You are on page 1of 9

Engineering Structures 32 (2010) 37253733

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

The projection gradient algorithm with error control for structural reliability
Frdric Duprat , Alain Sellier, Xuan Son Nguyen, Grard Pons
Universit de Toulouse, Laboratoire Matriaux et Durabilit des Constructions, LMDC UPS-INSA, 135 Avenue de Rangueil, 31077 Toulouse Cedex 4, France

article

info

Article history:
Received 22 February 2010
Received in revised form
24 June 2010
Accepted 9 August 2010
Available online 16 September 2010
Keywords:
Reliability
Probabilistic methods
RackwitzFiessler algorithm
Projection gradient
Design point
Error control

abstract
Nowadays, probabilistic approaches are frequently used in the design of new civil engineering structures
and the durability analysis of existing constructions. HasoferLinds reliability index is one of the most
popular reliability measures due to its relevance and ease of use, and is now referred to in many structural
design codes. This index can be determined by several algorithms dealing with minimization under
constraint such as the well known RackwitzFiessler algorithm based on the projection gradient method.
The drawback of this algorithm lies in the estimation of the gradient vector of the limit-state function,
which is often carried out by finite differences for non-explicit functions, resulting from the Finite Element
Method for instance. If the perturbation chosen in the estimation of the gradient vector gives a variation
of the output lower than the numerical accuracy of the limit-state function, the algorithm could lead to
erroneous results or even not converge. In order to circumvent this problem, an original technique is
suggested in this paper called Projection gradient with error control. The principle of the proposed
technique is to attach to RackwitzFiesslers algorithm a procedure for judiciously determining the
perturbation in the calculation of the gradient vector by finite differences, accounting for the numerical
precision of the limit-state function. The efficiency and interest of the proposed procedure is emphasized
through various examples.
2010 Elsevier Ltd. All rights reserved.

1. Introduction
The design of civil engineering structures is based on procedures and models integrating the design constraints such as
loads, material properties and geometry. The uncertainties affecting the design parameters are of a random nature. Probabilistic approaches are consequently suitable tools for assessing structural
risk and reliability as they include a probabilistic description of the
random uncertainties. A common practical measure of the structural reliability is HasoferLinds reliability index denoted [1].
This index is defined in the standardized space of reduced normal
and independent variables as the minimum distance from the origin to a point of the failure surface, the so-called design point P .
The reliability index can be determined by several algorithms
dealing with minimization under constraint such as the RackwitzFiessler algorithm [2]. This algorithm, like those derived from
it to improve its robustness and efficiency, or other gradient-based
algorithms, requires computation of the gradient vector composed
of the partial derivatives of the failure function with respect to the
standardized variables. When the failure function is defined from
a numerical model that only provides approximations of the exact solution to the mechanical problem, the partial derivatives are

Corresponding author. Tel.: +33 561 559 930; fax: +33 561 559 900.
E-mail address: frederic.duprat@insa-toulouse.fr (F. Duprat).

0141-0296/$ see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engstruct.2010.08.017

in turn computed approximately, which can lead to errors or even


non-convergence in some cases. Although fruitful developments
have been achieved in the finite element field for calculating the
sensitivity of model response to design variables, the finite difference method remains a practical and general tool for evaluating the
partial derivatives. The aforementioned problem may arise due to
poor choice of the perturbation applied to variables in this evaluation. In order to overcome this drawback, a simple and efficient
procedure is proposed for choosing the perturbation accounting for
the numerical precision of the model.
2. Determining the reliability index
The determination of HasoferLinds index is a constrained optimization problem. The function to be minimized is the Euclidean
distance u in the standardized space under the constraint G(u) =
0, where G() is the failure function. If u is the solution of the optimization with = OP = u , then
u = arg min(u)G(u)=0 .

(1)

G(u) = 0 defines the failure surface, the frontier between the


safe side with G(u) > 0 and the unsafe side with G(u) < 0.
According to design codes, G() defines a limit-state function, which
is expected to be continuous in the safe side and the unsafe side,
at least in the neighbourhood of the design point P . This explains
the reason why most optimization algorithms used in structural

3726

F. Duprat et al. / Engineering Structures 32 (2010) 37253733

reliability apply to differentiable constraint functions, although


other non-gradient-based algorithms exist that also aim to solve
Eq. (1), such as the simplex algorithm [3], tailored for structural
reliability analysis [4] or, more recently, the bundle algorithm [5].
Rackwitz and Fiessler proposed an iterative method which
solves Eq. (1) by generating a series of points converging towards
the optimal solution [2]. Every new point u(k+1) of the series at
iteration (k + 1) is determined from the previous point u(k) as
follows
u(k+1) =

G(u(k) )
(uT G(u(k) ) G(u(k) ))
G(u(k) )2 (k)

(2)

where G(u(k) ) is the gradient vector of G(u(k) ).


Iterations stop when the expected accuracy p is reached on two
successive coordinates for every axis i

|ui(k+1) ui(k) | P .

(3)

The descent or search direction vector reads:


d(k) =

G(u(k) )
(uT G(u(k) ) G(u(k) )) u(k) .
G(u(k) )2 (k)

(4)

Thus the new iteration point can be expressed as:


u(k+1) = u(k) + (k) d(k)

(5)

where (k) is the step size, constant and equal to 1 in the RackwitzFiessler algorithm. The main improvements made to this algorithm by researches have focused on how to optimize the step
size in order to make the convergence more robust and stable. Liu
et al. [6], Zhang et al. [7] suggested adding a line search scheme in
which the step size is selected to minimize a merit function.
Alternative algorithms to those derived from the gradient projection method exist, such as PolakHes algorithm [8] tailored for
reliability purposes [9] or algorithms based on the Lagrangian expression associated with Eq. (1):
L(u, ) = f (u) + G(u)

(6)

where f (u) is the objective function. The solution that minimizes


L(u) can be found by using a second order Taylor expansion of the
gradient of L(u) together with the optimality conditions L(u) = 0
and G(u) = 0. Finding this root sequentially is the principle of the
Sequential Quadratic Programming method in which an updating
scheme for the Hessian matrix of G(u) is needed. An iterative and
efficient scheme was proposed by Liu and Der Kiureghian [10] with
the objective function f (u) = u2 /2. Abdo and Rackwitz [11] suggested using the augmented Lagrangian
L(u, ) = f (u) + G(u) + cG(u)2 /2

(7)

as a merit function for optimizing the step size in the Rackwitz


Fiessler algorithm, with the objective function f (u) = u2 . This
choice considerably simplifies the updating scheme for the Hessian matrix.
The algorithms shortly described here are globally convergent and are expected to be more robust than the original RackwitzFiessler algorithm. However, a common feature among all the
algorithms is that they use the gradient vector G(u) of the limitstate function to define the search direction, and their efficiency
depends on the quality of the estimate of G(u). As the limit-state
function is defined in the physical space, it is necessary to calculate
partial derivatives in the standardized space from partial derivatives in the physical space as follows:

G(u)
x
T 1 ( u)
= G(u)i = G(x)T
= G(x)T
ui
ui
ui
G(u)

(8)

where u is the partial derivative of the limit-state function with


i
respect to the standardized variable ui , G(x) is the gradient vector

Fig. 1. Problem posed by the minimal perturbation value.

with respect to physical random variables and u = T (x) is the Nataf


transformation.
In practical problems where G(x) is computed from the finite
element method, analytical expressions of the partial derivatives
are rarely available. The direct differentiation method involves development of analytical derivatives of the finite element response
and its implementation as a part of the finite element code. Some
implementations have been carried out for static and dynamic inelastic problems, where response sensitivities were obtained with
respect to material parameters, nodal coordinates, cross-sectional
geometry and nodal loads for a variety of material models and element formulations [7,9,12,13]. However, for complex mechanical
models combining several types of element and problem (ageing
chemicalmechanical behaviour of concrete for example), these
implementations become somewhat cumbersome. For the sake of
simplicity, general purpose reliability programs often resort to the
finite difference method for computing the gradient vector [4,14
17], which remains attractive and easily practicable:

G(x)
G(x + 1xi ei ) G(x)

(9)
xi
1xi
where 1xi is the perturbation and ei is the unit vector for the axis xi .
The truncation error in Eq. (9) results from the neglected terms
in the Taylor series expansion of the perturbed function G(x), and
can be reduced by using a small perturbation 1xi . The perturbation
size should nevertheless be chosen in compliance with the mechanical model precision. The finite element response of a nonlinear mechanical model is twofold: one part is deterministic and
one part is random, resulting from the computational inaccuracies
accepted in the solving procedures (convergence threshold) and
the inevitable numerical round-off errors (machine code precision). It can be reasonably stated that the random part, denoted
1G(x) in Fig. 1, is proportional to the value of the response itself
with a working computational precision imposed by the engineers
knowledge as the proportionality factor. Hence, if the perturbation
1xi is too small, say 1xi < 1xmin in Fig. 1, it may lead to a difference G(x + 1xi ei ) G(x) that is smaller than the random part
1G(x) of the response, and thereafter lead to poor efficiency, even
non-convergence of the search algorithm whatever the underlying
method used for finding the design point.
In order to get a best estimate of a finite difference, it is necessary to impose a judicious value of the perturbation: sufficiently
large to make the random part of the mechanical response as
negligible as possible and small enough to avoid significant truncation error. Furthermore, as iterations of the search algorithm
proceed, the partial derivative for the axis xi should be estimated
using Eq. (9) only if the coordinate of the new iteration point is far
enough from the previous one.
To avoid the pitfalls of calculating the gradient vector response,
the response surface methods can alternatively be employed. By iteratively constructing an explicit approximation of the limit-state
function and combining it with a classical search algorithm for
finding the design point, these methods can give satisfactory results: the approximating function generally behaves very much
like the true limit-state function in the neighbourhood of the design point. Nevertheless, these methods often become dramatically
costly in computation resources for high dimension problems [18].

F. Duprat et al. / Engineering Structures 32 (2010) 37253733


Table 1
Input variables example #1.

3. Projection gradient method with error control


The above mentioned considerations regarding the problem
posed by the minimal perturbation value have been taken into consideration in the improvement to the RackwitzFiessler algorithm
proposed in this paper. The integrated procedure combines the
gradient vector calculated at the previous iteration and the working computational precision of the mechanical model in order to
state the new perturbation value. It is assumed that the limit-state
function of the structure is defined as the difference between a resource variable R(x) and a demand variable S (x):
G(x) = R(x) S (x).

(10)

Alternative expressions are sometimes used, for instance G(x)


= R(x)/S (x) 1. In most engineering fields however it is easier to
handle with limit-state functions or safety margins having a physical meaning and unity.
For the first iteration, as the sensitivity of the mechanical model
to variables is not known yet, a large perturbation is applied:

1xi = c1 (R + S )xi

(11)

where c1 is a magnifying factor (c1 > 1), R and S are working


computational precisions of both models. For the ongoing iteration
(k), 1xi is given by:

(1xi )(k)

1G(x(k1) )

= c2
G(x(k1) )i

R R(x(k1) ) + S S (x(k1) )

= c2

G(x(k1) )i

(12)

where c2 is a magnifying factor (c2 > 1), 1G(x(k1) ) = R R


(x(k1) ) + S S (x(k1) ) is the maximal numerical inaccuracy of the
G(x

(k1)
is the ith compolimit-state function and G(x(k1) )i =
xi
nent of the gradient vector at the previous iteration (k 1).
The working computational precisions are stated according to
the features of the limit-state function. If G() is analytical a usual
choice is R = S = 108 (for double precision floating numbers).
If G() results from an iterative numerical scheme, R and S must be
chosen according to the employed convergence criteria. From the
experience of the authors, relevant values of the magnifying factors
c1 and c2 lie between 5 and 10. The influence of these factors on the
results obtained is emphasized hereafter.
As previously mentioned, the distance between two consecutive points given by the iterative search procedure for finding the
design point should be a criterion for updating the partial derivative for the axis under consideration. Consequently the following
condition applies for updating the component G(x(k) )i :

((xi )(m) (xi )(m1) ) (1xi )(k0 )

m=k +1

3727

(13)

where ((xi )(m) (xi )(m1) ) is the algebraic distance between two
successive points along axis i, for which no update of the component G(x(k) )i has been achieved yet, (k0 ) denotes the iteration
number corresponding to the last update of the partial derivative and (k) is the ongoing iteration (see Fig. 1). (1xi )(k0 ) is the
minimum required perturbation stated by Eq. (12). The condition in Eq. (13) allows the overall computational cost to be reduced by avoiding useless computation of the partial derivatives
for axes where the optimum has already been reached. This is
similar to the use of omission sensitivity factors. The omission
sensitivity factor for a mean value is defined as the ratio of the reliability index determined, with the variable in question taken at
its mean value, to the original reliability index. An omission sensitivity factor very close to one exhibits that the variable could
be omitted in determining the reliability index, say fixed to its

Variable

Distribution

Mean

Stand. dev.

x1
x2
x3
x4

Lognormal
Normal
Normal
Lognormal

10
25
5
0.0625

5
0.8
0.2
0.0625

mean value. Based on approximate omission sensitivity factors,


an implementation of these factors in the search algorithm was
proposed by Madsen [19], requiring the number of variables to
be modified as iterations proceed. In PGEC, the sensitivity criterion given by Eq. (13) is equivalent to the use of omission sensitivity factors. It is finally found that the proposed improvement
of determining the perturbation 1xi as a function of the working
computational precision without systematically updating the partial derivatives in the course of iterations leads to the convergence
of the RackwitzFiessler algorithm being facilitated.
4. Examples
Five examples are presented below in order to illustrate the efficiency of the proposed method and the computational cost benefit
brought by not systematically updating the partial derivatives.
The two first examples deal with explicit limit-state functions
and detailed comparison is made with other frequently employed
algorithms. The results of interest are the value of the reliability
index , the coordinates of design point P , the value of the limitstate function at this point with respect to its value at the median
point (space origin) |G(u )/G(u0 )|, which is expected to be zero if
P lies exactly on the failure surface, and the number of runs of the
limit-state function Nr . Nr reflects the efficiency of the methods
with respect to the computational cost. The proposed technique is
subsequently referred to as the PGEC method (projection gradient
with error control).
The three last examples concern implicit failure functions:
some comparisons are made with results from the literature when
available.
4.1. Example #1
The limit-state function is a polynomial used as a response surface for reliability analysis of a pipeline [6]:
G(x) = 1.1 0.00534x1 0.0705x2 0.226x3 + 0.998x4

0.00115x1 x2 0.0149x1 x3 + 0.0717x1 x4


+ 0.0135x2 x3 0.0611x2 x4 0.558x3 x4
+ 0.00117x21 + 0.00157x22 + 0.0333x23 1.339x24 .

(14)

The following algorithms are under consideration: the LiuDer


Kiureghian algorithm (SQP) and the ZhangDer Kiureghian algorithm (iHLRF), implemented in OpenSees [10,11], the Abdo
Rackwitz algorithm (RFLS), the RackwitzFiessler algorithm (HLRF),
and the NelderMead algorithm (SPLX), implemented in Comrel [13].
The non-correlated random variables used in this example are
defined in Table 1. Convergence criteria in PGEC were p = 5
103 and R = S = 108 . Convergence criteria are also stated in
Comrel. In OpenSees the convergence criteria are:

|G(u )/G(u0 )| 1

(15)

checking for the proximity of P to the failure surface, and

u T u 2

(16)

checking that the gradient vector at P points towards the space

3728

F. Duprat et al. / Engineering Structures 32 (2010) 37253733

Table 2
Reliability index and convergence criteria example #1.
Technique

Nr

|G(u )/G(u0 )|

u T u

max |ui ui(k1) |

SQP
iHLRF
RFLS
HLRF
SPLX
PGEC (c1 = c2 = 1)
PGEC (c1 = c2 = 5)
PGEC (c1 = c2 = 104 )

1.5296
1.5296
1.5296
1.5296
1.5294
1.5296
1.5296
1.5298

57
49
25
25
55
25
23
20

4.24 106
2.40 105
2.10 107
2.10 107
1.14 106
2.10 107
1.84 107
1.44 106

8.71 104
6.95 104
n.a.
n.a.
n.a.
6.40 107
6.51 107
1.67 106

n.a.
n.a.
<5 103
<5 103
<5 103
1.70 103
1.71 103
7.37 104

Table 3
Design point example #1.
Technique
SQP
iHLRF
RFLS
HLRF
SPLX
PGEC (c1 = c2 = 5)

u1

u2

u3

1.499
1.499
1.496
1.496
1.496
1.496

1.152 10
1.142 101
1.147 101
1.147 101
1.187 101
1.148 101
1

u4

8.040 10
7.775 102
7.957 102
7.957 102
7.601 102
7.976 102
2

2.675 101
2.706 101
2.845 101
2.845 101
2.849 101
2.845 101

Table 4
Input variables example #2.
Variable

Description

Distribution

Mean value

Stand. dev.

x1
x2
x3
x4
x5
x6
x7

Bending moment (MN m)


Effective depth (m)
Steel yielding strength (MPa)
Reinforcing steel area (m2 )
Block factor
Width (m)
Concrete strength (MPa)

Normal
Normal
Lognormal
Normal
Normal
Normal
Lognormal

0.015
0.30
550
308 106
0.5
0.20
38

0.00525
0.015
55
30.8 106
0.05
0.01
5

Table 5
Reliability index and convergence criteria example #2.
Technique

Nr

|G(u )/G(u0 )|

u T u

max |ui ui(k1) |

SQP
iHLRF
RFLS
HLRF
PGEC (c1 = c2 = 1)
PGEC (c1 = c2 = 10)
PGEC (c1 = c2 = 5000)

4.2829
4.2829
4.2173
4.2174
4.2176
4.2176
4.2177

55
50
40
40
40
34
33

1.31 105
1.60 105
7.79 107
7.57 107
1.60 108
1.66 108
1.08 107

2.15 104
5.53 104
n.a.
n.a.
1.60 107
1.60 107
1.53 107

n.a.
n.a.
<5 103
<5 103
1.62 103
1.62 103
5.00 104

origin. Default values for these criteria are 1 = 2 = 102 . The


results are reported in Tables 2 and 3.
As it can be seen in Table 2, the influence of the magnifying
factors c1 and c2 remains limited in the case of an analytical limitstate function. Since the assessment of the gradient vector is fairly
accurate in such case, the sole expected effect of the magnifying
factors is globally to save runs of the limit-state function, that is
obtained for c1 = c2 5.
The reliability index results are almost the same for all the
techniques employed. As far as computational cost is concerned,
the algorithms that appear the most relevant are those in which the
optimization of the step size in the descent direction is made with
minimum effort, or even not made at all, say RFLS, HLRF and PGEC.
Five iterations are needed in these algorithms for convergence to
a design point very close to the failure surface to be achieved. By
saving two calculations of partial derivatives (with c1 = c2 =
5), PGEC is the least costly algorithm. It can be noted that the
convergence criterion used in PGEC (Eq. (3)) is as, or more, severe
than those in iHLRF and SQP (Eqs. (15) and (16)). For the latter,
requiring supplementary runs of gradient evaluation to optimize
the step size of the search direction, it is worth noting that the
computational cost is highly dependent on the perturbation value
used in finite difference for estimating the partial derivatives. In

OpenSees, the perturbation value is proportional to the standard


deviation of random variables: the results presented above were
obtained with proportionality factors of 1/1000 for SQP and
1/1500 for iHLRF. With the default value 1/1000 for iHLRF, the
number of computations jumps to Nr = 70 for identical reliability
results.
4.2. Example #2
The limit-state function is the safety margin of a reinforced
concrete section subjected to a bending moment:
G(x) = x2 x3 x4

x5 x23 x24
x6 x7

x1 .

(17)

The random variables are listed in Table 4. The same convergence criteria and magnifying factors were imposed here as in the
previous example. The results are reported in Tables 5 and 6, exhibiting again the influence of the magnifying factors c1 and c2 . As
already noticed this effect is the reduction of the computational
cost while maintaining the same accuracy of results.
As far as reliability outputs are concerned, values of the reliability index are slightly different when the SQP and iHLRF algorithms
are compared with the RFLS, HLRF and PGEC algorithms. For the

F. Duprat et al. / Engineering Structures 32 (2010) 37253733

3729

Table 6
Design point example #2.
Technique
SQP
iHLRF
RFLS
HLRF
PGEC (c1 = c2 = 10)

u1

u2

u3

3.074
3.076
3.016
3.013
3.011

9.854 10
9.855 101
9.560 101
9.535 101
9.526 101

1.610
1.609
1.718
1.716
1.715

Table 7
Computation of partial derivatives example #2.

u5

u6

2.307
2.305
2.194
2.201
2.206

4.939 10
4.937 102
4.774 102
4.710 102
4.665 102
2

u7

2.484 10
2.484 102
2.391 102
2.368 102
2.356 102
2

6.501 102
6.498 102
6.256 102
6.201 102
6.198 102

Table 9
Frame element properties example #3.

Table 8
Omission sensitivity factors example #2.
Variable

x1

x2

x3

x4

x5

x6

x7

i (i )

6.5159
1.545

4.3241
1.025

4.6407
1.100

4.8411
1.148

4.2179
1.000

4.2177
1.000

4.2190
1.000

Fig. 2. Structural system example #3 (Unit: m).

latter, the closeness of the design point to the failure surface and
the computational cost are more satisfactory. It is finally found that
the PGEC algorithm is the least costly (with c1 = c2 = 10): by
saving six computations of the partial derivatives with respect to
variables x5 x7 at iterations 4 and 5 (see Table 7) according to the
criterion given by Eq. (13), PGEC allows useless runs of the limitstate function to be avoided. The omission sensitivity factors have
been computed for this example. Results for these are reported in
Table 8. It can be seen that variables x5 x7 are the least influential.
4.3. Example #3
The case of a multi-storey multi-span metallic portal frame is
dealt with (see Fig. 2). The limit-state function is the margin with
respect to a prescribed value max of the displacement at the top
of the structure:
G(x1 , . . . , x21 ) = max (x1 , . . . , x21 ).

u4

(18)

The x-vector contains 21 basic variables translating different


properties of structural components. The structural data and the
statistical parameters are reported in Tables 9 and 10. Some
variables are assumed to be correlated. All loads are correlated by
a coefficient of correlation F = 0.95. All cross section properties
are correlated by AiAj = IiIj = AiIj = 0.13. The two different
modulus of elasticity E1 and E2 are correlated by E = 0.9. All other

Element

Modulus of elasticity

Moment of inertia

Cross section

1
2
3
4
5
6
7
8

E1
E1
E1
E1
E2
E2
E2
E2

I5
I6
I7
I8
I1
I2
I3
I4

A5
A6
A7
A8
A1
A2
A3
A4

variables are assumed to be uncorrelated. The internal language


of the CAST3M free finite element software [20] was used in this
study, as for the following examples, for both finite element and
probabilistic calculations. Convergence criteria in PGEC were p =
5 102 , R = 108 and S = 104 . The results are reported in
Table 11 for several values of max and c1 = c2 = 10.
It is worth noting that the convergence criteria are fulfilled satisfactorily, whatever the value of the reliability index is, which is
increased with larger value of max . In Table 11 Ns is the number
of saved runs of the finite element model resulting from the criterion given by Eq. (13). As seen in Table 12, this criterion allows
the gradient to be updated only when required as the computation goes along. The absence of update of a partial derivative is
hence not definitive in the course of the search procedure. This interesting feature gives the PGEC algorithm some adaptability in the
search procedure for finding the design point. The study of the effect of the magnifying factors c1 and c2 reveals in Table 13 that the
range [510] for these ones leads to a fairly good compromise between the precision and the saving of runs.
4.4. Example #4
In this example, the risk of depassivation of the reinforcements
embedded in concrete is addressed. Reinforcing bars are depassivated when pH decreases due to the ingress of pollutants such
as chlorides or carbon dioxide. Once depassivated, steel is prone
to corrosion if sufficient moisture and oxygen are available in the
neighbourhood of reinforcing bars. The case of a concrete beam exposed to carbonation is considered here. As cracks due to the applied load facilitate the ingress of carbon dioxide, the mechanical
behaviour is taken into account together with the diffusion process
and the chemical aspects. The overall modelling is detailed in [21].
The limit-state function for the durability limit-state is defined as
follows:
G(x) = x1 d(x2 , x3 , x4 , x5 ).

(19)

The carbonation depth is given by the function d() which is


computed from a finite element model. The steep slope profile
of the calcium content in concrete leads to strong non-linearity
in the function d(). In this model, a fine mesh in both space
and time was used in order to ensure satisfactory convergence of
computation procedures. The carbonation depth profiles obtained
for several exposure times can be seen in Fig. 3. In function d(), the
carbonation depth is given along the line AB.
The random variables are listed in Table 14. Convergence
criteria in PGEC were p = 5 102 , R = 107 and S = 104 .

3730

F. Duprat et al. / Engineering Structures 32 (2010) 37253733

Table 10
Input variables example #3.
Variable

Unit

Distribution

Mean value

Standard deviation

F1
F2
F3
E1
E2
I1
I2
I3
I4
I5
I6
I7
I8
A1
A2
A3
A4
A5
A6
A7
A8

kN
kN
kN
kN/m2
kN/m2
m4
m4
m4
m4
m4
m4
m4
m4
m2
m2
m2
m2
m2
m2
m2
m2

Gumbel max
Gumbel max
Gumbel max
Normal
Normal
Positive normal
Positive normal
Positive normal
Positive normal
Positive normal
Positive normal
Positive normal
Positive normal
Positive normal
Positive normal
Positive normal
Positive normal
Positive normal
Positive normal
Positive normal
Positive normal

133.454
88.97
71.175
2.173752 107
2.379636 107
0.813443 102
1.150936 102
2.137452 102
2.596095 102
1.081076 102
1.410545 102
2.327853 102
2.596095 102
0.312564
0.3721
0.50606
0.55815
0.253028
0.29116825
0.37303
0.4186

40.04
35.59
28.47
1.9152 106
1.9152 106
1.08344 103
1.298048 103
2.59609 103
3.028778 103
2.596095 103
3.46146 103
5.624873 103
6.490238 103
0.055815
0.07442
0.093025
0.11163
0.093025
0.10232275
0.1209325
0.195375

Table 11
Reliability index and convergence criteria example #3.

max (cm)

3
6
9

Nr

1.342
3.398
4.573

52
91
119

Ns

|G(u )/G(u0 )|

u T u

max |ui ui(k1) |

14
63
79

4.29 10
1.72 105
1.74 104

1.69 10
2.00 105
1.55 104

3.63 102
2.58 102
4.93 102

Table 12
Computation of partial derivatives example #3 (max = 6 cm).

Table 13
Influence of c1 and c2 factors example #3 (max = 9 cm).
c1 , c2

Nr

Ns

|G(u )/G(u0 )|

u T u

max |ui ui(k1) |

1
2
4
6
8
10
12
50

4.573
4.573
4.573
4.573
4.573
4.573
4.573
4.558

178
156
139
135
123
119
121
140

42
64
81
85
75
79
99
168

1.91 106
6.95 106
2.89 105
2.80 105
3.27 105
1.74 104
7.43 104
9.17 104

3.59 105
3.96 105
2.79 105
2.93 105
1.24 104
1.55 104
5.07 104
6.59 104

2.81 102
2.73 102
2.24 102
2.33 102
4.87 102
4.93 102
3.92 102
1.54 102

Table 14
Input variables example #4.
Variable

Description

Distribution

Mean value

Stand. dev.

x1
x2
x3
x4
x5

Concrete cover (mm)


Coefficient of diffusion (m2 /s)
Tortuousity
Concrete strength (MPa)
Live load (kN/m2 )

Lognormal
Lognormal
Uniform (0.10.9)
Lognormal
Gumbel max

20
1 108
0.5
35
1.04

4
0.8 108
0.23
5
0.4

The results for an exposure time of 5 years are reported in Tables 15


and 16.

The reliability indices and coordinates of the point P supplied


by PGEC and by the response surface method used in [20] are

F. Duprat et al. / Engineering Structures 32 (2010) 37253733

(a) 1 month.

(b) 5 years.

(c) 20 years.

3731

(d) 35 years.

Fig. 3. Carbonation fields at various times (half cross-section at mid-span).


Table 15
Reliability index and convergence criteria example #4.
Technique
RSM [20]
PGEC
HLRFa
a

3.998
3.923

Nr

|G(u )/G(u0 )|

u T u

max |ui ui (k 1)|

43
26

3.17 10
4.48 104

8.70 10
5.53 104

9.62 103
1.54 102

Not converged.

Table 16
Design point example #4.
Technique
RSM [20]
PGEC

u1

u2

3.601
3.076

1.312 10
7.814 102

u3
1

similar even though some differences exist. These latter probably


result from the approximation of the true failure function given
by the response surface, which might not be sufficiently accurate
in the vicinity of point P . In contrast, it is not surprising that
even satisfactory convergence results with PGEC are less tight than
with RSM. This is due to the fact that the approximate failure
function handled by RSM is explicit as are the gradients as well.
Nevertheless, PGEC exhibits an advantage regarding the numerical
cost: 10 runs of the mechanical model are saved as shown in
Table 17.
Like PGEC, the original HLRF algorithm was implemented in
the CAST3M software. The perturbation value for evaluating the
gradient vector in the HLRF algorithm was set at 1xi = i /1000,
where i is the standard deviation of the variable xi . It can be
seen in Table 18 that convergence is not reached after 6 iterations
(not even beyond); the value of reliability index is not stabilized
and the value of the limit state function is far from zero. This
arises because the perturbation value does not suit the numerical
inaccuracy of limit state function adequately, and is below the
optimum value. The non-convergence is not caused by an irregular
or discontinuous profile of the failure surface.
4.5. Example #5
The case of a wide prestressed footbridge is considered in this
example. The longitudinal cross-section and cross-section over the
support can be seen in Fig. 4. The structure was designed according to Eurocode 2 [22]. The prestressing force of 23 MN was determined in such a way that only the minimum area of reinforcing
steel was required: under the characteristic combination of loads
and for the upper and lower values of the prestressing force, the
tensile stress in concrete is limited to the average tensile strength.

9.212 10
2.547 102
4

u4

u5

1.731
2.026

1.601 102
4.837 102

Table 17
Computation of partial derivatives example #4.

Table 18
Comparison between PGEC and HLRF example #4.
Iteration

PGEC

|G(u )/G(u0 )|

3.298
0.375

5.210
0.220

3.951
0.026

3.984
0.016

3.926
0.0006

3.923
0.0004

HLRF

|G(u )/G(u0 )|

3.741
1.032

3.852
0.316

1.893
0.126

1.608
1.011

3.687
0.695

4.901
0.548

The structural durability is affected by the concrete cracking and


the requirement for durability is therefore that the effective stress
does not exceed the effective tensile strength, taking into account
the concrete damage. The limit state function is hence
G(x1 , . . . , x9 ) =

x1
1 dtp (x2 , . . . , x9 )

M (x2 , . . . , x9 )
1 dM (x2 , . . . , x9 )

(20)

where dtp () is the concrete damage on reaching the maximum


strength (variable x1 ), dM () is the effective concrete damage, and
M () is the normal stress, both computed at point M in Fig. 4.

3732

F. Duprat et al. / Engineering Structures 32 (2010) 37253733

Table 19
Input variables example #5.
Variable

Description

Distribution

Mean value

Stand. dev.

x1
x2
x3
x4
x5
x6
x7
x8
x9

Tensile strength (MPa)


Modulus of elasticity (MPa)
Modulus of compressibility (MPa)
Characteristic strain for the consolidation of spherical viscosity
Viscosity for the reversible creep (MPa/d)
Moisture pressure (MPa)
Drying creep stress (MPa)
Chemical factor
Pedestrian live load (kN/m2 )

Lognormal
Lognormal
Lognormal
Lognormal
Lognormal
Lognormal
Lognormal
Lognormal
Gumbel max

3
37 600
20 8890
104
62 668
58
11.7
3.1 104
2.437

0.6
9400
41 778
3.2 105
12 534
14.21
3.1
9.748 104
0.9748

Table 20
Reliability index and convergence criteria example #5.
Exposure time (y)

Nr

Ns

|G(u )/G(u0 )|

u T u

max |ui ui (k 1)|

10
50
90

3.261
2.582
1.897

55
38
40

15
2
10

1.99 105
6.90 104
9.01 104

4.22 104
6.45 103
1.50 103

8.60 103
3.48 102
3.01 102

Fig. 4. Prestressed footbridge studied (units: cm). Main design inputs: the concrete strength class is C40/50, the tensile strength of strands is 1860 MPa, and the pedestrian
live load is 3.71 kN/m2 .

Creep and shrinkage of the concrete are factors that reduce


the prestressing force, and have been included in the functions
M (), dtp () and dM () which are actually outputs of a global finite
element model, geometrically reduced to a quarter of the structure
(see Fig. 5). This global model comprises several comprehensive
coupled sub-models for creep, shrinkage and mechanical damage
of the concrete, accounting for the main parameters that influence
creep and shrinkage: the ambient humidity, the dimensions of
the element, the composition of the concrete, the maturity of
the concrete when the load is first applied and the duration and
magnitude of the loading. The mathematical formulation of the
sub-models derives from the damage theory for a visco-elastic
material submitted to consolidation. The modelling is detailed
in [23].
The random variables are listed in Table 18. Very few statistical
data exist in the literature for the creep and shrinkage of concrete:
most of the surveys are incomplete. Data are hence missing for
the parameters of the model, which are expected to be uncertain
in a broad range. From the assumption that creep and shrinkage
functions given in EC2 present an uncertainty of 20%30% [22],
and from available collected experimental data, propositions have
still been derived in Table 19. From a physical point of view, and
owing to possible numerical issues, lognormal distributions were
suggested.
Convergence criteria in PGEC were p = 5 102 , R = 105
and S = 104 . The results obtained for several exposure times are
reported in Table 20.

Fig. 5. Stresses just after tensioning.

It can be seen that the number of runs of the limit state function
remains reasonably low. The convergence criterion, Eq. (3), on
the distance between two successive search points, which is the
only one active in the procedure, allows the other convergence
criteria, Eqs. (15) and (16), to be satisfactorily fulfilled. The update
for estimating the coordinates of the gradient vector, as stated by
Eq. (13), again leads to a saving of runs of the limit state function
for this example, as shown in Table 21.
5. Conclusion
In this paper, an improvement of the RackwitzFiessler algorithm has been proposed for cases where the model associated

F. Duprat et al. / Engineering Structures 32 (2010) 37253733


Table 21
Computation of partial derivatives example #5 (10 year exposure).

3733

Acknowledgements
The CEA (french Atomic Energy Commission) is thanked for the
provision of the Cast3M software to the LMDC in its development
version for education and research.
References

with the limit-state function presents some random computational


inaccuracies. These are part of the outputs supplied by the Finite
Element Method when resorting to non-linear solving procedures.
If the computational inaccuracies are not accounted for in estimating the coordinates of the gradient vector by finite differences, the
iterative search procedure for finding the design point may converge erroneously or not at all. In order to circumvent this drawback, a dedicated procedure for calculating the perturbation to be
used in finite differences was achieved and attached to the RackwitzFiessler algorithm. Simultaneously, the procedure takes into
account the working computational precision assigned by the engineers knowledge and depending on the finite element modelling, and allows the number of runs to be reduced thanks a
criterion for the updating of the partial derivatives with respect
to the input variables. The overall algorithm, called the PGEC
algorithm, was compared to enhanced algorithms commonly implemented in dedicated software. From examples considered with
explicit failure functions, PGEC exhibits satisfactory behaviour: for
similar convergence constraints, the results obtained are very close
to those supplied by enhanced algorithms in terms of the reliability index and the coordinates of the design point. Furthermore,
PGEC is the least costly algorithm numerically speaking. Some examples with strongly non-linear finite element modelling were
also addressed, dealing with durability issues of concrete structures exposed to carbonation, or subjected to creep and shrinkage
effects. Beyond efficiently fulfilling the convergence criteria, the
PGEC algorithm allows a significant number of runs to be saved.
The proposed technique could also give rise to improvements in
the other search algorithms derived from the projection gradient
or gradient-based and resorting to finite differences for estimating
the gradient vector.

[1] Hasofer AM, Lind NC. An exact and invariant second moment code format.
J Eng Mech 1974;100:11121.
[2] Rackwitz R, Fiessler B. Structural reliability under combined random load
sequences. Comput Struct 1979;9:48994.
[3] Nelder JA, Mead R. A simplex method for function minimization. Comput J
1965;7:30813.
[4] COMREL & SYSREL users manual. RCP Consult. Munich; 19872004.
[5] Bonnans JF, Gilbert JC, Lemarechal C, Sagastizabal C. Optimisation numrique:
aspects thoriques et pratiques. In: Collection Mathmatiques et Applications.
Springer; 1999.
[6] Liu P-L, Der Kiureghian A. Optimization algorithms for structural reliability.
Struct Saf 1991;9:16178.
[7] Zhang Y, Der Kiureghian A. Finite element reliability methods for inelastic
structures. Report No. UCB/SEMM-97/05. Berkeley (CA): Department of Civil
and Environmental Engineering, University of California; 1997.
[8] Polak E. Optimization algorithms and consistent approximations. Applied
mathematical sciences, vol. 124. Springer Verlag; 1997.
[9] Haukaas T, Der Kiureghian A. Strategies for finding the design point in nonlinear finite element. Struct Saf 1991;9:16178; Probab Eng Mech 2006;21:
13347.
[10] Liu PL, Der Kiureghian A. Finite element reliability of geometrically non-linear
uncertain structures. J Eng Mech 1991;17:180625.
[11] Abdo T, Rackwitz R. A new -point algorithm for large time-invariant and
time-variant reliability problems. In: Der Kiureghian A, Thoft-Christensen P,
editors. Proc. of the 3rd IFIP WG 7.5; 1991.
[12] Haukaas T, Der Kiureghian A, Fujimura K. Structural reliability software at the
University of California, Berkeley. Struct Saf 2006;28:4467.
[13] Reh S, Beley JD, Mukherjee S, Khor EH. Probabilistic finite element analysis
using ANSYS. Struct Saf 2006;28:1743.
[14] Schuller GI, Pradlwarter HJ. Computational stochastic structural analysis
(COSSAN) a software tool. Struct Saf 2006;28:6882.
[15] Gollwitzer S, Kirchgner B, Fischer R, Rackwitz R. PERMAS-RA/STRUREL
system of programs for probabilistic reliability analysis. Struct Saf 2006;28:
10829.
[16] Lemaire M, Pendola M. PHIMECA-SOFT. Struct Saf 2006;28:13049.
[17] Tvedt L. Proban probabilistic analysis. Struct Saf 2006;28:15063.
[18] Nguyen XS, Sellier A, Duprat F, Pons G. Adaptive response surface method
based on a double weighted regression technique. Probab Eng Mech, available
online 27 April 2008.
[19] Madsen HO. Omission sensitivity factors. Struct Saf 1988;5:3545.
[20] CAST3M. Finite element software. Commissariat lEnergie Atomique.
http://www-cast3m.cea.fr.
[21] Duprat F, Sellier A. Probabilistic approach to corrosion risk due to carbonation
via an adaptive response surface method. Probab Eng Mech 2006;21:20716.
[22] Eurocode 2: EN 1992-2. ICS: 91.080.40; 93.040. Design of concrete structures,
part 2: concrete bridges.
[23] Sellier A, Buffo-Lacarrire L. Vers une modlisation simple et unifie du fluage
propre, du retrait et du fluage en dessiccation du bton. Eur J Environ Civil Eng
2009;13:116182.

You might also like