Professional Documents
Culture Documents
Yi-Kwei WEN a n d H o w a r d T. P E A R C E
Professor and Research Assistant, respectively, University of lllinois at Urbana, IL 61801, USA
A large number of loadings on structures, particularly nuclear structures, are random and dynamic in nature and may be
generated from common or related sources (e.g., storms, earthquakes, and accidents). The combination of load effects which
may be nonlinear, dynamic and statistically dependent is therefore an important issue in the assessment of safety and
performance of a nuclear structure over its lifetime. This paper summarizes the latest developments in this area with emphasis
on the versatile method of analysis based on a consideration of load coincidence. The validity and accuracy of this method is
established by comparison with other methods and Monte Carlo simulations. Parametric studies are carried out and effects of
dynamic response and load dependencies are examined. The accuracies of some of load combination rules are also studied and
ranges of validity of each rule are identified such that some risk-consistency can be achieved if such rules are to be used.
0029-5493/83/$03.00 © 1983 N o r t h - H o l l a n d
180 Y.K. Wen, H.T. Pearce / Combined dynamic effects of correlated load processes
Sl(t) L ~
(2)
in which/~x, °x and r are mean response, r.m.s, response
o, x Occurrence Time Point Process and apparent frequency, conditional on the excitation
sz( t ) intensity Y = y . Note that the excitation intensity varies
from pulse to pulse and is treated as a random variable,
e.g., the maximum mean wind velocity in a storm, the
maximum ground acceleration in an earthquake.
3.1. Linear combination accuracy depends mainly on those of eqs. (2) and (4)• It
is found that the errors introduced by using the simple
For linear combination of two random load effects, Poisson approximation in eqs. (2) and (4) are quite
the lifetime combined maximum R m, has the following insignificant because of the usually much larger macro-
distribution function, scale uncertainties [5], and in the study of structural
safety only high threshold levels are of concern. Eq. (5)
FR.,( r ) = exp[ -A,TF~,( r) - ~2TF~2( r ) - hl2TF~q2( r) ] generally gives good, conservative results.
(3) A method based on the upcrossing rate (U.R.) has
also been proposed for Poisson pulse processes by Lar-
in which F~., = 1 - Fx. The first two terms are contribu- rabee and Cornell [2] and for the combination of I.C.P.'s
tions from individual process (eqs. (1) and (2)). The by Winterstein [6] and Shinozuka and Tan [7]. The
third term is the contribution from the coincidence of accuracies of these two methods have been compared
events, and A]2 ~-'/~lX2(~d, +~d2) , the mean rate of for Poisson rectangular and triangular pulse processes
coincidence. For rectangular pulse processes, Xl2 = X l for a wide range of load parameters and no appreciable
+ X2; for I.C.P.'s differences were found. For combination of I.C.P.'s, one
F:,2(r) can show that the two methods coincide as r--* 0o.
However, for finite r, long T and long occurrence or
coincidence durations (e.g., hours instead of seconds),
the U.R. method has a tendency to give unduly con-
servative results. This is due to the fact that in the L C .
method, through eqs. (2) and (4), failure (upcrossings) is
X fr, ( Y, )fv2( Y2)d y,d y2, (4)
counted at most once in each occurrence (pulse), where
in which ul2, #x,2 and Ox,2 are respectively the apparent in a U.R. method they may be counted a large n u m b e r
frequency, mean response and r.m.s, response of the of times even at very high threshold levels• The dif-
combined process given the coincidence and the excita- ference is specially important when the variation of
tion intensities being Y1 =Yl and Y2 =Y2. Eqs. (2) and
(4) can be modified to include the within-occurrence
intensity variation with time. /~d,~= the mean duration
Threshold r
of the coincidence of two loads• = I~a,t*d2/(l~d, + /~d2)" 3 4 5 6
For combination of multi-(n) load effects,
Nonlinear combination of load effects arises when Dynamic Linear ~ Linear (4)
(macro- and Nonlinear (5)
the limit state function of a linear structure is nonlinear
micro-scale) Nonlinear - - Nonlinear (6)
(e.g., failure of a beam-column) and where the structure
response becomes nonlinear (e.g., plastic collapse of a '~ Scale of fluctuation with time with respect to the fundamen-
structure frame). In either case, the performance func- tal natural period of the structure.
tion (or failure surface) is nonlinear, and the reliability
problem in general cannot be easily formulated as the
resistance exceeded by the maximum combined load
effects as in the linear combination. analysis. When there are a large number of failure
A method which is analogous to the First Order modes, such as the static plastic collapse problem, one
approximation for the time invariant problems by can use a probability network (PNET) method by Ang
Rackwitz [9] is to linearize the failure surface at the and Ma [10] or a bounding technique to find approxi-
point where the mean crossing rate out of the tangent mate solutions or bound to P;] and P,j~, etc. For
plane by the vector load effect process reaches a sta- category (5), since the system is still linear, the re-
tionary value as indicated in Pearce and Wen [5]. A sponses can be modeled, at least approximately, as the
more practical and computationally more efficient ap- components of a Gaussian vector process, methods by
proach is to generalize the load coincidence method, Veneziano et al. [11] can be used for this purpose. When
since the nonlinearity of the problem does not affect the the structural response is both dynamic and nonlinear
time domain behavior of the processes, hence the coinci- (and inelastic for most structures), category (6), meth-
dence analysis remains valid. The probability that a ods based on a linearization of the structural system,
nonlinear limit-state E being exceeded in (0, t) is e.g., Baber and Wen [12] appear to be most tractable. In
short, such a formulation utilizing the most efficient
Fig. 4. Accuracy of load coincidence method for nonlinear 4.1.1. Dependence between intensity and duration
combination. For example, storms with longer duration usually
have higher intensity, therefore, these two parameters
may be correlated within each load. The occurrence
times are assumed to be statistically independent, i.e.,
modeled by Poisson processes, therefore the occurrence
safe domain by the L.C. method is shown in fig. 4. and coincidence rates in eq. (5) remain unchanged.
Similar results were found in a study of lifetime plastic However, a longer duration causes a higher coincidence
collapse probability in Wen [13] where the safe domain rate and at the same time a higher probability of maxi-
is a polygon. m u m response since the load intensity is more likely to
be higher. Therefore, ;kij and Fx," cannot be separated,
also terms F~(, and F~,j will be affected by the correla-
4. Modeling and effects of load dependencies tion since the maximum dynamic response increases
with duration. These interdependencies of the load
The L.C. method has been generalized to include the parameters can be easily included in the L.C. method.
dependencies between load parameters into considera- For example, eq. (2) is replaced by a double integration
tion for pulse processes [3,4]. The same approach is used over the intensity and duration. The coincidence term in
here for the combination of dependent I.C.P., i.e., the eq. (3) is replaced by
macro-time parameters (occurrence time, duration and
intensity) are modeled as correlated random variables.
The within-occurrence r.m.s, intensity is assumed to be ~klX2~lldl [ l -fo~L(rlYl)Mtfr,(yl)dyl]
proportional to the intensity random variable (e.g., Y~in
eq. (4)). Examples are the r.m.s, fluctuation of wind
velocity being approximately proportional to the mean
+XlX2Fd2(l - fo~L(rly2)M2fr2(y2)dy2], (7)
wind in a storm and the r.m.s, acceleration to the
in which
maximum acceleration in an earthquake. The micro-
scale variabilities, therefore, are also correlated in an
indirect way. It is a reasonable model for the real L ( r l y , ) = fo g S(rlyl,y2)fr2(y2)dy2, (8)
physical process. Details of the model and analysis
procedures can be found in ref. [5]. The features of the where S(r I Yl, Y2) = probability distribution of the max-
184 Y.K. Wen, H.T. Pearce / Combined dynamic effects of correlated load processes
°o, ( Yl - t~Y, ]
m 1 = 1 + PD,,Y~-- - - " (9) 0.999
or, \ oy~ !
//
,/
p = 0,99
• • •, etc.
The duration D and excitation intensity Y are as-
0.995
sumed to be bivariate normal and po, v = correlation
coefficient. The results of comparison with the indepen-
0.99
dent case (PD.Y = 0) are shown in figs. 5 to 7. It is seen
that this correlation causes a higher maximum com- 0,98
bined response. However, the difference is small enough
that it would not significantly affect any design deci-
0.95
sion.
0.9
4.1.2. Occurrence dependence (clustering)
A c o m m o n phenomenon of occurrence dependence
0.8
is clustering; examples are main and after shocks of
0.7 / w : .0
earthquake, a large number of tornadoes spawned by a
single storm and a sequence of S R V loads, etc. In ref.
0.6
~,////// ,/ V. : 000137
O.S
1/ - - INOEPENDENT
[3], such dependence in occurrence time for static re- 0.4 !i :iill .........
0.3
0.2
0.1
0.05
0.01
0.9!
sponse is modeled by a point process with clusters of
0.98 random length; within each cluster the occurrence fol-
0.95
// lows a Poisson law. Comparisons of results with inde-
pendent occurrence time indicated that the overall ef-
fect of clustering can be described as only moderate and
it causes lower lifetime combined maximum• Since the
i,J=3.0 within-occurrence dynamic response is not affected by
X =0.5
0.8 I~ : 0 . 0 0 0 3 4 2 the occurrence time. The conclusion would hold for the
- - - - INDEPENDENT
dynamic case as well.
0.7
0.6 / .
. .
.
.
-
.
.
. .
.
. .
WITHIN
BETWEEN
LOAD dNTENS[TY
LOAD I~TENSITY
001
/, rence.
This dependence between load pulses in a single
process is achieved by imbedding a Gauss Markov
3.0 4.0 50
sequence in the renewal pulse process. The value of the
Fig. 5. Effect of dependencies in intensity and duration on mean component of one pulse is dependent on the value
probability of combined dynamic response (h/~ d = 0.00017). of the mean of the previous pulse as follows:
Y.K. Wen, H.T. Pearce / Combined dynamic effects of correlated load processes 185
//i0!0.99
0.999
(see eq. (2)), etc. The joint density fy2y,(Y2' Y l) is derived
from the product of a conditional density and a margi-
nal density. The conditional density is obtained from
the relationship given in eq. (10). Considering the Pois-
son character of the occurrence of load pulses, the
0.995
//' probability the maximum of a single process not exceed-
0.99
ing a threshold r in T years is approximately given by
[3]:
0.98
0.6
m u m of the coincidence part can be derived in a similar
0.5 manner. The results are then combined to obtain the
0.4 // ~ = 0.0027 overall lifetime maximum probabilty [5].
0.3
For correlation coefficients less than 0.9 there is no
0.2
0.1 visible difference between the dependent and indepen-
0.05 dent cases. As the coefficient goes from 0.9 towards 1.0
0.01 the difference increases at low thresholds but the inde-
pendent case is always conservative. For perfect correla-
4,0 5,0 6.0 7.0
tion (p = 1.0) all the pulses in the process have the same
Fig. 7. Effect of dependencies in intensity and duration on static component and the exceedance probability is
probability of combined dynamic response (~/1 d = 0.022). greatly reduced at low thresholds (figs. 5 to 7).
X = 1 /YR /
FR,2(r) = exp[ - XtzTF?2 (r)]. (19) x~ = 4 /,,R /
I~, = 20.0 s
The maxima due to the noncoincident and the coinci- % : I0,0 s
dent parts of the combined process are then assumed
F's : 0.7 / /
independent to compute the overall maximum.
The results indicate that when I.C.P.'s are correlated
in intensity both within load and between loads, two /, J
//
/
2 3 4 5 6 2 3 4 5
I ~ ~ " , ~~% - 1 I 1
163 _
IC J l J I ;6 5 l t I I
3 4 5 6 7 2 3 4 5 6
I L ~\ ~ , \ f i i "~ r-,,,, i [ r
\ N "\ LC • LC
\\
'°-"\ \\ ~ \ -----SRSS
,d'i
\ ~,
\\ -.-~Rss
~. .,. .T~
"\ • 0.00017
, \\ .o.,o \ , .50
- ,~,. \
163 --
"~ - I I I I
4 5 6 7 3 4 5 6 7
l\~ ~1 ~,, I I
\\ \ ~,,,.oooo,7
-, wo • 3.0
10-2 _ ,6' - \~ \
,6
•
-
'\ \\
\,\ ,
,o-,- 5::i!
X/= = 0.022
10-4 wo = 3.0 .>2 \\ ~.
io~ , , , s\~\~
\\ I~5 jJ I \~\\
' \\,,\, \,
can show that the coincidence rate is Turkstra [15] suggested a simple method for combin-
ing loads, i.e.,
Xes = XeXs ( ~e _{_~ls) .+_X e p s ~e (20)
/~e + P~O ' Max[ S~(t) + S z ( t ) ] ~ M a x [ R l + $2, R 2 + Sl] , (22)
in which h e, ),~ = occurrence rate of earthquake and in which S, are the arbitrary-point-in-time values. This
independent SRV load. /x¢, /~s and #0 are the mean procedure has been used in recent code studies. See,
values of earthquake duration, SRV load duration and e.g., Ellingwood et al. [16]. In general, the accuracy of
SRV lag time. The first term is the coincidence due to the rule depends on load and structural system parame-
the independent SRV and the second term is that tri- ters as well as probability level. Typical comparisons of
ggered by earthquakes, #e, /~s and ~0 are of the same results are shown in fig. 9, for systems with various
order of magnitude, i.e., seconds. Therefore, the second natural frequencies and damping factor ~"= 0.05. The
term usually dominates. For example, for h e = l / y r , load intensities are modeled by gamma variates with
~s = 4 / y r , Ps = 0.1, /% = 20 s, /to = 10 s and /~s = 5 s, /%. = 10 and or. = 0.3. The r.m.s, value of the dynamic
Xe~ = 0.066/yr. If the dependence is not considered, oscillation is proportional to the intensity. Because most
Xes = 2 × 1 0 - 6 / y r . stochastic dynamic loadings are infrequent and brief,
The conditional probability of a given threshold only small values of X/% are considered. T = 50 yrs. It is
being exceeded by the combined response given coinci- seen that the SRSS rule is generally very conservative,
dence can be evaluated analytically regarding the SRV perhaps too much so when X~d is very small (loads
effect as time-dependent threshold. The probabilities of which are extremely infrequent and brief). But, in gen-
lifetime maximum combined earthquake and SRV re- eral, it gives comparatively better results than in the
sponse is compared in fig. 8. It is seen that the probabil- static load effect combination [13]. The effect of chang-
ity of exceedance increases significantly due to occur- ing natural frequency can be seen from the results. As
rence dependence. structural stiffness increases, SRSS is less conservative
When both intensity and occurrence dependence ex- and may become unconservative at low risk level, say
ist between loads, the effects are multiplicative which 10 4. The TR, on the other hand, consistently gives a
can cause a much higher probability of exceedance at lower bound estimate which is good only when ~'~d is
the high threshold levels. extremely small and does not seem sensitive to changes
in structural system parameters.
The conclusions given in the foregoing are valid for
5. Error of load combination rules combination of load or load effect processes which are
statistically independent of one another. For dependent
Simple rules for load combination are useful and loads, as shown in the foregoing, inter-(between) load
perhaps necessary in routine safety checking and code dependencies may significantly increase the probability
formulation. However, since such rules are mostly judg- of threshold level being exceeded by the combined load.
ment-based, structures designed according to such pro- Since all load combination rules do not consider effects
cedures may not have the intended safety level. It is of such dependencies, they would give results more on
therefore important to examine the accuracies of such the unconservative side when such dependencies exist.
rules. This is done by comparing the probability of the
combined load effect implied by such rules with analyti-
cal results. For the combination of dynamic responses, 6. Conclusions
the rules examined here are the Square Root of the Sum
of the Squares (SRSS) and Turkstra's Rule (TR). The conclusions reached in this study are sum-
The popular SRSS rule for combination of modal marized in the following
responses in structural dynamics has been suggested for (1) The pulse process and intermittent continuous
combination of loadings on nuclear structures, e.g., by process are efficient models for the macro- and micro-
Mattu [14]. The maximum combined load effect is scale variabilities of time varying loads and load effects.
calculated according to They capture the essential features of most loadings and
structural responses including dependencies and yet al-
Max[Sl(t ) + S2(t)] = ~+ R2 , (21) low tractable analytical treatments of the probabilistic
combination problem.
in which S i l t ) are load effect processes and R~ are their (2) The load coincidence method is well suited for
maximum value over a time period T. probabilistic combination of dynamic load effects, in-
Y.K. Wen, 11. T. Pearce / Combined dynamic effects of correlated load processes 189
cluding nonlinear c o m b i n a t i o n a n d load effects which [5] H.T. Pearce and Y.K. Wen, A method for the combination
m a y be correlated in occurrence time, d u r a t i o n and of time varying stochastic load effects, Structural Research
intensity. Uses of m e a n d u r a t i o n a n d Poisson crossing Series, Civil Engineering Studies, University of Illinois,
a s s u m p t i o n are f o u n d to be generally sufficiently accu- Urbana, Illinois (June 1983).
[6] S.R. Winterstein, Combined dynamic response, extreme
rate a p p r o x i m a t i o n as far as the lifetime probability is
and fatigue damage, Department of Civil Engineering,
concerned.
M.I.T., Report No. R80-46, Order No. 688 (December
(3) The effect of load dependencies or c o m b i n e d 1980).
d y n a m i c response in general follow the trends observed [7] M. Shinozuka and R. Tan, Probabilistic load combination
for c o m b i n a t i o n of static load effects. The types of and crossing rates, Proc. Symposium on Probabilistic
correlation which need most scrutiny are the between Methods in Structural Engineering, ASCE, St. Louis, Mis-
load intensity dependence, between load occurrence souri, October 26-27, 1981.
clustering a n d to a lesser extent, duration-intensity cor- [8] C.A. Cornell and S. Winterstein, Load combination analy-
relation. A c o m b i n a t i o n of these factors could have very sis for fatigue and fracture, Paper Presented at ASCE
serious consequences. National Meeting, New Orleans, LA (October 1982).
(4) C u r r e n t rules for c o m b i n a t i o n of d y n a m i c re- [9] R. Rackwitz, Principles and methods for a practical prob-
abilistic approach to structural safety, Subcommittee for
sponses are not satisfactory in terms of giving consistent
First-Order Reliability Concepts for Design Codes of the
results for all probability levels a n d c o m b i n a t i o n s of
Joint CEB-CECM-CIB-FIP-IABSE Committee on Struct-
load type a n d structural parameters. A l t h o u g h the re- ural Safety, CEB Bulletin, 2 (112) (July 1976).
sults are better t h a n in the static load effects case. The [10] A.H-S. Ang and H.F. Ma, On the reliability of structural
inaccuracy is primarily due to their failure to properly systems, Proc. 3rd International Conference on Structural
account for the coincidence effect which generally plays Safety and Reliability, Trondheim, Norway, June 23-25,
a d o m i n a n t role at the low risk (or high response) levels. 1981.
[11] D. Veneziano, M. Grigoriu and C.A. Cornell, Vector-pro-
cess models for system reliability, J. Engineering Mecha-
nics Division, Proc. 2, ASCE, 103 (EM3) (June 1977) pp.
Acknowledgment
441-460.
[12] T.T. Baber and Y.K. Wen, Random vibration of hysteretic
This study is supported by N a t i o n a l Science F o u n d a - degrading systems, J. Engineering Mechanics Divisions,
tion u n d e r G r a n t N S F C E E 82-07590. Proc. ASCE, (EM6) (December 1981).
[13] Y.K. Wen, Method for reliability of structures under
multiple time varying loads, Nucl. Engrg. Des. 60 (Oc-
References tober 1980).
[14] R.K. Mattu, Methodology for combining dynamics re-
[1] Y.K. Wen, Statistical combination of extreme loads, J. sponses, NUREG-0484, Rev. 1, Office NRR, U.S. NRC
Structural Division, ASCE, 103 (ST5) (May 1977). (May 1980).
[2] R.D. Larrabee and C.A. Cornell, Upcrossing rate solution [15] C.J. Turkstra, Theory of structural design decision, Solid
for load combinations, J. Structural Division, ASCE, 105 Mechanics Study, No. 2, University of Waterloo, Waterloo,
(ST1) (January 1979) pp. 125-132. Ontario, Canada (1970).
[3] Y.K. Wen and H.T. Pearce, Stochastic models for depen- [16] B. Ellingwood, T.V. Galambos, J.G. MaGregor and C.A.
dent load processes, Structural Research Series No. 489, Cornell, Development of a probability based load criterion
UILUENG-81-2002, Civil Engineering studies, University for American National Standard A58, National Bureau of
of Illinois, Urbana, Illinois (March 1981). Standards, NBS Special Publication 577, Washington, D.C.
[4] Y.K. Wen, A clustering model for correlated load (June 1980).
processes, J. Structural Division, Proc. of ASCE, 107 (ST5)
(May 1981) pp. 965-983.