You are on page 1of 9

The stochastic finite element method in structural

reliability

Armen Der Kiureghian and Jyh-Bin Ke

Department of Engineering, University of California, Berkeley, CA 94720, USA

First-order reliability and finite element methods are used to develop a methodology for reliability
analysis of structures with stochastically varying properties and subjected to random loads. Two
methods for discretization of random fields are examined and the influence of the correlation
length of random property or load fields on the reliability of example structures are investigated.
It is found that the correlation length of load fields has significant influence on the reliability
against displacement or stress limit states. The correlation length of property fields is significant
for displacement limit states, but may not be significant for stress limit states. Examples studied
include a fixed ended beam with stochastic rigidity and a plate with stochastic elasticity.

INTRODUCTION advantages of that formulation in the context of the


simulation approach.
The term stochastic or probabilistic finite element (SFE or
The preceding works are all concerned with second-
PFE) is used to refer to a finite element method which
moment analysis of the response, i.e., computing means,
accounts for uncertainties in the geometry or material
properties of a structure, as well as the applied loads. Such variances, and covariances. Such a characterization of the
uncertainties are usually spatially distributed over the structural response provides useful insight into the
manner structural uncertainties propagate, but it does
region of the structure and should be modelled as random
or stochastic fields. The interest in this area has grown not provide adequate information for reliability analysis.
from the perception that in some structures the response Since most civil engineering structures are designed with
is strongly sensitive to the material properties or high levels of reliability, one is interested in the tails of the
geometry of the structure, and that even small probability distributions, which are poorly defined by the
uncertainties in these characteristics can adversely affect first and second moments. SFE methods that aim at
the structure reliability. computing the reliability (or the failure probability) have
Applications of SFE or similar discrete methods of been proposed by Veneziano et al. 2s and by Der
Kiureghian et al. 6"7"11. The former uses first-order
analysis appear to have been initiated in the early 1970's.
Early applications used the Monte Carlo simulation reliability methods in conjunction with response surface
method to investigate the effect of uncertain variability in fitting, experimental design, and simulation techniques to
structural properties z. Later, first-order Taylor series estimate the probability distribution of a response
expansion or perturbation methods were used to quantity. The SFE method for reliability analysis
compute second-moment statistics of response developed by the authors is based on well established,
quantities in structural or geotechnical applications a-5,12. first-order reliability methods ~,21 and is, at least
This method was extended to second-order by Hisada conceptually, applicable to any problem where the failure
and Nakagiri, who investigated a variety of structural criterion is specified in terms of a vector of random
problems13 16,23. A similar approach has been used variables.
recently by Liu, Belytschko and Mani with applications In this paper, the first-order SFE reliability approach
to elasto-plastic and nonlinear dynamics problems zl. for linear structures under static loads is presented
Vanmarcke and Grigoriu used essentially the same together with two example applications. The examples
approach and demonstrated the use of spatial averaging demonstrate the effects of spatial variability in material
of random fields in SFE with an application to a beam properties or loads on the reliability of structures.
with randomly fluctuating rigidity 27. Recently, Applications dealing with stochastic dynamics problems
Lawrence 19 has used basis random variables to describe are presented elsewhere ~s. Extension of the methodology
random fields and a Galerkin approach to develop a to nonlinear structural problems is currently in progress.
second-moment SFE procedure. Also, Shinozuka and
N o m o t o 25 and Shinozuka and Dasgupta 24 have used
Neumann expansion in conjunction with Monte Carlo THE FIRST-ORDER RELIABILITY METHOD
simulation techniques and have demonstrated Structural reliability problems of interest here are based
on two fundamental assumptions: (1) the state of the
Taken from the Proceedings,US-Austria Joint Seminar on Stochastic structure is defined in the outcome space of a vector of
Structural Mecahnics, Florida Atlantic University, Boca Raton, random variables; (2) the structure can be in one of two
Florida, May 1987 states, the safe state or the failure state. The boundary

© 1988 Computational Mechanics Publications Probabilistic Engineering Mechanics, 1988, Vol. 3, No. 2 83
Stochastic finite element method: A. Der Kiureghian and J-B. Ke
between the two states in the outcome space is known as exists for continuous variables and is specified in terms of
the limit-state surface. the distribution of X (Refs 17 and 10). The mapping of the
Failure criteria in mechanics are normally defined in limit-state surface onto the standard normal space is
terms of load effects, such as stresses and deformations. symbolically described by
Let S denote a vector of such effects and g(s) = 0 denote
g(s) = g(s(x(y))) = G(y)= 0 (6)
the limit-state surface in the outcome space of S. Without
loss of generality, the function g(. ) is formulated such that The probability of failure, then, is
9(s) > 0 defines the safe state and g(s) ~<0 defines the failure
state. The probability of failure of the structure, then, is p¢= ~ (p(y)dy (7)

pf= ( fs(s)ds (1)


g(s) ~ 0
where ~b(y) denotes the standard normal density of Y.
The reasons for the probability transformation are two
where fs(s)ds denotes the joint probability density useful properties of the standard normal space. One is
function (PDF) of S. As a simple example of such a that the probability density in that space is rotationally
formulation, consider the failure of a structure defined as symmetric about the origin (the mean point) and decays
yielding at a critical point in accordance with the Von exponentially with the square of the distance from the
Mises criterion. The limit-state function, for a plane stress origin. As a result, most of the contribution to p f in the
condition, then is integral in equation (7) comes from the neighbourhood of
the point(s) on the limit-state surface which is(are) nearest
g(S) = S~ - (S~x + S~y- SxxSyy + 3S~y) (2)
to the origin. This is particularly true for civil engineering
where Sx,,, Syy,and Sxy are the components of stress at the structures for which the limit-state surface is far from the
point and S o is the yield stress, which may all be random origin. This property suggests that good approximations
variables due to randomness in loads and/or in material to the integral in equation (7) can be constructed by
properties. This formulation of the reliability is not replacing the integration boundary (i.e., the limit-state
limited to static problems. For example, for the first- surface) with an approximating boundary which has a
passage failure of an uncertain structure under close fit at the point(s) nearest to the origin. The second
(stochastic) dynamic loads, the limit-state function may useful property of the standard normal space is that the
be described by probability contents for some simple subsets in arbitrary
dimension are available. These two properties are the
9(s)=So-S~ (3) bases for the first and second-order reliability methods.
where So is the threshold (possibly a random variable) In the first-order reliability method, the limit-state
and Sin, a random variable, is the maximum of the surface in the standard normal space is replaced with the
stochastic response process over the duration of the tangent plane at the point with minimum distance from
excitation. Applications with this formulation are the origin. The first-order estimate of the probability of
reported in Ref. 18. failure, then, is
A fundamental difficulty in reliability analysis is that pjl = o ( - f l ) (8)
the statistics of load effects (i.e., the PDF fs(s)) are not
directly available. This is because measurements and where fl, denoted reliability index, is the minimum
observations are usually made on basic quantities, such as distance from the origin and (I)(.) denotes the standard
loads and material property constants, rather than on normal cumulative probability. In the second-order
load effects which are derived quantities. This, in fact, is method, the limit-state surface in the standard normal
the chief reason for the need for finite element methods in space is fitted with a second-order surface, usually a
reliability. paraboloid 9, and in situations where the limit-state
Let the vector X denote the set of all basic random surface has several local minimum-distance points,
variables pertaining to a structure and assume the joint multiple fittings with plane or paraboloid surfaces may be
PDF fx(x) is known. Included in this vector are variables used to improve the reliability approximation. In the
defining loads, material properties, member sizes, etc. For present paper, only first-order approximations with
spatially varying quantities, discretization in space is single or multiple fittings are employed. Extension to
employed, as described later in this paper. In deference second-order approximation is relatively straight-
to S, we denote the outcome space of X as the load space. forward, but may require considerable additional
These two vectors are related through the mechanical computation (see Ref. 9).
transformation The main effort in the first or second-order reliability
method is in finding the minimum-distance point(s),
S=S(X) (4) denoted y* (x* in X space, s* in S space). This is
For all but trivial structures, this transformation is formulated as a constrained optimization problem:
available only in an algorithmic sense, e.g., a finite-
element code. minimize ly[ (9)
For reliability analysis, it is convenient to transform the subject to G(y)=0
variables X into the standard normal space:
Standard solution algorithms for this problem are
Y=Y(X) (5) available 1°. Starting from an initial point Y0, these
where the elements of Y are statistically independent and algorithms typically generate a sequence of points Yl, Y2,
have the standard normal density. Such a transformation, . . . . which converge at a minimum-distance point,
denoted herein as the probability transformation, always provided the algorithm is stable. Experience shows that

84 Probabilistic Engineering Mechanics, 1988, Vol. 3, No. 2


Stochastic finite element method: A. Der Kiureghian and J-B. Ke
for structural engineering applications convergence is (8) are given by
achieved often in less than ten and seldom in more than
fifty steps. Most efficient algorithms utilize the gradient Vap/, = - ~b(fl)Vpfl (15)
vector of the limit-state function, where q~(. ) is the standard normal PDF. These sensitivity
measures are easy to compute since they do not involve
[-t3G(y) t3G(y)] the mechanical transformation. It is obvious that these
VyG=LT, ' 3 measures can be extremely useful in design and analysis of
structures.
to compute the sequence. By the chain rule of
differentiation, this vector is FINITE E L E M E N T I M P L E M E N T A T I O N
Vy a : VsgJs.xJ .,y (10) Implementation of the first-order reliability method in a
where Vsg is the gradient of the limit-state function with finite element code requires two main algorithmic
respect to the load effect variables, J,,, is the Jacobian of features. The first is an ability to compute the load effects
the mechanical transformation, and Jx,y = Jy~x1, where Jy,x for the sequence of points selected in the standard normal
is the Jacobian of the probability transformation. The space for solution of the optimization problem, equation
first vector is easily computed (either analytically or (9). Each point requires a single conventional run of the
numerically) for any limit-state function which is finite element code with the load space variables
expressed in terms of the load effects. The Jacobian of the determined from the inverse of the probability
probability transformation and its inverse are also easy to transformation, equation (5). (As stated earlier, this
computer, as the transformations only depends on the transformation is usually in a triangular form, which
probability distribution of X and usually can be expressed facilitates its inversion.) The second is a capability to
in a triangular form 1°'17. A major difficulty, however, lies effÉciently compute the Jacobian of the mechanical
in computing the Jacobian of the mechanical transformation, Js,x. Since the size of X is usually large, a
transformation, Js,,. The elements of this matrix are the straightforward finite-difference approach is obviously
partial derivatives of load effects with respect to load- impractical, as it would require repeated runs of the code
space variables. Such quantities are not computed in for each element of the Jacobian.
ordinary finite element codes and their formulation and For linear elastic structures under static loads, the
implementation in an efficient manner is the key to the conventional finite-element solutions for the nodal
SFE reliability method described here. displacements, u, strains, e, and stresses, a, are expressed
by
u=K-1P e=Bu a=De (16)
MEASURES OF SENSITIVITY
where K is the stiffness matrix, P is the nodal load vector,
An important feature of the first-order reliability method B is the displacement-to-strain transformation matrix,
is the facility to compute reliability sensitivity measures and D is the elasticity matrix. It is easily verified that the
with respect to any set of desired parameters. The simplest Jacobians of u, ~, and o with respect to the load-space
such measure 1S the unit vector ~t=-VyG/]VyG I variables are 11
computed at the linearization point y*, which represents
Ju,x = K- 1(Je,x- JK,xU) (17)
the sensitivity of fl with respect to variations in the
linearization point L22, i.e., L,x=BS.,x (18)
Vy.fl=~ (11) J,,x = DIKlu,x + Ja,xBu (19)
The sensitivity vector with respect to the linearization where Jp,x and JK.~ are the Jacobians of the load vector
point in the load space is given by 7 and the stiffness matrix with respect to x, respectively, and
it is assumed that the finite element mesh and shape
Vx.fl = ctJy,x (12) functions, and therefore the matrix B, remain fixed. Due
where the Jacobian is computed at the linearization point. to the linearity of the structure, these Jacobians can be
Furthermore, let 0 and ~/ denote sets of parameters computed by assembling partial derivative load vectors
defining the probability distribution of X and the limit- and stiffness matrices of all elements. It is important to
state function, respectively, i.e., fx(X,0) and g(S,t/). (0 note that this procedure requires no additional inversions
might include such parameters as means, standard Of the stiffness matrix for computing the Jacobians.
deviations, correlation coefficients, etc., and ~/ might Efficient algorithms for solution of these Jacobians are
include such parameters as allowable stresses or given, for example, in Ref. 29.
deformations, deterministic parameters defining The above procedure is coded in a FORTRAN
geometry, etc.) It can be s h o w n 22 that sensitivities of fl program with truss, beam, three-node constant strain,
with respect to these parameters are given by: and four-node quadrilateral elements. The program
includes routines for computing the partial derivative
Vofl=~Jy,o (13) stiffness matrix and load vector for each element, which
1 are then assembled using the conventional direct stiffness
V.fl = ~ Vng(s*, F/) (14) approach to compute the global partial derivative
IVUI matrices. The variables X considered are magnitudes of
nodal or distributed loads, material property variables
where, again, the derivatives are computed at the such as elastic moduli and Poisson's ratio, and geometric
linearization point. Finally, for any set of parameters p, characteristics such as thickness or moment of inertia. In
the sensitivities of the first-order probability in equation storage allocation and multiplication of matrices,

Probabilistic Engineering Mechanics, 1988, Vol. 3, No. 2 85


Stochastic finite element method: A. Der Kiureghian and J-B. Ke

advantage is taken of the fact that the partial derivative Then, for each random field, a separate mesh is
matrices are highly sparsed. considered which is equal to or coarser than the finite
element mesh, such that each random field element is a
block of one or more finite elements. This is done so as to
RANDOM FIELD DISCRETIZATION reduce the number of random variables, and to avoid
Structural properties or loads varying randomly in space near perfect correlation between variables which causes
are modelled here as multi-dimensional random numerical instability in the probability transformation.
processes, or random fields. For finite element The effect of the random field mesh size is examined in the
implementation, it is necessary to discretize such fields following example.
into random vector representations. Two methods of
discretization are investigated here. In one method, the EXAMPLE 1 - BEAM W I T H STOCHASTIC
field value over an element is represented by its value at RIGIDITY
the midpoint of the element. In the second method, the
value for an element is represented by the spatial average The example is a 32-feet long, fixed ended beam with
of the process over the element, as originally suggested by stochastic flexural rididity, El(x), which is subject to a
Vanmarcke et al. 27. For a two-dimensional process distributed load with stochastic intensity W(x), as shown
X(x,y), the above representations for an element i are in Fig. 1. Both processes are assumed to be homogeneous
and Gaussian with the means and coefficients of variation
X i= X (xi, yi) (20) listed in Table 1. The autocorrelation coefficient functions
are assumed to be of the form
and
; IAxl (22)
Xi=~ X(x,y)dAi (21)
i

respectively, where xi, yi are the coordinates of the point


in the element that transforms onto the centroid of the
,ww(Ax)=exp(_l xl) (231
• \ awL/
square element in the normalized coordinate system, and
Ai is the area. The moments (i.e., mean vector and where Ax is the distance between two points, L = 32 ft is
covariance matrix) of X~ are obtained in terms of the the beam span, and ae~ and aw are dimensionless
mean and covariance functions of the process from well measures of the correlation lengths. Note that aE~ = ~ or
known results in random field theory z6. In particular, the aw = ~ reduces the corresponding process to a random
variance of X~ defined by the midpoint discretization variable.
method (equation (20)) is the same as the point variance The beam is modelled with 32 finite elements, but the
of the process, whereas the variance of Xi defined by the number of random field elements for each process is
spatial averaging method (equation (21)) is the point varied depending on its correlation length. Two limit-
variance multiplied by the 'variance reduction' factor state criteria are considered: the exceedance of the
which depends on the area A~. In general, the midpoint midspan deflection above the threshold of 0.6 in, denoted
discretization method tends to over-represent the displacement lmit state, and the exceedance of the left-end
variability within the element, whereas the spatial bending moment above the threshold of l l00k-ft,
averaging method tends to under-represent the same denoted moment limit state.
variability. Therefore, the two methods tend to bracket Fig. 2 shows the convergence of fl as the number of
the exact result for any given mesh, which provides a random field elements is increased for the two limit-state
useful means for checking the convergence of the solution functions and for both the midpoint (dashed curves) and
with the mesh size. the spatial average (solid curves) discretization methods.
If the random field is Gaussian, the joint distribution of The two methods coincide when the random field mesh is
Xi for various elements remains Gaussian for both the sufficiently fine in relation to the correlation length. For a
midpoint and spatial averaging methods. However, for coarse random field mesh, the spatial averaging method
non-Gaussian fields the distribution of X~ defined by
equation (21) is difficult or impossible to obtain. An
approximate description of the distribution for that case
has been suggested elsewhere 8. Here, attention is
restricted to Gaussian random fields so that a fair
comparison between the two methods of discretization X

can be made.
An important issue in SFE is the selection of the mesh
size. Two separate factors should be considered for this
selection. One is the expected gradient of the stress field,
and the other is the expected rate of fluctuation of each Fig. 1. Example 1 - beam with stochastic rigidity
random field, as measured, for example, by the
corresponding correlation length. The two requirements Table 1. Statistics ]br example 1
do not necessarily coincide in each region of the structure
and, hence, it is usually necessary to consider two or more Variable Mean Coefficient of v a r i a t i o n
separate meshes. In the present application, the finite
W, k/ft 8.0 0.30
element mesh is selected such that both sets of EI, k-ft 2 1.125 × l 0 s 0.20
requirements in each region of the structure are satisfied.

86 Probabilistic Engineering Mechanics, 1988, Vol. 3, No. 2


Stochastic finite element method: A. Der Kiureghian and J-B. Ke
4.2 averaging methods give identical results. For the
aEt = 0.25, a w = oo
displacement limit state, the correlation length of each
/ random field has significant influence on the reliability -
3.B / the reliability decreases with increasing correlation length
/ (i.e., as the random field approaches a random variable).
I ..-- aEt = 1.00,a w = This implies that random fluctuations in the rigidity of the
3.0 /
aw=0.25, aE/=oo
beam or the distributed load tend to decrease the midspan
air = 1.00,act = oo
deflection. The same result is observed for the bending
moment limit state, except that there appears to be a lack
E.4
Displacement L i m i t State of sensitivity to fluctuations in the rigidity.
Reliability sensitivities with respect to the local mean
I I I I I I I |
and standard deviation of each random field element for
i.B
the case with aE1= aw = 0.25 are shown in Fig. 4. These
4.2- measures are scaled by the corresponding standard
Moment L i m i t S t a t e
deviations (i.e., ael or aw of the process) to make them
a w = 0.25, aEI = oo dimensionless. This scaling also has the effect of making
3.B the parameter variations equally likely in a statistical
r
sense. For the displacement limit state, the reliability is
l
I most sensitive to the means and standard deviations of
3.0 l random rigidity elements in the central region as well as
# l the two ends of the beam. Sensitivity to the means and
I standard deviations of random load elements is relatively
E.4 i/~~g. = 1.00, aEi = oO
insignificant and is only visible in the central region of the
/ aEl = 1.00, aw = oo aEi = 0.25, a W, = o o
beam. For the bending moment limit state (for the left end
. . . . . J of the beam), the reliability is most sensitive to the means
t.B and standard deviations of random rigidity elements at
'* i '2 t '8 2'0 2'4 2'0 a'2
the two ends of the beam and to the means and standard
NUMBER OF RANDOM FIELD ELEMENTS
deviations of random load elements in the left-central
Fig. 2. Convergence of [3 with number of random field region of the beam. If sensitivities with respect to the
elements global mean or standard deviation of a random field
(rather than individual element) are desired, they are
provides a better accuracy than the midpoint method. As simply obtained by algebraicly summing the sensitivity
the mesh becomes finer, the two methods tend to bracket values for the individual elements.
the exact result. As indicated before, this is because the The above results are all in concurrence with our
spatial averaging method generally under-represents the intuitive expectations and merely quantify what we
variability within an element, whereas the midpoint already know or can guess. However, for more complex
method tends to over-represent the same variability.
Thus, the two methods provide a useful means for 5.0
checking the convergence of fl with the random field mesh ~ = a , allDisplacement Limit State
size. From the results in Fig. 2, it is apparent that =all.=(/

convergence in fl is effectively achieved when the random 4.0


"=CC
field element size is one-quarter to one-half of the
correlation length parameter, aE~ or aw. Similar results
have been obtained by Hisada et al. 15 for a one- 3.0
dimensional beam, where they numerically investigated # \
the dimensional invariance of fl for the analysis of
2.0
continuum.
As noted earlier, when the correlation length of a
random field is large, a fine random field mesh causes l.O I I I I
numerical instability, as the correlation matrix of the
discretized field tends to be near singular and the inverse 5.0
Moment L i m i t S t a t e
of the probability transformations in equation (5)
becomes unstable. For example, in Fig. 2, the cases with
aE~= 1 or aw = 1 could not be analysed with 16 or more
random field elements. Of course, in these situations a fine
random field mesh is not necessary, and this merely
reiterates the need for a separate mesh for each random #
4.0

3.0
~ 1 =all' =a

field. Such separate meshes are also helpful in reducing


the number of random variables representing each aEi a = ~ °°
2.0
random field. ~aEi =a, all, = c¢
Fig. 3 demonstrates the influence of the correlation
lengths aE~ and aw on the reliability with respect to the 1.0.0
two limit states. The number of random field elements • t7o t:5
used in each case depends on the correlation length, but is CORRELATION LENGTH. a
sufficiently large such that the midpoint and spatial Fig. 3. Influence of correlation length on reliability

Probabilistic Engineering Mechanics, 1988, Vol. 3, No. 2 87


Stochastic finite element method." A. Der Kiureghian and J-B. Ke

0.8 Because of the stochastic properties of the plate,


Displocement Limit State
symmetry cannot exist in a strict sense. Nevertheless, for
0.4 simplicity, stochastic symmetry is assumed and only one
quarter of the plate is analysed. The quarter plate is
at~ 0.0 J modelled with 30 elements in the region of material 1 and
9 elements in the region of material 2. The random field
-0.4 and finite element meshes are assumed to be identical.
t
The limit-state criterion for the plate is selected to be the
iL t exceedance of the principal tensile stress at element 32
-O.B t °'El
r above the threshold of 30 ksi.
(o) -. As expected, the reliability of the plate with respect to
-1.2 I I I I I I
the stress criterion is strongly influenced by the difference
between the properties of the two materials. Fig. 6a shows
0.2
Moment Limit Stote the variation of fl with the ratios of the means of the bulk
and shear moduli, r=#r,/~K2=#c,/l~o~, when the
0.1 correlation length a is infinite (i.e., when K I and Gt are
random variables) and the cross correlations p K , K =
po,G = 0 . The reliability is seen to decrease with
~N 0.0 -. ... _. . . . . .-" . . . . . increasing difference between the mean properties of the
two materials. Scaled sensitivities of fl with respect to the
means and standard deviations of the random variables
are shown in Figs 6b and 6c, respectively. The reliability is
(b) most sensitive to the mean and standard deviation of the
-0.2 I I J I I l load intensity, W. Among the property variables,
0 5 I0 15 20 25 30
HORIZONTAL COORDINATE, ft sensitivities to the mean values of K~, G~, and K2 and
standard deviations of Ka and K2 appear to be significant
Fig. 4. Reliability sensitivity measures with respect to for certain values of r.
element means and standard deviations, for aE1= a w ---- 0.25 At around r = 1, a sharp change in the sensitivities of fl
to the material property variables is observed in Figs 6b
structures, the reliability index and the associated and 6c. The reason for this can be inferred from the data in
sensitivity measures provide valuable insight into the Table 3, where the mean values and linearization points
behaviour of the structure and the significance of various for the cases r = 1 and r = 1.25 are listed. In the present
uncertainties, thus providing a resourceful means for
rational decision making in the design or analysis stage.

Y
EXAMPLE 2 - PLATE WITH STOCHASTIC
ELASTICITY
W
The example structure is one quarter of a 1 in thick, 32 ×
32 in / square plate with distributed edge loads, as shown
in Fig. 5. The plate is made of two materials: a 4-inch
diameter core material, and an outside material. The bulk
and shear moduli K1 and G1, respectively, of the outside
material are modelled as homogeneous Gaussian random
fields. The moduli K2 and G2 of the core material and the
load intensity W are considered to be Gaussian random
variables. The assumed means and coefficients of Material 1:
variation are listed in Table 2. The shear and bulk moduli KI I G I /
for each material are assumed to be statistically Elemen 32
independent, but correlation between the same moduli X
for the two materials, i.e., PK,K~ and PG,o:, is considered. Material 2: K 2,
The autocorrelation coefficient functions for the moduli
K1 and G1 are assumed to be identical and of the form --~ 16"
Fig. 5. Example 2 - plate with stochastic elasticity
= exp[ ( ,,x ?_(
(24)
Table 2. Statistics for example 2
where Ax and Ay are the differences between the x and y Variable Mean Coefficient
of variation
coordinates of any two points, L = 1 6 i n is the side
dimension of the quarter plate, and ax and ay are K1, ksi 5000 0.15
dimensionless measures of the correlation length in the Gl, ksi 3750 0.15
Kz, ksi variable 0.15
two directions. For the present analysis a x = a y = a is G2, ksi variable 0.15
assumed, which implies an isotropic correlation W, k/in 16.0 0.20
structure.

88 Probabilistic Engineering Mechanics, 1988, Vol. 3, No. 2


Stochastic finite element method: A. Der Kiureghian and J-B. K e
5.0 where the second fl is much larger than the first) the first-
order reliability estimate in equation (8) is sufficiently
4.0
accurate. However, when the two fl's are nearly equal,
multiple fittings of the limit-state surface are necessary to
3.0
improve the accuracy. Using hyperplane fittings at the
two local minimal-distance points and employing well
2.0
known results for the probability content of polyhedral
sets 22, the generalized fl for various values of r are
1.0
computed and are shown in Fig. 6a as a dashed curve.
(The generalized reliability index is defined as fl=
0.0
• - 1 ( 1 - p ) , where 1 - p is the probability content of the
1.0 polyhedral safe set.) It is observed that the contribution of
the second failure mode is only significant near r = 1.
KI
Fig. 7a shows the variation of/3 with respect to the
0.5
cross-correlation coefficient P = PK,K~ = P~,c~, for
identical mean values, i.e., r = 1, and a = ~ . The reliability
~' a~-uB0,0 is seen to decrease as the properties of the two materials
become independent. In Fig. 7b, scaled sensitivity
-0.5
oi/ ~/°/~ measures of fl with respect to the standard deviations are
shown. It is observed that the reliability is most sensitive
(b) "X__..~.~ to the uncertainty in the load and that the sensitivity to
-1.0 i i 1111111 i i fill,, I I , I;Illl i I J lll,H i the uncertainty in the material properties diminishes with
increasing correlation between the property variables.
1.0 This is expected, as with increasing correlation the plate
becomes statically determinate and the stress distribution
0.0 becomes independent of the moduli.
Fig. 8 shows the variation of fl with the correlation
length parameter a for a deterministic load W= 22, with
a# _ 1.0
r = 1 and p = 0, and for both the midpoint and spatial
averaging discretizations of the random property fields.
-2.0 For the selected value of r, the limit-state surface has two
local minimal distance points, which, for certain values of
a, have nearly equal fl values. Thus, the dashed curves in
-3,0 i i 111111 I i ] i iiiill i i i iiiilj i i i iiiiij

10 -' 10 -' I 10 10 '


the figure show the fl value for the single dominant mode,
RATIO OF M E A N S . r whereas the solid curves show the generalized reliability
Fig. 6. Influence o f ratio o f mean properties on (a)
reliability index, (b) sensitivities to means, and (c) 5.0
sensitivities to standard deviations, for a = ~ and p = 0

4.5
Table 3. Failure modes f o r example 2 (a= ~ , p =0)

r=l.00 r = 1.25 ~ 4.0

Variable Mean x* Mean x*


3.5
K1 5000 3801 5000 5448
Gl 3750 3685 3750 3990 (o)
Kz 5000 5708 4000 3369 .3.0 i I ] ] i I J I r I
G2 3750 3887 3000 2375
W 16.00 26.76 16.00 25.71
1,0 --

//G1, G2
0.0
K2 /
case with Gaussian variables, the latter coordinates -I.0 K1/
represent the most likely realizations of the random
variables given failure has occurred. It is observed in "a'--g
a# - 2 . 0
Table 3 that the most likely failure mechanism is
-3.0
fundamentally different for the two cases. That is, for r = 1
the most likely mode of failure is with K1 and Ga below -4.0
their respective means and K 2 and G2 above their (b)
respective means (i.e., weak plate, strong core), whereas --5.0 ' I ; I , I , I , I

with r = 1.25 the situation is reversed. In fact, it turns out 0.0 0.2 0.4 0.6 0.8 1.0
CORRFLATION COEFFICENT, p
that for each value of r there are two local minimal
distance points on the limit-state surface in the standard Fig. 7. Influence o f correlation coefficient pK,K2=
space, which are associated with these two modes of po,o=p on (a) reliability index, (b) sensitivities to
failure. In situations where one mode dominates (i.e., standard deviations, for r = 1 and a =

Probabilistic Engineering Mechanics, 1988, Vol. 3, No. 2 89


Stochastic finite element method: A. Der Kiureghian and J-B. Ke
Table 4. Results for example 2 (a =0.25, p =0)

W=22 W=N(16,3.2)

fl 3.656 3.4957
Pyl 0.000128 0.000236

Variable x* ~ a x* a a --
9~ 9a ~ 9a

K1 6245 0.027 - 1.71 5882 - 0.083 -0.639


Gl 4458 -0.246 - 0.550 4072 - 0.113 -0.118
K2 3383 0.590 - 1.27 4208 0.276 -0.226
G2 3384 0.178 - 0.116 3590 0.076 - 0.020
W 22.0 25.4 - 0.838 - 2.45

5.0 K1 and G1 than the standard deviations of the variables


Spat]c~l A v e r o g e
ldethod K 2 and G2, respectively.

4.54.0 ~ a ~ - " . . . ~. . . . . . . .- " ' [


SUMMARY AND CONCLUSION
A first-order stochastic finite element method for
3.5 reliability analysis of complex structures is described. The
method is based on the first-order reliability approach
/ System #
3.O and is applicable to any limit-state criterion that is
...... Dominont Mode fl
prescribed in terms of random variables. Measures of
2..5 , I , I , , reliability sensitivity with respect to any set of parameters
0.0 1.0 2.0 3.0 are easily computed. The main effort for finite element
CORRELATION LENGTH, a
implementation is in computing the gradient of response
Fig. 8. Influence of correlation length on reliability for variables in terms of basic load, material property or
W = 2 2 , r = l , and p=O geometry variables. Such an implementation for linear
elastic structures under static loads is described. Two
index considering both failure modes of the system. It is alternate methods of random field discretization are
seen that the second mode of failure is significant only for investigated and their respective advantages are
0.5~<a~<2. discussed.
As stated earlier, the true value of fl usually lies in Two example applications are described in detail,
between the results based on the midpoint and the spatial which address such issues as the convergence of the
averaging methods, but usually closer to the latter. Thus, reliability index for the two methods of random field
it appears in Fig. 8 that the selected finite element and discretization, the effect of correlation length on the
random field mesh is adequate for values of the convergence, the influence of correlation length on
correlation length parameter a/> 0.25. It is also observed reliability, reliability sensitivity with respect to
that the reliability tends to increase with increasing distribution parameters, and multiplicity of failure
correlation length of the random property fields. This modes. These examples clearly demonstrate the
applicability and usefulness of the first-order reliability
result, however, is overshadowed if the uncertainty in the
load is also included, which typically dominates the approach in conjunction with the finite element method
reliability with respect to stress limit states. for reliability analysis of complex structures.
The significance of uncertainty in the load can be
observed in Table 4, which lists the values of the reliability
ACKNOWLEDGEMENT
index, failure probability, linearization point, and scaled
sensitivity measures with respect to the means and The research work presented herein was supported by the
standard deviations. The results are for both National Science Foundation Grant No. CEE-8205049,
deterministic and random values of the load and for r = 1, which support is gratefully acknowledged.
p = 0 , and a=0.25. For the processes K1 and G~, the
linearization point listed is for element 32 and the
sensitivities are with respect to the process means and REFERENCES
standard deviations rather than those of the individual
element. It is interesting to note that for the selected 1 Ang, A. H-S. and Tang, W. H. Probability Concepts in
Enoineerin9 Planning and Design - Volume II: Decision, Risk, and
values of the parameters the reliability is more sensitive to Reliability, John Wiley and Sons, New York, NY, 1984
the mean o f K 2 than the mean ofK~. This is attributed to 2 Astill, J., Nosseir, C. J. and Shinozuka, M. Impact Loading on
the random fluctuations in K~ and the cancellation of the Structures with Random Properties, Journal of Structural
sensitivities arising from the various elements of the plate. Mechanics, 1972, 1(1), 63-77
3 Baecher, G. B. and Ingra, T. S. Stochastic F E M in Settlement
Such cancellations do not occur for the sensitivities with Predictions, Journal of the Geotechnical Engineering Division,
respect to the standard deviation of the process, as they April 1981, 107(GT4), 449~,63
are negative for all elements. Hence, the reliability index is 4 Cambou, B. Application of First Order Undertainty Analysis in
more sensitive to the standard deviations of the processes the Finite Element Method in Linear Elasticity in Proceedings,

90 Probabilistic Engineering Mechanics, 1988, Vol. 3, No. 2


Stochastic finite element method: A. Der Kiureghian and J-B. K e

Second International Conference on Applications of Statistics Conference on Structural Mechanics in Reactor Technology,
and Probability in Soil and Structural Engineering, London, Chicago, Illinois, August 1983, M, 199-206
England, 1971 17 Hohenbichler, M. and Rackwitz, R. Non-Normal Dependent
5 Dendrou, B. A. and Houstis, E. N. Uncertainty Finite Element Vectors in Structural Safety, Journal of the Engineering
Dynamic Analysis, Report CSD-TR 271, Department of Civil Mechanics Division, December 1981, 107(EM6), 1227-1238
Engineering, Purdue University, West Lafayette, Indiana, July 18 Igusa, T. and Der Kiureghian, A. Response of Uncertain Systems
1978 to Stochastic Excitation, Journal of Engineering Mechanics, to
6 Der Kiureghian, A. Finite Element Methods in Structural Safety appear
Studies in Proceedings, ASCE Symposium on Structural Safety 19 Lawrence, M. A. A Basis Random Variable Approach to
Studies, Denver, CO, April-May 1985, 40-52 Stochastic Structural Analysis, p. Thesis, submitted in partial
7 Der Kiureghian, A. Finite Element Based Reliability Analysis of fulfillment of the requirements for the degree of Doctor of
Frame Structures in Proceedings, Fourth International Philosophy in Civil Engineering, University of Illinois, Urbana,
Conference on Structural Safety and Reliability, Kobe, Japan, Illinois, 1986
May 1985, I, 395-404 20 Liu, P-L. and Der Kiureghian, A. Optimization Algorithms for
8 Der Kiureghian, A. Multivariate Distribution Models for Structural Reliability Analysis, Report No. UCB/SESM-86/09,
Structural Reliability in Transactions, International Conference Department of Civil Engineering, Division of Structural
on Structural Mechanics in Reactor Technology, Lausanne, Engineering and Structural Mechanics, University of California,
Switzerland, August 1987 Berkeley, California, July 1986
9 Der Kiureghian, A., Lin, H.-Z. and Hwang, S.-J. Second Order 21 Liu, W. K., Belytschko, T. and Mani, A. Probabilistic Finite
Reliability Approximations, Journal of Engineering Mechanics, Elements for Nonlinear Structural Dynamics, Computer
to appear Methods in Applied Mechanics and Engineering, 1986, (56), 61-86
10 Der Kiureghian, A. and Liu, P-L. Structural Reliability Under 22 Madsen, H. O., Krenk, S. and Lind, N. C. Methods of Structural
Incomplete Probability Information, Journal of Engineering Safety, Prentice-Hall Inc., Englewood Cliffs, NJ, 1986
Mechanics, January 1986, 112(1), 85 104 23 Nakagiri, S. and Hisada, T. A Note on Stochastic Finite Element
11 Der Kiureghian, A. and Taylor, R. L. Numerical Methods in Method (Part 6) - An Applic "ion in Problems of Uncertain
Structural Reliability in Proceedings, Fourth International Elastic Foundation, Seisan- nkyu, Institute of Industrial
Conference on Applications of Statistics and Probability in Soil Science, University of Tokyc Fokyo, Japan, January 1983,
and Structural Engineering, Florence, Italy, June 1983, 769-784 35(1), 20-23
12 Handa, K. and Anderson, K. Application of Finite Element 24 Shinozuka, M. and Dasgupta, G. Stochastic Finite Element
Methods in the Statistical Analysis of Structures in Proceedings, Methods in Dynamics in Proceedings, ASCE-EMD Specialty
Third International Conference on Structural Safety and Conference on Dynamic Response of Structure, Los Angeles,
Reliability, Trondheim, Norway, June 1981, 409-417 March 1986, 44-54
13 Hisada, T. and Nakagiri, S. Stochastic Finite Element Method 25 Shinozuka, M. and Nomoto, T. Response Variability Due to
Developed for Structural Safety and Reliability in Proceedings, Spatial Randomness, Report No. CU-CEEM-MS-80-1,
Third International Conference on Structural Safety and Department of Civil Engineering and Engineering Mechanics,
Reliability, Trondheim, Norway, June 1981, 395-408 Columbia University, August 1980
14 Hisada, T. and Nakagiri, S. Role of the Stochastic Finite Element 26 Vanmarcke, E. Random Fields: Analysis and Synthesis, MIT
Method in Structural Safety and Reliability in Proceedings, Press, Cambridge, Mass., 1983
Fourth International Conference on Structural Safety and 27 Vanmarcke, E. and Grigoriu, M. Stochastic Finite Element
Reliability, Kobe, Japan, May 1985 Analysis of Simple Beams, Journal of Engineering Mechanics,
15 Hisada, T., Nakagiri, S. and Mashimo, M. A Note on Stochastic October 1983, 109(5), 1203-1214
Finite Element Method (Part 10) - On Dimensional Invariance 28 Veneziano, D., Casciati, F. and Faravelli, L. Method of Seismic
of Advanced First-Order Second-Moment Reliability Index in Fragility for Complicated Systems in Proceedings,Second CSNI
Analysis of Continuum, Seisan-Kenkyu, Institute of Industrial Speciality Meeting on Probabilistic Methods in Seismic Risk
Science, University of Tokyo, Tokyo, Japan, 1985, 37(3), 111 Assessment for Nuclear Power Plants, Livermore, California,
114 May 1983
16 Hisada, T., Nakagiri, S. and Nagasaki, T. Stochastic Finite 29 Arora, J. S. and Huag, E. J. Methods of Design Sensitivity
Element Analysis of Uncertain Intrinsic Stresses Caused by Analysis in Structural Optimization, AIAA Journal, 1979, 17,
Structural Misfits in Transactions, Seventh International 970-974

Probabilistic Engineering Mechanics, 1988, Vol. 3, No. 2 91

You might also like