Probabilistic Engineering Mechanics 28 (2012) 164168

Contents lists available at SciVerse ScienceDirect

Probabilistic Engineering Mechanics

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

Reliability analysis of large structural systems

A. Naess a, , B.J. Leira b , O. Batsevych c
a

Department of Mathematical Sciences & Centre for Ships and Ocean Structures, Norwegian University of Science and Technology, NO-7491 Trondheim, Norway

Department of Marine Technology, Norwegian University of Science and Technology, NO-7491 Trondheim, Norway

Kongsberg Maritime, Horten, Norway

article

info

Article history:
Received 10 June 2011
Received in revised form
14 July 2011
Accepted 17 August 2011
Available online 7 September 2011
Keywords:
System reliability
Large scale system
Monte Carlo simulation
Failure probability
Statistical estimation

abstract
Brute force Monte Carlo simulation methods can, in principle, be used to calculate accurately the reliability
of complicated structural systems, but the computational burden may be prohibitive. A new Monte
Carlo based method for estimating system reliability that aims at reducing the computational cost is
therefore proposed. It exploits the regularity of tail probabilities to set up an approximation procedure
for the prediction of the far tail failure probabilities based on the estimates of the failure probabilities
obtained by Monte Carlo simulation at more moderate levels. In this paper, the usefulness and accuracy
of the estimation method is illustrated by application to a particular example of a structure with several
thousand potentially critical limit state functions. The effect of varying the correlation of the load
components is also investigated.
2011 Elsevier Ltd. All rights reserved.

1. Introduction
In general, it is very difficult to calculate the reliability of
realistic structural systems by using conventional theoretical
reliability techniques. This is usually caused by the high number
of basic random variables needed for modeling the uncertain
elements of the problem and the large number of safety margins
that are used to describe the system, which will generally
consist of a series system of parallel subsystems. At least, in
principle, the reliability of complicated structural systems can
be accurately predicted in a straightforward manner by standard
Monte Carlo simulation methods. However, the computational
burden may be prohibitive for highly reliable systems. Motivated
by this observation, the authors have initiated the development
of simulation based methods for calculating the reliability of
structural systems that aims at reducing the computational cost.
This is achieved by introducing a cascade of systems depending
on a parameter varying between zero and one, where the original
system is obtained when this parameter equals one. When the
parameter value is zero, the system is highly prone to failure, and
for the small to intermediate values of the parameter, the failure
probability can be estimated with high accuracy by Monte Carlo
simulation with moderate computational efforts. By exploiting the

Corresponding author.
E-mail addresses: arvidn@math.ntnu.no (A. Naess), bernt.leira@ntnu.no
(B.J. Leira), oleksandr.batsevych@kongsberg.com (O. Batsevych).
0266-8920/$ see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.probengmech.2011.08.024

regularity of the failure probability as a function of the parameter,

an approximation procedure for predicting the failure probability
at parameter value one has been proposed in [1]. This method
was applied to small structural systems in [1,2] and it was shown
that good results could be quickly obtained on a standard laptop
computer.
In this paper, the specific case of a 3D beam structure (grillage)
will be studied. It consists of 40 transverse and 40 longitudinal
beams, creating 40 40 equidistant and intersecting main beams.
As discussed in the section on numerical examples, the number
of basic random variables in the model will be 4880 and the
number of limit state functions will be 6540. At the present stage
of development, this system would represent a huge challenge
for any standard reliability method for system analysis. It will be
shown that the proposed method can handle also this system with
feasible computational efforts.
2. Component reliability
The basic element in the construction of the cascade of systems
mentioned in the Introduction is already seen for the case of
component reliability. Let us consider a safety margin M which
represents the failure mode under consideration. It is assumed that
M is expressed in terms of n basic random variables X1 , . . . , Xn as
M = G(X1 , . . . , Xn ), where the limit state function G can be rather
complicated. To calculate the failure probability pf = Prob(M
0), the method proposed in [1] is based on introducing a cascade
of safety margins M () = M M (1 ), where M = E[M ]

A. Naess et al. / Probabilistic Engineering Mechanics 28 (2012) 164168

and the parameter [0, 1]. Putting pf () = Prob(M () 0),

then pf = pf (1). It is seen that E[M (0)] = 0, which means that the
corresponding component is highly prone to failure when = 0.
Hence, the failure probability pf () will decrease from a high value
at = 0 to the small target value at = 1. As proposed in [1], we
shall now make the following assumption about the behavior of
the failure probability pf () of the safety margin M ():
pf () q() exp a( b)c ,

(1)

where the function q() is slowly varying compared with the

exponential function exp{a( b)c }.
The practical importance of this relation, if it applies, is that the
target failure probability pf = pf (1) can be obtained from values
of pf () for < 1. Our focus, in this paper, is on methods for
estimating pf by Monte Carlo simulation. The observation above
may then be significant, as it is usually easier to estimate pf ()
for < 1 more accurately than the target value since they are
larger. Fitting the parametric form for pf () to the estimated
values would then allow us to provide an estimate of the target
value by extrapolation. The viability of this proposal will be
demonstrated by a numerical example below. However, there
could be exceptional cases, e.g., for variables which are uniformly
distributed on a finite interval, where application of a functional
form like Eq. (1) might not be accurate for all probability levels.
It is believed that such discrepancies can be identified at an early
stage of the probabilistic analysis.
3. System reliability

165

For a sample of size N of the vector of basic random variables

X = (X1 , . . . , Xn ), let Nf () denote the number of samples in the
failure domain of M (). The estimate of the failure probability is
then
p f () =

Nf ()

(5)

N
The coefficient of variation of this estimator is

CV (pf ()) =

1 pf ()
pf ()N

(6)

A fair approximation of the 95% confidence

interval

for the value

pf () can be obtained as CI0.95 () = C (), C + () , where
C () = p f () 1 1.96CV (pf ()) .

(7)

The problem of finding the optimal values of the parameters

q, a, b, c can carried out by optimizing the fit on the log level by
minimizing the following mean square error function [1],
F (q, a, b, c ) =

2
wj log p f (j ) log q + a(j b)c ,

where 0 1 < < M < 1 denotes the set of values where

the failure probability is empirically estimated. wj , j = 1, . . . , M,
denote weight factors that put more emphasis on the more reliable
data points, alleviating the heteroscedasticity of the estimation
problem at hand. The choice of weight factors is to some extent

arbitrary. In this paper, we use wj = log C + (j ) log C (j )

with = 1 and 2, combined with a LevenbergMarquardt least
squares optimization method [6]. This usually works well provided
reasonable, initial values for the parameters are chosen. In this
paper, = 2 has been used. Note that the form of wj puts some
restriction on the use of the data. Usually, there is a level j
beyond which wj is no longer defined. Hence, the summation in
Eq. (8) has to stop before that happens. Also, the data should be
preconditioned by establishing the tail marker 0 in a sensible way.
Although the LevenbergMarquardt method, as described
above, generally works well, it may be simplified by exploiting
the structure of F . It is realized by scrutinizing Eq. (8) that if b
and c are fixed, the optimization problem reduces to a standard
weighted linear regression problem. That is, with both b and c
fixed, the optimal values of a and log q are found using closed form
weighted linear regression formulas in terms of wj , yj = log p f (j )
and xj = (j b)c .
It is obtained that the optimal values of a and q are given by the
relations,

We shall consider structural systems that can be described

as combinations of component failure modes. Each failure mode
is assumed to be represented by a safety margin. Let Mj =
Gj (X1 , . . . , Xn ), j = 1, . . . , m, be a set of m given safety margins
expressed in terms of the n basic random variables that are
involved in specifying the structural system we consider. The
cascade of systems we shall use in our method for estimating the
failure probability is then obtained from the class of safety margins
defined as Mj () = Mj j (1 ), where j = E[Mj ].
The basic systems are series systems and parallel systems. If the
system at hand is a series system, the failure probability will be
given as,

pf () = Prob
{Mj () 0} .

(2)

j =1

In general, any system can be written as a series system of

parallel subsystems [35]. The failure probability would then be
given as,

pf () = Prob
{Mi () 0}

(3)
a (b, c ) =

j=1 iCj

where Cj {1, . . . , m}, j = 1, . . . , l, denote the index sets

defining the parallel subsystems.
To use the proposed method for failure probability estimation,
we then make the assumption that pf () can be represented as in
Eq. (1) also for the system reliability problem.

The method to be described in this section is based on the

assumption expressed by Eq. (1). As argued in [1], for practical
applications it is implemented in the following form,
pf () q exp a( b)c ,

(4)

for 0 1 for a suitable value of 0 , where q is now assumed

to be a constant. An important part of the method is therefore to
identify a suitable 0 so that the rhs of Eq. (4) represents a good
approximation of pf () for [0 , 1].

wj (xj x)(yj y)

j=1
M

(9)

wj (xj x)2

j =1

and
log q (b, c ) = y + a (b, c )x,

(10)

where x =
j=1 wj xj /
j=1 wj , y =
j=1 wj yj /
j=1 wj .
The LevenbergMarquardt method may now be used on the
function F (b, c ) = F (q (b, c ), a (b, c ), b, c ) to find the optimal
values b and c , and then the corresponding a and q can be
calculated from Eqs. (9) and (10).
For estimation of the confidence interval for a predicted value of
the failure probability provided by the optimal curve, the empirical
confidence band is reanchored to the optimal curve. The range of
fitted curves that stay within the reanchored confidence band will
determine an optimized confidence interval of the predicted value.
This is obtained by constrained nonlinear optimization. As a final

4. Reliability estimation by optimized fitting

(8)

j=1

166

A. Naess et al. / Probabilistic Engineering Mechanics 28 (2012) 164168

point, it has been observed that the predicted value is not very
sensitive to the choice of 0 provided it is chosen with some care.
5. Numerical example
A numerical example with a large number of limit state
functions is considered. Each of the limit state functions have a
quite simple form, and they are explicitly defined in terms of the
basic random variables. However, since the number of such limit
state functions is very large, the computation time becomes an
issue. If the proposed method was to be used in combination with
computationally demanding procedures involving, e.g., a large
Finite Element Model for calculation of the failure probability,
effective sampling strategies would need to be selected. Presently,
the number of samples for each value of the parameter is based
on direct sampling without any refinements such as importance
sampling.
The example structure is a 3D beam system (grillage) with a grid
of 40 40 equidistant and intersecting main beams as shown in
the figure below. Each of the main beams are clamped at both ends.
The 3D structure is modeled by means of beam elements, and it
is subjected to a vertical load at each of the intersections (i.e., at the
nodes of the corresponding finite element model). The beams have
a rectangular cross-section of dimension 0.05 m times 0.12 m, and
the length of each beam is 0.5 m. The mean yield stress is equal to
380 MPa, with a standard deviation of 19 MPa (i.e., a coefficient of
variation equal to 0.05). A lognormal distribution is assumed.
All the vertical load components are Gaussian distributed
with a mean value equal to 420 N and a standard deviation of
126 N, corresponding to a coefficient of variation equal to 30%.
Furthermore, all the load components are pairwise equi-correlated
(i.e., they have pairwise the same correlation coefficient). The effect
of varying this correlation coefficient is also investigated below. For
the system model in the present example there are 4880 random
variables and 6540 safety margins.
At each node of the finite element model there are three
degrees-of-freedom, i.e., one vertical displacement and two
rotations (around two orthogonal horizontal axes). All the degreesof-freedom are assembled in the system displacement vector r.
Correspondingly, all the load components are collected in the load
vector R, which has the same dimension as the displacement
vector. Every third entry of this load vector is non-zero and
corresponds to the vertical load components. The intermediate
components are equal to zero (which correspond to the two
external bending moments at each node). For a given realization
of the random load vector components, the corresponding
displacement vector of the system is obtained by inverting the
system stiffness matrix K:
r = K1 R.

(11)

For each beam element, the corresponding internal reaction

forces at the two end nodes, which are collected in a sixdimensional element vector S, can be obtained by extracting the
proper element displacements from the system vector and collecting them in an element-specific vector v. These are subsequently
multiplied by the element stiffness matrix, ki :
Si = ki vi ,

(12)

where the index i refers to the particular element number which is

being considered.
For each element, there will now be one limit state function
at each of the two nodes. This corresponds to the requirement
that the sum of the normalized bending stress, the normalized
vertical shear stress and the normalized torsion shear stress
must be smaller than 1.0 in order for the structure to be safe.
The normalization of each stress component is based on the

Fig. 1. Horizontal projection of 3D beam system (grillage) with 40 40 main

members.

corresponding stress value at initial yielding. All the limit state

functions will then have the following form:

B V T
+
+
,
B,cr V ,cr T ,cr

gj X = 1

(13)

where the index j runs from 1 to the total number of limit state
functions (i.e., two times the number of beam elements). The vector
of random variables, X, contains the computed bending stress at
each cross section, B , the corresponding vertical shear stress, V ,
the torsion shear stress, T , and the yield stress, Yield . The three
critical stresses (i.e., B,cr , V ,cr , T ,cr ) are all proportional to the
yield stress (since initial yielding is considered). They are hence
fully correlated for a given cross-section.
Examples of the bending moment diagrams (i.e., two bending
moments B1 and B2) in the 3D beam system for the case of fully
correlated loads are shown in Fig. 2. Plots of the collection of limit
state function values for one of the simulations (corresponding to a
load correlation length of 1.5 m, see below) are shown in Fig. 3. The
upper part of the figure shows the values for the second node of all
the transverse elements. The lower part shows the corresponding
values for all the longitudinal elements. It is seen that no failures
are observed for this particular sample. Furthermore, the minimum
values of the limit state functions (which correspond to maximum
values of the utilization) occur along the clamped boundaries for
both of the orthogonal directions. This reflects that the bending
stresses (i.e., which are obtained from the absolute values of the
bending moments in Fig. 2) dominate the three terms of the limit
state function in Eq. (13).
Simulations are first performed for a load correlation length
of 1.5 m. The yield stress correlation length is 0.1 m for all the
examples which are considered. (This is equal to 1/5 of the length
of each beam.) The results which correspond to a number of
samples which is equal to 4000 50 = 200 000 are shown in
Fig. 4. For each series of 4000 samples, the failure probability pf ()
is calculated for each value of . On the basis of all 50 values of
pf (), the mean value and standard deviation is calculated. From
this can be calculated the final estimate of the failure probability
pf () and the 95% confidence interval for this estimate for each
value of . On the basis of these results, it is obtained that the
predicted failure probability which corresponds to the scaling level
= 1 is obtained as pf (1) = 6.21 1012 . The corresponding
95% confidence interval is obtained as (2.61 1012 , 1.51 1011 ).
The parameters of the fitted curve are, respectively, a = 10.37, b =
0.417, c = 2.81 and d = 1.79.

A. Naess et al. / Probabilistic Engineering Mechanics 28 (2012) 164168

(a) Bending moment around x-axis.

167

(b) Bending moment around y-axis.

Fig. 2. Bending moments around two orthogonal axes for the 3D beam system (lower figure is rotated by 90 compared to the upper one).

(a) Limit state function values for Node 2 of transverse beam elements in
Fig. 1.

(b) Limit state function values for Node 2 of longitudinal beam elements in
Fig. 1.

Fig. 3. Examples of limit state function values for selected cross-sections of 3D beam system. Load correlation length is 1.5 m.

Fig. 4. Plot of the probability of failure pf () versus for load correlation length
1.5 m: Monte Carlo (); fitted optimal curve ( ); reanchored empirical confidence
band ( ); fitted confidence band ( ).

Fig. 5. Plot of the probability of failure pf () versus for load correlation length
2 m: Monte Carlo (); fitted optimal curve ( ); reanchored empirical confidence
band ( ); fitted confidence band ( ).

The effect of increasing the correlation length to 2.0 m is next

investigated. The results are shown in Fig. 5 for the same sample
size of 200 000. The predicted target failure probability at = 1.0

is pf (1) = 3.73 107 , while the confidence interval extends from

2.40 107 to 5.72 107 . This is roughly an increase by an order
of 105 as compared to a correlation length of 1.5 m. This implies

168

A. Naess et al. / Probabilistic Engineering Mechanics 28 (2012) 164168

probability for = 1 being significantly higher for the present

value of the load correlation length.
The computational burden which is involved in using the
proposed method on the specific example studied in this paper
would typically range between half an hour and one hour on a
standard laptop computer (for a sample size of 200 000).
6. Concluding remarks

Fig. 6. Plot of the probability of failure pf () versus for load correlation length
2.5 m: Monte Carlo (); fitted optimal curve ( ); reanchored empirical confidence
band ( ); fitted confidence band ( ).

that the sensitivity with respect to the correlation length is very

strong. The parameters of the fitted curve are now, respectively,
a = 18.27, b = 0.168, c = 1.42 and d = 0.718. By increasing
the sample size to 2 000 000, the confidence interval becomes
4.91 107 to 6.0 107 which is significantly more narrow. The
corresponding target failure probability is now predicted to be
pf (1) = 5.42 107 .
By increasing the correlation length to 2.5 m, the target failure
probability is predicted to be pf (1) = 1.29 105 , with a predicted
95% confidence interval of (9.23 106 , 1.69 105 ) for a sample size
of 200 000. This is an increase of the failure probability by roughly
a factor of around 40 for the lower bound and a factor of around
30 for the upper bound as compared to a correlation length of 2.0.
The resulting sample and the fitted curve are shown in Fig. 6. The
parameters of this curve are now a = 13.34, b = 0.142, c = 1.42
and d = 0.524.
It is seen that for the present case, the number of samples could
have been significantly reduced since the degree of extrapolation
is much less than for the previous cases. This is due to the failure

In this paper, we have described a Monte Carlo-based

method for estimating the reliability of structural systems. It
has been shown that the method may provide good estimates
for the reliability of large structural systems with a moderate
computational effort.
It has been pointed out that the use of Monte Carlo methods for
system reliability analysis has several very attractive features, the
most important being that the failure criterion is usually relatively
easy to check almost irrespective of the complexity of the system
and the number of basic random variables.
Acknowledgments
The financial support from the Research Council of Norway
(NFR) through the Centre for Ships and Ocean Structures (CeSOS) at
the Norwegian University of Science and Technology is gratefully
acknowledged. The third author was financially supported by the
Department of Mathematical Sciences at NTNU.
References
[1] Naess A, Leira BJ, Batsevych O. System reliability analysis by enhanced Monte
Carlo simulation. Structural Safety 2009;31:34955.
[2] Naess A, Leira BJ, Batsevych O. Estimation of system reliability by Monte
Carlo simulation. In: Proceedings 28th international conference on offshore
mechanics and arctic engineering. OMAE-2009, New York: ASME; 2009.
Paper no. 79623.
[3] Madsen HO, Krenk S, Lind NC. Methods of structural safety. Englewood Cliffs:
Prentice-Hall Inc.; 1986.
[4] Ditlevsen O, Madsen HO. Structural reliability methods. Chichester (UK): John
Wiley & Sons, Inc.; 1996.
[5] Melchers RE. Structural reliability analysis and prediction. second ed. New York:
John Wiley & Sons, Inc.; 1999.
[6] Gill P, Murray W, Wright MH. Practical optimization. London: Academic Press;
1981.