Professional Documents
Culture Documents
Ahstraet-An algorithm for the calculation of structural reliability under combined loading is formulated. Loads or
any other actions upon structuresare modelkd as independentrandom sequences.The rekvant limit state criterion
is pointwise approximated by a tangent hyperplane. The combination of time-variant actions then reduces to the
calculation of the maximum of a sum of random variables which is facilitated through proper, discrete ap-
proximation of extreme value and other non-normal distribution functions by normal distributions.The iteration
algorithm searches for an approximation point on the limit state criterion where the probabiity content of the
failure domain limited by the tangent hyperpkne reaches its maximum. Any type of continuouslimit state criterion
and any distribution type for the loads can be dealt with. The method is illustrated for a section of a wall without
tensile strengthloaded by a bendingmoment and a normal force.
zk = maxT{maxTk{. . . . .max,,_,
Imax,_,&:, . Xl+ PA-, . X,-,}
+.....‘+g;+! 'xk+,}+g;'xk) (5)
the number of tee-v~ant actions exceeds, say 3, the easily calculated means and variances. Also, given a
computation of eqn (6) becomes an extremely laborious point in the distribution of a sum, it can be decomposed
task. into two vahres of two other normal variables where
However, approximate results can be obtained by each value is obtained by weighing with the relative
transforming non-normal distribution functions into contribution of its variance to the variance of the orig-
normal ~s~bution functions. in general, the exact inal sum.
t~sfo~tion z = Cp-‘[FQ)] where F(x) is an arbitrary
distribution function and @-‘[.I is the inverse standard ~TmrlTmtATloN~
normal integral, causes prohibitive computational For convenience, the formulation is given in the orig-
d~cuities. The authors, therefore, suggested in [7J an inal space of basic variables in which the eqns (7) are to
approximate mappingsuch that in a given point both the be applied. Let
original and the approximating normal distribution
functions and their densities have the same vahres, &XI, Xi?,. . .‘ X”) = c (8)
respectively. This is equivalent to a linear expansion of
the exact ~ansformation. It yields the following be the state function, describing the st~ctur~ behavior
parameters for the approximatingnormal distribution, such that for C > 0 safe states are defined whereas C < 0
corresponds to failure. Expand this expression in linear
&’= x* - o’ * @-‘IF( (7al Taylor series which should equal zero, if the limit state
criterion is fulfilled:
gtx:,
x: ,..‘, xX)+f:gf.(x*-x2)=0.
1-I
(9)
where x* is the approximation point, Ff.) and ft.1 the
~s~bution function and density function of the non- As mentioned, the problem is to find the point
normal distribution, respectively, and &) the standard cx:, x: , . . .,x3 where 6 becomes minimal.The deriva-
normal density. Clearly, this approximationmay become tives g; may first be evaluated at some arbitrary point
more and more inaccurate if the original distribution (x1”, xz(O), * . ., xi’)& According to the first order reiiabii-
becomes increasingly skewed. The same applies for dis- ity theory the limit state criterion is only f&lied if (see
tributions of the type given in eqn (2) if the occurrence Refs. [I], (21or Pff, the point lies on the failure surface.
probability p becomes small. In this case one might Therefore, we set:
approximate by a normal distribution having the same
median and same exceedance probability at the ap- *JO’
, _
-- fixi - ai * p * irxi mt
proximation point x*. Then, it is, e.g. from eqn (2)
for variables with re~tition number ri = I and where the
1-f3w” sensitivity factors oi are given by?
fit = r; = Jy-t
[
I_
P I (7cI
(11)
and
X*-&L’
g’=
cb-'IF(x*)I and the reliability index @ is chosen to fulfill the limit
state criteri0n.S
wherein tbe notations are as before and F*(.) is the Assume for the moment that ail basic variables are
inverse of eqn (2) containing also the case n = p = 1. It normally distributed. To take account of time-variant
appears that there is a whole not yet fully investigated variables one has in starting from the variable X. with
hierarchy of pointwiseapproximationprocedures (see [7] mean ccx, .and standard deviation ox. to form a new
for fu~her results), which admittedly are only meaning- variable 2, related to X,, by:
ful as long as the first order reliability concept used
herein is considered a feasible method. On the other & =g;*x. VW
hand the mathemat~caiand numerical problems might
become hopelessly ~ompiicated-for non-continuous with mean
distribution functions even impossible-when using the &?”= g: * PX” Wb)
exact transformation within this concept. Then, ap-
plication of more involved methods, e.g. the results on and standard deviation
sums of random processes are more in order which
should be available also for more complex stochastic oz. = [WI * OX”11’“. mcf
models but, at present, are limited to only a few eiemen-
tary shapes of the limit state criterion. This is, however, Then, the extreme-value distribution F- ,._,(z’~‘)
no restriction for the algorithm to be described which can be generated from the distribution of Z” =
subsequently. It simply makes use of the fact, that sums g: - X,, using tqn (2). Next, this dis~bution must be
of normally distributed variables are again normal with approx~ated by a normal distribution with mean hi,
and standard deviation 05~ in the pre-selected point
*It can be shown that these sensitivity factors are related to (zn”’ = gl: +x,“$ The new variable Z, now can be
the direction cosines of the approximating hypet@ane. combined with the variable X,-, with distribution
Sit can be shown that in fact. the initial approximation point @(x,-t; c4x,_,,ux._,) according to
need not necessarily lie on the failure surface. This assumption,
however. facilitates greatly the following presentation. Zn-l = Z” +&L.* . x,-t Wa)
492 R. RACKWITZ
and B. RESSLER
Similarly, further decomposition can be made using eqns Consider a joint below a supporting structural element
(16)-(V) as recurrence relations (substitute the indices k as shown in Fig. 4. The joint is loaded as indicated in Fig.
and k + 1 by k + 1 and (k + 1)+ 1, respectively) up to the 4 and has no tensile strength. Then, its limit state condi-
variable X.. Then, the iteration may start again with the tion can be written as:
computation of new derivatives g; at the new point
{XI”‘,x:“, . . ., X”“3.
The generalization of the expressions (8)-(17) for any
variable and interaction loop is straight-forward and is
x3 1
given in the flow dii (Pig. 3). The flow dii
also generalized to arbitrary distributional assumptions
is
forx,+x,%
for the basic variables which simply consists in substi-
tuting JLX,and OX, by pii and oh, in eqns (8)-(17) by
using the expressions (7). If the limit state criterion is at
leaat lo&y continuous and monotonic, the iteration
process converges at least towards a local minimum in p
as given by eqn (15) which is the safety index of the
problem. for
1,
6 m-(2
x3 I
Structural reliability under combined random load sequences 493
I START I
I’
Zl = z1- go
initial approxima-
tion point Z1-p6
B--x
e.g. xi = c,, i=l...n
s
I
calculation of yee
i=i+l
go = g(x, . . X”)
i>n?
i=n+l
I no
4-h
@,=--
6
I
xi = kr
- a,, * Bi . 4,
treme-value distribution of
no
normal distribution
a,, = 9; OYUZ,
’ I 02,
=z,+t= uz,.,
4 output:
[-G-I
x:, /I
SUMMARY
A technique to compute the reliability of structures
under combined loading is presented. It is assumed that
loads or other actions can be modelled by independent.
stationary random sequences. The procedure may be
used directly in probabilistic design or to check more
elementary approaches to load combination.
Jomt
Acknowledgements-The authors take pleasure in acknowledging
the intense discussions with E. Paloheimo whose pilot results
initiated this study. They are also grateful to 0. Ditlevsen and K.
Skov for valuable comments received while the first author was
Fig. 4. Design situation of example. on leave at the Technical University of Denmark. Support for
this research was provided by the Deutsche Forschungsgemein-
schaft.
described in the above are given: REFEBENCl?S
1. First Order Reliability Concepts for Design Codes. Joint
Committee on Structural Safety. CEB Bull. I I?. Paris (1976).
Table 2. 2. Common Unified Rules for Different Types of Construction
and Material. In the Int. System of Unified Standard Codes of
r2 h P3 B p,=l-da)
Practice for Structures, Vol. I. CEB Bull. 116. Paris (1976).
3. J. Ferry Borges and M. Castanheta, Srnrisricol Dejnition and
I 1 I 2.78 2.72. IO-’
Combination of Lands, pp. 43-62. AC1 Publication SP 31.
IO loo 1 2.17 1.50. 10-z
Detroit (1971).
10 1.000 1 1.95 2.53 . IO-’
4. E. Paloheimo, A Generalisation of the Weighted Froctiles
10 1o.ooo I 1.76 3.89. lO-2
Method. Memorandum presented to the Joint Committee on
10 100.000 I 1.59 5.60~lo-2
Structural Safety, Wiesbaden (19751.
100 1.000 I 1.95 2.55 ’ 10-z
5. 0. Ditlevsen, A Collection of Notes Concerning Structural
100 ItXl.00 I 1.59 5.63. IO-*
Reliability in Second International Workshop of Definition.
100 100.000 lo-’ 2.17 1.51 10-2
Classification and Evolution of Code Formats, Mexico City.
loo 100.000 10+ 2.80 2.58. IO-’
Jan. 1976, Dialog 2-76, Danmarks Ingenicarakademi, Lyngby
(1976).
6. C. Turkstra, Probabilistic design formats. In Theory of Struc-
rum1 Design Decisions. Sol. Mech. Div., University of Water-
Sufficient convergence has been reached after 3-5 loo, Studies Series No. 2. Waterloo (1970).
iteration loops. Checking with Monte-Carlo studies 7. R. Rackwitz and B. Fiessler, Note on Discrcre Safety Checking
suggests that the calculated failure probabilities in the when Using Non-normal Stochastic Models for Basic Vari-
last column are at least as accurate as indicated in the ables. Berichte zur Sicherheitstheorie der Bauwerke, Lab. fiir
table. Konstr. Ingb.. Milnchen. H. 14 (1976).