m3).
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15
For personal use only.

The meaning of reliability index in original are simply viewed as forming a tilted ellipsoid centered at the
space mean in original space, as discussed by Low (1996) and Low
and Tang (1997).
The above example of deterministic analysis does not reflect The quadratic form (x − m)T C−1(x − m) in eq. 1 appears
the uncertainty in the input parameters. Ambiguities could also also in the negative exponent of the multivariate normal dis-
arise from different definitions (e.g., eq. j vs. eq. k in Fig. 3) tribution (e.g., Johnson and Wichern 1992, chapter 4). The
given to the factor of safety of reinforced embankments on soft quadratic form is an ellipsoid. For example, the one standard
ground, as discussed in Low et al. (1990). A more rational deviation (1σ) ellipse in Fig. 1 corresponds to β = 1. One may
approach is to evaluate the second moment reliability index. regard the gradually expanding ellipsoidal surfaces as repre-
The matrix formulation (e.g., Veneziano 1974; Ditlevsen senting the contours of decreasing probability density. There-
1981; Madsen et al. 1986) of the Hasofer-Lind index β is fore, for normally distributed variables, to minimize β in eq. 1
 T 1/2
(or β2 in the multivariate normal distribution) is to maximize
[1] β = min (x − m) C−1 (x − m)  the value of the multivariate normal probability density func-
xe F   tion; and to find the smallest ellipsoid that is tangent to the
failure surface is equivalent to finding the most probable fail-
where x is a vector representing the set of random variables, m ure point. This perspective is consistent with Shinozuka (1983)
their mean values, C the covariance matrix, and F the failure that “the design point x* is the point of maximum likelihood
region. The above index is traditionally interpreted as the if x is Gaussian, whether or not its components are uncorrelated.”
shortest distance from the origin of the normalized variates to Another perspective is to regard the Hasofer-Lind reliabil-
the failure surface in transformed space (right sketch, Fig. 1). ity index β as the axis ratio of the ellipsoid that just touches
The spreadsheet optimization scheme described in the next the failure surface to the 1σ dispersion ellipsoid, as shown in
section operates directly on eq. 1 in the original space of the Fig. 2 of Low and Tang (1997). A little reflection will show
random variables. Instead of rotating the frame of reference and that this perspective and that in Fig. 1, namely β = minimum
scaling the rotated coordinate axis, correlated random variables [R(θ)/r(θ)], are equivalent.

© 1997 NRC Canada


Low and Tang 677

Fig. 5. Example reliability analysis of reinforced embankment on soft ground.


Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15
For personal use only.

Reliability analysis: ellipsoid approach via optimization iterations.) The coefficient of variation is as-
spreadsheet sumed to be 0.15 for each of the six cu values.
The diagonal terms of the correlation matrix are 1. The
In the spreadsheet method one literally sets up a tilted ellipsoid embankment cohesion (cm) and friction angle (φm) are as-
in the spreadsheet and minimizes the ellipsoid subject to the sumed to be positively correlated with the unit weight (γ) of
constraint that it be tangent to the failure surface. For this the embankment (ρ12 = ρ13 = 0.5) and negatively correlated
purpose eq. 1 is rewritten as with each other (ρ23 = ρ32 = –0.3). The undrained shear
shrength profile in the soft foundation is in this case charac-
 T 1/2

  xi − mi   xi − mi   terized by six cu values at different depths, with correlation
[2] β= min    [R]−1    coefficients (lower right quadrant of the correlation matrix)
xe F 
  σi   σi   obeying the following negative exponential model of spatial
continuity:
 |[depth (i) − depth (j)]| 
in which [R]–1 is the inverse of the correlation matrix. This [3] ρij = exp − 
equation will be used to set up the multidimensional ellipsoid  3 
in spreadsheet since the correlation matrix R displays the cor- where the depths are in metres.
relation structure more explicitly than the covariance matrix C. The autocorrelation distance of 3 (metre) in the exponent
The mean values, the standard deviations σ, and the corre- implies that the correlation coefficient is 0.368 for two points
lation matrix [R] are assumed known and entered, as shown in 3 m apart, and 0.05 for 2 points 9 m apart, and so on.
the spreadsheet in Fig. 5. The mean value of “lever” is entered The subsequent steps are as follows:
as an equation, as shown. The standard deviations σ are also (1) Formulas are entered for the column vector [(xi – mi)/σi],
entered as formulas, equal to the respective mean values mul- where xi, mi, and σi represent the “xvalue”, mean, and
tiplied by their coefficients of variation; these coefficients are standard deviation, respectively, of a random variable, as
0.05, 0.15, 0.1, 0.1, 0.15, and 0.3 for γ, cm, φm, T, lever, and hc listed in Fig. 5. The “xvalues” are set equal to the mean
respectively. (Hence both the mean value and standard devia- values initially (by copying the “Mean” column and past-
tion of “lever” will vary with z and D during the spreadsheet’s ing over the “xvalues” column).

© 1997 NRC Canada


678 Can. Geotech. J. Vol. 34, 1997

Table 1. Using GoalSeek and Solver repeatedly to obtain the mean value of T required for a β index of 3.0.
Successive trials
1 2 3 4
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15

GoalSeek for β = 3.0 3.000 3.000 3.000 3.000


Required mean of T 229.1 243.9 247.8 248.3
Minimum β using Solver 2.799 2.957 2.995 3.000

Fig. 6. Programmed iterative calling of Solver routine, as an attributable to their being positively correlated to the unit
alternative to the procedure shown in Table 1. weight of the embankment γ and also reflects their rela-
tively minor roles in affecting the reliability for the case in
hand. (If the positive correlation 0.5 in matrix entries ρ12, ρ13,
ρ21, and ρ31 are replaced by zero, the reliability index β is
reduced to 2.415, and γ, cm, and φm at tangency point have
values 19.90, 9.835, and 29.79, respectively, the last two being
smaller than their respective mean values.)
Although the correlation coefficient between two random
variables has a range –1 ≤ ρij ≤ 1, one is not totally free in
assigning any values within this range for the correlation ma-
trix. This was explained in Ditlevsen (1981, p. 85) where a 3 × 3
correlation matrix with ρ12 = ρ13 = ρ23 = –1 implies a contra-
diction and is hence inconsistent. The covariance matrix has
to be positive definite. For the present approach, an inconsis-
For personal use only.

tent correlation matrix will manifest itself by causing Solver


to break down due to “error encountered” when Solver tries to
(2) Select the entire correlation matrix, and define a name for evaluate the square root of a negative quantity—the right-hand
it, say “crmatrix,” using the menu command insert/name/de- side of eq. 2—during its iterative searching for the minimum
fine. Likewise, define the column vector [(xi – mi)/σi] as dispersion ellipsoid and the corresponding reliability index β.
“nxvector”, say. For the case in hand the correlation matrix in Fig. 5 becomes
(3) The formula (eq. 2) of the reliability index, β = singular when ρ23 = ρ32 = –0.5. This means that given their
sqrt{[(x – m)/σ]T[R]–1[(x – m)/σ]}, is typed in a cell: “ = positive correlation (ρ21 = ρ31 = 0.5) with γ, the correlation
sqrt(mmult(transpose(nxvector), mmult(minverse(crmatrix), between cm and φm has to be greater than –0.5.
nxvector))),” followed by pressing “Enter” while holding For the example analysis in Fig. 5, an additional random
down the “Ctrl”and “Shift” keys. (mmult, transpose, variable M can be added to represent model uncertainty,
and minverse are Microsoft Excel’s built-in spreadsheet thereby giving a 13 × 13 correlation matrix and a somewhat
functions for matrix operations.) smaller reliability index. One source of model uncertainty is
(4) The formula of the performance function, g(x) = MR + MT – the possible discrepancy between the present stability model
(Mo + ∆Mo), is entered based on “xvalues.” and a more rigorous one (e.g., one that satisfies force equilib-
(5) Initial values are set for x, z, and D, e.g., x = l/2 = 4 m, z = rium in addition to moment equilibrium or one that accounts
2H = 8 m, D = ½ clay thickness = 6 m. (The remarks in for three-dimensional effect).
the earlier section “Deterministic Analysis...” regarding The critical slip surface (as indicated by the values of x, z,
reasonable initial values of x, z, and D also apply here.) and D) in Fig. 5 differs little from that in the deterministic
(6) Solver is invoked, to “Minimize” β, “By Changing” the 12 analysis. Further analyses show that the difference between the
“xvalues” and also (x, z, D), “Subject To” the constraints critical slip surface of the deterministic analysis and that of the
PerFunc = 0, D > = 0, and D < = 12. reliability analysis is a function of the dispersion of the random
The solution obtained by Solver is as shown in Fig. 5. The variables. For example, if the coefficient of variation (c.o.v.)
spreadsheet approach is simple and intuitive because it works of the cu values is 0.25 instead of 0.15, Solver obtains a reli-
in the original space of the variables; it does not involve the ability index β of 1.692 (instead of 2.467), and a critical slip
orthogonal transformation of covariance matrix or user-input surface defined by (x, z, D) = (4, 10.12, 8.01). The decrease in
partial derivatives, both of which are required in the traditional reliability is mainly due to the increased uncertainty of the cu
FORM procedure. values (which a deterministic safety analysis will not be able
The 12 “xvalues” as obtained by Solver in Fig. 5 represent to reflect).
the most probable point on the failure surface. It is the point A noteworthy feature of the proposed spreadsheet reli-
of tangency of the 12-dimensional dispersion ellipsoid with the ability evaluation procedure is that it automatically searches
failure surface. At this most probable point the values of cu, T, for the critical slip surface based on reliability criteria. It
and “lever” are, as expected, smaller than their corresponding does not assume that the critical slip surface is the same
mean values; but cm and φm at tangency point have slightly higher as the deterministic critical slip surface—an assumption
values than their respective mean values. This peculiarity is common in reliability analysis hitherto due to the difficulty

© 1997 NRC Canada


Low and Tang 679

Fig. 7. Deterministic analysis of an unreinforced test embankment at Nong Ngoo Hao (units: m, kN/m, kN⋅m/m, kN/m2, kN/m3).
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15
For personal use only.

of performing reliability evaluation for each slip surface by of 3.0 is 248.3 kN/m. Had the coefficient of variation been
the traditional method. higher, the required mean of T would also be larger.
The above iterative procedure is needed because the “xval-
ues” and the critical slip surface (x, z, D) will be different when
Design considerations based on reliability the mean of T is changed to achieve a desired reliability index.
index
For design it would be of interest to know what changes need Further simplifications using macros and
be made to the mean value or standard deviation of a random notes on constraints of changing cells
variable so as to increase the reliability index to a desired
value. For example, in Fig. 5, the mean value of the strength of The above problem of finding the required mean value of a
the reinforcement (T) is 200 kN/m, with a coefficient of vari- random variable that will yield a target reliability index can be
ation equal to 0.1. One may ask what higher mean value of T further automated (if desired) by writing a simple macro using
(by increasing the amount of geosynthetic reinforcement, for the VBA programming language that resides in Microsoft Ex-
example) would increase the reliability index from 2.467 to cel. The short code is shown in Fig. 6. Each iteration loop
3.0, given the same coefficient of variation for T. The answer invokes the Solver routine twice: the first time to set the β cell
can be obtained by using both the GoalSeek tool and the Solver to the target value 3 by changing only the mean value of T;
tool of the spreadsheet repeatedly. With the values of β and the second time to minimize the β value by changing all the
mean T (2.467 and 200, respectively) as shown in Fig. 5, the 12 “xvalues” and also (x, z, D). The answer obtained after six
GoalSeek tool was invoked to find the mean of T that would iteration loops is the same as that obtained by the repeated
set β equal to the design index value of 3.0. The answer given manual use of the GoalSeek and Solver tools described in the
by GoalSeek is 229.1 kN/m, as shown in Table 1. This was previous section.
followed by using the Solver tool to minimize β by changing In using the spreadsheet’s optimization tool, it is sometimes
all the “xvalues” of the 12 random variables, and x, z, and D, necessary to specify as constraints the legitimate bounds for
subject to the “xvalues” being on the failure surface (i.e., Per- each of the “Changing Cells.” For reliability index computa-
Func = 0 in Fig. 5). The minimized β is 2.799. Again GoalSeek tion, an intrinsic constraint exists in the sense that the final
was invoked to set β to 3.0 by changing the mean value of T, “xvalues” (Fig. 5) tend to be at most a few standard deviations
and the process continues. Convergence by this successive ap- from the mean values. Hence, as a guide, one may start with
proach is fast, as shown in Table 1. For a coefficient of vari- only the essential constraints (e.g., that the Performance Func-
ation 0.1, the required mean value of T to achieve a beta index tion be equal to zero in Fig. 5). If Solver found a solution in

© 1997 NRC Canada


680 Can. Geotech. J. Vol. 34, 1997

Fig. 8. Examples of reinforcement scenarios for embankment at Nong Ngoo Hao (units: m, kN/m, kN⋅m/m, kN/m2, kN/m3).
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15
For personal use only.

which all the final values of the “Changing Cells” are within be done more efficiently via the proposed spreadsheet reliabil-
their permissible bounds (e.g., all positive), additional constraints ity evaluation procedure than Monte Carlo simulation.
are not necessary. (Solver can handle up to 500 constraints—one
upper limit and one lower limit for each changing cell, plus Dependency of deterministic solution on
100 additional constraints—according to Microsoft Excel 7’s the definition of Fs
on-line help. Each changing cell can be constrained by a number,
another cell or range reference, or formula). In the design of a reinforced embankment on soft ground, it is
often necessary to estimate the magnitude of the reinforcement
force (T) that is required to achieve a desired factor of safety,
Probability of failure and comparison with or, given a value of reinforcement force, to determine the fac-
Monte Carlo simulations tor of safety (Fs). The solution obtained (whether T or Fs)
depends on the definition of Fs. This ambiguity was noted in
For linear performance function and normally distributed random Hoek and Bray (1981) for reinforced rock slopes. The same
variables, the mathematical probability of failure (Pf) is related to ambiguity exists for deterministic limit equilibrium analysis of
reliability index β by the equation Pf = 1 – Φ(β), where Φ(.) is the reinforced embankments on soft ground, as both eqs. j and k in
cumulative distribution function of the standard normal variate. Fig. 3 have been used in practice, as discussed in Low et al.
The value of Pf can be obtained using the spreadsheet function (1990). It was shown that seemingly different T or Fs solutions
normsdist in the formula “ = 1 – normsdist(beta).” For β = 1, 2, to the same reinforced embankment can be reconciled if the
3, and 4, the values of Pf are 0.16, 0.023, 0.001, respectively, different definitions of Fs used are accounted for.
and 3E–05, respectively. The reliability index of 2.467 in It seems logical that eq. j be used if the T force is the maximum
Fig. 5 implies a Pf value of 0.0068. mobilizable force (Tp), and eq. k be used if the T force is either
Monte Carlo simulation of the problem in Fig. 5 with 10 000 the allowable force or the actual mobilized force (Ta). In the for-
realizations using the popular spreadsheet add-in program @Risk mer case the computed Fs applies to both the reinforcement
produces a Pf value of 0.0072. Further runs with 100 000 reali- strength and the soil shear strength; it is a global Fs of the rein-
zations result in a Pf value of 0.0065. The 95% confidence forced soil system. In the latter case the computed Fs is that with
interval based on 100 000 realizations is estimated to be respect to the shear strength of soil only, the factor of safety with
0.0060 < Pf < 0.0070, which agrees well with the 0.0068 based respect to reinforcement being given by the ratio Tp/Ta.
on reliability index β. It may be inferred that the limit state The ambiguity arising from the different formulations of Fs
surface of the problem in Fig. 5 is not highly nonlinear. does not exist if the reliability index as defined by Hasofer and
The time required for Monte Carlo simulation is at least an Lind (1974) is used, because at the limit state surface (Fs = 1),
order of magnitude higher than the evaluation of reliability index eqs. j and k of Fig. 3 are identical.
using spreadsheet. If sensitivity information or a parametric study The assumed lever arm also affects the solution of Fs or T.
is desired, or a reliability-based design is to be performed, it can In this paper, the following has been assumed: (1) if Fs is

© 1997 NRC Canada


Low and Tang 681

Fig. 9. Deterministic analysis of reinforced embankment on Bangkok clay (units: m, kN/m, kN⋅m/m, kN/m2, kN/m3).
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15
For personal use only.

defined by eq. j of Fig. 3, then “lever arm” = z + D/2, that is, surface corresponding to a minimum factor of safety. The fac-
the line of action is between horizontal and tangential; (2) if tor of safety was expressed in terms of the stresses from finite
Fs is defined by eq. k, then “lever arm” = z, that is, the line of element analysis. The undrained shear strength (cu) profile of
action is horizontal. The former scenario was the case in Fig. 4. the soft clay beneath the embankment was obtained by apply-
For reliability analysis the mean lever arm is assumed to be ing a reduction factor of 0.7 to a vane shear strength (suv)
equal to (z + D/2). As shown in Fig. 5, the xvalue of “lever” profile. The dense sand fill of the embankment was assumed
on the limit state surface is smaller than its mean (i.e., 10.65 < to have a unit weight of 18 kN/m3 and an average friction angle
12.63) when T and “lever” are uncorrelated. of 37.5°. Deformation properties were also needed for the fi-
nite element analysis. A factor of safety equal to 1.00 was
Comparisons with published case histories reported at the 3.4 m failure height of the embankment. It was
concluded that the finite element method has the advantage
In this section the stability formulas of this paper are applied over limit equilibrium methods in that the strain- and time-
to three published full-scale trial embankmets on soft ground. dependent behaviour of soil and the stage construction of
In the first two cases, the embankment height was increased slopes can be modelled.
until failure occurred. In the third case, the reinforced embank- To compare the present formulations (eqs. in Fig. 3) with Zou
ment was stable at final height. et al.’s (1995) finite element method (FEM)-based search for
critical slip surface, the same assumptions regarding cu profile and
Case 1: Unreinforced test embankment at Nong Ngoo Hao embankment soil properties are used, as shown in the spreadsheet
The Nong Ngoo Hao test embankments on Bangkok clay were in Fig. 7. It is further assumed that the ground-water table (GWT)
described in Zou et al. (1995). One of the test embankments is at the ground surface of the embankment. The Fs obtained by
was 100 m long by 30 m wide, with side slopes 1V:2H. The the spreadsheet’s optimization tool is 0.987, or practically 1.00.
embankment was built rapidly to failure using sand fill. Failure The critical slip surface extends to a depth D = 5.26 m in the soft
occurred when the unreinforced embankment reached a height foundation, as shown in the figure.
of 3.4 m. To check the accuracy of the formulations proposed in this
Zou et al. (1995) conducted a finite element analysis to de- paper, the critical slip surface obtained in Fig. 7 can be ana-
termine the mobilized shear stress within the slope and used a lyzed using the Chen and Morgenstern’s (1983) extended
dynamic programming method to search for the critical slip Morgenstern-Price generalized method of slices. The value of

© 1997 NRC Canada


682 Can. Geotech. J. Vol. 34, 1997

Fig. 10. Deterministic analysis of Muar reinforced embankment in Malaysia. Embankment is stable. (Units: m, kN/m, kN⋅m/m, kN/m2, kN/m3.)
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15
For personal use only.

Fs obtained is 0.986 (compared with 0.987 in Fig. 7) if GWT is embankments. The reinforced embankment failed at a height
at the ground surface of the embankment and 1.065 if GWT is at of 6 m; the unreinforced control embankment failed at 4 m.
the base of the embankment. For the latter scenario the present Based on back-analysis of the unreinforced embankment, Bergado
approach obtains a value of Fs = 1.060. Hence, at least for the et al. (1994) suggested a correction factor 0.8 for the field vane
unreinforced embankment in hand, the formulations in this paper shear strength of the soft Bangkok clay.
yield practically the same Fs as the Chen-Morgenstern method. This reinfoced embankment is analyzed here as follows.
It should be mentioned that, because of the approximate nature The total height at failure is H = 8 m, which includes the 2 m
of eq. h (Fig. 3) for the coefficient λ, some inaccuracy in the deep trench excavation (Fig. 9). The 2 m thick in situ weath-
computed Fs is incurred by the present approach if the ratio ered clay forms part of the embankment. The average cor-
D/H is smaller than 0.5. The error in computed Fs is generally rected undrained shear strength for this 2 m weathered clay is
insignificant for D/H ≥ 0.5. cu = (20 + 48)/2 = 34 kPa. The 6 m fill has a cohesion of 10 kPa
The analytical model used in deterministic Fs analysis is and a φ angle of 30°, as reported by Bergado et al. (1994) based
also used to define the limit state surface (Fs – 1 = 0) in reli- on large direct shear tests. In the present analysis, a weighted
ability analysis. Hence the need exists for checking the accu- average of cm = 16 kN/m2 is computed for the two-layer em-
racy of the proposed stability model, as is done in the above bankment material, as described in Fig. 9. Since the friction
and subsequent paragraphs. angle of 30° applies only in the 6 m fill of the 8 m total height,
Figure 8 shows the Fs values and critical slip surfaces, de- by the nature of eq. f in Fig. 3 one should multiply H by 6/8
termined by the spreadsheet’s optimization tool for three hy- and θhR by 6/8, which is the same as multiplying λγ by 0.563.
pothetical reinforcement scenarios of the Nong Ngoo Hao The value of Fs computed using the present approach is 0.99,
embankment. The tangent depth (D) of the critical slip surface which is consistent with failure condition. The actual slip surface,
increases with the magnitude of the T force. as inferred by Bergado et al. (1994), is narrower and shallower
than the one in Fig. 9.
Case 2: Geotextile-reinforced embankment on Bangkok The analysis in Fig. 9 assumes a unit weight of 19.2 kN/m3
clay for the entire 8 m high slope based on a reported dry density
Bergado et al. (1994) described an embankment reinforced of 17 kN/m3 and water content of 13% at the time of failure.
with one layer of nonwoven geotextile of tensile strength Had the lower 2 m in situ weathered clay been assumed to have
200 kN/m. Another unreinforced embankment was constructed a unit weight of 15 kN/m3, thereby giving an average γ of
nearby as a control embankment. A trench 2 m deep and 18.2 kN/m3 for the 8 m high slope, a factor of safety Fs = 1.038
7.5 m wide at the bottom was excavated along the side of both would have been computed.

© 1997 NRC Canada


Low and Tang 683

Fig. 11. Reliability analysis of Muar reinforced embankment in Malaysia.


Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15
For personal use only.

Case 3: Reinforced embankment on Muar clay deposit definition of Fs is used to describe the limit state surface
Chai and Bergado (1993) conducted finite element analysis Fs = 1 because eqs. j and k in Fig. 3 are identical when Fs = 1.
and stability analysis on two geogrid reinforced embankments (2) The mean value of T is the expected maximum mobilizable
on Muar clay deposit in Malaysia. The soft foundation had force of the reinforcement. It is an average strength value,
been installed with vertical drains. Finite element results re- not a mobilized value. This is consistent with the mean
garding geometry change, strength increase of soft soil, and values of cu, cm, and φm used in the reliability analysis,
mobilized reinforcement force were used in stability analysis. which are also strength values. The uncertainties in these
One embankment of thickness 6.07 m (net height about 5.25 m average strength values are reflected in their respective
due to settlement) was found to have a factor of safety 1.22 standard deviations and correlations.
when a mobilized T force of 11 kN/m was used. The increase (3) The uncertainty regarding whether the reinforcement force
over the unreinforced Fs was 0.012. If the peak strength T acts in a horizontal direction (with lever arm z) or acts
(80 kN/m) of the geogrid was fully mobilized, the increase tangential to the slip surface (with lever arm (z + D)) is
(over the unreinforced Fs) was 0.077 if the T force acts in a reflected in the reliability analysis by taking the mean lever
horizontal direction and 0.113 if the T force acts tangential to arm to be an average between z and (z + D), and using a
the slip surface. cofficient of variation (c.o.v.), which in this case is equal
The Muar reinforced embankment is analyzed using the to 0.15. The c.o.v. of 0.15 means that the standard devia-
present approach in Fig. 10. A multilinear cu profile is used tion of lever arm is 0.15 × (mean value of lever arm). At
instead of the published multilayer cu profile. The Fs values the most probable failure point on the limit state surface of
are 1.215, 1.226, and 1.294 for T equal to 0, 11, and 80 kN/m, the random variables, the xvalues of both T and “lever”
respectively. The increase over the unreinforced Fs values are will always be smaller than their respective mean values,
0.011 and 0.079 respectively. If hc = hw = 2 m, an Fs value of if they are uncorrelated or positively correlated.
1.217 is obtained for T = 80 kN/m. Thus, some of the ambiguities and confusions in the deter-
An example reliability analysis for the Muar embankment ministic Fs analysis of reinforced embankments on soft ground
is shown in Fig. 11. A tension crack of mean value 2 m and are either mitigated or overcome in the reliability analysis,
c.o.v. of 0.30 has been used. The following may be noted: apart from the latter’s ability to reflect the uncertainty of the
(1) The same reliability index is obtained regardless of which input parameters. The latter advantage is further illustrated in

© 1997 NRC Canada


684 Can. Geotech. J. Vol. 34, 1997

Fig. 12. Same factor of safety, same covariance, yet different numerically the depth-dependent undrained shear strength
reliability index. values. In one example, the probability of failure based on
reliability index was compared with Monte Carlo simulation;
good agreement was obtained. The reliability index β de-
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15

creases from 2.467 to 1.692 when the coefficient of variation


(c.o.v.) of the undrained shear strength in the soft ground in-
creases from 0.15 to 0.25. The factor of safety based on mean
values does not change. The reliability index is considered to
be a more rational indicator of safety than the conventional
factor of safety because it reflects the degree of uncertainty of
the underlying parameters.
The ease of implementation, the conceptual simplicity, and
the versatility of the ellipsoid approach via spreadsheet may
contribute to reducing the language barrier in geotechnical re-
liability noted by several pioneer researchers. The spreadsheet
method of reliability evaluation can be extended to nonnormal
variates, with or without correlation (Low and Tang 1997).

Acknowledgements
This work was performed while the first author was on sabbatical
leave at the Hong Kong University of Science and Technology,
Fig. 12, which shows two hypothetical cases with the same Fs (1.6) from September to December 1996. The second author would
at their respective mean values of x1 and x2. Both cases also have like to acknowledge the support of a research grant No.
the same standard deviations of x1 and x2. If judged by the determi- HKUST 722/96E sponsored by the Research Grant Council of
nistic factor of safety, they are equally safe. However, a reliability Hong Kong. Graduate student Raymond Cheung of HKUST
For personal use only.

analysis will reveal that case B is less reliable than case A. performed the Monte Carlo simulation mentioned in the paper.
The three case histories mentioned above represent only a
few of the published cases of reinforced embankments on soft References
ground. Additional references can be found in Rowe et al.
(1996), which compares results of finite element analysis with Ang, H.S., and Tang, W.H. 1984. Probability concepts in engineering
planning and design, volume ii—decision, risk, and reliability.
field measurements of a geotextile-reinforced test embankment.
John Wiley & Sons, New York.
Bergado, D.T., Long, P.V., Lee, C.H., Loke, K.H., and Werner, G.
Summary and conclusions 1994. Performance of reinforced embankment on soft Bangkok
clay with high-strength geotextile reinforcement. International
A limit equilibrium stability model of reinforced embankments Journal of Geotextiles and Geomembranes, 13: 403–420.
on soft ground has been presented. The model allows for a tension Chai, J., and Bergado, D.T. 1993. Performance of reinforced embank-
crack in an embankment, tensile reinforcement at the base of ment on Muar clay deposit. Soils and Foundations, 33(4): 1–17.
embankment, and a nonlinear undrained shear strength profile in Chen, Z., and Morgenstern, N.R. 1983. Extensions to the generalized
the soft foundation. An example analysis is given that demon- method of slices for stability analysis. Canadian Geotechnical
strates the use of a widely available spreadsheet optimization tool Journal, 20(1): 104–119.
to facilitate both the deterministic factor of safety evaluation and Chowdhury, R.N., and Xu, D.W. 1995. Geotechnical system reliabil-
the Hasofer-Lind reliability index evaluation. ity of slopes. Reliability Engineering and System Safety, 47:
The proposed reliability evaluation procedure via spreadsheet 141–151.
derives from an alternative interpretation of the Hasofer-Lind in- Christian, J.T., Ladd, C.C., and Baecher, G.B. 1994. Reliability ap-
plied to slope stability analysis. Journal of Geotechnical Engineer-
dex in the original space of the random variables. Correlated ing, ASCE, 120(12): 2180–2207.
random variables are viewed as forming a tilted ellipsoid centered Dai, S.H., and Wang, M.O. 1992. Reliability analysis in engineering
at the mean. The computed reliability index is the same as that applications. Van Nostrand Reinhold, New York.
based on the first order reliability method (FORM) except that Ditlevsen, O. 1981. Uncertainty modeling: with applications to multi-
the spreadsheet automated search is in the original space of the dimensional civil engineering systems. McGraw-Hill, New York.
random variables. Correlated variables are dealt with easily with- Duncan, J.M. 1996. State of the art: limit equilibrium and finite-
out the need for orthogonal transformation of the covariance ma- element analysis of slopes. Journal of Geotechnical Engineering,
trix. The meaning of the reliability index is transparent. The ASCE, 122(7): 577–596.
performance function need not be explicit. Reliability-based de- Haldar, A., and Mahadevan, S. 1995. First-order and second-order
sign can also be performed with ease. reliability methods. In Probabilistic structural mechanics handbook.
Edited by C. (Raj) Sundararajan. Chapman & Hall, New York,
Example and actual reinforced embankments on soft ground pp. 27–52.
have been analyzed using the ellipsoid approach via spreadsheet. Hasofer, A.M., and Lind, N.C. 1974. Exact and invariant second-
The 12 random variables include six autocorrelated cu values that moment code format. Journal of Engineering Mechanics, ASCE,
define the nonlinear undrained shear strength profile of the foun- 100(1): 111–121.
dation soil. The performance function is nonexplicit due to the Hoek, E., and Bray, J. 1981. Rock slope engineering. 3rd ed. Institute
need to search for the critical slip surface and to integrate of Mining and Metallurgy, London.

© 1997 NRC Canada


Low and Tang 685

Hoek, E., Kaiser, P.K., and Bawden, W.F. 1995. Support of underground Morgenstern, N.R. 1995. Managing risk in geotechnical engineering.
excavations in hard rock. A.A. Balkema, Rotterdam, Netherlands. Proceedings, Pan American Conference, International Society
Johnson, R.A., and Wichern, D.W. 1992. Applied multivariate statis- for Soil Mechanics and Foundation Engineering, Guadalajara,
tical analysis. 3rd ed. Prentice Hall, Englewood Cliffs, N.J. Mexico.
Kitch, W.A., Wright, S.G., and Gilbert, R.B. 1995. Probabilistic analysis National Research Council. 1995. Probabilistic methods in geotech-
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by ET- Adama Science & Technology University Library (PERI) on 08/05/15

of reinforced soil slopes. Proceedings, 10th Conference on Engineer- nical engineering. National Academy Press, Washington, D.C.
ing Mechanics, ASCE, New York, Vol. 1, pp. 325–328. Oka, Y., and Wu, T.H. 1990. System reliability of slope stability.
Li, K.S., and Lumb, P. 1987. Probabilistic design of slopes. Canadian Journal of Geotechnical Engineering, ASCE, 116(8): 1185–1189.
Geotechnical Journal, 24: 520–535. Rowe, R.K., Gnanendran, C.T., Landva, A.O., and Valsangkar A.J.
Low, B.K. 1989. Stability analysis of embankments on soft ground. 1996. Calculated and observed behaviour of a reinforced embank-
Journal of Geotechnical Engineering, ASCE, 115(2): 211–227. ment over soft compressible soil. Canadian Geotechnical Journal,
Low, B.K. 1996. Practical probabilistic approach using spreadsheet. 33: 324–338.
Proceedings, Uncertainty in Geologic Environment: From Theory Shinozuka, M. 1983. Basic analysis of structural safety. Journal of
to Practice. American Society of Civil Engineers, Geotechnical Structural Engineering, ASCE, 109(3): 721–740.
Special Publication No. 58, Vol. 2, pp. 1284–1302. Tang, W.H. 1993. Recent developments in geotechnical reliability.
Low, B.K., and Tang, W.H. 1997. Efficient reliability evaluation us- Probabilistic methods in geotechnical engineering. Edited by Li
ing spreadsheet. Journal of Engineering Mechanics, ASCE, and Lo. A.A. Balkema, Rotterdam, pp. 3–27.
123(7). Tichy, M. 1993. Applied methods of structural reliability. Kluwer
Low, B.K., Wong, K.S., Lim, C., and Broms, B.B. 1990. Slip circle Academic, Dordrecht, Boston.
analysis of reinforced embankments on soft ground. International Vanmarcke, E.H. 1977. Reliability of earth slopes. Journal of
Journal of Geotextiles and Geomembranes, 9: 165–181 Geotechnical Engineering, ASCE, 103(11): 1247–1266.
Luckman, P.G., Der-Kiureghian, A., and Sitar, N. 1987. Use of sto- Veneziano, D. 1974. Contributions to second moment reliability. De-
chastic stability analysis for Bayesian back calculation of pore partment of Civil Engineering, Massachusetts Institute of Tech-
pressures acting in a cut slope at failure. Proceedings, ICASP5, the nology, Cambridge, Research Report No. R74–33.
5th International Conference on Applications of Statistics and Whitman, R.V. 1984. Evaluating calculated risk in geotechnical engi-
Probability in Soil and Structural Engineering, Vancouver, B.C., neering. Journal of Geotechnical Engineering, ASCE, 110(2):
Vol. 2, pp. 922–929. 143–188.
Madsen, H.O., Krenk, S., and Lind, N.C. 1986. Methods of structural Zou, J., Williams, D.J., and Xiong, W. 1995. Search for critical slip
safety. Prentice Hall, Englewood Cliffs, N.J. surfaces based on finite element method. Canadian Geotechnical
For personal use only.

Melchers, R.E. 1987. Structural reliability: analysis and prediction. Journal, 32: 233–246.
Ellis Horwood Ltd., Chichester, West Sussex, England.

© 1997 NRC Canada

You might also like